content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Robots have been designed to interact with unknown environments and act on behalf of humans to minimize the risk of accidents or injuries. Thanks to their rapid deployment and relatively low cost, ground robots as well as Unmanned Aerial Vehicles (UAVs) have recently emerged as alternatives to address emergency and mission-critical scenarios~\cite{SARDO}~\cite{Public_safety}.
Such use-cases drive the evolution of simple remote-controlled robots into moving platforms equipped with dedicated operating systems, advanced computing capabilities and multiple communication modules, to support autonomous navigation and robot control tasks, which can be also aided by Artificial Intelligence (AI) based solutions to perform more accurate decisions thanks to real-time multi-sensor data streams.
However, a trade-off arises when pursuing higher degrees of robot autonomy. Moving robots require sufficient energy to operate in unknown environments throughout their mission. While mobility still represents the major cause of battery consumption~\cite{Swanborn2020}, computing and processing of multi-sensor information are increasing their impact on available energy resources. At present, most of the real-world learning problems faced by robots can only be tackled by feeding dense data directly to more capable infrastructure and dedicated computing units~\cite{HIRO_NET}.
Simultaneous localization and mapping (SLAM) and object recognition are only some example of robot applications that may be needed during rescue operations and deployment of first-aid support~\cite{Multi_Robot_SLAM}.
Most of the modern solutions require the setup of dedicated local networks, e.g., WiFi-based, which limits the applicability of robot technology to only indoor environments~\cite{Past_Present_SLAM}.
In this context, the next generation of mobile networks (5G) is envisioned as a key-enabler to provide the outdoor wireless communication and enhance the autonomy of robotic applications~\cite{5GvsLTE}. The ubiquitous connectivity, high bandwidth and the low latency access provided by the modern mobile technologies, together with the possibility to seamlessly exploit multi-access edge computing platforms (MEC) deployed at the edge of the networks, and cloud computing to host processing and analytic services as virtual or container-based instances, would enable unprecedent levels of flexibility in the deployment of robotic applications which may efficiently off-load computing tasks to the edge and alleviate their energy consumption~\cite{Mission_Critical_Edge}\cite{E2E_Reliability}.
In the 5G context, an orchestrating entity acts as a mediator in charge of guaranteeing the most efficient use of the infrastructure resources, while pursuing the satisfaction of heterogeneous communication requirements. Similarly, an orchestration entity \xl{can be adopted in the robot domain to control} the multitude of operational modules and sensors installed on the on-board robot platform, e.g., to optimize energy consumption and resource utilization.
\xl{However, in the state-of-the-art both orchestration entities work independently, with little to no awareness of each other during the operational phases. On the one hand, the robot domain assumes ubiquitous and unlimited resource availability, both from the networking and cloud resource perspective, which leads to performance drops when dealing with wireless and virtualized environments, especially in heterogeneous multi-robot scenarios. On the other hand, the 5G domain operates without knowledge of the actual resource requirements from the robot domain, mainly by over-provisioning the resources. This introduces inefficient networking resource usage and, perhaps more importantly, waste of energy in the robots, making it hard for such energy-constrained devices to fulfill their tasks.}
\xl{To fill this gap, in this work we propose} to combine the orchestration logic from the network infrastructure and the robot domains,
enabling information exchange between the robots and the hosting infrastructure. In this way, not only energy-aware decisions performed in the robotic domain may be tuned according to real-time infrastructure conditions, but also infrastructure resources can be re-configured to meet robot communication and energy requirements in a dynamic and efficient manner. The main contributions of this paper include:
\begin{itemize}
\item We propose an optimization approach combining the orchestration of the 5G mobile infrastructure including mobile edge jointly with the energy-aware optimization of the on-board robot sensor applications.
\item We design a framework for the deployment of robotic applications with the 5G infrastructure, defining the interfaces and the interactions between the proposed joint orchestration logic with the individual robot orchestrator and the 5G orchestrator.
\item We evaluate our proposed approach by means of an exhaustive simulation campaign, showcasing the achievable gains deriving from the joint orchestration of the 5G infrastructure and robot applications, both in terms of exploration time as well as the resulting energy savings.
\end{itemize}
The remainder of the paper is structured as follows.
Section~\ref{sec:related} summarizes related works in the field.
Section~\ref{sec:framework} presents the main building blocks of our architecture and describes the interaction among the different modules.
Section~\ref{sec:problem} formulates our optimization problem detailing our model assumptions.
Section~\ref{sec:perf_eval} validates the design principles of our solution by means of a comprehensive simulation campaign highlighting the main benefits deriving from our approach.
Finally, Section~\ref{sec:conclusion} concludes this paper.
\section{Related Work}
\label{sec:related}
Several works in the literature investigate the adoption of mobile networks to control robots in outdoor scenarios. In~\cite{Voigtlnder2017} the authors propose a framework for the offloading of time-critical and computational exhaustive operations onto a distributed node architecture, where the communication between the robot and the cloud server is done via 5G.
Despite offloading computational intensive task would help robots to save energy, in realistic environments the limited energy availability provided by on-board batteries still represents a major limitation. Ideally, robots would require the largest batteries in order to extend their mobility range. At the same \cd{time}, heavier batteries would impact on their energy consumption rate, thus introducing a design trade-off~\cite{Albonico2021}.
Swanborn \textit{et al.} identify robot navigation as the main energy consumer~\cite{Swanborn2020} at runtime. Additionally, they identify secondary sources of energy consumption, such as inefficient hardware, inefficient management algorithms, idle times, operational inefficiencies (e.g., poor quality software that leads to unnecessary stops and/or turns, as well as sharp acceleration and deceleration), processing energy, and finally, unnecessary communication and wasting of sensor data acquisition. We highlight that our solution will positively affect the last two drawbacks.
To optimize the exploration of unknown areas, several energy-aware management schemes have been proposed in the literature. \cite{Rappaport2016} proposes an approach for energy efficient path planning during autonomous mobile robot exploration. The idea is for a unique robot to efficiently explore the environment and periodically return to the starting point of the exploration for recharging its battery. The periodicity depends on an adaptive threshold that concurrently considers the movement of the robot, its power consumption and the current state of the environment. The authors focus on minimizing the overall travel path in order to minimize the energy consumption.
Since exploring an area poses hard limits in a single robot settings, in their follow up work they extended the idea to teams of coordinated robots sharing a limited number of Charging Points (CPs) while exploring a structured, unknown environment with unknown obstacles. The objective in this case aims at exploring the area as fast as possible~\cite{Rappaport2017}. Since it is infeasible to pre-compute an optimized schedule due to a limited time horizon, an energy-aware planner is used for adaptive decision making on when and where to recharge. However, their results are limited to numerical simulations.
Similarly, Benkrid \textit{et al.} investigate the problem of multi-robot exploration in unknown environment\cd{s}. In their work, they propose a decentralized coordination approach to minimize the exploration time while considering the total motion energy saving of the mobile robots \cite{Benkrid2019}. The exploration target is defined as a segment of the environment including the frontiers between the unknown and the explored areas. Each robot evaluates its relative rank, and compare it with the other robots of the team considering the energy consumption to reach this exploration target. As a result, the robot is assigned to the segment for which it has the lowest rank.
They evaluate their proposal through simulation experiments as well as ROS-enabled robots. However, inaccuracies during the robot localization and the map generation are not considered. Additionally, they only consider scenarios with unlimited energy availability, or limited energy without the possibility of recharging.
\xl{Nevertheless, none of the above works considers networking aspects in their optimization, nor adopt an NFV approach for their deployments. Recent works such as~\cite{DEEP_Platform} and~\cite{5G-Dive}, started exploring the benefits of adopting 5G and cloud-native deployment in robotic applications exploiting the offloading of computational tasks to fog, edge or cloud systems so as to build a cloud-to-things continuum. Despite their promising results, the current research is mainly focusing on virtualization aspects, and on the development of orchestration platforms to automate the deployment
and allocation of both networking and computing resources over the 5G infrastructure.
However, such orchestration platforms only control the infrastructure resources, and neither consider the internal logic of robotic applications nor decide any actions for the robots. To the best of our knowledge, this is the first work to jointly consider robots and infrastructure resource orchestration, pursuing robot energy consumption optimization.}
\begin{figure}[t!]
\centering
\includegraphics[trim = 2cm 1.9cm 4.5cm 0cm , clip, width=\columnwidth ]{Figures/scenario}
\caption{\small Overview of the architectural building blocks.}
\label{fig:scenario}
\vspace{-4mm}
\end{figure}
\section{Framework Overview}
\label{sec:framework}
As depicted in Fig.~\ref{fig:scenario}, we consider a set of ground robots deployed in an unknown outdoor environment, where wireless communication is provided by means of a 5G network.
We assume a 5G Radio Access Network (RAN) composed by a set of base stations (gNBs) providing radio coverage over the area of interest. We also assume the presence of an edge and a remote cloud platforms, to host robotic applications and the 5G core functionalities, respectively, running as virtual or container-based instances within a computing infrastructure. A non-exhaustive list of robot applications include motion planning, video processing, etc., while the 5G core accounts for the set of user authentication, connection setup and mobility management functionalities proper of a mobile network. We assume the robot controller running in the edge premises, and rely on the presence of a User Plane Function (UPF) to seamlessly route the 5G data plane traffic generated by the robots to the edge platform, \lz{therefore favouring low-latency communication and wider bandwidth availability}.
In this work, we focus on a surveillance (public protection and disaster relief) use-case. Cloud-based robots can perform 24/7 security inspections, replacing security personnel, reducing cost and storing all data needed. They collect video and images and send them to the cloud for real-time identification of suspicious people and activity. Similar robots are already being used at airports and in outdoor rescue scenarios~\cite{airport_robots}.
\subsection{Robot Orchestration}
In this paper we build on the Robot Operating System (ROS) design and specifications~\cite{ROS} for the control and orchestration of robots and the onboarding of robotic applications. ROS is an open-source robotics middleware that provides common functionality (e.g., read sensor data, navigation, planning, etc.) over general hardware abstraction using low-level device control. It contains a collection of tools, libraries, APIs and conventions that simplify the task of creating complex and robust robot behaviour across a variety of robotic systems~\cite{Quigley2009}.
A system built using ROS consists of a number of processes, potentially on a number of different hosts, connected at runtime in a peer-to-peer topology. The ROS topology is supported by a lookup mechanism to allow processes to find each other in real time. The mechanism is implemented through ROS Masters and ROS nodes.
\cd{While a ROS Master represents a stateless entity that coordinates the ROS nodes, the
ROS nodes communicate with each other by passing messages, which are defined with a strictly typed data structure and published through topics. }
ROS was initially designed for standalone robotic applications, with the centralized design for serving a single robot in a local area network.
As such, it is not suitable in upcoming multi-robot cloud-based applications, which demand for a distributed environment containing a variety of networking and cloud resources to be concurrently coordinated in order to connect and control multiple robots in potentially distributed areas.
In its latest releases~\cite{Maruyama2016}, ROS mitigates the issue of real-time topic sharing over a distributed platform by means of a novel Data Distributed Service (DDS), while supporting more flexible container-based deployments~\cite{ROS_Container}.
In order to seamlessly control robots and related robotic applications, we envision the \emph{Robot Orchestrator} entity as composed by
a layered architecture consisting of three layers: application layer, ROS client layer, and ROS middleware layer. The application layer hosts a variety of robotic applications offered with run-time application programming capabilities. The ROS client layer provides a set of ROS client APIs~\cite{ROS_RCL_APIs} based on the built-in ROS client libraries to the developer interface supporting different languages such as C, C++, Python. The ROS middleware layer offers a set of APIs~\cite{ROS_RWP_APIs} to enable compatibility with different interchangeable low-level communication protocols, and support distributed data and service sharing. Through these provided APIs, the robot orchestrator is able to translate the application logic into a set of instructions to control and coordinate groups of robots via ROS command messages dispatched through its Southbound Interface (SBI).
\subsection{5G Orchestration}
In its simplest definition, a 5G Orchestrator is in charge of the allocation and management of the 5G infrastructure resources, including those required to enable robots communication and transmit their application data, as well as the set of computing resources to host and run the robotic control plane applications. \lz{For example, a 5G orchestration solution may decide on the amount of radio resources to be allocated in order to support both the robot control plane (e.g., navigation and velocity commands), and the robot data plane (e.g., video and sensor data) communications.} \lz{Besides, the robot applications can be deployed in a virtualized environment, e.g., being containerized in a local or a cloud computing infrastructure. In such case, the 5G orchestrator is also in charge of the life-cycle management of such container-based robotic application instances, including their on-boarding, instantiation and termination, automatic scaling and self-healing operations. At last but not least, the 5G orchestrator can also determine a proper placement strategy of placing robotic applications which, thanks to softwarelized and cloud-native approaches, may be deployed locally, i.e., onto the robot computing infrastructure, or remotely, in edge and cloud platforms, or even adopting hybrid approaches.
Nevertheless, an accurate placement strategy demands for proactive resource allocation in order to decide optimal amount of computing, memory and storage resources to support the provisioning of various robotic applications, both onto the robots and edge/cloud platforms.}
The 5G orchestrator can be built relying on existing open source orchestrator platforms such as Open Source MANO (OSM) or leveraging on open-source orchestration platforms developed for instance in \cite{5GT_arch} \cite{5Growth-commag}. From a system architecture perspective, the 5G orchestrator consists of three layers: service layer, orchestration layer and resource layer. The service layer defines an \textit{intent engine} to receive and process the application requests, translating the application requirements and mapping them to \emph{network slices} in the form of network slice template as defined by the 3GPP TS 28.531~\cite{ts28531} and the network slice resource model (3GPP TS 28.541~\cite{3GPPTS}).
The orchestration layer consists of
\xl{a \textit {Management and Network Orchestration (MANO)} stack to enforce the OROS{} 5G policy on the allocation of resources, placement and life-cycle management of the robotic applications. These include the instantiation and releasing of computing resources to host drivers and processing instances managing the corresponding sensors on the robots, and their associated applications on a computing infrastructure when activating or deactivating sensors.} On the bottom is the resource layer. It includes a \textit{Virtual Infrastructure Manager (VIM)}, which interacts with the underlying physical infrastructure and offers unified abstractions over the heterogeneous set of mobile resources. It carries out monitoring, allocation and management of resources across the infrastructure and exposes this information to the orchestrator engine to guide its tasks.
\subsection{OROS{}}
So far, major effort has been taken in the definition of orchestrating platforms in each corresponding domain, e.g., OSM~\cite{OSM}, ONAP~\cite{ONAP} in the 5G domain, and ROS~\cite{ROS} in the robotic environment. \xl{However, the orchestration tasks work independently, and lack of dedicated data models and interfaces to interconnect them.}
\xl{Traditionally, robot navigation and path planning processing are performed by a local robot controller deployed within the same premises.
When considering modern smart multi-robot platforms instead, a dedicated controller entity is generally located in the cloud.
On the one hand, to enable the communication from the central controller entity to the individual robots, mobile network resources should be allocated and also adapted to guarantee the bandwidth and latency requirements, especially when computing offloading is also part of the robot task.}
\xl{On the other hand, enabled by the NFV technologies, modern robot sensing devices not only produce (and/or consume) sensor data, but also can be controlled by dedicated software applications \xl{(e.g., based on ROS)} running on top of a shared computing infrastructure.} Without a joint orchestration solution, these software instances will always remain active \xl{and consume the infrastructure resources, impacting on the overall energy consumption.}
In this scenario, we advocate for the adoption of \xl{novel application-oriented orchestration solutions} to guide the overall life-cycle management of cloud computing resources, provisioning of dedicated resources, and context-aware robot motion planning, pursuing energy savings strategies.
To fill this gap, in this work we propose OROS{}, a solution for the joint orchestration of the robotic and 5G ecosystem, to control ROS-driven collaborative connected robots in 5G networks. The architecture design of the proposed OROS{} solution is depicted in Fig.~\ref{fig:OROS}. OROS{} seamlessly connects the orchestrating entities of the two domains, and coordinates their operations via a \xl{joint Robots-5G Orchestration module, as the central brain of our solution.}
\begin{figure}[t!]
\centering
\includegraphics[clip, trim = 5cm 0cm 8cm 1cm, width=\columnwidth]{Figures/OROS.pdf}
\caption{\small Architecture overview of the OROS{} solution.}
\vspace{-4mm}
\label{fig:OROS}
\end{figure}
During the operation phase, assuming that the robotic applications are already instantiated and running, the joint Robots-5G orchestration module makes high-level joint orchestration decisions, namely \textit{Robotic Policies} and \textit{5G Policies}, and sends to each domain-specific orchestrator. \xl{It relies on the input taken from the monitoring metrics exposed by the Robot Orchestrator via its Northbound Interface (NBI) (e.g., ROS topics detailing velocity, location, and power consumption of the robots), and those exposed by the 5G Orchestrator (e.g., radio resource availability, and computing resources on the MEC and robot platforms for running the robotic applications, etc.).}
Upon receiving the robotic policy, the Robot Orchestrator is in charge of reconfiguring the related robotic applications and translating the required corresponding actions to robot command messages, through exploiting the provided ROS client and middleware APIs. Examples of these include instructing the robot to move towards its new navigation goal, or switch off a specific sensor to save the energy. In the 5G domain, once receiving a new 5G policy from the join optimization module, the Intent Engine will translate the received policy to update the corresponding network slices based on the new requirements. The 5G orchestration engine will further process the slice update request and optimize the reallocation of 5G resources (e.g., RAN, core, and MEC). \xl{This includes the instantiation and release of the networking and computing resources associated to the robotic applications, which can be switched on/off depending on the decision of OROS{}}. Moreover, the 5G orchestration decides on the migration of the robotic applications towards the MEC and robot platforms. These decisions will be forwarded to the MANO to execute the operation workflows, and consequently to the VIM in order to enforce the configurations of the resources on the robot platforms and the 5G infrastructure via the SBI following ETSI IFA 005 specifications~\cite {ifa005}. In the following Section~\ref{sec:problem}, we detail the energy-aware mathematical formulation that guides the orchestration decisions while interacting with the specific domain orchestrators.
\begin{table*}[t]
\caption{Model parameters}
\label{tab:optParam}
\centering
\begin{tabular}{l|l}
\hline
\textbf{Parameter} & \textbf{Definition} \\
\hline
$\mathcal{T} = \{t_1, \dots, t_{|\mathcal{T}|} \}$ & Set of time instants; index $t$ refers to time instant $t_t$\\
$\mathcal{R}=\{r_1, \dots, r_{|\mathcal{R}|}\}$ & Set of robots; index $r$ refers to task $r_r$\\
$A \times B$ & Geometric dimensions of the area of interest\\
$\mathcal{G} =\{g_{a,b}, \forall (a,b) \in (A,B) \}$ & Grid representing the area to be explored \\
$m_{a,b,a',b'}$ & Terrain-velocity constant for moving from position $g_{a,b}$ to $g_{a',b'}$ \\
$u_{r,t}$ & Binary decision variable indicating if the charging station is being used at time $t$ by robot $r$\\
$l_{r,t,a,b}$ & Binary decision variable indicating if robot $r$ is in position $g_{a,b}$ at time instant $t$\\
$e_{t,a,b}$ & Binary variable indicating if the unit of area $g_{a,b}$ has been explored at time $t$\\
$b_{r,t}$ & Continuous variable indicating the battery level of robot $r$ at time instant $t$, where $0 \leq b_{r,t} \leq B_{max}$ \\
$CR$ & Charging rate provided by the charging station \\
$P_{move_{a,b,a',b'}}$ & Power consumed by moving from position $g_{a,b}$ to position $g_{a',b'}$\\
$P_{RX}$ & Power consumed for receiving a packet \\
$P_{TX,a,b}$ & Power consumed for transmitting a packet \\
$P_{SEN}$ & \lz{Power consumed by activating sensors, local data processing and computing infrastructure
} \\
\hline
\end{tabular}
\vspace{-5mm}
\end{table*}
\section{Problem Formulation}
\label{sec:problem}
\subsection{OROS{} - Joint Robots-5G Orchestration Optimization}
\label{sec:optimization}
Hereafter, we present our assumptions, notation and problem formulation
to address the Energy-Aware Optimization problem. All variables and system parameters are resumed in Table~\ref{tab:optParam} to allow faster referencing.\\
\textbf{Input variables}
Let us consider a discrete set of time instants denoted by the set $\mathcal{T} = \{t_1, \dots, t_{|\mathcal{T}|} \}$, and a set of robot devices $\mathcal{R}=\{r_1, \dots, r_{|\mathcal{R}|} \}$. Each robot $r \in \mathcal{R}$ is equipped with a rechargeable battery characterized by a limited capacity $B_{max}$, $\forall r \in \mathcal{R}$,
whose charging status $b_{r,t}$ varies over time depending on the robot operational activities, i.e.,
$0 \leq b_{r,t} \leq B_{max}, \forall t \in \mathcal{T}$. Each robot is also equipped with multiple sensors, such as cameras, lidar, etc.
We assume robots to be deployed in an outdoor environment covered by a mobile infrastructure, as to enable 5G connectivity.
Without loss of generality, let us define the area of interest with dimensions $A \times B$ meters, and decompose the 2D surface into a grid $\mathcal{G} =\{g_{a,b}, \forall (a,b) \in (A,B) \}$, where each element $g_{a,b} \in \mathcal{G}$ needs to be explored by at least one robot during the operational phase. The dimension of $g_{a,b}$, i.e., $|g_{a,b}|$, depends on the maximum field of view of the adopted robot camera and sensors, and we assume such cameras/sensors to be able to provide a $360^{\circ}$ view of the surrounding environment.
Each robot needs to receive periodical updates from the radio interface, as well as upload
sensed environmental information.
For doing so, each robot adapts the Modulation and Coding Schemes (MCS) used according to the perceived channel quality, which we assume proportional to the instantaneous distance of the robot to the serving base station, as will be later detailed in Section \ref{sec:perf_eval}.
We assume the serving base station located at position $(g_{a_{BS},b_{BS}} ) \in \mathcal{G}$.
\begin{figure}[t!]
\centering
\includegraphics[trim = 0cm 0.5cm 0cm 0cm , clip, width=0.99\columnwidth ]{Figures/20211221_Problem_Example.png}
\caption{\small Example of \cd{the} robot exploration problem solved with multiple robots.}
\label{fig:example}
\vspace{-6mm}
\end{figure}
To keep track of the multi-robot exploration, we introduce $e_{t,a,b}$ as a binary variable indicating if the 2D area unit $g_{a,b}$ has been already explored at time $t$, or not.
Exploration of unknown areas may require charging stations to extend the operational time of rescue robots. We assume the presence of a charging station (CS) located in position $(g_{a_{CS},b_{CS}}) \in \mathcal{G}$, which is known a-priori to all robots, and that can be either available or unavailable (due to an already charging robot) at a certain time instant. The CS provides a charging rate $CR$
to recover the robot battery level.
We define $\Delta t$ as the period between two instants in $\mathcal{T}$ (i.e., $\Delta t = |t_t - t_{t-1}|$), and assume the spatial mobility of every robot $r \in \mathcal{R}$ limited to a unit of distance $\Delta x$ for every $\Delta t$, where $\Delta x$ is the distance between the centers of two neighbour area units $g_{a,b}$ and $g_{a',b'}$ as depicted in Fig.~\ref{fig:example}.
Therefore, it turns that each robot moves at a velocity \( \frac{\Delta x}{\Delta t} \) if moving into the left/right or up/down directions, but moves at velocity \( \frac{\sqrt{2}\Delta x }{\Delta t} \) when travelling in diagonal directions. Clearly, if the velocity increases, so does the corresponding energy cost.
We consider this by introducing the terrain-velocity constant $m_{a,b,a',b'} \quad \forall a, a' \in A, \quad \forall b, b' \in B$, which weights the energy consumption of the robots from moving from position $g_{a,b}$ to position $g_{a',b'}$. The value of $m_{a,b,a',b'}$ is set to infinite in presence of an obstacle, therefore influencing its mobility in the corresponding unit of area.
Whenever the robot moves, we can derive its power consumption by adapting Equation~1 of Rappaport~\cite{Rappaport2016}, where instead of only computing the velocity, we consider the terrain-velocity constant to account for obstacles and terrain conformity form as:
\begin{equation}
\label{eq:Pmove}
P_{move_{a,b,a',b'}} = 0.29 + 7.4 m_{a,b,a',b'}.
\end{equation}
As mentioned before, robots exploit an existing mobile infrastructure to communicate with the orchestrating entity. We introduce $P_{TX,a,b}$ as a variable representing the energy consumed by the robot for transmitting data. The actual robot transmit power should be adapted as to compensate the radio path-loss. Therefore, the value of $P_{TX,a,b}$
depends on the current robot location $(g_{a,b})$ and its distance from the serving base station.
Similarly, robots consume energy for receiving data. In this case, we introduce $P_{RX}$ as a constant value independent of the robot location.
Finally, since robots need to map and explore the terrain, they need to use their camera and sensors, as well as process those data to avoid inefficient raw data transmission. We collect the energy consumption derived by all these activities in the variable $P_{SEN}$, which represents the energy consumed by robot sensors and corresponding local data processing. \lz{We assume this variable also to include the energy consumption coming from the computing infrastructure locally running on the robot}.
\textbf{Decision variables}
Let $l_{r,t,a,b}$ be a binary decision variable to control the robot mobility. Its value gets positive if the robot $r$ is at position $g_{a,b}$ at time instant $t$.
Similarly, we introduce $u_{r,t}$ as a binary decision variable performing decisions on which robot $r$ has to recharge its battery at time $t$.
\textbf{Objective}
To explore an unknown area as fast as possible, and therefore increase the chances of detecting the target object (or person), we need to maximize the explored area within the given time period $|\mathcal{T}| \times \Delta t$, therefore we can write our objective function as:
\begin{equation}
\label{eq:maximiza}
\max \sum_{t \in \mathcal{T}} \sum_{(a,b) \in (A,B)} e_{t,a,b}
\end{equation}
\textbf{Constraints}
We assume that the charging station can only recharge one robot $r \in \mathcal{R}$ at every time instant $t \in \mathcal{T}$, therefore we introduce the following constraint:
\begin{equation}
\label{eq:const1}
\sum_{r \in \mathcal{R}} u_{r,t} \leq 1 \quad \forall t \in \mathcal{T},
\end{equation}
\noindent additionally, a robot $r \in \mathcal{R}$ can be charged at some time $t \in \mathcal{T}$ only if the robot is in the charging station at that specific time instant
$t$:
\begin{equation}
\label{eq:const6}
u_{r,t} \leq l_{r,t,a_{CS},b_{CS}} \quad \forall t \in \mathcal{T}, \forall r \in \mathcal{R}.
\end{equation}
\noindent Clearly, the duration of the recharge period may comprise multiple time intervals $t$, to allow more energy to fill the battery.
With the following constraint, we ensure that each robot $r \in \mathcal{R}$ can only be in one place in every time instant $t \in \mathcal{T}$:
\begin{equation}
\label{eq:const7}
\sum_{(a,b) \in (A,B)} l_{r,t,a,b} = 1 \quad \forall r \in \mathcal{R}, \forall t \in \mathcal{T}.
\end{equation}
\noindent In order to keep track of the exploration progress among multiple robots, if any robot $r \in \mathcal{R}$ visited an area unit $g_{a,b} \in (A, B)$ at some earlier time, or if it is exploring such area unit at the current time $t$, that area becomes explored at time $t$ and we update the variable $e_{t,a,b}$ accordingly:
\begin{equation}
\label{eq:const3}
e_{t,a,b} \leq e_{t-1,a,b} + \sum_{r \in \mathcal{R}} l_{r,t,a,b} \quad \forall t \in \mathcal{T}, \forall (a,b) \in (A,B),
\end{equation}
\begin{equation}
\label{eq:const4}
e_{t,a,b} \geq e_{t-1,a,b} \quad \forall t \in \mathcal{T}, \forall (a,b) \in (A,B),
\end{equation}
\begin{equation}
\label{eq:const5}
|\mathcal{R}| e_{t,a,b} \geq \sum_{r \in \mathcal{R}} l_{r,t,a,b} \quad \forall t \in \mathcal{T}, \forall (a,b) \in (A,B).
\end{equation}
\noindent With the following constraint we define the mobility boundaries of the robots, and ensure that for each $\Delta t$ a robot $r \in \mathcal{R}$ can only move to a neighbour area unit, or stay in the same position:
\begin{gather}
l_{r,t+1,a,b} \leq l_{r,t,a,b} + l_{r,t,a-1,b} + l_{r,t,a+1,b} + l_{r,t,a,b-1} + \notag \\
l_{r,t,a,b+1} + l_{r,t,a-1,b-1} + l_{r,t,a+1,b+1} + l_{r,t,a-1,b+1} + \notag \\
l_{r,t,a+1,b-1} \quad \forall r \in \mathcal{R} , \forall t \in \mathcal{T}, \forall (a,b) \in (A,B). \label{eq:const8}
\end{gather}
\noindent Equation~\ref{eq:const8} affects the robot velocity. In fact, as mentioned before, the velocity and therefore the power consumption depends on the direction of the robot movement. This is taken into account in the constant $P_{move_{a,b,a',b'}}$.
Conversely, if the robot is at the charging station and it is being charged (i.e., $u_{r,t} = 1$), the energy level of its battery increases with a constant charging rate $CR$. Furthermore, in every time instant the robot is not charging, it sends or receives data. When transmitting, the consumed power depends on the distance to the base station (according to $P_{TX,a,b}$), while in case of reception it depends on $P_{RX}$ as previously discussed. Finally, if a robot has never been in an area unit, its sensors, camera, processing units and transmission elements should be active. However, if the robot is in an already explored area, an advanced orchestration solution may turn them off for saving energy. Robots are expected to be reachable by the managing entity in every time instant. For this reason, we assume robots never switch off their receiving antennas. Taking this into account, the battery level $b_{r,t}$ gets updated in every time instant by the following equation:
\begin{gather}
b_{r,t+1} = b_{r,t} + CR \times u_{r,t+1} - P_{RX} \times (1 - u_{r,t+1}) - \notag \\
\sum_{(a,b) \in (A,B)} \sum_{(a',b') \in (A,B)} l_{r,t,a,b} \times l_{r,t+1,a',b'} \times P_{move_{a,b,a',b'}} \notag \\
- P_{SEN} \times \sum_{(a,b) \in (A,B)} (1 - e_{t,a,b}) \times l_{r,t+1,a,b} - \label{eq:const9a}\\
\sum_{(a,b) \in (A,B)} P_{TX,a,b} \times (1 - e_{t,a,b}) \times l_{r,t,a,b} \quad \forall t \in \mathcal{T} , \forall r \in \mathcal{R}. \notag
\end{gather}
\subsection{Energy-Aware SLAM{} Comparison Benchmark}
\label{sec:comparison}
In order to compare our energy-aware algorithm, we introduce a benchmark that does not allow to dynamically turn off the robot sensors, even if robots are travelling within an already explored area. This means that whenever the robot is not charging, its camera and processing units are enabled, and sensor data is flowing through the closest base station. To ensure fair comparison, we maintain the same objective function (see Equation \ref{eq:maximiza}), as well as the same set of constraints, whereas we replace Equation~\ref{eq:const9a} to enforce the \cd{energy-aware SLAM} scenario as the following:
\begin{gather}
b_{r,t+1} = b_{r,t} + CR \times u_{r,t+1} - (P_{RX} + P_{SEN} )\times (1 - u_{r,t+1}) \notag \\
- \sum_{(a,b) \in (A,B)} \sum_{(a',b') \in (A,B)} l_{r,t,a,b} \times l_{r,t+1,a',b'} \times P_{move_{a,b,a',b'}} \notag \\
- \sum_{(a,b) \in (A,B)} P_{TX,a,b} \times (1 - u_{r,t+1}) \times l_{r,t+1,a,b} \notag \\
\quad \forall t \in \mathcal{T} , \forall r \in \mathcal{R}.
\label{eq:const20}
\end{gather}
Finally, we can introduce our benchmark scenario as:\\
\noindent \textbf{Problem}~\texttt{SLAM{}}:
\label{prob:always_On}
\begin{flalign}
\quad\quad & \text{max} \sum_{t \in \mathcal{T}} \sum_{(a,b) \in (A,B)} e_{t,a,b} \nonumber & &\\
\quad\quad & \text{subject to:} \nonumber\\
\quad\quad & \quad\quad (\ref{eq:const1}) (\ref{eq:const6}) (\ref{eq:const7}) (\ref{eq:const3}) (\ref{eq:const4}) (\ref{eq:const5}) (\ref{eq:const8}) (\ref{eq:const20}); \nonumber
\end{flalign}
We highlight that this approach is also energy-aware, meaning that it allows robots to follow the best path to explore the area as fast as possible, taking into account the available energy on the battery, and the possibility of recharging it.
\section{Performance Evaluation}
\label{sec:perf_eval}
In this section, we evaluate the performance of the energy-aware optimization module when dealing with collaborative robot scenarios. We first introduce the scenario setup. Then, in order to evaluate the orchestration benefit, we compare our OROS{} energy-aware optimization module that is capable of managing the sensing applications (as seen in Section \ref{sec:optimization}), against the standard SLAM{} approach, which does not allow to dynamically turn on and off the sensing modules when robots are in already explored areas (see Section \ref{sec:comparison}).
\subsection{Scenario Setup and Methodology}
\label{sec:scenarioSetup}
In our evaluation, we consider a variable set of robots that need to be jointly orchestrated as to explore an unknown area in the fastest possible way. Table~\ref{tab:setup} resumes the experimental parameters used in our evaluation.
All robots start their exploration from the same point within the area of interest, co-located with the charging station.
We set $\Delta t = 10$s and $\Delta x = 10 m$, which translates into $|g_{a,b}| = 100m^2$. This allows to derive the velocity of the robot as $\frac{\Delta_x}{\Delta_t} = 1 m/s$ if it moves into the up/down/left/right direction, or at
$\frac{\sqrt{2}\Delta_x}{\Delta_t} = \sqrt{2} m/s$ when moving in diagonal, which is in line with the maximum speed of $2 m/s$ provided by commercial robot devices~\cite{Robot_velocity}.
Each robot is equipped with a fully charged battery at the beginning of the exploration phase.
About the charging station, in our experiment we consider the worst case scenario with a single charging station providing the lowest charging rate among the possible commercial options, i.e., $CR=9.24 J/s$, as in the case of TurtleBot3\footnote{https://www.robotis.us/lipo-battery-charger-lbc-010/}.
\begin{table}[t]
\caption{Experimental Setup}
\label{tab:setup}
\centering
\begin{tabular}{ll|ll}
\hline
\textbf{Definition} & \textbf{Value} & \textbf{Definition} & \textbf{Value} \\
\hline
$|\mathcal{T}|$ & 15 & $|\mathcal{R}|$ & 1,2,3\\
$A \times B$ & $40\times40$ $m^2$ & $|g_{a,b}|$ & $10\times10$ $m^2$ \\
$B_{max}$ & $5000$ J & $\Delta t$ & $10$ s \\
$CR$ & $9.24$ J/s & $v$ & $1$, $\sqrt{2}$ m/s \\
$Tx Power$ & $20$ dBm & $Receiver Gain$ & $-20$ dB \\
$Power Noise$ & $-104$ dBm & $P_{RX}$ & $4$ W \\
$P_{SEN}$ & $12$ W & & \\
\hline
\end{tabular}
\vspace{-4mm}
\end{table}
We characterize the energy consumed by the robot for locomotion, sensor operation and radio wireless communication as follows.
First, the power consumed by the locomotion $P_{move_{a,b,a',b'}}$ (see Equation~\ref{eq:Pmove}) depends on the terrain-velocity constant $m_{a,b,a',b'}$, which also takes into account detected obstacles. This constant can be defined as:
\begin{equation}
m_{a,b,a',b'} =
\begin{cases}
0 & \text{if $a' = a$ and $b' = b$}\\
\infty & \text{if there is an obstacle}\\
& \text{between $(a,b)$ and $(a',b')$ }\\
v = \{1,\sqrt{2}\} & \text{if $a' \neq a$ or $b' \neq b$}
\end{cases}
\end{equation}
where $v$ is the velocity in m/s.
Note that we assume no power consumption for turning the robot in the desired direction.
Second, we characterize the power consumed by the sensors $P_{SEN}$, which includes processing all the generated data by the camera, lidar and other sensors, \lz{as well as the power consumed by the local computing infrastructure}. Based on the values from~\cite{Mei2005}, we set it to take a value of $12 W$.
Finally, we consider each robot is equipped with a 5G New Radio (NR) antenna module consuming energy whenever transmitting or receiving data.
The power dissipated by the robot depends on its distance from the serving base station and on the corresponding Signal to Noise Ratio (SNR) which, in turn, affects the adoption of a particular MCS.
The processing power consumed to encode radio packets is also affected by the instantaneous MCS in use. A detailed characterization of the power consumption in virtual base station as function of the MCS can be obtained from ~\cite{Ayala2021}. In our evaluation we set a transmission power of 20 dBm, a receiver gain of $-20 dB$, a Power Noise of $-104 dBm$ and a carrier frequency of $3.5 GHz$, as determined by 3GPP ~\cite{3gppstandard}. In case of downlink transmission, we consider an average value of $4W$ according to the measurements of~\cite{Ayala2021}, and assume the receiving module always on, to enable reachability of the robot devices with control messages at any time. We use the Gurobi solver~\cite{gurobi} to address the optimization problem described in Section~\ref{sec:problem}.
\begin{figure}[t!]
\centering
\includegraphics[width=.95\columnwidth]{Figures/20210730_RAExplored_12_4_0_n.PNG}
\caption{\small Area Explored when robots do not charge.}
\label{fig:20210713_RAExplored_12_4_0}
\vspace{-4mm}
\end{figure}
\subsection{Assessing the gain of the orchestrator}
Hereafter we compare our OROS{} against the SLAM{} approach.
As our work targets mission-critical use-cases, we advocate for the adoption of the area exploration rate as main metric of interest. This would allow to compare different approaches, and identify the one that maximises such objective in the shortest time.
Starting from an obstacle-free scenario of $40\times40 m^2$, Fig.~\ref{fig:20210713_RAExplored_12_4_0} depicts the resulting discovery performances for the two approaches when deploying an increasing number of robots, namely $1$, $2$ or $3$.
For each strategy, we detail at the top of each bar the time spent (in epoch) to reach the corresponding explored area.
In this first example we do not consider the presence of a recharging station, however when using 3 robots we are able to explore the full area within the time boundaries.
It is straightforward to understand that the concurrent deployment of multiple robots allows to explore additional areas in the same amount of time. Therefore, in order to focus the attention on our main findings, in this work we do not consider the adoption of an extreme number of robot devices.
In particular, in the worst case scenario, i.e. when only a single robot is deployed (and there is no option to recharge), it can only explore the $31.25$\% of the area when not using OROS{}, while reaching a rate of $37.5$\% with our solution, highlighting the effects of a more efficient battery usage.
Such improvement gets even more evident when considering an increasing number of robots. In the two robots scenario for example, we move the overall exploration rate from $56.25$\% to $68.75$\%, at the cost of $1$ additional time epoch. Finally, when $3$ robots are simultaneously active in the target area, the benefit of our approach is maximized, moving the ratio of explored area from $75$\% to $100$\%, while also using $1$ additional time epoch as in the previous case.
Fig.~\ref{fig:20210713_RAExplored_15_4_0} depicts the overall evolution of the exploration rate over time, for each of the two approaches and for an increasing number of robots.
After an initial exploration phase, whose rate linearly grows with the number of rescue devices deployed in the field, robots need to follow their path back to the charging station. This, together with the amount of time to recharge the batteries leads to a non-increasing exploration rate. In the $1$ robot - SLAM{} scenario this is particularly evident. After recharging, the robot restarts its exploration task which translates into the exploration rate growth during the $15$-th epoch. Obviously, this time may be shortened by employing more recharge stations, or making use of a faster charging rate.
\begin{figure}[!t]
\centering
\includegraphics[width=.90\columnwidth]{Figures/20210730_AExplored_15_4_0.PNG}
\caption{\small Explored Area over time when one robot is able to charge.}
\label{fig:20210713_RAExplored_15_4_0}
\vspace{-4mm}
\end{figure}
\subsection{Sensitivity Analysis}
In the following, we perform a sensitivity analysis on several key-parameters of our model formulation assessing the impact of their variation onto the overall system performances.
First, we evaluate the influence of different sensing power consumption, dubbed as $P_{SEN}$ in our formulation, due to increasing or decreasing computation requirements, e.g., local video processing, or novel robot sensors.
\lz{In particular, Fig.~\ref{fig:20210720_PsensRAExplored_12_4_0} displays the achieved explored area by two rescue robots when the power consumed by the sensors and by the underlying robot computing infrastructure varies, e.g., due to larger volume of traffic or intense processing.}
In this case, additionally to the reference value of $12$W resumed in Table~\ref{tab:setup}, we consider both less-demanding scenarios and two extremely energy binding scenarios.
From the picture, we can appreciate the influence of $P_{SEN}$ on the robot exploration performances and how, in general, the exploration rate is reduced for an increasing sensor power consumption value. Interestingly, the gap between the SLAM{} approach and our proposed OROS{} follows an opposite trend.
This is due to the fact that our smart orchestration allows turning off the sensors when the robot is in an area that has already been explored, therefore limiting the overall energy consumption.
Despite the higher exploration gap is achieved in the worst case scenario $P_{SEN}=36W$, we also notice how extreme sensor energy requirements actually affect the overall robot mobility, finally leading to fully discharged batteries within a shorter number of time epochs, e.g., $4$ and $5$ for the SLAM{} and OROS{} approaches, respectively.
Conversely, in Fig.~\ref{fig:20210720_BatRAExplored_12_4_0} we analyze the robot performances for a variable battery capacity, namely $B_{max}$. We remark that its value affects the battery level of the rescue robots, as we assume them to start their task in a fully-charged state. As before, we consider a two robot scenario and the reference value reported in Table~\ref{tab:setup} (i.e, $5000$J), as well as additional testing values accounting for greater (i.e., $10000$J and $20000$J) and more constrained capacities (i.e., $2500$J and $1250$J).
If we provide the robots with an energy reserve greater than $10000$J, 2 robots are enough to completely explore the area of interest within the time boundaries. In this case, the impact of a smart resource orchestration over the exploration rate is negligible. Nevertheless, OROS{} allows maintaining a higher charge level in the robot's batteries at the end of the run. In particular, for the $10000$J case when OROS{} is in place robots end with $1784.88$J and $2051.76$J respectively, while their remaining energy is $1445.88$J and $1882.26$J when adopting the SLAM{} approach.
On the other side, despite the presence of a recharge station, if we decrease the battery capacity we limit the robot mobility and reduce the achievable exploration rate. Robots can explore less than a third of the overall area of interest (exactly $31.25$\%) with limited $2500$J batteries. In case of $1250$J, this further reduces to $18.75$\% in $2$ epochs when using OROS{}, while without it,
robots are capable of exploring only $12.5$\% of the space, thus needing
$12$ epochs.
\begin{figure}[t!]
\centering
\includegraphics[width=.92\columnwidth]{Figures/20210730_PsensRAExplored_12_4_0.png}
\caption{\small Explored Area for variable $P_{SEN}$ values.}
\label{fig:20210720_PsensRAExplored_12_4_0}
\vspace{-4mm}
\end{figure}
\setcounter{figure}{7}
\begin{figure}[t!]
\centering
\includegraphics[clip, trim=5cm 7.5cm 6cm 3cm, width=.92\columnwidth]{Figures/obstaclesMovements_v2.pdf}
\caption{\small Example of robot navigation paths for $3$, $7$ and $9$ obstacles.}
\label{fig:obstaclesMovements}
\vspace{-4mm}
\end{figure}
\subsection{Area with obstacles}
In the previous subsections, we considered barrier-free scenarios and focus our study on the sensitivity analysis. Hereafter, we consider a more realistic scenario and introduce not surmountable obstacles along the robot path, and evaluate the overall system performances with constrained robot mobility.
Obstacles are introduced as horizontal lines at the edge of two adjacent grid elements, and they are placed as to avoid unreachable sub-areas.
For the sake of visualization, in Fig.~\ref{fig:obstaclesMovements} we consider a small-scale scenario and depict all possible robot navigation paths for an increasing number of obstacles, namely $3$, $7$ and $9$. For the $3$ obstacles, the number of possible movements is actually much higher than when there are $9$ obstacles.
Fig.~\ref{fig:20210727_Obstacles_12_4_0} builds on the scenario depicted in Fig.~\ref{fig:obstaclesMovements}, and summarizes the
explored area in presence of obstacles, comparing the performances for an increasing number of collaborative robots and the two
discussed
approaches.
To ease the comparison, each bar provides an incremental view with respect to the reference scenario with a larger number of robots. For example, in absence of obstacles, by adopting $3$ robots at the end of the experiment about $20$\% more of the target area was explored when compared with the $2$ robots scenario.
It can be noticed how performances remain constant when a single robot is adopted into the research task. This is in line with our previous results considering the limited energy availability ($5000$J): an increasing number of obstacles has no impact due to a poor mobility range.
Conversely, in multi-robot scenarios, obstacles limit the available paths and induce lower exploration rates. In particular, due to the special geometry of our scenario, obstacles are mostly affecting the initial exploration phase, leading the different robots to follow the same path in their initial steps.
When increasing to the maximum number of robots, either $3$ or $7$ obstacles have the same impact on performances, while in case of $9$ obstacles, the achievable exploration rate is highly reduced. Interestingly, in this case using $3$ or $2$ robots provides the same performances due to the constrained navigation paths when using the SLAM{} approach, while OROS{} outperforms the benchmark in all the considered tests, confirming that the adoption of a smart orchestration solution provides significant benefits in collaborative exploration scenarios.
\setcounter{figure}{6}
\begin{figure}[t!]
\centering
\includegraphics[width=.92\columnwidth]{Figures/20210730_BatRAExplored_12_4_0.png}
\caption{\small Explored Area for variable $B_{max}$ values.}
\label{fig:20210720_BatRAExplored_12_4_0}
\vspace{-2mm}
\end{figure}
\setcounter{figure}{8}
\begin{figure}[t!]
\centering
\includegraphics[width=.94\columnwidth]{Figures/20210730_Obstacles_12_4_0.png}
\caption{\small Effects of an increasing number of obstacles.}
\label{fig:20210727_Obstacles_12_4_0}
\vspace{-4mm}
\end{figure}
\section{Conclusions}
\label{sec:conclusion}
Due to limited computing and energy resource availability, cloud-based robot deployments rely on mobile infrastructure to enable collective robotic intelligence exchange and increase their efficiency when performing tasks. However, current solutions are limited by the fact that robot operating systems and ICT computing and communication platforms, do not have means to interact with each other.
In this paper, we propose a joint optimization framework for the concurrent orchestration of the robotic, computing and communication infrastructure domains. Our results show that mission-critical collaborative robot operations would benefit from the adoption of OROS{}, a joint orchestration solution which significantly improves tasks completion duration by reducing overall energy consumption.
\lz{Future work will focus on the implementation of the proposed OROS{} framework in real deployments, comprising off-the-shelf outdoor robots and 5G commercial infrastructure.
}
\section*{Acknowledgment}
The research leading to these results has been supported by the Spanish Ministry of Economic Affairs and Digital Transformation and the European Union - NextGeneration EU (project no. TSI-063000-2021-6), by the CERCA Programme from the Generalitat de Catalunya, and by the European Union's H$2020$ 5G ERA Project (grant no. $101016681$).
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Deep learning has achieved tremendous progress in various fields including image classification \cite{he2016deep}, object detection \cite{he2017mask}, image segmentation \cite{long2015fully}, etc. However, many security vulnerabilities hinder the deployment of deep neural networks in some risk-sensitive domains like self-driving. Attacks against the robustness of DNNs can be grouped into two categories: training phase and inference phase. Adversarial example attack \cite{carlini2017towards,zhong2021undetectable} is a notorious threat to DNNs, which happens in the inference phase. Nowadays, the backdoor attack \cite{zhoumulti,wang2021backdoorl} is another severe threat to DNNs which happens in the training phase. BadNets \cite{gu2017badnets} is a seminal study to investigate the vulnerability of DNNs during the training phase. The trigger pattern is a conspicuous square in the BadNets. We name the benign inputs stamped with the trigger as malicious inputs. Then attackers alter the label of malicious inputs to the target label and mix the benign and malicious samples to create a new training dataset. The victims training the model under the new training dataset obtain a compromised model, which behaves normally for benign inputs yet returns the target label when the trigger appears.
In the subsequent backdoor studies, researchers focus on the visual distortion of the trigger \cite{li2020backdoorinvisible,li2020invisible,NguyenT21}. In the early studies, the trigger is conspicuous which results in poor visual quality and can be easily removed by human inspection. Li et al. \cite{li2020invisible} propose a novel invisible trigger that resorts to the image steganography technique to ensure the visual quality of the trigger. Nguyen et al. \cite{NguyenT21} propose using image affine transformation as a generator to create a unique trigger for each benign image. Malicious images are warped from the clean images. To the best of our knowledge, ISSBA \cite{li2020backdoorinvisible} is the state-of-the-art invisible backdoor attack that defeats most state-of-the-art defences. ISSBA inspired by deep learning-based steganography employs an encoder proposed in \cite{2019stegastamp} to generate the trigger for each benign input.
Although existing invisible approaches have achieved satisfactory visual quality for humans, they cannot resist statistical detection \cite{zeng2021rethinking}. As a countermeasure for backdoor attacks, backdoor defences also develop rapidly. As our aforementioned description, backdoor attacks stamp the trigger onto the benign input to induce the compromised model to return the target label. The trigger inevitably changes the benign inputs. Therefore, defenders can track the trace left by the trigger to reject the malicious inputs. Zeng et al. \cite{zeng2021rethinking} propose to detect the trace of the trigger from the frequency perspective and thwart various invisible backdoor attacks. Apart from trigger detection, defenders also can diagnose the well-trained model directly. Neural Cleanse \cite{wang2019neural} is a well-known model diagnose-based defence against backdoor attacks. It reversely constructs the potential trigger pattern for each label. The size of the potential trigger pattern of the target label is significantly smaller than those of clean labels. Network Pruning \cite{liu2018fine} is an alternative effective countermeasure against backdoor attacks. Network Pruning deletes dormant neurons for benign inputs in the penultimate layer. For modern DNNs, there are a lot of dormant neurons for benign inputs, whereas they are activated when the trigger appears. Compromised models return target labels without regard to the semantic information of inputs and only depend on the trigger. Dormant neurons are activated when the trigger appears in the feature representation space. Therefore, Network Pruning can purify the compromised model by cutting dormant neurons.
In this paper, we consider the stealthiness of the backdoor attack from two perspectives: input space and feature representation space. We focus on the image classification tasks in this paper, and the input space is the spatial image. We employ a noise following multinomial distribution as the trigger. The parameters of the distribution are generated by each benign image, i.e., each trigger is exclusive to its corresponding benign image. We minimize the cost function of the backdoor attack to update the generator to create the optimal trigger. In terms of feature representation space, we make the feature representations of the malicious images tightly entangled with the benign ones. The defences based on the separateness of feature representation are ineffective to our attack.
The main contributions of this paper are as follows: \textbf{(1)} We provide a novel invisible backdoor attack, which is imperceptible to both human inspection and state-of-the-art statistical detection. The trigger generation is based on a multinomial distribution whose parameters are controlled by each benign image. \textbf{(2)} We consider the separateness of feature representation space caused by the backdoor attacks and focus on the feature representations of malicious inputs to be as identical to the benign ones as possible. \textbf{(3)} We conduct extensive experiments including different datasets and network structures to demonstrate the effectiveness and stealthiness of our approach.
\section{Related Work}
\paragraph{Backdoor Attack.} BadNets \cite{gu2017badnets} is the first seminal study to investigate that DNNs are vulnerable to backdoor attacks during the training phase. First, attackers need to design a trigger pattern, which is a conspicuous square in BadNets. Then, attackers select a small part of benign images to be used as malicious samples whose labels are changed to the target label. Besides changing the label, the trigger pattern (a conspicuous square) is stamped onto the benign images. Afterwards, attackers use the new training dataset containing malicious images to train a compromised model which behaves normally on benign samples yet returns the target label when the trigger appears. ISSBA \cite{li2020backdoorinvisible} is the latest invisible backdoor attack, which employs a well-trained steganography encoder to generate a unique trigger pattern for each image. In the previous studies, the design of the trigger is not trivial. If the trigger pattern is too conspicuous, it can be easily removed by human inspection. However, the backdoor is hardly implanted into the compromised model if the perturbation of the trigger is too slight. DNN-based steganography encoder is suitable to generate the trigger and ISSBA achieves satisfactory performance under the evaluation of multiple defences.
\paragraph{Backdoor Defence.} Since backdoor attacks pose a severe threat to machine learning security fields, backdoor defences \cite{wang2019neural,chen2019deepinspect,wu2021adversarial,li2020neural,zeng2021rethinking} are also rapidly developing. We give a brief introduction about backdoor defences. We roughly divide backdoor defences into two categories: input diagnose-based defences and model diagnose-based defences. Input diagnose-based defences scrutinize the inputs of DNNs and analyze whether it contains the trigger. To the best of our knowledge, FTD proposed by Zeng et al. \cite{zeng2021rethinking} is the state-of-the-art detection. It first conducts Discrete Cosine Transformation (DCT) to transfer the spatial pixels to the frequency domain. Since Zeng et al. observe that various kinds of malicious images (i.e., various backdoor attacks) show a consistent abnormality in the frequency domain from the benign images, they propose using DCT as the first preprocessing to enlarge the trace of the trigger. Then, they employ a DNN-based discriminator to conduct binary classification tasks to determine whether the input image contains the trigger.
In terms of model diagnose-based defences, these approaches directly analyze whether the suspicious model contains backdoors. Neural Cleanse \cite{wang2019neural} is the most well-known defence in this category. Neural Cleanse reversely constructs a potential trigger. This potential trigger can lead the model to return an identical label for all inputs when it is stamped onto the benign inputs. For a classifier that has $N$ categories, the defender constructs $N$ potential trigger. If the suspicious model is clean, the size of the potential trigger is similar. Nonetheless, the size of the potential trigger for the target label in the compromised model will be significantly smaller than other labels. Then Neural Cleanse utilizes a statistical anomaly detection to determine whether the model contains backdoors. Another well-known defence is Network Pruning \cite{liu2018fine}. Since the feature representations of malicious inputs and benign ones are separable, Network Pruning cuts dormant neurons for benign inputs to alleviate the impacts of backdoor attacks.
\begin{figure*}[t]
\centering
\includegraphics[width=7in,clip,trim=0 0 0 0]{fig/fig2}
\caption{The framework of our backdoor attack scheme.}\label{fig_1}
\end{figure*}
\section{Proposed Method}
\subsection{Threat Model}
In this paper, we describe our attack under image classification tasks with $N$ categories. First, we give the details of the scenario of our attack. Nowadays, Machine Learning as a service (MLaaS) is more and more popular. Users (also dubbed as victims in this paper) may not own enough computing resources. They resort to MLaaS to acquire enough computing resources to satisfy their computing requirement. Users upload their training dataset and network structures, and MLaaS returns a well-trained classifier.
\paragraph{Adversary's Capacities.} Attackers can manipulate the process of the training phase. They can alter the label of images and stamp the trigger onto the benign images. However, they cannot change the structure of the classifier which is determined by users.
\paragraph{Adaversary's Goals.} Stealthiness is an overarching requirement in the backdoor attack scenario. If users perceive the backdoor in the classifier, they can discard it and retrain a new classifier. The stealthiness of backdoor attacks should be considered from two perspectives: trigger stealthiness and compromised classifier stealthiness. Another important goal of backdoor attacks is attack effectiveness. Attackers aim to implant the backdoor into the classifier without degrading the performance of the classifier over the benign inputs, i.e., the performance of the clean classifier and compromised one for the benign inputs is as identical as possible. Finally, the attacker hopes that the attack success rate is as high as possible, i.e., the possibility that a compromised classifier returns the target label when the trigger appears.
\subsection{Attack Overview}
Fig. \ref{fig_1} illustrates the framework of our attack. We consider the stealthiness of the backdoor attack from two perspectives: trigger stealthiness and compromised model stealthiness. For the first part, we employ a U-Net-like \cite{ronneberger2015u} generator to obtain a pair of $\pm$1 modification probability matrices (also can be named as the parameters of the multinomial distribution). Then, we use a SampleNet based on a simple MLP (Muti-Layer Perception) network to sample a concrete trigger for the benign image. The elements of the trigger are only $-$1, $+$1 or 0, which is hard to be perceived by humans and statistical detection. Then we change the label of the malicious image like previous studies \cite{li2020backdoorinvisible}. For model stealthiness, we design a feature representation entanglement algorithm to ensure the feature representations of malicious images are not separable from the clean ones.
\subsection{Stealthiness of Input Space}
In this subsection, we describe the design of the trigger in detail. As aforementioned input-based defences, we hope that the number of changed pixels and the modification magnitude in the benign image is as little as possible. The less modification, the more stealthy the trigger is. Therefore, we set the elements of the trigger as $-1$, $+1$, or 0, that is, the maximum modification magnitude is 1. This trigger also can be seen as a sample that randomly samples from a multinomial distribution which has three possibilities $-1$, $+1$, and 0. We describe our trigger design in a concrete formula manner. For a benign image $x_{benign}$ whose corresponding label is $y_{ori}$, we use a U-Net like generator (named as $G(\cdot)$) to create the parameters of the multinomial distribution $t\sim PN(t_{+1},t_{-1},t_{0})$. The output of generator is restricted between 0 and 0.5 by $0.5\times sigmoid(\cdot)$ function. There are two parameters in this multinomial distribution, i.e., the possibility of $+$1 and $-$1. Due to the restriction of distribution definition, the possibility of unchanged $t_{0}$ is $1-t_{+1}-t_{-1}$. Then we sample from the multinomial distribution to obtain a concrete trigger and we name this procedure as $S(\cdot)$. The malicious image can be expressed as $S(G(x))$ and their label is changed as target label $y_{tgt}$. A natural question may come here:
\textit{Why do not we employ a generator to create the trigger directly? What are the benefits of creating triggers by sampling from a multinomial distribution?}
The pixels of images are all integers whose range is between 0 and 255. If we employ a generator to create the trigger, the elements of the trigger are float-point numbers. Although we can use the round operation to transfer the float-point numbers into integers, the round error cannot be ignored when the value is small like -1 to 1 in our setting. We resort to the multinomial distribution to circumvent this problem. We employ a generator to determine the parameters of the multinomial distribution which implicitly determine the final trigger.
Sampling is an important procedure in our approach. Sampling a multinomial distribution can be expressed by the following equation
$$
{t_{i,j}} = \left\{ {\begin{array}{*{20}{c}}
{ - 1}&{if}&{{n_{i,j}} < t_{i,j}^{ - 1}}\\
1&{if}&{{n_{i,j}} > 1 - t_{i,j}^{ + 1}}\\
0&{otherwise}&{}
\end{array}} \right., \eqno{(1)}
$$
where $n_{i,j}$ is a random number in the interval of [0,1]. We obtain the trigger by equation (1). However, sampling function equation (1) is a non-differentiable step function. To better conduct the back-propagation algorithm, we use a simple MLP-based network (named as SampleNet) to simulate the equation (1). The structure of SampleNet is shown in Table \ref{tab1}. It is solely trained before training the compromised classifier and its parameters are frozen.
\begin{table}[]
\begin{center}
\begin{tabular}{ccc}
\hline
Layer type & Input channel & Output channel \\
\hline
Full connection + Relu & 3 & 16 \\
Full connection + Relu & 16 & 32 \\
Full connection + Tanh & 32 & 1 \\
\hline
\end{tabular}
\caption{The structure of simulation for equation (1).}
\label{tab1
\end{center}
\end{table}
We define the cost function as equation (2)
$$
L_{cls} = \mathcal{L}(f_{\theta }(x_{benign},y_{ori})) + \mathcal{L}(f_{\theta }(x_{malicious},y_{tgt})), \eqno{(2)}
$$
$$
x_{malicious}=S(G(x_{benign}),n), \eqno{(3)}
$$
where $\mathcal{L}(\cdot )$, $f_{\theta}$, $n$ mean the cross-entropy loss, classifier and a random matrix sampled from a uniform distribution $n$$\sim$$U(0,1)$, respectively.
\subsection{Stealthiness of Feature Representation}
Previous studies show that the feature representations of malicious images and benign ones are separable which results in poor resistance against model-based backdoor defences.
\begin{figure}[b]
\centering
\includegraphics[width=3.2in,clip,trim=70 0 80 0]{fig/fig3}
\caption{The illustration for the usage of feautre represenetation entanglement.}\label{fig_2}
\end{figure}
For previous studies like BadNets, although a compromised classifier returns the target label for both benign image (whose original label is the target label) and malicious image, their feature representations are significantly separable. We aim to make the feature of malicious images entangled with benign ones. We design a regularization item as (5) to achieve the above goal. Fig. \ref{fig_2} depicts the usage of the entanglement regularization.
$$
L_{etg} = (f_{benign}-f_{malicious})^2, \eqno{(5)}
$$
where $f_{malicious}$ is the feature representation of malicious images. $f_{benign}$ is the average of the benign images whose original label is equal to the target label. $f_{benign}$ is an alternative updated after updating the parameters of the generator and classifier. Through the entanglement regularization $L_{etg}$, we make sure that the feature of benign images and malicious images are inseparable.
\subsection{Implement Details}
Thanks to the restriction of the $\pm$1 modification probability matrices, the maximum changed magnitude is 1 in the trigger. We add an extra loss item to further decrease the number of changed pixels. The total number of changed pixels can be expressed as (6),
$$
L_{num} = \sum_{i = 1}^{w} \sum_{j = 1}^{h} (| trigger_{i,j} \vert ), \eqno{(6)}
$$
where $w$ and $h$ are the sizes of the trigger (benign image).
We describe our scheme from the input space to the feature representation space, and the total cost function is expressed as (7)
$$
L_{tot} = L_{cls} + \alpha \cdot L_{etg} + \beta \cdot L_{num}, \eqno{(7)}
$$
where hyperparameter $\alpha$ and $\beta$ controls the balance between cross-entropy loss $L_{cls}$ , entanglement loss $L_{etg}$ and loss $L_{num}$.
In previous studies, they only set one target label in their experiments. However, in our scheme, we conduct our backdoor attack against all labels simultaneously. For instance, there are $N$ categories for the classifier, we use a generator to create $N$ pairs of $\pm$1 modification probability matrix. Then, we obtain $N$ malicious images which are corresponding to the $N$ target label. We also calculate $N$ $f_{benign}$ for each category. Our attack scheme can be seen as an extension of choosing one target label. During the attack phase, attackers can make the compromised classifier return an arbitrary label by using the corresponding trigger.
In the attack phase, attackers feed the benign into the well-trained generator and obtain a pair of $\pm$1 modification matrix (multinomial distribution). Then, attackers use a random matrix to sample the multinomial distribution to obtain the trigger. Note that although attackers may obtain different triggers due to different random matrices, the attack success rate is very similar. During the experiments, we find that the trigger sampled from a random matrix or calculated by the expectation (average) of multiple sampling results achieves a very similar attack success rate.
\begin{table*}[htbp]
\centering
\begin{tabular}{ccccccc}
\toprule
Dataset$\rightarrow$ & \multicolumn{3}{c}{GTSRB} & \multicolumn{3}{c}{CelebA} \\
\hline
\multirow{2}[0]{*}{Aspect$\rightarrow$} & \multicolumn{2}{c}{Effectiveness} & Distortion & \multicolumn{2}{c}{Effectiveness} & Distortion \\
& BA (\%) & ASR (\%) & $L_1$-norm & BA (\%) & ASR (\%) & $L_1$-norm \\
\hline
Standard Training & 98.06 & $\backslash$ & $\backslash$ &79.70 & $\backslash$ & $\backslash$ \\
BadNets & 98.07 & 100 & 0.1954 & 79.06 & 100 &0.2020\\
ISSBA & 98.04& 99.98& 4.9572& 79.10& 99.88& 5.8129\\
Ours-one-target &97.61& 99.87& 0.0073& 79.17& 99.96& 0.0217\\
Ours-all-targets & 97.61& 99.79& 0.0076& 79.17& 99.99& 0.0213\\
\bottomrule
\end{tabular}%
\caption{Experimental results for attack effectiveness. BA and ASR mean the accuray of benign images and attack success rate, respectively.}
\label{tab2}%
\end{table*
\section{Experiment Results}
\subsection{Experimental Setup}
\paragraph{Datasets.} We employ ResNet-18 \cite{he2016deep} as the classifier, which is widely used in previous studies \cite{li2020backdoorinvisible}. We adopt two different datasets including GTSRB \cite{houben2013detection} and CelebA \cite{liu2015faceattributes}. GTSRB is a traffic signal recognition dataset with 43 categories. CelebA dataset contains 40 independent binary attribute labels. We follow the configuration proposed by previous studies \cite{NguyenT21} and choose the top three most balanced attributes including Smiling, Mouth Slightly Open, and Heavy Makeup. These attributes are concatenated to create eight classification categories. All images are resized into 128$\times$128. The number of training samples and test samples are 39209 and 12630 for GTSRB, and 162084 and 40515 for CelebA, respectively.
\paragraph{Baseline Selection.} We compare our attack with BadNets \cite{gu2017badnets} and ISSBA \cite{li2020backdoorinvisible}. BadNets is a well-known backdoor attack and is usually set as a baseline in previous studies. We employ a colourful square (6$\times$6) as the trigger in BadNets. ISSBA is a state-of-the-art invisible backdoor attack, which evades various defences. The trigger of ISSBA is generated by the official implementation released on Github. The backdoor rate of baselines is set as 0.1. We choose the label “0” as the target label for BadNets and ISSBA. Note that our approach can generate $N$ triggers for each label simultaneously (named as Ours-all-targets). We also show the result of choosing the label “0” as the target label (named as Ours-one-target).
\paragraph{Training Details.} The batch size and learning rate are set as 16 and 1e-3, respectively. The hyperparameter $\alpha$ and $\beta$ are set as 0.3 and 0.1, respectively. We keep $\alpha$ unchanged during the training process. We multiply $\beta$ by 2 every 20 epochs. The total epochs are 110 and 50 for GTSRB and CelebA, respectively. We adopt Adam optimizer and all experiments are conducted with Pytorch 1.10 version with an NVIDIA RTX3090.
\begin{figure}[b]
\centering
\includegraphics[width=3.2in,clip,trim=220 240 270 160]{fig/fig4}
\caption{Attack success rate for each label.}\label{fig_3}
\end{figure}
\subsection{Attack Effectiveness and Visualization}
For classification tasks, we employ accuracy (test set) as the metric to measure the performance of the compromised classifier. We find that all approaches achieve similar attack effectiveness in Table \ref{tab2}. Actually, most existing backdoor attacks are very similar in the aspect of attack effectiveness. The ASR is very close to 100\% and the accuracy degradation of benign images is less than 1\%. Fig. \ref{fig_3} depicts the concrete ASR over each label, and ASR is more than 98\% in most cases.
In terms of image distortion, our image distortion $L_1$-norm is significantly smaller than baselines. $L_1$-norm is calculated by $sum(abs(x_{benign}-x_{malicious}))/(channel\times height \times width)$. The $L_1$-norm is equivalent to the number of modified pixels since the maximum modification magnitude is only 1 in our approach. Attack effectiveness of ours almost achieves perfect results. Furthermore, we visualize the trigger of ours and two baselines in Fig. \ref{fig_4}. We find that our approach achieves the best visual quality. There is a conspicuous colourful square in the top right corner of the malicious image of BadNets.
To further illustrate the modification of the trigger, we show the histogram of the trigger in Table \ref{tab3}. As depicted in Table \ref{tab3}, the number of changed pixels and magnitude is significantly smaller than the baselines. Such a small modification contributes that our approach can evade state-of-the-art trigger detection. Although the number of changed pixels of BadNets is very small, the magnitude of changed pixels is much larger than ours. Note that the practical value of modification magnitude 1,2 and 3 of BadNets is slightly larger than 0\% (around 0.0016\%). The reason is that there exists a very small proportion of pixels whose original value equals $\pm$1,$\pm$2, or $\pm$3 of the trigger. We omit these results for conciseness.
\begin{figure}[t]
\centering
\includegraphics[width=3.2in,clip,trim=320 220 320 210]{fig/fig5}
\caption{Visual comparison of various attacks.}\label{fig_4}
\end{figure}
\begin{table}[b]
\centering
\begin{tabular}{ccccccc}
\toprule
MM & 0 & 1 & 2 & 3 & 4 & $>=$5 \\
\hline
BadNets & 99.78 & 0 & 0 & 0 & 0 & 0.21 \\
ISSBA & 9.25 & 16.83 & 14.29 & 11.56 & 9.16 & 38.89 \\
Ours & 99.27 & 0.73 & 0 & 0 & 0 & 0 \\
\hline
BadNets & 99.78 & 0 & 0 & 0 & 0 & 0.22 \\
ISSBA & 7.83 & 14.43 & 12.66 & 10.7 & 8.9 & 45.45 \\
Ours & 97.87 & 2.13 & 0 & 0 & 0 & 0 \\
\bottomrule
\end{tabular}%
\caption{The histogram of the modification magnitude of the triggers. MM means the modification magnitude. The values in the table mean the proportion (\%) of corresponding modification magnitude in the total number of changed pixels. The top three rows and the bottom three rows mean the datasets of GTSRB and CelebA, respectively.}
\label{tab3}%
\end{table}%
\subsection{Defences}
\begin{table}[htbp]
\centering
\begin{tabular}{ccc}
\toprule
Dataset & Attack & Acc(\%) \\
\hline
\multirow{3}[0]{*}{GTSRB} & BadNets & 96.03 \\
& ISSBA & 94.76 \\
& Ours & 49.73 \\
\hline
\multirow{3}[0]{*}{CelebA} & BadNets & 99.74 \\
& ISSBA & 85.34 \\
& Ours & 49.9 \\
\bottomrule
\end{tabular}%
\caption{Detection accuray of FTD against various attacks.}
\label{tab4}%
\end{table}%
In this part, we evaluate the resistance of our approaches against multiple backdoor defences. First, We employ state-of-the-art trigger detection FTD \cite{zeng2021rethinking} to scan above attacks. The details of FTD have been introduced in related work. Table \ref{tab4} shows the results of FTD against various backdoor attacks. We can see that FTD is ineffective to our approach but easily detects the other two baselines. The accuracy of FTD against ours is only around 50\% which is equivalent to the random guess for a binary classification task. We only alter less than 1\% pixels with $\pm$1 modification magnitude. The trace of the trigger is too small to be detected by FTD.
Apart from trigger detection, we also employ two model diagnose-based defences: Neural Cleanse \cite{wang2019neural} and Network Pruning \cite{liu2018fine}. Neural Cleanse returns an anomaly index for the suspicious classifier. If the anomaly index is more than 2, the classifier is seen as a compromised classifier. Our approach can bypass Neural Cleanse as shown in Fig. \ref{fig_5}.
\begin{figure}[t]
\centering
\includegraphics[width=3.2in,clip,trim=210 255 210 130]{fig/fig7}
\caption{The experimental results of Neural Cleanse. }\label{fig_5}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=3.2in,clip,trim=240 45 310 35]{fig/fig8}
\caption{The experimental results of Network Pruning. }\label{fig_6}
\end{figure}
Then, we evaluate our attack against Network Pruning. Thanks to the entanglement regularization, our attack performs more resistant than baselines under two different datasets. Network Pruning cannot alleviate the backdoor without decreasing the accuracy of benign accuracy. The experimental results shown in Fig. \ref{fig_6} demonstrate that the accuracy of malicious images is entangled with benign ones. Horizatontal axis means the number of cut neurons.
\subsection{Ablation Studies}
\begin{figure}[t]
\centering
\includegraphics[width=3.2in,clip,trim=265 230 350 162]{fig/fig9}
\caption{The experimental results of removing entanglement regularization. (CelebA). }\label{fig_7}
\end{figure}
In this part, we investigate the impacts of hyperparameter $\alpha$ and $\beta$.
We first set $\alpha$ as 0, i.e., remove the entanglement regularization. Then, we conduct the network Pruning on the compromised model without the entanglement regularization. The experiments are shown in Fig. \ref{fig_7}. We find that the accuracy of malicious and benign can be separated like baselines. Specifically, when we cut 504 neurons, the accuracy of benign images only drops by around 3\%, whereas the accuracy of malicious images only 23.46\% (drops by around 80\%). We also conduct experiments with large $\alpha$ ($\alpha$=0.5 or 1). The results are similar to $\alpha$=0.3.
In terms of hyperparameter $\beta$, it minimizes the number of changed pixels. We initialize $\beta$ as 0.1 in previous experiments. We conduct experiments with a large $\beta$ (more than 0.2) and find that the cross-entropy for malicious images cannot converge.
When we set $\beta$ as 0, the proportion of changed pixels is close to 50\%. FTD also identifies our attack with more than 90\% accuracy.
\section{Conclusions}
In this paper, we propose a novel imperceptible backdoor attack. We analyze the stealthiness of backdoor attacks from input space to feature representation. We elaborate the trigger through the sampling from a multinomial distribution which contains three probabilities +1, -1 and 0. Thanks to the elaborated trigger, we achieve both visual and statistical invisibility. In terms of the feature representation, we design the entanglement regularization to make sure the feature representations of malicious and benign images are inseparable. Extensive experiments demonstrate the effectiveness and stealthiness of our approach.
\section*{Acknowledgements}
This work was supported by the National Natural Science Foundation of China (U20B2051, U1936214).
\bibliographystyle{named}
|
\section{Fitting strategy to determine the flow-based scale}
\label{app:t1_all_flow}
In this appendix we discuss the fitting strategy employed to determine the scales based on the gradient flow and present a detailed account of the extraction of $T_1$.
As mentioned in section~\ref{s:tekflow}, our results show that one can use a universal fitting function to describe the flow-time dependence of the energy density, allowing to extract the scale even in the cases where $T_1$ can only be reached by extrapolation.
The parameterization of the flow-time dependence we have used, to be described below, relies on the connection between the infinite volume flow and the gradient flow renormalized coupling constant $\lambda_{\rm gf}$, c.f. eq.~\eqref{gf-coupling}~\footnote{For the rest of this section and for simplicity we will drop the sub-index ${\rm gf}$ and refer to the gradient flow coupling as $\lambda$.}.
Starting from the renormalization group equation defining the $\beta$ function and integrating it between two different reference scales $t_s$ and $t$ one arrives at:
\begin{equation}
\int_{\lambda(t_s)}^{\lambda(t)} \frac{dx}{\beta(x)} = \frac{1}{2} \log \left ( \frac{t_s}{t} \right).
\label{betaint}
\end{equation}
The left hand side of this equation can be easily integrated using a general parameterization of the $\beta$-function
inspired by the Novikov-Shifman-Vainshtein-Zakharov (NSVZ) $\beta$-function~\cite{Novikov:1985rd}:
\begin{equation}
\beta(\lambda) = - \frac{b_0 \lambda^2}{1 - \sum_{k=0}^{n_b} a_k \lambda^{k+1}},
\end{equation}
with coefficients chosen to reproduce the universal expansion of the $\beta$-function to second order in $\lambda$, c.f.
$a_0$ is set to $b_1/b_0$, where $b_0$ and $b_1$ stand for the first two universal coefficients of the $\beta$-function of the $\mathcal{N}=1$ SUSY Yang-Mills theory, c.f. eqs.~\eqref{eqn:b0pietro},~\eqref{eqn:b1pietro} with $N_f=1/2$.
\begin{table}
\centering
\begin{tabular}{c c c c c c }
\toprule
$b$ & $\kappa_a$ & $T_1 (N=289)$ & $T_1 (N=361)$ &$T_1^0 (N=289)$ & $T_1^0 (N=361)$ \\
\midrule
0.340 & 0.185000 & 2.878(2) (46) & 2.883(2) (58) &-- & --\\
0.340 & 0.187500 & 3.209(3) (68) & 3.188(3) (90) & 3.2014(12) & 3.1818(10)\\
0.340 & 0.189000 & 3.514(3) (65) & 3.488(3) (88) & 3.5108(20) & 3.4785(13)\\
0.340 & 0.191000 & 4.049(4) (35) & 4.042(4) (58) & 4.0484(81) & 4.0420(44)\\
0.340 & 0.192067 & 4.568(6) (58) & --&-- &-- \\
0.340 & 0.193000 & 5.273(7)(160) & 5.244(8)(175) &-- & -- \\
\midrule
0.345 & 0.180000 & 3.166(3) (74) & 3.145(2) (93) & 3.1737(10) & 3.1639(7)\\
0.345 & 0.184000 & 3.664(4) (77) & 3.645(3) (98) & 3.6698(33) & 3.6552(19) \\
0.345 & 0.186800 & 4.294(5) (40) & 4.274(5) (55) & 4.2804(65)& 4.2709(44) \\
0.345 & 0.189600 & 5.559(8)(154) & 5.614(8)(201)& -- & -- \\
\midrule
0.350 & 0.177500 & 3.730(3) (97) & 3.737(3)(105) & 3.7409(36)& 3.7536(20) \\
0.350 & 0.180000 & 4.109(4) (75) & 4.003(4)(109) & 4.0898(59)& 4.0116(37) \\
0.350 & 0.182500 & 4.634(6) (70) & 4.516(5) (64) & 4.5889(50)& 4.4948() \\
0.350 & 0.185000 & 5.364(8)(144) & 5.323(7) (65) & -- & 5.3155()\\
0.350 & 0.186378 & 5.909(9)(191) & -- & --&-- \\
0.350 & 0.187500 & 6.608(11)(249) & 6.582(12)(287)& --& -- \\
\midrule
0.360 & 0.176000 & 5.77(1)(32) & -- &-- &-- \\
0.360 & 0.178000 & 6.32(1)(31) & -- &-- & --\\
0.360 & 0.180000 & 7.05(1)(31) & -- &-- & --\\
0.360 & 0.182000 & 7.85(2)(35) & -- &-- & --\\
0.360 & 0.183172 & 8.68(2)(44) & -- &-- &-- \\
0.360 & 0.184000 & 9.16(2)(52) & -- &-- & --\\
\bottomrule
\end{tabular}
\caption{Values of the scale $T_1$ determined from the flow data with $N=289$ and $N=361$. Results labelled as $T_1$ come from a fit to eq.~\eqref{fit-pert} with 3 free parameters of the $\beta$-function on top of the two universal ones, as described in the text. Values labelled as $T_1^0$ are directly determined by interpolation in the cases in which $T_1$ falls within the scaling window given by eq.~\eqref{fit-window} with $\gamma=0.28$.}
\label{tab:t1_all_flow}
\end{table}
After integration, one arrives at the following identity:
\begin{equation}
{\cal H} (t) \equiv \frac{1}{ \lambda(t)}- \frac{1}{\lambda(t_s) } + a_0 \log \left (\frac{\lambda(t)}{\lambda(t_s)}\right) + \sum_{k=0}^{n_b} \frac{a_k}{k}
\left(\lambda^k (t) - \lambda^k(t_s) \right ) + \frac{b_0}{2} \log \left (\frac{t}{t_s} \right )=0.
\label{fit-pert}
\end{equation}
One can use this identity to fit the numerical results for the flow, leaving as free parameters the non-universal coefficients of the $\beta$-function and the $t_s$ scale.
In the continuum, at infinite volume, and for one masless Majorana fermion, this relation is universal and dictated by the non-perturbative $\beta$-function of the $\mathcal{N}=1$ SUSY Yang-Mills theory in the gradient flow scheme. Departures from universality arise due to lattice artefacts, finite volume effects and the fermion mass-dependence of the flow. Our results, presented in section~\ref{s:tekflow}, indicate that violations of universality remain small, within the numerical accuracy we have been able to attain, as long as one stays within the scaling window given by eq.~\eqref{fit-window}.
To implement the fitting procedure and determine the lattice scale $T_1(b,\kappa_a,N)$, we set $\lambda(t_s)=0.05/\mathcal{K}(\infty)$ and minimize the $\chi^2$ function defined as:
\begin{equation}
\chi^2 = \sum_{T} \left ( \frac{ \mathcal{H} (T) }{\delta\mathcal{H}(T)} \right) ^2,
\end{equation}
where:
\begin{equation}
\delta\mathcal{H}(T) = b_0 \, \delta \lambda (T) \, |\beta|^{-1},
\end{equation}
in a procedure analogous to the one used to parameterize the step scaling function in
ref.~\cite{DallaBrida:2019wur}.
Our final determination of $T_1$ is obtained by fitting simultaneously in this way all our simulations at fixed value of $N$, all values of $b$ and all values of the hopping parameter $\kappa_a$. For the final fit, the $\beta$-function has been parameterized with 3 free coefficients, in addition to the two universal ones. Fits were performed restricting the data for different values of $b$ and $\kappa_a$ to the corresponding scaling window $T\in[1.25, \gamma^2 N/8]$, with $\gamma=0.28$, as shown in figure~\ref{fig:t1_all_flow}.
Our final results are given in table~\ref{tab:t1_all_flow}.
We give separately results of the scale determined from $N=289$ and $N=361$ simulations, whose compatibility serves as a test of the absence of finite size effects.
In addition to the values obtained from the universal fitting function, we also provide the values $T_1^0$ obtained by interpolation in those cases in which the scale falls well within the scaling window. We have assigned a systematic error to the final result that covers for the difference between them as well as for other types of fits and fitting ranges. Among those, we have included fits to eq.~\eqref{fit-pert} with only one and two free parameters and also a second-degree polynomial fit performed in a reduced fitting window corresponding to $\gamma=0.22$. In addition, separate fits have also been performed, including joint fits to all datasets at fixed value of the bare coupling $b$. The final systematic error quoted in the table remains in general below a $3\%$ relative error, going up to about 6$\%$ for the, in physical units, smaller lattices.
\section{Finite size effects in the gradient flow}
\label{ap:flow-imp}
In this appendix we summarize the basic ingredients for the derivation of the improved expression of the flow presented in
eq.~\eqref{improved-phi} and provide also explicit formulas for computing the normalization constant $\mathcal{N}_L (x,N)$ on the lattice.
The perturbative expansion of the infinite volume flow in terms of the ${\overline{\rm MS}}$ 't Hooft coupling
at scale $\mu=1/\sqrt{8t}$ is given by~\cite{Luscher:2010iy}:
\begin{equation}
\Phi_\infty (t,N) = \mathcal{K}(N) \lambda_{\overline{\rm MS}}(\mu) \left( 1 + c_1 \lambda_{\overline{\rm MS}} (\mu) \right)\Big |_{\mu=1/\sqrt{8t}},
\label{eq:pt-flow-inf}
\end{equation}
where:
\begin{equation}
\mathcal{K}(N) = \frac{3 (N^2 -1)}{128 \pi^2 N^2}.
\label{eq:norm-inf}
\end{equation}
where $N$ stands for the number of colours.
This expression is used to define the \textit{gradient flow} (GF) coupling constant~\cite{Luscher:2010iy}:
\begin{equation}
\lambda_{\rm gf} (\mu=1/\sqrt{8t}) \equiv \frac{ \Phi_\infty(t) } {\mathcal{K}(N)}=\lambda_{\overline{\rm MS}}(\mu) \left( 1 + c_1 \lambda_{\overline{\rm MS}} (\mu) \right) ,
\end{equation}
showing that $c_1$ stands for the finite one-loop renormalization constant that relates the $\Lambda$ parameters in the two schemes.
One can derive an analogous expression for the finite volume flow. In our case, on a box of size $l^4$ with twisted boundary conditions as the ones used in this work, one obtains at second order in the coupling~\cite{Bribian:2019ybc}:
\begin{equation}
\Phi(t,l, N) = \mathcal{N} \left(c(t), N\right) \lambda_{\overline{\rm MS}}(\mu) \left \{1 + c_1 \lambda_{\overline{\rm MS}}(\mu) + \mathcal{C}\left (c(t), N \right) \lambda_{\overline{\rm MS}}(\mu) \right \}|_{\mu=1/\sqrt{8t}},
\end{equation}
where finite volume effects depend on the dimensionless variable:
\begin{equation}
c(t)=\frac{\sqrt{8 t}}{\tilde{l}}, \quad \tilde{l}\equiv \sqrt{N}l,
\end{equation}
and are encoded in the functions $\mathcal{N} (c(t), N)$ and $\mathcal{C}(c(t), N )$.
The former has a simple expression in terms of Jacobi $\theta_3$ functions
and reads:
\begin{equation}
\mathcal{N} (x,N) = \frac{3x^2}{128} \left ( \theta_3^4(0,i \pi x)- \theta_3^4(0,i \pi N x) \right),
\end{equation}
with:
\begin{equation}
\theta_3(0,ix) = \sum_{m\in\mathbb{Z}} e^{-\pi x m^2} \equiv \frac{1}{\sqrt{x}} \sum_{m\in\mathbb{Z}} e^{-\pi m^2/x}
.
\end{equation}
In the large volume limit, taken by sending $l$ to infinity at fixed flow time, i.e. by sending $c(t)$ to zero, this normalization factor tends to the infinite volume one, c.f. eq.~\eqref{eq:norm-inf}.
The function $\mathcal{C}(c(t), N )$ goes to zero in this limit and the infinite volume expansion of the flow presented above is recovered.
Let us also mention that, on account of volume independence, the same expressions can be obtained for $SU(\infty)$
by taking the $c(t)$ to zero limit in a different way, i.e. by sending $N$ to infinity at fixed torus size $l$ (in the particular case of the one-point lattice, $l=a$). In that limit, the explicit $N$ dependence of the normalization $\mathcal{K}( N )$ factor disappears.
This exercise indicates a simple way to correct the flow at leading order in the coupling.
We introduce the quantity
\begin{equation}
\hat {\lambda}(t,l,N) = \frac{1}{\mathcal{N}(c(t),N)} {\Phi}(t,l,N),
\end{equation}
which, at leading order in the coupling, has the correct perturbative expansion
and we define a modified flowed energy density given by:
\begin{equation}
\hat {\Phi}(t,l,N) = \mathcal{K}(N) \hat {\lambda}(t,l,N). \end{equation}
This removes finite size effects at tree level. At second order in $\lambda$, the remnant volume dependence comes from the $c(t)$ dependence of the function $\mathcal{C}\left (c(t), N \right)$, which has not yet been computed in perturbation theory for the case envisaged in this work (with a dynamical Majorana fermion and the symmetric twist). As discussed in section~\ref{s:analysis-flow}, within the numerical accuracy we have achieved, this correction is very small for values of $c(t)\lesssim 0.3$.
On the lattice, an analogous correction, which in addition takes into account lattice artefacts, is derived by computing the tree level factor $\mathcal{N}(c(t),N)$ in lattice perturbation theory. For our choice of the clover discretization of the energy density and
for the one-point lattice~\cite{GarciaPerez:2014azn}:
\begin{equation}
\mathcal{N}_L(x,N) = \frac{x^4}{128} \sum_{\mu \ne \nu} \sum'_q e^{- \frac{N x^2}{4} \widehat q^2}
\, \, \sin^2 (q_\nu) \cos^2 (q_\mu/2) \, \frac{1}{\widehat q^2},
\label{lattnorm}
\end{equation}
where the lattice momentum is given by $ \widehat q_\mu = 2 \sin(q_\mu/2)$, with $q_\mu$ taking values:
\begin{equation}
q_\mu = \frac{2 \pi m_\mu}{\sqrt{N}}\, ,
\end{equation}
for $m_\mu = 0, \cdots, \sqrt{N}-1$. The prime in the sum excludes the zero momentum mode, with $m_\mu = 0$, $\forall \mu$.
With all this, our final formula for the discretized version of the flow is given by eq.~\eqref{improved-phi}.
\section{Simulation Parameters and Statistics}
\label{app:simulation}
In this appendix we collect extra information regarding the
simulation in addition to the one described in
section~\ref{s:generationofconfs}. In particular, the full set of model
parameters and the number of configurations generated for them are
collected in table~\ref{tab:ParameterAndStatistics}. The hyper-parameters of the RHMC algorithm and the statistics for each ensemble
are also included in the table.
Each configuration is separated with five trajectories with a
trajectory length of $\tau= 1$.
The number of the MD time steps in the generalized multiple-time step
integrator is denoted by $N_{\mathrm{G}}$ for the gauge action,
and $N_{\mathrm{UV}}$ ($N_{\mathrm{IR}}$) for the UV (IR) part of the
pseudo-fermion action, respectively.
Table~\ref{tab:HMCAccAndLmin} shows the HMC acceptance rate,
the expectation value of the plaquette, and
the expectation value $\qty|\lambda_{\mathrm{min}}|$, with $\lambda_{\mathrm{min}}^2$ the minimum eigenvalue
of $Q_W^2=(\gamma_5 D_W)^2$.
The choice of
hopping parameters have been performed having in mind the ultimate goal
of taking the massless limit. Thus, our $\kappa_{a}$ values
approach the critical value from below. On the opposite edge, for small
hopping parameter the theory becomes equivalent to the TEK model.
This model exhibits a first order phase transition at a certain
$b_c \sim 0.346$ for $(N,k)=(169,5)$ and below this point one enters a strong coupling region of
the model. Indeed, this phase transition extends into a first order phase transition line
in the current model in the $b$-$\kappa_{a}$ plane.
Hence, we first surveyed the phase diagram for the weak coupling region from which the proper continuum limit can be approached,
and the model parameters in the table are chosen to stay in the weak coupling region.
The RHMC algorithm program was primarily written in Fortran 2003/2008 language and uses OpenMP thread parallelization.
To speed up the generation of configurations for the cases with $N=361$, the pseudo-fermion force computation in the MD evolution
is offloaded to the corresponding GPU kernel written in the CUDA language.
To further speed up the configuration generation we made several replica with the same parameter set
but with different random number sequence so that several computational resources are concurrently
available on multi-node systems. We discarded the first 200--500 trajectories in each replica for the thermalization,
where the seed configurations are chosen in such a way that the system remains in the weak coupling region avoiding the phase transition.
\begin{center}
\begin{longtable}{cccccc}
\hline
$(N,k)$ &
$b$ &
$\kappa_{a}$ &
$(N_R, N_{\mathrm{split}}, a, b)$ &
$(N_{\mathrm{G}},N_{\mathrm{UV}},N_{\mathrm{IR}})$ &
{\# of configs} \\
\hlin
(361,7) & 0.350 & 0.1775000 & ( 12, 5, 0.001700, 4.630) & ( 200, 100, 90) & 600 \\
& & 0.1800000 & ( 12, 5, 0.000700, 4.700) & ( 200, 100, 90) & 600 \\
& & 0.1825000 & ( 14, 5, 0.000500, 4.780) & ( 200, 100, 90) & 600 \\
& & 0.1850000 & ( 12, 6, 0.000200, 4.890) & ( 200, 100, 90) & 607 \\
& & 0.1875000 & ( 14, 6, 0.000020, 4.960) & ( 200, 100, 90) & 600 \\
& 0.345 & 0.1800000 & ( 10, 5, 0.002000, 4.645) & ( 200, 100, 90) & 600 \\
& & 0.1840000 & ( 10, 5, 0.000800, 4.780) & ( 200, 100, 90) & 712 \\
& & 0.1868000 & ( 10, 6, 0.000300, 4.875) & ( 200, 100, 90) & 600 \\
& & 0.1896000 & ( 12, 6, 0.000030, 4.980) & ( 200, 100, 90) & 600 \\
& 0.340 & 0.1850000 & ( 14, 6, 0.000700, 4.790) & ( 200, 100, 90) & 640 \\
& & 0.1875000 & ( 14, 6, 0.000400, 4.860) & ( 200, 100, 90) & 638 \\
& & 0.1890000 & ( 14, 6, 0.000300, 4.910) & ( 200, 100, 90) & 604 \\
& & 0.1910000 & ( 12, 6, 0.000080, 4.990) & ( 200, 100, 90) & 660 \\
& & 0.1930000 & ( 12, 6, 0.000005, 5.070) & ( 200, 100, 90) & 625 \\
\hlin
(289,5) & 0.360 & 0.1780000 & ( 12, 5, 0.000500, 5.050) & ( 200, 80, 64) & 516 \\
& & 0.1800000 & ( 12, 5, 0.000200, 4.800) & ( 200, 80, 64) & 800 \\
& & 0.1760000 & ( 12, 5, 0.000800, 4.700) & ( 200, 80, 64) & 600 \\
& & 0.1820000 & ( 12, 5, 0.000080, 4.850) & ( 200, 80, 64) & 720 \\
& & 0.1831720 & ( 12, 5, 0.000100, 4.863) & ( 200, 80, 64) & 600 \\
& & 0.1840000 & ( 12, 5, 0.000020, 4.900) & ( 200, 80, 64) & 800 \\
& 0.350 & 0.1800000 & ( 12, 5, 0.001000, 4.700) & ( 200, 80, 64) & 680 \\
& & 0.1850000 & ( 12, 5, 0.000150, 4.900) & ( 200, 80, 64) & 600 \\
& & 0.1775000 & ( 12, 5, 0.001900, 4.620) & ( 200, 80, 64) & 600 \\
& & 0.1825000 & ( 12, 5, 0.000500, 4.800) & ( 200, 80, 64) & 600 \\
& & 0.1863780 & ( 12, 5, 0.000100, 4.905) & ( 200, 80, 64) & 600 \\
& & 0.1875000 & ( 12, 5, 0.000040, 5.000) & ( 200, 80, 64) & 680 \\
& 0.345 & 0.1800000 & ( 12, 5, 0.002000, 4.650) & ( 200, 80, 64) & 624 \\
& & 0.1840000 & ( 10, 5, 0.000800, 4.785) & ( 200, 80, 64) & 600 \\
& & 0.1868000 & ( 12, 5, 0.000300, 4.880) & ( 200, 80, 64) & 600 \\
& & 0.1896000 & ( 12, 5, 0.000050, 4.980) & ( 200, 80, 64) & 600 \\
& 0.340 & 0.1850000 & ( 10, 5, 0.000900, 4.780) & ( 200, 80, 64) & 610 \\
& & 0.1875000 & ( 10, 5, 0.000400, 4.860) & ( 200, 80, 64) & 600 \\
& & 0.1890000 & ( 12, 5, 0.000200, 4.920) & ( 200, 80, 64) & 608 \\
& & 0.1910000 & ( 12, 5, 0.000080, 4.990) & ( 200, 80, 64) & 600 \\
& & 0.1920670 & ( 12, 5, 0.000050, 5.030) & ( 200, 80, 64) & 610 \\
& & 0.1930000 & ( 12, 5, 0.000005, 5.070) & ( 200, 80, 64) & 630 \\
\hlin
(169,5) & 0.350 & 0.1775000 & ( 10, 3, 0.001000, 4.700) & ( 100, 80, 40) & 600 \\
& & 0.1800000 & ( 10, 3, 0.000800, 4.750) & ( 100, 80, 40) & 600 \\
& & 0.1825000 & ( 10, 3, 0.000400, 4.820) & ( 100, 80, 40) & 600 \\
& & 0.1850000 & ( 10, 3, 0.000150, 4.890) & ( 100, 80, 40) & 700 \\
& & 0.1875000 & ( 10, 3, 0.000080, 4.980) & ( 100, 80, 40) & 600 \\
& 0.340 & 0.1850000 & ( 10, 3, 0.000900, 4.780) & ( 100, 80, 40) & 600 \\
& & 0.1875000 & ( 10, 3, 0.000300, 4.870) & ( 100, 80, 40) & 600 \\
& & 0.1890000 & ( 10, 3, 0.000250, 4.930) & ( 100, 80, 40) & 610 \\
& & 0.1910000 & ( 10, 3, 0.000050, 5.000) & ( 100, 80, 40) & 610 \\
& & 0.1930000 & ( 10, 3, 0.000020, 5.075) & ( 100, 80, 40) & 600 \\
\hlin
\caption{Data sample: For each value of $N$ and twist factor $k$, we list the bare coupling $b$, the adjoint hopping parameter $\kappa_a$, the set of RHMC hyper-parameters used in the generation of configurations and the final number of configurations for each ensemble.}
\label{tab:ParameterAndStatistics}
\end{longtable}
\end{center}
\begin{center}
\begin{longtable}{cccccc}
\hline
$(N,k)$ &
$b$ &
$\kappa_{a}$ &
{HMC acc.} & $\expval{P}$ &
$\expval{\qty|\lambda_{\mathrm{min}}|}$ \\ \hline
(361, 7) & 0.350 & 0.1775000 & 0.7723(98) & 0.561689(73) & 0.049554(65) \\
& & 0.1800000 & 0.766(10) & 0.564610(70) & 0.038933(51) \\
& & 0.1825000 & 0.7475(92) & 0.568020(55) & 0.028445(42) \\
& & 0.1850000 & 0.7400(86) & 0.571769(42) & 0.018722(79) \\
& & 0.1875000 & 0.745(11) & 0.576334(91) & 0.011781(77) \\
& 0.345 & 0.1800000 & 0.726(13) & 0.55072(10) & 0.049524(85) \\
& & 0.1840000 & 0.7340(80) & 0.556355(81) & 0.032503(57) \\
& & 0.1868000 & 0.7137(96) & 0.561253(71) & 0.020603(59) \\
& & 0.1896000 & 0.733(11) & 0.567106(91) & 0.011117(97) \\
& 0.340 & 0.1850000 & 0.7530(85) & 0.541414(99) & 0.040589(71) \\
& & 0.1875000 & 0.7402(91) & 0.546237(83) & 0.029619(64) \\
& & 0.1890000 & 0.720(11) & 0.549479(88) & 0.023011(64) \\
& & 0.1910000 & 0.7248(81) & 0.55419(10) & 0.014393(64) \\
& & 0.1930000 & 0.7341(85) & 0.560145(74) & 0.008346(73) \\
\hline
(289, 5) & 0.360 & 0.1760000 & 0.8060(78) & 0.583233(62) & 0.039755(62) \\
& & 0.1780000 & 0.814(10) & 0.585296(89) & 0.031968(77) \\
& & 0.1800000 & 0.8057(75) & 0.587629(60) & 0.024920(65) \\
& & 0.1820000 & 0.795(11) & 0.590255(64) & 0.019453(67) \\
& & 0.1831720 & 0.7820(85) & 0.591912(75) & 0.017702(68) \\
& & 0.1840000 & 0.8023(63) & 0.593188(51) & 0.017090(80) \\
& 0.350 & 0.1775000 & 0.8040(97) & 0.561674(93) & 0.050168(76) \\
& & 0.1800000 & 0.7726(76) & 0.564575(72) & 0.039760(56) \\
& & 0.1825000 & 0.769(13) & 0.568034(85) & 0.029459(77) \\
& & 0.1850000 & 0.7843(79) & 0.57191(10) & 0.020167(95) \\
& & 0.1863780 & 0.776(11) & 0.57421(10) & 0.016132(79) \\
& & 0.1875000 & 0.7956(74) & 0.576383(73) & 0.01427(11) \\
& 0.345 & 0.1800000 & 0.7917(93) & 0.55045(12) & 0.050209(83) \\
& & 0.1840000 & 0.7477(81) & 0.55636(10) & 0.033087(77) \\
& & 0.1868000 & 0.785(10) & 0.56126(11) & 0.021689(91) \\
& & 0.1896000 & 0.7403(99) & 0.567256(95) & 0.013094(99) \\
& 0.340 & 0.1850000 & 0.7141(84) & 0.541611(88) & 0.040868(72) \\
& & 0.1875000 & 0.752(10) & 0.546422(99) & 0.029998(74) \\
& & 0.1890000 & 0.7638(88) & 0.54981(14) & 0.02343(10) \\
& & 0.1910000 & 0.7347(89) & 0.554363(99) & 0.015480(69) \\
& & 0.1920670 & 0.7793(93) & 0.557248(99) & 0.011992(77) \\
& & 0.1930000 & 0.7803(86) & 0.56039(14) & 0.010557(92) \\
\hline
(169, 5) & 0.350 & 0.1775000 & 0.832(10) & 0.56191(18) & 0.05322(12) \\
& & 0.1800000 & 0.8357(72) & 0.56479(16) & 0.04353(10) \\
& & 0.1825000 & 0.845(10) & 0.56834(15) & 0.034286(94) \\
& & 0.1850000 & 0.8503(68) & 0.57194(14) & 0.02676(11) \\
& & 0.1875000 & 0.8333(86) & 0.57680(12) & 0.022537(83) \\
& 0.340 & 0.1850000 & 0.8060(78) & 0.54165(23) & 0.04309(15) \\
& & 0.1875000 & 0.8367(85) & 0.54666(23) & 0.03270(17) \\
& & 0.1890000 & 0.8292(80) & 0.55007(19) & 0.02691(11) \\
& & 0.1910000 & 0.8085(91) & 0.55495(17) & 0.020784(92) \\
& & 0.1930000 & 0.8447(74) & 0.56003(28) & 0.017592(78) \\
\hline
\caption{For each value of $N$, $b$ and $\kappa_a$ we list the HMC acceptance rate, the expectation value of the plaquette $\expval{P}$, and $\expval{\qty|\lambda_{\mathrm{min}}|}$, with $\lambda^2_{\mathrm{min}}$ the lowest eigenvalue of $Q_W^2= (D_W \gamma_5)^2$.}
\label{tab:HMCAccAndLmin}
\end{longtable}
\end{center}
\section{Meson correlators in the reduced model}
\label{app:meson_correlator}
In this appendix we give technical details about computing
meson correlators in both the
fundamental and adjoint representation and their subsequent use to
obtain the masses of the lowest state with the corresponding quantum
numbers. The underlying ideas and derivations are given in the text
and references supplied.
For the case of fundamental representation fermions there have been
extensive studies done previously~\cite{Gonzalez-Arroyo:2015bya,Perez:2020vbn} using the same techniques
applied in the paper. For that reason we will mostly focus on the adjoint
representation formulas.
The meson correlation functions are expectation values of products of
fermion bilinears separated in time and averaged over space. The correlation function
is computed as a Fourier transform in time
\begin{equation}\label{eqn:correlatorij}
\mathcal{C}_{AB}^{i j}(n_0)= \sum_{q_0} \eu^{-\iu q_0 n_0 }
\expval{\Tr\qty[ \mathbf{O}_A^{(i)} D_W^{-1}(q_0) \mathbf{O}_B^{(j)}
D_W^{-1}(0)]},
\end{equation}
where the trace is over spin, space and colour degrees of freedom.
In this work we take the time period to be $2\sqrt{N}$ so that the
temporal momentum takes values $q_0= \pi m/\sqrt{N}$ with integer $m$.
The symbols $A$ and $B$ specify the spin-parity quantum numbers of the
operator and the indices $i$ and $j$ run over a family of operators
with the same quantum numbers. The symbol $D_W^{-1}(0)$ denotes the
inverse of the Wilson-Dirac operator of the corresponding fermion (fundamental or adjoint).
For the reduced model this inverse greatly simplifies. In the case of
the adjoint the operator is simply the one appearing in
eq.~\eqref{eq:AdjWDmatrix}. The inversion is performed using the BiCGStab algorithm or the Conjugate Gradient algorithm whenever the former does not converge.
One can define $D_W(p)$ as the corresponding
operator with the substitution
\begin{equation}
U_\mu^{\mathrm{adj}} \longrightarrow U_\mu^{\mathrm{adj}} e^{i p_\mu}.
\end{equation}
These operators and their inverses
can be used to allow these valence fermions to live in an arbitrary
lattice. One has simply to average over the
corresponding family of Fourier modes $p$
\begin{equation}
\mathcal{C}_{AB}^{i j}(n_0)= \sum_{q_0} \eu^{-\iu q_0 n_0 }
\frac{1}{|\Lambda_p|}\sum_{p \in \Lambda_p}
\expval{\Tr\qty[ \mathbf{O}_A^{(i)}(p) D_W^{-1}(\vec{p},p_0+q_0)
\mathbf{O}_B^{(j)}(p)
D_W^{-1}(p)]}.
\end{equation}
If $|\Lambda_p|$, the number of elements in $\Lambda_p$, is large, this
might imply many inversions. In practice, what we do is to perform this
average stochastically: For each configuration we generate $p$ randomly
and use it to perform the inversion of $D_W(p)$.
For the case of fundamental fermions things are very similar except for
the expression of the Wilson-Dirac operator. We refer the reader to
the literature for further details~\cite{Gonzalez-Arroyo:2015bya,Perez:2020vbn}.
Now we have to specify the selection of operators $\mathbf{O}_A^{(i)}$
used to obtain the masses. This is based in applying Wuppertal
smearing~\cite{Gusken:1989ad,Gusken:1989qx,Bali:2016lva} to the bilinear operator combined with
three-dimensional APE smearing to the link variables~\cite{APE:1987ehd}. The particular
application to the reduced model in the adjoint representation
is explained below.
\subsection*{Wuppertal fermion smearing in adjoint representation}
\label{subsec:wuppertalsmear}
This amounts to replacing the fermion bilinear associated to an
element of the Clifford algebra $\gamma_A$ as follows
\begin{equation}\label{eqn:fsmearing}
\overline{\Psi}\gamma_A\Psi \longrightarrow \overline{\Psi}
\mathbf{O}_A^{(i)} \Psi \equiv
\overline{\Psi}\gamma_A (\tilde{M}(\bm{p})) ^{s_i} \Psi,
\end{equation}
where the single step smearing operator is given by
\begin{equation}
\tilde{M}(\bm{p})= \dfrac{1}{1+6 c}
\qty[\mathbf{I} + c\sum_{k=1}^{3}\qty[\eu^{\iu p_k } \bar{U}^{\mathrm{adj}}_k +
\eu^{-\iu p_k } (\bar{U}^{\mathrm{adj}}_k )^\dag] ],
\label{eq:LapH}
\end{equation}
and the $s_i$ are integers. In this work we have used $c=0.5$ and the following
list of values:
\begin{table}[H]
\centering
\begin{tabular}{ccccccccccc}
\toprule
$i$ & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9\\\midrule
$s_i$ & 0 & 1 & 4 & 16 & 36 & 64 & 100 & 144 & 196 & 256 \\
\bottomrule
\end{tabular}
\end{table}
In the case of adjoint fermion we only used up to 6 operators, while for fundamental ones we used up to 9.
The symbol $\bar{U}_k$ appearing in eq.~\eqref{eq:LapH} corresponds to the
10 times APE-3d smeared link to be explained below.
\subsection*{APE-3d link smearing}
\label{subsec:ape}
This smearing procedure is an iterative one which in our case maps three spatial SU($N$) matrices onto new ones:
\begin{equation}
U_i^{(s)} \longrightarrow U_i^{(s+1)}=
\mathcal{P}\qty[(1-f)U^{(s)}_i +
\frac{f}{4}\sum_{j\neq i }
\qty(z_{ji}^{*} U^{(s)}_j U^{(s)}_i U_j^{{(s)}\dagger} +
z_{ji} U_j^{{(s)}\dagger}
U^{(s)}_i U^{(s)}_j)],
\end{equation}
where $\mathcal{P}$ is an operator projecting on SU($N$). The starting
point of the iteration are the link matrices $U_i^{(0)}=U_i$. In this
paper, for the case of adjoint fermion correlators, we chose $f=0.081$ and stopped after 10 iterations
$\bar{U}_i=U^{(10)}_i$, which is then transformed into the adjoint
representation and replaced in eq.~\eqref{eq:LapH}.
For the adjoint fermion, the projection to SU($N$) matrix is not
required and it is enough to project onto U($N$) as follows
\begin{equation}
\mathcal{P}(W)= W (W^\dag W)^{-1/2}.
\end{equation}
For the case of fundamental meson correlators we applied the same method using $f=0.15$.
\subsection*{GEVP methodology}
We expect the signal to decay in time as an (infinite) sum of exponentials, corresponding to the contribution of the ground state plus other heavier excited states. Given the typical hierarchy in the mass spectrum, we expect the excited states to decay faster than the ground state, whose signal dominates for large-enough time separation. We want to extract the mass of the ground state from an exponential fit in a region where excited states give no systematic contribution. In order to do so, one has to maximize the projection onto the ground state to have a single-exponential decay setting in early on. Fermion smearing in eq.~\eqref{eqn:fsmearing} should provide an improvement in this regard since it increases the overlap onto the ground state wave function. Beyond that, we make use of a variational approach, in the same philosophy of ref.~\cite{Perez:2020vbn}. Using the definition of the correlator matrix given in eq.~\eqref{eqn:correlatorij}, where the index $i$ and $j$ run over the smearing levels, we applied the so-called GEVP method.
Given two timeslices $\tau_0$ and $\tau_1$ (with $\tau_0<\tau_1$) (not to be confused with gradient flow scales) we solve numerically the generalized eigenvalue problem (GEVP)
\begin{equation}
\mathcal{C}^{ij}(\tau_1) {\bm v}_j^{(n)} = \lambda^{(n)} \mathcal{C}^{ij}(\tau_0) {\bm v}_j^{(n)} \quad\text{at fixed $\tau_0, \tau_1$},
\end{equation}
where ${\bm v}^{(n)}$ and $\lambda^{(n)}$ are the eigenvectors and the eigenvalues, respectively for a given choice of $\tau_0$ and $\tau_1$ (in order to lighten the notation, we also omitted the indices $A$ and $B$ that appeared in eq.~\eqref{eqn:correlatorij}). In this work we used $\tau_0=a$ and $\tau_1=2a$. Among the basis composed by the eigenvectors {${\bm v}^{(n)}$}, we choose the one whose corresponding eigenvalue is the biggest among the others. Let us denote this maximum eigenvector with ${\bm v}^\text{max}$, which we use to define an \textit{optimal} operator by rotating the original correlator-matrix
\begin{equation}
\mathcal{C}_\text{opt}(n_0,\tau_1,\tau_0) = {{\bm v}_i^\text{max}}^* \mathcal{C}^{ij}(n_0,\tau_1,\tau_0) {\bm v}_j^\text{max}.
\end{equation}
The ground state mass is extracted from the exponential decay at large time of this correlator.
\subsection{The lattice spacing and the \texorpdfstring{$\beta$}{beta}-function in the SUSY limit}
After having completed the determination of the scale with good precision for our massive gluino values,
we will here attempt achieving our main goal of determining the scale for the supersymmetric theory. This will be done by extrapolation of the scale to the massless-gluino limit. Given that we have used different scale setting observables and units which are mutually compatible, we will here focus on the flow unit $\sqrt{8t_1}$ since it covers all our simulation points and is both precise and relatively insensitive to finite $N$ corrections. Hence, we extrapolate the lattice spacing $a$ in those units to the limit of vanishing $m_\text{pcac}$ always expressed in physical units. The resulting plot is shown in figure~\ref{fig:chirallimit}.
In the plot, different markers were used to display the points belonging to $N=289$ and $N=361$. By looking at the plot, it is visible by eye that the points corresponding to different markers are compatible within errors, showing that we were able to control finite-volume effects.
The points are well fitted by a straight line and no higher-order polynomial terms are necessary to perform the extrapolation, which is remarkable, taking into account the wide range of $m_\text{pcac}$ values covered.
\begin{table}[t]
\centering
\begin{tabular}{lcccl}
\toprule
$b$ & $\frac{a^\chi}{\sqrt{8t_1}}$ & $\frac{a^\chi}{\sqrt{8t_0}}$ & $\frac{a^\chi}{w_0}$ & $P(b)$ \\
\midrule
0.34 & 0.1702(31) & 0.1046(37) & 0.2904(72) & 0.5620(3) \\
0.345 & 0.1423(46) & 0.0875(39) & 0.2428(89) & 0.57027(15) \\
0.35 & 0.1292(35) & 0.0794(33) & 0.2205(71) & 0.57810(15) \\
0.36 & 0.1048(45) & 0.0644(34) & 0.1788(83) & 0.5934(1) \\
\bottomrule
\end{tabular}
\caption{Lattice spacing in units of $\sqrt{8t_1}$ for the supersymmetric theory. For comparison with other authors we also convert to $\sqrt{8t_0}$ and $w_0$ units using the conversion factor determined in the previous section table~\ref{tab:flow_scales}. The last column displays the plaquette expectation value extrapolated to the massless gluino limit.}
\label{tab:achi}
\end{table}
The values of the scale extrapolated to the massless-gluino limit are reported in table~\ref{tab:achi} for different values of the gauge coupling $b$. Having these values greatly simplifies future studies in both selecting the parameters in which to simulate and in expressing the results in physical units.
Our results provide the value of the lattice spacing as a function of the inverse lattice 't Hooft coupling $\lambda$. This is precisely the dependence that follows from the $\beta$-function of the theory as given by the following formula:
\begin{equation}\label{eqn:loga}
\log{a(\lambda)} + \log\Lambda = -\int^\lambda \frac{\dd x}{\beta(x)},
\end{equation}
where $\Lambda$ is just an integration constant. Thus, by assuming some functional form for the $\beta$-function, integrating it and comparing it with our data values we can determine the parameters appearing in this functional form. This $\beta$-function is the one describing the bare coupling and depends on the discretization procedure, and on the definition of this bare coupling. However, perturbation theory predicts the leading behaviour of the function up to next to leading order. For Yang-Mills theory coupled to $N_f$ flavours of adjoint Dirac fermions this gives
\begin{equation}\label{eqn:ptbeta}
\beta(\lambda) = - b_0 \lambda^2 - b_1 \lambda^3 + \order{\lambda^4} = -\dv{\lambda}{\log{a}},
\end{equation}
with the coefficients $b_0$ and $b_1$ given by
\begin{align}
b_0 &= \frac{1}{8\pi^2}\frac{11-4N_f}{3} \label{eqn:b0pietro}, \\
b_1 &= \frac{1}{128\pi^4}\frac{34-32N_f}{3} \label{eqn:b1pietro}.
\end{align}
There is one particular scheme in which the $\beta$-function is known to all orders. This is the so-called Novikov-Shifman-Vainshtein-Zakharov (NSVZ) $\beta$-function~\cite{Novikov:1983uc}:
\begin{equation}
\beta(\lambda) = -\frac{b_0\lambda^2}{1-\frac{b_1}{b_0}\lambda}.
\end{equation}
This functional form is particularly well-suited for performing the integration of the inverse of the $\beta$-function which is what is needed for eq.~\eqref{eqn:loga}. One gets
\begin{equation}\label{eqn:logapar}
-\log{a} = \frac{1}{b_0\lambda} + \frac{b_1}{b_0^2}\log{\lambda} + \log\Lambda.
\end{equation}
Thus, although we are certainly not in NSVZ scheme, this formula incorporates nicely the universal part of the $\beta$-function and provides a suitable parametrization which allows adding extra terms proportional to the coupling and to higher powers of it.
Now let us apply these ideas to our data. It is well-known that for lattice QCD the naive lattice coupling $\sfrac{1}{b}$ is not particularly well-suited for comparison with the perturbative predictions of scaling. Different authors have proposed improved couplings that do behave much better in this respect. We will then consider these same definitions for the Supersymmetric Yang-Mills theory. One possible choice is the one given in ref.~\cite{Allton:2007py}:
\begin{equation}
\lambda_I = \frac{1}{b P(b)} \label{eqn:allton},
\end{equation}
where $P(b)$ represents the average value of the plaquette extrapolated to the massless-gluino limit, whose values are reported in table~\ref{tab:achi}.
Fitting the NSVZ $\beta$-function (with $N_f=\sfrac{1}{2}$) to our 4 data points gives a chi-square per degree of freedom $\chi^2/\#\mathrm{dof}=1.42$, and the single fitted parameter is $(\sqrt{8t_1}\Lambda)^{-1}=45.2(7)$. If we modify the fit to include a higher order term in the $\beta$-function the fit gives a worse $\chi^2/\#\mathrm{dof}$ and the additional parameter comes out compatible with zero. We conclude that our data does not have enough sensitivity to determine modifications to the NSVZ $\beta$-function. However, our data does have sensitivity to the leading coefficients of the $\beta$-function. A two-parameter fit leaving $N_f$ and $\Lambda$ free having $\chi^2/\#\mathrm{dof}=1.38$, gives $N_f=0.31(20)$ ($24\pi^2 b_0=9.76(80)$) to be compared with the perturbative result $N_f=0.5$ ($24\pi^2 b_0=11-4N_f=9$).
One can repeat the procedure with another choice of improved coupling constant, like the one proposed in refs.~\cite{Martinelli:1980tb,Edwards:1997xf}:
\begin{equation}
\lambda'_I = 8(1-P(b)).
\label{eqn:other}
\end{equation}
The one-parameter fit to the NSVZ $\beta$-function gives $\chi^2/\#\mathrm{dof}=1.42$ with a best fit $(\sqrt{8t_1}\Lambda)^{-1}=472(7)$. Again a two-parameter fit leaving also $N_f$ free gives $N_f=0.30(22)$ ($24\pi^2 b_0=9.78(87)$), completely compatible with that obtained for the other improved coupling. In figure~\ref{fig:betafunction} we display side-by-side the logarithm of the lattice spacing as a function of both improved couplings together with the lines corresponding the fits described before.
\begin{figure}
\centering
\hspace*{-1.cm}
\includegraphics[width=\linewidth+2cm]{pictures/conclusions/betafunction.pdf}
\caption{Dependence of the logarithm of the lattice spacing as a function of the improved coupling $\lambda_I$ defined in eq.~\eqref{eqn:allton} (left panel), and $\lambda_I'$ defined in \eqref{eqn:other} (right panel).
Points have been fitted with eq.~\eqref{eqn:logapar} leaving only $\Lambda$ free (solid orange line), and with the same analytical form leaving also $N_f$ as a free parameter (dashed green line).
}
\label{fig:betafunction}
\end{figure}
In summary, we emphasize that the behaviour of the scale in the range explored in our study is certainly not far and even compatible with the dependence predicted by perturbation theory. The data also shows the tendency expected by the addition of an adjoint Majorana fermion since $N_f$ comes out larger than the value 0 corresponding to the pure gauge theory and not incompatible with the value 0.5, expected for the SUSY Yang-Mills theory at asymptotic small values of the coupling. This implies that the leading perturbative coefficient agrees within a $10\%$ with the expected value $b_0=9/(24\pi^2)$.
\subsection{Summary}
The purpose of this paper was to study the large $N$ limit of the $\mathcal{N}=1$ SUSY Yang-Mills theory on the lattice. More precisely our goal was the determination of the dependence of the lattice spacing on the bare coupling constant, a necessary first step in connecting lattice to continuum results. The paper includes dealing with many aspects of a technical nature, although some of them have been moved to appendices to facilitate readability. Also with this idea in mind we consider appropriate to add this short final summary to point to our main results.
We have used redundancy to ensure the robustness of our final conclusions. Hence, we have computed the lattice spacing using three different and rather independent observables. First we did this for all our different simulation runs having various values of the lattice coupling and the gluino mass. The results are collected in table~\ref{tab:6scales}. Since each observable expresses the lattice spacing in its associated unit, compatibility demands that the values are proportional to each other with a proportionality constant given by the ratio of units. This is actually happening as can be deduced from the table or more visually from figure~\ref{fig:comparison}. The proportionality is in principle only exact in the continuum limit, but our results show that this holds for our lattice sample within errors.
The last step is to extrapolate the results to the massless gluino limit giving the desired value of the lattice spacing for each bare lattice coupling (table~\ref{tab:achi}). Remarkably, the resulting dependence of the beta function for two common improved lattice couplings comes close to the prediction of perturbation theory (figure~\ref{fig:betafunction}). This also shows that the range of bare couplings that we have explored is within the precocious scaling regime.
\subsection{Rational Hybrid Monte Carlo algorithm}
To generate the ensemble of $U_\mu$ according to the partition function eq.~\eqref{eq:partitionfunction}
via Markov chain Monte Carlo methods, we have to transform the integrand to be suitable for Monte Carlo simulations taking into account that the
Pfaffian could be negative.
By using the identity
\begin{align}
\qty|\mathrm{Pf}\qty(CD_W)| = \qty|\det\qty[Q_W^2]|^{1/4},
\end{align}
where $Q_W \equiv D_W \gamma_5$, we transform the partition function as follows:
\begin{align}
\mathcal{Z} &= \int \prod_{\mu=1}^{4} dU_{\mu} \, \mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W))
\qty|\det\qty[Q_W^2]|^{1/4} e^{-S_\mathrm{TEK}(U)},
\label{eq:RHMC0}
\end{align}
and employ the RHMC algorithm~\cite{Kennedy:1998cu,Clark:2003na,Clark:2005sq,Clark:2006wq,Clark:2006wp}
for the probability weight:
\begin{align}
P(U)=\qty|\det\qty[Q_W^2]|^{1/4} e^{-S_\mathrm{TEK}(U)}.
\end{align}
Using the pseudo-fermionic integral, the factor $\qty|\det\qty(Q_W^2)|^{1/4}$ is represented by:
\begin{align}
\qty|\det\qty(Q_W^2)|^{1/4} &= \int \dd{\phi} \dd{\phi}^{\dag} e^{-S_Q(U,\phi)},\quad
S_Q(U,\phi) = \Tr\qty[\phi^{\dag} R_{N_R}^{(-1/4)}\qty(Q_W^2) \phi],
\label{eq:pfaction}
\end{align}
where $\phi$ is the pseudo-fermion field in bi-fundamental form for the adjoint representation
and $\Tr$ is the trace over the colour index.
The matrix $R_{N_R}^{(p)}\qty(Q_W^2)$ ($p=-1/4$) in eq.~\eqref{eq:pfaction} is an approximation
for $\qty(Q^2_W)^{p}$
defined through the $N_R$-th order rational polynomial approximation to $x^p$ for a real number $x \in [a,b]$:
\begin{align}
x^p \underset{x\in[a,b]}{\simeq} R_{N_R}^{(p)}\qty(x) \equiv \alpha_0^{(p)} + \sum_{j=1}^{N_R}\dfrac{\alpha_j^{(p)}}{x-\beta_j^{(p)}}.
\end{align}
Using the Remez algorithm, sets of the coefficients $\qty{\alpha_{j=0,\dots,N_R}^{(p)}, \beta_{j=1,\dots,N_R}^{(p)}}$
have been prepared and tabulated for various cases: the power $p=1/8,-1/4$, the dynamic range $b/a$ and the order of the approximation $N_R$.
To generate the pseudo-fermion field $\phi$ from the probability $\exp\qty(-S_Q)$ in eq.~\eqref{eq:pfaction},
we set:
\begin{align}
\phi &= R^{(1/8)}_{N_R}\qty(Q_W^2)\eta,
\end{align}
where $\eta$ is
drawn from a Gaussian distribution $\exp\qty(-\Tr\qty(\eta^{\dag}\eta))$
at the beginning of the molecular dynamics (MD) evolution in the RHMC algorithm.
The traceless condition is imposed as $\Tr\qty(\eta)=0$ and $\Tr\qty(\phi)=0$ to remove the unwanted U(1) contribution.
The sequence of $\qty(Q_W^2 - \beta^{(p)}_j I)^{-1}\phi$, required for the rational polynomial approximation,
is evaluated using the multi-shift conjugate gradient algorithm.
To apply $D_W$ to the pseudo-fermion field $\phi$ in bi-fundamental form, we use
\begin{align}
D_W \phi =
\phi
- \kappa_{a} \sum_{\mu}\qty[ \qty(1-\gamma_\mu) U_\mu \phi U_\mu^{\dag}
+\qty(1+\gamma_\mu) U_\mu^{\dag} \phi U_\mu ].
\end{align}
The coefficients $\qty{\alpha_{j=0,\dots,N_R}^{(p)}, \beta_{j=1,\dots,N_R}^{(p)}}$ in the RHMC algorithm are determined so as to minimize the metric:
\begin{align}
\Delta_R = \underset{x \in [a,b]}{\min{\max{}}}\qty|x^p - R_{N_R}^{(p)}(x)|,
\end{align}
where the interval $[a,b] $ is set by the lowest and highest eigenvalues of $Q_W^2$.
At the beginning and the end of each MD trajectory,
a highly accurate approximation for the pseudo-fermionic field $\eta$ and action $S_Q$
is needed in the HMC Metropolis test to have the correct weight, c.f. eq.~\eqref{eq:pfaction}. For this purpose a set of coefficients satisfying the condition: $\Delta_R < \mathrm{tol}=10^{-14}$, with the smallest $N_R$, is chosen from the table of set-parameters.
The smallest and largest eigenvalues of $Q_W^2$ are
calculated using a thick restart type Lanczos algorithm that is optimized to the matrix $Q_W^2$,
having the property $J\qty(Q_W^2)J^{-1}=\qty(Q_W^2)^{t}$ with $J=C\gamma_5$~\cite{Ishikawa:2020xac}.
For the coefficient set used during the MD evolution, which should be invariant during the period of generation of configurations,
we fix a lower accuracy $\Delta_R < 10^{-5}$--$10^{-8}$
and select a parameter set for ($N_R$, $a$, $b$), with $a$ and $b$ tuned to cover the eigenvalue range determined during the thermalization period.
In addition, we employ the generalized multiple-time step MD integrator scheme~\cite{HAAR2017113,KAMLEH20121993}
with the fourth order Omelyan-Mryglod-Folk scheme~\cite{Takaishi:2005tz,OMELYAN2003272}.
Three different time step sizes are assigned to the gauge action, and to the UV and IR parts of the pseudo-fermion action.
The pseudo-fermion action in eq.~\eqref{eq:pfaction} is separated at an order $N_{\mathrm{split}} \in (1,N_R)$ as
\begin{align}
S_Q(U, \phi)
&= \Tr\qty[\phi^{\dag}\qty(\alpha_0^{(-1/4)}
+ \sum_{j=1}^{N_{\mathrm{split}}-1}\alpha_j^{(-1/4)}\qty(Q_W^2 - \beta_j^{(-1/4)})^{-1})\phi]
\notag\\
&\quad
+ \Tr\qty[\phi^{\dag}\qty( \sum_{j=N_{\mathrm{split}}}^{N_R}\alpha_j^{(-1/4)}\qty(Q_W^2 - \beta_j^{(-1/4)})^{-1})\phi],
\end{align}
where $\beta_j^{(-1/4)}$ are ordered as: $\beta_{N_R}^{(-1/4)} < \dots < \beta_{2}^{(-1/4)} < \beta_{1}^{(-1/4)}< 0$.
This action defines the Molecular Dynamics force.
Since the coefficients $\alpha_{j}^{(-1/4)}$ align as $0< \alpha_{0}^{(-1/4)} < \alpha_{1}^{(-1/4)} < \dots < \alpha_{N_R}^{(-1/4)}$,
we assign the first portion to the IR part and the last one to the UV part,
resulting in a hierarchy in the magnitude of the MD forces~\cite{Clark:2005sq}.
The finest MD time step size is assigned to the gauge action,
the next finer one to the UV part of the pseudo-fermion,
and the coarsest step size to the IR part, respectively.
\newcommand{0.045}{0.045}
\begin{figure}[t]
\centering
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n169b0340k5.png}
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n169b0350k5.png}\\
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n289b0340k5.png}
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n289b0345k5.png}
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n289b0350k5.png}
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n289b0360k5.png}\\
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n361b0340k7.png}
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n361b0345k7.png}
\includegraphics[clip,scale=0.045]{pictures/4/FigProc_eval_all_n361b0350k7.png}
\caption{We display in the complex plane and for a subset of our simulation parameters at various values of $N$, the 100 eigenvalues of the Wilson-Dirac operator $D_W$ closest to $-0.1$. }
\label{fig:complexeigen}
\end{figure}
\subsection{Eigenvalue spectrum and the sign of the Pfaffian}\label{ss:pfaffian}
As mentioned in the previous section and the preceding subsection our
strategy has been to deal with the effect of the sign of the Pfaffian
$\mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W))$
in eq.~\eqref{eq:RHMC0} by means of the reweighting method.
In this context the expectation value of an observable $\expval{O}$ is evaluated with
\begin{equation}
\expval{\mathcal{O}} =\dfrac{\expval{\mathcal{O}\,\mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W))}_{P(U)}}
{\expval{ \mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W))}_{P(U)}},
\end{equation}
where $\expval{\dots}_{P(U)}$ denotes the expectation value with the non-negative weight $P(U)$
which is statistically evaluated with the ensemble generated with the RHMC algorithm.
To compute the sign of the Pffafian for our configurations we make use
of the results of refs.~\cite{Wuilloud:2010wxi,Bergner:2011wf,Bergner:2011zp,Piemonte:2015,Bergner:2015adz},
according to which, the sign can be determined by counting the number of negative real
eigenvalues of $D_W$. To perform this search, we carried out a complete
determination of all the complex eigenvalues of $D_W$ near the origin of the complex plane, using the \texttt{ARPACK} library~\cite{lehoucq1997arpack}.
The shift invert mode with the shift parameter $\sigma=-0.1$, which
computes the extreme eigenvalues of $(D_W - \sigma)^{-1}$,
is used so that the eigenvalues having negative real part can be captured.
Figure~\ref{fig:complexeigen} shows, for all configurations, the 100 complex eigenvalue of $D_W$
closest to $z=-0.1$ in the complex plane.
Given the symmetry of $\qty(CD_W)^{t}=-\qty(CD_W)$, each point represents
a two-folded eigenvalue.
For heavy gluino masses we are not expecting flips of sign of the Pfaffian. Hence, our analysis focused on the lightest adjoint fermion masses at each $b$ and $N$. We did not observe any negative-real eigenvalues in the spectrum. This absence was also checked for several heavier fermion masses. Therefore we conclude that the sign of the Pfaffian is always positive for the model parameters we employed, simplifying the reweighting method by validating the distribution using the absolute value of the Pfaffian. Similar results showing that the negative sign of the Pfaffian are rare even with moderately light adjoint masses has been observed in lattice SUSY models~\cite{Demmouche:2010sf,Ali:2018dnd}.
\subsubsection{Methodology}
So far our discussion of the flow observables in subsection~\ref{s:general-flow} has been in the continuum.
A generalization to the lattice is rather straightforward, one has to select a discretization of the energy density
and of the flow equations. We adopt the clover version of the field strength which, in our one-site reduced lattice,
leads to a discretized version of eq.~\eqref{energydensity} given by:
\begin{equation}
\hat{E} = -\frac{1}{128} \sum_{\mu,\nu}
\Tr\left [ z_{\mu\nu} \left (U_\nu U_\mu U^\dagger_\nu U^\dagger_\mu
+ U_\mu U^\dagger_\nu U^\dagger_\mu U_\nu
+ U^\dagger_\nu U^\dagger_\mu U_\nu U_\mu
+ U^\dagger_\mu U_\nu U_\mu U^\dagger_\mu \right )
- \text{h.c.}\right]^2.
\label{clover}
\end{equation}
As for the flow, the flow time is discretized in units of the lattice spacing as: $t=Ta^2$, where we will from now on use capital letters
to denote lattice, dimensionless quantities. We employ the so-called Wilson flow~\cite{Luscher:2010iy} and integrate the discretized flow
equations by using a 3rd order Runge-Kutta integrator with constant time interval $\Delta T = 0.03125$.
We have checked that changing the time step does not produce any sizeable difference in the integrated flowed observable, resulting only in a different computational cost for different lattice spacings.
In terms of $\hat{E}(T)$, the naive dimensionless flowed energy density can be estimated on the lattice from:
\begin{equation}
\expval{\frac{T^2 \, \hat{E}(T)}{N}}.
\label{naive-phi}
\end{equation}
The lattice equivalent of the normalization factor $\mathcal{N}(c(t),N)$, corresponding to a concrete discretized definition of the energy density, is also easily determined by a
tree level calculation in lattice perturbation theory.
For our choice of the clover discretization and
for the one-point lattice, an explicit expression has been computed in ref.~\cite{GarciaPerez:2014azn} and is provided in Appendix~\ref{ap:flow-imp}.
One advantage of using the lattice determined instead of the continuum norm is that one corrects finite lattice artefacts on top of finite size effects at tree level.
With all this, our final formula for the discretized version of the flow is given by:
\begin{equation}
\hat {\Phi}_L(T,b,N) = \frac{3}{128 \pi^2 \mathcal{N}_L(\sqrt{8T/N},N)} \expval{\frac{T^2 \, \hat{E}(T)}{N}},
\label{improved-phi}
\end{equation}
with $\mathcal{N}_L(x,N)$ given by eq.~\eqref{lattnorm}.
\begin{figure}[t]
\centering
\begin{subfigure}{.5\textwidth}
\centering
\raggedleft
\includegraphics[width=1.\linewidth]{pictures/naive-phi_b34.pdf}
\captionsetup{width=.9\textwidth}
\caption{Naive flow (eq.~\eqref{naive-phi}).}
\label{fig:finiteN_naive}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\raggedleft
\includegraphics[width=1.\linewidth]{pictures/improved-phi_b34.pdf}
\captionsetup{width=.9\textwidth}
\caption{Improved flow (eq.~\eqref{improved-phi}).}
\label{fig:finiteN_imp}
\end{subfigure}
\caption{Finite $N$ effects on the dimensionless flowed energy density. The coloured bands in the right plot represent the scaling windows for each value of $N$, c.f. $ T \in [ 1.25, \gamma^2 N/8]$, with $\gamma=0.28$.}
\end{figure}
In order to illustrate the kind of improvement
attained with the use of eq.~\eqref{improved-phi} we present in figure~\ref{fig:finiteN_naive} the dependence on $T$ of the naive
expression for three different values of $N$, corresponding to $N=169$, 289 and 361.
As can be observed, finite $N$ corrections are a sizable source of systematics.
The plot of figure~\ref{fig:finiteN_imp} displays instead the flowed energy density obtained using the improved observable $\hat \Phi$. In this case, the window in which the three curves collapse to a single one extends over a much larger range.
Generically, this window is set by the ratio $c(t)= \sqrt{8T/N}$, determining the fraction of the box occupied by the smearing radius.
An empirical observation is that the correction is efficient over a
scaling window given by:
\begin{equation}
T \in [ 1.25, \gamma^2 N/8], \text{ with } \gamma\lesssim 0.3,
\label{fit-window}
\end{equation}
where the lower end of the interval is set to have under control the remaining lattice artefacts.
\subsubsection{Results}
\label{s:analysis-flow}
Before presenting our results, there is an additional consideration that has to be made. To have a precise determination of the scale it is important that the value of $T_s\equiv t_s/a$ appearing in the lattice counterparts of eqs.~\eqref{tc} and~\eqref{wc} falls well within the scaling window given by eq.~\eqref{fit-window} and can therefore be reached by interpolation. With our set of parameters this is best attained by choosing $s=0.05$ instead of the commonly used value of $s=0.1$; the corresponding lattice scale will be denoted by $T_1$ from now on. However, on our smaller lattices, particularly those corresponding to $b=0.36$, even this value leads to a scale that can only be reached by extrapolation.
We address this issue by simultaneously fitting our data, for all values of $b$ and hopping parameter $\kappa_a$, to a single universal curve depending on $T/T_1(b,N,\kappa_a)$. Our combined data covers a window that runs from the perturbative small flow-time region up to values around $T_0/T_1$. In playing this game we are neglecting violations of universality that may come from lattice artefacts, remnant finite size effects or a dependence of the flow on the gluino mass. As we will see below, these assumptions are well satisfied by our results.
In appendix~\ref{app:t1_all_flow} we give full details of the form of the universal fitting functional used to describe the flow, here let us just indicate that we have selected it so as to describe appropriately the perturbative domain for small values of $T/T_1$. For that one can use that the infinite volume flow defines a renormalized coupling constant and the small $T/T_1$ regime is therefore well described by using the two-loop perturbative $\beta$-function. As we will see below, this two-loop function describes very well the time dependence of the flow in a remarkably large window of flow times.
\begin{figure}[t]
\centering
\hspace*{-0.8cm}
\includegraphics[width=\linewidth+0.8cm]{pictures/6/fit_flow_N289.pdf}
\caption{We display $\hat \Phi(t)$ as a function of $t/t_1$ for all our datasets with $N=289$. The width of the coloured bands indicates the error on the different quantities. The black line on the plot corresponds to the two-loop infinite $N$ (infinite volume) perturbative prediction for the flow-time dependence of $\hat \Phi(t)$. }
\label{fig:t1_all_flow}
\end{figure}
To serve the purpose of illustrating how well universality holds, we display in figure~\ref{fig:t1_all_flow} the dependence of $\hat \Phi(T)$ on $T/T_1$ for $N=289$.
The different data displayed in the plot correspond to different values of $b$ and $\kappa_a$ restricted to the scaling window eq.~\eqref{fit-window} with $\gamma=0.28$.
The values of $T_1$ have been obtained from a universal fit to the data as the one described by eq.~\eqref{fit-pert} with three parameters of the $\beta$-function, in addition to the two universal ones corresponding to $\mathcal{N}=1$ Supersymmetric Yang-Mills, and gives a $\chi^2$ per degree of freedom of 1.2 (we obtain $\chi^2/\# {\rm dof}=3.1$ for the datasets with $N=361$). As becomes evident from the plot, the advantage of the joint fit is that it allows to constrain the time dependence of the flow in a region of scales much larger than the actual fitting window and permits to determine $T_1$ even when it falls outside it. Finally, we also display for comparison the prediction of two loop perturbation theory at infinite $N$, given by the black line in the plot, which, as mentioned previously, describes quite well our results in a large window of flow times.
Let us now present our results. In order to have an additional check on the effective reduction of finite $N$ effects, we have fitted separately the $N=289$ and 361 data sets, obtaining compatible results within errors. Our final values for $T_1$ are given in table~\ref{tab:6scales}, the first quoted error is statistical and the second systematic. The latter is determined so as to cover various determinations of the scale corresponding to different fitting functional and ranges as detailed in appendix~\ref{app:t1_all_flow}.
We end this section by discussing the relation of the scale $t_1$ to those more common in the literature, as $t_0$ or $w_0$. We have already mentioned that in most of our simulations $t_0$ falls out of the scaling window. Nevertheless, and under the assumption of scaling, we can use the universal fitting functional describing the flow to obtain a determination of the ratio $R=\sqrt{T_0/T_1}$. The result is
$R=1.624(50)$ and $R=1.631(70)$ for $N=361$ and 289 respectively. This is in perfect agreement with the ratios obtained at the few cases where we can determine $T_0$ directly by interpolation. The error quoted in all cases covers for the systematics in the fitting functional and fitting ranges following the same procedure used to determine $t_1$.
Finally, we have also determined the scales $w_1$ and $w_0$ derived by solving the implicit equation eq.~\eqref{wc} with $s=0.05$ and $s=0.1$ respectively. The strategy to determine these scales is very similar to the one used for $t_1$. We rely on the universality of the flow and fit $t d \hat \phi(t)/dt$ as a function of $t/a^2$ simultaneously for all our datasets within the scaling window corresponding to $\gamma=0.28$. In this case we use a degree-two polynomial fit, with the systematic error obtained by varying the fitting range. The resulting scales can be compared to $t_1$.
Excluding the sets at $b=0.36$, which have very large systematic errors and give nevertheless results consistent within errors, and restricting to the cases with $m_{\text{pcac}}\sqrt{8t_1}>0.3$, a fit of the dimensionless ratio of scales to a constant gives: $w_1/\sqrt{8t_1}= 0.4535(49) $ and $w_0/\sqrt{8t_1}=0.586(10) $ with $\chi^2$ per degree of freedom equal to 0.24 and 0.15 respectively.
We collect our final results for the ratio of $\sqrt{8 t_0}$, $w_0$ and $w_1$ to $\sqrt{8t_1}$ in table~\ref{tab:flow_scales}. These ratios can be used to convert all the results given in this section, in particular the values of the lattice spacing as a function of the bare coupling and gluino mass, to the other more standard units used in the literature.
\begin{table}[t]
\centering
\begin{tabular}{ccc}
\toprule
$\sqrt{t_0/t_1}$ & $w_1/\sqrt{8 t_1}$&$w_0/\sqrt{8 t_1}$ \\
\midrule
1.627(50) & 0.4535(49) & 0.586(10) \\
\bottomrule
\end{tabular}
\caption{Different scales determined from the flow expressed in units of $\sqrt{8t_1}$. Scales $t_1(t_0)$ and $w_1(w_0)$ are derived respectively from eqs.~\eqref{tc} and~\eqref{wc} setting $s=0.05(0.1)$. }
\label{tab:flow_scales}
\end{table}
\subsection{Determination from the eigenvalues of \texorpdfstring{$D_W \gamma_5$}{D\_W gamma\_5}}
As already mentioned, the first method that we have used to determine the point of SUSY restoration is based on the eigenvalues of the Dirac matrix. In the continuum limit the minimum eigenvalue of the massive Dirac operator at infinite volume goes to zero linearly with the fermion mass. On the lattice one expects a similar behaviour for the operator $D_W/\kappa_a$ which tends to the continuum one up to a renormalization factor. As explained earlier the lowest lying spectra of $Q_W=D_W \gamma_5$ has been determined for all our configurations. This includes $|\lambda_{\text{min}}|$, whose average value is listed in table~\ref{tab:HMCAccAndLmin} for all our datasets. Notice that this quantity has very small errors, making it a perfect observable for the determination of the zero mass point. Indeed, the aforementioned continuum behaviour implies that $\expval{\qty|\lambda_{\mathrm{min}}|}/\kappa_a$ should depend linearly on $1/(2\kappa_a)$ and vanish at $\kappa_a=\kappa_a^c$, signalling the point of vanishing gluino mass. Since finite $N$ corrections amount to finite volume effects, this behaviour is only expected to hold at infinite $N$. Building on this analogy we would expect a dependence of the following form:
\begin{align}
\qty(\dfrac{\expval{\qty|\lambda_{\mathrm{min}}|}}{\kappa_{a}})^2
&=A \, \qty( \dfrac{1}{2 \kappa_{a}} - \dfrac{1}{2 \kappa_a^c})^2 + \delta(1/N^2).
\label{eq:LminFit}
\end{align}
where the function $\delta(1/N^2)$ should vanish at large $N$. A correction of this type was observed earlier~\cite{GarciaPerez:2015rda} for the large $N$ reduced model coupled to two flavours of adjoint fermions. In our case the formula describes our data very well as
shown in figure~\ref{fig:LminVsKappa}. The dashed lines gives our best fit with $\delta(1/N^2)=B/N^2$ and the dotted line the extrapolation to infinite $N$. The resulting critical hopping parameters $\kappa_a^c$ are presented in table~\ref{tab:kappa_a_c}.
In the next subsection we will present an alternative method for determining the critical hopping values.
\begin{table}
\centering
\begin{tabular}{c|c| c}
\toprule
b & $\kappa_a^c$ from $\expval{\qty|\lambda_{\mathrm{min}}|}$ &
$\kappa_a^c$ from $m_{\text{pcac}}$ \\
\midrule
0.340 & 0.19359(5)& 0.19365(5)\\
0.345 & 0.19095(6) & 0.19100(6)\\
0.350 & 0.18857(5) & 0.18845(2)\\
0.360 & 0.18415(4) & 0.18418(2)\\
\bottomrule
\end{tabular}
\caption{Values of the adjoint critical hopping parameter $\kappa_a^c$ determined from the vanishing, as a function of $1/(2\kappa_a)$, of $\expval{\qty|\lambda_{\mathrm{min}}|}$ or the $m_{\text{pcac}}$ mass.}
\label{tab:kappa_a_c}
\end{table}
\subsection{Determination from the PCAC mass}
An alternative way of extracting the point of SUSY restoration is based on the vanishing of the non-singlet adjoint-pion correlator and the associated PCAC mass. Contrary to what happens in standard lattice simulations, and due to the form in which it is computed in the reduced setup, the signal-to-noise ratio in the adjoint-pion correlation function deteriorates significantly at large time~\cite{Gonzalez-Arroyo:2015bya}. For this reason, it turns out to be more convenient to do the tuning in terms of the PCAC mass which can be determined with very good precision, and this is the choice we have used in this work. As we will see below, the results obtained in this way turn out to be consistent with the less precise determination from the adjoint-pion mass.
\renewcommand{0.045}{0.55}
\begin{figure}[t]
\centering
\includegraphics[scale=0.045]{pictures/4/mpcac_ball.pdf}
\caption{Dependence of the PCAC adjoint fermion mass on $1/(2 \kappa_a)$ for all the values of the bare coupling $b$ and the largest value of $N$ within out set of simulations. The bands are linear fits at fixed value of $b$ used to determine the critical value of the adjoint hopping parameter $\kappa_a^c$, with the width indicating the error in the fit. The $\chi^2$ per degree of freedom of each of the fits is indicated in the legend. }
\label{fig:mpcac-adj}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.045]{pictures/4/mpi_ball.pdf}
\caption{Dependence of the adjoint pion mass on $1/(2 \kappa_a)$. The bands are fits to the data at fixed value of $b$ with the critical hopping parameter fixed to the one extracted from the vanishing of the PCAC mass, as detemined from figure~\ref{fig:mpcac-adj}.}
\label{fig:mpion-adj}
\end{figure}
In this and other sections we will be using correlators of meson operators made of fundamental quarks or adjoint gluinos. The methodology is
simple as it amounts to allowing the fermions to travel in a lattice of varying sizes, including infinite, in the background field of the reduced model gauge field. This has been presented in
ref.~\cite{Gonzalez-Arroyo:2015bya} and used
extensively in ref.~\cite{Perez:2020vbn}. For other representations, like the adjoint, the
philosophy is the same (details will be given elsewhere ref.~\cite{AGA_mesons_new}). Some
specific features about the implementation in this work are contained in appendix~\ref{app:meson_correlator}.
On the lattice, the PCAC mass can be determined from a discretized version of eq.~\eqref{eqn:mpcaccont}. For that, we consider an ultralocal version of the axial vector current and an interpolating operator for the pion with optimized projection onto the pseudoscalar channel ground state, obtained as detailed in appendix~\ref{app:meson_correlator}. The PCAC mass is then determined by fitting to a constant the ratio of correlation functions:
\begin{equation}\label{eq:mpcac}
a m_\text{pcac} = \frac{\mathcal{C} (x_0+a; \gamma_0\gamma_5,\gamma_5^{\rm opt})-\mathcal{C} (x_0-a; \gamma_0\gamma_5,\gamma_5^{\rm opt}) }{4\, \mathcal{C}(x_0;\gamma_5,\gamma_5^{\rm opt} )},
\end{equation}
where $\mathcal{C}(x_0; A,B)$ stands for the two-point correlation function of the operators corresponding to channels $A$ and $B$.
The values of the PCAC masses extracted in this way are provided in table~\ref{tab:mpcac_adj}. We include in the table results for our datasets with $N=289$ and 361. The general agreement, within errors, of the results for the two different values of $N$ gives an indication of the small influence of finite $N$ effects on this quantity.
In the chiral limit the PCAC mass is expected to be proportional to the fermion mass. This implies that close to this limit $m_{\text{pcac} }$ should depend linearly on $1/\kappa_a$ and vanish at the critical value. Indeed, in figure~\ref{fig:mpcac-adj} we display our results, for the datasets with larger value of $N$, together with a fit to a linear function of $1/(2 \kappa_a)$. We see that the straight line provides a very good qualitative description of the data. Given the small errors, an additional quadratic term is needed in some cases to get a fit with $\chi^2$ per degree of freedom of the order of one, if we include also the largest masses. In any case, for the purpose of determining the point of vanishing of the PCAC mass both fits give identical results.
\begin{table}
\centering
\begin{tabular}{c c c c c c }
\toprule
$b$ & $\kappa_a$ & $am_{\text{pcac}} (N\!=\!289)$ & $am_{\text{pcac}} (N\!=\!361)$ &$am_\pi (N\!=\!289)$ & $am_\pi (N\!=\!361)$\\
\midrule
0.340& 0.1850& 0.1957(21)& 0.2018(14) & 0.970 (6)& 0.977 (4)\\
0.340& 0.1875& 0.1426(14)& 0.1424(15) & 0.814 (7)& 0.819 (7)\\
0.340& 0.1890& 0.1041(12)& 0.1083 (7) & 0.699 (5)& 0.719 (4)\\
0.340& 0.1910& 0.0603 (9)& 0.0619 (9) & 0.534 (6)& 0.540 (5) \\
0.340& 0.192067& 0.0344 (8)& -- & 0.397 (4)& -- \\
0.340& 0.1930& 0.0120 (3)& 0.0138 (6) & 0.258 (6)& 0.263(13)\\
\midrule
0.345& 0.1800& 0.2390(60)& 0.2352(32) &1.040(17)& 1.043 (6) \\
0.345& 0.1840& 0.1477(22)& 0.1507(11) &0.814(16)& 0.821 (5) \\
0.345& 0.1868& 0.0901 (7)& 0.0913(18) &0.630 (7)& 0.631 (6) \\
0.345& 0.1896& 0.0275 (5)& 0.0289 (5) &0.349 (6)& 0.353 (5) \\
\midrule
0.350& 0.1775& 0.2363(29)& 0.2353(19) &1.013(12)& 1.001 (7) \\
0.350& 0.1800& 0.1805(25)& 0.1826(16) &0.878 (9)& 0.883 (7) \\
0.350& 0.1825& 0.1242(24)& 0.1242(15) &0.719(13)& 0.733 (6) \\
0.350& 0.1850& 0.0733(14)& 0.0712(10) &0.534 (6)& 0.540 (6) \\
0.350& 0.1864& 0.0427 (9)& -- &0.433 (7)& --\\
0.350& 0.1875& 0.0192 (7)& 0.0195 (4) &0.291 (8)& 0.293 (6) \\
\midrule
0.360& 0.1760& 0.1684 (9)& -- &0.804(7)& --\\
0.360& 0.1780& 0.1227(17)& -- &0.689(15)& --\\
0.360& 0.1800& 0.0837 (8)& -- &0.541(15)& --\\
0.360& 0.1820& 0.0420 (7)& -- &0.370 (7)& --\\
0.360& 0.1832& 0.0202 (6)& -- &0.270(19)& --\\
0.360& 0.1840& 0.0038 (5)& -- &0.249(17)& --\\
\bottomrule
\end{tabular}
\caption{Values of the PCAC and adjoint-pion masses determined for all values of the bare coupling and adjoint hopping parameter in our set of simulations with $N=289$ and $N=361$.}
\label{tab:mpcac_adj}
\end{table}
Our final results for the critical value of $\kappa_a$, corresponding to the largest value of $N$ simulated in each case, are collected in table~\ref{tab:kappa_a_c}. The values obtained are in remarkable agreement with the determination based on the eigenvalues of the Dirac operator. In the rest of this work, the vanishing of the PCAC mass will be used as the criteria to tune all other quantities to the zero gluino mass limit.
Finally, to check for the consistency of our approach, we have as well determined the adjoint-pion mass and analyzed its dependence on the hopping parameter $\kappa_a$. Results are collected in table~\ref{tab:mpcac_adj}.
Figure~\ref{fig:mpion-adj} displays $m_\pi^2$ vs $1/(2 \kappa_a)$. The coloured lines shown in the plot correspond to linear fits where, for each value of the bare coupling $b$, the critical value of the hopping parameter is fixed to the one determined from the PCAC mass; only the slope is left as a free parameter. Linearity works very well in all cases, giving $\chi^2$ per degree of freedom of the order of one, except on the smallest lattice with bare coupling $b=0.36$, where a deviation is observed for the lighter masses (the two lightest ones are excluded from the fit). This difference can be due to finite size effects, which come out more pronounced for the pion mass than for the PCAC mass. For this reason it is also preferable to use the latter in order to determine the point of SUSY restoration.
\subsubsection{Methodology}\label{sssec:methodology}
The meson correlators in time of fundamental quarks are expectation values
\begin{equation}
{\cal C}_{A B}(n_0)= \langle \mathbf{O}_A(0) \mathbf{O}_B^\dagger(n_0) \rangle,
\end{equation}
where $\mathbf{O}_{A}$ and $\mathbf{O}_{B}$ are gauge invariant bilinear quark operators projected to zero spatial momentum. As explained previously, for the purpose of this paper we restrict ourselves to pseudoscalar and vector-meson operators, as for example: $\sum_{\vec{n}}\mathbf{\overline{\Psi}}(n_0,\vec{x}) \, \gamma_5 \, \mathbf{\Psi}(n_0,\vec{n}) $ and $\sum_{\vec{n}}\mathbf{\overline{\Psi}}(n_0,\vec{n}) \, \gamma^\mu \, \mathbf{\Psi}(n_0,\vec{n}) $. However, there are infinitely many operators with the same quantum numbers and this is an essential advantage that is used by us and other researchers. In practice we will be considering spatially non-local operators obtained by applying two well-known algorithms: the Wuppertal smearing of fermion bilinears~\cite{Gusken:1989ad,Gusken:1989qx,Bali:2016lva} and the three-dimensional APE smearing~\cite{APE:1987ehd}. The explicit expressions have been used in previous papers and are recalled in appendix~\ref{app:meson_correlator}. The advantage of using these operators is that their coupling to the lowest mass state in each channel is enhanced with respect to excited states. In any case, it is useful to consider linear combinations of these operators designed to optimize the coupling to the ground state in each channel relative to other low mass states. This variational procedure is by now quite standard and goes by the name GEVP~\cite{Michael:1985ne,Luscher:1990ck}.
As an example of the outcome of the method, we display in figure~\ref{fig:corr_example} the correlators of the optimal operators in the pseudoscalar and vector channels for the case $b=0.35$, $\kappa_a=0.1825$, $\kappa_{f}=0.1525$, with $\kappa_f$ the fundamental hopping parameter. The time-dependence of these correlators shows a clear exponential decay from which the mass
of the $\pi$-meson (pseudoscalar) and $\rho$-meson (vector) can be extracted. The numbers obtained and the $\chi^2$ of the fits to an exponential are also displayed on the figure. We also show the signal that allows one to extract the fundamental PCAC mass $m^f_\text{pcac}$, defined as in eq.~\eqref{eq:mpcac}.
Summing up, the procedure to obtain the final results is the one described in the example. It amounts to determining the optimal operator and then to fit the corresponding correlator in a certain interval to an exponential (rather to a hyperbolic cosine function, taking into account the periodic nature of the temporal direction). There are of course some specific details which reflect the selection of the operator and the choice of fitting interval that affect the final numerical value for the mass. In the end, variations of these types which are of similar statistical significance are accounted for as a systematic error of the determination. The specific details of the GEVP method that we have used will be collected in appendix~\ref{app:meson_correlator}. Here we will comment briefly on the choice of the fitting interval. The main points to be taken into account in this selection are finite volume effects, lattice artefacts and the contamination of excited states. Typically, finite-size effects are more relevant close to the chiral limit, as the pion mass goes to zero and its Compton wavelength becomes comparable to the effective volume. To avoid too severe effects our selected values of the hopping parameter should stay sufficiently far from the chiral limit. In particular in our data the lightest cases still had $m_\pi a \sqrt{N}\sim 3.4$.
Nonetheless, finite-size effects may still reflect in the appearance of a constant term in the correlator arising from the propagation of quarks wrapping around the finite extent of the lattice~\cite{Umeda:2007hy}. Although this effect disappears in the large volume (large $N$) limit, we have observed that in some cases the addition of a small constant to the hyperbolic cosine was required to obtain a good fit.
Finally, we should comment about the lower limit of the fitting interval. A smaller time separation implies a smaller relative error at the expense of a larger contamination of excited states. A balance is then necessary, our choice has been to fix the lower limit of the interval in physical units setting it equal to $R\sqrt{8t_1}$. We have taken in all cases $R=0.95$ except for $b=0.36$ for which, to increase the signal-to-noise ratio, a value $R=0.8$ was preferred. As for the upper limit of the interval, it was set in all cases so as to have at least 6 lattice points in the fitting range.
\begin{table}[H]
\centering
\small
\begin{tabular}{ccllll}
\toprule
$b$ & $\kappa_a$ & $\kappa_f$ & $am^f_\text{pcac}$ & $am_\pi$ & $am_\rho$ \\
\midrule
\multirow{15}*{0.34} & \multirow{4}*{0.189} & 0.15500 & 0.09711(88) & 0.507(12) & 0.549(20) \\
& & 0.15700 & 0.05755(89) & 0.381(17) & 0.434(31) \\
& & 0.15800 & 0.0392(11) & 0.350(14) & 0.421(27) \\
& & 0.15920 & 0.0149(11) & 0.248(39) & 0.336(56) \\
\cline{2-6}
& \multirow{4}*{0.191} & 0.15250 & 0.1253(11) & 0.545(16) & 0.570(21) \\
& & 0.15500 & 0.07464(76) & 0.423(17) & 0.450(29) \\
& & 0.15580 & 0.05929(86) & 0.370(14) & 0.416(22) \\
& & 0.15700 & 0.03601(89) & 0.345(33) & 0.373(54) \\
\cline{2-6}
& \multirow{3}*{0.192} & 0.15380 & 0.08564(87) & 0.430(20) & 0.458(27) \\
& & 0.15550 & 0.05090(88) & 0.368(21) & 0.405(32) \\
& & 0.15650 & 0.03086(75) & 0.336(21) & 0.379(34) \\
\cline{2-6}
& \multirow{4}*{0.193} & 0.15240 & 0.1000(12) & 0.443(21) & 0.450(30) \\
& & 0.15380 & 0.06932(92) & 0.318(24) & 0.317(30) \\
& & 0.15530 & 0.03904(60) & 0.301(21) & 0.299(27) \\
& & 0.15630 & 0.01982(96) & 0.288(34) & 0.283(46) \\
\midrule
\multirow{22}*{0.35} & \multirow{4}*{0.1775} & 0.15000 & 0.1730(11) & 0.701(16) & 0.741(22) \\
& & 0.15250 & 0.11939(90) & 0.558(13) & 0.599(20) \\
& & 0.15500 & 0.06896(73) & 0.425(15) & 0.480(27) \\
& & 0.15625 & 0.04288(45) & 0.328(18) & 0.382(28) \\
\cline{2-6}
& \multirow{4 }*{0.18 } & 0.15000 & 0.15806(93) & 0.608(12) & 0.626(15) \\
& & 0.15250 & 0.10532(64) & 0.488(13) & 0.514(19) \\
& & 0.15500 & 0.05327(78) & 0.343(14) & 0.377(24) \\
& & 0.15625 & 0.02986(60) & 0.323(18) & 0.409(33) \\
\cline{2-6}
& \multirow{5 }*{0.1825} & 0.14700 & 0.2093(16) & 0.719(18) & 0.731(21) \\
& & 0.15000 & 0.1457(13) & 0.560(19) & 0.582(25) \\
& & 0.15250 & 0.08904(91) & 0.443(22) & 0.463(34) \\
& & 0.15500 & 0.03765(92) & 0.306(30) & 0.303(50) \\
& & 0.15580 & 0.02157(81) & 0.314(22) & 0.319(43) \\
\cline{2-6}
& \multirow{5 }*{0.185} & 0.14930 & 0.1410(19) & 0.533(26) & 0.537(33) \\
& & 0.15100 & 0.1025(13) & 0.430(24) & 0.432(32) \\
& & 0.15250 & 0.0704(12) & 0.351(25) & 0.346(34) \\
& & 0.15380 & 0.0428(10) & 0.288(30) & 0.277(42) \\
& & 0.15500 & 0.0176(11) & 0.259(42) & 0.262(78) \\
\cline{2-6}
& \multirow{5 }*{0.1875} & 0.14180 & 0.1538(23) & 0.496(32) & 0.495(39) \\
& & 0.15000 & 0.1106(17) & 0.447(28) & 0.452(35) \\
& & 0.15100 & 0.0882(14) & 0.397(29) & 0.401(36) \\
& & 0.15250 & 0.0547(12) & 0.323(34) & 0.323(45) \\
& & 0.15380 & 0.0258(11) & 0.253(52) & 0.249(68) \\
\midrule
\multirow{3}*{0.36} & \multirow{3 }*{0.1831} & 0.14750 & 0.1194(20) & 0.413(28) & 0.406(33) \\
& & 0.15000 & 0.0630(14) & 0.282(39) & 0.282(46) \\
& & 0.15100 & 0.0430(12) & 0.289(55) & 0.282(67) \\
\bottomrule
\end{tabular}
\caption{The values of $m^f_\text{pcac}$, $m_\pi$ and $m_\rho$ for fermions in the fundamental representation and for each $(b,\kappa_a,\kappa_f)$ value at $N=289$ are given. The fitting procedure is the one described in sec.~\ref{sssec:methodology}, from which we obtained $\chi^2$ per degrees of freedom in general smaller than one. }
\label{tab:fundmeson}
\end{table}
\subsubsection{Results}
The results we obtained for the PCAC mass, the pion mass and the vector meson mass of fundamental fermions are reported in lattice units in table~\ref{tab:fundmeson}.
We use these results to explore the (fundamental) chiral limit of the theory and extract an alternative scale to the ones presented in subsections~\ref{s:tekflow} and~\ref{s:creutz}.
\begin{figure}[t]
\centering
\hspace*{-1.cm}
\includegraphics[width=\linewidth+2cm]{pictures/6/mpcac_vs_k.pdf}
\caption{Extrapolation of $am^f_\text{pcac}$ to 0. In the labels we report the corresponding value of $\kappa_{f}^{(c)}$ extracted from the fit and the corresponding $\chi^2$ per degrees of freedom. Errors are calculated using standard jackknife techniques. }
\label{fig:mpcac_vs_k}
\end{figure}
The first check we do is to analyze the dependence of the PCAC mass on the fundamental hopping parameter $\kappa_{f}$. In order to determine the critical hopping parameter where the fundamental fermions become massless we follow the same strategy as for gluinos, i.e. we analyze the dependence of the PCAC mass on $\kappa_{f}$ and determine the critical hopping parameter as the point where the fundamental PCAC mass vanishes.
In figure~\ref{fig:mpcac_vs_k} we plot $am^f_{\text{pcac}}$ as a function of $1/(2\kappa_{f})$.
Performing separate linear fits for each value of $(b,\kappa_a)$, we extracted in each case the critical value of the hopping parameter $\kappa_{f}^{(c)}$. As signalled by the $\chi^2$ per degree of freedom reported in the legend of the plot, the observed linear dependence is very good and confirms what one would expect from chiral symmetry restoration in the limit of vanishing quark masses for Wilson fermions.
\begin{table}[t]
\centering
\begin{tabular}{clcccc}
\toprule
$b$ & $\kappa_a$ & $\frac{\sqrt{8t_1}}{a}$ (N=289)& $\frac{\sqrt{8t_1}}{a}$ (N=361) & $\frac{\bar{r}(1,0.65)}{a}$ & $\frac{1}{am_\rho^\chi}$ \\
\midrule
\multirow{6}*{0.34} & 0.185 & 2.878(2)(46) & 2.883(2)(58) & 3.362(36) & - \\
& 0.1875 & 3.209(3)(68) & 3.188(3)(90) & 3.658(37) & - \\
& 0.189 & 3.514(3)(65) & 3.488(3)(88) & 3.959(40) & 3.10(13) \\
& 0.191 & 4.049(4)(37) & 4.042(4)(58) & 4.682(48) & 3.53(17) \\
& 0.192067 & 4.568(6)(58) & - & - & 3.54(28) \\
& 0.193 & 5.273(7)(161) & 5.244(8)(175) & 6.014(78) & 4.86(51) \\
\midrule
\multirow{4}*{0.345} & 0.18 & 3.166(3)(74) & 3.145(2)(93) & 3.635(37) & - \\
& 0.184 & 3.664(4)(77) & 3.645(3)(98) & 4.208(41) & - \\
& 0.1868 & 4.294(5)(40) & 4.274(5)(55) & 4.939(51) & - \\
& 0.1896 & 5.559(8)(154) & 5.614(8)(201) & 6.649(111) & - \\
\midrule
\multirow{6}*{0.35} & 0.1775 & 3.730(3)(97) & 3.737(3)(105) & 4.377(43) & 3.08(14) \\
& 0.18 & 4.109(4)(75) & 4.003(7)(109) & 4.485(43) & 3.65(17) \\
& 0.1825 & 4.634(6)(70) & 4.516(5)(64) & 5.135(56) & 3.94(26) \\
& 0.1850 & 5.364(8)(144) & 5.323(7)(65) & 6.187(108) & 5.06(37) \\
& 0.186378 & 5.909(9)(192) & - & - & - \\
& 0.1875 & 6.608(11)(249) & 6.582(12)(287) & 7.681(180) & 5.34(71) \\
\midrule
\multirow{6}*{0.36} & 0.1760 & 5.77(1)(32) & - & - & - \\
& 0.1780 & 6.32(1)(31) & - & - & - \\
& 0.18 & 7.05(1)(31) & - & - & - \\
& 0.1820 & 7.85(2)(35) & - & - & - \\
& 0.183172 & 8.68(2)(44) & - & - & 6.99(87) \\
& 0.184 & 9.16(2)(52) & - & - & - \\
\bottomrule
\end{tabular}
\caption{Values of the inverse lattice spacing in units of the three different scales determined in this work. For flow related scales, the first error is statistical, the second is systematic.}
\label{tab:6scales}
\end{table}
We finally come to the determination of the scale based on the vector meson mass.
To determine the chiral extrapolation of this quantity we perform a global fit of $am_\rho$ as a linear function of $am^f_\text{pcac}$ for each value of $(b,\kappa_a)$, by imposing a common slope and extracting $am_\rho^\chi(b,\kappa_a)$ as the intercept at vanishing PCAC mass. We obtain a $\chi^2$ per degree of freedom of 0.56.
The final values of the inverse lattice spacing in units of $m_\rho^\chi$ obtained in this way are reported in our summary table~\ref{tab:6scales}. As we will discuss in the next section, the ratio between this scale and $\sqrt{8t_1} $ turns out to be quite close to one.
\subsection{Setting the scale with the flow}
\label{s:tekflow}
\input{body/6-flow}
\subsection{Setting the scale with Wilson loops (Creutz ratios) }
\label{s:creutz}
\input{body/6-creutz}
\subsection{Setting the scale with fundamental meson spectrum}
\label{s:mesons}
\input{body/6-mesons}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{pictures/6/creutzR.pdf}
\includegraphics[width=\linewidth]{pictures/6/mesonR_ultimo.pdf}
\caption{Comparison of scales as a function of the adjoint hopping parameter $\kappa_a$ for different gauge couplings. In the top panel it is displayed the ratio $\sqrt{8t_1}/\bar{r}(1,0.65)$, in the bottom one $m_\rho^\chi\sqrt{8t_1}$. The red bands represent the average weighted over the errors, while their width represent the statistical uncertainty over the average ratio. When needed, points corresponding to the same $\kappa_a$ have been slightly shifted in the x-axes to avoid overlapping.}
\label{fig:ratio}
\end{figure}
\subsection{Scale comparison}
\label{s:comparison}
We dedicate this last part of the section to the comparison of the results obtained with the three different scale setting methods which are summarized in table~\ref{tab:6scales}. In each method we have determined the lattice spacing in terms of a different unit: $a/\sqrt{8t_1}$, $a/\bar{r}$ and $am_\rho^\chi$. Although each of these quantities changes considerably when we change $b$ (the inverse lattice 't Hooft coupling) and the hopping parameter $\kappa_a$ (related to the gluino mass), scaling dictates that the ratio should stay constant and be given by the ratio of the corresponding units of energy. In figure~\ref{fig:ratio} we display the two independent ratios
$\sqrt{8t_1}/\bar{r}(1,0.65)$ and $m_\rho^\chi\sqrt{8t_1}$ for all the cases in which it is available. For the case of $a/\sqrt{8t_1}$ we have used an average of the results of $N=289$ and $N=361$ with errors that are dominated by the systematic ones. The results are compatible with being a constant within errors. From the best fit we estimate that the conversion factor between the two units $\sqrt{8t_1}$ and $\bar{r}(1,0.65)$ is $0.8708(54)$, while between $\sqrt{8t_1}$ and $1/m_\rho^\chi$ it is $1.144(23)$.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{pictures/6/comparison_new.pdf}
\caption{The inverse lattice spacing for the different values of $(b,\kappa_a)$. The quantities $\bar{r}(1,0.65)/a$ and $1/am_\rho^\chi$ have been rescaled with the conversion factor indicated in the legend to match the value of $\sqrt{8t_1}/a$.}
\label{fig:comparison}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{pictures/6/mrho_vs_mpcac.pdf}
\caption{Dependence of $m_\rho\sqrt{8t_1}$ on $m^f_\text{pcac} \sqrt{8t_1}$ for all the data samples included in table~\ref{tab:fundmeson}. }
\label{fig:mrho_vs_mpcac}
\end{figure}
A more visual impression of how scaling works can be seen in figure~\ref{fig:comparison}
where the three lattice spacing determinations of the inverse of the lattice spacing are displayed side by side after applying the conversion factors determined earlier. All datasets for which $\bar{r}(1, 0.65)$ was available are displayed. The figure shows how the three scales change considerably within all the datasets following the same trend in a consistent way. One can also notice the relative size of the errors of the three determinations of the lattice spacing. It looks as if the scale based on the Creutz ratios is the most precise, but the $N$ dependence had to be corrected for and systematic errors might be underestimated. Furthermore, our flow-based scale is the one that has been determined for all our datasets and hence, it will be used for the determination of the scale in the Supersymmetric limit to be presented in the next section.
Concerning the rho mass, although also consistent, it is much less precise than the other two. Indeed, the scaling behaviour also holds for the meson masses built from massive fundamental quarks. This can be seen in figure~\ref{fig:mrho_vs_mpcac} in which the rho mass in $1/\sqrt{8 t_1}$ units is plotted against the fundamental PCAC mass in the same units. Although with large errors all the data points seem to follow the same trend.
\section{Reduced model for $\mathcal{N}=1$ SUSY Yang-Mills at large-$N$}
In this section we specify the model which is used to generate our results. The starting point is the basic Wilson
discretization of the $\mathcal{N}=1$ SUSY Yang-Mills action for arbitrary $N$.
For the pure gauge part this is given by Wilson plaquette action:
\begin{equation}
S_g= b N \sum_n \sum_{\mu \ne \nu} \Tr \left [ \mathbf{I} - V_\mu (n) V_\nu(n+\hat \mu) V_\mu^\dagger (n + \hat \nu) V_\nu^\dagger (n) \right] ,
\end{equation}
where the link variables $V_\mu(n)$ are SU($N$) matrices (in the fundamental representation) and $b$ is the inverse of the lattice 't Hooft coupling.
For the fermionic part it is given by the Wilson-Dirac action for one adjoint Majorana fermion:
\begin{equation}
S_f= \frac{1}{2} \sum_{n,m} \Psi^t(n) C D_W(n,m) \Psi (m) ,
\end{equation}
where $D_W(n,m)$ stands for the Wilson-Dirac operator in the adjoint representation:
\begin{align}\label{eqn:WilsonDiracop}
&D_W(n,m) = \delta(n,m) \, \mathbf{I} \\
&- \kappa_a \sum_\mu \left [ (1 -\gamma_\mu) V_\mu^\mathrm{adj} (n) \delta (n+\hat \mu, m)
+
(1 + \gamma_\mu) \left(V_\mu^\mathrm{adj} (n-\hat \mu) \right)^\dagger \delta (n-\hat \mu, m) \right].
\nonumber
\end{align}
where $\kappa_a$ is the adjoint Majorana-fermion hopping parameter and $V_\mu^\text{adj}$ is the link variable in the adjoint representation.
The charge conjugation matrix $C$ satisfies the following properties
\begin{align}
&\gamma_\mu^t C = -C \gamma_\mu \label{eq:cgamma},\\
& C^t= -C.
\label{eq:ct}
\end{align}
ensuring the antisymmetric character of the $C D_W$ matrix.
In the large $N$ limit the twisted reduction prescription~\cite{Gonzalez-Arroyo:1982hyq} implies that the model is equivalent to a matrix model
obtained as follows (see also refs.~\cite{Eguchi:1982ta,Aldazabal:1983ec}). One introduces 4 SU($N$) constant matrices $\Gamma_\mu$ satisfying
\begin{equation}
\Gamma_\mu \Gamma_\nu = z_{\mu \nu} \Gamma_\nu \Gamma_\mu,
\end{equation}
with
\begin{equation}
z_{\mu \nu} = e^{2 \pi i n_{\mu \nu} /N},
\end{equation}
given in terms of the twist tensor $n_{\mu \nu}$, which is an antisymmetric tensor of integers modulo $N$.
The displacement operator by one lattice point in the direction $\mu$ is then replaced by the adjoint action by the matrix $\Gamma_\mu$. For the link matrices
this is equivalent to the substitution
\begin{equation}
V_\nu(n+\hat \mu)= \Gamma_\mu V_\nu(n) \Gamma_\mu^\dagger.
\end{equation}
After doing these substitutions, dropping the $n$ label and changing variables to $U_\mu=V_\mu \Gamma_\mu$, the pure gauge action converts
into the one of the twisted Eguchi-Kawai (TEK) model~\cite{Gonzalez-Arroyo:1982hwr,Gonzalez-Arroyo:1982hyq}:
\begin{equation}
S_\mathrm{TEK}(U) = b N \sum_{\mu\ne\nu} \mathrm{Tr}\left[ I - z_{\mu\nu}^* U_{\mu}U_{\nu}U_{\mu}^{\dag}U_{\nu}^{\dag}\right].
\label{eq:TEKaction}
\end{equation}
The same prescription applies to other fields in the adjoint representation such as the gluino fields. There are two ways to represent the
adjoint fermion fields. If we represent them as traceless $N\times N$ matrices the prescription is similar to that of the links
\begin{equation}
\Psi (n+\hat \mu) = \Gamma_\mu \Psi(n) \Gamma_\mu^\dagger.
\end{equation}
If we represent these fields as $N^2-1$ dimensional vectors the prescription is the equivalent one:
\begin{equation}
\Psi (n+\hat \mu) = \Gamma_\mu^\mathrm{adj} \Psi (n).
\end{equation}
In this way one reaches the reduced form of the Wilson-Dirac operator:
\begin{equation}
D_W = \mathbf{I} - \kappa_a \sum_\mu \left [ (1 -\gamma_\mu) U_\mu^\mathrm{adj} + (1 + \gamma_\mu) \left(U_\mu^\mathrm{adj}\right)^\dagger \right].
\label{eq:AdjWDmatrix}
\end{equation}
This procedure was used already in ref.~\cite{Gonzalez-Arroyo:2013bta} in constructing the SU($N$) gauge theory coupled to several flavours of adjoint Dirac fermions.
After integration over the fermion fields, we finally arrive to the model used in our simulations, whose partition function is given by
\begin{equation}
{\cal Z} =\int \prod_{\mu=0}^4 dU_{\mu} \, \mathrm{Pf}\left(CD_W \right) e^{-S_\mathrm{TEK}(U)},
\label{eq:partitionfunction}
\end{equation}
where the integration over the group variables $U_{\mu}$ is given by the Haar measure, and $\mathrm{Pf}(M)$
denotes the Pfaffian of the antisymmetric matrix $M$. We still need to specify the choice of the twist tensor $n_{\mu \nu}$. In our case,
we have chosen the so-called symmetric twist given by:
\begin{equation}
n_{\mu \nu} = \sqrt{N} k \left ( \theta(\nu-\mu) - \theta(\mu-\nu)\right),
\label{eq:symtw}
\end{equation}
with $\sqrt{N}$ and $k$ coprime integers and with $\theta$ representing the Heaviside theta-function.
The presence of the Pfaffian in eq.~\eqref{eq:partitionfunction} poses the question of whether this defines a positive definite probability distribution.
Indeed, the general properties of the Dirac operator in the continuum ensures its positivity, but some of these properties do not hold for the lattice Wilson-Dirac operator (they do hold for the overlap~\cite{Neuberger:1997bg}). This brings in the menace of the so-called sign problem. Since the lack of positivity could only be driven by lattice artifacts we expect it not to be too severe. We take advantage of the studies done by the Desy-Jena-Regensburg-Munster collaboration which deals with the same situation~\cite{Ali:2019agk,Ali:2018fbq,Ali:2018dnd,Munster:2014cja}. We can resort to the well-known reweighting method to deal with the sign-problem. This amounts to defining a probability density given by substitution of the Pfaffian by its absolute value:
\begin{equation}
P(U) = |\mathrm{Pf}\left(CD_W \right)| e^{-S_\mathrm{TEK}(U)}.
\end{equation}
The expectation value of an observable $\expval{O}$ in the original
model is then evaluated as follows
\begin{align}
\expval{O}=
\dfrac{\int \prod_{\mu=1}^{4} \dd{U_{\mu}} \, O(U) \,
\mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W)) P(U)}
{\int \prod_{\mu=1}^{4} \dd{U_{\mu}}
\mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W)) P(U)}
=\dfrac{\expval{O\,\mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W))}_{P(U)}}
{\expval{
\mathrm{sign}\qty(\mathrm{Pf}\qty(CD_W))}_{P(U)}},
\end{align}
The right-hand side involves the ratio of two expectation values
with respect to the probability density $P(U)$.
In the next section we describe our methodology to generate
configurations according to this distribution.
There is one important consideration that has to be made in interpreting the results of our work. It concerns finite $N$ effects. The reduction mechanism applies at infinite $N$, while the numerical results necessarily are performed at finite although very large $N$. Thus, it is very important to understand the nature and size of finite $N$ effects. The perturbative proof of reduction shows that planar diagrams in the reduced model are identical to those of the ordinary model on a finite lattice of size $(\sqrt{N})^4$. This is not the only difference between the finite $N$ versions of the ordinary and reduced model, but very often is the dominant effect. Hence, it is useful to regard $\sqrt{N}$ as the effective lattice size and finite $N$ errors acquire the form of finite volume effects. Our simulations which use $N=169$, $289$ and $361$ correspond to an effective lattice size of $13^4$, $17^4$ and $19^4$. From that viewpoint extended observables are more affected by this type of errors as compared to more local ones. For example, for Wilson loops the effects grow with the size of the loop and for smeared or flowed quantities with the corresponding smearing radius. A detailed study of these finite $N$ errors on Wilson loops has been done both perturbatively~\cite{Perez:2017jyq} and non-perturbatively~\cite{Gonzalez-Arroyo:2014dua}.
The finite-$N$/finite-effective volume identification is crucial in interpreting our choice of parameters and our results. For example, a large value of $b$ implies a small value of the lattice spacing $a$ in physical units. This translates into an effective lattice box of side $a \sqrt{N}$ and this should be kept larger than the characteristic correlation lengths of the theory. This limits the maximum values of $b$ in much the same way as the lattice size does for the ordinary lattice simulations. In conclusion, it is clear that going to larger values of $N$ is of course desirable but this is limited by the computational effort involved.
\section{Scale setting}
In this section we will present our results on the scale setting for our
lattice model. By extrapolation to the massless gluino limit we will
later achieve our main goal of determining the scale for the $\mathcal{N} = 1$ Supersymmetric gauge theory.
Scale setting is a standard and rather crucial step one has to face in lattice
gauge theories. Different procedures start by defining what observable
is going to be used as a unit of energy. On the lattice this
observable will appear as a dimensionless quantity given by the
product of the continuum quantity times the lattice spacing. Hence, by
measuring this particular lattice observable we are actually measuring
the lattice spacing in units of the inverse of that physical quantity.
It is obvious that there are infinitely many ways to fix the scale
corresponding to the infinitely many observables that one can choose.
In theories with a massive spectrum one could simply take one of the
masses as unit. For QCD or Yang-Mills theory the lowest glueball mass, the
vector meson mass or the square root of the string tension are typical
units of this kind. However, in practice making the right choice of a
unit is a crucial one. A unit must be of the appropriate size to the
objects to be measured, it should be precise, easy to compute and as
insensitive as possible to typical statistical or systematic errors.
From that point of view the natural type of units mentioned earlier might not
be the optimal, since they involve taking asymptotic limits. Other
possible family of choices involve dimensionless monotonous functions of an
energy scale. The unit of energy is taken to be the value at which the function
takes a particular numerical value. To this class belong some of the most common
scales used nowadays by the lattice community such as the Sommer scale ~\cite{Sommer:1993ce},
based on the quark-antiquark potential, and those relying on
the use of the gradient flow such as $t_0$~\cite{Luscher:2010iy} or
$w_0$~\cite{Borsanyi:2012zs}. For a review of these and other common
methods to fix the scale, the interested reader can consult ref.~\cite{Aoki:2021kgd}.
For theories like pure Yang-Mills or $\mathcal{N}=1$ SUSY Yang-Mills with
a single relevant operator the ratio of two scales is just a constant,
given by the ratio of the corresponding units. On the lattice this is
only true close enough to the critical point and the phenomenon is
called scaling.
In this work we will study three different methods to fix the scale.
This will allow us to measure the lattice spacing in each of the three
units, and in addition to test whether our data shows scaling
at the values of $b$ at which we are simulating. Obviously, scaling is
expected to work better for the larger values of $b$. In making our
choices of units we have taken into account three factors. First of all
the precision of the corresponding observable. Second its
insensitivity to the gluino mass. And third the insensitivity to
finite $N$ effects. The latter, as mentioned in
Section~\ref{s:reducedmodels}, are equivalent to finite volume effects, and represent
a very important source of systematic errors.
Here we will mention briefly what the methods are about and in the
next three subsections develop in more detail what are the
corresponding observables. The technical details of its implementation
on the lattice as well as the final results will be covered in the
following section.
The first method is a modification of gradient flow based methods devised to reduce finite volume effects in the determinations of $t_0$ and $w_0$,
equal in spirit to the one introduced for periodic boundary conditions in~\cite{Fodor:2014cpa}, but making use in this case of twisted boundary conditions
that have well known advantages from the point of view of perturbation theory. The second, a variant of which has been already used in large $N$ pure
Yang-Mills theory~\cite{Gonzalez-Arroyo:2012euf}, is analogous to the Sommer scale but involves Wilson loops of fixed aspect ratio. The advantage of this is that the
extrapolation to asymptotically large time needed to compute the quark-antiquark potential is no longer required.
Finally, the third method involves using the spectra of mesons made of
quarks in the fundamental representation to set the scale. In the large
$N$ limit fundamental fermion loops are suppressed and the quenched approximation is exact.
The computation of the meson spectrum can therefore be determined at no additional
cost on our set of dynamical adjoint fermion configurations. These
meson masses depend on an additional scale, which is the fundamental
quark mass, but in the zero mass limit this dependence disappears
and the mass of the lightest non-singlet vector meson becomes a natural scale for SUSY Yang-Mills.
We emphasize that the methods presented in the following subsections
can also be used at finite $N$ and for other gauge theories and are
particularly well suited when finite size effects are an important
source of concern.
\subsection{Gradient flow observables}
\label{s:general-flow}
The gradient flow~\cite{Narayanan:2006rf,Lohmayer:2011si,Luscher:2009eq,Luscher:2010iy} is a smoothing technique that has received much
attention in recent years. The idea is to replace the original gauge fields $A_\mu(x)$ by a set of flow time-dependent fields obtained by solving the gradient flow equations: $\partial_t B_\mu (x;t) = D_\nu G_{\mu\nu}(x;t)$ (with $B_\mu (x;0) =A_\mu(x)$) and leading to an effective smearing of the gauge fields over a length scale $\sqrt{8t}$. The
flow time, having physical dimensions of a length squared, induces natural candidates for scale setting; one just has to find a dimensionless, flow time dependent, quantity and determine the flow time at which it equals a particular pre-fixed value. The most common choices used for this purpose are based on the quantity:
\begin{equation}\label{flowedenergydensity}
\Phi(t) = \expval{\frac{t^2 E(t)}{N}},
\end{equation}
where $E(t)$ stands for the \textit{flowed energy density}:
\begin{equation}\label{energydensity}
\expval{E(t)} = \frac{1}{2}\expval{\Tr G_{\mu\nu}(x,t)G_{\mu\nu}(x,t)}.
\end{equation}
Particular examples are the ones obtained by solving either of the two following implicit equations:
\begin{align}
\eval{\Phi(t)}_{t=t_s } &= s,
\label{tc}\\
\eval{t \dv{t}\Phi(t)}_{t=w_s^2} &= s,
\label{wc}
\end{align}
where $s=0.1$ corresponds to the standard choices in the literature denoted by $t_0$~\cite{Luscher:2010iy} and $w_0$~\cite{Borsanyi:2012zs}~\footnote{
Note that these definitions differ from the standard ones used in lattice QCD simulations due to the extra, $N$-dependent, factor, required to have a well
defined large $N$ limit of these quantities.}.
All these considerations take place in infinite volume. However, numerical simulations are implemented on a $L$-site lattice
of finite physical size $l=a L$, a fact that leads to finite size effects in the determination of the scale whenever the flow radius extends over
a significant fraction of the box size. Our proposal, analogous to the one in ref.~\cite{Fodor:2014cpa}, is to take an alternative version of the observable $\Phi(t)$ with the correct infinite
volume limit but with reduced finite size effects. The idea makes use of the fact that in infinite volume one can define a renormalized \textit{gradient flow} (GF) coupling constant proportional to $\Phi(t)$~\cite{Luscher:2010iy}:
\begin{equation}
\lambda_{\rm gf} (\mu=1/\sqrt{8t}) = \frac{1} {\mathcal{K}(N)}\, \Phi(t),
\label{gf-coupling}
\end{equation}
with a flow-time independent proportionality constant $\mathcal{K}(N)$
fixed by imposing that the GF coupling equals the bare one at leading order in perturbation theory.
On a finite torus of size $l$ with twisted boundary conditions, this proportionality factor, denoted from now on by $\mathcal{N}(c(t),N)$, acquires a volume and flow time dependence parameterized in terms of the dimensionless variable $c(t)\equiv \sqrt{8 t /N l^2}$, for further details see appendix~\ref{ap:flow-imp}. We then define the finite volume quantity:
\begin{equation}
\hat \lambda (t,l,N) = \frac{1}{\mathcal{N}(c(t),N)}\, \, \Phi(t,l,N),
\label{improved-gf-coupling}
\end{equation}
which tends to the gradient flow coupling in the infinite volume limit, and, by analogy, a modified flowed energy density:
\begin{equation}
\hat {\Phi}(t,l,N) = \, \mathcal{K}(N) \hat \lambda(t,l,N).
\label{improved-phi-cont}
\end{equation}
As we will see in subsection~\ref{s:tekflow}, finite volume (finite $N$ in our set-up) effects are considerably reduced with this choice. Our scale definitions are then obtained by replacing $\Phi$ by $\hat{\Phi}$ in eqs. \eqref{tc} and \eqref{wc}, which at infinite volume coincide with the standard ones.
Let us finally mention that the
twisted finite volume normalization factor, $\mathcal{N}(c(t),N)$, has been computed and used before for a series of step scaling studies~\cite{Ramos:2014kla,GarciaPerez:2014azn,Bribian:2021cmg};
explicit formulas are given in
appendix~\ref{ap:flow-imp}, where we also give some more details on the perturbative expansion and the relevant references.
This procedure also accounts for correcting lattice artefacts at leading order if the normalization factor is computed instead in lattice perturbation
theory.
In subsection~\ref{s:tekflow} we will discuss the specific lattice implementation of this method and the methodology applied to extract the scale from the implicit eqs.~\eqref{tc} and \eqref{wc}.
\subsection{Wilson loops}
\label{s:general-creutz}
We move now onto the discussion of the second alternative to fix the scale, based on Wilson loops. As already mentioned, a particular version of this procedure was used to set the scale in large $N$ pure Yang-Mills theory~\cite{Gonzalez-Arroyo:2012euf}. Standard
physical quantities such as the string tension or
glueball masses are difficult observables to measure as they involve taking asymptotic limits. This makes them more affected by corrections and systematic uncertainties. It is in this spirit that Sommer scale~\cite{Sommer:1993ce}, which is based on the quark-antiquark potential but not at infinite separation, is superior. However, the Sommer scale
still involves the study of loops that are asymptotically long in
time. Our proposal is based on fixed aspect ratio Wilson
loops and hence it involves no limit.
Ultimately all gluon observables are based on Wilson loop expectation values. However, the Wilson loops themselves are UV divergent quantities and thus not suitable observables. Our proposal is based upon the following observable function
\begin{equation}
F(r,r') = -\pdv{\log \mathcal{W}(r,r')}{r}{r'},
\label{F-of-r}
\end{equation}
where $\mathcal{W}(r,r')$ are expectation values for rectangular
$r'\times r$ Wilson loops. The double derivative gets rid of perimeter and corner singularities
and one gets a well-defined continuum quantity depending on two scales.
We can reduce it to a single scale by fixing the aspect ratio $r'/r$.
Different choices give different definitions. Notice that Sommer scale
involves the limit $r\gg r'$. Here we will restrict
ourselves to the opposite limit given by symmetric loops $r=r'$ (a restriction taken after the
derivative is evaluated).
Indeed we claim that $F(r,r)$ is a very interesting physical quantity for SU($N$) Yang-Mills theory, which has been computed in
ref.~\cite{Gonzalez-Arroyo:2012euf} both for finite and infinite $N$. Obviously the
string tension is given by
\begin{equation}
\sigma = \lim_{r \longrightarrow \infty} F(r,r).
\end{equation}
However, it is better to fix the scale in a different way. For that
purpose one notices that $F(r,r)$ has dimensions of energy square, so we consider the dimensionless observable $G(r)=r^2 F(r,r)$. A physical scale $\bar{r}(f_0)$ can be defined, \'{a}-la Sommer, as follows
\begin{equation}
G(\bar{r}(f_0))= \eval{r^2 F(r,r)}_{r=\bar{r}} = f_0,
\end{equation}
where $f_0$ is some numerical value that can be chosen arbitrarily.
This is essentially the method used in ref.~\cite{Gonzalez-Arroyo:2012euf} to fix the
scale with $f_0=1.65$. Here we will consider a variant of the method that makes use of the gradient flow and results advantageous from the point of view of the lattice implementation.
One can use the gradient flow to
construct flow-time dependent Wilson loops $ \mathcal{W}_t (r,r')$
and use them to define a function $F_t (r,r)$, in analogy to eq.~\eqref{F-of-r}. Our new proposal is to use the flowed functions at non-zero flow time to set the scale. In order to define a function of a single energy scale in the problem, we fix the flow smearing radius to be proportional to the loop size:
$\sqrt{8 t}=\sqrt{8\bar{t}(r,s)}\equiv s r$. This amounts to defining Wilson loops having {\em fat} edges with thickness proportional to the size of the loop. This choice defines a dimensionless function of a single length $r$
\begin{equation}\label{Ghatdef}
\hat{G}(r;s)=r^2 F_{\bar{t}(r,s)}(r,r).
\end{equation}
Applying now the, by now well-known, prescription we can define a physical scale $\bar{r}(f_0,s)$ as follows
\begin{equation}\label{creutzcond}
\hat{G}(\bar r(f_0,s);s)= f_0,
\end{equation}
Any choice of $f_0$ and $s$ defines a different scale, but they should all be proportional to each other.
The concrete choice of $s$ and $f_0$ is dictated by practical reasons of accessibility and insensitivity of the corresponding lattice observable to different error sources. From that viewpoint the choice used previously $\bar{r}(1.65,0)$ is not the most adequate here.
\subsection{Fundamental meson spectrum}
\label{s:general-mesons}
In the large $N$ limit, fermions in the fundamental representation play a completely different role than adjoint ones. If the limit is taken á-la 't Hooft, sending $N$ to infinity while keeping fixed the number of fermion flavours $N_f$, fundamental-quark loops are suppressed and the quenched approximation is exact. As already mentioned, these fundamental quarks give rise to a meson spectrum, that can be used to determine an additional scale for the $\mathcal{N}=1$ SUSY Yang-Mills theory at large $N$.
Even though the fundamental spectrum introduces an additional scale, the fundamental fermion mass,
this scale is removed in the chiral limit (not to be confused with the massless gluino limit described in previous sections). It is in this particular limit when the masses of non-Goldstone fundamental mesons can be considered a natural scale of the SUSY theory.
For the purpose of this paper we have selected the lightest vector meson state to determine this additional scale.
In subsection~\ref{s:mesons} we will present our results for the determination of the lattice spacing using this method.
\section{Introduction}
\label{s:introduction}
\input{body/1-introduction}
\section{\texorpdfstring{$\mathcal{N}=1$}{N=1} SUSY Yang-Mills at large-\texorpdfstring{$N$}{N} on the lattice}
\label{s:reducedmodels}
\input{body/2-reducedmodels}
\section{Generation of configurations}
\label{s:generationofconfs}
\input{body/3-generationofconfs}
\section{SUSY restoration limit}
\label{s:chirallimit}
\input{body/4-chirallimit}
\section{Scale setting: choosing the observables}
\label{s:scalesetting}
\input{body/5-scalesetting}
\section{Scale setting: lattice determination of the scale}
\label{s:results}
\input{body/6-results}
\section{Final results and Conclusions}
\label{s:conclusion}
\input{body/8-conclusion}
\section*{Acknowledgments}
This work is partially supported by grant PGC2018-094857-B-I00 funded by MCIN/AEI/ 10.13039/501100011033 and by “ERDF A way of making Europe”, and by the Spanish Research Agency (Agencia Estatal de Investigación) through grants IFT Centro de Excelencia Severo Ochoa SEV-2016-0597 and No CEX2020-001007-S, funded by MCIN/AEI/10.13039
/501100011033. We also acknowledge support from the project H2020-MSCAITN-2018-813942 (EuroPLEx) and the EU Horizon 2020 research and innovation programme,\linebreak
STRONG-2020 project, under grant agreement No 824093. M.O. is supported by JSPS KAKENHI Grant Numbers 21K03576.
K.-I.I. is supported by MEXT as ``Program for Promoting Researches on the Supercomputer Fugaku''
(Simulation for basic science: from fundamental laws of particles to creation of nuclei, JPMXP1020200105) and JICFuS.
This work used computational resources of SX-ACE (Osaka U.) and Oakbridge-CX (U. of Tokyo)
through the HPCI System Research Project (Project ID: hp220011, hp210027, hp200027, hp190004) and Subsystem B of ITO system (Kyushu U.).
We acknowledge the use of the Hydra cluster at IFT and HPC resources at CESGA (Supercomputing Centre of Galicia).
|
\section{Introduction}
\input{sections/introduction.tex}
\section{Background}\label{sec:background}
\input{sections/background.tex}
\section{Secure Domain Rollback}\label{sec:system_design}
\input{sections/system_design.tex}
\section{Prototype Implementation}\label{sec:implementation}
\input{sections/software_implementation.tex}
\section{Case Study: Memcached}\label{sec:case_study}
\input{sections/case_study.tex}
\section{Security Evaluation}\label{sec:security_ev}\label{securitydiscussion}
\input{sections/security_eval.tex}
\section{Discussion}\label{sec:discussion}\label{sec:extensions}\label{sec:applicability}
\input{sections/discussion.tex}
\section{Related Work}\label{sec:related_work}
\input{sections/related_work.tex}
\section{Conclusion \& Future Work}
\label{sec:conclusion}\label{sec:future_work}\label{sec:availability}
\label{futurework}
\input{sections/conclusion.tex}
\begin{acks}
We thank Ilhan Gurel, Michael Liljestam, Eddy Truyen, Sini Ruohomaa and Sava Nedeljkovic for their feedback,
which helped improve the paper.
We further thank Stijn Volckaert and his team at KU Leuven -- Ghent for providing the infrastructure
to run our experiments, and for his feedback on our work. This research is partially funded by the
Research Fund KU Leuven, by the Flemish Research Programme Cybersecurity.
This research has received funding under EU H2020 MSCA-ITN action 5GhOSTS,
grant agreement no. 814035.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
\section{Supplementary Measurements}
\label{app:throughput}
\Cref{tbl:rollbacklatency} shows the detailed results for our rollback latency measurements for Memcached. The \emph{Rollback latency} column gives the mean latency in $\mathit{\mu{}s}$ over 1000 rollback iterations and the \emph{$\sigma$} column the standard deviation. We measured both the rollback that destroys the offending domain but leaves the contents of its memory intact and a version of the rollback that "scrubs" (zeroes) the full contents of domain memory before it can be re-allocated, as indicated in the \emph{Zeroing of domain data column}. The latter estimates the upper bound for maintaining confidentiality guarantees for nested domains using a 4GB heap pool and 4MB domain stack thats store sensitive information as discussed in \Cref{sec:discussion}. As rollbacks are exceptional events, we deem the $0.2s$ latency reasonable for use cases with confidentiality requirements. The latency could be be reduced by tracking and scrubbing only allocations that contain sensitive data, or reducing the reserved memory confidential domains.
\Cref{tbl:memoryconsumption} shows the detailed results of our memory consumption measurements. The \emph{Maximum resident set size} column reports the mean maximum resident size (RSS) reported by the Bash shell builtin \texttt{time} command over five iterations of the YCSB benchmark loading phase for the unmodified baseline (\emph{Baseline}) and Memcached equipped with SDRoB\xspace (\emph{SDRoB}). The columns marked \emph{$\sigma$} give the relative standard deviation of the RSS. We used used four worker threads for this experiment as indicated in the \emph{\#Thr} column.
\Cref{tbl:throughput} shows the detailed result of our throughput measurements of the YCSB benchmark for Memcached using different numbers of threads (\emph{\#Thr}) that were summarized in \Cref{sec:performance}. The \emph{Throughput} column gives the throughput in operations / seconds of three versions of Memcached: {\captionenumber{ \protect\item The unmodified baseline (\emph{Baseline}), \protect\item Memcached using the TLSF allocator (\emph{TLSF}), and \protect\item Memcached equipped with SDRoB\xspace (\emph{SDRoB}).}} The columns marked \emph{$\sigma$} give the relative standard deviation of the throughput over ten benchmark runs as percentage for the aforementioned versions. The \emph{Throughput degradation} column gives the degradation of throughput in percentage of {\captionenumalpha{ \protect\item Memcached using the TLSF allocator compared to the baseline (\emph{TLSF/Baseline}), \protect\item Memcached equipped with SDRoB\xspace compared to Memcached using the TLSF allocator (\emph{SDRoB\xspace/TLSF}), and \protect\item Memcached equipped with SDRoB\xspace compared to baseline (\emph{SDRoB\xspace/Baseline}).}}
\onecolumn
\section{OpenSSL Example}\label{sec:openssl_example}
\begin{lstlisting}[style=CStyle, label={lst:OpenSSLExample}, caption={\texttt{gcm\_encrypt()} is an example of an OpenSSL caller function taken from \url{https://wiki.openssl.org/index.php/EVP_Authenticated_Encryption_and_Decryption}. Lines 7--22 have been added to initialize the rollback functionality. The call to the EVP functions have been wrapped to execute in domain \texttt{OPENSSL\_UDI}.}]
int gcm_encrypt(...)
{
EVP_CIPHER_CTX *ctx;
int len;
int ciphertext_len;
/* Creating a persistent domain for OpenSSL */
if(sdrob_init(OPENSSL_UDI, SDROB_EXECUTION_DOMAIN | SDROB_ISOLATED_DOMAIN | SDROB_RETURN_TO_CURRENT) != SDROB_SUCCESSFUL_RETURNED)
handleErrors(); /* On an abnormal domain exit of OpenSSL domain, alternative action can be taken,
* i.e return the caller */
/* Creating a data domain shared between current domain and OpenSSL domain */
if(sdrob_init(OPENSSL_DATA_UDI, SDROB_DATA_DOMAIN) != SDROB_SUCCESSFUL_RETURNED){
sdrob_destroy(OPENSSL_UDI, SDROB_NO_HEAP_MERGE) /* If the shared domain initialization fails, don't continue,
* OpenSSL domain is destroyed */
handleErrors();
}
/* Grant access of data domain to domain OpenSSL */
if(sdrob_dprotect(OPENSSL_UDI, OPENSSL_DATA_UDI, WRITE_ENABLE | READ_ENABLE) != SDROB_SUCCESSFUL_RETURNED){
sdrob_destroy(OPENSSL_UDI, SDROB_NO_HEAP_MERGE); /* If granting shared domain access to OpenSSL fails, don't continue,
* OpenSSL domain and shared data domain is destroyed */
sdrob_destroy(OPENSSL_DATA_UDI, SDROB_NO_HEAP_MERGE)
}
/* Create and initialize the context */
if(!(ctx = EVP_CIPHER_CTX_new()))
handleErrors();
/* Initialize the encryption operation. */
if(1 != EVP_EncryptInit_ex(ctx, EVP_aes_256_gcm(), NULL, NULL, NULL))
handleErrors();
/* Set IV length if default 12 bytes (96 bits) is not appropriate */
if(1 != EVP_CIPHER_CTX_ctrl(ctx, EVP_CTRL_GCM_SET_IVLEN, iv_len, NULL))
handleErrors();
/* Initialize key and IV */
if(1 != EVP_EncryptInit_ex(ctx, NULL, NULL, key, iv))
handleErrors();
/*Provide any AAD data. This can be called zero or more times as required */
if(1 != EVP_EncryptUpdate(ctx, NULL, &len, aad, aad_len))
handleErrors();
/*Provide the message to be encrypted, and obtain the encrypted output.
* EVP_EncryptUpdate can be called multiple times if necessary */
if(1 != EVP_EncryptUpdate(ctx, ciphertext, &len, plaintext, plaintext_len))
handleErrors();
ciphertext_len = len;
/* Finalize the encryption. Normally ciphertext bytes may be written at
* this stage, but this does not occur in GCM mode */
if(1 != EVP_EncryptFinal_ex(ctx, ciphertext + len, &len))
handleErrors();
ciphertext_len += len;
/* Get the tag */
if(1 != EVP_CIPHER_CTX_ctrl(ctx, EVP_CTRL_GCM_GET_TAG, 16, tag))
handleErrors();
/* Clean up */
EVP_CIPHER_CTX_free(ctx);
return ciphertext_len;
}
\end{lstlisting}
\subsection{Vulnerabilities and Countermeasures}\label{sec:vuln_count}
Prominent vulnerabilities caused by memory-safety issues include buffer
overflows, use-after-free, and format string
vulnerabilities~\cite{Szekeres13}. Today, countermeasures such as
W$\oplus$X~\cite{Schwartz11}, ASLR~\cite{Larsen14}, stack
canaries~\cite{Cowan98} and CFI~\cite{Abadi09} aimed at mitigating memory
vulnerabilities are widely deployed by all major operating systems.
However, the current state-of-practice in runtime defenses focuses on
detecting attacks and terminating the offending processes. This is effective in
preventing adversaries from leveraging memory vulnerabilities as a stepping
stone for privilege escalation, remote code execution, or exfiltration of sensitive
information, but disregards availability concerns in favor of disrupting the
attack's kill chain. This approach may be acceptable to protect end users from
security threats, e.g., restarting a browser is a minor inconvenience
for the user. However, in high-availability applications any measure which may
cause service disruption necessitates overall system resilience to be provided
through redundancy and load balancing.
\ifnotabridged
It may also be desirable to collect diagnostics data from the application when
an attack is detected to aid in root-cause analysis. Instant termination of an
application hinders the collection of diagnostics data, except perhaps for a core
dump, but collecting data from a process which is under the attacker's control
poses yet another challenge.
\fi
\subsection{Checkpoint \& Restore}\label{sec:check_restore}
Application checkpoint \& restore~\cite{restoring} is a technique for increasing
system resilience against failure. It involves saving the state of a running
process periodically or before a critical operation, so that a failed process
can later be restarted from the checkpoint. The cost of this depends on the
amount of data that is needed to capture the system's state and the
checkpointing interval.
Several studies have focused on optimizing
checkpointing~\cite{optimizechekpoint, Young1974AFO, checkpointrestart}.
\ifnotabridged
Checkpoints can be created at system level or application level. At system
level, checkpointing needs to capture a complete reproduction of the
application's memory as well as other attributes, such as sockets, open files,
and pipes. At application level, checkpointing requires additional functionality
inserted into the application itself to facilitate checkpoint \& restore.
\fi
Secure checkpointing schemes~\cite{seccheck} generally leverage cryptography to
protect the integrity and confidentiality of checkpoint data at rest. However
they generally do not consider attacks that may tamper with checkpointing
code at the application level. Furthermore, the cost of bulk encryption of
checkpoint data is too high for latency-sensitive applications, e.g., network traffic
processing or distributed caches unless the system provides load balancing and
redundancy.
In this work, we avoid the pitfalls of checkpoints that reproduce process memory
by leveraging hardware-assisted fault isolation to partition an application
process intro distinct, isolated domains. This compartmentalization facilitates
secure rollback of application state by isolating the effects of memory errors.
This enables rollback to application states that precede the point of failure in
the application's call graph and are unaffected by a caught and contained error.
\subsection{Memory Protection Keys}\label{sec:mpk}
Memory protection keys (MPK) provide an access control mechanism that augments
page-based memory permissions. MPK allows memory access permissions to be
controlled without the overhead of kernel-level modification of page table
entries (PTEs). On 64-bit x86 processors, protection keys for userspace code
(PKU) are supported since Intel's Skylake~\cite{Intel64} and AMD's Zen 3~\cite{AMD-AMD64}
microarchitectures. Future Intel processors will add support for memory
protection keys for supervisor mode (PKS)~\cite{Intel64}. Similar hardware
mechanisms are also available in ARMv8-A~\cite{ARMv8-A}, IBM
Power~\cite{IBM-StorageProtectKeys}, HP PA-RISC~\cite{HP-PA-RISC} and
Itanium~\cite{Intel-IA64} processor architectures.
\Cref{fig:PKU} illustrates the component of PKU on 64-bit x86 processors.
Each memory page is associated with a 4-bit protection key stored in the page's
PTE \ding{182}\xspace. The access rights to memory associated with each
protection key are kept in a \emph{protection key rights register} (PKRU)~\ding{183}\xspace.
The PKRU allows write-disable (WD) and access-disable (AD) policies to be
configured for protection keys. These policies are enforced by hardware on
each memory access.
Unlike MPK mechanisms for architectures such as ARMv8-A and IBM Power, which
limit access to the PKRU to privileged code, the PKRU
in 64-bit x86 is configurable from userspace. This allows domain transitions to
occur efficiently without involving the OS kernel. However, it also means that
PKU on its own cannot effectively enforce secure in-process isolation, but has
to be combined with mechanisms such as W$\oplus$X and CFI that can limit PKRU
access to code which is trusted to manage isolation. Existing work has shown
that compiler-based code rewriting~\cite{Koning17} or binary
inspection~\cite{Vahldiek-Oberwagner19} when combined with system call
filtering~\cite{Voulimeneas22}, or non-invasive hardware
extensions~\cite{Schrammel20}, provide sufficient security for PKRU access to
preclude bypassing PKU policies.
\begin{figure}[tb]
\centering
\includegraphics[scale=0.35]{figures/PKU.png}
\caption{Overview of 64-bit x86 PKU components.}
\label{fig:PKU}
\end{figure}
\subsection{Memcached with SDRoB\xspace}\label{sec:memcached_sdrob}
We propose that each client event should be handled in a nested domain. In case
of memory corruption, the abnormal domain exit occurs in the nested domain, we
discard the related nested domain contents and come back to the root domain\xspace
securely. Memcached closes the related connection, and it can continue its
execution, handling another client request without restarting.
\Cref{fig:seq_diagram} shows a sequence diagram of Memcached with SDRoB\xspace.
We configure SDRoB\xspace with a partially isolated multithreading configuration (see
in \Cref{sec:multithreading}), because the main thread needs to communicate with
the worker threads. Each event is handled using the \texttt{drive\_machine()}
function with a corresponding connection buffer. We isolate this function using
the SDRoB\xspace API, along the lines of \Cref{lst:sdrob_call}. Recall
from \Cref{sec:domainlifecycle} that nested domains by default only have read
access to data that belongs to the parent domain. Nevertheless, certain
subroutines, such as \texttt{drive\_machine()}, need to update shared state
residing in a parent domain, e.g., the connection buffer. As a solution, the
event handler that calls
\texttt{drive\_machine()} initializes a non-isolated, nested domain $D$
(\textbf{3}) and makes a \emph{deep copy of the connection buffer} that is made
available to \texttt{drive\_machine()} (\textbf{4}). It then enters $D$
(\textbf{5}) and calls \texttt{drive\_machine()} (\textbf{6}) to handle the
client request, working on a copy of the connection buffer. After successfully
handling the request, it exits from $D$ (\textbf{7}), the original connection
buffer in the parent domain is updated with any changes present in the shared
copy (\textbf{8}). On abnormal domain exit, the copied connection buffer is
discarded. Since the event handler returns after handling the request, we need
to invalidate the saved execution context of $D$. As \texttt{drive\_machine()}
does not allocate any persistent state in $D$, we could use the transient domain
pattern and destroy $D$. However, as an optimization, we retain the copy of the
connection buffer used by the domain, hence \texttt{sdrob\_deinit()}
(\textbf{10}) is used.
The \texttt{drive\_machine()} function also needs both read and write access to
the hash table and database to perform look-ups, insertions, and updates. To
protect database and hash table integrity, we store them in separate data
domains and grant read-only access to the workers. Since domain heaps grow
dynamically, we do not need preallocation as other PKU-based Memcached
adaptations~\cite{Park19,Kjellqvist2019}.
In order to allow inserts and updates, we wrap the \texttt{slabs\_alloc()}
function, that normally returns a pointer to a memory area in the database, to
return a copy of the memory area to insert the key-value pair. Similarly, we
wrap the \texttt{store\_item()} function which stores new data and updates the
hash table. Each event handler first performs its operation on a copy of the
corresponding item. On a normal exit from the nested domain, we insert the
key-value pair to the database, and update the hash table (\textbf{9}). On an
abnormal domain exit (\textbf{11-12}), the corrupt key-value pair is discarded
along with all other memory of the domain. Note that this solution delays
updates to the database. However, due to the atomic nature of the Memcached
requests, consistency is not affected.
Memcached uses a shared mutex for synchronization of worker threads. In this
case our copying mechanism for shared data does not work, because it would hide
concurrent accesses to the mutex and break the synchronization. We opted to
create a separate data domain for the mutex that every worker can
access. See \Cref{sec:applicability} for a security discussion of this scheme.
\subsection{Performance Evaluation}
\label{sec:performance}
Two aspects of the secure rollback scheme's performance are evaluated:
\begin{inparaenum}[1)]
\item rollback latency on an abnormal exit,
\item performance impact of the isolation mechanism.
\end{inparaenum}
We run our experiments on Dell PowerEdge R540 machines with 24-core
MPK-enabled Intel(R) Xeon(R) Silver 4116 CPU (2.10GHz) having 128 GB RAM and
using Ubuntu 18.04, Linux Kernel 4.15.0.
\paragraph{Rollback Latency}
We reproduced CVE-2011-4971~\cite{CVE} to verify the SDRoB\xspace rollback mechanism
and compiled Memcached v1.4.5 with SDRoB\xspace. This CVE causes denial of
service by crashing Memcached via a large body length value in a packet. Simply
put, it creates a heap overflow but SDRoB\xspace ensures that this overflow is limited
to current execution domain, hence it triggers the domain violation and an
abnormal domain exit occurs. We measured the latency of abnormal domain exit
starting with catching \texttt{SEGFAULT} until after we close the corresponding
connection. The mean latency is $3.5\mathit{\mu s}$
($\sigma$=$0.9\mathit{\mu{}s}$).
For comparison, in our experiments the restart and loading time for 10GiB of
data into Memcached was about 2 minutes. Thus, an attacker that successfully
launches repeated attacks could completely hobble the Memcached service without
secure rollback. While this is clearly dominated by the loading time, even
applications without such volatile state, but ultra-reliable low-latency
requirements, can profit from rollback. For reference, we measured the mean
latency to restart the Memcached container automatically at about
$0.4s$ ($400000\mathit{\mu s}$, $\sigma$={$19000\mathit{\mu{}s}$).
\paragraph{Performance Impact}
We used the Yahoo! Cloud Service Benchmark (YCSB)~\cite{YCSB} to test the impact
of SDRoB\xspace on Memcached performance. YCSB has two phases: a loading phase that
populates the database with key-value pairs, and a running phase which perform
read and update operations on this data. We used workloads with sizes of 1KiB,
with a read/write distributions of 95/5. For our measurements, we stored $1
\times 10^7$ key-value pairs (1KiB each) and performed $1 \times 10^8$
operations on those pairs. Operations were performed with a Zipfian
distribution over the keys.
We compiled Memcached v1.6.13 with \texttt{-O2} optimizations, \texttt{-pie}
(for ASLR), \texttt{-fstack-protector-strong}, and \texttt{-fcf-protection} and
evaluated the performance of Memcached with the TLSF allocator and Memcached
with equipped with SDRoB\xspace as described in \Cref{sec:memcached_sdrob}. We compare
the results against YCSB on unmodified Memcached. \Cref{fig:performance} shows
the load and running phase throughput (operations/second) of the three versions
for 1, 2, 4, and 8 workers threads over 10 benchmark runs each. Each thread was
pinned to separate CPU cores. We used 32 YCSB clients with 16 threads pinned to
separate cores for each test. We fully saturated Memcached cores
for 1, 2, and 4 threads but were unable to reach full saturation
for 8 threads. We concluded that TLSF has negligible impact on throughput in all
our tests (<1\%). For Memcached augmented with SDRoB\xspace the load and running
phase overhead is $3.0\%$ / $3.7\%$ respectively for 4 threads, and $4.4\%$ /
$4.9\%$ for 2 threads. SDRoB\xspace introduced a worst-case overhead of $6.6\%$ /
$7.3\%$ for a single thread. We measured a performance degradation of $<1.7\%$ for
8 threads but lack confidence in the soundness of that result as the CPU was not
fully saturated (cf.~\Cref{tbl:throughput} in \Cref{app:throughput} for all
figures).
We measured the memory overhead of SDRoB\xspace from the maximum resident set size
(RSS) after the YCSB load phase and comparing the RSS of Memcached with SDRoB\xspace to
the baseline. The mean RSS increase is $0.4\%$ ($\sigma$=171kb).
\begin{figure}[t]
\centering
\includegraphics[scale=0.300]{figures/perfeval.png}
\caption{Throughput of different Memcached instrumentation for different numbers of threads.}
\label{fig:performance}
\end{figure}
\subsection{SDRoB\xspace API}
\label{sec:api}
Developers can use the SDRoB\xspace API calls shown in \Cref{Tab:sdrobapi} to enhance
their application with a secure rollback mechanism in a flexible way, accounting
for the different design patterns described above.
Domains are initialized by \texttt{sdrob\_init()} \ding{192}\xspace where the developer
chooses a unique index by which the domain is referenced in future API
calls. \emph{Execution} and \emph{data} domains may be created, where the latter may
hold shareable data but cannot execute code. For execution domains we further
distinguish isolated and non-isolated domains, and whether an abnormal exit from
the domain should be handled in the current domain or its parent domain. A
domain can only be initialized once per thread (unless it is deinitialized or
destroyed before by the programmer) and the point of initialization for
execution domains marks the execution context to which control flow returns in
case of an abnormal domain exit.
The API call's return value fulfills two roles. When the domain is first
initialized, it returns \emph{OK} on success or an error message, e.g.,
if the domain was already initialized in the current
thread. On abnormal domain exit, control flow returns another time from the
init function and the return value signifies the index of the nested domain that
failed and was configured to return to this point. This means that error
handling for abnormal domain exits needs to be defined in a case split on the
return value of the \texttt{sdrob\_init()} function.
After initialization, memory in an execution or data
domain can be managed using \texttt{sdrob\_malloc()} \ding{193}\xspace and \texttt{sdrob\_free()}
\ding{194}\xspace, e.g., to be able to pass arguments into the domain. Note that this
is only allowed for child domains of the current domain that are
non-isolated. For isolated domains, a shared data domain needs to be used to
exchange data. Using \texttt{sdrob\_dprotect()} \ding{195}\xspace access permissions to a
data domain can be configured for child domains.
An execution domain initialized in the current domain can be entered
and exited using \texttt{sdrob\_enter()} \ding{196}\xspace and \texttt{sdrob\_exit()}
\ding{197}\xspace. This switches the stack and heap to the selected domain and back, and
changes the memory access permissions accordingly. Currently, the SDRoB\xspace
prototype does not copy local variables on such domain transitions. Such
variables need to be passed via registers or heap.
To support the transient domain design pattern, child domains can be deleted
using \texttt{sdrob\_destroy()} \ding{198}\xspace with the option to either discard the
domain's heap memory or to merge it to the current domain. The latter is
forbidden for isolated domains. The persistent domain pattern can then be
implemented by simply not destroying the domain after exiting it, so that it can
be entered again.
An important requirement is that a nested execution domain needs
to be destroyed before the function that initialized that domain
returns. Otherwise, the stored execution context to which to return to would
become invalid as it would point to a stack frame that no longer exists. To
provide more flexibility, \texttt{sdrob\_deinit()} \ding{199}\xspace
allows to just discard a child domain's execution context but leave its memory
intact. Before entering the domain again, it needs to be re-initialized, setting
a new return context for abnormal exits.
Finally, we provide \texttt{sdrob\_call()} \ding{200}\xspace which
implements the domain life cycle of \texttt{enter\_domain()} shown
in \Cref{fig:sdroblifecycle} for a function $F$ that receives a pointer to an
object in memory as input and returns an integer-sized value. To illustrate the
API, \Cref{lst:sdrob_call} shows how \texttt{sdrob\_call()} can be implemented.
\begin{lstlisting}[style=CStyle, float, belowskip=-0.8 \baselineskip, escapechar=@@, label={lst:sdrob_call}, caption={Pseudocode showing a possible implementation of \texttt{sdrob\_call(udi\_F,F,arg,size,ret)} using other API calls (orange).}]
int err = sdrob_init(udi_F, EXECUTION_DOMAIN |
NONISOLATED | RETURN_HERE);
if (err == OK ){
// prepare passing return value and argument
register int r asm ("r12");
register void *adr asm ("r13");
adr = sdrob_malloc(udi_F, size);
if (!adr && size>0) { return MALLOC_FAILED; }
if (size>0) { memcpy(adr, arg, size); }
sdrob_enter(udi_F);
// invoke F on copy of argument and save return value
r = F(adr);
sdrob_exit();
if (ret) { *ret = r; }
sdrob_free(adr);
sdrob_destroy(udi_F, NO_HEAP_MERGE);
}
return err;
\end{lstlisting}
In the example, we first initialize a new non-isolated execution domain for
function \emph{F}. If an abnormal exit occurs in that domain,
control returns here, so we save the error code in \texttt{err}. If
initialization succeeded, we allocate a local variable \texttt{r} in a register
to retrieve the return value later. We also allocate memory for the input
argument at \texttt{adr} in the new domain. Since that domain is non-isolated,
we can copy the argument directly from the parent domain. Afterwards, we enter
the nested domain and invoke \emph{F} on the copy of the argument, saving the
return value. After exiting, we are back in the parent domain and can copy the
return value to the desired location (which is inaccessible to the nested
domain). We free all temporary memory and destroy the nested domain, freeing its
remaining memory.\footnote{Technically, the \texttt{sdrob\_free()} call is
redundant here because \texttt{sdrob\_destroy()} would free this memory as well
with the \texttt{NO\_HEAP\_MERGE} option.} Finally, we
return \texttt{OK} in case of normal domain exit, or the error code
otherwise. Users of \texttt{sdrob\_call()} can then define their own error
handling depending on this return value as shown earlier.
\begin{lstlisting}[style=CStyle, float=*, belowskip=-0.8 \baselineskip, label={lst:openssl_wrapper}, caption={excerpt from wrapper function for \texttt{EVP\_EncryptUpdate()} that executes OpenSSL in a persistent nested domain. Data is passed between the parent and nested domain via a distinct shared data domain (\texttt{OPENSSL\_DATA\_UDI}). Error handling is omitted in excerpt for brevity.}]
int __wrap_EVP_EncryptUpdate(EVP_CIPHER_CTX *ct
register evp_encrypt_update_args_t *args asm ("r12")
args = sdrob_malloc(OPENSSL_DATA_UDI, sizeof(evp_encrypt_update_args_t));
args->ctx = ctx;
args->inl = inl;
if (out != NULL && inl >= 0) { // inl + cipher_block_size is upper bound for yet unknown output size
args.out = sdrob_malloc(OPENSSL_DATA_UDI, inl + cipher_block_siz
} else { args.out = NULL; }
if (in != NULL && inl >= 0) {
args->in = sdrob_malloc(OPENSSL_DATA_UDI, (size_t)inl);
memcpy(args->in, in, inl);
} else { args.in = NULL; }
sdrob_enter(OPENSSL_UDI); // execute real EVP_EncryptUpdate in isolated domain
args->ret = __real_EVP_EncryptUpdate(args->ctx, args->out, &(args->outl), args->in, args->inl);
sdrob_exit();
*outl = args->outl;
if (out != NULL) { // copy out encrypted data from shared data domain
memcpy(out, args->out, (size_t)*outl);
}
sdrob_free(OPENSSL_DATA_UDI, args->out); // data domain persists past the lifetime of wrapper function so
sdrob_free(OPENSSL_DATA_UDI, args->in); // allocated memory must be freed to avoid memory leak
sdrob_free(OPENSSL_DATA_UDI, args);
}
\end{lstlisting}
\Cref{lst:openssl_wrapper} shows a \texttt{EVP\_EncryptUpdate()} wrapper for
OpenSSL that implements the persistent domain
pattern introduced in \Cref{sec:domainlifecycle}. Here, OpenSSL allocates it's
data, such as its context (\texttt{ctx}) \ding{182}\xspace from a domain which is strongly
isolated from its parent. While the caller can hold a pointer to \texttt{ctx},
the object itself is unaccessible to the parent domain.
Functions that enter the domain through such wrappers must receive their input and write their
output via an intermediate shared data domain. \ding{183}\xspace and \ding{184}\xspace show how data
is allocated and copied in, and \ding{185}\xspace how it is copied out from the
intermediate data domain. This persistent domain can be
combined with a transient domain as shown in \Cref{fig:sdrobrollback} to
\begin{inparaenum}[1)]
\item encapsulate the pointer to \texttt{ctx} within an outer domain, and
\item to protect the root domain from errors in the caller, such as an \texttt{out} buffer of insufficient size.
\end{inparaenum}
\Cref{lst:OpenSSLExample} in \Cref{sec:openssl_example} shows another example usage of the wrapper.
\subsection{Implementation Overview}
In a nutshell, SDRoB\xspace is implemented using four components:
\begin{inparaenum}[1)]
\item a hardware mechanism for enforcing in-process memory protection,
\item an isolated \emph{monitor data domain} that houses control data for managing execution and data domains,
\item initialization code that is run at the beginning of each application that is linked to the SDRoB\xspace library, setting up the monitor domain and memory protection, as well as
\item trusted reference monitor code that realizes the SDRoB\xspace API calls, having exclusive access to the monitor data domain.
\end{inparaenum}
Below we provide more details about these components.
\paragraph{Memory Protection}
SDRoB\xspace uses PKU protection keys (cf.~\Cref{sec:mpk}) as a hardware-assisted SFI
mechanism to create different isolated domains within an application governed by
different memory access policies. When a domain is created, a unique protection
key is assigned for it. At each domain transition, PKRU is updated to grant
access to memory areas as permitted for the newly entered domain, and to prevent
access to other memory areas.
Our evaluation platform supports Intel PKU, hence it allows us to manage up to
15 isolated domains at a time for each process. Software abstractions for
MPK, like libmpk~\cite{Park19}, increase the number of available domains.
\paragraph{SDRoB\xspace Control Data}
SDRoB\xspace stores global control data in the monitor data domain for keeping track of,
e.g., the registered domain identifiers and the protection key usage. It also
stores per-thread information for domains such as stack size, heap size, parent
domain, and memory access permissions. To support abnormal domain exits, the
currently executing domain and the saved execution contexts are stored here as
well.
\paragraph{Initialization} An application is compiled with the SDRoB\xspace
library to use the rollback mechanism. The library provides a
constructor function that is executed before \texttt{main()} to assign all
application memory to the initial isolated domain as a root domain\xspace
associated with one of the PKU protection keys. It then initializes SDRoB\xspace global
control data where the default stack and heap size for domains is configurable
through environment variables. Furthermore, it sets the root domain\xspace as active
domain, to be updated at domain transitions by the reference monitor, and
finally initializes a signal handler.
For the multithreading scenario, SDRoB\xspace has a thread constructor function as
well, that is executed before the thread start routine function to assign a
thread memory area to a domain and associate it with one of the protection keys.
\paragraph{Reference Monitor}
The reference monitor is responsible for book-keeping of domain information
in SDRoB\xspace control data. It performs domain initialization, domain
memory management, and secure domain transitions, including updating the memory
access policy, and saving and restoring the execution state of the
calling domain. Only the reference monitor has access to the monitor data
domain by updating the PKRU register accordingly. The monitor code is executed
using the stack of the nested domain that invoked it.
\paragraph{Rollback} The secure rollback from a domain is achieved by the reference
monitor saving the execution context of the parent domain into SDRoB\xspace control data
when that domain is initialized. SDRoB\xspace uses a \texttt{setjmp()}-like functionality
to store the stack pointer, the instruction pointer, the values of other
registers, and the signal mask for the context to which the call
to \texttt{sdrob\_init()} returns. Note that we cannot simply
call \texttt{setjmp()} within \texttt{sdrob\_init()} because that execution context
would become invalid as soon as the initialization routine returns. On an
abnormal domain exit, the saved parent execution state can be used by the
recovery process to restore the application’s state to the initialization point
prior to entering the nested domain by using \texttt{longjmp()}. This rollback
allows the application to continue from that last secure point of execution that
is now redirected to the developer-specified error handling code. To simplify
programming under these non-local goto semantics~\cite{setjmp}, we only allow to set the return point once per domain and
thread. Moreover, the convention that a domain needs to be destroyed or
deinitialized before the function that initialized it returns, ensures that the
saved execution context is always valid.
\subsection{Memory Management and Isolation}\label{sec:sdrobisolation}
The secure rollback mechanism is achieved by creating different domains within
an application and its security objective is that a memory defect within a
domain must only affect that domain's memory, not the memory of others. Using
the underlying SFI mechanism based on PKU, each domain is isolated from other
domains.
\paragraph{Global Variables}
We modify the linker script to ensure that global variables are allocated in a
page-aligned memory region that can be protected by PKU.
At application initialization, all global variables are assigned to the
root domain\xspace; consequently they are not accessible to nested domains. As a pragmatic
solution to the problem, we make the root domain\xspace by default read-only
for all nested domains. Write access to global data may then be achieved by
allocating it on the heap of a shared data domain, referenced by a global
pointer. Note that this approach breaks the confidentiality of the root domain\xspace
towards nested domains. As our main goal is integrity, and confidential data can
still be stored and processed in separate domains, we find it a reasonable
trade-off for our prototype.
\paragraph{Stack Management}
SDRoB\xspace creates a disjoint stack for each execution domain to ensure that the code
running in a nested domain cannot affect the stacks of other domains. The stack
area is allocated when first initializing a domain and protected using the
protection key assigned to that domain. As an optimization, we never unmap the
stack area, even when the domain is destroyed, but keep it for reuse, i.e., when
a new domain is initialized. At each domain entry, we change the stack pointer
to the nested domain stack pointer and push the return address of
the \texttt{sdrob\_enter()} call, so that the API call returns to the call site
using the new stack. Then we update the PKRU register according to memory access
policy for that domain. A similar maneuver is performed when switching back to
the parent domain's stack via the \texttt{sdrob\_exit()} API call.
\paragraph{Heap Management}
\addtocounter{footnote}{-1}
\footnotetext{The \texttt{args} pointer is kept in a callee-saved register (\texttt{r12}) to ensure it remains accessible to the nested domain after \texttt{sdrob\_enter()} (line 17) changes the domain stack}
\addtocounter{footnote}{+1}
\footnotetext{The \texttt{args.out} buffer requires room to store \texttt{inl} + \texttt{cipher\_block\_size} bytes of data. The \texttt{cipher\_block\_size} is known from \texttt{ctx}.~\cite{opensslman}}
\ifnotabridged
In order to manage a domain's heap allocations as described
in \Cref{sec:domainlifecycle} the underlying allocator must have the
ability to differentiate between allocations that occur in different domains and
ensure that the underlying memory is chosen from an address range in a
particular domain's reserved heap area. Traditionally the heap is set up to be
one large continuous memory area. Modern \texttt{malloc()}
implementations, including the GNU Allocator in
glibc~\cite{glibc} have the ability to maintain multiple disjoint heap areas, typically
for the purpose of optimizing memory access patterns in multi-threaded
applications. For example, the GNU Allocator internally maintains one or more
memory areas, referred to as \emph{arenas}, that are reserved
via \texttt{mmap()} to provide the backing memory for the application's initial
heap and subsequently allocated thread heaps. Concurrent allocations require
threads to obtain a lock on the arena structure that \texttt{malloc()} operates
on. Consequently, by assigning different arenas to different threads the GNU
Allocator enables memory allocations in different threads to occur concurrently
without interfering with each other. Unfortunately the GNU Allocator's design
does not guarantee thread-isolation between arenas: if a thread fails to
allocate memory from the arena attached to it, the \texttt{malloc()}
implementation continues the search for a suitable large block of memory to
satisfy the allocation from the application's other
arenas~\cite{glibcmalloc}.
\fi
Because heap isolation in SDRoB\xspace requires memory management with strict
guarantees that allocations within a domain are satisfied only from memory
reserved for that domain, we opted to use an allocator that natively supports
fully disjoint heap areas instead of the default glibc GNU
Allocator\ifabridged~\cite{glibcmalloc}\fi.
For our proof-of-concept implementation, we chose the Two-Level Segregated Fit
(TLSF)~\cite{tlsf} allocator~\cite{mattcontetlsf}.
TLSF is a "good-fit", constant-time allocator that allocates memory blocks from
one or more pools of memory. Each free block in a pool is linked in two
different doubly linked lists:
\begin{inparaenum}[1)]
\item a free list of blocks belonging to the same size class, and
\item a list ordered by physical address.
\end{inparaenum}
The TLSF control structure contains a \emph{free list bitmap} that describes the
availability of free memory blocks in different size ranges. TLSF uses
processor bit instructions and the bitmap to locate a corresponding linked list
of suitable-sized free blocks.
Each SDRoB\xspace domain is an assigned its own TLSF control structure and memory pool that correspond to the domain's subheap.
The size of the initial pool assigned to domains is configurable via an environmental variable.
Each individual TLSF pool is limited to 4GB in size~\cite{tlsf}.
Beyond that, a domain's memory is increased by reserving additional pools for the domain's TLSF allocator.
\paragraph{Lazy Heap Initialization}
In order to reduce the memory footprint when domains don't make heap allocations,
the control structure and initial memory pool are not initialized when the
domain is created. Instead, the reference monitor initializes the domain's TLSF
instance the first time the domain allocates heap memory with
the \texttt{malloc()} family of functions.
We interpose such functions with wrappers by placing the SDRoB library before libc in the library load order.
Upon initialization the associated memory pool is protected by associating it with the domain's protection key.
\paragraph{Subheap Merging}
Recall from \Cref{sec:api} that a domain's subheap is either discarded or merged with
the parent domain's subheap when \texttt{sdrob\_destroy()} is called, depending on
the option parameter provided. To enable subheap merging we extended the TLSF
implementation to associate a partially consumed pool to
a pre-existing TLSF control structure. When a subheap is merged, its
associated protection key is updated to match the parent domain's protection
key. The parent's TLSF control structure is then updated as follows:
\begin{inparaenum}[1)]
\item detect used and unused blocks in the memory pool being merged,
\item link all these blocks to the pre-existing block lists in the parent domain,
\item update the parent domain's TLSF free list bitmap for unused blocks, and
\item delete the child domain's TLSF control structure.
\end{inparaenum}
If the heap is not be merged when the domain is destroyed any allocations are
freed and the underlying memory pools can be reused by new domains.
\paragraph{Error Detection}
Memory access violations are generally reported to userspace software either via
\begin{inparaenum}[1)]
\item a \texttt{SEGFAULT} signal, e.g. when a domain tries to access memory
past the confines of the domain’s memory area, and
\item calls to runtime functions inserted by instrumentation, e.g., GCC's
stack protector calls \texttt{\_\_stack\_chk\_fail()} if a stack guard
check fails.
\end{inparaenum} The \texttt{\_\_stack\_chk\_fail()} function is typically
provided by glibc and terminates the application if called. During process
initialization SDRoB\xspace sets up its own signal handler for the \texttt{SEGFAULT}
signal that attributes the segmentation fault to its cause using the signal code
(\texttt{si\_code}) available via a \texttt{siginfo\_t} structure~\cite{sigaction}
provided by the runtime to the signal handler. For instance, violations of PKU
access rules are reported by \texttt{SEGV\_PKUERR} signal code.
In Linux the \texttt{SEGFAULT} signal is always delivered to the thread that generated it.
If the SDRoB\xspace signal handler detects that a violation occurs in a nested domain
it can trigger an abnormal domain exit. If a segmentation fault occurs in a
root domain\xspace, or the fault is attributed to a cause the SDRoB\xspace signal handler
is not prepared to handle the process is still terminated. SDRoB\xspace also provides
its own implementation of \texttt{\_\_stack\_chk\_fail()} that replaces the
default glibc implementation to respond to stack guard violations.
SDRoB\xspace can be extended to incorporate other pre-existing run-time error detection
mechanisms, such as Clang CFI~\cite{cficlang}
or heap-based overflow protections (e.g., heap red zones \cite{Serebryany19}),
improving the recovery capabilities. Probabilistic and passive protections such
as ASLR hinder the exploitation of memory safety violations, but cannot detect
them. Nevertheless, our rollback mechanism is compatible with ASLR, as domains
are created at run-time.
\subsection{Threat Model and Requirements}\label{sec:threatmodel}
\paragraph{Assumptions.} %
In this work, we assume that the attacker has arbitrary access to process
memory, but is restricted by the following assumptions about the system:
\begin{enumerate}[label=\textbf{A{\arabic*}}, leftmargin=*]
\item\label{as:wxorx} A W$\oplus$X policy restricts the adversary from
modifying code pages and performing code-injection.
\item\label{as:hardening} The application is hardened against run-time attacks
and can detect when the attack is in progress, but not necessarily prevent
the attacker from corrupting process memory. For the purposes of the
security analysis in \Cref{sec:security_ev} we assume that general Linux
protections outlined in \Cref{sec:vuln_count} are in place.
\end{enumerate}
We limit the scope of \ref{as:hardening} to software-level attacks.
Transient execution~\cite{Xiong21} and hardware-level attacks, e.g., fault
injection~\cite{Shepherd21}, rowhammer~\cite{Mutlu20} etc., which are generally
mitigated at hardware, firmware, or kernel level, are therefore out of scope.
\newcommand{\reqRecovery}[0]{The mechanism must allow the application to
continue operation after system defenses (\ref{as:hardening}) detect an attack.}
\newcommand{\reqMemoryIntegrityA}[0]{The mechanism must ensure that the
integrity of memory after recovering from the detected attack is maintained.}
\newcommand{\reqMemoryIntegrityB}[0]{To facilitate \ref{req:recovery} and
\ref{req:memory_integrity} the application is compartmentalized into isolated
domains with the following requirements:}
\newcommand{\reqMemoryIntegrityC}[0]{Run-time attacks that affect one domain
must not affect the integrity of memory in other domains}
\newcommand{\reqIsolationMaintained}[0]{The attacker must not be able to
tamper with components responsible for maintaining
isolation between domains, transitions between domains, or data
used as part of the rollback process.}
\paragraph{Requirements.} %
Our goal is to improve the resilience of a protected application against active
run-time attack that may compromise the integrity of the application's memory by
introducing a mechanism for \emph{secure rollback}. We define the following
requirements:
\begin{enumerate}[label=\textbf{R{\arabic*}}, leftmargin=*]
\item\label{req:recovery} \reqRecovery{}
\item\label{req:memory_integrity} \reqMemoryIntegrityA{}
\end{enumerate}
\reqMemoryIntegrityB{}
\begin{enumerate}[label=\textbf{R{\arabic*}}, leftmargin=*,resume]
\item\label{req:domain_isolation} \reqMemoryIntegrityC{}
\item\label{req:isolation-maintained} \reqIsolationMaintained{}
\end{enumerate}
\subsection{A Vulnerable C Program}\label{sec:vulnprogram}
Latent memory vulnerabilities may exist undiscovered within applications until
they are set off by input that triggers the software defect.
When a defect is triggered, it is highly likely that it
causes the application's memory to become corrupt in unexpected ways. As a
motivating example, consider the program in \Cref{lst:example_overflow} that
contains a "classic" buffer overflow vulnerability. The memory vulnerability~\ding{192}\xspace
is triggered by input that exceeds the size of the input buffer
\texttt{buf}. As the input overflows the buffer it will corrupt the surrounding
stack frame, and eventually overflow into the main function's stack frame
where it will corrupt the local variable \texttt{sum}.
\begin{lstlisting}[float, style=CStyle, label={lst:example_overflow}, caption={A vulnerable C application.}]
int get_number() {
char buf[BUFSIZE]; // BUFSIZE = 8
gets(buf);
return atoi(buf); // returns number as integer
} // if conversion is possible,
// otherwise zero
void main(void) {
int sum = 0; // sum of all inputs
for (;;) {
sum += get_number();
printf("The sum so far:
}
}
\end{lstlisting}
Modern C compilers guard against this particular vulnerability
\ifabridged
by emitting stack canaries that detect if a buffer overflow corrupts the stack.
\else
in two ways:
\begin{inparaenum}[1)]
\item by deprecating the \texttt{gets()} function in favor of \texttt{fgets()}
with explicit array bounds checks,
\item by emitting stack canaries at stack frame boundaries which indicate if a
buffer overflow corrupts a function's stack frame.
\end{inparaenum}
\Cref{lst:stack_smashing} shows how the hardened application is terminated at
run-time when the overflow is detected.
\fi
However, by the time the overflow is
detected the application's memory has already been corrupted, rendering the
application process unrecoverable.
\ifnotabridged
\begin{lstlisting}[float, style=ConsoleStyle, label={lst:stack_smashing}, caption={A stack overflow is detected at run-time.}]
./a.out
AAAAAAAAA
*** stack smashing detected ***: terminated
Aborted (core dumped)
\end{lstlisting}
\fi
\begin{lstlisting}[float, style=CStyle, label={lst:example_simple_domain}, caption={The loop from \Cref{lst:example_overflow} equipped for rollback.}]
int ret; // to hold return value
for(;;) {
// The return value of get_number is written
// to ret on normal domain exit
if(enter_domain(&get_number, &ret) == OK) {
sum += ret; // normal domain exit
printf("The sum so far:
} else { // abnormal domain exit
printf("ERROR! Bad Input");
}
}
\end{lstlisting}
\subsection{High-level Idea}\label{sec:highlevelidea}
The objective of the secure rollback mechanism is to recover the application's
execution state after a memory defect has triggered to a prior state before the
application's memory has been corrupted (\ref{req:memory_integrity}). Thus, the
application can resume its execution and continue to provide its services
without interruption (\ref{req:recovery}). To facilitate this, the application is compartmentalized
into separate, isolated domains that each execute in distinct memory
compartments allocated from the process's memory space. Should the execution of
code inside a domain fail due to a memory defect, the memory that belongs to the
domain may be corrupt. However, since the effects of the memory defect are
isolated to the memory belonging to the failing domain
(\ref{req:domain_isolation}, \ref{req:isolation-maintained}) the application's
execution can now be recovered by:
\begin{inparaenum}[1)]
\item discarding any affected memory compartments,
\item unwinding the application's stack to a state prior before the offending
domain began its execution, and
\item performing an application-specific error handling procedure that avoids
triggering the same defect again, e.g., by discarding the potentially
malicious input that caused it.
\end{inparaenum}
\Cref{lst:example_simple_domain} shows how the program from
\Cref{lst:example_overflow} is modified to benefit from secure rollback. The
call to \texttt{get\_number()} has been wrapped by a call to the
\texttt{enter\_domain()} function~\ding{182}\xspace. This wrapper function
is part of secure rollback instrumentation, here shown with simplified
arguments. It performs the following actions:
\begin{itemize}[leftmargin=*]
\item Save information about the calling environment such as register
values, including the stack and instruction pointers, and signal mask (in a
manner similar to C \texttt{setjmp()}~\cite{setjmp})
for later use by the rollback mechanism
\item Reserve a portion of the application's memory for a new domain. This
memory area contains the domain's stack and heap and persists for the
duration the domain remains active (see \Cref{sec:domainlifecycle}).
\item Update the memory access policy enforced by hardware, granting access to
memory areas assigned to new domain, and preventing access to
all other memory areas. Global data is set to read-only.
\item Finally, invoke the function inside the domain.
\end{itemize}
The newly spawned domain can end its execution in one of two ways:
\begin{inparaenum}[1)]
\item \emph{normal domain exit}, and
\item \emph{abnormal domain exit}
\end{inparaenum}
A normal domain exit~\ding{183}\xspace occurs when the application's execution flow
returns naturally from the code invoked inside the domain to the call site. This
means that the isolated code has completed successful and application
execution is resumed outside the domain. An abnormal domain exit occurs~\ding{184}\xspace
if the isolated code tries to access memory past the
confines of the domain's memory area, or a possible run-time attack is
detected. In an abnormal domain exit, the execution of the domain is halted, the
domain's memory is discarded, and the application execution is resumed by
restoring the calling environment from the information stored prior to invoking
the offending domain, effectively rolling back application state to the point
before the domain started executing. In \Cref{lst:example_simple_domain}, the
effects of a buffer overflow inside \texttt{get\_number()} are limited to the
newly created domain; the main function's stack is unaffected by the
overflow, and the rollback mechanism can thus transfer control back to \texttt{main()}.
The caller learns a domain's exit status
from \texttt{enter\_domain()}'s return value. On abnormal domain exit the application is
expected to take some alternate action to avoid the conditions that lead to the
previous abnormal domain exit before retrying the operation. For example, a
service-oriented application can close the connection to a potentially malicious
client. The rollback of application state is limited
to the state of the application's memory.
Operations that have side-effects on the application's environment, e.g.,
reading from a socket, are still visible to the application after rollback.
In the following we explain design patterns for secure
rollback that apply to different software architectures. We
discuss considerations regarding the applicability of our approach
in \Cref{sec:applicability}.
\subsection{Domain Life Cycle} \label{sec:domainlifecycle}
As part of process initialization, all application memory, including stack,
heap, and global data, are assigned to the \emph{root domain\xspace}
which forms the initial isolated domain where an application executes.
Application subroutines are compartmentalized into \emph{nested domains}
to create multiple recovery points from which
rollbacks can be performed at any point during the application’s execution. As
the names suggests, domains can be nested in the sense that several domains can
be entered subsequently starting from the root domain\xspace, each with a dedicated
rollback procedure (cf.~\Cref{sec:domainnesting}). We envision two
flavors of isolation with rollback for application subroutines:
\begin{itemize}[leftmargin=*]
\item \textbf{Protecting the application from a subroutine:} Code that may
have undetected memory vulnerabilities, e.g., third-party software
libraries, can be executed in a nested domain by instrumenting calls to the
functionality so that they execute in their own domain and may be rolled
back in case memory safety violations are detected, as shown in the
previous section.
\item \textbf{Protecting a subroutine from its caller:} parts of the
application that operate on persistent sensitive application data such as
cryptographic keys can be isolated from vulnerabilities in its callers,
preventing the leak and loss of such data. For example, functions for
encryption, decryption, or key derivation from the OpenSSL library can be
isolated in their own nested domain to protect the application's
cryptographic keys if a fault occurs in a calling nested domain.
\Cref{lst:openssl_wrapper} in \Cref{sec:api} shows a concrete example of isolating OpenSSL.
\end{itemize}
When the application's execution flow enters any isolated subroutine, a nested
domain is created and assigned a substack and subheap area. To support the two
application scenarios described above, we identify two design patterns for
nested domains. The type of a domain determines how it continues its
life cycle, in particular what happens upon domain exit.
\paragraph{Persistent Domains}
A persistent domain retains all assigned memory areas even after the
application’s execution flow returns from the persistent nested domain to the
parent domain. Another code path may enter the persistent domain again, at which
point access is granted to memory areas that belong to the persistent domain.
Modules that maintain state information across invocations should be isolated in
a persistent domain so that their state is not lost after a normal domain
exit. For instance, some software libraries may encapsulate such state by
creating a "context" object that decouples domain-specific data from business
logic. One strategy for compartmentalizing such libraries is to ensure each
distinct context is allocated in different persistent domains. A good example of
this pattern is OpenSSL which can be instantiated multiple times within an
application using different contexts. Assigning one persistent domain per
context therefore ensures cryptographic keys associated in memory with one
context remain isolated from other domains.
On abnormal exits from any domain, the rollback mechanism is triggered and all
state of the domain is discarded. Note that for persistent domains this may have
serious repercussions for an application, if the program state depends on the
persistent state of the isolated library. When, for example, the abnormal exit
leads to the loss of session keys for a TLS connection, the application may need
to recover by re-initializing the affected context and close connections that
were handled in the lost context.
Depending on the application, the parent domain may or may not be given access
to a persistent nested domain's memory. For instance, in the case of
cryptographic libraries, access to the persistent domain's memory from any other
domain should be blocked to protect sensitive data stored by the library. In
such cases data cannot be directly passed between the caller and callee in
distinct domains and shared data, e.g., call arguments and results need to be
copied between the nested domain and its caller via a designated shared memory
area. This is similar to how data is passed between different protection
domains in hardware-enforced trusted execution environment, e.g., Intel SGX
enclaves~\cite{Intel-SGX}.
\paragraph{Transient Domains}
Memory areas assigned to transient nested domains persist until the
application’s execution flow returns from such a domain to the parent domain at
which point the stack as well as unused heap memory areas assigned to the
transient nested domain are discarded. For this the rollback mechanism needs to
keep track of memory allocations in a nested domain. Any allocated memory in a
transient nested domain’s heap area can be merged back to the parent domain’s
heap area or discarded upon a normal domain exit, depending on the specific
application scenario.
In the example in~\Cref{lst:mergingarea}, execution enters a nested domain
\ding{182}\xspace where \texttt{get\_name()} \ding{183}\xspace dynamically allocates a buffer in
heap memory \ding{184}\xspace and returns it as a result to \texttt{main()}. In a normal
domain exit \ding{185}\xspace, the active allocation of the transient domain is merged to
the root domain\xspace where \texttt{main()} lives. This means that control of the
buffer allocated in the transient domain and pointed to by \texttt{ret} is
transferred. The developer can free this memory later \ding{186}\xspace. Unused
heap area of the nested domain and the stack are discarded automatically by the
secure rollback mechanism. On an abnormal domain exit \ding{187}\xspace substack area and
subheap area assigned to the transient nested domain are discarded by the
rollback mechanism already and no further clean-up by the developer is
necessary.
\begin{lstlisting}[style=CStyle, label={lst:mergingarea}, caption={Domain Merging Area Example}]
void get_name() {
void *buf;
buf = malloc(BUFSIZE);
gets(buf);
return buf;
}
void main(void) {
void *ret; // holds the return value
for(;;) {
if(enter_domain(&get_name, ret) == OK){
printf(
free(ret);
} else { // abnormal domain exit
printf("ERROR! Bad Input");
}
}
\end{lstlisting}
While read-only access from any nested domain to data in the parent domain is
allowed, writable access must never be allowed in order to contain any memory
safety violations to the nested domain.
\begin{figure*}[t]
\centering \includegraphics[scale=0.42]{figures/sdrob_domain_life_cycle.png} \caption{Domain
life cycle for calling an internal or library function \emph{F} in a nested
domain from a parent domain. Dotted arrows represent execution of user space
instructions. Steps \includegraphics[scale=0.33]{figures/unicONE.png}{} to \includegraphics[scale=0.33]{figures/unicTHREE.png}{} initialize the domain. An
argument \emph{arg} of size \emph{size} is copied into the nested domain in
step \includegraphics[scale=0.33]{figures/unicFOUR.png}{} and the result is returned in variable \emph{ret} in
step \includegraphics[scale=0.33]{figures/unicEIGHT.png}. Deleting the domain in steps \includegraphics[scale=0.33]{figures/unicNINE.png}{} and \includegraphics[scale=0.33]{figures/unicFOURTEEN.png}{}
reverses the initialization and frees all of the nested domain's memory. Code
for steps \includegraphics[scale=0.33]{figures/unicTEN.png}{} and \includegraphics[scale=0.33]{figures/unicFIFTEEN.png}{} is provided by the
programmer.} \label{fig:sdroblifecycle}
\end{figure*}
The overall life cycle of a nested domain in the transient style is depicted
in~\Cref{fig:sdroblifecycle}, highlighting the different steps the rollback
mechanism has to perform to isolate execution in a nested domain from a parent
domain. As an example, we consider a call to a function or library \emph{F} that
takes one in-memory argument and returns a value. The call \includegraphics[scale=0.33]{figures/unicSIX.png}{} is wrapped
by \texttt{enter\_domain()}, which creates a new domain \includegraphics[scale=0.33]{figures/unicONE.png}, allocates
separate stack and heap memory \includegraphics[scale=0.33]{figures/unicTWO.png}, saves the caller context
\includegraphics[scale=0.33]{figures/unicTHREE.png}, and copies the input argument onto the new heap \includegraphics[scale=0.33]{figures/unicFOUR.png}. The
domain transition step \includegraphics[scale=0.33]{figures/unicFIVE.png}{} performs mainly two actions:
\begin{inparaenum}[1)]
\item reconfiguring the memory protection mechanism to restrict or grant
memory access policies for the entered domain, \item switch between the
stacks of the two domains.
\end{inparaenum}
If a fault is caught in the nested domain \includegraphics[scale=0.33]{figures/unicELEVEN.png}, the abnormal domain
exit is triggered (\includegraphics[scale=0.33]{figures/unicTWELVE.png}-\includegraphics[scale=0.33]{figures/unicFOURTEEN.png}), discarding the contents of the faulty domain, and executing
the custom error handling code \includegraphics[scale=0.33]{figures/unicFIFTEEN.png}. The normal domain exit (\includegraphics[scale=0.33]{figures/unicSEVEN.png}-\includegraphics[scale=0.33]{figures/unicNINE.png}) stores the result and deletes the domain including freeing its memory. In the ``Handle Result'' step \includegraphics[scale=0.33]{figures/unicTEN.png}, control is transferred to the developer-provided handler code for normal exits. In both exit cases, after the handler code is executed, regular program execution resumes in the parent domain.
\subsection{Domain Nesting and Rollback} \label{sec:domainnesting}
\begin{figure}[t]
\includegraphics[width=0.4\textwidth]{figures/sdrob_rollback_example}
\caption{Example of deeply nested domains. Arrows indicate:
\includegraphics[scale=0.4]{figures/arrow_domain_enter} entering domain,
\includegraphics[scale=0.4]{figures/arrow_domain_exit_normal} normal domain exit,
\includegraphics[scale=0.4]{figures/arrow_domain_exit_abnormal} abnormal domain exit.
Normal domain exits \ding{193}\xspace, \ding{192}\xspace match the path of entered domains in reverse order. Abnormal domain exits may deviate from this path, e.g., the persistent domain at \ding{194}\xspace has the root domain\xspace as rollback target, same as the transient domain \ding{195}\xspace.}
\label{fig:sdrobrollback}
\end{figure}
Domain nesting means creating new isolated domains within an application. Each
nested domain has exactly one parent domain, which is responsible for creating
the nested domain.
All domains may have zero or more nested child domains, i.e., additional nested
domains may be created by already nested domains.
Transient and persistent-style domains can be nested with each other. One
example of a domain nesting configuration is illustrated in
\Cref{fig:sdrobrollback}. Here, the first level of nesting domain is
transient and its subsequent nested domain can be persistent. Such a setup can
allow the developer to effectively \emph{"hide"} errors in the more deeply
nested persistent domain by specifying that rollback always occur to the
recovery point established for the outer domain. At the event of an abnormal
domain exit, the rollback can occur from any nesting level to a lower nesting
level according to application requirement, i.e., the rollback can be configured
to occur from ``Nesting level 2'' to ``Nesting level 1'', or from ``Nesting
level 2'' to ``Nesting level 0'' as seen in \Cref{fig:sdrobrollback}. The
rollback mechanisms can achieve rollback from any nested domain, but not from
the root domain\xspace.
\subsection{Multithreading}\label{sec:multithreading}
The secure rollback mechanism supports POSIX threads. In most cases, threads
need to communicate with each other using shared memory, hence they need to
access root domain\xspace memory. Consequently, it would not be possible to
isolate two threads completely from each other or from the main process.
Nevertheless, it is still possible to isolate partial code paths within the
thread to separate domains, i.e., each thread can still create nested domains
with stacks and heaps that are isolated from the root domain\xspace and other nested
domains. Hence, each thread may recover via rollback from errors in such nested
domains. If one of the threads suffers an abnormal exit from the root domain\xspace,
the rollback mechanism cannot recover other threads
and the application must be terminated.
Threads have shared access to global and root domain heap memory, to per-thread
stack areas and thread-local storage. It is possible to strengthen the isolation
by configuring exclusive memory access to each thread's stack. However, the
security benefits are arguably marginal when heap access is still shared. A
shared root domain\xspace allows a higher number of parallel threads to be supported,
if the available domains provided by the underlying memory protection mechanism
is limited, as only threads that instantiate nested domains consume domain
slots. However, one could allow the developer to configure stricter, non-uniform
access privileges to other threads.
|
\section{Introduction}
The development of machine learning models for medical use cases often requires collecting large amounts of sensitive patient data. Medical datasets are usually scattered across multiple sites and underlie rigorous privacy constraints of both ethical and regulatory nature \cite{rieke2020FutureDigital}. The effectiveness of anonymization to enable data sharing is dependent on the type of data and cannot always prevent re-identification \cite{kaissis2020SecurePrivacypreserving}. Federated learning gains increasing attention as a method for training machine learning models on distributed data in a privacy-preserving manner. In a federated learning setting, holders of sensitive data can make their data available for machine learning without sharing it with other parties. In several iterations, a central server distributes an initial model to several clients holding the data, e.g., medical institutions, which then individually train their models and provide them to the server for aggregation.
Federated learning provides a basic level of privacy by the principle of data minimization, i.e., data collection and processing are restricted to a necessary minimum. However, it cannot by itself formally guarantee privacy \cite{kairouz2021AdvancesOpen}. It has been shown that input data can successfully be reconstructed from model gradients \cite{geiping2020InvertingGradients, zhu2019DeepLeakage}. In addition to the threat of data reconstruction, attacks disclosing the presence of a specific data sample or property in the training data imply a serious privacy risk for individual contributors~\cite{naseri2021RobustnessPrivacy}.
Measures to prevent privacy breaches of machine learning models are subject to ongoing research. Differential privacy is a concept actively explored in this field. Intuitively, the goal of differential privacy is to limit the impact of a single data sample or a subset of the data on the outcome of a function computed on the data, thereby providing a guarantee that no or little information can be inferred about individual samples~\cite{naseri2021RobustnessPrivacy}. However, the application of differential privacy is known to decrease the utility of the machine learning model, characterized by a trade-off between utility and privacy specific to each use case~\cite{abadi2016DeepLearning,kairouz2021AdvancesOpen,kaissis2021EndtoendPrivacy,li2019PrivacyPreservingFederated}. Despite the potential of differentially private federated learning in healthcare, there has been an increased research interest only recently on selected use cases.
Vast amounts of medical image data are currently produced in daily medical practice. Chest X-rays play an essential role in diagnosing a variety of diseases such as pneumonia~\cite{rajpurkar2017CheXNetRadiologistLevel}, and also recently in studying COVID-19~\cite{feki2021FederatedLearning}. Automatic diagnosis assistance may substantially support the work of radiologists, which is particularly of interest in the face of ongoing medical specialist shortages~\cite{rimmer2017RadiologistShortage}. Digital support systems may also mitigate the impact of error sources in human assessment that occur systematically, e.g., due to increased workload and varying professional experience~\cite{itri2018FundamentalsDiagnostic}. Increased costs for the healthcare system and potentially fatal misdiagnoses can thereby be avoided.
We evaluate the potential of privacy-preserving federated learning for the use case of disease classification on chest X-ray images. As a key contribution, we directly compare two popular image classification model architectures, DenseNet121 and ResNet50, in terms of the effects of differentially private training on model performance and privacy preservation. Extending previous work, we introduce a federated environment that is subject to data heterogeneity and imbalance. We demonstrate that the basic federated learning setting is vulnerable to privacy violation through the successful application of reconstruction attacks. We specifically compare the vulnerability to privacy breach and the effect of differential privacy on a previously unconsidered complex model, DenseNet121, with the previously studied ResNet architecture. Our results endorse the conjecture that reconstruction attacks pose a realistic threat within the federated learning paradigm, even for large and complex model architectures. We integrate Rényi differential privacy into the federated learning process and investigate how it affects the utility-privacy trade-off for our use case. Two measures of privacy are addressed: The privacy budget $\varepsilon$ as part of the formal differential privacy guarantee and the susceptibility of the local models to reconstruction attacks. Our results suggest that the DenseNet121 is a promising architecture for feasible privacy-preserving model training on X-ray images. This novel insight may direct future research and applications in that area.
This paper is structured as follows: In Section~\ref{sec:related_work}, we briefly present previous work related to privacy-preserving federated learning for the task of X-ray classification. We introduce the used datasets in Section~\ref{sec:data} and explain our federated learning setup in Section~\ref{sec:FL}. We provide background information on the \textit{Deep Leakage from Gradients} attack (Section~\ref{sec:attack}) and on the integration of differential privacy into the training of neural networks (Section~\ref{sec:background_dp}). We present the results on model performance in a basic federated learning setting (Section~\ref{sec:fl_results}), demonstrate the susceptibility of our federated learning models to reconstruction attacks (Section~\ref{sec:results_attack}), and finally evaluate the impact of differential privacy on model performance and attack vulnerability (Section~\ref{sec:results_dp}). We discuss and summarize our findings in Sections~\ref{sec:discussion} and~\ref{sec:conclusion}.
\section{Related Work}
\label{sec:related_work}
The healthcare sector especially profits from privacy-preserving machine learning due to the natural sensitivity of the underlying patient data~\cite{qayyum2021SecureRobust, rieke2020FutureDigital, shah2021MaintainingPrivacy}. A wide range of applications demonstrate that federated learning is a potential fit for leveraging diverse types of medical data, including electronic health records~\cite{brisimi2018FederatedLearning}, genomic data~\cite{li2016VERTIcalGrid}, and time-series data from wearables~\cite{chen2020FedHealthFederated}. Examples related to medical image classification include brain tumor segmentation~\cite{li2019PrivacyPreservingFederated, sheller2019MultiinstitutionalDeep}, classification and survival prediction on whole slide images in pathology~\cite{lu2020FederatedLearning}, classification of functional magnetic resonance images (fMRI)~\cite{li2020MultisiteFMRI}, and breast density classification from mammographic images~\cite{roth2020FederatedLearning}. One large research area is concerned with the classification of chest X-ray images. \citet{calli2021deep} provided an overview of recent deep learning advances in this field, but do not consider federated learning. The feasibility of federated learning on chest X-rays has previously been benchmarked for both the CheXpert~\cite{irvin2019CheXpertLarge} and the Mendeley~\cite{kermany2018LabeledOptical} dataset. \citet{chakravarty2021FederatedLearning} enhance a ResNet18 architecture with a graph neural network for federated learning on CheXpert data with site-specific data distributions. \citet{nath2020EmpiricalEvaluation} deploy a DenseNet121 model for a real-world, physically distributed implementation of federated learning on CheXpert. \citet{banerjee2020MultidiseasesClassification} determine the ResNet18 architecture as superior for federated learning on Mendeley data in comparison with ResNet50, DenseNet121, and MobileNet.
\begin{specialtable}[h]
\centering
\caption{Overview of related works evaluating deep neural networks on the CheXpert or Mendeley datasets using DenseNet or ResNet architectures. The mentioned models are not necessarily exhaustive, some papers evaluate more ResNet and DenseNet architectures. We also include our paper at the bottom for comparison with related work.\\ \emph{(non-)IID} corresponds to a (not) independent and identical data distribution.\\
\emph{DP} corresponds to the use of differential privacy ($\varepsilon=6$.)}
\label{table:sc1_chexpert_comp}
\begin{minipage}{\textwidth}
\begin{tabular}{c l l l l}
\toprule
\textbf{Data} & \textbf{Reference} & \textbf{Model} & \textbf{Federated Learning} & \textbf{AUC}\\
\midrule
\multirow{7}{*}{\rotatebox[origin=c]{90}{CheXpert}} & \citet{irvin2019CheXpertLarge} & DenseNet121 & no & $0.889$ \\[0.25em]
& \multirow{2}{*}{\citet{bressem2020ComparingDifferent}} & DenseNet121 & \multirow{2}{*}{no} & $0.869$ \\
& & ResNet50 & & $0.881$ \\[0.25em]
& \multirow{2}{*}{\citet{ke2021CheXtransferPerformance}} & DenseNet121 & \multirow{2}{*}{no} & $0.859$\\
& & ResNet50 & & $0.859$\\[0.25em]
& \citet{chakravarty2021FederatedLearning} & ResNet18 & 5 sites, non-IID & $0.782$\\[0.25em]
& \citet{nath2020EmpiricalEvaluation} & DenseNet121 & 5 sites, IID & $0.803$\\[0.25em]
\midrule
\multirow{4}{*}{\rotatebox[origin=c]{90}{Mendeley}} & \citet{banerjee2020MultidiseasesClassification} & ResNet50 & 3 sites, non-IID \footnote{The data distribution between the three hospitals is $30:32:38$, with slightly varying class distributions.} & $0.976$\footnote{No AUC given, the value corresponds to the binary accuracy.}\\[0.25em]
& \multirow{3}{*}{\citet{kaissis2021EndtoendPrivacy}} & \multirow{3}{*}{ResNet18} & no & $0.93$ \\
& & & 3 sites\footnote{No information on the data distribution available.} & $0.92$ \\
& & & 3 sites$^c$, DP & $0.89$ \\
\midrule
\multirow{4}{*}{\rotatebox[origin=c]{90}{Both}} & \multirow{4}{*}{This paper} & \multirow{2}{*}{DenseNet121} & 36 sites, non-IID & $0.935$\\
& & & 36 sites, non-IID, DP & $0.937$ \\
& & \multirow{2}{*}{ResNet50} & 36 sites, non-IID & $0.938$\\
& & & 36 sites, non-IID, DP & $0.764$ \\
\bottomrule
\end{tabular}\\[0.25em]
\end{minipage}
\end{specialtable}
Surveys on current developments in the field of privacy-preserving machine learning and federated learning describe potential threat models and privacy attacks~\cite{enthoven2020OverviewFederated, kairouz2021AdvancesOpen, lyu2020ThreatsFederated}. \citet{zhu2019DeepLeakage} originally proposed the \textit{Deep Leakage from Gradients} (DLG) attack, which allows a malicious server instance to reconstruct complete data samples from received model gradients. Subsequent improvements of the idea include analytical label reconstruction~\cite{zhao2020IDLGImproved}, improved loss functions for gradient matching~\cite{geiping2020InvertingGradients, wang2020SAPAGSelfAdaptive} and an extension towards larger batch sizes~\cite{yin2021SeeGradients}. DLG and other privacy attacks have been identified as a severe threat to federated learning. \citet{wei2020FrameworkEvaluating} evaluate the impact of attack initialization, optimization method, and training parameters including batch size, image resolution, and activation function on DLG attack success for a small network.
Ensuring privacy and protecting against reconstruction in a practicable manner is not yet fully explored and remains an open problem for the federated learning paradigm~\cite{kairouz2021AdvancesOpen, li2020FederatedLearning, pfitzner2021FederatedLearning, rieke2020FutureDigital}. A key method that finds wide use among federated learning research is \textit{differential privacy}, first proposed by \citet{dwork2006DifferentialPrivacy} in the context of database systems. Generally, it describes the addition of carefully crafted noise into a system to prevent learning too much about single data instances and measuring the remaining risk. \citet{mironov2017RenyiDifferential} introduced the variant of \textit{Rényi differential privacy}, defining a tighter bound on the privacy loss. Differential privacy in federated learning is often achieved using \textit{differentially-private stochastic gradient descent} (DP-SGD)~\cite{abadi2016DeepLearning, li2020SecureFederated, truex2020LDPFedFederated}, an algorithm that determines the appropriate noise scale and how to clip the model parameter. The combination of federated learning and differential privacy has been explored in multiple medical use cases including prediction of mortality and adverse drug reactions from electronic health records~\cite{choudhury2020DifferentialPrivacyenabled}, brain tumor segmentation~\cite{li2019PrivacyPreservingFederated}, classification of pathology whole slide images~\cite{lu2020FederatedLearning}, detection of diabetic retinopathy in images of the retina~\cite{malekzadeh2021DopamineDifferentially}, and identification of lung cancer in histopathologic images~\cite{adnan2021FederatedLearning}.
Most similarly to this work, \citet{kaissis2021EndtoendPrivacy} demonstrate a framework for the implementation and evaluation of privacy-preserving machine learning in a federated learning setting on the Mendeley dataset evenly distributed among three clients. They combine a ResNet18 model with a secure multi-party computation protocol and differential privacy and compare the success of reconstruction attacks on centralized and federated learning models. We extend their setting by considering larger networks, simulating a scenario with heterogeneous data unevenly distributed among a larger number of clients, and evaluating the impact of different parameters on model performance and the model's vulnerability to reconstruction attacks.
\section{Materials and Methods}
In this section, we first provide information about the used datasets. Then, we go over the three central pieces of our article: the federated learning baseline and our heterogeneous data distribution, the reconstruction attack, and finally the introduction of differential privacy as a defence against the attack.
\subsection{Data}
\label{sec:data}
CheXpert \cite{irvin2019CheXpertLarge} comprises 224,316 images of 65,240 adult patients in total, where 234 images are labeled by professional radiologists for use as a validation set. We only considered frontal view images as this accounts for the higher prevalence of frontal view images in the clinical setting and ensures compatibility with the Mendeley dataset. Each image is labeled with one or more of 13 classes referring to a medically relevant finding, or with "No Finding". Following previous work \cite{lenga2020ContinualLearning, mitra2020SystematicSearch}, uncertain labels were considered as negative (U-Zeroes method).
The Mendeley chest X-ray dataset version 3 \cite{kermany2018LabeledOptical} contains 5,856 images of pediatric patients and is split into original training and test sets with 5,232 and 624 images. Each image is labeled as either "Normal", with "Viral Pneumonia", or "Bacterial Pneumonia". For convenience, we assume that "Normal" in the Mendeley dataset corresponds to "No Finding" in the CheXpert dataset. To ensure compatibility between the dataset labels, our primary setting is a binary classification task based on the "No Finding" or "Normal" label, indicating the presence or absence of a medically relevant condition.
\subsection{Federated Learning}
\label{sec:FL}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{images/fl_overview.png}
\caption{In the federated learning setup, the server first initializes a model and distributes the model parameters to its clients. Over several iterations, each client trains the model individually on its data for a defined number of local epochs, sends the parameters of its locally trained model back to the server for aggregation, and receives a global model, aggregated from all trained local models.}
\label{fig:fl_overview}
\end{figure}
Successful training of a deep learning model usually relies on the availability of a single large, high-quality dataset, requiring prior data collection and curation, potentially associated with great expense in time and resources. Despite such efforts, data transfer or direct access to the data can still often not be granted due to patient privacy concerns. Federated learning enables model training on scattered data that remains at the participants' sites at all times \cite{mcmahan2017CommunicationEfficientLearning}.
A typical federated learning system consists of a central server that orchestrates the training procedure, and several clients that communicate with the server (Fig. \ref{fig:fl_overview}). The server initializes a model and distributes the model parameters to its clients. In parallel, a subset of clients trains the model individually on their data for a defined number of epochs, which is equivalent to local stochastic gradient descent (SGD) optimization. The clients send their local models back to the server, where they are aggregated through \textit{federated averaging} \cite{mcmahan2017CommunicationEfficientLearning}. The new global model is again distributed among the clients, and the process is repeated until convergence or until a defined number of communication rounds has been reached.
\subsubsection{Experimental Setup}
In real-world use cases, the expectation is that datasets between clients in a federated learning setting show some variety. We reflect this in our simulation of a federated environment by combining two public X-ray datasets representing heterogeneous target populations, adult and pediatric patients.
Our federated learning setup comprises 36 clients that each hold a subset of X-ray images from either the CheXpert or the Mendeley dataset. The clients represent hospitals or other medical institutions that provide their collected X-rays for the development of a classification model. The sizes of the clients' datasets are chosen such that they create a highly imbalanced setting including clients with very few data points, representing small institutions that make their limited amount of data available as soon as they are collected.
We simulated five clients with large subsets of the original CheXpert training set, and 31 clients with small subsets of either the original CheXpert validation set or the original Mendeley training set. We randomly split the patients whose images are part of the original CheXpert training set into five equal parts and assign each part randomly to one of five clients. Table \ref{table:dataset_splits} shows the distribution of the CheXpert validation data and Mendeley data among the remaining 31 clients, split in training, validation and test set sizes. These clients are used as targets for the reconstruction attacks in Sections~\ref{sec:results_attack} and \ref{sec:results_dp_attack}. Each client's dataset was further split into a dedicated training, validation, and test set, consisting of 70\%, 15\%, and 15\% of the client's data, respectively. Clients' datasets that are smaller than 50 images were split equally among the subsets. Datasets comprising less than ten images were used solely for training, omitting local validation or testing. All splits were performed randomly. No specific label distribution was enforced. We ensured that there was no patient overlap between clients and between training, validation, and test splits within each client's dataset.
\begin{specialtable}[h]
\caption{Number of images from the Mendeley training dataset (\subref{table:mendeley_split}) and Chexpert validation dataset (\subref{table:chexpert_split}), distributed among 14 and 17 clients, respectively. We specify how many clients are included that hold the respective amount of data. The last row shows the total of previous rows, taking into account the number of clients. }
\label{table:dataset_splits}
\centering
\captionsetup{justification=centering}
\begin{subtable}[t]{.5\linewidth}
\centering
\caption{Mendeley clients.}
\label{table:mendeley_split}
\begin{tabular}{c c c c | c}
\toprule
\makecell{\textbf{No.}\\\textbf{Clients}} & \textbf{Train} & \textbf{Val.} & \textbf{Test} & \textbf{Total}\\
\midrule
2 & 350 & 75 & 75 & 500\\
2 & 140 & 30 & 30 & 200\\
2 & 70 & 15 & 15 & 100\\
2 & 10 & 10 & 10 & 30\\
2 & 4 & 3 & 3 & 10\\
2 & 2 & 0 & 0 & 2 \\
2 & 1 & 0 & 0 & 1\\
\midrule
14 & 1,686 & 1,154 & 266 & 266
\\
\bottomrule
\end{tabular}
\end{subtable}%
\begin{subtable}[t]{.5\linewidth}
\centering
\caption{CheXpert clients.}
\label{table:chexpert_split}
\begin{tabular}{c c c c | c}
\toprule
\makecell{\textbf{No.}\\\textbf{Clients}} & \textbf{Train} & \textbf{Val.} & \textbf{Test} & \textbf{Total}\\
\midrule
2 & 10 & 10 & 10 & 30\\
5 & 4 & 3 & 3 & 10\\
5 & 2& 0 & 0 & 2\\
5 & 1 & 0 & 0 & 1\\
\midrule
17 & 125 & 55 & 35 & 35\\
\bottomrule
\end{tabular}
\end{subtable}
\end{specialtable}
\subsubsection{Model Training}
\label{sec:training}
We compared a densely connected network (DenseNet) \cite{huang2017DenselyConnected} and a residual network (ResNet) \cite{he2016DeepResidual} because both architectures have proven especially successful for the task of X-ray image classification~\cite{baltruschat2019ComparisonDeep, ke2021CheXtransferPerformance}. We monitored the local models' performance during training on their client's validation set. The global, aggregated model was evaluated using the average performance over the clients' validation sets. The client's test sets were held back for unbiased, internal evaluation of the final global model after training has finished. As the performance metric, we used the area under the receiver operating characteristics curve (AUC).
Both DenseNet121 and ResNet50 models were initialized with parameters pre-trained on ImageNet data. A fully connected layer with sigmoid activation and the adjusted number of output neurons replaced the original final classification layer. We modified the model architectures to accept one channel grayscale instead of three-channel RGB inputs to reduce unnecessary model complexity. To still leverage pre-trained model parameters, we summed the three-channel parameters of the first model layer to obtain new weights for the one-channel input. Images were resized to 224\texttimes224 pixels and normalized with ImageNet parameters adapted to grayscale color encoding by averaging over the input channels, yielding normalization parameters $\mu=0.449$ and $\sigma=0.226$. We did not apply any data augmentation methods.
Because multiple training rounds on the same dataset increase the risk of privacy leakage, we did not perform hyperparameter tuning and settled on standard hyperparameters. The training ran for at most 20 communication rounds. Each client participated in every round. We set the local batch size to ten and adapted it accordingly for clients with fewer than ten data points. To avoid overfitting, clients performed a single local epoch~\cite{kaissis2020SecurePrivacypreserving, sheller2019MultiinstitutionalDeep}. Early stopping was applied if the AUC value of the global model did not improve for five consecutive rounds. We minimized the binary cross-entropy loss using SGD with an initial learning rate of $1\mathrm{e}\text{-}2$. The learning rate was reduced by a factor of $0.1$ when reaching a performance plateau, i.e., after the AUC of the global model has not improved for three consecutive rounds. The global model with the highest mean AUC across all clients was selected as the best final model.
In private training, the privacy loss is difficult to track for some layer types. This includes active batch normalization layers, which are part of both DenseNet and ResNet architectures, as they create arbitrary dependencies between samples within a single batch \cite{ioffe2015BatchNormalization}. We experimented with different model layer freezing techniques to avoid training batch normalization layers resulting in intractable privacy loss. We refer to rendering model layers untrainable as \textit{layer freezing}. We considered full model training (no layer freezing), freezing batch normalization layers, and freezing all layers but the final classification layer.
\subsection{Reconstruction Attack}
\label{sec:attack}
Federated learning enables model training on distributed data without the need for direct data sharing. However, while federated learning satisfies the principle of data minimization by eliminating the need for data transfer, it is not by itself sufficiently privacy-preserving. Sensitive information about the training data can be inferred from shared models, which has been demonstrated in a variety of privacy attacks including inference of class representatives~\cite{hitaj2017DeepModels}, property inference~\cite{melis2019ExploitingUnintended}, membership inference~\cite{shokri2017MembershipInference}, and sample reconstruction~\cite{zhu2019DeepLeakage}.
We assume the server to be an honest-but-curious adversary with full knowledge of the federated as well as the local training procedures~\cite{kairouz2021AdvancesOpen}. It correctly orchestrates and executes the required computations. However, it has white-box access to shared model parameters and can passively investigate them without interfering with the training process.
Reconstruction attacks aim at recovering data samples from trained model parameters. A disclosure implies a serious privacy risk as X-ray images may reveal information about the patient's identity~\cite{packhauser2021MedicalChest} and sensitive properties such as patient age~\cite{sabottke2020EstimationAge}. Since reconstruction attacks can be conducted with little auxiliary information and in a passive manner, it is a relevant vulnerability within our threat model.
The \textit{Deep Leakage from Gradients} (DLG) attack enables pixel-wise reconstruction of training images from the model gradients obtained during SGD~\cite{zhu2019DeepLeakage}. The attack comprises the following steps:
\begin{enumerate}
\item Randomly initialize some dummy input data $x'$ and dummy label $y'$.
\item Fit the given initial model with the dummy data and obtain dummy gradients $\nabla \theta'$.
\item {Quantify the difference between the original and the dummy gradient by using the Euclidean ($\ell_2$) distance as the cost function:
\begin{equation} \label{eq:g_diff}
d_{grad}=\| \nabla \theta'-\nabla \theta\|^2
\end{equation}
}
\item Iteratively minimize the distance between the dummy and original gradients by adjusting the dummy input and label using the following objective:
\begin{equation}\label{g_obj}
x'^*, y'^*= \argmin_{x',y'}\| \nabla \theta'-\nabla \theta\|^2
\end{equation}
\item End the optimization process when the loss is sufficiently small, indicating complete reconstruction of the input data, or when reaching a maximum number of iterations.
\end{enumerate}
Following subsequent work, we used an improved version of the attack. We assume that labels can be reconstructed analytically \cite{zhao2020IDLGImproved} and restrict the optimization to the image data. We used a loss function based on the cosine similarity between original and dummy gradients and the Adam optimizer as proposed by \citet{geiping2020InvertingGradients}. The cosine similarity loss is defined as follows:
\begin{equation} \label{eq:cosine_loss}
d_{grad}=1-\frac{\langle \nabla \theta', \nabla \theta \rangle}{\| \nabla \theta'-\nabla \theta\|} + \alpha TV(\mathbf{x'}).
\end{equation}
\noindent $TV(\mathbf{x'})$ is the total variation of the dummy image $\mathbf{x'}$, with factor $\alpha$ as a small prior. The loss is minimized based on the sign of its gradient.
To evaluate the vulnerability of the local models to the DLG attack in our federated learning setting, we simulated an adversarial server that applies the reconstruction attack to model updates received from individual clients. We chose an arbitrary client holding one image from the Mendeley dataset to evaluate the impact of model layer freezing and attack time on image reconstruction quality. We then attacked other clients holding up to ten training images to demonstrate that they are also susceptible to a privacy breach. We conducted three trials per attack, initializing dummy images from a random normal distribution. We determined the best result as the trial with the lowest cosine similarity loss. The initial learning rate of the Adam optimizer for the attack was $1\mathrm{e}\text{-}1$. We adopted the strategy from \citet{geiping2020InvertingGradients} and reduced the learning rate by a factor of $0.1$ after $3/8$\textsuperscript{th}, $5/8$\textsuperscript{th}, and $7/8$\textsuperscript{th} of the maximum number of iterations. Each trial ran for 20,000 optimization steps. The total variation factor $\alpha$ for the cosine similarity loss was $1\mathrm{e}\text{-}2$. We inferred the model gradients by computing the absolute difference between original model parameters and local model parameters after local training.
For quantitative evaluation of attack success, we used the peak signal-to-noise ratio (PSNR), measured in the unit of decibels (dB):
\begin{equation}
PSNR=20\cdot\log_{10} \left(\frac{MAX_I}{\sqrt{MSE}}\right),
\end{equation}
where $MAX_I$ is the difference between the minimum and the maximum possible pixel value and $MSE$ is the mean squared error (MSE) between two images.
In addition to quantifying attack success with the PSNR measure, we demonstrate to what degree sensitive patient information can be derived from reconstructed X-ray images. Even if an individual cannot always be identified directly from a particular image, statistical knowledge about demographic information and other sensitive properties in a given dataset may lead to unwanted conclusions about individuals. We compare the performance of auxiliary models that predict demographic patient information from original and reconstructed X-rays. Because there is no demographic patient information available for the Mendeley data, we focus our evaluation on clients holding parts of the CheXpert dataset. We centrally trained two auxiliary ResNet50 models with the original CheXpert training data to predict patient sex and age. Sex was encoded as a binary category. The corresponding loss function for model training was binary cross-entropy. The loss function for age prediction was the MSE in years between the true and the predicted age. The sigmoid activation function in the age prediction model was replaced by a rectified linear unit (ReLU). Validation was carried out on a dedicated part of the CheXpert training data. The models achieved a validation AUC of $0.97$ on sex classification and a mean absolute error (MAE) of $6.0$ on age prediction. We applied the auxiliary classification on reconstructed images from clients that hold subsets of the original CheXpert validation data, thus ensuring that the model was only used for inference on images it has not been trained with.
\subsection{Differential Privacy}
\label{sec:background_dp}
\citet{dwork2006DifferentialPrivacy} originally proposed the notion of differential privacy in the context of database systems. Differential privacy guarantees that the amount of information revealed about any individual record during a query remains unchanged regardless of whether the record is included in the database at the time of the query or not. Put differently, the probability of receiving a specific output from a query on a database should be almost the same when an individual record is part of the database or not. \textit{Almost} means that the probabilities do not differ by more than a specific factor, which is captured by the \textit{privacy budget} or \textit{privacy loss} $\varepsilon$. In the context of machine learning, we regard model training as a function of a dataset equivalent to a query that runs on a database. Intuitively, differential privacy applied to machine learning means that training a model on a dataset should likely result in the same model that would be obtained when removing a single sample from the dataset.
The formal definition of ($\varepsilon, \delta$)-differential privacy is as follows:
\begin{equation}
Pr[\mathcal{M}(x) \in \mathcal{S}] \leq \exp({\varepsilon}) \cdot Pr[\mathcal{M}(y) \in \mathcal{S}] + \delta,
\end{equation}
where $\mathcal{M}(x)$ is the (randomized) query or function, $x$ and $y$ are parallel databases that differ in at most one entry, $\mathcal{S} \subseteq Range(\mathcal{M})$, and $\delta$ is a small term relaxing the guarantee, usually interpreted as the probability that it fails. A randomized mechanism $\mathcal{M}(x)$ can be obtained by adding noise to the original function drawn from a statistical random distribution, e.g., the Laplacian or the Gaussian distribution. The amount of noise necessary to achieve ($\varepsilon, \delta$)-differential privacy is scaled to the \textit{$\ell_2$-sensitivity} of the function, which is the maximum distance between the outputs of a function run on two parallel databases. Differential privacy has two important qualities important to its application to machine learning. The output of a differentially private random mechanism remains differentially private during the application of another data-independent function (closure under post-processing). The privacy loss can be analyzed cumulatively over several applications of a mechanism on the same database (composability). We use the variant of Rényi differential privacy, based on the Rényi divergence, in combination with a Gaussian noise mechanism that allows for a tighter estimate of the privacy loss over composite mechanisms than ($\varepsilon, \delta$)-differential privacy \cite{mironov2017RenyiDifferential}.
Differentially-private stochastic gradient descent (DP-SGD) is commonly deployed for integrating differential privacy into model training \cite{abadi2016DeepLearning}. DP-SGD adds two main steps to the SGD algorithm:
\begin{enumerate}
\item Bounding the function's sensitivity by clipping per-sample gradient $\ell_2$-norms to a clipping value $C$.
\item Adding Gaussian noise to the gradient, scaled to the sensitivity enforced by Step 1.
\end{enumerate}
We applied DP-SGD locally during training at the clients' sites. Private training was limited to at most ten communication rounds. A privacy accountant tracked the $\varepsilon$-guarantees for a specified list of orders $\alpha$ of the Rényi divergence over communication rounds. This yields the optimal $(\alpha, \varepsilon)$-pair at the end, where $\varepsilon$ is the lowest bound on the privacy loss in combination with the respective $\alpha$. Because the differentially private mechanism is closed under post-processing, aggregation of private model parameters yields a private global model that does not incur a larger privacy loss on individual clients' data than upper bounded by local DP-SGD.
To investigate the relationship between privacy and model performance for our use case, we limited the privacy budget to $\varepsilon \in \{1,3,6,10\}$. We tracked $\alpha$ values in $[1.1, 10.9]$ in steps of $0.1$, and values in $[12,63]$ in steps of $1$.
If $\delta$ is equal to or greater than the inverse of the size of the dataset, it would allow for leakage of a whole record or data sample without violation of the privacy constraint~\cite{dwork2014AlgorithmicFoundations}. As this is unacceptable, $\delta$ should be smaller than the inverse of the dataset size \cite{kaissis2021EndtoendPrivacy}. Because the size of individual client's dataset varies, we determined $\delta$ as follows:
\begin{equation}\label{eq:delta_client}
\delta_k = \mathrm{min}(\frac{1}{\|x_k\|_1} \cdot 0.9, 10^{-2}),
\end{equation}
\noindent where $\|x_k\|_1$ is the number of data samples in the training dataset of client $k$. Because some clients' datasets are very small, which would lead to high probabilities for the privacy guarantee to be violated, we defined a minimum value of $\delta=10^{-2}$.
We bounded the sensitivity of the training function by clipping per-sample gradients. An effective bound is a compromise between excessive clipping, which leads to biased aggregated gradient estimates that do not adequately represent the underlying true gradient values, and a loose clipping bound that forces to add an exaggerated amount of noise to the gradients. We employed global norm clipping, i.e., gradients were clipped uniformly over the course of training. Abadi et al. propose to use the median of unclipped gradient $\ell_2$ norms~\cite{abadi2016DeepLearning}. We could not obtain unclipped gradient norms directly because the clients' datasets were not considered available for non-private training. As a solution, we ran a few epochs of non-private, centralized training on an auxiliary chest X-ray dataset with the same training parameters as in our federated learning scenario \cite{kaissis2021EndtoendPrivacy}. We used the original Mendeley test set, which was not part of any of the clients' datasets. We randomly picked 5\% of the dataset as validation and test sets to validate the training procedure. Because centralized training on the Mendeley test set converged quickly, we tracked the medians over the first three epochs. We obtained median gradient norms of $0.42$ (DenseNet121) and $0.62$ (ResNet50) for models with frozen batch normalization layers, and $1.24$ (DenseNet121) and $0.72$ (ResNet50) for models with all layers frozen but the final layer.
\subsection{Implementation}
The implementation of all experiments is based on PyTorch\footnote{\url{https://pytorch.org/}. Last accessed December 3, 2021.} version 1.9. It is available under \url{https://github.com/linev8k/cxr-fl-privacy}. For differentially private model training, we used Opacus\footnote{\url{https://opacus.ai/}. Last accessed November 28, 2021.} version 0.14.0. Our privacy attacks follow the implementation published by Geiping et al.\footnote{\url{https://github.com/JonasGeiping/invertinggradients}. Last accessed December 3, 2021.} Our modified version is available under \url{https://github.com/linev8k/invert-gradients-cxr}.
\section{Results}
This section follows a similar structure as the previous one. We first show the results of the federated learning baseline. Then, we go over the evaluation of the reconstruction attack on the system, where we analyze different factors that impact attack success. Finally, we assess the impact of differential privacy, first on the federated learning effectiveness, and then on the reconstruction attack.
\subsection{Federated Learning Baseline}
\label{sec:fl_results}
We trained both models on the binary classification of the „No Finding“ label. The best global models achieved an AUC value of
$0.935$ (DenseNet121) and $0.938$ (ResNet50). The average AUC was larger on clients holding Mendeley data
($0.96$ for DenseNet121 and $0.95$ for ResNet50) compared to clients with CheXpert data
($0.85$ for DenseNet121 and $0.87$ for ResNet50). These results confirm the ability of deep learning models to reach a high classification performance on the Mendeley dataset \cite{banerjee2020MultidiseasesClassification, kaissis2021EndtoendPrivacy}, even when trained in a heterogeneous setting.
Given the implications of different layer freezing techniques for privacy, we compared the outcomes of full model training
(no layer freezing), freezing batch normalization layers, and freezing all layers but the final classification layer (Table~\ref{table:sc3_test_results}).
For both models, the performance after full model training and training with frozen batch normalization layers was similar with a maximum difference in AUC of $0.022$ on the test sets. We conclude that freezing batch normalization layers did not impede model training in our setting. In contrast, rendering all layers untrainable except for the final layer significantly decreased performance. This confirms outcomes from previous work where this transfer learning technique was found to be inferior to including more layers in training updates \cite{baltruschat2019ComparisonDeep}.
\begin{specialtable}[h]
\centering
\caption{Mean AUCs of the best global DenseNet121 and ResNet50 models, evaluated on the clients' test sets. \textit{Batch norm.} refers to freezing of batch normalization layers, \textit{All but last} to freezing all parameters except for the final classification layer. Training with frozen batch normalization layers delivered similar results to full model training.}
\label{table:sc3_test_results}
\begin{tabular}{l c c c}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{3}{c}{\textbf{AUC}} \\
\cmidrule{2-4}
& No freezing & Batch norm. & All but last\\
\midrule
DenseNet121 & $0.947$ & $0.935$ & $0.714$ \\
ResNet50 & $0.916$ & $0.938$ & $0.813$ \\
\bottomrule
\end{tabular}
\end{specialtable}
\subsection{Reconstruction Attack}
\label{sec:results_attack}
We attacked local models of arbitrary clients with varying layer freezing techniques, attack time points and batch sizes.
\subsubsection{Impact of Layer Freezing}
We applied the reconstruction attack to a single client's local model with a batch size of one during the first communication round. Table~\ref{table:fl_attack_results_freezing} reports the mean PSNR and sample standard deviation over three trials per experiment. Fig. \ref{fig:freezing_mend_attack} shows the reconstructed images of the best attack trials. The attack was only successful in the case of batch normalization layer freezing, indicated by larger mean PSNR values of $12.29$ (ResNet50) and $10.98$ (DenseNet121). Training the full model as well as fine-tuning only the output layer prevented the recovery of any useful image features in this setting. We further observed that the DenseNet121 seems to be more robust to leakage from gradients in this example, although the ResNet50 is the larger architecture in terms of parameter count, containing more than three times as many trainable parameters as the DenseNet121.
\begin{specialtable}[h]
\centering
\caption{Impact of layer freezing on the attack success during early training. We report the mean PSNR and sample standard deviation (STD) over all images obtained from three attack trials per setting. The batch size is kept constant at one. The attack was only successful on models with frozen batch normalization layers.}
\label{table:fl_attack_results_freezing}
\begin{tabular}{l c c c}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{3}{c}{\textbf{PSNR $\pm$ STD}} \\
\cmidrule{2-4}
& None & Batch norm. & All but last \\
\midrule
ResNet50
& $9.73 \pm 0.09$ & $12.29 \pm 0.71$ & $8.50 \pm 0.12$ \\
DenseNet121 & $8.16 \pm 0.03$ & $10.98 \pm 0.18$ & $8.07 \pm 0.03$ \\
\bottomrule
\end{tabular}
\end{specialtable}
The results highlight that shared model updates with partial layer freezing are practically relevant targets for privacy violation. Cases of attack failure, however, do not provide a formal privacy guarantee. Other factors such as the privacy-breaking properties of active batch normalization layers in the case of full model training need to be considered for a comprehensive assessment of model privacy.
\begin{figure}[h]
\centering
\begin{subfigure}[c]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/cl9_gt.png}
\caption{Original}
\end{subfigure}
\hspace{1em}
\rotatebox{90}{ResNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/resnet_less_cl9.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/resnet_bn_cl9.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/resnet_fc_cl9.png}
\end{subfigure}
\vspace{-3\baselineskip}
\hspace{0.2\linewidth}
\hspace{1em}
\rotatebox{90}{DenseNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/densenet_less_cl9.png}
\caption{None}
\label{fig:freezing_none}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/densenet_bn_cl9.png}
\caption{Batch norm.}
\label{fig:bn_cl9}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/densenet_fc_cl9.png}
\caption{All but last}
\label{fig:freezing_all_but_last}
\end{subfigure}
\caption{Best reconstructed images with varying layer freezing techniques. We attacked the locally trained model from a client holding a single Mendeley image. \textit{None} (\subref{fig:freezing_none}) refers to full model training, \textit{Batch norm.} (\subref{fig:bn_cl9}) to freezing batch normalization layers, and \textit{All but last} (\subref{fig:freezing_all_but_last}) to only training the output layer.}
\label{fig:freezing_mend_attack}
\end{figure}
\subsubsection{Impact of Training Stage}
We applied the attack during the initial communication round and after four rounds of training. We refer to the settings as an attack in \textit{early} and \textit{late} training stages, respectively. Fig. \ref{fig:stage_mend_attack} compares the images obtained from early and late attacks. The late attack was significantly more successful on both ResNet50 (mean PSNR $18.42 \pm 5.25$) and DenseNet121 (mean PSNR $11.7 \pm 2.23$). At the same time, the variation between trials was greater for the late attack.
\begin{figure}[h]
\centering
\begin{subfigure}[c]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/cl9_gt.png}
\caption{Original}
\end{subfigure}
\hspace{1em}
\rotatebox{90}{ResNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/resnet_bn_cl9.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_stage/resnet_bn_late.png}
\end{subfigure}
\vspace{-3\baselineskip}
\hspace{0.2\linewidth}
\hspace{1em}
\rotatebox{90}{DenseNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_layer/densenet_bn_cl9.png}
\caption{Early training}
\label{fig:bn_early}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_stage/densenet_bn_late.png}
\caption{Late training}
\label{fig:bn_late}
\end{subfigure}
\caption{Best reconstructed images after the first (\subref{fig:bn_early}) and after the fourth (\subref{fig:bn_late}) communication round. Reconstruction quality increased significantly after several rounds of training.}
\label{fig:stage_mend_attack}
\end{figure}
The observation that the attack was more successful as training progressed does not confirm previous evidence, which suggests that reconstruction is less successful from pre-trained models \cite{geiping2020InvertingGradients} and during later training stages \cite{kaissis2021EndtoendPrivacy}. Attack success has been associated with the magnitude of the gradients' $\ell_2$-norms, which are usually largest at the beginning when the model starts training on previously unseen data \cite{kaissis2021EndtoendPrivacy}. In Fig. \ref{fig:l2norm_training}, we investigate how the $\ell_2$-norms of our models' gradients changed as training progressed. We show the exemplary case of the pre-trained DenseNet121. Results were similar for the ResNet50, for which we refer to Appendix Section \ref{app:norm_resnet}. For each layer of every client's local model, we tracked the median $\ell_2$-norm during training. We display the per-layer mean values of all tracked medians over the local models. For both model architectures, the norms were greater during the first round of training than in the following iterations. Subsequent changes are more subtle and lack continuity. We validated that the attacked client's model did not pose an exception to this behavior.
Since our attacks were more successful during late training and we observed overall smaller gradient norms as training progressed, we could not associate larger gradient norms with increased attack success.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{images/gradient_norm_dense.pdf}
\caption{For each model layer of the DenseNet121, we tracked the median $\ell_2$-norms during training of every local model. Each dot represents the mean of one layer-median over all local models. The dashed line depicts the overall mean of all per-layer $\ell_2$-norm medians. Most layers' $\ell_2$-norms were greater during the first round of training than in later stages. In our experiments, image reconstruction was better on models from later rounds, suggesting that the magnitude of gradient $\ell_2$-norms is not a primary indicator for attack success.}
\label{fig:l2norm_training}
\end{figure}
\subsubsection{Impact of Batch Size}
We investigated the impact of the training batch size in the setting where the attack was most successful, i.e., on models trained with frozen batch normalization layers attacked during late training. We attacked clients for which the considered batch size was equal to the available number of training images. The setting is equivalent to clients with larger datasets sharing model updates after every processed batch. A batch size of ten reduced attack success as the mean PSNR values over the batch decreased to $9.01$ (ResNet50) and $8.15$ (DenseNet121) (Table \ref{table:fl_attack_results_batch}). While the quality of the reconstructions varied for individual images within a batch, at least one image out of each batch became recognizable. We note that the order of images in a batch may not be preserved in the reconstruction of larger batches, preventing a direct comparison between original and reconstructed data points. To assign a reconstructed image to its original for evaluation, we first obtained the PSNR of each original image with each reconstructed image. We then determined the first original-reconstruction pair as the one with the largest PSNR value. The next best pair was determined considering the PSNR values between the remaining original and reconstructed images. We iterated the procedure until all images have been assigned.
Fig. \ref{fig:batch_mend_attack} shows the best-reconstructed images out of each batch, demonstrating that all considered batch sizes permit severe privacy breaches on individual data samples.
\begin{specialtable}[h]
\centering
\caption{Impact of batch size on attack success during late training. We report the mean PSNR and sample standard deviation (STD) over all images obtained from three attack trials per setting. Models were trained with frozen batch normalization layers (cf. Fig.~\ref{fig:freezing_mend_attack}). Attack success deteriorated with a batch size of ten, but not significantly with smaller batch sizes.}
\label{table:fl_attack_results_batch}
\begin{tabular}{l c c c c}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{4}{c}{\textbf{PSNR $\pm$ STD}} \\
\cmidrule{2-4}
& 1 & 2 & 4 & 10\\
\midrule
ResNet50
& $18.42 \pm 5.25$ & $11.79 \pm 1.2$ & $14.60 \pm 2.86$ & $9.01 \pm 3.13$\\
DenseNet121
& $11.7 \pm 2.23$
& $12.47 \pm 3.24$
& $12.67 \pm 2.24$
& $8.15 \pm 2.82$\\
\bottomrule
\end{tabular}
\end{specialtable}
\begin{figure}[h]
\centering
\rotatebox{90}{ResNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_stage/resnet_bn_late.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_batch/resnet_bn_two.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_batch/resnet_bn_four.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_batch/resnet_bn_ten.png}
\end{subfigure}
\vspace{0.2cm}
\rotatebox{90}{DenseNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_stage/densenet_bn_late.png}
\caption{Batch size 1}
\label{fig:bn_one}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_batch/densenet_bn_two.png}
\caption{Batch size 2}
\label{fig:bn_two}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_batch/densenet_bn_four.png}
\caption{Batch size 4}
\label{fig:bn_four}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_batch/densenet_bn_ten.png}
\caption{Batch size 10}
\label{fig:bn_ten}
\end{subfigure}
\caption{Best reconstructed images out of each batch of the specified size. While other samples from those batches were not affected by the attack, the privacy of these examples' original X-rays has been severely breached, regardless of the batch size.}
\label{fig:batch_mend_attack}
\end{figure}
\subsubsection{Inference of Demographic Properties}
\label{subsec:inference_properties}
Finally, to investigate the leakage of sensitive patient information from reconstructed images, we applied the attack to 15 clients holding CheXpert validation data subsets. We included five clients each, holding one, two, and four training images, yielding 35 images in total. The setting was the same as for the attacks on Mendeley clients. We attacked models trained with frozen batch normalization layers during late training. Then, we predicted the patients' age and sex from the original X-rays and from the reconstructed images using auxiliary models to demonstrate that the images leak sensitive information.
Table \ref{table:classif_chexpert} summarizes the auxiliary model predictions. The low baseline performance of the auxiliary models on original images compared to the classifier validation estimate is probably due to the small sample size of $35$ images. Superior results on images reconstructed from the ResNet50 in the case of sex prediction suggest an increased susceptibility to privacy violation of this architecture compared to the DenseNet121.
\begin{specialtable}[h]
\centering
\caption{Performance of the auxiliary models for predicting patient sex and age from X-ray images. We compare the classification/regression of original images, and images reconstructed from local ResNet50 and DenseNet121 models. All attacked clients provided $35$ images in total. Metrics reported are AUC for sex prediction and the mean absolute error (MAE) in years for age regression.}
\label{table:classif_chexpert}
\begin{tabular}{l c c}
\toprule
\multirow{2}{*}{\textbf{Attacked Model}} & \multirow{2}{*}{\shortstack{\textbf{Sex}\\(AUC)}} & \multirow{2}{*}{\shortstack{\textbf{Age}\\(MAE)}}\\ \\
\midrule
- & $0.71$ & $11.51$\\
ResNet50 & $0.69$ & $15.42$\\
DenseNet121 & $0.56$ & $15.22$\\
\bottomrule
\end{tabular}
\end{specialtable}
\subsection{Differentially Private Federated Learning}
\label{sec:results_dp}
As a countermeasure to the reconstruction attack, we evaluate the introduction of local differential privacy into our training process. The following sections detail the implications that come with that added protection.
\subsubsection{Model Performance}
Table \ref{table:privacy_model_results} reports the models' performance with privacy budgets $\varepsilon \in \{1,3,6,10\}$. We include the non-private baseline performance for comparison. Batch normalization layer parameters were not updated during model training. We report the exact privacy budget spent by each local model as optimal $(\alpha, \varepsilon)$-pairs in Appendix Section \ref{app:eps_alpha}.
\begin{specialtable}[h]
\centering
\caption{Mean AUC of global DenseNet121 and ResNet50 models, evaluated on the clients' test sets for non-private training and private training with varying $\varepsilon$ values. Stronger privacy guarantees decreased model performance.}
\label{table:privacy_model_results}
\begin{tabular}{l c c c c c}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{5}{c}{\textbf{AUC}} \\
\cmidrule{2-6}
& - & $\varepsilon=10$ & $\varepsilon=6$ & $\varepsilon=3$ & $\varepsilon=1$ \\
\midrule
DenseNet121 & $0.935$ & $0.925$ & $0.937$ & $0.854$ & $0.800$\\
ResNet50 & $0.938$ & $0.861$ & $0.764$ & $0.711$ & $0.622$\\
\bottomrule
\end{tabular}
\end{specialtable}
We compare the utility-privacy trade-off between the two model architectures in Fig.~\ref{fig:dp_priv_perf}. The DenseNet121 performed better than ResNet50 for all considered privacy budgets. As expected, a stronger privacy guarantee claimed a higher cost in accuracy for both models. The degradation was more pronounced in the ResNet50 with an AUC difference of $0.24$ between $\varepsilon=10$ and $\varepsilon=1$. The private DenseNet121 performed equally well compared to its non-private counterpart for both $\varepsilon=10$ and $\varepsilon=6$, suggesting that a increasing the privacy budget beyond $\varepsilon=6$ does not benefit model performance. For $\varepsilon=6$, the DenseNet121 achieved an AUC of $0.937$, the ResNet50 only $0.764$.
\begin{figure}[h]
\centering
\includegraphics[width=0.75\linewidth]{images/dp_priv_vs_perf.pdf}
\caption{Model performance, evaluated on the clients' test sets in dependence on the privacy budget~$\varepsilon$. Baselines mark the peak performance of global non-private models. Stronger privacy guarantees degraded model accuracy.}
\label{fig:dp_priv_perf}
\end{figure}
We expect that model evaluation in our setting with imbalanced data distribution tends to be unreliable on clients with less data. Incidental good results on those clients may bias the global model's performance estimate. To provide a more meaningful assessment of the model performance under privacy conditions, we investigated the performance of the best global private DenseNet121 model on individual clients compared to the best non-private model. We visualize the comparison for $\varepsilon=6$ in Fig.~\ref{fig:clients_comp_eps}. We provide the figures for other considered $\varepsilon$-values in Appendix Section \ref{app:clients_eps}. Private training demanded a systematic cost in performance for clients holding large amounts of CheXpert data. AUC values on those clients' datasets decreased by $0.03$ ($\varepsilon=10$ and $\varepsilon=6$) and $0.06$ ($\varepsilon=3$) on average from non-private to private training.
We conclude that the impact of private training on model accuracy, also at moderate privacy budgets, needs to be carefully assessed on the client level. Further potential weak points of the resulting model, such as performance on underrepresented patient subgroups, require additional consideration.
\begin{figure}[h]
\centering
\includegraphics[width=0.75\linewidth]{images/clients_comp_dense_eps6.pdf}
\caption{Comparison of per-client AUC values achieved by the best global DenseNet121 model between private ($\varepsilon=6$) and non-private training. The bottom left circle markers belong to CheXpert clients with large datasets. Privacy demanded a higher cost in model accuracy on CheXpert clients. The figure does not consider AUC values below $0.79$. Markers may overlap.}
\label{fig:clients_comp_eps}
\end{figure}
\subsubsection{Additional Training Techniques}
We evaluate the effect of additional training techniques on model performance: Training only the final layer (\textit{All but last} layer freezing), client subsampling, and employment of layer-wise gradient clipping. All experiments were carried out with a privacy budget of $\varepsilon=10$. We found that none of the techniques introduced an advantage for private model training. When restricting training to the final layer, the performance of the DenseNet121 decreased significantly compared to \textit{Batch norm.} freezing (AUC $0.707$ vs. $0.925$) and that of the ResNet50 remained similar (AUC $0.871$ vs. $0.861$).
In a separate experiment, we introduced a client subsampling procedure where the maximum number of global communication rounds was set to ten and the maximum number of rounds that each client can be selected to five. The fraction of clients chosen each round was $0.3$, resulting in eleven clients selected per round. This way, less of the available privacy budget was effectively spent during the clients' local training because each client participated in fewer training rounds in total. However, there was no improvement in model performance. The DenseNet121 and ResNet50 achieved AUC values of $0.836$ and $0.819$, respectively. A potential explanation is that clients with small datasets got selected frequently during subsampling, but could not contribute as effectively to the global model as clients with larger datasets. Model accuracy degraded more heavily after a few rounds during the subsampling experiment, indicating stronger local overfitting which was amplified by the lower number of contributing clients.
Finally, instead of uniformly clipping the norm of each gradient value, we specified an individual clipping bound for each model layer. We utilized the per-layer median gradient norms from the auxiliary training experiment on the Mendeley test set (Section \ref{sec:background_dp}). The models' AUC values converged to $0.5$, indicating that model training failed for our use case when employing layer-wise gradient clipping. The variation between individual clipping values may be too large, preventing the model parameters to retain any information that is usable in combination with other layers' parameters.
\begin{figure}
\centering
\rotatebox{90}{ResNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/resnet_dp_eps10.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/resnet_dp_eps6.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/resnet_dp_eps3.png}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/resnet_dp_eps1.png}
\end{subfigure}
\vspace{0.2cm}
\rotatebox{90}{DenseNet}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/densenet_dp_eps10.png}
\caption{$\varepsilon=10$}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/densenet_dp_eps6.png}
\caption{$\varepsilon=6$}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/densenet_dp_eps3.png}
\caption{$\varepsilon=3$}
\end{subfigure}
\hspace{1em}
\begin{subfigure}[t]{0.2\linewidth}
\includegraphics[width=\linewidth]{images/mend_priv/densenet_dp_eps1.png}
\caption{$\varepsilon=1$}
\end{subfigure}
\caption{Best reconstructed images under different privacy constraints. Private training successfully prevents the leakage of any visible features.}
\label{fig:priv_rec_images}
\end{figure}
\subsubsection{Vulnerability to Reconstruction Attack}
\label{sec:results_dp_attack}
We attempted to reconstruct the training image from the local model shared by a Mendeley client during private training. We performed the attack on models with frozen batch normalization layers during late training. Table \ref{table:priv_attack_results} compares the mean PSNR over three trials between non-private and private training. The PSNR on all images from private models was significantly smaller than in the non-private setting. Fig. \ref{fig:priv_rec_images} confirms that the reconstructed images from both model architectures did not leak any visual parts of the training images. Differentially private training under all considered privacy budgets therefore successfully prevented the attack.
\begin{specialtable}[h]
\centering
\caption{Mean PSNR over three attack trials on non-private and private local models from a Mendeley client holding one training image. The attack failed for all considered $\varepsilon$ values.}
\label{table:priv_attack_results}
\begin{tabular}{l c c c c c}
\toprule
\multirow{2}{*}{\textbf{Model}} & \multicolumn{5}{c}{\textbf{PSNR $\pm$ STD}} \\
\cmidrule{2-6}
& - & $\varepsilon=10$ & $\varepsilon=6$ & $\varepsilon=3$ & $\varepsilon=1$ \\
\midrule
DenseNet121 & $10.98 \pm 0.18$
& $6.58 \pm 0.07$
& $6.50 \pm 0.08$
& $6.46 \pm 0.07$
& $6.41 \pm 0.03$ \\
ResNet50 & $12.29\pm 0.71$
& $7.22 \pm 0.04$
& $7.13 \pm 0.16$
& $7.12 \pm 0.17$
& $8.49 \pm 0.32$ \\
\bottomrule
\end{tabular}
\end{specialtable}
To validate that no sensitive information was leaked, we applied the auxiliary models (first introduced in Section~\ref{subsec:inference_properties}) to predict patient age and sex from images reconstructed from private models. Table \ref{table:classif_chexpert_priv} compares their performance on original and recovered images in private and non-private settings. We attacked the model with the weakest privacy guarantee of $\varepsilon=10$. The AUC values of $0.49$ and $0.47$ on sex prediction indicate that the classifier's performance was equivalent to random label assignment in the private setting. The age predictions deviated around $19$ years on average from the true patients' age. Differentially private model training prevented both auxiliary models to predict usable information about the patients' demographic properties.
\begin{specialtable}[h]
\centering
\caption{Performance of the auxiliary models for predicting patient sex and age from X-ray images. We compare the predictions on original images, and images reconstructed from local ResNet50 and DenseNet121 models in the non-private and private setting with $\varepsilon=10$. Images reconstructed from private models leaked no usable information about the selected properties.}
\label{table:classif_chexpert_priv}
\begin{tabular}{l c c c}
\toprule
\multirow{2}{*}{\textbf{Attacked Model}} &
\multirow{2}{*}{$\varepsilon$} &
\multirow{2}{*}{\shortstack{\textbf{Sex}\\(AUC)}} & \multirow{2}{*}{\shortstack{\textbf{Age}\\(MAE)}}\\\\
\midrule
- & - & $0.71$ & $11.51$ \\
\multirow{2}{*}{ResNet50} & - & $0.69$ & $15.42$ \\
& $10$ & $0.49$ & $19.23$\\
\multirow{2}{*}{DenseNet121} & - & $0.56$ & $15.22$\\
& $10$ & $0.47$ & $18.82$\\
\bottomrule
\end{tabular}
\end{specialtable}
We conclude, that in our federated learning setting, differential privacy is an effective countermeasure against sample reconstruction from gradients, and no sensitive information could be inferred from the reconstructed images.
\section{Discussion}
\label{sec:discussion}
In our federated learning setup, the effectiveness of model aggregation is limited by data heterogeneity and imbalance. The federated averaging algorithm weights the local model updates with respect to the clients' dataset size in relation to the overall amount of available data \cite{mcmahan2017CommunicationEfficientLearning}. This led to a strong emphasis on model updates from clients with large CheXpert subsets in our case, while updates from clients with fewer images contributed less to model aggregation. One option to mitigate data imbalance is to aggregate models after a specified number of batches instead of local epochs \cite{kaissis2021EndtoendPrivacy, wang2019AdaptiveFederated}. However, sharing intermediate models more frequently will increase the susceptibility to reconstruction attacks since the updates are obtained on small batches rather than the client's whole dataset. Improving the aggregation process under consideration of privacy costs is left for future work.
The applied attack has shown that the two considered deep machine learning models are susceptible to reconstruction of sensitive data from gradients. Most notably, and contrary to previous work, we found that the attack was more successful in later training stages and for pre-trained models.
Our privacy evaluation framework is limited by the choice of the \textit{Deep Leakage from Gradients} attack as a qualitative measure for model vulnerability. Even though we found reconstruction not successful under certain conditions, including full model training, restricting training to the final layer, and attacking the DenseNet121 at an early training stage, it cannot be assumed that model training would be privacy-preserving in these cases. Minor modifications of the attack scheme may improve attack success even in supposedly safe settings. Moreover, reconstruction attacks are only one example among a range of deliberate privacy breaches that neural networks are vulnerable to. Extending our privacy evaluation framework to include other privacy threats, e.g., property inference without data reconstruction, will provide further insights into potential vulnerabilities of the federated learning paradigm. Since the main limitation of \textit{Deep Leakage from Gradients} is its restriction to small datasets, it will be particularly valuable to capture the consequences of privacy breaches for clients with large amounts of data. From a security perspective, demonstrating that these attacks are practically feasible, albeit under limited circumstances, is sufficient for considering the machine learning process vulnerable to privacy violation. Countermeasures must constantly be re-evaluated for their effectiveness as a better understanding of privacy threats evolves.
Our privacy evaluation was further constrained to a limited choice of privacy budgets. While choosing $\varepsilon=6$ delivered the best utility-privacy trade-off for our use case, which is in line with previous work \cite{kaissis2021EndtoendPrivacy}, it may not be the optimal lower bound. We specifically suggest empirically examining choices of $\varepsilon$ in the range $[3,6]$ to potentially improve upon our results in future work. We also note that while all considered privacy constraints prevented the success of reconstruction attacks, smaller $\varepsilon$ values still formally provide stronger privacy guarantees that offer protection against threats beyond the limited case of the reconstruction attack.
A key implication of our results is that the DenseNet121 architecture proved more robust against private model training with regard to performance than the ResNet50. This observation is potentially related to the greater ability of the DenseNet121 to withstand reconstruction attacks. Although the model contains overall fewer parameters than the ResNet50, its dense structure may, to a certain degree, offer a natural defense against reconstruction from trained parameters as well as perturbation of parameter updates during private training. This outcome suggests substantial differences in the suitability of individual model types for privacy-preserving machine learning, which requires further validation.
In the medical context, fairness is crucial for the safe deployment of machine learning algorithms. Rare diseases or conditions must be reliably detected despite the restricted availability of representative data. Furthermore, a model should perform with equal accuracy on all patient subgroups. We uncovered performance differences between individual clients' data in our federated learning baseline, revealing that the classification produced better results on Mendeley than on CheXpert data. This potentially reflects the ability of deep learning models to recognize pneumonia as an abnormal finding particularly well since the pathologic X-rays in the Mendeley dataset only include cases of pneumonia. More thorough investigations are required to reveal other potential biases, e.g., with respect to patient subgroups. It is further known that underrepresented classes and population subgroups are potentially affected more strongly by model performance degradation when applying differential privacy~\cite{bagdasaryan2019DifferentialPrivacy}. Because model performance evaluation on Mendeley clients was less reliable due to smaller amounts of data, it remains an open question how exactly these clients were affected by the integration of differential privacy. For practical applications, it is mandatory to thoroughly investigate how privacy mechanisms affect the model's performance on different types of data to identify a potential underlying bias.
We did not consider the practical implementation of federated learning between different institutions with regard to communication time, required infrastructure, costs and validation of correct computational execution. The focus of our paper lies in analyzing the threat of data reconstruction and the effectiveness of differential privacy against it. While simulated use cases like ours are vital to prepare for leveraging differential privacy in real-world cases where sensitive data is involved, further case studies are required to investigate aspects of practicability for privacy-preserving federated learning on a large scale.
\section{Conclusions and Future Work}
\label{sec:conclusion}
We simulated a collaborative machine learning use case in which 36 institutions provide their diverse chest X-ray data collections for the development of a classification model. Two main concerns in this scenario are the physical separation of the data sources and the privacy of patients to whom the data belongs. We employed the paradigm of federated learning as a solution for machine learning on dispersed data. Throughout our experiments, we compared two large network architectures: DenseNet121 and ResNet50. Extending previous evidence, we demonstrated that individual X-rays can be reconstructed from shared model updates within the federated learning setting from those networks using the \textit{Deep Leakage from Gradients}. The attack is especially successful during later training stages.
As a step towards privacy-preserving distributed learning, we integrated Rényi differential privacy with a Gaussian noise mechanism into the federated learning process. The DenseNet121 achieved the best utility-privacy trade-off with a mean AUC of $0.937$ for $\varepsilon=6$, where we identified an expected cost in accuracy of $0.03$ in terms of the AUC on CheXpert clients' data compared to the non-private baseline. The results suggest that $\varepsilon \in [3,6]$ are suitable candidates for private model training depending on the specific demands on model privacy and performance for the respective application. Overall, we found the DenseNet121 model superior to ResNet50 with regard to private model training for all considered $\varepsilon$ values.
The adverse impact of differential privacy on model performance must be carefully considered, particularly for medical use cases. Our results endorse that differentially private federated learning is feasible at a small cost in model accuracy for the classification of heterogeneous chest X-ray data. As real-world medical use cases become more complex in practice, future work may elaborate on the potential of differentially private federated learning for multi-label X-ray classification where heterogeneous data from a broader range of sources is effectively integrated under consideration of an improved bound on the privacy budget. We identified the DenseNet121 as a robust model architecture suitable for differentially private training. Further comparison with other neural network architectures may reveal key indicators for the suitability of different model types and provide guidance in the choice of models for privacy-preserving machine learning. We further suggest to extend our evaluation framework in future work to consider the vulnerability to other types of privacy breaches, enabling a comprehensive qualitative assessment of model privacy. Finally, other variants of differential privacy, e.g. Gaussian differential privacy \cite{dong2019GaussianDifferential}, may offer suitable alternatives to the application of Rényi differential privacy providing yet tighter bounds on the privacy loss.
\vspace{6pt}
\clearpage
\authorcontributions{Conceptualization, J.Z., B.P., H.S., A.S. and B.A.; methodology, J.Z.; software, J.Z.; validation, J.Z. and B.P.; formal analysis, J.Z.; investigation, J.Z.; resources, J.Z., B.P. and B.A.; data curation, J.Z.; writing---original draft preparation, J.Z.; writing---review and editing, J.Z., B.P., H.S., A.S. and B.A.; visualization, J.Z.; supervision, B.P., H.S., A.S. and B.A.; project administration, B.A., B.P. and J.Z.; funding acquisition, B.A. \\All authors have read and agreed to the published version of the manuscript.}
\funding{This research received no external funding.}
\dataavailability{The CheXpert dataset \cite{irvin2019CheXpertLarge} is a public chest X-ray dataset available at \url{https://stanfordmlgroup.github.io/competitions/chexpert/}. The Mendeley dataset version 3 can be found at \url{https://data.mendeley.com/datasets/rscbjbr9sj/3}.}
\acknowledgments{Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Projektnummer 491466077}
\abbreviations{The following abbreviations are used in this manuscript:\\
\noindent
\begin{tabular}{@{}ll}
AUC & Area under the receiver operating characteristics curve\\
DLG & Deep Leakage from Gradients\\
DP-SGD & Differentially private stochastic gradient descent\\
MAE & Mean absolute error\\
MSE & Mean squared error\\
PSNR & Peak signal-to-noise ratio\\
ReLU & Rectified linear unit\\
ROC & Receiver operating characteristic\\
SGD & Stochastic gradient descent
\end{tabular}}
|
\section{Introduction}
\label{sec:Introduction}
Multi-Robot Task Allocation (MRTA) involves coordinating a set of tasks among a team of cooperative robotic systems such that the decisions are free of conflict and optimize a quantity of interest~\cite{gerkey2004formal}. The potential real-world applications of such MRTA problems are immense considering that multi-robotics is one of key emerging directions of robotics research and development, and task allocation is fundamental to most multi-robotic or swarm-robotic operations. Examples include disaster response~\cite{ghassemi2018decmata}, last-mile delivery~\cite{aurambout2019last}, environment monitoring~\cite{espina2011multi}, and reconnaissance~\cite{olson2012progress}.
We specifically focus on a class of MRTA problems that falls into the Single-task Robots, Single-robot Tasks, and Time-extended Assignment (SR-ST-TA) class defined in \cite{gerkey2004formal,nunes2017taxonomy}, that include characteristics such as tasks with time deadlines and workload, and robots with constrained work capacity. Based on iTax taxonomy as defined in \cite{gerkey2004formal}, these problems fall into the In-schedule Dependencies (ID) category.
The MRTA problems considered here involves \textit{mildly-heterogeneous} robots (with different task completion rates), and where the task selections are assumed to take place in a decentralized manner, under full observability of the task environment and peers' decisions.
There exists a notable body of work (mostly non-learning approaches) in solving MRTA problems related to these applications, e.g., graph-based methods~\cite{ghassemi2018decmata}, integer-linear programming (ILP) approaches~\cite{nallusamy2009optimization,toth2014vehicle}, and auction-based methods~\cite{dias2006market}. However, critical challenges continue to persist, especially regarding the scalability with number of robots and tasks, and the ability to adapt to complex problem characteristics without tedious hand-crafting of underlying heuristics.
Existing solutions to multi-Traveling Salesman Problem (mTSP) and Vehicle Routing Problem (VRP) in the literature~\cite{bektas2006multiple,braekers2016vehicle} have addressed analogical problem characteristics of interest to MRTA, albeit in a disparate manner; these characteristics include tasks with time deadlines, and multiple tours per vehicle, with applications in the operations research and logistics communities~\cite{azi2010exact,wang2018multi}.
Integer linear programming or ILP based mTSP-type formulations and solution methods have also been extended to task allocation problems in the multi-robotics domain~\cite{jose2016task}. Although the ILP-based approaches can in theory provide optimal solutions, they are characterized by exploding computational effort as the number of robots and tasks increases~\cite{toth2014vehicle,cattaruzza2016vehicle}.
For example, for the studied SR-ST problem, the cost of solving the exact ILP formulation of the problem, even with a linear cost function (thus an ILP), scales with $\mathcal{O}(n^3m^2h^2)$, where $n$, $m$, and $h$ represent the number of tasks, the number of robots, and the maximum number of tours per robot, respectively~\cite{ghassemi2019decmrta}.
As a result, most online MRTA methods, e.g., auction-based methods~\cite{dias2006market,schneider2015auction,otte2020auctions}, metaheuristic methods \cite{Mitiche2019, VANSTEENWEGEN20093281}, and bi-graph matching methods~\cite{ismail2017decentralized,ghassemi2019decmrta,Ghassemi2020}, use some sort of heuristics, and often report the optimality gap at least for smaller test cases compared to the exact ILP solutions.
In recent years, learning approaches based on Graph Neural Networks or GNN (which can model non-Euclidean data) are being increasingly used to solve planning problems with a Combinatorial Optimization (CO) formulation, e.g., TSP, VRP, Max-Cut, Min-Vertex, and MRTA~\cite{Kool2019, barrett2019exploratory, khalil2017learning, Kaempfer2018LearningTM, mittal2019learning, li2018combinatorial, nowak2017note, 9116987, Tolstaya2020MultiRobotCA, Sykora2020, Dai2017}.
These existing studies are however limited in three key aspects: \textbf{1)} They address simplified problems that often exclude common real-world factors such as resource and capacity constraints \cite{Kool2019, Kaempfer2018LearningTM, khalil2017learning, Tolstaya2020MultiRobotCA}). \textbf{2)} They are mostly focused on smaller sized problems ($\leq$ 100 tasks)~\cite{Paleja2020,Malcolm2005, 9116987,Sykora2020}, with their scalability remaining unclear. \textbf{3)} They rarely provide evidence of generalizing to problem scenarios that are larger in size than those used for training. This latter capability is particularly critical since real-world MRTA problems often involve simulating episodes whose costs scale with the number of tasks and robots, making re-training efforts burdensome. Moreover the time-sensitive nature of many real-world applications of MRTA \cite{lin2018efficient}, such as in disaster response \cite{Ghassemi2020}, call for a near real-time performance, while involving a large-sized decision space \cite{9435208}. Therefore, near real-time computable good feasible solutions are of essence.
Along with the need to address scalability, we posit that a sequential decision-making approach is favored considering environmental and communication uncertainties and the possibility of new tasks to emerge during the operation \cite{8206001, 9272626}. Therefore, we formulate the MRTA problem such that the decision taken by the robots on visiting the next node is performed in a sequential, decentralized, and asynchronous manner.
To enable scalable policies, we design a novel \textbf{encoder-decoder} policy network, where the \textbf{encoder} is based on Graph Capsule Convolutional Neural Networks (GCAPCN) \cite{Verma2018}, which is hypothesized to incorporate local and global structural information with permutation invariance. The \textbf{decoder} is based on a Multi-head Attention mechanism (MHA)~\cite{Kool2019, VaswaniSPUJGKP17} which fuses the encoded information and problem-specific \textbf{context} using matrix multiplication, in order to enable decentralized sequential decisions. The proposed network architecture is named Capsule Attention Mechanism or CapAM.
\textbf{Contributions:}
The main contributions of this paper can be summarized as \textbf{1)} Formulate the general SR-ST class of MRTA problems as a Markov Decision Process or MDP over graphs with the multi-robot system's state embedded as the \textit{context} portion of the policy model, such that the task allocation policy can be learned using an RL approach. \textbf{2)} Explore how a policy network that learns the structural information of task-graphs results in better decision-making. \textbf{3)} Demonstrate this learning framework's ability to generalize to larger-sized problems without the need to retrain. \vspace{-.1cm}
\textbf{Paper outline: }
The next section provides a brief overview of MRTA and it's MDP formulation. Section \ref{sec:Learning_framework} then describes our proposed new graph learning architecture that operates over this MDP. Section \ref{sec:Experimental_evaluation} presents the settings and outcomes of numerical experiments on a large number of task allocation problems with time and capacity constraints \cite{Mitiche2019}, which are used to evaluate this new approach to MRTA. This evaluation study includes comparative analysis w.r.t. non-learning methods as well as a well-known attention mechanism based reinforcement learning method, and further parametric and scalability analysis. Finally, Section \ref{sec:Conclusion} provides concluding remarks and discusses the potential future extensions for our new graph learning-based approach to MRTA.
Throughout this paper, \textbf{tasks} are also referred by the terms ``\textbf{nodes}'' or ``\textbf{vertices}''.
\section{MRTA: Problem definition and formulation
\label{sec:problem_formulation}
MRTA problems can be formulated as Integer Linear or Non-Linear Programming (ILP or INLP) problems. When tasks are defined in terms of location, the MRTA problem becomes analogical to the Multi-Traveling Salesmen Problem (mTSP)~\cite{khamis2015multi} and its generalized version, the Vehicle Route Planning (VRP) problem~\cite{dantzig1959truck}, albeit with additional constraints and operation-specific objective function. \vspace{-.2cm}
\subsection{MRTA problem description and formulation}
\label{sec:mrta_optimization}
The exact solution of the MRTA problem can be obtained by formulating it as the following INLP problem:
\vspace{-.2cm}
\begin{align}
\label{eq:objectiveFunction}
\min~ f_\text{cost} = \sum_{i = 1}^{N} r_{i}, \
\begin{cases}
r_{i} = \frac{t_{i}^{f}}{d_{i}},& \text{if } \ t_{i}^{f} > d_{i}\\
0, & \text{otherwise}
\end{cases}
\end{align}
\ \ \ \ \ \text{subject to~}
\begin{align}
\label{eq:constraint1}
& s_{i} \in S \; \forall \; i \in [1..N] \\
\label{eq:constraint2}
& s_{i} \neq s_{j} \; \forall \ i \neq j
\end{align}
\noindent Here $t_{i}^{f}$ is the time at which task $i$ was completed, $d_{i}$ is the time deadline of task $i$, and $S = [s_{1}, s_{2}, ... s_{N}]$ is the sequence of all the tasks/nodes that were visited, and $N$ is the number of tasks available. The minimum cost function that can be achieved using Eq. \ref{eq:objectiveFunction} is $0$, which corresponds to the case where all the tasks are successfully completed.
A detailed formulation of the exact ILP constraints that describe the MRTA problem can be found in this recent work on MRTA ~\cite{Mitiche2019}. Note that, in our paper, we use a slightly different objective/cost function as compared to \cite{Mitiche2019}, but the intention of the objective function of both of these works remains the same, which is to maximize the number of successfully completed tasks (thereby allowing fair comparison with methods reported in the literature).
Here, we craft the objective function (Eq.~\eqref{eq:objectiveFunction}) such that only missed tasks ($t_{i}^{f} > d_{i}$) contribute to the cost function. This objective function can be tailored according to user/application preferences, without affecting the proposed learning framework.
The constraints in Eqs. \ref{eq:constraint1} and \ref{eq:constraint2} are such that each task must be visited exactly once by any robot.
\begin{figure}[!t]
\centering
\includegraphics[width=.9\linewidth]{Figures/Sequential_decision_making.pdf}
\vspace{-0.4cm}
\caption{Deployment of an MRTA policy based on the CapAM architecture. a) Robot-1 at $t_0$. b) Robot-2 at $t_1$; here, the output for previously selected task (e.g., task 2 in (b)) is set at 0.}
\label{fig:sequential_decision_making}
\vspace{-0.6cm}
\end{figure}
\subsection{An ``MDP over Graph" Representation of MRTA}\label{sec:ProblemFormulation}
The task space of an MRTA problem can be represented as a complete un-directed graph $\mathcal{G} = (V, E, \textbf{A})$, which contains a set of nodes/vertices ($V$), a set of edges ($E$) that connect the vertices to each other, and the weighted adjacency matrix $\textbf{A}$ that gives the extent to which two nodes are connected. Each node is a task, and each edge connects a pair of nodes. For MRTA with $N$ tasks, the number of vertices and the number of edges are $N$ and $N(N-1)/2$, respectively. Node $i$ is assigned a 3-dimensional feature vector denoting the task location and time deadline, i.e., $\delta_i=[x_i,y_i,d_i]$. Here we consider a weighted adjacency matrix without self-loop ($\alpha_{ii} = 0, \alpha_{i,j} \in \textbf{A}$, where $ i \ {\text{and}} \ j \in [1, N]$).
The weights in the adjacency matrix $\alpha_{ij} = 1/(1+ |\delta_i - \delta_j|)$ take a real value between 0 and 1, in a manner such that if the features of two nodes $i$ and $j$ are relatively closer to each other, then $\alpha_{ij}$ is higher.
Since we formulate MRTA as an undirected graph, the adjacency matrix $A$ is symmetric. In this paper, the elements of $\textbf{A}$ ($\alpha_{i,j}$, $i \neq j$) are computed as the inverse of the normalized Euclidean distance of the node properties.
The MDP can be defined in a decentralized manner to capture the myopic task-assignment process of each individual robot, which can be expressed as a tuple $<\mathcal{S}, \mathcal{A}, \mathcal{P}_a, \mathcal{R}>$. Figure \ref{fig:sequential_decision_making} shows how this MDP is executed to perform the sequential task selection by each robot. As evident from Fig. \ref{fig:sequential_decision_making}, the components of the MDP can then be defined as follows: \textbf{State Space ($\mathcal{S}$):} A robot at any decision-making instance uses a state $s\in\mathcal{S}$, which contains the following information:
1) elapsed mission time, 2) robot's current location, 3) constraints of this robot such as its work capacity, 4) the planned tasks of its peers, and 5) the constraints of its peers. The state space also includes the environment, defined by the location and time deadline of tasks, which are expressed here as node features in the graph. Here we assume that each robot can broadcast its information, which includes the task it is currently engaged in and its work capacity, to its peers without the need for a centralized system for communication; this is well aligned with modern communication capabilities \cite{Sykora2020}.
\noindent\textbf{Action Space ($\mathcal{A}$):} The set of actions is represented as $\mathcal{A}$, where each action $a\in\mathcal{A}$ is defined as the index of the selected task, $\{1,\ldots,N\}$. In MRTA problems that include a depot, task $0$ can be considered as the depot. The task status, i.e., \textit{active}, \textit{completed}, and \textit{missed} (i.e., deadline is passed), is used to properly decode the actions generated by the policy model, with the latter being in the form of probability of selecting each available task. Note that even under full observability, as the action space of a single robot includes only its own degree of affinity for selecting each available task, the decentralized decision-making formulation differs from an ideal centralized formulation, and remains readily scalable. \textbf{Transition ($\mathcal{P}_a$):} The transition is an event-based trigger. An event is defined as the condition that a robot reaches its selected task or visits the depot. As environmental uncertainties and communication issues (thus partial observation) are not considered in this paper, only deterministic state transitions are allowed.
\textbf{Reward ($\mathcal{R}$):} A delayed reward is considered here, estimated at the end of the multi-robot simulation, i.e, when there are no more active tasks. The actual reward function is typically application-dependent. For example, in some MRTA problems, task completion is more important, while in another, decreasing the total effort (e.g., distance traveled) by robots is more important, which can then be estimated using the weight matrix.
\vspace{-.1cm}
\section{Learning framework}
\label{sec:Learning_framework}
\vspace{-.2cm}
In this work, we implement an RL algorithm on an \textbf{encoder-decoder} architecture to learn the optimal MRTA policies over the graph formulation described earlier. An illustration of this architecture is shown in Fig. \ref{fig:encoder_decoder}. The \textbf{encoder} represents each node as a continuous vector infusing both its own properties as well as its local and global structural properties within the graph representation of the task space. The decoder sequentially computes output probabilities for all the available tasks/nodes using the information from the encoder and the current state (\textbf{context}), by applying the MHA (as shown in left part of Fig. \ref{fig:encoder_decoder}). The sequentially computed output probabilities for each node indicate the relative goodness of selecting that node as the next one to be visited by the concerned robot. A detailed description of the architecture of the encoder and decoder is discussed in the following sections using the representative SR-ST type MRTA problem described in section \ref{sec:ProblemFormulation}.
\begin{figure}[!t]
\centering
\includegraphics[width=0.46\textwidth]{Figures/GCAPCN_NIPS.pdf}
\vspace{-0.3cm}
\caption{CapAM architecture including context, encoder \& decoder.}
\label{fig:encoder_decoder}
\vspace{-0.7cm}
\end{figure}
\subsection{Graph Capsule Convolutional Network based Encoder}
\label{Sec:Encoder}
The main purpose of the encoder is to represent useful information related to a node/task as a learnable continuous vector or tensor, which can then be used by the learning algorithm. For a graph node, the information includes the properties of the node itself (e.g., the coordinates and time deadline of the node), the local neighborhood information for the node, and also the global structural information. General solutions to CO problems such as CVRP, MTSP and MRTA should be permutation invariant, which means that the order by which each node is numbered should not affect the optimal solution. Hence the node encoding must also be permutation invariant. In this work, we are exploring how a Graph Capsule Convolutional Neural Network (GCAPCN) can be implemented for learning local and global structures with the node properties, with permutation invariant node embedding. GCAPCN is a class of Graph Neural Networks (GNN), introduced in \cite{Verma2018} to address the drawbacks (e.g., permutation invariance) of Graph Convolutional Neural Networks (GCN), and to enable the encoding of global information based on \textbf{capsule networks} presented in \cite{Hinton}. The advantage of GCAPCN lies in capturing more local and global information, compared to conventional aggregation operations used in GNN such as summation or standard convolution operations.
\subsubsection{Graph Capsule Convolutional Neural Networks}
\label{Sec:GCAPCN}
Let $X \in \mathbb{R}^{N \times |\delta_{i}|}$, be the node feature matrix, where $|\delta_{i}|$ is the input dimension for each node $i$. The standard graph Laplacian is defined as $L = D - A \in \mathbb{R}^{N \times N}$, where $D$ is the degree matrix and $A$ is the adjacency matrix of the graph. A capsule vector is computed using a Graph Capsule function based on different order of statistical moments, as shown in the equations later in this section.
We first compute a feature vector $F_{0i}$ for each node by linear transformation of the node properties $\delta_{i}$, as $ F_{0i} = \delta_{i}.W_{0} + b_{0}$ for all $i \in [1,N]$, where
$W_{0} \ \varepsilon \ \mathbb{R}^{h_{0} \times |\delta_{i}|}, \ \ b_{0} \ \varepsilon \ \mathbb{R}^{h_{0} \times 1}$, and $h_{0}$ is the length of the feature vector. For an MRTA problem with time constraints, $\delta_{i} = [x_{i}, y_{i}, d_{i}]$, where $x_i$, $y_i$, and $d_i$ are the $x$ coordinate, $y$ coordinate, and the time deadline of location/node $i$. Note that this method for encoding is not limited to a specific MRTA problem, and can also be extended or modified for MRTA problems with nodes having other features such as workload.
Each feature vector $F_{0i}, i \in [1,N]$ is then passed through a series of Graph capsule layers, where the output from the previous layers is used to compute a matrix $f_{p}^{(l)} (X, L)$ using a graph convolutional filter of polynomial form as given by:
\vspace{-.3cm}
\begin{equation}
\label{Eq:GraphCapsuleFunction}
f_{p}^{(l)} (X, L) = \sigma(\sum_{k=0}^{K}L^{k}(F_{(l-1)}(X,L)^{\circ p})W_{pk}^{(l)}) \vspace{-.1cm}
\end{equation} \vspace{-.2cm}
\noindent Here $L$ is the graph Laplacian, $p$ is the order of the statistical moment, $K$ is the degree of the convolutional filter, $F_{(l-1)}(X,L)$ is the output from layer $l -1$, $F_{(l-1)}(X,L)^{\circ p}$ represents $p$ times element-wise multiplication of $F_{(l-1)}(X,L)$.
Here, $F_{(l-1)}(X,L) \in \mathbb{R}^{N \times h_{l-1}p}$, $W_{pk}^{(l)} \in \mathbb{R}^{h_{l-1}p \times h_{l}}$. The variable $f_{p}^{(l)} (X, L) \in \mathbb{R}^{N \times h_{l}}$ is a matrix where each row is an intermediate feature vector for each node $i \in [1,N]$, infusing nodal information from $L_{e} \times K$ hop neighbors, for a value of $p$. The output of layer $l$ is obtained by concatenating all $f_{p}^{(l)} (X, L)$, as given by:
\vspace{-.2cm}
\begin{equation}
\label{Eq:LayerOutput}
F_{l}(X,L) = [f_{1}^{(l)} (X, L), f_{2}^{(l)} (X, L), ...f_{P}^{(l)} (X, L)]
\end{equation}
\noindent
Here $P$ is the highest order of statistical moment, and $h_{l}$ is the node embedding length of layer $l$. We consider all the values of $h_{l}$ (where $l \in [0, L_{e}]$) to be the same for this paper.
Equations \ref{Eq:GraphCapsuleFunction} and \ref{Eq:LayerOutput} were computed for $L_{e}$ layers, where each layer uses the output from the previous layer ($F_{l-1}(X,L)$). Adding more layers helps in learning the global structure, however, this can affect the performance by increasing the number of learnable parameters (compared to the size of the problem), leading to over-fitting.
The output from the final layer is then passed through a feed-forward layer so that the final feature vector has the right dimension ($h_l$) to be fed into the decoder as shown in Fig. \ref{fig:encoder_decoder}.
\vspace{-.2cm}
\subsection{Attention-based Decoder and Context}
\label{Sec:Decoder}
The main objective of the MHA-based decoder is to use the information from the encoder and the current state as context or query, and thereof choose the best task by calculating the probability value of getting selected for each (task) node. In this case, the first step is to feed the embedding for each node (from the encoder) as \textbf{key-values} ($\mathcal{K}$, $V$), since inputs for MHA are key-value pairs \cite{DBLP:journals/corr/VaswaniSPUJGKP17}. The key $K$ and value $V$ for each node is computed by two separate linear transformations of the node embedding obtained from the encoder. The next step is to compute a vector representing the current state, also known as the \textbf{context} (as shown in bottom left of Fig. \ref{fig:encoder_decoder}).
In our case studies, the context for the multi-head attention or MHA layer in the decoder consists of the following five features: \textbf{1)} elapsed mission time; \textbf{2)} work capacity of the robot taking decision; \textbf{3)} current location of the robot taking decision; \textbf{4)} current destination of robot's peers; and \textbf{5)} work capacity of peer; all concatenated to a single vector of length $h_{q}$, which then undergoes a linear transformation to get a vector of length $h_{l}$ also called the query $Q$. It is important to note that the \textbf{context} can be modified to account for other features such as remaining ferry-range and payloads of robots, thus promoting the flexibility to apply this architecture to a wider range of MRTA problems than those specifically studied in this paper.
Figure \ref{fig:encoder_decoder} illustrates the structure of the decoder.
Now the attention mechanism can be described as mapping the query ($Q$) to a set of key-value ($\mathcal{K},V$) pairs. The inputs, which are the query ($Q$) is a vector, while $\mathcal{K}$ and $V$ are matrices of size $h_{l} \times N$ (since there are $N$ nodes). The output is a weighted sum of the values $V$, with the weight vector computed using the compatibility function expressed as:
\vspace{-.2cm}
\begin{align}
\label{attention_function}
\text{Attention}(Q,\mathcal{K},V) = \text{softmax}(Q^{T}\mathcal{K}/\sqrt{h_{l}})V^{T}
\vspace{-.2cm}
\end{align}
Here $h_{l}$ is the dimension of the key of any node $i$ ($k_{i} \in \mathcal{K}$).
In this work, we implement a MHA layer in order to determine the compatibility of $Q$ with $\mathcal{K}$ and $V$. The MHA implemented in this work is similar to the decoder implemented in \cite{Kool2019} and \cite{DBLP:journals/corr/VaswaniSPUJGKP17}.
As shown in \cite{DBLP:journals/corr/VaswaniSPUJGKP17} the MHA layer can be defined as:
\vspace{-.2cm}
\begin{equation}
\label{mha}
\text{MHA}(Q,\mathcal{K},V) = \text{Linear}(\text{Concat}(\text{head}_{1}\dots \text{head}_{h_{e}}))
\vspace{-.2cm}
\end{equation}
Here $\text{head}_{i} = \text{Attention}(Q,\mathcal{K},V)$ and $h_{e}$ (taken as 8 here) is the number of heads.
The feed-forward layer is implemented to further process the mapping that results from the MHA layer, and transform it to a dimension that is coherent with the number of nodes in the task-graph ($N$). The interjecting batch normalization layers serve to bound values of a specific batch using the mean and variance of the batch.
The final \texttt{softmax} layer outputs the probability values for all the nodes. Here, the next task to be done is then chosen based on a greedy approach, which means that the node with the highest probability will be chosen. The nodes which are already visited will be masked (by setting their probability as $0$) so that these nodes are not available for selection in the future time steps of the simulation of the multi-robot operation.
\vspace{-.2cm}
\subsection{Training Algorithm}
\label{Sec:LearningAlgorithm} \vspace{-.1cm}
The training algorithm used here is REINFORCE.
Other relevant details for training are shown in Table \ref{table:training_algo_info_MRTA}. This training process can readily be further advanced in the future through the adoption of advanced state-of-the-art policy gradient algorithms such as Proximal Policy Optimization \cite{schulman2017proximal} and the Actor-Critic method \cite{Bahdanau2017}.
For each epoch of the training, two sets of data (MRTA operation scenarios) are used, which are the training set and the validation set. The training data set is used to train the model ($\theta_\text{CapAM}$) while the validation set is used to update the baseline model ($\theta_\text{CapAM}^{BL}$). Both the training data and validation data generated from the distribution of MRTA scenarios are explained later in section \ref{Sec:Baseline_method_dataset}.
The size of the training data and the validation data used in this paper is mentioned in Table~\ref{table:training_algo_info_MRTA}.
\vspace{-.1cm}
\section{Experimental evaluation}
\label{sec:Experimental_evaluation}
\vspace{-.2cm}
\subsection{Comparison with baselines}
\label{Sec:BaselineComparison}
\subsubsection{Details on dataset for training and testing}
\label{Sec:Baseline_method_dataset}
\begin{wraptable}[10]{R}{4.7cm}
\vspace{-.9cm}
\footnotesize
\caption{Settings for model training in CapAM and AM-RL}
\vspace{-0.7cm}
\label{table:training_algo_info_MRTA}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\footnotesize
\begin{tabular}{|l|c|}
\toprule
Details & Values \\
\midrule
\textit{Algorithm} & \textit{REINFORCE} \\
\textit{Baseline} & \textit{Rollout} \\
\textit{Epochs} & 100\\
\textit{Training samples} & 500,000 \\
\textit{Validation samples} & 10,000 \\
\textit{Optimizer} & \textit{Adam} \\
\textit{Learning step size} & 0.0001 \\
\textit{Training frequency} & 500 samples \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\vspace{-0.4cm}
\end{wraptable}
We consider the (NP-hard) MRTA problem with time and capacity constraints also known as Task Allocation Problem with Time and Capacity (TAPTC), as described in \cite{Mitiche2019}.
Each training sample has 100 tasks, and are located randomly within a 100 $\times$ 100 grid map. Each task $i$ has a time deadline $50 \leq d_{i} \leq 600$, and a workload $10 \leq w_{i} \leq 30$. Each sample has $n_r$ number of robots where $2 \leq n_r \leq 7$. The initial positions of the robots in a sample is also chosen randomly within the grids.
Each robot $j$ has a work capacity of $c_{j}$ where $1 \leq c_{j} \leq 3$. All the robots move at a speed of $1$ unit. A task $i$ is considered to be completed only if a robot $j$ visits node $i$ and spends a time of $w_{i}/c_{j}$. All the training samples are generated such that all the associated variables (mentioned above) follow a uniform distribution within their respective bounds.
The average training time per epoch for CapAM is $~\sim$10 mins.
In order to evaluate and compare the generalizability of CapAM-trained models with that of baselines,
we used the dataset in \cite{Mitiche2019} (found in \cite{dataset}). This dataset, which has a similar distribution of scenarios as used in training CapAM, consists of 96 test cases with 100 tasks and a varying number of robots ($2, 3, 5, 7$ robots), and the speed of every robot is considered to be the same (1 unit/s). The 96 cases are divided into ones with tight and slack deadlines. Each category can be further divided into 4 sub-categories based on the fraction of tasks ($25\%$, $50\%$, $75\%$, and $100\%$) that have normally distributed tasks deadlines.
For example, $25\%$ indicates that there are 25 tasks with deadlines normally distributed between the limits $d_{\texttt{low}}$ and $d_{\texttt{high}}$, while the remaining 75 tasks have a deadline of $d_{\texttt{high}}$.
For group 1 (tight deadline), the value of $d_{\texttt{high}}$ is half of that for group 2 (slack deadline). Further description of the test cases can be found in \cite{Mitiche2019}.
\subsubsection{Test for Generalizability}
\label{Sec:TestForGeneralizability}
We compare our method with the following four baseline methods over the stated test cases:
\textbf{i)} \textbf{Iterated Local Search (ILS)}: This is an online metaheuristic iterated search algorithm \cite{VANSTEENWEGEN20093281}, where the output of one iteration is partially used as the input to the next iteration. During each iteration, the best solution is improved by a perturbation step, followed by a local search.
\textbf{ii)} \textbf{Enhanced Iterated Local Search (EILS)}: EILS is also an online metaheuristic iterated search method \cite{Mitiche2019}, with an improved perturbation step as compared to \cite{VANSTEENWEGEN20093281}.
\textbf{iii)} \textbf{Bi-Graph MRTA (BiG-MRTA)}: BiG-MRTA~\cite{ghassemi2019decmrta} is an online method based on the construction and matching of a bipartite graph. In this method, a bipartite graph is constructed to connect robots to tasks, with the weights of connecting edges determined by an incentive model as a function of the tasks' features and robots' states. This decomposes the problem and yield a measure of robot-task pairing suitability. A maximum weighted matching problem is then solved by each robot to identify the optimal task assignments that maximize a net incentive for the team.
\textbf{iv)} \textbf{Attention Mechanism based RL (AM-RL)}: The AM-RL method \cite{Kool2019} consists of an encoder-decoder architecture purely based on attention mechanism. To implement this method for our problem, it is adapted to a multi-robot setting by making the following changes: 1) The node properties defined in Section~\ref{sec:ProblemFormulation} are used here; 2) The context for the MHA in the decoder is modified to be the same as that in CapAM; and 3) The cost function used for training is changed to that in Eq.~\eqref{eq:objectiveFunction}. We set the parameters which correspond to the best performing model for a CVRP with 100 locations. This AM-RL model is trained using the same sample distribution and settings as respectively described in Sections \ref{Sec:Baseline_method_dataset}. The average training time per epoch is found to be $\sim$11 mins for AM. The architectural difference between AM-RL and CapAM is mainly in the encoder, with their performance comparison intended to show the impact of better encoding for learning over graphs.
\begin{wraptable}[10]{r}{3.6cm}
\vspace{-.7cm}
\footnotesize
\caption{Comparison of computing time to generate entire solution sequence, averaged over 96 cases}
\vspace{-0.75cm}
\label{table:computation_time}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\footnotesize
\begin{tabular}{|l|c|}
\toprule
Method & Time (sec) \\
\midrule
\textit{ILS} & ~1.342 \\
\hline
\textit{EILS} & ~0.992 \\
\hline
\textit{BiG-MRTA} & ~0.301 \\
\hline
\textit{AM-RL} & ~0.086 \\
\hline
\textit{CapAM} & ~0.085 \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vspace{-.1cm}
\end{wraptable}
The computation time to generate the entire sequence of task assignments, as required by the non-learning baselines and the learnt policies of CapAM and AM, are compared in Table \ref{table:computation_time} -- reported as an average over the 96 test cases. The real-time performance advantage of the learnt models, which are $>$10 times faster than the non-learning online methods (ILS/EILS/BiG-MRTA) is readily evident here.
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{Figures/gini_test_case_comparison.pdf}
\vspace{-0.3cm}
\caption{Comparison of the performance (on test cases) of CapAM with other baseline methods. Here CapAM has $K$ = 2, $P$ = 3, and $L_{e}$ = 1.}
\label{fig:gini_test_cases}
\vspace{-0.7cm}
\end{figure}
Figure \ref{fig:gini_test_cases} shows the box plots for task completion performance over the 96 test cases divided into 4 categories (each with 24 cases) based on the number of robots (2,3,5, and 7). As seen from Fig. \ref{fig:gini_test_cases}, CapAM obtains a better performance compared to the baseline methods. To provide statistical evidence of the advantage of CapAM over the baseline methods, we performed a pairwise T-test of CapAM \textit{vs} each baseline on the 24 cases in each category, with results summarized in Table \ref{table:ttest_pvals}; here, the null hypothesis is that the difference between the values of the two sets has a mean equal to $0$. Considering a 5$\%$ significance level, Table \ref{table:ttest_pvals} shows that the \textbf{p-value} from the T-test is generally less than $0.05$, which indicates the rejection of the null hypothesis -- CapAM's performance on the test cases is thus significantly better than of the baselines.
\begin{table}[H]
\vspace{-.4cm}
\footnotesize
\centering
\caption{\textbf{p-value} for pairwise (CapAM vs. baselines) T-Test corresponding to Fig. \ref{fig:gini_test_cases}.} \vspace{-.3cm}
\label{table:ttest_pvals}
\begin{tabular}{|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{\# robots}} & \multicolumn{4}{c|}{\textbf{CapAM vs}} \\ \cline{2-5}
& \textbf{ILS} & \textbf{EILS} & \textbf{BiG-MRTA} & \textbf{AM-RL} \\ \hline
\textbf{2} & 2.03e-4 & 0.002 & 0.003 & 1.6e-11 \\ \hline
\textbf{3} & 0.0307 & 0.079 & 0.038 & 2.4e-8 \\ \hline
\textbf{5} & 0.024 & 0.163 & 8.8e-4 & 0.42 \\ \hline
\textbf{7} & 0.025 & 0.049 & 0.003 & 0.031 \\ \hline
\end{tabular}%
\vspace{-.4cm}
\end{table}
\subsection{Parametric and Scalability Analysis}
\label{Sec:Parametric-ScalabilityAnalysis}
\vspace{-.2cm}
Here, parametric analyses is performed w.r.t. the key parameters of $P$ and $L_{e}$ in the encoder of CapAM. Higher values of $P$ promotes greater information infusion for a given neighborhood; with increasing $L_{e}$ or $K$ (and fixed $P$), information from (great number of) $L_{e} \times K$ hop neighbours is considered in the node embeddings. Increasing $K$, $P$ and $L_{e}$ however increases the number of learnable parameters, which can lead to over-fitting and deteriorating performance for unseen cases.
Our numerical experiments with higher values of $K$ and $L_{e}$ corroborated this expectation, and hence we fixed $K=2$ and $L_{e}=1$.
Hence this section focuses on the impact of $P$, particularly in generalizing, and scaling to larger-sized problems beyond training -- both of which are premised as major advantages of learning local structural information.
To this end, we train 3 different models on the 100-task problem, with varying $P$ from 1 to 3, and fixed $K$ and $L_{e}$.
Then the CapAM (and for comparison AM-RL) models that are trained for 100 tasks and 2-7 robots are implemented on two sets of unseen problems -- one with the same size as in training, and another with numbers of tasks and robots scaled up by a factor of 2 (training/testing tasks are drawn from the same distribution).
Table \ref{table:CAM_P_var} shows the effect of $P$ on scalability of CapAM, which is also compared to AM-RL. Results are reported in terms of task completion $\%$ averaged over 24 samples in each category (based on \# tasks and robots). It can be seen that all CapAM models with any $P$ perform significantly better than AM-RL. For the problems of the same size as training (100-task cases), the task completion \% gap between CapAM and AM-RL is between $3\%$ and $13.5\%$. For the problems of double the size as training (200-task cases), the task completion \% gap between CapAM and AM-RL is between $1.9\%$ and $17.6\%$. In both cases, the gap (CapAM vs. AM-RL) is the highest when task-to-robot ratios are high.
These observations directly demonstrate the effectiveness of the novel encoder used in CapAM, which is designed to capture local structural information much better than in attention mechanisms.
On comparing the CapAM models with different $P$ (in Table \ref{table:CAM_P_var}), it can be seen that ones with $P = 1$ perform slightly poorer compared to corresponding ones with $P=2$ and $P=3$, with the difference between the latter two being marginal. Thus increasing $P$ beyond 2 may not provide further performance improvements.
\begin{table}[H]
\vspace{-.35cm}
\footnotesize
\caption{Performance of CapAM (in terms of \% of task completed) with $P=1,2,3$ for unseen problem sets with the same and twice the number of tasks and robots as in training.}
\vspace{-.3cm}
\label{table:CAM_P_var}
\resizebox{.48\textwidth}{!}{%
\footnotesize
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\textbf{\# tasks} &
\multicolumn{1}{l|}{\textbf{\# robots}} &
\multicolumn{1}{l|}{\textbf{CapAM($P$=3)}} &
\multicolumn{1}{l|}{\textbf{CapAM($P$=2)}} &
\multicolumn{1}{l|}{\textbf{CapAM($P$=1)}} &
\multicolumn{1}{l|}{\textbf{AM-RL}} \\ \hline
\multirow{4}{*}{\textbf{100}} & \textbf{2} & 34.6 & \textbf{36.2} & 32.6 & 22.7 \\ \cline{2-6}
& \textbf{3} & \textbf{51.1} & 50.3 & 49.4 & 38.0 \\ \cline{2-6}
& \textbf{5} & \textbf{74.9} & \textbf{74.9} & 73.2 & 70.2 \\ \cline{2-6}
& \textbf{7} & \textbf{94.8} & 94.1 & 91.2 & 85.6 \\ \hline
\multirow{4}{*}{\textbf{200}} & \textbf{4} & \textbf{28.2} & 27.1 & 24.4 & 10.7 \\ \cline{2-6}
& \textbf{6} & 46.7 & \textbf{47.0} & 45.2 & 29.4 \\ \cline{2-6}
& \textbf{10} & 67.8 & \textbf{68.6} & 66.5 & 61.4 \\ \cline{2-6}
& \textbf{14} & \textbf{77.8} & 77.5 & 76.2 & 74.3 \\ \hline
\end{tabular}%
}
\vspace{-.4cm}
\end{table}
\footnote{The codes for this paper can be found in \url{https://github.com/adamslab-ub/CapAM-MRTA}}.
\section{Conclusion}
\label{sec:Conclusion}\vspace{-.2cm}
In this paper, we proposed a new graph neural network architecture called CapAM, to learn policies for MRTA problems involving tasks with time deadlines and robots with work capacity. This new architecture incorporates an encoder based on capsule networks and a decoder based on the attention mechanism, along with a context module to feed the state of the robots and the environment. To learn the features of the encoder and decoder, the problem has been posed as an RL problem and solved using the policy gradient algorithm, REINFORCE. In addition, the proposed architecture is found to provide effective MRTA policies over varying task size and robot-team size. The new CapAM architecture demonstrated better performance compared to other state-of-the-art baselines, which include both a learning based method (AM-RL) and non-learning based methods (ILS, EILS, BiG-MRTA).
The CapAM architecture with its capsule based node embedding showed that learning local and global structural information of the task graph results in better generalizability over unseen test cases, as observed from comparing its performance with AM-RL (that uses a different node embedding). The computational cost analysis showed that the trained CapAM model takes only a few milliseconds to yield a task-assignment decision, making it highly suited for operations that require time-sensitive online decisions. The advantage of using local neighborhood information was also empirically evident from the scalability analysis on the defined MRTA problem, where CapAM demonstrated superior performance when applied to graphs with a larger number of tasks/nodes and robots in the team.
\textbf{Future directions:} Firstly, computational analysis of training feasibility when episodes must be simulated over more realistic virtual robotics environment remains a open direction of research across all learning-based approaches. Further, to enable transition of these MRTA methods to application, in the future we should consider dynamic tasks, environment uncertainties, and partially observable state spaces within the CapAM architecture.
\bibliographystyle{./bibliography/IEEEtran}
\section{Introduction}
\label{sec:Introduction}
Multi-Robot Task Allocation (MRTA) involves coordinating a set of tasks among a team of cooperative robotic systems such that the decisions are free of conflict and optimize a quantity of interest~\cite{gerkey2004formal}. The potential real-world applications of such MRTA problems are immense considering that multi-robotics is one of key emerging directions of robotics research and development, and task allocation is fundamental to most multi-robotic or swarm-robotic operations. Examples include disaster response~\cite{ghassemi2018decmata}, last-mile delivery~\cite{aurambout2019last}, environment monitoring~\cite{espina2011multi}, and reconnaissance~\cite{olson2012progress}.
We specifically focus on a class of MRTA problems that falls into the Single-task Robots, Single-robot Tasks, and Time-extended Assignment (SR-ST-TA) class defined in \cite{gerkey2004formal,nunes2017taxonomy}, that include characteristics such as tasks with time deadlines and workload, and robots with constrained work capacity. Based on iTax taxonomy as defined in \cite{gerkey2004formal}, these problems fall into the In-schedule Dependencies (ID) category.
The MRTA problems considered here involves \textit{mildly-heterogeneous} robots (with different task completion rates), and where the task selections are assumed to take place in a decentralized manner, under full observability of the task environment and peers' decisions.
There exists a notable body of work (mostly non-learning approaches) in solving MRTA problems related to these applications, e.g., graph-based methods~\cite{ghassemi2018decmata}, integer-linear programming (ILP) approaches~\cite{nallusamy2009optimization,toth2014vehicle}, and auction-based methods~\cite{dias2006market}. However, critical challenges continue to persist, especially regarding the scalability with number of robots and tasks, and the ability to adapt to complex problem characteristics without tedious hand-crafting of underlying heuristics.
Existing solutions to multi-Traveling Salesman Problem (mTSP) and Vehicle Routing Problem (VRP) in the literature~\cite{bektas2006multiple,braekers2016vehicle} have addressed analogical problem characteristics of interest to MRTA, albeit in a disparate manner; these characteristics include tasks with time deadlines, and multiple tours per vehicle, with applications in the operations research and logistics communities~\cite{azi2010exact,wang2018multi}.
Integer linear programming or ILP based mTSP-type formulations and solution methods have also been extended to task allocation problems in the multi-robotics domain~\cite{jose2016task}. Although the ILP-based approaches can in theory provide optimal solutions, they are characterized by exploding computational effort as the number of robots and tasks increases~\cite{toth2014vehicle,cattaruzza2016vehicle}.
For example, for the studied SR-ST problem, the cost of solving the exact ILP formulation of the problem, even with a linear cost function (thus an ILP), scales with $\mathcal{O}(n^3m^2h^2)$, where $n$, $m$, and $h$ represent the number of tasks, the number of robots, and the maximum number of tours per robot, respectively~\cite{ghassemi2019decmrta}.
As a result, most online MRTA methods, e.g., auction-based methods~\cite{dias2006market,schneider2015auction,otte2020auctions}, metaheuristic methods \cite{Mitiche2019, VANSTEENWEGEN20093281}, and bi-graph matching methods~\cite{ismail2017decentralized,ghassemi2019decmrta,Ghassemi2020}, use some sort of heuristics, and often report the optimality gap at least for smaller test cases compared to the exact ILP solutions.
In recent years, learning approaches based on Graph Neural Networks or GNN (which can model non-Euclidean data) are being increasingly used to solve planning problems with a Combinatorial Optimization (CO) formulation, e.g., TSP, VRP, Max-Cut, Min-Vertex, and MRTA~\cite{Kool2019, barrett2019exploratory, khalil2017learning, Kaempfer2018LearningTM, mittal2019learning, li2018combinatorial, nowak2017note, 9116987, Tolstaya2020MultiRobotCA, Sykora2020, Dai2017}.
These existing studies are however limited in three key aspects: \textbf{1)} They address simplified problems that often exclude common real-world factors such as resource and capacity constraints \cite{Kool2019, Kaempfer2018LearningTM, khalil2017learning, Tolstaya2020MultiRobotCA}). \textbf{2)} They are mostly focused on smaller sized problems ($\leq$ 100 tasks)~\cite{Paleja2020,Malcolm2005, 9116987,Sykora2020}, with their scalability remaining unclear. \textbf{3)} They rarely provide evidence of generalizing to problem scenarios that are larger in size than those used for training. This latter capability is particularly critical since real-world MRTA problems often involve simulating episodes whose costs scale with the number of tasks and robots, making re-training efforts burdensome. Moreover the time-sensitive nature of many real-world applications of MRTA \cite{lin2018efficient}, such as in disaster response \cite{Ghassemi2020}, call for a near real-time performance, while involving a large-sized decision space \cite{9435208}. Therefore, near real-time computable good feasible solutions are of essence.
Along with the need to address scalability, we posit that a sequential decision-making approach is favored considering environmental and communication uncertainties and the possibility of new tasks to emerge during the operation \cite{8206001, 9272626}. Therefore, we formulate the MRTA problem such that the decision taken by the robots on visiting the next node is performed in a sequential, decentralized, and asynchronous manner.
To enable scalable policies, we design a novel \textbf{encoder-decoder} policy network, where the \textbf{encoder} is based on Graph Capsule Convolutional Neural Networks (GCAPCN) \cite{Verma2018}, which is hypothesized to incorporate local and global structural information with permutation invariance. The \textbf{decoder} is based on a Multi-head Attention mechanism (MHA)~\cite{Kool2019, VaswaniSPUJGKP17} which fuses the encoded information and problem-specific \textbf{context} using matrix multiplication, in order to enable decentralized sequential decisions. The proposed network architecture is named Capsule Attention Mechanism or CapAM.
\textbf{Contributions:}
The main contributions of this paper can be summarized as \textbf{1)} Formulate the general SR-ST class of MRTA problems as a Markov Decision Process or MDP over graphs with the multi-robot system's state embedded as the \textit{context} portion of the policy model, such that the task allocation policy can be learned using an RL approach. \textbf{2)} Explore how a policy network that learns the structural information of task-graphs results in better decision-making. \textbf{3)} Demonstrate this learning framework's ability to generalize to larger-sized problems without the need to retrain. \vspace{-.1cm}
\textbf{Paper outline: }
The next section provides a brief overview of MRTA and it's MDP formulation. Section \ref{sec:Learning_framework} then describes our proposed new graph learning architecture that operates over this MDP. Section \ref{sec:Experimental_evaluation} presents the settings and outcomes of numerical experiments on a large number of task allocation problems with time and capacity constraints \cite{Mitiche2019}, which are used to evaluate this new approach to MRTA. This evaluation study includes comparative analysis w.r.t. non-learning methods as well as a well-known attention mechanism based reinforcement learning method, and further parametric and scalability analysis. Finally, Section \ref{sec:Conclusion} provides concluding remarks and discusses the potential future extensions for our new graph learning-based approach to MRTA.
Throughout this paper, \textbf{tasks} are also referred by the terms ``\textbf{nodes}'' or ``\textbf{vertices}''.
\section{MRTA: Problem definition and formulation
\label{sec:problem_formulation}
MRTA problems can be formulated as Integer Linear or Non-Linear Programming (ILP or INLP) problems. When tasks are defined in terms of location, the MRTA problem becomes analogical to the Multi-Traveling Salesmen Problem (mTSP)~\cite{khamis2015multi} and its generalized version, the Vehicle Route Planning (VRP) problem~\cite{dantzig1959truck}, albeit with additional constraints and operation-specific objective function. \vspace{-.2cm}
\subsection{MRTA problem description and formulation}
\label{sec:mrta_optimization}
The exact solution of the MRTA problem can be obtained by formulating it as the following INLP problem:
\vspace{-.2cm}
\begin{align}
\label{eq:objectiveFunction}
\min~ f_\text{cost} = \sum_{i = 1}^{N} r_{i}, \
\begin{cases}
r_{i} = \frac{t_{i}^{f}}{d_{i}},& \text{if } \ t_{i}^{f} > d_{i}\\
0, & \text{otherwise}
\end{cases}
\end{align}
\ \ \ \ \ \text{subject to~}
\begin{align}
\label{eq:constraint1}
& s_{i} \in S \; \forall \; i \in [1..N] \\
\label{eq:constraint2}
& s_{i} \neq s_{j} \; \forall \ i \neq j
\end{align}
\noindent Here $t_{i}^{f}$ is the time at which task $i$ was completed, $d_{i}$ is the time deadline of task $i$, and $S = [s_{1}, s_{2}, ... s_{N}]$ is the sequence of all the tasks/nodes that were visited, and $N$ is the number of tasks available. The minimum cost function that can be achieved using Eq. \ref{eq:objectiveFunction} is $0$, which corresponds to the case where all the tasks are successfully completed.
A detailed formulation of the exact ILP constraints that describe the MRTA problem can be found in this recent work on MRTA ~\cite{Mitiche2019}. Note that, in our paper, we use a slightly different objective/cost function as compared to \cite{Mitiche2019}, but the intention of the objective function of both of these works remains the same, which is to maximize the number of successfully completed tasks (thereby allowing fair comparison with methods reported in the literature).
Here, we craft the objective function (Eq.~\eqref{eq:objectiveFunction}) such that only missed tasks ($t_{i}^{f} > d_{i}$) contribute to the cost function. This objective function can be tailored according to user/application preferences, without affecting the proposed learning framework.
The constraints in Eqs. \ref{eq:constraint1} and \ref{eq:constraint2} are such that each task must be visited exactly once by any robot.
\begin{figure}[!t]
\centering
\includegraphics[width=.9\linewidth]{Figures/Sequential_decision_making.pdf}
\vspace{-0.4cm}
\caption{Deployment of an MRTA policy based on the CapAM architecture. a) Robot-1 at $t_0$. b) Robot-2 at $t_1$; here, the output for previously selected task (e.g., task 2 in (b)) is set at 0.}
\label{fig:sequential_decision_making}
\vspace{-0.6cm}
\end{figure}
\subsection{An ``MDP over Graph" Representation of MRTA}\label{sec:ProblemFormulation}
The task space of an MRTA problem can be represented as a complete un-directed graph $\mathcal{G} = (V, E, \textbf{A})$, which contains a set of nodes/vertices ($V$), a set of edges ($E$) that connect the vertices to each other, and the weighted adjacency matrix $\textbf{A}$ that gives the extent to which two nodes are connected. Each node is a task, and each edge connects a pair of nodes. For MRTA with $N$ tasks, the number of vertices and the number of edges are $N$ and $N(N-1)/2$, respectively. Node $i$ is assigned a 3-dimensional feature vector denoting the task location and time deadline, i.e., $\delta_i=[x_i,y_i,d_i]$. Here we consider a weighted adjacency matrix without self-loop ($\alpha_{ii} = 0, \alpha_{i,j} \in \textbf{A}$, where $ i \ {\text{and}} \ j \in [1, N]$).
The weights in the adjacency matrix $\alpha_{ij} = 1/(1+ |\delta_i - \delta_j|)$ take a real value between 0 and 1, in a manner such that if the features of two nodes $i$ and $j$ are relatively closer to each other, then $\alpha_{ij}$ is higher.
Since we formulate MRTA as an undirected graph, the adjacency matrix $A$ is symmetric. In this paper, the elements of $\textbf{A}$ ($\alpha_{i,j}$, $i \neq j$) are computed as the inverse of the normalized Euclidean distance of the node properties.
The MDP can be defined in a decentralized manner to capture the myopic task-assignment process of each individual robot, which can be expressed as a tuple $<\mathcal{S}, \mathcal{A}, \mathcal{P}_a, \mathcal{R}>$. Figure \ref{fig:sequential_decision_making} shows how this MDP is executed to perform the sequential task selection by each robot. As evident from Fig. \ref{fig:sequential_decision_making}, the components of the MDP can then be defined as follows: \textbf{State Space ($\mathcal{S}$):} A robot at any decision-making instance uses a state $s\in\mathcal{S}$, which contains the following information:
1) elapsed mission time, 2) robot's current location, 3) constraints of this robot such as its work capacity, 4) the planned tasks of its peers, and 5) the constraints of its peers. The state space also includes the environment, defined by the location and time deadline of tasks, which are expressed here as node features in the graph. Here we assume that each robot can broadcast its information, which includes the task it is currently engaged in and its work capacity, to its peers without the need for a centralized system for communication; this is well aligned with modern communication capabilities \cite{Sykora2020}.
\noindent\textbf{Action Space ($\mathcal{A}$):} The set of actions is represented as $\mathcal{A}$, where each action $a\in\mathcal{A}$ is defined as the index of the selected task, $\{1,\ldots,N\}$. In MRTA problems that include a depot, task $0$ can be considered as the depot. The task status, i.e., \textit{active}, \textit{completed}, and \textit{missed} (i.e., deadline is passed), is used to properly decode the actions generated by the policy model, with the latter being in the form of probability of selecting each available task. Note that even under full observability, as the action space of a single robot includes only its own degree of affinity for selecting each available task, the decentralized decision-making formulation differs from an ideal centralized formulation, and remains readily scalable. \textbf{Transition ($\mathcal{P}_a$):} The transition is an event-based trigger. An event is defined as the condition that a robot reaches its selected task or visits the depot. As environmental uncertainties and communication issues (thus partial observation) are not considered in this paper, only deterministic state transitions are allowed.
\textbf{Reward ($\mathcal{R}$):} A delayed reward is considered here, estimated at the end of the multi-robot simulation, i.e, when there are no more active tasks. The actual reward function is typically application-dependent. For example, in some MRTA problems, task completion is more important, while in another, decreasing the total effort (e.g., distance traveled) by robots is more important, which can then be estimated using the weight matrix.
\vspace{-.1cm}
\section{Learning framework}
\label{sec:Learning_framework}
\vspace{-.2cm}
In this work, we implement an RL algorithm on an \textbf{encoder-decoder} architecture to learn the optimal MRTA policies over the graph formulation described earlier. An illustration of this architecture is shown in Fig. \ref{fig:encoder_decoder}. The \textbf{encoder} represents each node as a continuous vector infusing both its own properties as well as its local and global structural properties within the graph representation of the task space. The decoder sequentially computes output probabilities for all the available tasks/nodes using the information from the encoder and the current state (\textbf{context}), by applying the MHA (as shown in left part of Fig. \ref{fig:encoder_decoder}). The sequentially computed output probabilities for each node indicate the relative goodness of selecting that node as the next one to be visited by the concerned robot. A detailed description of the architecture of the encoder and decoder is discussed in the following sections using the representative SR-ST type MRTA problem described in section \ref{sec:ProblemFormulation}.
\begin{figure}[!t]
\centering
\includegraphics[width=0.46\textwidth]{Figures/GCAPCN_NIPS.pdf}
\vspace{-0.3cm}
\caption{CapAM architecture including context, encoder \& decoder.}
\label{fig:encoder_decoder}
\vspace{-0.7cm}
\end{figure}
\subsection{Graph Capsule Convolutional Network based Encoder}
\label{Sec:Encoder}
The main purpose of the encoder is to represent useful information related to a node/task as a learnable continuous vector or tensor, which can then be used by the learning algorithm. For a graph node, the information includes the properties of the node itself (e.g., the coordinates and time deadline of the node), the local neighborhood information for the node, and also the global structural information. General solutions to CO problems such as CVRP, MTSP and MRTA should be permutation invariant, which means that the order by which each node is numbered should not affect the optimal solution. Hence the node encoding must also be permutation invariant. In this work, we are exploring how a Graph Capsule Convolutional Neural Network (GCAPCN) can be implemented for learning local and global structures with the node properties, with permutation invariant node embedding. GCAPCN is a class of Graph Neural Networks (GNN), introduced in \cite{Verma2018} to address the drawbacks (e.g., permutation invariance) of Graph Convolutional Neural Networks (GCN), and to enable the encoding of global information based on \textbf{capsule networks} presented in \cite{Hinton}. The advantage of GCAPCN lies in capturing more local and global information, compared to conventional aggregation operations used in GNN such as summation or standard convolution operations.
\subsubsection{Graph Capsule Convolutional Neural Networks}
\label{Sec:GCAPCN}
Let $X \in \mathbb{R}^{N \times |\delta_{i}|}$, be the node feature matrix, where $|\delta_{i}|$ is the input dimension for each node $i$. The standard graph Laplacian is defined as $L = D - A \in \mathbb{R}^{N \times N}$, where $D$ is the degree matrix and $A$ is the adjacency matrix of the graph. A capsule vector is computed using a Graph Capsule function based on different order of statistical moments, as shown in the equations later in this section.
We first compute a feature vector $F_{0i}$ for each node by linear transformation of the node properties $\delta_{i}$, as $ F_{0i} = \delta_{i}.W_{0} + b_{0}$ for all $i \in [1,N]$, where
$W_{0} \ \varepsilon \ \mathbb{R}^{h_{0} \times |\delta_{i}|}, \ \ b_{0} \ \varepsilon \ \mathbb{R}^{h_{0} \times 1}$, and $h_{0}$ is the length of the feature vector. For an MRTA problem with time constraints, $\delta_{i} = [x_{i}, y_{i}, d_{i}]$, where $x_i$, $y_i$, and $d_i$ are the $x$ coordinate, $y$ coordinate, and the time deadline of location/node $i$. Note that this method for encoding is not limited to a specific MRTA problem, and can also be extended or modified for MRTA problems with nodes having other features such as workload.
Each feature vector $F_{0i}, i \in [1,N]$ is then passed through a series of Graph capsule layers, where the output from the previous layers is used to compute a matrix $f_{p}^{(l)} (X, L)$ using a graph convolutional filter of polynomial form as given by:
\vspace{-.3cm}
\begin{equation}
\label{Eq:GraphCapsuleFunction}
f_{p}^{(l)} (X, L) = \sigma(\sum_{k=0}^{K}L^{k}(F_{(l-1)}(X,L)^{\circ p})W_{pk}^{(l)}) \vspace{-.1cm}
\end{equation} \vspace{-.2cm}
\noindent Here $L$ is the graph Laplacian, $p$ is the order of the statistical moment, $K$ is the degree of the convolutional filter, $F_{(l-1)}(X,L)$ is the output from layer $l -1$, $F_{(l-1)}(X,L)^{\circ p}$ represents $p$ times element-wise multiplication of $F_{(l-1)}(X,L)$.
Here, $F_{(l-1)}(X,L) \in \mathbb{R}^{N \times h_{l-1}p}$, $W_{pk}^{(l)} \in \mathbb{R}^{h_{l-1}p \times h_{l}}$. The variable $f_{p}^{(l)} (X, L) \in \mathbb{R}^{N \times h_{l}}$ is a matrix where each row is an intermediate feature vector for each node $i \in [1,N]$, infusing nodal information from $L_{e} \times K$ hop neighbors, for a value of $p$. The output of layer $l$ is obtained by concatenating all $f_{p}^{(l)} (X, L)$, as given by:
\vspace{-.2cm}
\begin{equation}
\label{Eq:LayerOutput}
F_{l}(X,L) = [f_{1}^{(l)} (X, L), f_{2}^{(l)} (X, L), ...f_{P}^{(l)} (X, L)]
\end{equation}
\noindent
Here $P$ is the highest order of statistical moment, and $h_{l}$ is the node embedding length of layer $l$. We consider all the values of $h_{l}$ (where $l \in [0, L_{e}]$) to be the same for this paper.
Equations \ref{Eq:GraphCapsuleFunction} and \ref{Eq:LayerOutput} were computed for $L_{e}$ layers, where each layer uses the output from the previous layer ($F_{l-1}(X,L)$). Adding more layers helps in learning the global structure, however, this can affect the performance by increasing the number of learnable parameters (compared to the size of the problem), leading to over-fitting.
The output from the final layer is then passed through a feed-forward layer so that the final feature vector has the right dimension ($h_l$) to be fed into the decoder as shown in Fig. \ref{fig:encoder_decoder}.
\vspace{-.2cm}
\subsection{Attention-based Decoder and Context}
\label{Sec:Decoder}
The main objective of the MHA-based decoder is to use the information from the encoder and the current state as context or query, and thereof choose the best task by calculating the probability value of getting selected for each (task) node. In this case, the first step is to feed the embedding for each node (from the encoder) as \textbf{key-values} ($\mathcal{K}$, $V$), since inputs for MHA are key-value pairs \cite{DBLP:journals/corr/VaswaniSPUJGKP17}. The key $K$ and value $V$ for each node is computed by two separate linear transformations of the node embedding obtained from the encoder. The next step is to compute a vector representing the current state, also known as the \textbf{context} (as shown in bottom left of Fig. \ref{fig:encoder_decoder}).
In our case studies, the context for the multi-head attention or MHA layer in the decoder consists of the following five features: \textbf{1)} elapsed mission time; \textbf{2)} work capacity of the robot taking decision; \textbf{3)} current location of the robot taking decision; \textbf{4)} current destination of robot's peers; and \textbf{5)} work capacity of peer; all concatenated to a single vector of length $h_{q}$, which then undergoes a linear transformation to get a vector of length $h_{l}$ also called the query $Q$. It is important to note that the \textbf{context} can be modified to account for other features such as remaining ferry-range and payloads of robots, thus promoting the flexibility to apply this architecture to a wider range of MRTA problems than those specifically studied in this paper.
Figure \ref{fig:encoder_decoder} illustrates the structure of the decoder.
Now the attention mechanism can be described as mapping the query ($Q$) to a set of key-value ($\mathcal{K},V$) pairs. The inputs, which are the query ($Q$) is a vector, while $\mathcal{K}$ and $V$ are matrices of size $h_{l} \times N$ (since there are $N$ nodes). The output is a weighted sum of the values $V$, with the weight vector computed using the compatibility function expressed as:
\vspace{-.2cm}
\begin{align}
\label{attention_function}
\text{Attention}(Q,\mathcal{K},V) = \text{softmax}(Q^{T}\mathcal{K}/\sqrt{h_{l}})V^{T}
\vspace{-.2cm}
\end{align}
Here $h_{l}$ is the dimension of the key of any node $i$ ($k_{i} \in \mathcal{K}$).
In this work, we implement a MHA layer in order to determine the compatibility of $Q$ with $\mathcal{K}$ and $V$. The MHA implemented in this work is similar to the decoder implemented in \cite{Kool2019} and \cite{DBLP:journals/corr/VaswaniSPUJGKP17}.
As shown in \cite{DBLP:journals/corr/VaswaniSPUJGKP17} the MHA layer can be defined as:
\vspace{-.2cm}
\begin{equation}
\label{mha}
\text{MHA}(Q,\mathcal{K},V) = \text{Linear}(\text{Concat}(\text{head}_{1}\dots \text{head}_{h_{e}}))
\vspace{-.2cm}
\end{equation}
Here $\text{head}_{i} = \text{Attention}(Q,\mathcal{K},V)$ and $h_{e}$ (taken as 8 here) is the number of heads.
The feed-forward layer is implemented to further process the mapping that results from the MHA layer, and transform it to a dimension that is coherent with the number of nodes in the task-graph ($N$). The interjecting batch normalization layers serve to bound values of a specific batch using the mean and variance of the batch.
The final \texttt{softmax} layer outputs the probability values for all the nodes. Here, the next task to be done is then chosen based on a greedy approach, which means that the node with the highest probability will be chosen. The nodes which are already visited will be masked (by setting their probability as $0$) so that these nodes are not available for selection in the future time steps of the simulation of the multi-robot operation.
\vspace{-.2cm}
\subsection{Training Algorithm}
\label{Sec:LearningAlgorithm} \vspace{-.1cm}
The training algorithm used here is REINFORCE.
Other relevant details for training are shown in Table \ref{table:training_algo_info_MRTA}. This training process can readily be further advanced in the future through the adoption of advanced state-of-the-art policy gradient algorithms such as Proximal Policy Optimization \cite{schulman2017proximal} and the Actor-Critic method \cite{Bahdanau2017}.
For each epoch of the training, two sets of data (MRTA operation scenarios) are used, which are the training set and the validation set. The training data set is used to train the model ($\theta_\text{CapAM}$) while the validation set is used to update the baseline model ($\theta_\text{CapAM}^{BL}$). Both the training data and validation data generated from the distribution of MRTA scenarios are explained later in section \ref{Sec:Baseline_method_dataset}.
The size of the training data and the validation data used in this paper is mentioned in Table~\ref{table:training_algo_info_MRTA}.
\vspace{-.1cm}
\section{Experimental evaluation}
\label{sec:Experimental_evaluation}
\vspace{-.2cm}
\subsection{Comparison with baselines}
\label{Sec:BaselineComparison}
\subsubsection{Details on dataset for training and testing}
\label{Sec:Baseline_method_dataset}
\begin{wraptable}[10]{R}{4.7cm}
\vspace{-.9cm}
\footnotesize
\caption{Settings for model training in CapAM and AM-RL}
\vspace{-0.7cm}
\label{table:training_algo_info_MRTA}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\footnotesize
\begin{tabular}{|l|c|}
\toprule
Details & Values \\
\midrule
\textit{Algorithm} & \textit{REINFORCE} \\
\textit{Baseline} & \textit{Rollout} \\
\textit{Epochs} & 100\\
\textit{Training samples} & 500,000 \\
\textit{Validation samples} & 10,000 \\
\textit{Optimizer} & \textit{Adam} \\
\textit{Learning step size} & 0.0001 \\
\textit{Training frequency} & 500 samples \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\vspace{-0.4cm}
\end{wraptable}
We consider the (NP-hard) MRTA problem with time and capacity constraints also known as Task Allocation Problem with Time and Capacity (TAPTC), as described in \cite{Mitiche2019}.
Each training sample has 100 tasks, and are located randomly within a 100 $\times$ 100 grid map. Each task $i$ has a time deadline $50 \leq d_{i} \leq 600$, and a workload $10 \leq w_{i} \leq 30$. Each sample has $n_r$ number of robots where $2 \leq n_r \leq 7$. The initial positions of the robots in a sample is also chosen randomly within the grids.
Each robot $j$ has a work capacity of $c_{j}$ where $1 \leq c_{j} \leq 3$. All the robots move at a speed of $1$ unit. A task $i$ is considered to be completed only if a robot $j$ visits node $i$ and spends a time of $w_{i}/c_{j}$. All the training samples are generated such that all the associated variables (mentioned above) follow a uniform distribution within their respective bounds.
The average training time per epoch for CapAM is $~\sim$10 mins.
In order to evaluate and compare the generalizability of CapAM-trained models with that of baselines,
we used the dataset in \cite{Mitiche2019} (found in \cite{dataset}). This dataset, which has a similar distribution of scenarios as used in training CapAM, consists of 96 test cases with 100 tasks and a varying number of robots ($2, 3, 5, 7$ robots), and the speed of every robot is considered to be the same (1 unit/s). The 96 cases are divided into ones with tight and slack deadlines. Each category can be further divided into 4 sub-categories based on the fraction of tasks ($25\%$, $50\%$, $75\%$, and $100\%$) that have normally distributed tasks deadlines.
For example, $25\%$ indicates that there are 25 tasks with deadlines normally distributed between the limits $d_{\texttt{low}}$ and $d_{\texttt{high}}$, while the remaining 75 tasks have a deadline of $d_{\texttt{high}}$.
For group 1 (tight deadline), the value of $d_{\texttt{high}}$ is half of that for group 2 (slack deadline). Further description of the test cases can be found in \cite{Mitiche2019}.
\subsubsection{Test for Generalizability}
\label{Sec:TestForGeneralizability}
We compare our method with the following four baseline methods over the stated test cases:
\textbf{i)} \textbf{Iterated Local Search (ILS)}: This is an online metaheuristic iterated search algorithm \cite{VANSTEENWEGEN20093281}, where the output of one iteration is partially used as the input to the next iteration. During each iteration, the best solution is improved by a perturbation step, followed by a local search.
\textbf{ii)} \textbf{Enhanced Iterated Local Search (EILS)}: EILS is also an online metaheuristic iterated search method \cite{Mitiche2019}, with an improved perturbation step as compared to \cite{VANSTEENWEGEN20093281}.
\textbf{iii)} \textbf{Bi-Graph MRTA (BiG-MRTA)}: BiG-MRTA~\cite{ghassemi2019decmrta} is an online method based on the construction and matching of a bipartite graph. In this method, a bipartite graph is constructed to connect robots to tasks, with the weights of connecting edges determined by an incentive model as a function of the tasks' features and robots' states. This decomposes the problem and yield a measure of robot-task pairing suitability. A maximum weighted matching problem is then solved by each robot to identify the optimal task assignments that maximize a net incentive for the team.
\textbf{iv)} \textbf{Attention Mechanism based RL (AM-RL)}: The AM-RL method \cite{Kool2019} consists of an encoder-decoder architecture purely based on attention mechanism. To implement this method for our problem, it is adapted to a multi-robot setting by making the following changes: 1) The node properties defined in Section~\ref{sec:ProblemFormulation} are used here; 2) The context for the MHA in the decoder is modified to be the same as that in CapAM; and 3) The cost function used for training is changed to that in Eq.~\eqref{eq:objectiveFunction}. We set the parameters which correspond to the best performing model for a CVRP with 100 locations. This AM-RL model is trained using the same sample distribution and settings as respectively described in Sections \ref{Sec:Baseline_method_dataset}. The average training time per epoch is found to be $\sim$11 mins for AM. The architectural difference between AM-RL and CapAM is mainly in the encoder, with their performance comparison intended to show the impact of better encoding for learning over graphs.
\begin{wraptable}[10]{r}{3.6cm}
\vspace{-.7cm}
\footnotesize
\caption{Comparison of computing time to generate entire solution sequence, averaged over 96 cases}
\vspace{-0.75cm}
\label{table:computation_time}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\footnotesize
\begin{tabular}{|l|c|}
\toprule
Method & Time (sec) \\
\midrule
\textit{ILS} & ~1.342 \\
\hline
\textit{EILS} & ~0.992 \\
\hline
\textit{BiG-MRTA} & ~0.301 \\
\hline
\textit{AM-RL} & ~0.086 \\
\hline
\textit{CapAM} & ~0.085 \\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vspace{-.1cm}
\end{wraptable}
The computation time to generate the entire sequence of task assignments, as required by the non-learning baselines and the learnt policies of CapAM and AM, are compared in Table \ref{table:computation_time} -- reported as an average over the 96 test cases. The real-time performance advantage of the learnt models, which are $>$10 times faster than the non-learning online methods (ILS/EILS/BiG-MRTA) is readily evident here.
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{Figures/gini_test_case_comparison.eps}
\vspace{-0.3cm}
\caption{Comparison of the performance (on test cases) of CapAM with other baseline methods. Here CapAM has $K$ = 2, $P$ = 3, and $L_{e}$ = 1.}
\label{fig:gini_test_cases}
\vspace{-0.7cm}
\end{figure}
Figure \ref{fig:gini_test_cases} shows the box plots for task completion performance over the 96 test cases divided into 4 categories (each with 24 cases) based on the number of robots (2,3,5, and 7). As seen from Fig. \ref{fig:gini_test_cases}, CapAM obtains a better performance compared to the baseline methods. To provide statistical evidence of the advantage of CapAM over the baseline methods, we performed a pairwise T-test of CapAM \textit{vs} each baseline on the 24 cases in each category, with results summarized in Table \ref{table:ttest_pvals}; here, the null hypothesis is that the difference between the values of the two sets has a mean equal to $0$. Considering a 5$\%$ significance level, Table \ref{table:ttest_pvals} shows that the \textbf{p-value} from the T-test is generally less than $0.05$, which indicates the rejection of the null hypothesis -- CapAM's performance on the test cases is thus significantly better than of the baselines.
\begin{table}[H]
\vspace{-.4cm}
\footnotesize
\centering
\caption{\textbf{p-value} for pairwise (CapAM vs. baselines) T-Test corresponding to Fig. \ref{fig:gini_test_cases}.} \vspace{-.3cm}
\label{table:ttest_pvals}
\begin{tabular}{|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{\# robots}} & \multicolumn{4}{c|}{\textbf{CapAM vs}} \\ \cline{2-5}
& \textbf{ILS} & \textbf{EILS} & \textbf{BiG-MRTA} & \textbf{AM-RL} \\ \hline
\textbf{2} & 2.03e-4 & 0.002 & 0.003 & 1.6e-11 \\ \hline
\textbf{3} & 0.0307 & 0.079 & 0.038 & 2.4e-8 \\ \hline
\textbf{5} & 0.024 & 0.163 & 8.8e-4 & 0.42 \\ \hline
\textbf{7} & 0.025 & 0.049 & 0.003 & 0.031 \\ \hline
\end{tabular}%
\vspace{-.4cm}
\end{table}
\subsection{Parametric and Scalability Analysis}
\label{Sec:Parametric-ScalabilityAnalysis}
\vspace{-.2cm}
Here, parametric analyses is performed w.r.t. the key parameters of $P$ and $L_{e}$ in the encoder of CapAM. Higher values of $P$ promotes greater information infusion for a given neighborhood; with increasing $L_{e}$ or $K$ (and fixed $P$), information from (great number of) $L_{e} \times K$ hop neighbours is considered in the node embeddings. Increasing $K$, $P$ and $L_{e}$ however increases the number of learnable parameters, which can lead to over-fitting and deteriorating performance for unseen cases.
Our numerical experiments with higher values of $K$ and $L_{e}$ corroborated this expectation, and hence we fixed $K=2$ and $L_{e}=1$.
Hence this section focuses on the impact of $P$, particularly in generalizing, and scaling to larger-sized problems beyond training -- both of which are premised as major advantages of learning local structural information.
To this end, we train 3 different models on the 100-task problem, with varying $P$ from 1 to 3, and fixed $K$ and $L_{e}$.
Then the CapAM (and for comparison AM-RL) models that are trained for 100 tasks and 2-7 robots are implemented on two sets of unseen problems -- one with the same size as in training, and another with numbers of tasks and robots scaled up by a factor of 2 (training/testing tasks are drawn from the same distribution).
Table \ref{table:CAM_P_var} shows the effect of $P$ on scalability of CapAM, which is also compared to AM-RL. Results are reported in terms of task completion $\%$ averaged over 24 samples in each category (based on \# tasks and robots). It can be seen that all CapAM models with any $P$ perform significantly better than AM-RL. For the problems of the same size as training (100-task cases), the task completion \% gap between CapAM and AM-RL is between $3\%$ and $13.5\%$. For the problems of double the size as training (200-task cases), the task completion \% gap between CapAM and AM-RL is between $1.9\%$ and $17.6\%$. In both cases, the gap (CapAM vs. AM-RL) is the highest when task-to-robot ratios are high.
These observations directly demonstrate the effectiveness of the novel encoder used in CapAM, which is designed to capture local structural information much better than in attention mechanisms.
On comparing the CapAM models with different $P$ (in Table \ref{table:CAM_P_var}), it can be seen that ones with $P = 1$ perform slightly poorer compared to corresponding ones with $P=2$ and $P=3$, with the difference between the latter two being marginal. Thus increasing $P$ beyond 2 may not provide further performance improvements.
\begin{table}[H]
\vspace{-.35cm}
\footnotesize
\caption{Performance of CapAM (in terms of \% of task completed) with $P=1,2,3$ for unseen problem sets with the same and twice the number of tasks and robots as in training.}
\vspace{-.3cm}
\label{table:CAM_P_var}
\resizebox{.48\textwidth}{!}{%
\footnotesize
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\textbf{\# tasks} &
\multicolumn{1}{l|}{\textbf{\# robots}} &
\multicolumn{1}{l|}{\textbf{CapAM($P$=3)}} &
\multicolumn{1}{l|}{\textbf{CapAM($P$=2)}} &
\multicolumn{1}{l|}{\textbf{CapAM($P$=1)}} &
\multicolumn{1}{l|}{\textbf{AM-RL}} \\ \hline
\multirow{4}{*}{\textbf{100}} & \textbf{2} & 34.6 & \textbf{36.2} & 32.6 & 22.7 \\ \cline{2-6}
& \textbf{3} & \textbf{51.1} & 50.3 & 49.4 & 38.0 \\ \cline{2-6}
& \textbf{5} & \textbf{74.9} & \textbf{74.9} & 73.2 & 70.2 \\ \cline{2-6}
& \textbf{7} & \textbf{94.8} & 94.1 & 91.2 & 85.6 \\ \hline
\multirow{4}{*}{\textbf{200}} & \textbf{4} & \textbf{28.2} & 27.1 & 24.4 & 10.7 \\ \cline{2-6}
& \textbf{6} & 46.7 & \textbf{47.0} & 45.2 & 29.4 \\ \cline{2-6}
& \textbf{10} & 67.8 & \textbf{68.6} & 66.5 & 61.4 \\ \cline{2-6}
& \textbf{14} & \textbf{77.8} & 77.5 & 76.2 & 74.3 \\ \hline
\end{tabular}%
}
\vspace{-.4cm}
\end{table}
\footnote{The codes for this paper can be found in \url{https://github.com/adamslab-ub/CapAM-MRTA}}.
\section{Conclusion}
\label{sec:Conclusion}\vspace{-.2cm}
In this paper, we proposed a new graph neural network architecture called CapAM, to learn policies for MRTA problems involving tasks with time deadlines and robots with work capacity. This new architecture incorporates an encoder based on capsule networks and a decoder based on the attention mechanism, along with a context module to feed the state of the robots and the environment. To learn the features of the encoder and decoder, the problem has been posed as an RL problem and solved using the policy gradient algorithm, REINFORCE. In addition, the proposed architecture is found to provide effective MRTA policies over varying task size and robot-team size. The new CapAM architecture demonstrated better performance compared to other state-of-the-art baselines, which include both a learning based method (AM-RL) and non-learning based methods (ILS, EILS, BiG-MRTA).
The CapAM architecture with its capsule based node embedding showed that learning local and global structural information of the task graph results in better generalizability over unseen test cases, as observed from comparing its performance with AM-RL (that uses a different node embedding). The computational cost analysis showed that the trained CapAM model takes only a few milliseconds to yield a task-assignment decision, making it highly suited for operations that require time-sensitive online decisions. The advantage of using local neighborhood information was also empirically evident from the scalability analysis on the defined MRTA problem, where CapAM demonstrated superior performance when applied to graphs with a larger number of tasks/nodes and robots in the team.
\textbf{Future directions:} Firstly, computational analysis of training feasibility when episodes must be simulated over more realistic virtual robotics environment remains a open direction of research across all learning-based approaches. Further, to enable transition of these MRTA methods to application, in the future we should consider dynamic tasks, environment uncertainties, and partially observable state spaces within the CapAM architecture.
\bibliographystyle{./bibliography/IEEEtran}
|
\section{Introduction}
\label{sec:intro}
The Standard Model (SM) has been experimentally verified with high accuracy and very successful. In the SM, the Higgs field plays a role in the mass generation of particles.
The potential for the Higgs field is given to break the electroweak (EW) symmetry, and particle masses are originated from the non-vanishing vacuum expectation value (VEV).
This mechanism has been tested by the EW precision observables, the 125-GeV Higgs measurements, and so on, and the predictions correspond reasonably well with the experimental results.
This success, however, poses a question about the origin of the Higgs potential. In particular, the mass squared of the Higgs field is negative and requires a severe fine-tuning to cancel the very large radiative corrections. In order to avoid the fine-tuning,
many models beyond the SM have been proposed so far; {\em e.g.}, supersymmetry, composite Higgs, little Higgs, top partner, extra dimensions, and gauge-Higgs unification.
These extended models generally predict additional particles around TeV scale, that can be tested in the experiments at the large hadron collider (LHC). Among the new particles, additional scalar fields
are often suggested as good candidates to validate the models.
The additional scalars, therefore, have been studied widely in both model-dependent and model-independent ways.
Interestingly, such a new particle, that interacts with the SM particles through the weak interaction and/or Yukawa interactions, can still take mass of $\mathcal{O}(100)\GeV$~\cite{ATLAS:2021moa,ATLAS:2022rme}, if the couplings with light quarks and leptons are suppressed.
The possibility of an additional scalar boson lighter than the 125\GeV~Higgs still remains.
Recently,
the CMS collaboration has reported a new excess in the di-tau final states within all the $\tau$ decay modes (leptonic and hadronic) by using the Run 2 full data \cite{CMS:2022rbd},
which is an extension of the previous searches \cite{CMS:2011lzj,ATLAS:2012ube,ATLAS:2014vhc,ATLAS:2016ivh,ATLAS:2017eiz,CMS:2018rmh}. The excess can be interpreted as a resonance of a new particle.
The local and global significance of the excess are $2.6\sigma $ and $2.3\sigma$ at the invariant mass of $m_{\tau\tau}=95\GeV$, respectively.\footnote{%
They have also reported a $2.8\sigma$ (local) and $2.4\sigma$ (global) excess at $m_{\tau\tau}=1.2\TeV$. It is, however, excluded by the same search by the ATLAS collaboration \cite{ATLAS:2020zms}.
} At $m_{\tau\tau}=100\GeV$, these values are
$3.1\sigma$ and $2.7\sigma$.
By introducing an additional single neutral narrow resonance $\phi$,
the best fit values for the excess are
\cite{CMS:2022rbd}
\begin{align}
\sigma(gg \to \phi)\times \mathrm{BR}(\phi \to \tau \bar\tau) = 7.7^{+3.9}_{-3.1}~\mathrm{pb}\quad \text{for~}m_{\phi} = 95\GeV\,,
\label{eq:tata95}
\end{align}
or
\begin{align}
\sigma(gg \to \phi)\times \mathrm{BR}(\phi \to \tau \bar\tau) = 5.8^{+2.4}_{-2.0}~\mathrm{pb}\quad \text{for~}m_{\phi} = 100\GeV\,.
\label{eq:tata100}
\end{align}
These cross sections are comparable to the $125\GeV$ SM Higgs boson, $h$ \cite{Anastasiou:2016cez},
\begin{align}
\sigma(gg \to h)\times \mathrm{BR}(h \to \tau \bar\tau) = 3.1\pm0.2~\mathrm{pb}.
\label{eq:SMtata}
\end{align}
Note that there is no excess in the $b$-tagging category,
which implies that the $b$-associated $\phi$ production is disfavored.
Interestingly,
in the same mass region,
a different excess has also been reported in the di-photon final states by the CMS collaboration
based on the full Run 1 \cite{CMS:2015ocq} and the first Run 2 (35.9\,fb$^{-1}$) data \cite{CMS:2018cyk}.
The local and global significance are $2.8\sigma$ and $1.3\sigma$
at $m_{\gamma \gamma} = 95.3\GeV$.
This excess can be interpreted as a new resonance, $\phi$, that decays to two photons, {\em i.e.}, $gg \to \phi\to\gamma \gamma$ (see Eq.~\eqref{eq:CMS_diphoton}).
Although the similar analysis has been performed by the ATLAS collaboration \cite{ATLAS:2018xad},
the sensitivity is not good enough to check the consistency \cite{Heinemeyer:2018wzl}.
Moreover, another mild excess has been reported
in the LEP experiment,
that can be interpreted as $e^+e^+\to Z \phi \to Z b\bar{b}$ \cite{LEPWorkingGroupforHiggsbosonsearches:2003ing}.
The signal corresponds to $2.3\sigma$ local significance at $m_\phi=98\GeV$.
Therefore, it is very interesting to consider the possibility that these excesses are caused by the same new particle $\phi$.
A variety of new physics interpretations have been discussed
in Refs. \cite{Cao:2019ofo,Aguilar-Saavedra:2020wrj,Hollik:2018yek,Liu:2018xsw,Domingo:2018uim,Biekotter:2017xmf,Haisch:2017gql,Fox:2017uwr,Biekotter:2019kde,Biekotter:2020cjs,Biekotter:2021qbc,Heinemeyer:2021msz,Biekotter:2022jyr,Biekotter:2022abc,Benbrik:2022azi,Benbrik:2022dja}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=24.5em]{figs/100gev_tata}
\includegraphics[width=12.25em]{figs/100gev_tata_v2}
\caption{
\label{Fig:dia}
Representative Feynman diagrams for a narrow resonance $\phi$ exchange in
the gluon fusion ($gg \to \tau \bar \tau$) and
the top-quark associated processes ($gg \to t \bar{t} + \tau \bar \tau$).
}
\end{center}
\end{figure}
In this paper,
we point out that associated production of $\phi$ with two top quarks
can provide a simple and powerful way to verify the di-tau excess in Eqs.~\eqref{eq:tata95} and \eqref{eq:tata100}.
In Fig.~\ref{Fig:dia},
the relevant diagrams for $gg \to \tau\bar\tau$
and $gg\to t\bar{t}+\tau\bar{\tau}$ are shown.
As shown Eq.~\eqref{eq:SMtata},
the di-tau excess requires
the new resonance cross section almost comparable to the SM Higgs boson via the gluon fusion.
The $\tau\bar{\tau}$ resonance at $95$--$100\GeV$, however,
suffers from the huge $Z$ boson background which stems from the tree-level Drell-Yan (DY) process.
Such a $Z$ peak pollution will be mild when the top-quark associated production is selected.
This is because that
the production cross section of $t\bar{t}+Z$ and $t\bar{t}+h$ are the same size in the SM (see Sec.~\ref{sec:CPeven} for the explicit numbers), and furthermore, the new resonance is produced by the tree level,
which could amplify the sensitivity to probe the new resonance.
We will investigate the LHC sensitivity to the possible $95$--$100\GeV$ resonance in the top-associated process.
In Sec.\,\ref{Sec:setup}, we introduce the minimal setup to account for the $\tau\bar{\tau}$ excess and summarize the predictions of the top-associated productions.
In Sec.\,\ref{Sec:numerics}, the numerical analysis to derive the upper limits on the model-independent $t\bar{t}+\tau\bar{\tau}$ cross section is performed for the new resonance.
Section\,\ref{sec:Summary} is devoted to summary and discussion.
\section{Setup}
\label{Sec:setup}
In this section,
we introduce a simplified model that can account for the excesses in the 95--$100\GeV$ region.
We introduce an additional scalar boson which has large couplings with top and $\tau$, since the gluon fusion is dominated by the chiral heavy fermion loop \cite{Kumar:2012ww} and the resonance preferentially decays into $\tau$ leptons.
There are several ways to derive such a scalar effectively at low energy.
In this study, we do not specify models and simply assume that
extra neutral scalars, $H$ and $A$, couple to top and $\tau$ as follows:
\begin{align}
-\mathcal{L}_{\rm eff} =
\frac{\rho_{tt}^H}{\sqrt{2}}\,\bar{t} H t +\frac{\rho_{\tau\tau}^H}{\sqrt{2}}\,\bar{\tau} H \tau
\pm i \frac{\rho_{tt}^A}{\sqrt{2}}\,\bar{t} A \gamma_5 t+i \frac{\rho_{\tau\tau}^A}{\sqrt{2}}\,\bar{\tau} A \gamma_5 \tau\,.
\label{Eq:Lint}
\end{align}
The all couplings, $\rho^\phi_{tt}$ and $\rho^\phi_{\tau \tau}$ ($\phi=H, \, A$),
are real in our study.
Then, $H$ $(A)$ corresponds to a CP-even (CP-odd) scalar.\footnote{Although additional vector particle,
that couples to the top and $\tau$ \cite{Abdullah:2019dpu}, could also be considered as the candidate,
the gluon fusion (via the top loop) vanishes for the vector boson production when the fermion interactions are vector couplings \cite{Campbell:2011cu}. We will discuss the possibility of the vector particle interpretation in Sec.~\ref{sec:Summary}.
}
As we will see,
the $\gamma_5$ structure plays an important role in our analysis.
Note that the relative sign for the $A$ interaction depends on the UV theory.
In our study, this relative sign does not affect the conclusion.
The possible UV completion of the light scalar boson
could be the Two-Higgs-Doublet Model (2HDM) with a real singlet scalar~\cite{Biekotter:2019kde,Biekotter:2020cjs,Biekotter:2021qbc,Heinemeyer:2021msz,Biekotter:2022jyr,Biekotter:2022abc} and Generic 2HDM (G2HDM, type-III 2HDM)~\cite{Branco:2011iw,Cline:2015lqp}, axion-like particle (ALP)~\cite{Georgi:1986df}, NMSSM~\cite{Ellwanger:2009dp} and so on.
If one takes a G2HDM as an illustrative model,
the oblique corrections and the LHC constraint on a charged scalar \cite{ATLAS:2018gfm,ATLAS:2021upq}
do not allow to set both $H$ and $A$ masses around $100\GeV$ for the $\tau\bar{\tau}$ excess.
In this paper we do not discuss the detailed setup and consider the two cases
that either $H$ or $A$ resides around 100 GeV.
\subsection{CP-even scalar scenario}
\label{sec:CPeven}
First, we consider the case that the resonance reported by the CMS collaboration is a CP-even scalar $H$.
The gluon fusion production cross section of $H$ at $\sqrt{s}=13\TeV$ is predicted as follows
\begin{align}
&\sigma(pp\to gg\to H)= 87.2 \,(\rho_{tt}^H)^2~\mathrm{pb}\quad {\rm{for}}~ m_H=95\GeV\,,
\label{Eq:ggH95}\\
&\sigma(pp\to gg\to H)= 79.5 \, (\rho_{tt}^H)^2~\mathrm{pb}\quad {\rm{for}}~ m_H=100\GeV\,,
\label{Eq:coef_ggF_H}
\end{align}
which are evaluated by \texttt{SusHi\,v1.7.0}\,\cite{Harlander:2012pb,Harlander:2016hcx} at next-to-next-to leading order (NNLO).\footnote{%
It is noted that the next-to-NNLO (N$^3$LO) correction increases the cross section up to $3\%$.}
The di-tau excess in Eqs.~\eqref{eq:tata95} and \eqref{eq:tata100} can be accommodated by
\begin{align}
&\rho_{tt}^H \sqrt{\mathrm{BR}(H \to \tau \bar\tau)} = 0.30\pm0.07
\quad {\rm{for}}~ m_H=95\GeV\,,\\
&\rho_{tt}^H \sqrt{\mathrm{BR}(H \to \tau \bar\tau)} = 0.27\pm0.05
\quad {\rm{for}}~ m_H=100\GeV\,.
\end{align}
We find that $\rho^H_{tt}$ is sizable to explain the excess: $\rho^H_{tt} \gtrsim 0.22$.
The $\rho_{tt}^H$ interaction also contributes to
the production of $\phi$ in association with two top quarks, as shown in the middle and right diagrams in Fig.~\ref{Fig:dia}.
The production cross section at $\sqrt{s}=13\TeV$
is obtained as
\begin{align}
&\sigma(pp\to t\bar{t}+ H)= 1.07\,(\rho_{tt}^H)^2~\mathrm{pb}\quad {\rm{for}}~ m_H=95\GeV\,,
\label{Eq:ttH95}\\
&\sigma(pp\to t\bar{t}+ H)= 0.94\,(\rho_{tt}^H)^2~\mathrm{pb}\quad {\rm{for}}~ m_H=100\GeV\,,
\label{Eq:prediction_ttH}
\end{align}
at next-to leading order (NLO).
Here, we evaluate the production cross section at the leading order by \texttt{MadGraph5\_aMC@NLO} \cite{Alwall:2014hca}
and multiply the NLO $K$ factor of 1.29 for simplicity \cite{Frixione:2014qaa}.
Combining the above results, the explanation of the excess predicts
a sizable cross section of $pp \to t\bar{t} H \to t\bar{t} + \tau\bar{\tau}$:
\begin{align}
&\sigma(pp\to t\bar{t}+ H)\times \mathrm{BR}(H \to \tau \bar\tau) =[0.056,0.094,0.14]~\mathrm{pb} \quad{\rm{for}}~ m_H=95\GeV\,,
\label{Eq:prediction_H95}
\\
&\sigma(pp\to t\bar{t}+ H)\times \mathrm{BR}(H \to \tau \bar\tau) =[0.045,0.069,0.097]~\mathrm{pb}\quad {\rm{for}}~ m_H=100\GeV\,,
\label{Eq:prediction_H}
\end{align}
where numbers in parentheses indicate the $1\sigma$ range with its central value.
The top-associated production cross section of the SM Higgs boson
has been measured with 79.8 fb$^{-1}$ of the Run 2 data \cite{ATLAS:2018mme}:
$\sigma (pp\to t\bar{t} + h)=0.67\pm0.14\,\mathrm{pb}$.
This is consistent with the SM prediction, $\sigma (pp\to t\bar{t}+h)_{\rm SM}=0.51\,\mathrm{pb}$.
Combining the subsequent decay branching ratio of $\mathrm{BR}(h\to\tau\bar\tau)_{\rm SM}\sim6\%$,
$\sigma(pp\to t\bar{t} + h)_{\rm SM} \times\mathrm{BR}({h\to\tau\bar\tau})_{\rm SM} \simeq 0.03\,\mathrm{pb}$
is derived in the SM, and thus the predicted cross section of $H$ is larger than the SM Higgs by a factor of approximately three.
We comment on
the top-associated $Z$ production cross section.
It has been measured based on the full Run 2 data \cite{ATLAS:2021fzm};
$\sigma (pp\to t\bar{t}+Z)=0.99\pm0.09\,\mathrm{pb}$.\footnote{%
The cross section of $\sigma (pp\to t\bar{t}+Z)$ has been measured in $Z\to \ell\bar{\ell}$ with $\ell=e$ and $\mu$ channels.
}
This result is also consistent with the SM prediction, $\sigma (pp\to t\bar{t}+Z)_{\rm SM} = 0.84^{+0.09}_{-0.10}\,\mathrm{pb}$
at NLO QCD and EW accuracy \cite{Frixione:2015zaa}.
Since $\mathrm{BR}(Z\to \ell\bar{\ell})$ is about $3\%$,
we obtain $\sigma (pp\to t\bar{t}+Z)\times \mathrm{BR}(Z\to \tau\bar\tau) \simeq 0.03\,\mathrm{pb}$,
that is comparable to the top-associated production of $h$.
\subsection{CP-odd scalar scenario}
Next, we consider the CP-odd scalar $A$ interpretation.
The analysis in Sec.~\ref{sec:CPeven} is applied to this case, replacing $H$ with $A$. The sizes of the cross sections are, however, different because of the couplings in Eq.~(\ref{Eq:Lint}), so that the predictions are totally different.
The gluon fusion cross section of $A$ at $\sqrt{s}=13\TeV$ is
\begin{align}
&\sigma(pp\to gg\to A)= 201.7\,(\rho_{tt}^A)^2\,\mathrm{pb}\quad {\rm{for}}~ m_A=95\GeV\,,\\
&\sigma(pp\to gg\to A)= 184.4\,(\rho_{tt}^A)^2\,\mathrm{pb}\quad{\rm{for}}~ m_A=100\GeV\,,
\label{Eq:ggF_A}
\end{align}
at NNLO \cite{Harlander:2012pb,Harlander:2016hcx}.
Assuming $\rho^H_{tt}=\rho^A_{tt}$, these predictions are twice as large as them in the $H$ case
due to the different $\gamma_5$ structure in the top-quark loop.
The di-tau excess in Eqs.~\eqref{eq:tata95} and \eqref{eq:tata100} can be accommodated by
\begin{align}
&\rho_{tt}^A \sqrt{\mathrm{BR}(A \to \tau \bar\tau)} = 0.20\pm0.04
\quad {\rm{for}}~ m_A=95\GeV\,,\\
&\rho_{tt}^A \sqrt{\mathrm{BR}(A \to \tau \bar\tau)} = 0.18\pm0.03
\quad {\rm{for}}~ m_A=100\GeV\,.
\end{align}
The production cross section of the top-associated production is
\begin{align}
&\sigma(pp\to t \bar{t}+ A)=0.30\,(\rho_{tt}^A)^2\,\mathrm{pb}\quad{\rm{for}}~ m_A=95\GeV\,,\\
&\sigma(pp\to t\bar{t}+ A)=0.29\,(\rho_{tt}^A)^2\,\mathrm{pb}\quad{\rm{for}}~ m_A=100\GeV\,,
\end{align}
where the same NLO $K$ factor as in the CP-even scalar production is assumed \cite{Frederix:2011zi}.
In contrast to the gluon fusion cross section,
the results are three times as small as them in the $H$ case when $\rho^H_{tt}=\rho^A_{tt}$, as shown in Eqs.~\eqref{Eq:ttH95} and \eqref{Eq:prediction_ttH}.
This is again originated from the $\gamma_5$ structure. There is a destructive interference between the contribution of the middle diagram and that of the right diagram in Fig.~\ref{Fig:dia}
\cite{Djouadi:2005gj,Frederix:2011zi,Dolan:2016qvg}.
As a result, the cross section of
$pp\to t\bar{t}+A \to t\bar{t} + \tau\bar{\tau}$, that is consistent with the di-tau excess, is predicted to be smaller than that of the $H$ case:
\begin{align}
&\sigma(pp\to t\bar{t}+ A) \times \mathrm{BR}(A \to \tau \bar\tau)=[0.007,0.011,0.017]\,\mathrm{pb} \quad{\rm{for}}~ m_A=95\GeV\,,\label{Eq:ttA95}\\
&\sigma(pp\to t\bar{t}+ A)\times \mathrm{BR}(A \to \tau \bar\tau)=[0.005,0.009,0.013]\,\mathrm{pb} \quad{\rm{for}}~ m_A=100\GeV\,.
\label{Eq:prediction_A}
\end{align}
It is also known that the angular correlation and transverse momentum distributions of $pp\to t\bar{t}+A$ are different from $pp\to t\bar{t}+H$ due to the presence of $\gamma_5$ in Eq. (\ref{Eq:Lint}) \cite{Dolan:2016qvg}.
\section{Comparisons to the ATLAS data}
\label{Sec:numerics}
In this section,
we compare the CMS di-tau excesses in Eqs.~\eqref{eq:tata95} and \eqref{eq:tata100} to the ATLAS results that involve $\tau\bar{\tau}$ in the final states.
Since the range of the ATLAS data set for the exotic particle search decaying to $\tau\bar{\tau}$ is $m_{\tau\tau}\geq 200\GeV$ \cite{ATLAS:2020zms}, the $95$--$100\GeV$ region has not been covered.
Instead, we utilize the ATLAS Run 2 full data for the $h \to \tau\bar\tau$ decay channel where $h$ is the SM Higgs boson \cite{ATLAS:2022yrq}.
The $h\to\tau\bar{\tau}$ and $Z\to\tau\bar{\tau}$ events have been carefully studied by using the several production processes.
Therefore the additional narrow resonance $\phi$ that decays to $\tau\bar{\tau}$ can also be probed.
In the following, we analyze the data in the {\em boosted $\tau_{\rm h} \tau_{\rm h}$} and {\em $tt(0\ell) + \tau_{\rm h} \tau_{\rm h}$ categories}~\cite{ATLAS:2022yrq}, where $\tau_{\rm h}$ denotes $\tau$ that decays hadronically.
\subsection{Boosted \texorpdfstring{$\tau\bar{\tau}$}{tau tau} search}
\label{sec:boostedtautau}
The boost categories defined in Ref. \cite{ATLAS:2022yrq} consist of the events that fail to meet the criteria of the vector-boson fusion (VBF), vector-boson associated production (VH), and a pair of top-quark associated production. The events have high-$p_{\rm T}$ (boosted) Higgs candidates.
The more than $70 \%$ events, actually, come from the gluon fusion with large Higgs boson transverse momentum in the analysis of Ref.~\cite{ATLAS:2022yrq}.
Hence, the ATLAS data in this category should be a good comparison to the CMS di-tau excess.
We use the result in the {\em boost\_2~category}~corresponding to $200 < p_{\rm T}^{\tau\tau} < 300\GeV$ in Ref.~\cite{ATLAS:2022yrq}.\footnote{%
We found that the data in {\em boost\_3~category}~corresponding to $ p_{\rm T}^{\tau\tau} > 300\GeV$ is less sensitive in the following analysis due to the small amount of statistics.}
Using the the narrow width approximation,
we define a signal strength for the gluon fusion of $\phi$,
\begin{align}
\mu_\phi (\tau\bar\tau) \equiv
\frac{\sigma(pp\to gg \to \phi)\times \mathrm{BR}(\phi\to \tau\bar\tau)}{\sigma(pp\to gg \to h)_{\rm{SM}}\times \mathrm{BR}(h\to \tau\bar\tau)_{\rm{SM}}}\,,
\end{align}
where the denominator is values of the $125\GeV$ SM Higgs.
In the left panel of Fig.~\ref{Fig:95_mu2}, an expected histogram of the $\phi$ with $m_{\phi}=95\GeV$ and $\mu_{\phi}(\tau\bar{\tau})=2$ is shown by the blue shaded region.
Here the SM Background (Bkg), except for the SM Higgs boson, is subtracted from the data which corresponds to
the bottom-left panel of Fig.~20 ({\em boost\_2~category}) of Ref.~\cite{ATLAS:2022yrq}.
Note that the SM Higgs histogram (red shaded region) stands for $pp \to h \to \tau \bar{\tau}$ and is scaled by $1/0.93$ from the original figure.\footnote{%
We also subtract the non gluon-fusion contributions according as Table~11 of Ref.~\cite{ATLAS:2022yrq} to construct the histogram of $\phi$.
}
The $0.93\,(^{+0.13}_{-0.12})$ is a global fit result of the signal strength of $p p \to h \to \tau \bar{\tau}$ \cite{ATLAS:2022yrq}.
The dashed band represents the total uncertainty of the SM Bkg.
It is shown that there is a huge uncertainty around $m_{\tau\tau}=80$--$100\GeV$, which comes from the DY (boosted) $Z$-boson production.
\begin{figure}[t]
\begin{center}
\includegraphics[width=17em]{figs/boosttata_95_mu2}
\quad
\includegraphics[width=17em]{figs/tttata_95_mu2}
\caption{
\label{Fig:95_mu2}
The histograms of the additional resonance $\phi$ (blue)
and SM Higgs (red) event shapes are shown with the experimental data (cross) and their uncertainties (dashed band).
The left panel is for the boosted $\tau_{h}\tau_{h}$ and the right panel is
the $tt(0\ell) + \tau_{\rm h}\tau_{\rm h}$ categories, respectively.
The signal normalization is assumed to be $\mu_\phi=2$, and $m_\phi=95\,$GeV are set for the mass.
See the text for details.
}
\end{center}
\end{figure}
By using a $\chi^2$ test which will be explicitly defined in the next section,
we set $95\%$ confidence level (CL) upper limits on the signal strength,
\begin{align}
& \mu_\phi (\tau\bar\tau) < 3.78 \quad \text{for~}m_{\phi} = 95\GeV\,,
\\
& \mu_\phi (\tau\bar\tau) < 3.61 \quad \text{for~}m_{\phi} = 100\GeV\,,
\end{align}
or equivalently,
the upper limits on the production cross sections,
\begin{align}
&\sigma(gg \to \phi)\times \mathrm{BR}(\phi \to \tau\bar\tau) < 11.5
\,\mathrm{pb}\quad \text{for~}m_{\phi} = 95\GeV\,,
\\
&\sigma(gg \to \phi)\times \mathrm{BR}(\phi \to \tau\bar\tau) < 11.0
\,\mathrm{pb}\quad \text{for~}m_{\phi} = 100\GeV\,.
\end{align}
We find that the ATLAS data in the boosted $\tau_{\rm h}{\tau}_{\rm h}$ category
are consistent with the CMS di-tau excess in Eqs.~\eqref{eq:tata95} and \eqref{eq:tata100} even if the upper edge of the 1$\sigma$ is considered.
\subsection{\texorpdfstring{$t\bar{t}+\tau\bar{\tau}$}{t t + tau tau} search}
\label{sec:tttautau}
In this section, we compare the CMS di-tau excess with the ATLAS data in the $tt(0\ell) + \tau_{\rm h} \tau_{\rm h}$ category \cite{ATLAS:2022yrq}.
Due to the limited statistics,
the ATLAS collaboration has not performed a serious top reconstruction,
but just imposed either six jets including at least one $b$-tagged jet or five jets including at least two $b$-tagged jets
for the event selection.
As mentioned in Sec.~\ref{sec:intro},
the additional $t\bar{t}$ requirement is expected to improve the sensitivity to probe the new resonance, because the huge DY $Z$-boson Bkg can be vetoed and the production cross sections of $t\bar{t} + \{h,Z,\phi\} \to t\bar{t} + \tau\bar{\tau}$ are of the same size.
However, since the top-quark tagging has not been seriously imposed in this ATLAS data,
the SM Bkg is still dominated by the DY process ($pp \to Z+5\text{--}6\,$QCD~jets),
and still large Bkg remains even in the signal region ($m_{\tau\tau}\simeq 125\GeV$),
as one can see in Fig.~11 of Ref.~\cite{ATLAS:2022yrq}.
The region of interest in this paper is
$m_{\tau\tau}=95$--$100\GeV$,
and hence we have the large Bkg from $Z\to \tau\bar{\tau}$.
Nonetheless,
we use this result to derive the current experimental limit on the new resonance production.
It is naively expected that a severe top-quark tagging algorithm (by the mass reconstruction)
improves the sensitivity.
Since experimental analyses have used the boosted decision tree (BDT) techniques,
it is difficult to access the detailed information about the final kinematic cuts.
Instead, we utilize the data in {\em ttH\_1~category}, corresponding to the right panel of Fig.~11 of Ref.~\cite{ATLAS:2022yrq}, where the $t\bar{t}h$ events are optimised to be enhanced over $Z$ and $t\bar{t}$ Bkg events by the BDT.
In the plot,
$(\text{data~yields)}-(\text{SM~Bkg~except~for~the~SM~Higgs})$ is shown ($N_{\rm ob} \pm \Delta N_{\rm ob}$), as well as
uncertainty of the total SM Bkg ($\pm \Delta N_{\rm Bkg}$) and the SM Higgs histogram ($N_{h}$).
From this figure, we estimate the sensitivity to probe the new resonance,
supposing that the $m_{\tau\tau}$ distribution of $t\bar{t} +\phi \to t\bar{t} + \tau\bar{\tau}$ under the BDT is similar to the SM Higgs.
Similar to the previous section,
we define a signal strength for the top-associated production of $\phi$,
\begin{align}
\mu_\phi (t\bar{t}+\tau\bar\tau) \equiv
\frac{\sigma(pp\to t\bar{t}+\phi)\times \mathrm{BR}(\phi\to \tau\bar\tau)}{\sigma(pp\to t\bar{t}+h)_{\rm{SM}}\times \mathrm{BR}(h\to \tau\bar\tau)_{\rm{SM}}}\,.
\end{align}
In the right panel of Fig.~\ref{Fig:95_mu2},
an expected histogram of $\phi$ with $m_{\phi}=95\GeV$ and $\mu_{\phi}(t\bar{t}+\tau\bar\tau) =2$
is shown by the blue shaded region.
Since the each bin width in the data is $10\GeV$,
we simply shift the distribution of the SM Higgs histogram
by three bins to obtain the distribution of $m_\phi=95\GeV$.\footnote{%
See also Sec.~\ref{sec:Summary} for a discussion of this treatment.}
Since the width of the SM Higgs histogram stems from the experimental resolution,
it is expected that the width of the histogram of $\phi$ is roughly of the same size as the SM Higgs.
Moreover, if the resolution is proportional to the value of $m_{\tau \tau}$, the histogram of $\phi$ becomes sharpened.
Therefore, we just rescale the SM Higgs histogram to conservatively predict the $\phi$ contribution.
We represent the histogram of $\phi$ by $N_\phi(\mu_\phi)$.
Again, the SM Higgs histogram (red shaded region)
is scaled by $1/0.93$ from the original one.
We perform the following $\chi^2$ test,
\begin{align}
\chi^2 (\mu_\phi) = \text{max} \left[\chi^2_i(\mu_\phi) \right] \quad \text{and} \quad
\chi^2_i(\mu_\phi)=\sum_{j=i}^{i+2} \frac{\left[N_{\rm ob}^j -N_{h}^{j} - N_{\phi}(\mu_\phi)^j \right]^2}{(\Delta N_{\rm ob}^{j})^2+(\Delta N_{\rm Bkg}^j)^2}
\,,
\label{eq:chisq}
\end{align}
where $i$ and $j$ are indices of each bin.
Note that since correlations between the data in each bin are not available in Ref.~\cite{ATLAS:2022yrq}, we discard them for simplicity.
Due to the finite experimental resolution for the $m_{\tau\tau}$ distribution,
judging based on the single bin data is too aggressive.
Here, we use at least $3$ contiguous bins for each $\chi^2_i(\mu_\phi)$ evaluation.
The criterion for setting the upper limit on $\mu_\phi$
is $\chi^2(\mu_\phi) < \chi^2_{3\text{dof}, 95\%} \simeq 7.82$.
As a validation of this $\chi^2$ test,
we compare the upper limit on the SM Higgs production cross section in the ttH\_1~category.
We obtain the 95\% CL upper limit for the Higgs boson cross section,
\begin{align}
\mu_h = \frac{\sigma(pp\to t\bar{t}+h) \times \mathrm{BR} (h \to \tau\bar{\tau})}
{\sigma(pp\to t\bar{t}+h)_{\rm SM} \times \mathrm{BR}( h \to \tau\bar{\tau})_{\rm SM}}\le 2.65\,,
\end{align}
while $\mu_h \lesssim 2.96$ ($\mu_h=1.02^{+0.97}_{-0.81}$) has been set in Ref.~\cite{ATLAS:2022yrq}.
It is found that the $\chi^2$ test in Eq.~\eqref{eq:chisq} gives a slightly severe limit.
This could be attributed to the following reasons;
This $\chi^2$ test does not include the theoretical uncertainty of the signal events properly.
The total uncertainty (dashed band $\Delta N_{\rm Bkg}$) includes the theoretical uncertainty of $\mu_h \simeq 1$ and it should be inflated according as $\mu_h$.
Second, when we evaluate the $\chi^2_i$ value for at least not $3$ but $4$ contiguous bins, the resultant limit is weakened by $5$--$10\%$.
Therefore, we decide to weaken the obtained upper limit on $\mu_{\phi}$
by
$10\%$
to be conservative test.
In this prescription, we obtain $\mu_h < 2.92$.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.5}
\newcommand\TIMES{\!\times\!}
\newcommand{\bhline}[1]{\noalign{\hrule height #1}}
\rowcolors{2}{gray!15}{white}
\begin{tabular}{lccc}
\bhline{1pt}
&$m_{\phi}=95\GeV$&$m_{\phi}=100\GeV$&$m_{\phi}=105\GeV$\\
\hline
$\mu_{\phi}(\tau \bar{\tau})$ &$<3.78$&$<3.61$& $<2.00$\\
$\sigma(gg\to\phi)\TIMES\mathrm{BR}(\phi\to\tau\bar\tau)$ &$<11.5\,\mathrm{pb}$&$<11.0\,\mathrm{pb}$& $<6.08\,\mathrm{pb}$\\
$\mu_{\phi}(t\bar{t}+\tau\bar{\tau})$ &$<1.56$&--& $<1.10$\\
$\sigma(pp\to t\bar{t} +\phi)\TIMES\mathrm{BR}(\phi\to\tau\bar\tau)$ &$<0.050\,\mathrm{pb}$&--& $<0.035\,\mathrm{pb}$\\
\bhline{1pt}
\end{tabular}
\caption{\label{Tab:result_muphi}
The $95\%$ CL upper limits on the signal strengths ($\mu_\phi$) and the production cross sections for $m_{\phi}=95,\,100,\,105\GeV$ are summarized.
These limits are obtained from the ATLAS Run 2 full data \cite{ATLAS:2022yrq} in the boosted $\tau\bar{\tau}$ (boost\_2~category) and $t\bar{t} +\tau\bar{\tau}$ ($tt(0\ell) + \tau_{\rm h} \tau_{\rm h}$ category) searches, via the $\chi^2$ test defined in Eq.~\eqref{eq:chisq}. }
\end{table}
Using the $\chi^2$ test, we set the $95\%$ CL upper limit on the signal strength,
\begin{align}
\mu_\phi (t\bar{t}+\tau\bar{\tau}) < 1.56
\quad \text{for~}m_{\phi} = 95\GeV\,,
\label{eq:tttata95}
\end{align}
or equivalently,
\begin{align}
\sigma(pp \to t\bar{t} + \phi)\times \mathrm{BR}(\phi \to \tau\bar\tau) < 0.050
\,\mathrm{pb}\quad \text{for~}m_{\phi} = 95\GeV\,.
\end{align}
It is found that the CP-even scalar $H$ interpretation on the CMS di-tau is excluded
by the ATLAS $t\bar{t}+\tau\bar{\tau}$ search, while the CP-odd $A$ one is consistent with the ATLAS
(see Eqs.~\eqref{Eq:prediction_H95} and \eqref{Eq:ttA95}).
Note that
due to the fixed bin width,
we can not perform the same analysis for the $m_\phi=100\GeV$ case.\footnote{%
We performed the same $\chi^2$ test for $m_\phi = 105\GeV$, and the result is $\mu_{\phi}(t\bar{t} +\tau\bar{\tau}) < 1.10$, which is not far from Eq.~\eqref{eq:tttata95}.
This result implies that the upper limit for $m_\phi = 100\GeV$ would be the same size as the $m_\phi = 95\GeV$ case.}
Our results are summarized in Table~\ref{Tab:result_muphi}.
Finally, we project the obtained ATLAS limits
onto the the minimal setup in Eq.~\eqref{Eq:Lint}.
In Fig.~\ref{Fig:coupling},
we show the ATLAS limits from the data in the boosted $\tau\bar\tau$ and $t\bar{t}+\tau\bar\tau$ searches
by the gray shaded regions surrounded by the solid and dashed lines, respectively.
The CP-even (odd) scalar with $m_{H(A)}=95\GeV$ is considered in the left (right) panel.
The CMS di-tau excess in Eq.~\eqref{eq:tata95} can be explained in the yellow region.
For the branching ratio,
we calculate $H(A) \to \tau \bar{\tau}$ at the tree level and
$gg,\,\gamma\gamma$ and $Z\gamma$ at the one-loop level \cite{Gunion:1989we} for the decay channels.
Furthermore, the aforementioned CMS di-photon excess can be explained in
the blue region corresponding to \cite{CMS:2018cyk,Biekotter:2019kde}
\begin{align}
\sigma (gg \to \phi)\times \mathrm{BR}(\phi \to \gamma \gamma) = 0.058\pm0.019\,\mathrm{pb}\,,
\label{eq:CMS_diphoton}
\end{align}
where the \texttt{SusHi} and Ref.~\cite{Denner:2011mq} are used for the $95\GeV$ SM-like Higgs value.
It is clearly shown that an interesting parameter region that can explain both di-tau and di-photon excesses is excluded in the CP-even scalar case.
\begin{figure}[t]
\begin{center}
\includegraphics[width=17.3em]{figs/rhoH_result.pdf}
\quad
\includegraphics[width=17.3em]{figs/rhoA_result.pdf}
\caption{
\label{Fig:coupling}
The obtained ATLAS $95\%$ CL limits are shown on a $\rho_{\tau\tau}$--$\rho_{tt}$ plane (see Table~\ref{Tab:result_muphi}).
The gray shaded regions surrounded by
the solid and dashed lines are excluded
by the ATLAS boosted $\tau\bar{\tau}$ and $t\bar{t}+\tau\bar{\tau}$ searches, respectively.
The left panel is for the CP-even ($H$) case and the right one is for the CP-odd ($A$).
The CMS di-tau and di-photon excesses can be explained in the yellow and blue regions, respectively. The scalar boson mass is set to be $95\GeV$.
}
\end{center}
\end{figure}
\subsection{Comment on \texorpdfstring{$t\bar{t}+\gamma\gamma$}{t t + gamma gamma} search}
\label{Sec:ttgamgam}
Motivated by the CMS di-photon excess in Eq.~\eqref{eq:CMS_diphoton},
we point out that the above procedure can be repeated
to estimate the upper limit on $\sigma(pp \to t\bar{t}+\phi)\times \mathrm{BR}(\phi \to \gamma\gamma)$
by analysing $pp \to t\bar{t} + h \to t\bar{t} + \gamma \gamma$ process.
This process must be cleaner than the $pp\to t\bar{t} + \{h,Z\} \to t\bar{t}+\tau\bar{\tau}$.
This is because the $Z \to \gamma\gamma$ decay is forbidden (the Landau--Yang theorem \cite{Landau:1948kw,Yang:1950rg})
so much smaller SM Bkg is expected around $m_{\gamma\gamma} \sim 90\GeV$.\footnote{%
Still, there is a doubly photon-misidentified Bkg from $Z \to e^+e^-$ \cite{CMS:2018cyk}.}
Furthermore, $\sigma(pp \to t\bar{t}+\phi)\times \mathrm{BR}(\phi \to \gamma\gamma)$ is expected to be the same size as the
$\sigma(pp \to t\bar{t}+h)_{\rm SM}\times \mathrm{BR}(h \to \gamma\gamma)_{\rm SM}$.
Therefore, the search for the additional resonance in $pp \to t\bar{t} + \gamma\gamma $ is promising.
Indeed, such a check has been implicitly done in the CMS collaboration \cite{CMS:2018cyk}.
However,
we could not find any results of direct searches for $pp \to t \bar{t} + \gamma \gamma$ in a region of $m_{\gamma\gamma}\le105\GeV$ \cite{ATLAS:2018mme,ATLAS:2020ior,CMS:2021kom} (see, {\em e.g.}, Fig.~2 of Ref.~\cite{ATLAS:2020ior}).
Therefore, we would like to suggest
an experimental analysis of the $pp \to t \bar{t} + \gamma \gamma$ process with the low-mass region
to probe the additional resonance.
\section{Summary and discussion}
\label{sec:Summary}
Due to the nature of the high-energy proton collider
and its harsh hadron activity,
it is difficult to probe weakly interacting colorless new particles below $\mathcal{O}(100)\GeV$ mass region.
Very recently, the
CMS collaboration has reported a di-tau excess with a local significance of 2.6--3.1$\sigma$ around $m_{\tau\tau}=95$--$100\GeV$.
This excess can be interpreted as an additional scalar boson $\phi$ produced via the gluon-fusion process.
Interestingly,
in the same mass region, two other excesses have been reported;
a di-photon excess by the CMS collaboration and
an excess in $e^+e^- \to Z b\bar{b}$ from the LEP experiment.
In this paper, we focus on the CMS di-tau excess.
This excess can be explained by the light scalar $\phi$ which couples with the top and $\tau$ in the minimal setup.
First, we found that the minimal scalar model is still consistent with the ATLAS result for the similar boosted $\tau\bar{\tau}$ search, where a huge SM Bkg comes from the $Z$-boson DY production.
Second, we point out that the minimal scalar model predicts the inevitable correlation with the top-quark associated process ($gg \to t\bar{t} +\tau\bar\tau$), see the Feynman diagrams in Fig.~\ref{Fig:dia}.
Moreover, a requirement of an additional top-quark pair in the final state suppresses the huge $Z$-boson DY Bkg, so that the experimental sensitivity to probe $\phi$ is certainly better compared to the boosted $\tau\bar{\tau}$ search around $m_{\tau\tau}=95$--$100\GeV$, although the statistics are limited.
Based on the ATLAS data available in Ref.~\cite{ATLAS:2022yrq},
we obtain the $95\%$ CL upper limits on the gluon-fusion and top-associated production cross sections in the minimal setup, which are summarized in Table~\ref{Tab:result_muphi}.
One of important points is that the limits depend on the CP eigenstate of $\phi$, {\em i.e.}, whether $\phi$ is CP-even ($H$) or CP-odd ($A$).
We point out that
the gluon-fusion production cross section is twice larger in the $A$ case, while the top-associated production cross section is three times larger in the $H$ case including the QCD higher-order corrections.
This difference is caused by the $\gamma_5$ structure in the Yukawa interaction.
As the result, it is found that
the $H$ interpretation on the di-tau excess is excluded by the ATLAS $t\bar{t}+\tau\bar{\tau}$ search, while the $A$ case is allowed, see Fig.~\ref{Fig:coupling}.
The available ATLAS data (the $tt(0\ell) + \tau_{\rm h} \tau_{\rm h}$ category) is pure hadronic final states.
We also hope the ATLAS collaboration to stop categorizing
and to combine both leptonic \cite{ATLAS:2017ztq,ATLAS:2019nvo} and hadronic decay modes \cite{ATLAS:2022yrq}
to increase the sensitivity to probe a possible new particle.
The similar analysis could also be applied in $t\bar{t}+\gamma\gamma$ search to probe a possible light resonance that can accommodate the CMS di-photon excess.
Note that the LEP $b\bar{b}$ excess implies that $\phi$ is the CP-even state ($H$) because the $Z$-boson-associated $A$ production vanishes.
Therefore, once the $b\bar{b}$ excess is involved seriously,
the CP symmetry must be violated in the scalar model.
Such CP-violating parameters (Yukawa or scalar self interactions) would induce the electron electric dipole moment at two-loop level ({\em e.g.}, Ref.~\cite{Fuyuto:2015ida}).
Furthermore, a search for $t\bar{t}+b\bar{b}$ signature would be interesting in light of the $b\bar{b}$ excess.
Although Fig.~12 of Ref.~\cite{ATLAS:2017fak} would be helpful to check the $b\bar{b}$ excess,
the measurement suffers from the QCD jets.
The more statistics and dedicated study are necessary to make a clear conclusion.
An additional vector particle ($A'$) would be also a possibility for the CMS di-tau excess. The simplest possibility is the hidden-photon model \cite{Fayet:2007ua,Pospelov:2008zw}.
The production of $A'$ is the DY process and the decay branching ratio is fixed [$\mathrm{BR}(A'\to\tau\bar{\tau})\simeq 15\%$] \cite{Buschmann:2015awa}.
Currently there is no direct experimental bound on the $A'$ production with mass region of around 90\,GeV at the LHC \cite{CMS:2019buh}.
Instead, an indirect bound comes from measurements of the EW precision observables at the LEP and Tevatron experiments
\cite{Curtin:2014cca}.
It is, however, beyond the scope of this paper to examine whether the $A'$ interpretation is allowed.
We also comment on our prescription for the $\chi^2$ test in Eq.~\eqref{eq:chisq}.
Our prescription is intuitive and does not rely on realistic Monte-Carlo simulations.
Necessary cut information for the detailed analysis
is not available from the experimental papers
because the BDT algorithm is adopted.
Our prescription may receive additional effects from the BDT,
since the BDT variables include the sub-leading $p_{\rm T}^\tau$ and missing transverse momentum.
Since the current ATLAS data is consistent with the CMS di-tau excess if a light CP-odd scalar is introduced,
it is nice to consider specific new physics models.
A light CP-odd scalar emerges as a pseudo-Nambu--Goldstone (NG) boson which comes from the spontaneous symmetry breaking of some global symmetries.
Such a mass of the pseudo-NG boson does not lead to an additional fine-tuning problem.
Within the G2HDM, additional scalars appear in the same mass scale with moderate mass differences of $\mathcal{O}$(VEV) (see, Ref.~\cite{Iguro:2017ysu} for instance).
It is possible to predict the lighter CP-odd scalar in general.
Nevertheless the mass degenerated setup, where only $\rho_{tt}$ and $\rho_{\tau\tau}$ are assumed to be nonzero, is not compatible with the direct search for the charged scalar in $pp\to tb + H^\pm\to tb+\tau\nu~{\rm{or}}~tb$ \cite{ATLAS:2018gfm,ATLAS:2021upq}.
Therefore the heavier charged scalar is necessary while keeping the CP-odd scalar mass to be around 95$\GeV$.\footnote{The mass degeneracy of a heavier CP-even scalar and a charged scalar is required to satisfy the constraint from the electroweak precision test.}
Another possible solution to this dilemma is to put the additional coupling and open up a new decay mode of a charged scalar.
However, additional Yukawa couplings could also contribute to the scalar production and suffers from the direct searches.
The weakest constrained Yukawa coupling is a top-charm flavor violating coupling, $\rho_{tc}$, where top quark is a left handed and charm quark is right handed.
Since the top mass is heavier than 100\,GeV, the coupling does not reduce the BR($\phi\to\tau\bar{\tau}$).
On the other hand the $SU(2)_L$ rotation generates $H^-\to b\bar{c}$ decay when $\rho_{tc}$ is not vanishing.
It is recently pointed out that low-mass di-bottom jets would be sensitive to this coupling depending on the mass \cite{Iguro:2022uzz}.
It is worthwhile to comment that the non-zero product of $\rho_{tc}$ and $\rho_{\tau\tau}$ can enhance $R_{D^{(*)}}={\rm{BR}}(\bar{B}\to D^{(*)} \tau\bar{\nu})/{\rm{BR}}(\bar{B}\to D^{(*)} \ell\bar{\nu}$), where $\ell=e,\,\mu$ and the 3--4$\sigma$ discrepancy is reported \cite{Aoki:2021kgd,Iguro:2020cpg},
consistently with the $B_c$-meson lifetime \cite{Blanke:2022pjy}.\footnote{For $m_{H^{-}}\ge400$\,GeV, $\tau\nu$ resonance searches exclude the interpretation \cite{Iguro:2018fni}.}
The non-zero $\rho_{tc}$ with the large mass difference between neutral scalars induces the same sign top signal \cite{Iguro:2018qzf}. This is also interesting as well as the search for a light scalar in the double-scalar production at the LHC \cite{ATLAS:2021fet}.
The more quantitative and dedicated study is beyond this paper and will be given elsewhere.
\section*{Acknowledgements}
S.\,I. would like to thank Shigeki Hirose, Yuta Takahashi, Ulrich Nierste, Andreas Crivellin, and Hantian Zhang for fruitful comments and valuable discussion.
T.\,K. thanks
Junji Hisano, Kazuhiro Tobe, and Nodoka Yamanaka
for useful discussion.
The work of S.\,I.~is supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under grant 396021762-TRR\,257.
T.\,K.~is supported by the Grant-in-Aid for Early-Career Scientists (No.\,19K14706) from the Ministry of Education, Culture, Sports, Science, and Technology (MEXT), Japan.
The work of Y.\,O.~is supported by Grant-in-Aid for Scientific research from the MEXT, Japan, No.\,19K03867.
This work is also supported by
the Japan Society for the Promotion of Science (JSPS) Core-to-Core Program,
No.\,JPJSCCA20200002.
\bibliographystyle{utphys28mod}
|
\section{Introduction}
\label{sec:intro}
A sunspot penumbra is formed in the presence of a strong and inclined magnetic field. To explain its observed brightness, some kind of convection must take place within it. This implies a system of vertical and horizontal flows. Upflows are observed predominately in the inner penumbra, while downflows are concentrated at the outer penumbral boundary \citep[e.g.][]{Franz2009}. This results in a strong horizontal radial outflow, as discovered by \citet{Evershed1909}.
The magnetic field in the penumbra can be characterised by two interlacing systems of field lines: one more vertical and one more horizontal \citep{SolankiMont1993}. Radially elongated parts of the penumbra with a more vertical and stronger field, called spines \citep{Lites1993}, alternate with a nearly horizontal field in interspines. The more vertical field wraps around the more inclined one, and the spines seem to be a continuation of the umbral magnetic field \citep{Tiwari2015}.
\citet{Schlich1998ApJ, Schlich1998AA} proposed a dynamical model of a flux tube in the interspine with a hot plasma upflow at a place observed as a bright head of a penumbral filament, oriented towards the umbra: the penumbral grain \citep[PG;][]{Muller1973}. In the photosphere, the hot plasma inside the tube is cooled by radiation as it flows horizontally outwards. As long as the flowing plasma is hotter than its surroundings, it constitutes a bright radial filament; however, at a longer distance the filament becomes gradually dark thanks to the radiative cooling. The tube constitutes a channel that carries the Evershed flow.
The classical picture of sunspot penumbrae consisting of bright and dark fibrils suffered from a number of problems and disagreements, as described, for example, in the extended reviews by \citet{Solanki2003}, \citet{BorreroIchimoto2011}, and \citet{Tiwari2017}. Particularly, the relations of the intensity structures to magnetic field and Evershed flow channels were found to be weak and unclear.
Recently, \citet{Tiwari2013} presented a model of a penumbral filament derived from inversions of Hinode spectropolarimetric observations. This filament represents a magnetoconvective cell, where hot plasma from sub-photospheric layers enters into the filament's head (the PG), which is the hottest and brightest part of the filament, with a magnetic field weaker than that in its surroundings (spines) and moderately inclined. The mass upflow of about 5~km\,s$^{-1}$ in the head then becomes horizontal in the filament's body, parallel to the $\sim$1~kG horizontal magnetic field, and leaks laterally, producing narrow downflow regions on the sides of the filament. The radial flow along the filament's axis generates the Evershed effect. The filament's body becomes gradually darker than its surroundings due to radiative cooling. The flow turns down at the end of the filament, producing a strong downflow of 7~km\,s$^{-1}$ and a local heating in deep photospheric layers. Tiwari's picture of a sunspot penumbra consisting of convective filaments and spines resolves most of the past issues. However, the volution of filaments and related horizontal motions are missing from this picture due to the limited cadence of Hinode spectropolarimetric data.
Measurements of horizontal (or proper) motions, using high-cadence data, can complement the concept of penumbrae obtained from the spectropolarimetric observations discussed above. Different methods can be applied to time series of intensity images or Dopplergrams to measure speeds of horizontal motions: local correlation tracking \citep[LCT,][]{NovSim1988}, feature tracking \citep[e.g.][]{SBSIII1999}, and time slices. It is necessary to note that the proper motions of intensity or velocity structures may not directly correspond to real plasma flows.
The LCT method suffers from inherent space and time averaging, which reduce the observed velocities, and from its inability to distinguish between motions of bright and dark features. The LCT results of \citet{WangZ1992}, \citet{Denker1998}, and recently \citet{Vafa2015} have revealed that PGs and dark features in the inner penumbra move towards the umbra (inwards) at a typical speed of 0.5~km\,s$^{-1}$ with a maximum of 0.7~km\,s$^{-1}$, and the elements in the outer penumbra move outwards at approximately the same speed with a maximum of 2~km\,s$^{-1}$. The inward motion of PGs, reported already by \citet{Muller1973}, was explained in terms of the dynamical flux-tube model \citep{Schlich1998AA}. The rise of the flux tube causes the crossing point of the tube with the visible surface, observed as a PG, to move radially inwards. It is clear that this motion does not correspond to the upward- and outward-oriented plasma flow in the tube, so the observed inward motions of PGs are apparent.
The feature-tracking method makes it possible to distinguish between different types of features. It relies on a correct image segmentation that isolates the objects under study. \citet{SBSIII1999} and \citet{SSIV2001} tracked bright features identified as PGs in high-resolution ground-based observations. They found inward- as well as outward-moving PGs with mean speeds of 0.4~km\,s$^{-1}$ and 0.5--0.8~km\,s$^{-1}$, respectively. The inward-moving PGs dominated the inner penumbra, while the majority of outward-moving PGs appeared in the outer penumbra. These findings were later confirmed by \citet{Zhang2013}, who applied their feature-tracking routines to Hinode observations from space. They reported mean speeds 0.7~km\,s$^{-1}$ for the inward-moving PGs and 0.65~km\,s$^{-1}$ for the outward-moving ones. In principle, the feature-tracking method could also be utilised to measure motions of dark features, that is, the bodies of penumbral filaments. However, unlike PGs, dark features in the penumbra have diffuse, largely elongated shapes, and they do not show clearly defined intensity minima; these properties make the tracking unreliable.
Time slices are an alternative to the two above-mentioned methods. They can be easily made from a time series of images by fixing a certain spatial track in the image plane and mapping the time evolution of values along this track. Time slices are 2D images, where the horizontal usually shows positions along the spatial track and the vertical corresponds to elapsed time. Although the time slices were widely used to depict oscillatory and wave phenomena, they were seldom applied to the evolution of penumbral structure. \citet{Shineetal1994} used this method to study motions of PGs and places with an increased Doppler signal of Evershed flow (`Evershed clouds') along several penumbral filaments in series of continuum images and Dopplergrams. They detected the inward motion of PGs (0.5~km\,s$^{-1}$) in the inner penumbra, the fast outward motion ($\sim$4~km\,s$^{-1}$) of places with increased Doppler signal in the outer penumbra, and advection flows ($\sim$1~km\,s$^{-1}$) away from the penumbra in the neighbouring granulation.
Proper motions of Evershed clouds were also measured by \citet{Cabreraetal2007} in a time series of Dopplergrams. They obtained outward-oriented speeds in the range 2--4~km\,s$^{-1}$.
In this work we apply the time-slice method to a six-hour-long time series of intensity images with high spatial and temporal resolution to measure proper motions of PGs, dark structures in the filaments' bodies, and G-band bright points (GBPs) in the neighbouring granulation. The length of the series also makes it possible to measure the expansion or contraction of the penumbra--granulation border. We have to note that spectropolarimetric observations were not made in our observing run, so magnetic-field parameters, temperature maps, and Dopplergrams are not used in our study.
\section{Observations and data processing} \label{sec:obs}
We observed a large sunspot with the 1-metre Swedish Solar Telescope \citep{SchaSST2003}, equipped with adaptive optics \citep{SchaAO2003}, under excellent seeing conditions on 18 June 2004. The leading sunspot in AR 10634 was growing to the maximum phase of evolution, and it was quite stable during the observing period, which took place from 07:43 to 15:30 UT. It was located close to the centre of the solar disc (N13, E12, heliocentric angle $\vartheta = 13.5\degr$, $\cos{\vartheta} = 0.97$), so observed proper motions of sunspot structures were only slightly influenced by the projection effect and were practically equal to horizontal motions.
Broadband images of the spot were acquired simultaneously in the G band ($430.89 \pm 0.6$~nm), blue continuum ($450.8 \pm 0.5$~nm), and red continuum ($602.0 \pm 1.3$~nm). Exposure times were 11--14~ms, and the pixel size was 0\farcs 0405 $\times$ 0\farcs 0405. Image acquisition in the frame-selection mode resulted in an average time difference between two selected frames of 20.1~s, which is the time resolution of our observation. The intensities in each image were calibrated using dark and flat-field frames and then normalised to the average intensity, $I_{\rm phot}$, of a quiet granulation area near the spot to compensate for changes of transparency and exposure time. The images were corrected for scattered light in the Earth's atmosphere and telescope, and the instrumental profile of the diffraction-limited 1-metre telescope was deconvolved simultaneously with noise filtering, using Wiener filters. The noise suppression started at 0\farcs 13 for the G band, 0\farcs 11 for the blue continuum, and 0\farcs 14 for the red continuum. Regarding the wavefront correction done by the adaptive optics, these values characterise the spatial resolution in the best frames. The deconvolved frames were of very good quality, and further image reconstruction was not necessary. Then, the image rotation was compensated for and the frames were aligned and de-stretched. Five-minute oscillations and a residual jitter caused by the seeing were removed by a subsonic filter with a cutoff at 4~km\,s$^{-1}$. The data reduction has been described in detail by \citet{SobPusch2009}, who used this dataset for a study of fine structures in the umbra.
\begin{figure}[t]\centering
\includegraphics[width=0.49\textwidth]{fig1_slice_pos3.eps}
\caption{Penumbra and neighbouring granulation of the sunspot observed in the red continuum, approximately in the middle of the series. Straight lines depict positions of time slices, identified by numbers 1--30. The distance between two neighbouring lines is 0\farcs 4. The arrow points to the disc centre. The temporal evolution of structures in the red continuum and G band is shown in two movies, available online.
\label{fig:slicepos}}
\end{figure}
We selected the series of 1090 red-continuum and 1090 \mbox{G-band} images acquired from 08:05 to 14:10~UT for the study of horizontal motions in the penumbra. The G-band images were re-aligned to the red ones, and a field of view of $32'' \times 31''$ was selected to cover a part of the umbra, the whole width of the penumbra, and the neighbouring granulation (Fig.~\ref{fig:slicepos}). Movies in both wavelength bands were created and are available online.
\begin{figure*}[t]\centering
\includegraphics[width=0.9\textwidth]{fig2_examp_dslices_red.eps}
\caption{Examples of time slices. Slices 16 and 8 (red continuum) show traces of bright PGs and dark bodies of filaments, respectively. Slices 8 and 19 (G band) show traces of GBPs and the expansion and contraction of the penumbra--granulation border, respectively.
\mbox{{\it Top row}}: Original time slices. {\it Bottom row}: Time slices enhanced by unsharp masking together with red lines drawn manually to measure speeds and durations of features.
\label{fig:sliceexamp}}
\end{figure*}
In an ideal case, a spatial track in the images that defines the time slice should follow the shape of a filament. Penumbral filaments are usually curved, and even though it would be possible to make a curved track along the filament axis, the filaments evolve but the track is fixed in time. This means that different structures can enter the spatial track throughout the time series. Moreover, it is difficult to distinguish dark bodies of filaments from spines when the information about magnetic and Doppler velocity fields is missing \citep[cf.][]{Tiwari2013}. To circumvent these problems, we defined spatial tracks as groups of densely spaced, straight parallel lines oriented approximately parallel to the filaments' directions. Two such groups, each of 15 lines, are shown in Fig.~\ref{fig:slicepos}. The distance between two neighbouring lines is 0\farcs 4, which is half of the typical filament width according to \citet{Tiwari2013}.
The group of lines covers an area where different features, for example parts of filaments, are evolving. They coincide with individual lines more or less randomly, but, in sum, nearly all of them form traces in the set of time slices of the group. These traces look like inclined bright or dark paths, the inclination of which corresponds to the speed of motion, and the vertical extent characterises the time period when a feature coincides in position with the line. This duration of a feature in a time slice is usually shorter than its real lifetime. Although in many cases an individual trace gives incomplete information on the speed and lifetime of a given feature, a statistic of numerous measurements can provide plausible results.
We created the time slices from 3D data cubes of red and G-band intensity movies. To reduce noise, the intensities at each position along each line were averaged over an area of $3 \times 3$ pixels (0\farcs 12)$^2$, which is comparable to the resolution limit. The lines begin in the umbra, cross the whole width of the penumbra, and end in the granulation (Fig.~\ref{fig:slicepos}). Their lengths are 26\farcs 7 (19.4~Mm, step 29~km) for the group of horizontal slices, 1--15, and 25\farcs 5 (18.5~Mm, step 32~km) for the group of inclined slices, 16--30.
We measured the speeds of horizontal motions and durations (lifetimes) of four types of features in the time slices. The red-continuum series was used to examine (1) PGs and (2) structures in dark bodies of penumbral filaments. The G-band series was used for measurements of (3) motions of the outer penumbral border and (4) photospheric GBPs. All the features are seen and coincide very well in both wavelength bands, but (1) and (2) are more visible in the less noisy red band and (3) and (4) have a higher contrast in the G band.
The measurements were made visually using lines drawn over the position--time traces of the features in the time slices. We took only clearly defined traces into account and used images of the time slices enhanced by unsharp masking to draw the lines. Some examples are shown in Fig.~\ref{fig:sliceexamp}. The top row from left to right shows time slices 16 (red), 8 (red), 8 (G band), and 19 (G band), with well-marked traces of PGs, dark bodies of filaments, GBPs, and positional changes of the outer penumbral border, respectively. The corresponding enhanced images together with the lines used to measure speeds and durations are displayed in the bottom row. These examples show only one type of feature in each individual slice for better clarity, but we measured all types of features in each time slice if their well-defined traces were present. Comparing the values of spatial (0\farcs 12 $\simeq$ 90~km) and time (20.1~s) resolutions with typical travel distances and durations of features, the relative error of an individual measurement is estimated to be smaller than 10\,\%.
\section{Results} \label{sec:results}
\begin{table}\centering
\caption{Mean velocities and lifetimes.} \label{tab:tabvel}
\begin{tabular}{lrrc}
\hline\hline
Feature & $n$ & $v$ [km\,s$^{-1}$] & $t$ [minutes] \\
\hline
PGs, inner penumbra & 546 & $-0.7 \pm 0.3$ & $19 \pm 8$ \\
PGs, outer penumbra & 126 & $-0.1 \pm 0.7$ & $20 \pm 9$ \\
Dark bodies of filaments & 566 & $3.6 \pm 0.9$ & $7 \pm 3 $ \\
Border expansion & 74 & $0.5 \pm 0.2$ & $60 \pm 23$\\
Border stationary & 18 & $0.0 \pm 0.1$ & $68 \pm 32$\\
Border contraction & 20 & $-2.2 \pm 1.3$ & $19 \pm 9$ \\
G-band bright points & 241 & $0.8 \pm 0.7$ & $16 \pm 9$ \\
\hline
\end{tabular}
\tablefoot{Symbols: $n$ is the number of measurements, and $v$ and $t$ are the mean horizontal speed and duration of the features, respectively. A negative sign for $v$ indicates motion towards the umbra, positive away from the umbra.}
\end{table}
Mean values of horizontal speeds and durations of different features are summarised in Table~\ref{tab:tabvel}. Standard deviations characterise the scatter of individual measurements. We assigned negative values of speeds to the features moving towards the umbra (inwards) and positive to those moving away from it (outwards). The motions of PGs located in the inner half of the penumbra are directed inwards with an average speed of $-0.7$~km\,s$^{-1}$, while PGs in the outer half of the penumbra move inwards and outwards, showing a larger scatter of speeds. The mean duration time of PGs of approximately 20 minutes is the same in the inner and outer penumbrae. Dark features in the bodies of filaments always move outwards with speeds in the range 2--6~km\,s$^{-1}$, and the mean speed is 3.6~km\,s$^{-1}$. Their mean duration, 7 minutes, is substantially shorter than that of PGs. Their traces often appear repeatedly in time slices at approximately the same position, with quasi-periods of 10 and 20~minutes estimated from Fourier analysis. This effect was already found by \citet{Shineetal1994} in time slices of Dopplergrams.
It can be observed in the movies that the position of the penumbra--granulation boundary changes with time. The length of the time series of 365 minutes makes it possible to follow these long-term changes. The penumbral border tends to expand gradually, with a speed of 0.5~km\,s$^{-1}$, over one hour on average and then shrinks rapidly when a granule or several granules are formed nearby. This is demonstrated in the example shown in the rightmost column of Fig.~\ref{fig:sliceexamp}. Sometimes the border remains stationary. In Table~\ref{tab:tabvel} we present, separately, the mean values for the expansion of the penumbral border (speed $v > 0.2$~km\,s$^{-1}$), the contraction of the penumbral border ($v < -0.2$~km\,s$^{-1}$) , and the stationary case ($-0.2 < v < 0.2$~km\,s$^{-1}$).
The tracking of GBPs provides complementary information about motions of magnetic elements in the vicinity of a sunspot. It can be seen in the G-band movie that many bright points appear close to the penumbral border, sometimes at the ends of penumbral filaments \citep{Bonetal2004}, and they follow the expansion of the border. Other GBPs appear in intergranular lanes, and their motions are influenced by the evolution of neighbouring granules, including the general outflow of granules in the sunspot moat \citep{Muller1987, Bonetal2005}. Speeds of GBPs observed in the time slices show a large scatter due to the interaction with granules, but their mean value, 0.8~km\,s$^{-1}$, is consistent with previous reports.
Histograms of horizontal speeds and durations are displayed in the top and bottom panels of Fig.~\ref{fig:histo}, respectively. Speeds of inward-moving PGs in the inner half of the penumbra show a narrow peak around $-0.7$~km\,s$^{-1}$ (solid black line). In the outer half of the penumbra, PGs move inwards and outwards with speeds of approximately $-1$ to 1~km\,s$^{-1}$ (green line), and some of them are stationary. The broad peak (dash-dot black) with high positive speeds corresponds to structures in dark filaments' bodies. The histogram of displacements of the outer penumbral border (red) has a strong peak at 0.5~km\,s$^{-1}$, corresponding to expansion, and some scattered contraction speeds (red dots) in the range $-3.5$ to $-0.5$~km\,s$^{-1}$. The speeds of GBPs are also mostly concentrated around 0.5~km\,s$^{-1}$, but their histogram (purple) has a tail of higher speeds of up to 3--4~km\,s$^{-1}$.
The histograms of the duration of PGs (black and green) and GBPs (purple) have similar asymmetric shapes, with maxima at shorter durations of 15 and 10~minutes, respectively. The time periods of motions of the outer penumbral border have a very broad range, but the contractions (red dots) clearly take much less time than the expansion and stationary periods (red lines).
\begin{figure}[t]\centering
\includegraphics[width=0.5\textwidth]{fig3_histo.eps}
\includegraphics[width=0.5\textwidth]{fig3tim_histo.eps}
\caption{Histograms of horizontal speeds ({\it top}) and durations (lifetimes) of the features ({\it bottom}). Black lines represent PGs in the inner half of the penumbra (solid line) and structures in dark bodies of filaments (dash-dot), green lines represent PGs in the outer half of the penumbra, red lines represent the motions of the outer penumbral border, both expansion and stationary (solid) and contraction (dotted), and purple lines represent GBPs.
\label{fig:histo}}
\end{figure}
The speeds of horizontal motions can be expected to depend on the positions of features in the penumbra. We defined a scale of relative distances from the umbra--penumbra border, setting zero at the inner and one at the outer penumbral boundaries. We defined the positions of the two boundaries by means of intensities averaged over the whole time period of 365 minutes for each individual time slice, adopting the values of 0.35\,$I_{\rm phot}$ for the inner and 0.90\,$I_{\rm phot}$ for the outer penumbral borders. The relative distance of each feature was then calculated from an average of the starting and ending horizontal coordinates of the line drawn over the feature's trace in the time slice.
A scatter plot of horizontal speeds versus relative distances from the umbra--penumbra border is displayed in Fig.~\ref{fig:scatplot}. We show the mean values and standard deviations of speeds calculated in 0.1 wide sections of the relative distance, $d$ (green lines).
The majority of measured PGs (black diamonds) are located in the inner third of the penumbra. These PGs move inwards, and their speeds do not depend on $d$. The speeds of PGs observed in the outer penumbra have, despite a large scatter, a clear trend of increasing with increasing $d$. The mean speed increases from $-0.7$~km\,s$^{-1}$ ($d \leq 0.4$) to 0.4~km\,s$^{-1}$ at $d = 0.8$. The dominant direction of PG motions changes from inwards to outwards at $d \simeq 0.6$. The mean outward speed of 0.5~km\,s$^{-1}$ is typical for the expansion periods of the outer penumbral border (red crosses). The major part of GBPs (purple squares) move away from the spot with a mean speed of 0.6~km\,s$^{-1}$; outlying speeds higher than 1.6~km\,s$^{-1}$, caused by local granular motions, are not included in this average. Moving structures in the dark bodies of filaments form a completely different system of largely scattered speeds. They are detected in the whole penumbra, and the majority of them appear at $d > 0.6$. Their mean speed, $3.6 \pm 0.9$~km\,s$^{-1}$, is independent of the relative position in the penumbra.
\begin{figure}[t]\centering
\includegraphics[width=0.5\textwidth]{fig4b_scatplot.eps}
\caption{Scatter plot of horizontal speeds versus relative distances of features from the umbra--penumbra border. Penumbral grains are marked by diamonds, structures in dark bodies of filaments by plus signs, motions of the outer penumbral border by red multiplication signs, and GBPs by purple squares.
Green lines show average speeds together with standard deviations. Distances 0 and 1 correspond to time-averaged positions of the inner and outer penumbral borders, respectively.
\label{fig:scatplot}}
\end{figure}
\section{Discussion and conclusions} \label{sec:discuss}
We have measured horizontal motions in the penumbra of a large stable sunspot. Speeds and duration times of PGs, dark bodies of penumbral filaments, GBPs near the sunspot, and the repeated expansion and contraction of the outer penumbral border were measured visually in time slices created from red-continuum and G-band intensity movies of high spatial (90~km) and temporal (20.1~s) resolutions.
The presented time-slice method has a limitation because the measured features can move from one slice to another. It can be seen in the movies that when the penumbral filaments are curved, some features, particularly PGs in the inner penumbra, can cross two to four slices. This obviously results in an underestimate of the lifetimes but also of the speeds because the features do not move parallel to the slices. Nonetheless, the majority of filaments appear radially elongated with minimal deviations, so the statistical analysis of many measurements can provide realistic results. Also, the visual measurement is subjective to some extent because of the visual selection of measured traces of features and the manual drawing of lines over them. On the other hand, the time-slice method and visual recognition has the advantage of avoiding numerous false and noisy identifications of features, which often appear in the results of automated feature tracking.
The apparent horizontal-motion speeds and durations of PGs found in this work are in good agreement with those obtained previously from feature tracking \citep{Zhang2013, SBSIII1999, SSIV2001}, including the change of prevailing direction of motions from inwards to outwards at the relative distance $d \simeq 0.6$. The mean speed increases gradually from negative to positive values between $d = 0.4$ and 0.8 and matches the mean expansion speed of the outer penumbral border well. Likewise, the mean outflow speed of magnetic elements represented by GBPs in the sunspot moat is similar to the mean expansion speed of the penumbral border (Fig.~\ref{fig:scatplot}).
Speeds detected in the dark bodies of penumbral filaments are consistent with previous measurements of proper motions of Evershed clouds \citep{Shineetal1994, Cabreraetal2007}. We have to note that these speeds may be underestimated in our work because the subsonic filter used to suppress five-minute oscillations and residual seeing effects has a cutoff at 4~km\,s$^{-1}$, causing a reduction in observed speeds. The outward motions in the bodies of filaments can be expected to be associated with the Evershed flow.
According to \citet{Tiwari2013}, penumbral filaments represent magnetoconvective cells distributed everywhere in the penumbra and surrounded by a magnetic field of spines, the inclination of which gradually increases with increasing distance from the umbra. In the inner penumbra, the magnetic field in the filament's head (PG) is more inclined to the normal than that in its surroundings, but the opposite is true in the middle and outer penumbrae, as shown in Fig.~6 of \citet{Tiwari2013}. The same figure also shows that the magnetic field in a PG is weaker than the surrounding one in the inner penumbra but that both fields are comparable in the middle and outer penumbrae.
We suggest the following tentative scenario, which, however, needs to be proven by numerical simulations that are beyond the scope of this paper. Rising hot plasma surrounded by a stronger and less inclined magnetic field may adapt its trajectory to be more vertical, which would lead to the inward motion of PGs in the inner penumbra. Oppositely, in the outer penumbra the rising hot plasma in the filament's head is dragged by the surrounding, more horizontal magnetic field such that its crossing point with the visible surface (PG) moves outwards. The outer penumbral border is formed by the ends of filaments. Its repeated expansion with speeds comparable to that of outward-moving PGs can be tentatively explained if we assume that the filaments (magnetoconvective cells) in the outer penumbra move outwards as a whole. The periods of expansion are interrupted by the formation and growth of granules close to the penumbral border.
\begin{acknowledgements}
We thank the referee for constructive comments, which helped us to improve the paper.
We express our thanks to C. M\"ostl, R. Kever, and R. Henderson for assisting in the observations and J. Jur\v{c}\'{a}k for discussion. This work was done under the institutional support ASU:67985815 of the Czech Academy of Sciences. The Swedish 1-m Solar Telescope is operated on the island of La Palma by the Institute for Solar Physics of the Royal Swedish Academy of Sciences in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrof\' {\i}sica de Canarias.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\IEEEPARstart{I}{ndoor} human tracking with low power \mbox{millimeter-wave} (\mbox{mmWave}) radar sensors has been receiving considerable attention in the last few years, due to its wide applicability to the Internet of Things (IoT)~\cite{shah2019rf}.
The typical aim of these systems is to exploit the reflected signal from human subjects to infer their {\it state} in the physical space, e.g., their position and movement speed~\cite{knudde2017indoor,zhao2019mid, pegoraro2020multiperson, pegoraro2021real}.
\rev{In this paper, we address the limitations of widely used Bayesian tracking techniques, such as the extended Kalman filter (EKF), which require strong assumptions about the movement process, e.g., constant velocity. Despite being widely used in the literature \cite{zhao2019mid, pegoraro2021real, pegoraro2020multiperson}, these methods only work sufficiently well in practice because of the high frame rates of mmWave radar devices, but their capability of grasping the complexity of human movement is severely limited.
In real environments, people often follow random and unpredictable trajectories, which do not match standard radar target movement models. This causes large predicted uncertainties when using model-based Bayesian filtering approaches, reflecting the intrinsic limitations of legacy models in human movement analysis.}
To resolve this, we advocate the use of a \mbox{model-free} and \mbox{end-to-end} deep learning approach. In addition, and to the best of our knowledge, we are the first to introduce to the radar field the concept of \textit{heteroscedastic}, i.e., sample-varying, uncertainty estimation for neural network (NN) architectures. Modeling the uncertainty in the state estimates allows obtaining an error covariance matrix associated with the NN predictions: recently, this has been successfully applied to computer vision problems~\cite{kendall2017uncertainties}. Note that such covariances are most valuable in indoor radar systems to increase the performance of processing blocks such as {\it data association}, in the case of \textit{(i)} multiple subjects being tracked concurrently, or \textit{(ii)} multiple radars with overlapping fields of view. The contributions of our work are:
\textbf{1)} We design a maximum-likelihood convolutional-recurrent neural network (ML-CRNN), based on gated recurrent units (GRU)~\cite{cho2014learning}. This NN outputs an estimate of the current subject state (position and velocity) given an arbitrarily long sequence of past radar observations, without making any assumptions on the underlying movement process. The \mbox{ML-CRNN} handles both the \textit{vision} part of the problem, processing the raw data, and the \mbox{non-linear} target \textit{tracking} part.
\textbf{2)} The proposed ML-CRNN outputs an heteroscedastic error covariance matrix paired with each state prediction, which weighs the confidence level of the state estimates. This is achieved by making the NN output the error covariance matrix of the state estimate, and training it as a probabilistic model via a Gaussian negative log-likelihood (NLL) loss function.
\textbf{3)} We design \mbox{ML-CRNN} for \textit{end-to-end} training (no pre-processing). While in the literature denoising and clustering phases are customary, \cite{knudde2017indoor,zhao2019mid, pegoraro2021real}, ML-CRNN sequentially processes {\it raw} range-Doppler/range-azimuth radar images.
Numerical results are obtained on our own experimental data, using a Frequency Modulated Continuous Wave (FMCW) INRAS RadarLog device working in the $77-81$~GHz band. The evaluation scenario is challenging and realistic, with furniture and other humans, in addition to the tracked subject.
\section{FMCW mmWave radar signal model}
\label{sec:radar}
A multiple-input multiple-output (MIMO) FMCW radar allows the joint estimation of the distance, the angular position and the radial velocity of the target(s) with respect to the radar device. To achieve this, the radar transmits sequences of chirp signals and measures the frequency shift of the reflection at its receiving antennas. Next, we provide a brief overview of the FMCW radar signal model, detailing the parameters that are used in this work. For a more comprehensive description, the reader may refer to~\cite{patole2017automotive, winkler2007range}.
We use an INRAS RadarLog FMCW radar with one transmitting antenna and \mbox{$\Gamma=16$} receiving antennas, organized as a linear array. The frequency of the transmitted chirp signal (TX) is linearly increased from a base value of \mbox{$f_o = 77$~GHz} to a maximum \mbox{$f_1 = 81$~GHz} in \mbox{$T_c = 180$~$\mu$s} (a \textit{sweep}). We define the bandwidth of the chirp as \mbox{$B = f_1 - f_o = 4$~GHz}. The chirps are transmitted every \mbox{$T_{\rm rep} = 250$ $\mu$s} in sequences of \mbox{$P = 256$} sweeps. For each of the $16$ antenna elements, a mixer combines the attenuated and delayed received signal (RX) with the transmitted one, generating the intermediate frequency (IF) signal. The IF signal is sampled along three different dimensions. First, \emph{fast time} sampling returns \mbox{$N=1024$} points from each chirp. For the \emph{slow time} (or Doppler) sampling, $P$ samples, one per chirp from adjacent chirps, are taken with period $T_{\rm rep}$. Finally, the \emph{spatial} sampling relates to the $\Gamma$ receiving channels, spaced apart by a distance $d$, and enables the localization of the targets in the physical space. A discrete Fourier transform (DFT) is applied along each sampling dimension and the square of its magnitude is computed to extract the power density for each frequency component. The resulting 3-dimensional signal is referred to as range-Doppler-azimuth (RDA) map, and the position of the power peaks along each axis can be associated with the radial distance, the angular position and the radial velocity of the subjects~\cite{patole2017automotive}.
The RDA maps are outputted by the radar at every \mbox{time-step}, with a rate of $15$~fps, and have a dimension of \mbox{$1024\times 64 \times 64$} points, due to the resolution used for the DFT along the fast time, angular and slow time dimensions respectively.
\section{Method}
We define the state of a human subject at a certain \mbox{time-step} $t$ as the vector \mbox{$\mathbf{x}_{t} = \left[x_t, y_t, v^x_t, v^y_t\right]^T \in \mathbb{R}^4$}, containing the Cartesian coordinates of the subject in the space, $x_t$ and $y_t$, and the velocity components $v^x_t$ and $v^y_t$. Our aim is to track the subject, namely, to sequentially estimate their current state across time, using a sequence of past and present observations of the system (\textit{filtering} problem). To this end, we design a recurrent NN that extracts information from a sequence of $T$ consecutive radar RDA maps, identified by index \mbox{$t=1, \dots, T$}, and that performs a regression task producing an estimate of the state, $\hat{\mathbf{x}}_{t}$. In contrast to typical regression approaches based on NNs, we wish to estimate not only the state of the subject, but also the corresponding error covariance, \mbox{$\mathbf{\Sigma}_{t} = E\left[\left(\mathbf{x}_{t} - \hat{\mathbf{x}}_{t}\right)\left(\mathbf{x}_{t} - \hat{\mathbf{x}}_{t}\right)^T\right]$}.
\subsection{Learning from raw data}
Processing the raw RDA maps from the radar can be computationally very expensive given their size.
To mitigate this, we first select only the range interval of interest from the fast time dimension, i.e., the first $134$ points, that correspond to distances from $0$ to $5$~m. The resulting $134 \times 64 \times 64$ RDA map is then projected onto the range-Doppler (RD) plane by integrating along the azimuth dimension and onto the range-azimuth (RA) plane by integrating along the Doppler dimension.
In this way, at each time-step $t$ we obtain a pair of $134\times 64$ images, denoted by $\mathbf{M}_t^{\rm RD}$ and $\mathbf{M}_t^{\rm RA}$, see \fig{fig:radar-maps}. Images are normalized so that pixels intensities lie in the interval $[0,1]$.
\begin{figure}[t!]
\begin{center}
\subcaptionbox{RD image, $\mathbf{M}^{\rm RD}_t$.\label{fig:rd}}{\includegraphics[width=4.3cm]{figures/rd_ex.pdf}}
\subcaptionbox{RA image, $\mathbf{M}^{\rm RA}_t$.\label{fig:ra}}{\includegraphics[width=4.3cm]{figures/ra_ex.pdf}}
\caption{Example RD and RA images. The target corresponds to the peak in received power around 2~m.}
\label{fig:radar-maps}
\end{center}
\end{figure}
\subsection{Proposed neural network architecture}
The proposed ML-CRNN combines convolutional layers operating on the single time-step with a recurrent structure based on GRU layers \cite{cho2014learning}, as shown in \fig{fig:spawc-mlcrnn}.
We can identify a convolutional and a recurrent block, which are connected together forming the ML-CRNN model and which are trained jointly via a common loss function (see \secref{sec:cov-est}).
The \textbf{convolutional block} is a convolutional neural network (CNN) that takes as input two images per time step, namely the RD and RA projections of the RDA map, \mbox{$\mathbf{M}_t^{\rm RD} \in \left[0,1\right]^{134\times 64}$} and \mbox{$\mathbf{M}_t^{\rm RA}\in \left[0,1\right]^{134\times 64}$}.
Given this input, the CNN learns a non-linear composite function $C$ that maps $\mathbf{M}_t^{\rm RD}$ and $\mathbf{M}_t^{\rm RA}$ onto a vector \mbox{$\mathbf{o}_t \in \mathbb{R}^{16}$}, called \textit{compressed} observation. The function is based on two parallel network branches that extract features from each input image separately, and that are then combined into a single output. Each branch, denoted by \mbox{$i \in \left\{{\rm RD}, {\rm RA}\right\}$}, is the composition of $L$ functions, which are the \textit{layers} of the CNN, $f^i_L\left(\dots f^i_1\left( \mathbf{M}_t^{i}\right)\right)$,
with \mbox{$L = 4$}.
Each layer $\ell$ computes the elementwise $\rm ELU$ activation function \cite{clevert2015fast} of the sum between the convolution of the input $X$ with $d_{\ell}$ \mbox{$3\times 3$} learned kernels, $\mathbf{K}_{\ell}^i$, and a bias parameter $b_{\ell}^i$: \mbox{$f_{\ell}^i\left( X\right) = {\rm ELU}\left(\mathbf{K}_{\ell}^i * X + b_{\ell}^i\right)$}.
The term $d_{\ell}$ represents the number of \textit{feature maps} of each layer and is equal to $4, 8, 16, 4$ for layer $\ell=1,2,3,4$, respectively. The kernels are applied using \textit{stride} \mbox{$2\times 2$}, that means they are shifted by two positions at each step of the convolution, resulting in a dimensionality reduction of a factor $2$ at each layer\footnote{Zero-padding is applied to maintain the correct shape of the data.}. The output of each branch, is reshaped into a vector $\mathbf{y}^i_t $, and the two outputs are concatenated into $\mathbf{y}_t$. The final layer of the convolutional block processes $\mathbf{y}_t$ using a fully connected (FC) layer with \textit{dropout} probability $p=0.33$. Dropout refers to randomly setting to 0 the output of some NN nodes during training as a regularization method \cite{srivastava2014dropout}.
The FC layer applies the function \mbox{$\mathbf{o}_t = {\rm ELU}\left(\mathbf{W}_{\rm fc}\mathbf{y}_t + \mathbf{b}_{\rm fc}\right) $}, with parameters $\mathbf{W}_{\rm fc}, \mathbf{b}_{\rm fc}$. Input radar images are processed in sequences of $T$ frames, applying in parallel the CNN block to each pair of RD and RA maps and obtaining a sequence of compressed observations, denoted by $\mathbf{o}_{1:T}$.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=8.5cm]{figures/spawc_mlcrnn.pdf}
\caption{Block diagram of the ML-CRNN architecture. }
\label{fig:spawc-mlcrnn}
\end{center}
\end{figure}
The \textbf{recurrent block} is a recurrent layer featuring GRU cells \cite{cho2014learning}. GRU cells maintain a hidden state across \mbox{time-steps}, processing it together with the current input vector to learn temporal dependencies in the input sequence (see \cite{cho2014learning} for the detailed description of a GRU cell).
The recurrent block takes as input $\mathbf{o}_{1:T}$ and outputs a sequence of estimates of the unobservable target states $\hat{\mathbf{x}}_{1:T}$ and the corresponding error covariance matrices $\hat{\mathbf{\Sigma}}_{1:T}$.
The hidden states are denoted by $\mathbf{h}_{1:T}$ and
have dimension $128$ each.
At each time-step, they are further processed with $3$ FC output layers to compute the state estimate, $\hat{\mathbf{x}}_t \in \mathbb{R}^4$, and two vectors $\boldsymbol{\alpha}_t \in \mathbb{R}_+^4$ and $\boldsymbol{\beta}_t \in \mathbb{R}^6$, which are used to build the covariance matrix estimate $\hat{\mathbf{\Sigma}}_t$, as described in \secref{sec:cov-est}.
The expressions of the FC layers are the following
\begin{equation}\label{eq:state-est}
\hat{\mathbf{x}}_t = \mathbf{W}_x\mathbf{h}_t + \mathbf{b}_x,
\end{equation}
\begin{equation}\label{eq:sigma-est}
\boldsymbol{\alpha}_t = \exp\left(\mathbf{W}_{\alpha}\mathbf{h}_t + \mathbf{b}_{\alpha}\right),
\end{equation}
\begin{equation}\label{eq:rho-est}
\boldsymbol{\beta}_t = \tanh \left(\mathbf{W}_{\beta}\mathbf{h}_t + \mathbf{b}_{\beta}\right),
\end{equation}
where we denoted by $\mathbf{W}_x, \mathbf{b}_x, \mathbf{W}_{\alpha}, \mathbf{b}_{\alpha}, \mathbf{W}_{\beta}, \mathbf{b}_{\beta}$ the learned weights and biases. In the recurrent block, \textit{recurrent dropout} is applied with probability \mbox{$p=0.33$} as described in~\cite{gal2016theoretically}. The ML-CRNN contains a total of $66730$ trainable parameters.
\subsection{Maximum likelihood state and covariance estimation} \label{sec:cov-est}
To model the uncertainty on the state estimates, we assume that the posterior distribution of the state given the observation sequence of length $T$ is Gaussian with mean $\hat{\mathbf{x}}_t$:
\mbox{$p(\mathbf{x}_t|\mathbf{M}^{\rm RD}_{1:T}, \mathbf{M}^{\rm RA}_{1:T}) \sim \mathcal{N}(\hat{\mathbf{x}}_t, \hat{\mathbf{\Sigma}}_t)$}.
Moreover, we let $\hat{\mathbf{\Sigma}}_t$ depend on the time-step, in order to reflect the variable uncertainty that affects radar measurements due to many factors, like the range-dependent power attenuation, the clutter distribution and the variability in the movement process.
The covariance matrix $\mathbf{\Sigma}_t$ must be symmetric and positive definite, and can be modeled as the sum of an aleatoric and an epistemic component~\cite{kendall2017uncertainties, russell2021multivariate}, as detailed next.
\textbf{1)} \textbf{Aleatoric covariance}, $\mathbf{\Sigma}_t^a$, is the uncertainty related to the {\it intrinsic noise} in the state evolution and measurement processes. It is estimated directly from the vectors $\boldsymbol{\alpha}_t, \boldsymbol{\beta}_t$ outputted by the ML-CRNN, see \eq{eq:sigma-est} and \eq{eq:rho-est}, using the Cholesky decomposition \mbox{$\hat{\mathbf{\Sigma}}_t^a = \mathbf{L}_t\mathbf{L}^T_t$} where $\mathbf{L}_t$ is a lower triangular matrix \cite{watkins2004fundamentals}.
To ensure that the covariance matrix is positive semi-definite it is sufficient that the diagonal elements of $\mathbf{L}_t$ are all non-negative. Vector $\boldsymbol{\alpha}_t$ is obtained using an exponential activation function, therefore its elements are positive and can be used as the diagonal elements of $\mathbf{L}_t$, namely \mbox{$[\mathbf{L}_{t}]_{i,i} = \alpha_{i, t}$}.
The six off-diagonal elements of $\mathbf{L}_t$ correspond to the elements of vector $\boldsymbol{\beta}_t$, that are placed following an arbitrary (but consistent across iterations) order.
Once the aleatoric covariance is obtained applying the above transformations,
a maximum-likelihood (ML) approach is used to jointly optimize the state and the covariance estimates, interpreting the training phase as fitting a probabilistic model. In particular, we use the negative log-likelihood of a multivariate Gaussian as the loss function of the ML-CRNN
\begin{equation}\label{eq:loss}
\ell(\mathbf{x}_t, \hat{\mathbf{x}}_t, \hat{\mathbf{\Sigma}}^a_t) = \left(\mathbf{x}_t - \hat{\mathbf{x}}_t\right)^T (\hat{\mathbf{\Sigma}}^{a}_t)^{-1}\left(\mathbf{x}_t - \hat{\mathbf{x}}_t\right) + \ln |\hat{\mathbf{\Sigma}}^a_t|,
\end{equation}
where both $\hat{\mathbf{x}}_t$ and $\hat{\mathbf{\Sigma}}^a_t$ are outputted by the network at each time-step. The total loss on the sequence of $T$ frames is computed as \mbox{$\mathcal{L} = \sum_{t=1}^T\ell(\mathbf{x}_t, \hat{\mathbf{x}}_t, \hat{\mathbf{\Sigma}}^a_t)/T$}. Training the network by minimizing \eq{eq:loss} amounts to maximizing the likelihood that the predicted state and covariance actually represent the parameters of a Gaussian probabilistic model.
\textbf{2)} \textbf{Epistemic covariance}, $\mathbf{\Sigma}_t^e$, is due to the {\it uncertainty in the prediction} made by the deep learning model. It can be estimated using Monte-Carlo (MC) dropout~\cite{gal2016dropout}. This method consists in applying the dropout procedure during inference, making the NN output random even for a fixed input. MC dropout uses the last NN available at time $t$, and is not part of the NN parameter learning process.
\textbf{Total variance}. The ML-CRNN model at time $t$ can be run $M$ times for each input with MC dropout. In this way, $M$ different state and covariance samples are obtained for the same input. The time index is dropped here for convenience, as we operate within a single time-step. We respectively denote by $\hat{\mathbf{x}}_m$ and $\hat{\mathbf{\Sigma}}_m^a$, $m = 1, \dots, M$, the state and covariance predictions outputted by the NN, while we refer to their empirical averages over the $M$ samples as $\bar{\mathbf{x}} = \sum_m \hat{\mathbf{x}}_m / M$ and $\bar{\mathbf{\Sigma}}^a = \sum_m \hat{\mathbf{\Sigma}}_m^a / M$. Using the sample covariance estimator for $\mathbf{\Sigma}^e$, and $\bar{\mathbf{\Sigma}}^a$ as the sample mean estimate of $\mathbf{\Sigma}^a$, the total error covariance matrix can be expressed as \cite{russell2021multivariate}
\begin{multline}\label{eq:tot-cov}
\hat{\mathbf{\Sigma}} \approx \underbrace{\frac{1}{M}\sum_{m=1}^M (\hat{\mathbf{x}}_m - \bar{\mathbf{x}})(\hat{\mathbf{x}}_m - \bar{\mathbf{x}})^T}_{\textrm{epistemic}} + \underbrace{\bar{\mathbf{\Sigma}}^a}_{\textrm{aleatoric}}.\\
= \frac{1}{M}\sum_{m=1}^M \hat{\mathbf{x}}_m \hat{\mathbf{x}}_m^T -\bar{\mathbf{x}}\bar{\mathbf{x}}^T + \bar{\mathbf{\Sigma}}^a.\\
\end{multline}
$\bar{\mathbf{x}}$ is typically more precise than a single sample from the MC dropout, so the final system outputs at step $t$ are the averaged estimate \mbox{$\hat{\mathbf{x}}_t= \bar{\mathbf{x}}_t$} and its covariance $\hat{\mathbf{\Sigma}}_t$ from \eq{eq:tot-cov}. The complete procedure is summarized in \fig{fig:spawc-sys}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=8.5cm]{figures/spawc_sys.pdf}
\caption{Block diagram of the ML state and covariance estimation.}
\label{fig:spawc-sys}
\end{center}
\end{figure}
\section{Results}
\begin{figure*}[t!]
\begin{center}
\centering
\subcaptionbox{\label{fig:nll}}{\includegraphics[width=6.1cm]{figures/nll.pdf}}
\subcaptionbox{\label{fig:std-comp}}{\includegraphics[width=5.9cm]{figures/std-comp.pdf}}
\subcaptionbox{\label{fig:cal}}{\includegraphics[width=5.8cm]{figures/cal_plot_ukf.pdf}}
\caption{(a) Negative log-likelihood (NLL) on test data for ML-CRNN and UKF. (b) Estimate of $v_y$ (top) and its predicted standard deviation (bottom). (c) Comparison between the empirical distribution of the squared Mahalanobis distances and the theoretical one.}
\label{fig:results}
\end{center}
\end{figure*}
The proposed NN architecture is trained and tested on experimental measurements taken in a $8$~m~$\times~4$~m research laboratory equipped with a motion tracking system with six \mbox{infra-red} cameras. A total of $10$~minutes of training data and $1$~minute of test data were collected for a single subject moving inside a $4$~m~$\times~2$~m rectangle, i.e., the working area of the motion tracking system. These measurements were taken in realistic conditions, with furniture and other people inside the room, but outside the tracking area. This makes the radar images highly cluttered, with bursts of frames in which the target subject is undetectable. Ground truth data is concurrently acquired by the motion tracking system, which was \mbox{time-synchronized} with the radar
We trained the ML-CRNN using the loss in \eq{eq:loss} and the Adam optimizer \cite{kingma2014adam}, using a subset of the training data as the validation set. Training was stopped when the loss reached convergence on the validation set, and the epistemic covariance on the training set became negligible. The length of the temporal sequences that are fed to the NN \textit{during training} is \mbox{$T=10$} ($0.667$~s). During the evaluation, instead, the predictions are obtained by inputting a new radar frame in the ML-CRNN as soon as it becomes available. The ML-CRNN then uses the hidden state, $\mathbf{h}_t$, and the current input to compute the prediction, similarly to how Bayesian filtering methods operate. The following metrics were used to evaluate the tracking error: \textit{(i)}~root mean square error (RMSE) and \textit{(ii)}~localization error outage, \mbox{${\rm LEO}(0.2) = {\rm Prob}\left(||\mathbf{x} - \hat{\mathbf{x}}|| > 20 \mbox{ cm}\right)$}.
\subsection{Performance}
\begin{table}[t!]
\small
\begin{center}
\begin{tabular}{lccc}
\toprule
& \multicolumn{2}{c}{{\bf Position}} & {\bf Velocity} \\
\cmidrule(lr){2-4}
\textbf{Method} & RMSE [cm] & LEO$\left(0.2\right)$ [\%] & RMSE [cm/s]\\
\midrule
UKF & 32.8 & 47.1 & 56.8\\
MSE-CRNN & 12.8 & 7.30 & 20.1\\
ML-CRNN & 7.59 & 0.62 & 14.0\\
\bottomrule
\end{tabular}
\end{center}
\caption{Tracking error of UKF and the proposed CRNN network with ML (ML-CRNN) and MSE (MSE-S2S) loss criteria. \label{tab:comp}}
\end{table}
\textbf{Tracking --} In \tab{tab:comp}, we show the results obtained by the proposed CRNN model with the ML loss of \eq{eq:loss} (ML-CRNN), compared to the same model trained with standard MSE loss (MSE-CRNN), i.e., without the covariance estimation, and to an unscented Kalman filter (UKF), which is a widely adopted Bayesian filtering method for estimating the posterior state distribution in non-linear radar tracking~\cite{julier1997new}. The parameters of the UKF have been optimized via grid search on the same training dataset used for the ML-CRNN.
Both CRNN methods are clearly superior in tracking accuracy to the UKF, showing an RMSE reduction of about $0.25$~m in the location accuracy and $0.4$~m/s in the velocity estimation. Note that the UKF cannot perform tracking from high-dimensional raw data: denoising and clustering are needed to transform the RDA maps into vectors containing the range and angular position of the target. For this purpose, we implemented the clustering method used in \cite{pegoraro2020multiperson, pegoraro2021real, zhao2019mid}, based on the DBSCAN~\cite{ester1996density} algorithm.
A further important aspect is the effect of NN training by using the ML loss: in addition to getting an estimate of the prediction uncertainty, we also observed a considerable improvement of the tracking accuracy. The ML-CRNN learning process is indeed less affected by outliers in the radar measurements due to its probabilistic nature, assigning low importance to unlikely observations.
\textbf{Uncertainty estimation --} To gauge the quality of the obtained uncertainty estimates, we first compare the NLL from ML-CRNN against that of UKF, see \fig{fig:nll}, using $M=25$ MC samples in \eq{eq:tot-cov}. Note that, in practice, the value of $M$ can be tuned to trade off between computational complexity and quality of the resulting uncertainty estimation. We notice that ML-CRNN is slower to converge, due to its long-term dependency on past inputs, (\mbox{$T=10$} time-steps), but achieves much better performance (smaller NLL) after the initial transient, showing its superior capability of capturing the underlying human movement model.
In \fig{fig:std-comp}, we show the relation between the predicted uncertainty and the position estimates for ML-CRNN, focusing on the $y$ component of the velocity, $v_y$. We see that the uncertainty (standard deviation of $v_y$, bottom) shows a positive peak when $v_y$ changes rapidly (top).
A further way to investigate the quality of the covariance is to compare the empirical frequency of the squared Mahalanobis distance, $\xi_t = \left(\mathbf{x}_t - \hat{\mathbf{x}}_t\right)^T (\hat{\mathbf{\Sigma}}_t)^{-1}\left(\mathbf{x}_t - \hat{\mathbf{x}}_t\right)$, on the test measurements, with its theoretical probability distribution.
Due to the Gaussian posterior probability assumption for the state, it can be shown that $\xi_t$ should follow a $\chi^2$ distribution with 4 degrees of freedom (the state dimension)~\cite{bar2009probabilistic}.
In \fig{fig:cal}, we plot a comparison between the empirical frequency and the theoretical value of the probability distribution of $\xi_t$. An ideal calibration of the uncertainty would yield a perfect match between the two, as in the black line. From our experiment, we see that a clear improvement is obtained with ML-CRNN by using both the aleatoric and the epistemic components of the covariance, over using either of them in isolation. In particular, only using the epistemic component leads to severely underestimating the uncertainty, because it neglects the intrinsic variability of the movement process of the target.
On the other hand, the UKF shows inferior calibration quality, which denotes the limitations of the underlying movement model. Quantitatively, the calibration mean-squared errors between the ideal case (perfect calibration) and the predicted uncertainty are $9\cdot 10^{-4}$ and $5\cdot 10^{-3}$ for the ML-CRNN ad the UKF, respectively.
\section{Conclusions}
In this paper, we proposed a convolutional recurrent neural network to track human movement in indoor spaces by means of a mmWave MIMO FMCW radar. Our model estimates position and velocity of the subjects from raw radar data with superior accuracy with respect to \mbox{state-of-the-art} techniques, and without requiring any assumptions on the movement evolution process. The proposed neural network is trained as a probabilistic model using a maximum-likelihood loss function, obtaining explicit {\it uncertainty estimates} at its output, in the form of a time-varying error covariance matrices. This, besides allowing one to gauge the uncertainty in the tracking process, also leads to greatly improved performance against the best approaches from the literature, i.e., the unscented Kalman filter, lowering the average tracking error from $32.8$ to $7.59$~cm and from $56.8$ to $14$~cm/s in terms of position and velocity, respectively.
Future research work includes the integration of deep learning models for object detection and recognition in the ML-CRNN. This would allow simultaneously detecting multiple targets, obtaining a probability distribution of their position and recognizing the target nature, e.g., person, pet, vehicle, etc.
|
\section{Introduction}
Subjective tinnitus is the perception of sounds without real acoustic stimulus and can only be heard by the patients themselves~\cite{henry2008using,reavis2012temporary}. People with tinnitus may suffer from sleep problems, difficulty concentrating, and negative emotions that may progress to psychiatric disorders~\cite{nova2006tinitus, espinosa2014tratamiento, eggermont2004neuroscience}. Subjective tinnitus is a worldwide problem, affecting around 15\% of the population~\cite{heller2003classification,nondahl2010ten}. Therefore, it is essential to study the diagnoses and treatments of subjective tinnitus.
Recently, there have been substantial achievements in tinnitus diagnosis. Among them, methods based on electroencephalography (EEG) signals attract much attention because 1) many works prove that there exists a strong correlation between abnormal EEG and tinnitus~\cite{eggermont2015maladaptive,ibarra2020acoustic}; 2) EEG signals have objective evaluation compared with traditional methods that are based on the Visual Analog Scale (VAS) or questionnaires~\cite{haefeli2006pain,kersten2014pain,powers2015acoustic,adamchic2014reversing}. With the development of information technology, many efforts are made to utilize machine learning and deep learning algorithms to analyze EEG signals and classify patients' tinnitus status~\cite{sun2019multi,liu2021generalizable,li2016svm,saeidi2021neural,allgaier2021deep,zhou2021tinnitus}. For example, Liu et al.~\cite{liu2021generalizable} design a neural network with
Siamese structure to augment training EEG data.
Despite the success of these methods, few of them equip with generalization abilities, especially when applying the models to datasets where distributions are mismatched with the training datasets. The collection and annotation of EEG signals of tinnitus are labor-intensive and time-consuming, which makes labeled training data hard to acquire. However, a large amount of labeled data are fundamental to the success of deep learning models. Therefore, we hope to adapt models trained on labeled datasets to other unlabeled datasets. However, EEG signals of different datasets can be sufficiently different due to EEG signals' high sensitivity to non-physiological factors, e.g., biological or physical conditions of brains~\cite{ibarra2020acoustic} and experimental settings such as time duration, sampling frequency, electrode placement, and equipment. The distribution discrepancy prevents the model's cross-dataset tinnitus diagnosis. In other EEG-based fields, a few works explore unsupervised domain adaptation or transfer learning to overcome the distribution mismatch across datasets~\cite{chai2016unsupervised,cimtay2020investigating,xu2020cross,rayatdoost2018cross}. However, in tinnitus diagnosis, methods that can generalize to other datasets without labeled data are under-explored.
In addition, although unsupervised domain adaptation can relieve the distribution discrepancy and enable a cross-dataset diagnosis to some extent, the adaptation process inevitably causes loss of useful information and introduces noise that is irrelevant to or even harms tinnitus diagnosis. Also, the original EEG signals have low signal-to-noise ratio, i.e., the data contains noise even without adaptation to other datasets. Therefore, designing models that can perform adaptation to transfer knowledge across datasets and exclude tinnitus-irrelevant information is fundamental to improving the cross-dataset diagnosis performance.
Apart from noise caused during signal collection and domain adaptation, the side information of the ears, i.e., which ear the evoking acoustic signal is presented to, may also impair the classification. The right and left ears display different EEG patterns even they have the same condition as tinnitus or normal~\cite{mikkelsen2015eeg,lindenberg2016neural}. This derives from functional and anatomical inter-hemispheric differences in auditory cortex~\cite{shtyrov1999noise,rodrigo2008language,penhune1996interhemispheric,dorsaint2006asymmetries}. However, this is not noise that can be excluded. Thus, it is essential to adapt the information and patterns of the left and right ears at a subject level to eliminate the side influence.
We propose a novel model to perform Disentangled and Side-aware Unsupervised Domain Adaptation (DSUDA) for cross-dataset subjective tinnitus diagnosis to address the aforementioned problems. We propose Disentangled Auto-Encoder (DAE) to decouple tinnitus-related and tinnitus-unrelated information. The performance can be improved as noise is excluded and only pure information is fed into the classifier. In addition, we present the Side-aware Unsupervised Domain Adaptation (SUDA) to perform subject- and domain-level adaptation. The side-aware adaptation aligns features of left and right ears to avoid intrinsic left-right EEG differences harming the classification. The domain-level adaptation cooperates with DAE to extract domain variance, i.e., the tinnitus-unrelated information that must be excluded, and adapt the variance across domains. Then unsupervised domain adaptation can be achieved that the classifier trained on the source domain can be applied in the target domain without labeled data of the source domain. This is essential for EEG-based tinnitus diagnosis as it may reduce the high demand for labeled EEG signals, save cost, and find universal models shared across subjects, datasets, and different experimental settings.
In summary, we make three-fold contributions as follows:
\begin{itemize}
\item We propose a novel model aiming at disentangled and side-aware unsupervised domain adaptation (DSUDA) for cross-dataset subjective tinnitus diagnosis. DSUDA combines disentangled auto-encoder and side-aware unsupervised domain adaptation, thus retaining the discrimination power when applied to the target domain.
\item We design an adversarial training framework for the cooperation of the proposed disentangled auto-encoder and side-aware unsupervised domain adaptation. Training the two modules in an adversarial manner can distill class-related information and align data from different sides and domains.
\item We conduct extensive experiments and ablation studies to evaluate and analyze our DSUDA. Our model yields 8.6\%, 10\%, and 7.5\% improvements over state-of-the-art (SOTA) methods in both-ears, left-ear, and right-ear accuracy, and consistently outperforms SOTA in other criteria, e.g., Negative-F1 score, Positive-F1 score, and Negative Predictive Value, etc. The results demonstrate that our model's generalization ability on new domain.
\end{itemize}
\section{Related work}
\subsection{EEG-based Tinnitus Diagnosis and Treatment}
Subjective tinnitus is defined as the perception of sound that can only be perceived by patients in the absence of an external sound source ~\cite{henry2008using,reavis2012temporary}. Tinnitus, also known as phantom noise, can occur in people with or without hearing problems. Tinnitus has always been a worldwide health problem, with an overall incidence rate of about 11\% to 16\%~\cite{heller2003classification,nondahl2010ten}. Tinnitus can impair the life quality of a patient in many ways, which may cause sleep problems, difficulty concentrating, and may even develop into psychiatric disorders, such as lethargy, anxiety, and anger~\cite{ nova2006tinitus, espinosa2014tratamiento, eggermont2004neuroscience}. For the treatment and diagnosis of tinnitus, traditional investigators assess with a Visual Analog Scale (VAS) or a temporary questionnaire~\cite{haefeli2006pain,kersten2014pain,powers2015acoustic,adamchic2014reversing}. Haefeli et al.~\cite{haefeli2006pain} and Kersten et al.~\cite{kersten2014pain} designed VAS to subjectively quantify certain sensations of patients using a sliding scale from lack to extreme sensation, e.g., pain. Different to the sliding scale, Powers et al.~\cite{powers2015acoustic} and Adamchic et al.~\cite{adamchic2014reversing} proposed to use questionnaires with only three responses, i.e., yes, sometimes, and no, to identify the tinnitus conditions of sufferers. With the development of Electroencephalogram (EEG), EEG has become a valuable tool to diagnose and treat tinnitus. EEG is capable of recording brain activity and analyzing neural oscillations in real time~\cite{ibarra2020acoustic}. A lot of evidence shows a strong correlation between abnormal EEG and tinnitus. For example, Eggermont et al.~\cite{eggermont2015maladaptive}, and Ibarra-Zarate et al.~\cite{ibarra2020acoustic} suggested that tinnitus may be caused by abnormal restructurings in the tonotopic maps, higher synchrony or spontaneous firing rates in the neuron or the central auditory system. Nguyen et al.~\cite{nguyen2020abnormal} found that tinnitus sufferers have abnormal neural oscillations in the beta and gamma frequency bands.
Traditional EEG-based tinnitus diagnosis and treatment rely on the statistical models to quantify and analyze the provided neurofeedback~\cite{7yearsfeedbackbutwork,focusonsinglebetaband}. For example, Vanneste et al.~\cite{vanneste2015tinnitus} performed a quantitative EEG analysis of 154 patients with tinnitus and used a regression analysis to find that structural differences in gray matter may be associated with the distress and duration of tinnitus. Moazami-Goudarzi et al.~\cite{moazami2010temporo} used power spectra to analyze tinnitus and control groups in a group-wise manner. They found that tinnitus will lead to a power enhancement in the delta and theta band of EEG. Although previous researchers have made significant progress in diagnosing and analyzing tinnitus-related EEG signals, their statistical analysis was case-specific and prone to being trapped in the characteristics of small study populations. To ease such problem, more and more work uses deep learning methods to improve the generalization ability of models~\cite{allgaier2021deep,liu2021generalizable,sun2019multi,zhou2021tinnitus}. For example, Liu et al.~\cite{liu2021generalizable} adopted a Siamese network to augment training samples with contrastive relationships and thus enhance the generalization ability of their model. Allgaier et al.~\cite{allgaier2021deep} used noise reduction and down-sampling to generalize the learning process of their proposed deep learning method and achieved an accuracy of 75.6\% in diagnosing tinnitus. However, these methods are still limited to a single dataset. EEG signals are easy to be influenced by multiple non-physiological factors, e.g., biological or physical conditions of brains~\cite{ibarra2020acoustic}. The data distributions may significantly differ from subjects. Moreover, the different settings of the experimental environment, e.g., time duration and sampling frequency, may lead to deep learning models failing to be applied in a new dataset. Therefore, it is necessary to develop a model that is robust across datasets to handle more diverse and general EEG signals.
\subsection{Cross-dataset EEG Research In Related Fields}
Most current EEG research~\cite{wan2021review,santana2019gp,dai2019domain,zhang2017cross} only takes into account the data distributions across subjects, which still ignores the distribution shift between datasets. Some modern models~\cite{cimtay2020investigating,xu2020cross,rayatdoost2018cross,bird2020cross,lin2019constructing} have evolved to ease the distribution shifts across both subjects and datasets. In other words, these models can be applied in multiple datasets rather than a single dataset. For example, in the field of gesture classification, Bird et al.~\cite{bird2020cross} proposed to transfer knowledge between an EEG dataset recording brainwaves and an Electromyographic (EMG) dataset recording muscular waves based on the pre-trained networks. They achieved a high accuracy of 85.12\% and 93.82\% by transferring EEG-to-EMG and EMG-to-EEG knowledge, respectively. Similarly, for emotion recognition, Cimtay et al.~\cite{cimtay2020investigating} proposed to use a pre-trained neural network from a source dataset to the target dataset and achieved an impressive enhancement in the model performance. These models use pre-trained models to provide extra knowledge that may be meaningful across datasets. To provide a more reliable solution to learn meaningful information across datasets, Xu et al.~\cite{xu2020cross} proposed an online method to pre-align the data formats before training and classification. However, all these methods rely on the annotation of target dataset, while many real-world datasets may lack labels. Therefore, it is necessary to propose a work that can be applied in a more general condition, i.e., only a given source dataset is annotated, and the target dataset that needs classification remains unlabeled.
\subsection{Unsupervised Domain Adaptation}
A domain refers to feature space and the corresponding probability distributions related to a set of data, e.g., a dataset~\cite{pan2009survey}. Unsupervised domain adaptation focuses on the situation where the training and testing data are collected from different distributions, e.g., datasets. The training dataset with labeled data and the testing dataset without any annotation are called the source and target datasets, respectively. Conventional deep learning models may suffer from a significant decrease in the model performance due to the domain difference between source and target datasets~\cite{wilson2020survey}. Unsupervised
domain adaptation aims to ease such performance decrease by adapting the different domains~\cite{lu2015transfer,shao2014transfer,tan2018survey,weiss2016survey}. A lot of works~\cite{chai2016unsupervised,jimenez2020custom,yoo2021transferring,hang2019cross} have been done to adopt unsupervised domain adaptation in the field of EEG. Most of these works only view each subject as a domain and fail to ease the dataset-level domain difference. Limited research~\cite{chai2017fast,lan2018domain} has been done to adopt unsupervised domain adaptation for cross-dataset settings. For example, Lan et al.~\cite{lan2018domain} adopted domain adaptation to reduce both cross-subject and cross-dataset variance based on unsupervised domain adaptation and achieved an impressive improvement in emotion recognition. However, the previous work lacks domain-specific knowledge to handle domain differences related to tinnitus, e.g., the domain difference caused by the different sides of the ear of presentation of the evoking stimulus. Our work is tinnitus-specific domain adaptation that is capable of easing cross-ear and cross-dataset differences.
\subsection{Disentangled Representation}
To better serve specific tasks, many deep-learning-based generative adversarial networks~\cite{kim2018disentangling,tran2017disentangled,lee2018diverse,zhu2020s3vae} disentangle the representations from the given data to learn the specific information. For example, Li et al.~\cite{li2021generalized} used a disentangled auto-encoder to disentangle category-distilling and category-dispersing
factors from images for better category classification. Similar, to EEG signals, there exists more information or noise than tinnitus-related signals in the collected EEG. For example, Cai et al.~\cite{cai2020aberrant} adopted spatial smoothing to increase the signal-to-noise ratio and reduce the influence of signal noise. Joos et al.~\cite{joos2012disentangling} and Meyer et al.~\cite{meyer2014disentangling} used questionnaires to disentangle and study the correlations of depression and distress to EEG from tinnitus patients. Therefore, we propose to disentangle the signals to exclude unrelated factors to obtain the pure classification information for tinnitus diagnosis.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.8\textwidth]{figures/model_update.pdf}
\caption{Overview of DSUDA. DSUDA consists of two modules: DAE and DUDA. DAE is composed of Pure-info Encoder, Domain-variance Encoder, Classifier, and Decoder. The labeled data from source domain are first preprocessed to align the unlabeled data from target domain, then the data are normalized to fed into DAE. The two encoders of DAE disentangle the input data into domain-invariant features and domain variance, both of which are then combined as the input of the Decoder. The Classifier takes domain-invariant features as input to predict tinnitus status. DUDA comprises Side Discriminator to distinguish which ear the signal are collected from, and Domain Discriminator to discriminate which domain the data belong to. The two modules are trained together in an adversarial manner to optimize losses $L_{cls}$, $L_{side}$, $L_{ae}$, and $L_d$. In the inference stage, we can directly use the Pure-info Encoder to extract pure class information for prediction.}
\label{model}
\end{figure*}
\begin{figure}[htb]
\centering
\includegraphics[width=0.44\textwidth]{figures/datapreprocess.pdf}
\caption{Data preprocessing.}
\label{data preprocess}
\end{figure}
\section{Method}
\subsection{Problem Definition}
Suppose that dataset $\mathcal{S}=\{(x^{s}, y^{s}, a^{s})|x^{s}\in R^{1\times n^{s}},y^{s}\in\{0,1\},a^{s}\in\{0,1\}\}$ from source domain are given, where $x^{s}=\{o_{j}:j\in[1,n^{s}],o_{j}\in R\}$ is an EEG feature containing signals collected through $n^s$ time points, $y^{s}\in\{0,1\}$ indicates the subject's tinnitus status, and $a^{s}\in\{0,1\}$ implies which ear the signals are collected from. More specifically, $y^{s}=1$ denotes that the subject has tinnitus, and $y^{s}=0$ represents control subject without tinnitus. For $a^{s}$, 0 or 1 denotes left or right ear, respectively. Then, given data from target domain $\mathcal{T}=\{(x^{t}, y^{t}, a^{t})|x^{t}\in R^{1\times n^{t}},y^{t}\in\{0,1\},a^{t}\in\{0,1\}\}$, where $x^{t}$ contains $n^t$ time points, and $y^t$ ($a^t$, respectively) indicates same meanings as $y^s$ ($a^s$, respectively). Our goal is to obtain a general classifier that can diagnose tinnitus for the target domain without labeled target data based on the labeled source data. That is, during training, only $\mathcal{S}$ from the source domain and $\{(x^{t},a^{t})\}$ of the target domain are given, and we need to implement unsupervised domain adaptation to achieve cross-datasets diagnosis.
\subsection{Pipeline Overview}
As Figure~\ref{model} shows, our proposed DSUDA integrates two modules: the Disentangled Auto-Encoder (DAE) and the Side-aware Unsupervised Domain Adaptation (SUDA). DAE (Section~\ref{dae}) disentangles input features from source and target domains into pure class information that are domain-invariant, and domain variance that are diverse noise across datasets caused by different experimental environments. SUDA (Section~\ref{suda}) consists of two-level domain adaptation: the subject-level side adaptation to align features from right and left ears to eliminate the side variance, and the domain-level variance adaptation to exclude domain-variance and noises from pure information for better classification. We accomplish tinnitus diagnosis, feature disentanglement, and domain adaptation based on three kinds of loss functions, namely: (1) reconstruction loss $L_{ae}$, (2) classification loss $L_{cls}$, and (3) domain discrimination loss $L_{side}$ and $L_d$.
\subsection{Data Preprocessing}
Due to the different acquisition equipment and settings of EEG signals, the data in the source and target domains may have significant discrepancies. The overall time duration and time steps of the collection may be different. And the ranges of signal strength or device noise differ across devices. Therefore, it is not feasible to directly apply a classifier suitable for the source domain to the target domain. We need to preprocess the source data to align the data format and scope.
As shown in Figure~\ref{data preprocess}, suppose that an EEG trial duration from source and target domains, i.e., $x^s$ and $x^t$, last for $d^s$ and $d^t$ milliseconds, respectively. We first adopt the strategy of sliding window to segment $x^s$ into slices $\{x^s_1,x^s_2,...\}$, where the window size equals $d^t$, and the sliding step is a hyper parameter. Then we use down sampling to compress $x^s_i=\{o^s_1,o^s_2,...\}$ to align the number of time points in $x^s_i$ (i.e., $n^w=\frac{d^t*n^s}{d^s}$) and $x^t$ (i.e., $n^t$) as follows:
\begin{gather}\label{slide_down_sample}
\bar{o}^{s}_{p}=\left\{\begin{matrix}
\frac{1}{l}\sum_{q\in [(p-1)*l+1,p*l]}o^{s}_{q} & p\in[1,m]\\
\frac{1}{l+1}\sum_{q\in [p*(l+1)-l-m,p*(l+1)-l]}o^{s}_{q} & p\in[m+1,n^{t}]
\end{matrix}\right.\\
s.t. \quad l = \left \lfloor
\frac{n^{w}}{n^{t}} \right \rfloor \quad m = n^{w}-l*n_{t}
\end{gather}
where $l=\left \lfloor
\frac{n^{w}}{n^{t}} \right \rfloor$, the floor of $\frac{n^{w}}{n^{t}}$, is the compressing step. In this way, we obtain a set of processed features $\{\bar{x}^s_1,\bar{x}^s_2,...\}$ from the original $x^s$, where $\bar{x}^s_i = \{\bar{o}^{s}_{1},\bar{o}^{s}_{2},...\}$ is consistent with $x^t$ in time duration and time point number. Note that we perform the data preprocessing based on the assumptions that $d^s \geq d^t$ and $n^w \geq n^t$, or we align target domain to source domain in reverse.
Then, we normalize data from source and target domains with MinMaxNorm separately as follows:
\begin{gather}\label{min_max_norm}
\hat{o}^{s}_{p}=\frac{\bar{o}^{s}_{p}-\min \bar{o}^{s}_{p}}{\max \bar{o}^{s}_{p} - \min \bar{o}^{s}_{p}}\\
\hat{o}^{t}_{p}=\frac{o^{t}_{p}-\min o^{t}_{p}}{\max o^{t}_{p} - \min o^{t}_{p}}
\end{gather}
where $o^t_p$ is the $p^{th}$ value in $x^t$.
\subsection{Disentangled Auto-encoder}\label{dae}
DAE aims to improve tinnitus diagnosis by extracting pure information that excludes categorically irrelevant factors. For this, we build a DAE using two parallel encoders. The purely informative encoder $f_{pe}$ (Pure-info Encoder) learns classification-related features, and the classifier $f_c$ supervises $f_{pe}$ to contain classification information. The domain variance encoder $f_{ve}$ learns class-independent features. The decoder $f_{de}$ constrains $f_{pe}$ to avoid overfitting the classification information, and restricts $f_{pe}$ and $f_{ve}$ from overfitting the source domain.
We take $\hat{\mathcal{S}}=\{(\hat{x}^s, y^s, a^s)\}$ and $\hat{\mathcal{T}}=\{(\hat{x}^t, a^t)\}$ as the input of DAE, where $\hat{x}^s=\{\hat{x^s_1}, \hat{x^s_2},...\}$, $\hat{x^s_1} = \{\hat{o}^{s}_{1},\hat{o}^{s}_{2},...\}$, and $\hat{x}^t = \{\hat{o}^{t}_{1},\hat{o}^{t}_{2},...\}$. Given $\hat{x}\in \mathcal{D}=\mathcal{S}\cup\mathcal{T}$, DAE disentangles $\hat{x}$ into class-related embedding $e^c$ and class-unrelated embedding $e^v$. Then $e^c$ is fed into the classifier $f_c$ for classification, and $e^c$ and $e^v$ are concatenated as input of the decoder $f_{de}$ for reconstruction. DAE is optimized by the classification loss $L_{cls}$ measured by CrossEntropy, and the reconstruction loss $L_{ae}$ calculated by MSE as follows:
\begin{gather}
L_{cls}=CrossEntropy(f_{c}(f_{pe}(\hat{x}^{s})),y^{s}) \label{equ1}
\\
L_{ae}=MSE(f_{de}(f_{pe}(\hat{x}),f_{ve}(\hat{x})),\hat{x}) \label{equ2}
\end{gather}
where $CrossEntropy$ and $MSE$ denote cross-entropy loss and mean square error, respectively.
Note that samples in $\hat{\mathcal{T}}$ are not labeled with tinnitus status. Thus, $L_{ae}$ is calculated based on all $\hat{x}\in \mathcal{D}$, while $L_{cls}$ is computed based on only $\hat{x}^{s} \in \hat{\mathcal{S}}$.
However, only loss $L_{cls}$ and $L_{ae}$ are not sufficient. With $L_{cls}$ and $L_{ae}$, we can only ensure that $f_c$ works for the source domain. To enable its feasibility in the target domain, we hope the extracted class-related information $e^c$ is domain-invariant to generalize to the target domain. Intuitively, after excluding the domain variance from the source domain, the remained information is domain-invariant. Therefore, we can consider the class-unrelated features $e^v$ that should be discarded as the domain variance, and adapt $e^v$ from the source domain to the target domain by the side-aware unsupervised domain adaptation, which will be discussed in Section~\ref{suda}.
\subsection{Side-aware Unsupervised Domain Adaptation}\label{suda}
The features learned by DAE may not fit target domain. To adapt DAE to the target domain, we propose the SUDA focusing on adapting domain variance from source to target domain. To avoid the confusion caused by different signal patterns of right and left ears, we further equip the SUDA with the side-aware adaptation to align information from both ears.
As shown in Figure~\ref{model}, SUDA is composed of the domain ($f_{dd}$) and side discriminator ($f_{sd}$), predicting which domain the domain-variant features belong to and which ear the domain-invariant features are collected from, respectively. Then, we design the loss functions of the domain and side discriminator as $L_d$ and $L_{side}$, which aims to optimize $f_{dd}$ and $f_{sd}$ to be able to distinguish source domain variance $\hat{x}^s$ as 0, target domain variance $\hat{x}^t$ as 1, left ear features as 0, and right ear features as 1:
\begin{gather}
L_{d}=\mathbb{E}_{\hat{x}\sim{\mathcal{T}}}[f_{dd}(f_{ve}(\hat{x}))]-\mathbb{E}_{\hat{x}\sim{\mathcal{S}}}[f_{dd}(f_{ve}(\hat{x}))] \label{equ3}
\\
L_{side}=\mathbb{E}_{\hat{x}\sim{\mathcal{D}^r}}[f_{sd}(f_{pe}(\hat{x}))]-\mathbb{E}_{\hat{x}\sim{\mathcal{D}^l}}[f_{sd}(f_{pe}(\hat{x}))] \label{equ4}
\\
L_{DS} = L_{d}+\eta*L_{side}
\end{gather}
where $\mathcal{D}^r$ and $\mathcal{D}^l$ denote sets of $\hat{x}$ that collected from right and left ears, respectively, $ L_{DS}$ is the overall discriminating loss, and $\eta$ is a hyper-parameter.
With losses $L_d$ and $L_{side}$, we can obtain reliable discriminators. Then, we can optimize DAE by optimizing $f_{ve}$ to confuse the domain discriminator to extract domain variance, and optimizing $f_{pe}$ to confuse the side discriminator to align side information. Intuitively, if the domain discriminators cannot distinguish class-irrelevant features from different domains, the domain variance is adapted successfully and $f_{pe}$ can disentangle class-related information from $\hat{x}$ for both domains. Then, we can apply the classifier trained on source domain to target domain.
We optimize DAE and SUDA in an adversarial manner as follows: 1) maximizing $L_{DS}$ to enable $f_{dd}$ and $f_{sd}$ to distinguish between different domains and ear sides; 2) optimizing $f_{ve}$, $f_{pe}$, $f_c$, and $f_{de}$ to minimize $L_{DAE}$ to improve the classifier and fool the discriminators. $L_{DAE}$ is the combined loss based on Equations~\ref{equ1}-\ref{equ4} as follows:
\begin{equation}
\begin{split}
L_{DAE}= &\alpha * L'_{d} + \beta*L'_{side} + L_{ae} +L_{cls}
\\
=&-\mathbb{E}_{\hat{x}\sim{\mathcal{T}}}[f_{dd}(f_{ve}(\hat{x}))]+\mathbb{E}_{\hat{x}\sim{\mathcal{D}^l}}[f_{sd}(f_{pe}(\hat{x}))]
\\&+ L_{ae} +L_{cls}
\end{split}
\end{equation}
where $\alpha$ and $\beta$ are two hyper-parameters, $L'_{d}$ and $L'_{side}$ are simplified versions of $L_{d}$ and $L_{side}$ without requirements of recognizing data from source domain or right ear. This is because that optimizing DAE to fit target domain to the source domain and align information from left to right side is sufficient for adaptation between two domains.
\subsection{Training and Inference}
\textbf{Training.} Due to the instability of adversarial training, we first train Pure-info Encoder, Classifier, and Decoder to provide a basic classification ability, and use the trained parameters of Pure-info Encoder as parameters of Domain-variance Encoder. Then, we consider the trained DAE as the initial model, and optimize DAE and SUDA based on $L_{DAE}$ and $L_{DS}$ adversarially.
\textbf{Inference.} We use Pure-info Encoder to extract class-related features from target domain data and use the Classifier to predict tinnitus status directly without modifying its parameters.
\begin{table*}
\centering
\caption{Overall comparisons on both ears with SOTA methods.}
\label{Main_exp_both}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccccc}
\toprule
\multirow{2}{*}{Model} & \multicolumn{7}{c}{Both Sides} \\
\cmidrule{2-8}
& NPV & TNR & N-F1 & PPV & TPR & P-F1 & Acc \\
\midrule
XGBoost & 0.467(0.000) & 0.525(0.000) & 0.494(0.000) & 0.457(0.000) & 0.400(0.000) & 0.427(0.000) & 0.463(0.000) \\
Nu-SVC & 0.467(0.000) & 0.350(0.000) & 0.400(0.000) & 0.480(0.000) & 0.600(0.000) & 0.533(0.000) & 0.475(0.000) \\
nCSP & 0.568(0.055) & 0.625(0.078) & 0.595(0.058) & 0.583(0.054) & 0.525(0.085) & 0.553(0.063) & 0.575(0.053) \\
DeepNet & 0.714(0.243) & 0.125(0.042) & 0.213(0.070) & 0.521(0.010) & 0.950(0.019) & 0.673(0.009) & 0.538(0.018) \\
ShalowNet & 0.581(0.044) & 0.625(0.147) & 0.602(0.112) & 0.595(0.033) & 0.550(0.110) & 0.571(0.040) & 0.588(0.035) \\
AE-XGB & 0.696(0.071) & 0.400(0.170) & 0.508(0.158) & 0.579(0.033) & 0.825(0.126) & 0.680(0.041) & 0.613(0.041) \\
EEGNET & 0.605(0.081) & 0.650(0.230) & 0.627(0.181) & 0.622(0.048) & 0.575(0.159) & 0.597(0.055) & 0.613(0.049) \\
AE & 0.660(0.047) & 0.775(0.077) & 0.713(0.032) & 0.727(0.104) & 0.600(0.154) & 0.658(0.143) & 0.688(0.056) \\
SAE & 0.703(0.029) & 0.650(0.063) & 0.675(0.018) & 0.674(0.027) & 0.725(0.076) & 0.699(0.034) & 0.688(0.013) \\
\midrule
UDA & 0.833(0.041) & 0.500(0.118) & 0.625(0.050) & 0.643(0.050) & 0.900(0.105) & 0.750(0.035) & 0.700(0.018) \\
SUDA & 0.744(0.022) & 0.725(0.051) & 0.734(0.017) & 0.732(0.030) & 0.750(0.054) & 0.741(0.022) & 0.738(0.012) \\
DSUDA & 0.789(0.031) & 0.750(0.057) & 0.769(0.024) & 0.762(0.032) & 0.800(0.057) & 0.780(0.023) & 0.775(0.018) \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}
\centering
\caption{Results on the left ear.}
\label{Main_exp_left}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccccc}
\toprule
\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\\Model\end{tabular}} & \multicolumn{7}{c}{Left Side} \\
\cmidrule{2-8}
& NPV & TNR & N-F1 & PPV & TPR & P-F1 & Acc \\
\midrule
XGBoost & 0.455(0.000) & 0.500(0.000) & 0.476(0.000) & 0.444(0.000) & 0.400(0.000) & 0.421(0.000) & 0.450(0.000) \\
Nu-SVC & 0.467(0.000) & 0.350(0.000) & 0.400(0.000) & 0.480(0.000) & 0.600(0.000) & 0.533(0.000) & 0.475(0.000) \\
nCSP & 0.591(0.088) & 0.650(0.109) & 0.619(0.090) & 0.611(0.078) & 0.550(0.114) & 0.579(0.091) & 0.600(0.081) \\
DeepNet & 0.500(0.245) & 0.100(0.053) & 0.167(0.085) & 0.500(0.016) & 0.900(0.026) & 0.643(0.017) & 0.575(0.029) \\
ShalowNet & 0.524(0.063) & 0.550(0.136) & 0.537(0.110) & 0.526(0.036) & 0.500(0.122) & 0.513(0.055) & 0.525(0.044) \\
AE-XGB & 0.615(0.143) & 0.400(0.179) & 0.485(0.169) & 0.556(0.049) & 0.750(0.146) & 0.638(0.065) & 0.575(0.060) \\
EEGNET & 0.600(0.115) & 0.600(0.240) & 0.600(0.196) & 0.600(0.054) & 0.600(0.169) & 0.600(0.068) & 0.600(0.062) \\
AE & 0.652(0.053) & 0.750(0.078) & 0.698(0.043) & 0.706(0.157) & 0.600(0.159) & 0.649(0.161) & 0.675(0.069) \\
SAE & 0.750(0.038) & 0.600(0.077) & 0.667(0.035) & 0.667(0.036) & 0.800(0.073) & 0.727(0.030) & 0.700(0.023) \\
\midrule
UDA & 0.900(0.063) & 0.450(0.137) & 0.600(0.075) & 0.633(0.072) & 0.950(0.111) & 0.760(0.051) & 0.700(0.047) \\
SUDA & 0.737(0.028) & 0.700(0.054) & 0.718(0.027) & 0.714(0.034) & 0.750(0.054) & 0.732(0.028) & 0.725(0.023) \\
DSUDA & 0.833(0.051) & 0.750(0.070) & 0.789(0.045) & 0.773(0.042) & 0.850(0.064) & 0.810(0.037) & 0.800(0.037) \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}
\centering
\caption{Results on the right ear.}
\label{Main_exp_right}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccccc}
\toprule
\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\\Model\end{tabular}} & \multicolumn{7}{c}{Right Side} \\
\cmidrule{2-8}
& NPV & TNR & N-F1 & PPV & TPR & P-F1 & Acc \\
\midrule
XGBoost & 0.478(0.000) & 0.550(0.000) & 0.512(0.000) & 0.471(0.000) & 0.400(0.000) & 0.432(0.000) & 0.475(0.000) \\
Nu-SVC & 0.467(0.000) & 0.350(0.000) & 0.400(0.000) & 0.480(0.000) & 0.600(0.000) & 0.533(0.000) & 0.475(0.000) \\
nCSP & 0.545(0.068) & 0.600(0.099) & 0.571(0.077) & 0.556(0.062) & 0.500(0.109) & 0.526(0.078) & 0.550(0.063) \\
DeepNet & 1.000(0.416) & 0.150(0.047) & 0.261(0.082) & 0.541(0.011) & 1.000(0.028) & 0.702(0.012) & 0.500(0.021) \\
ShalowNet & 0.636(0.049) & 0.700(0.166) & 0.667(0.125) & 0.667(0.045) & 0.600(0.104) & 0.632(0.033) & 0.650(0.043) \\
AE-XGB & 0.800(0.115) & 0.400(0.174) & 0.533(0.164) & 0.600(0.039) & 0.900(0.127) & 0.720(0.044) & 0.650(0.051) \\
EEGNET & 0.609(0.060) & 0.700(0.226) & 0.651(0.170) & 0.647(0.054) & 0.550(0.155) & 0.595(0.050) & 0.625(0.046) \\
AE & 0.667(0.049) & 0.800(0.098) & 0.727(0.037) & 0.750(0.110) & 0.600(0.160) & 0.667(0.140) & 0.700(0.055) \\
SAE & 0.667(0.038) & 0.700(0.063) & 0.683(0.027) & 0.684(0.037) & 0.650(0.090) & 0.667(0.054) & 0.675(0.031) \\
\midrule
UDA & 0.786(0.064) & 0.550(0.120) & 0.647(0.056) & 0.654(0.063) & 0.850(0.126) & 0.739(0.065) & 0.700(0.043) \\
SUDA & 0.750(0.031) & 0.750(0.060) & 0.750(0.023) & 0.750(0.042) & 0.750(0.070) & 0.750(0.036) & 0.750(0.023) \\
DSUDA & 0.750(0.046) & 0.750(0.074) & 0.750(0.037) & 0.750(0.056) & 0.750(0.082) & 0.750(0.047) & 0.750(0.037) \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{figure*}[htb]
\centering
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/both_side_auc.pdf}
\centering
\caption{ROC curve of both sides.}
\end{subfigure}\hfil
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/left_side_auc.pdf}
\centering
\caption{ROC curve of left side.}
\end{subfigure}\hfil
\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\textwidth]{figures/right_side_auc.pdf}
\centering
\caption{ROC curve of right side.}
\end{subfigure}\hfil
\caption{ROC curves of DSUDA. Class 0 and 1 represent subjects with or without subjective tinnitus, respectively.}\label{ROC_curve}
\end{figure*}
\begin{figure*}[htb]
\centering
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/fig1_raw_fig_both_side_10_subject.pdf}
\centering
\caption{Raw signals.}
\end{subfigure}\hfil
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/fig1_sudaembed_fig_both_side_10_subject.pdf}
\centering
\caption{Features learned by SUDA.}
\end{subfigure}\hfil
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/fig2_combine.pdf}
\centering
\caption{Raw signals.}
\end{subfigure}\hfil
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/fig2_dsudaembed1.pdf}
\centering
\caption{Features learned by $f_{pe}$ in DSUDA.}
\end{subfigure}\hfil
\caption{Visualization of SUDA and DSUDA. (a)-(b): Raw signals and features learned by SUDA of 10 subjects from target domain. Every subject has two signals, each for one ear. Signals belonging to the same subjected are marked in the same color. (c)-(d) Raw signals and disentangled domain-invariant features learned by DSUDA in target domain of 20 subjects. Tinnitus samples are in pink; control samples are in blue.}
\label{embedding}
\end{figure*}
\section{Experiments}
\subsection{Dataset and Experimental Setting}
We evaluate our model based on two Auditory brainstem responses (ABRs) datasets~\cite{schaette2011tinnitus,guest2017tinnitus} based on EEG sensors and consider the larger datasets~\cite{schaette2011tinnitus} as the source domain and the smaller one~\cite{guest2017tinnitus} as the target domain. ABRs are signals that originate from the early stages of the auditory pathway and evoked potentials recorded with EEG sensors~\cite{schaette2011tinnitus,guest2017tinnitus}. Schaette and McAlpine~\cite{schaette2011tinnitus} used presentation levels of 90 and 100 dB peSPL with an 11 click/s rate for the source domain approved by the University College London (UCL) ethics committee. The ethics ID number is 2039/002 and the declaration is the Helsinki declaration. Guest et al.~\cite{guest2017tinnitus} collected the target domain in response to 102 dB peSPL clicks presented with a rate of 7 clicks/s approved by the National Research Ethics Service Greater Manchester West Ethics Committee with REC reference 15/NW/0133 and IRAS project ID 168221. The signals are sampled with 50 kHz and bandpass-filtered before 100 Hz and 1500 Hz. The source dataset contains 408 EEG signals of 38 subjects recorded while an acoustic click train was presented over headphones to the affected ears for patients or either ear for control subjects, i.e., only one ear of each subject is recorded. Each trial lasts for 10 milliseconds resulting in 500-time points. The target dataset collects 80 EEG signals of 40 subjects with both of their ears recorded. The time duration is 8 milliseconds with 131-time points per trial.
This target dataset was approved by the University College London (UCL) ethics committee. The ethics ID number of our study is 2039/002. The declaration is the Helsinki declaration.
The 2017 study was approved by the National Research Ethics Service Greater Manchester West Ethics Committee, REC reference 15/NW/0133, IRAS project ID 168221.
We use Fully Connected (FC) Layers to build our model. $f_{pe}$ and $f_{ve}$ have the same structures consisting of 2 FC layers with Tanh as the activation function. Other model components also contain 2 FC layers while adopting Sigmoid as the activation function. During data preprocessing, we set the sliding step as 20; thus an original trial from the source domain can be sliced into 6 segments, which are further down-sampled. We set learning rates of DAE and SUDA as 1.0e-3 with different updating steps per batch, 10 and 1, respectively. We set the loss ratios $\eta$, $\alpha$, and $\beta$ as 1.0, 1.0, and 1.0e-6.
We choose 9 machine or deep learning methods as competitors: XGBoost~\cite{chen2016xgboost}, nu-SVC~\cite{v-svm}, nCSP~\cite{ncsp}, DeepNet \cite{DeepConvNet}, ShallowNet \cite{DeepConvNet}, AEXGB \cite{AEXGB}, EEGNet \cite{eegnet}, AE~\cite{sunilkumar2021bio}, and SAE~\cite{liu2021generalizable}. Unlike regular classification tasks, incorrect tinnitus diagnosis may cause severe consequences. Accuracy alone is not sufficient to evaluate the models. Therefore, we adopt Accuracy(Acc) along with Negative Predictive Value (NPV), True Negative Rate (TNR), Positive Predictive Value (PPV), True Positive Rate (TPR), F1-score for negative samples (N-F1), and F1-score for positive samples (P-F1) to evaluate the performance comprehensively.
\subsection{Experimental Results}
\subsubsection{Comparison with State-of-the-art Methods}
We present the results of our model and comparison in 3 different settings: both-side, left-side, and right-side, respectively, in Tables~(\ref{Main_exp_both}-\ref{Main_exp_right}). Among the compared methods, generative-based models, e.g., SAE, perform better than non-generative models, suggesting that the generative algorithms have a better generalization ability. DSUDA performs the best among all the algorithms, outperforming the second-best algorithm (including our variant) on N-F1, P-F1, accuracy by 3.5\%/7.1\%, 3.0\%/5.0\%, and 3.7\%/7.5\% in both-/left-side settings, respectively; 2.3\%, 3.0\%, and 5.0\% higher than the best comparison algorithm on the right-side setting, respectively. Our model shows better improvement in the left ear than in other settings, which indicates that the left ear presentations have more transferable knowledge on these two datasets than the right ear. UDA, SUDA, and DSUDA can improve performance on both left-side and both-side settings, indicating that domain adaptation, side-aware domain adaptation, and cross-dataset domain adaptation are effective in enhancing the model performance. On the right ear setting, the improvement on DSUDA is limited, which shows that the side-aware domain adaptation is more generalized and can fit more situations than the dataset domain adaptation.
\begin{table*}[h]
\centering
\caption{Parameter study of DSUDA on different updating steps of DAE.}
\label{Ablation_k2}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccc|ccccc|ccccc}
\toprule
\multirow{2}{*}{Updating Steps} & \multicolumn{5}{c|}{Both Sides} & \multicolumn{5}{c|}{Left Side} & \multicolumn{5}{c}{Right Side} \\\cmidrule{2-16}
& NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc \\
\midrule
1 & 0.725 & 0.725 & 0.725 & 0.725 & 0.725 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
2 & 0.730 & 0.675 & 0.698 & 0.750 & 0.713 & 0.750 & 0.600 & 0.667 & 0.800 & 0.700 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
3 & 0.681 & 0.800 & 0.758 & 0.625 & 0.713 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.654 & 0.850 & 0.786 & 0.550 & 0.700 \\
4 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.684 & 0.650 & 0.667 & 0.700 & 0.675 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
5 & 0.705 & 0.775 & 0.750 & 0.675 & 0.725 & 0.684 & 0.650 & 0.667 & 0.700 & 0.675 & 0.720 & 0.900 & 0.867 & 0.650 & 0.775 \\
6 & 0.750 & 0.675 & 0.705 & 0.775 & 0.725 & 0.800 & 0.600 & 0.680 & 0.850 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
7 & 0.707 & 0.725 & 0.718 & 0.700 & 0.713 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
8 & 0.707 & 0.725 & 0.718 & 0.700 & 0.713 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
9 & 0.732 & 0.750 & 0.744 & 0.725 & 0.738 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.727 & 0.800 & 0.778 & 0.700 & 0.750 \\
10 & 0.789 & 0.750 & 0.762 & 0.800 & 0.775 & 0.833 & 0.750 & 0.773 & 0.850 & 0.800 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[h]
\centering
\caption{Parameter study of DSUDA on different updating steps of SUDA.}
\label{Ablation_k1}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccc|ccccc|ccccc}
\toprule
\multirow{2}{*}{Updating Steps} & \multicolumn{5}{c|}{Both Sides} & \multicolumn{5}{c|}{Left Side} & \multicolumn{5}{c}{Right Side} \\\cmidrule{2-16}
& NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc \\
\midrule
1 & 0.725 & 0.725 & 0.725 & 0.725 & 0.725 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
2 & 0.673 & 0.825 & 0.774 & 0.600 & 0.713 & 0.667 & 0.800 & 0.750 & 0.600 & 0.700 & 0.680 & 0.850 & 0.800 & 0.600 & 0.725 \\
3 & 0.718 & 0.700 & 0.707 & 0.725 & 0.713 & 0.684 & 0.650 & 0.667 & 0.700 & 0.675 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 \\
4 & 0.718 & 0.700 & 0.707 & 0.725 & 0.713 & 0.684 & 0.650 & 0.667 & 0.700 & 0.675 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 \\
5 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.722 & 0.650 & 0.682 & 0.750 & 0.700 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 \\
6 & 0.743 & 0.650 & 0.689 & 0.775 & 0.713 & 0.765 & 0.650 & 0.696 & 0.800 & 0.725 & 0.722 & 0.650 & 0.682 & 0.750 & 0.700 \\
7 & 0.725 & 0.725 & 0.725 & 0.725 & 0.725 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
8 & 0.725 & 0.725 & 0.725 & 0.725 & 0.725 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
9 & 0.725 & 0.725 & 0.725 & 0.725 & 0.725 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 \\
10 & 0.711 & 0.800 & 0.771 & 0.675 & 0.738 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.692 & 0.900 & 0.857 & 0.600 & 0.750 \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[h]
\centering
\caption{Parameter study of DSUDA on different learning rates of DAE.}
\label{Ablation_lr2}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccc|ccccc|ccccc}
\toprule
\multirow{2}{*}{learning rate} & \multicolumn{5}{c|}{Both Sides} & \multicolumn{5}{c|}{Left Side} & \multicolumn{5}{c}{Right Side} \\\cmidrule{2-16}
& NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc \\
\midrule
6.0e-4 & 0.690 & 0.725 & 0.711 & 0.675 & 0.700 & 0.667 & 0.700 & 0.684 & 0.650 & 0.675 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
7.0e-4 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.652 & 0.750 & 0.706 & 0.600 & 0.675 \\
8.0e-4 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
9.0e-4 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1.0e-3 & 0.689 & 0.775 & 0.743 & 0.650 & 0.713 & 0.667 & 0.700 & 0.684 & 0.650 & 0.675 & 0.708 & 0.850 & 0.813 & 0.650 & 0.750 \\
1.1e-3 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1.2e-3 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1.3e-3 & 0.667 & 0.800 & 0.750 & 0.600 & 0.700 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 & 0.654 & 0.850 & 0.786 & 0.550 & 0.700 \\
1.4e-3 & 0.757 & 0.700 & 0.721 & 0.775 & 0.738 & 0.824 & 0.700 & 0.739 & 0.850 & 0.775 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 \\
1.5e-3 & 0.757 & 0.700 & 0.721 & 0.775 & 0.738 & 0.778 & 0.700 & 0.727 & 0.800 & 0.750 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[h]
\centering
\caption{Parameter study of DSUDA on different learning rates of SUDA.}
\label{Ablation_lr1}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccc|ccccc|ccccc}
\toprule
\multirow{2}{*}{learning rate} & \multicolumn{5}{c|}{Both Sides} & \multicolumn{5}{c|}{Left Side} & \multicolumn{5}{c}{Right Side} \\\cmidrule{2-16}
& NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc \\
\midrule
6.0e-4 & 0.680 & 0.850 & 0.800 & 0.600 & 0.725 & 0.708 & 0.850 & 0.813 & 0.650 & 0.750 & 0.654 & 0.850 & 0.786 & 0.550 & 0.700 \\
7.0e-4 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 & 0.667 & 0.700 & 0.684 & 0.650 & 0.675 & 0.696 & 0.800 & 0.765 & 0.650 & 0.725 \\
8.0e-4 & 0.718 & 0.700 & 0.707 & 0.725 & 0.713 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 \\
9.0e-4 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
1.0e-3 & 0.689 & 0.775 & 0.743 & 0.650 & 0.713 & 0.667 & 0.700 & 0.684 & 0.650 & 0.675 & 0.708 & 0.850 & 0.813 & 0.650 & 0.750 \\
1.1e-3 & 0.698 & 0.750 & 0.730 & 0.675 & 0.713 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1.2e-3 & 0.707 & 0.725 & 0.718 & 0.700 & 0.713 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
1.3e-3 & 0.673 & 0.875 & 0.821 & 0.575 & 0.725 & 0.680 & 0.850 & 0.800 & 0.600 & 0.725 & 0.667 & 0.900 & 0.846 & 0.550 & 0.725 \\
1.4e-3 & 0.725 & 0.725 & 0.725 & 0.725 & 0.725 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.750 & 0.750 & 0.750 & 0.750 & 0.750 \\
1.5e-3 & 0.711 & 0.675 & 0.690 & 0.725 & 0.700 & 0.722 & 0.650 & 0.682 & 0.750 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[h]
\centering
\caption{Parameter study of DSUDA on different $\alpha$.}
\label{Ablation_dratio}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccc|ccccc|ccccc}
\toprule
\multirow{2}{*}{ $\alpha$} & \multicolumn{5}{c|}{Both Sides} & \multicolumn{5}{c|}{Left Side} & \multicolumn{5}{c}{Right Side} \\\cmidrule{2-16}
& NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc \\
\midrule
0.01 & 0.673 & 0.875 & 0.821 & 0.575 & 0.725 & 0.654 & 0.850 & 0.786 & 0.550 & 0.700 & 0.692 & 0.900 & 0.857 & 0.600 & 0.750 \\
0.5 & 0.667 & 0.850 & 0.793 & 0.575 & 0.713 & 0.667 & 0.800 & 0.750 & 0.600 & 0.700 & 0.667 & 0.900 & 0.846 & 0.550 & 0.725 \\
0.1 & 0.732 & 0.750 & 0.744 & 0.725 & 0.738 & 0.789 & 0.750 & 0.762 & 0.800 & 0.775 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
0.3 & 0.689 & 0.775 & 0.743 & 0.650 & 0.713 & 0.667 & 0.700 & 0.684 & 0.650 & 0.675 & 0.708 & 0.850 & 0.813 & 0.650 & 0.750 \\
0.5 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.684 & 0.650 & 0.667 & 0.700 & 0.675 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
0.7 & 0.730 & 0.675 & 0.698 & 0.750 & 0.713 & 0.765 & 0.650 & 0.696 & 0.800 & 0.725 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 \\
0.9 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1 & 0.707 & 0.725 & 0.718 & 0.700 & 0.713 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[h]
\centering
\caption{Parameter study of DSUDA on different $\beta$.}
\label{Ablation_sae}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccc|ccccc|ccccc}
\toprule
\multirow{2}{*}{$\beta$} & \multicolumn{5}{c|}{Both Sides} & \multicolumn{5}{c|}{Left Side} & \multicolumn{5}{c}{Right Side} \\\cmidrule{2-16}
& NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc \\
\midrule
1.0e-6 & 0.698 & 0.750 & 0.730 & 0.675 & 0.713 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1.0e-5 & 0.667 & 0.800 & 0.750 & 0.600 & 0.700 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 & 0.654 & 0.850 & 0.786 & 0.550 & 0.700 \\
1.0e-4 & 0.707 & 0.725 & 0.718 & 0.700 & 0.713 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1.0e-3 & 0.707 & 0.725 & 0.718 & 0.700 & 0.713 & 0.700 & 0.700 & 0.700 & 0.700 & 0.700 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 \\
1.0e-2 & 0.667 & 0.800 & 0.750 & 0.600 & 0.700 & 0.652 & 0.750 & 0.706 & 0.600 & 0.675 & 0.680 & 0.850 & 0.800 & 0.600 & 0.725 \\
1.0e-1 & 0.643 & 0.675 & 0.658 & 0.625 & 0.650 & 0.650 & 0.650 & 0.650 & 0.650 & 0.650 & 0.636 & 0.700 & 0.667 & 0.600 & 0.650 \\
1.0 & 0.657 & 0.575 & 0.622 & 0.700 & 0.638 & 0.625 & 0.500 & 0.583 & 0.700 & 0.600 & 0.684 & 0.650 & 0.667 & 0.700 & 0.675 \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[h]
\centering
\caption{Parameter study of DSUDA on $\eta$.}
\label{Ablation_sdp}
\resizebox{\textwidth}{!}{%
\begin{tabular}{c|ccccc|ccccc|ccccc}
\toprule
\multirow{2}{*}{$\eta$} & \multicolumn{5}{c|}{Both Sides} & \multicolumn{5}{c|}{Left Side} & \multicolumn{5}{c}{Right Side} \\\cmidrule{2-16}
& NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc & NPV & TNR & PPV & TPR & Acc \\
\midrule
1.0e-6 & 0.681 & 0.800 & 0.758 & 0.625 & 0.713 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.654 & 0.850 & 0.786 & 0.550 & 0.700 \\
1.0e-5 & 0.707 & 0.725 & 0.718 & 0.700 & 0.713 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 \\
1.0e-4 & 0.696 & 0.800 & 0.765 & 0.650 & 0.725 & 0.714 & 0.750 & 0.737 & 0.700 & 0.725 & 0.680 & 0.850 & 0.800 & 0.600 & 0.725 \\
1.0e-3 & 0.711 & 0.675 & 0.690 & 0.725 & 0.700 & 0.765 & 0.650 & 0.696 & 0.800 & 0.725 & 0.667 & 0.700 & 0.684 & 0.650 & 0.675 \\
1.0e-2 & 0.667 & 0.850 & 0.793 & 0.575 & 0.713 & 0.680 & 0.850 & 0.800 & 0.600 & 0.725 & 0.654 & 0.850 & 0.786 & 0.550 & 0.700 \\
1.0e-1 & 0.673 & 0.825 & 0.774 & 0.600 & 0.713 & 0.682 & 0.750 & 0.722 & 0.650 & 0.700 & 0.667 & 0.900 & 0.846 & 0.550 & 0.725 \\
1.0 & 0.750 & 0.675 & 0.705 & 0.775 & 0.725 & 0.765 & 0.650 & 0.696 & 0.800 & 0.725 & 0.737 & 0.700 & 0.714 & 0.750 & 0.725 \\
\bottomrule
\end{tabular}
}
\end{table*}
\subsubsection{ROC Curve Analysis} In Figs.~\ref{ROC_curve} (a-c), we plot the ROC-AUC curves of DSUDA in both-ears, left-ear, and right-ear, respectively. When discriminating the left ear, DSUDA achieved the highest scores in all criteria, i.e., 0/1 classification, Micro- and Macro-average AUC scores of 0.73/0.73, 0.74 and 0.76, respectively. This may be because after aligning the data of the left ear to the data of the right ear, the right ear data may be too aligned and close to each other. Then, the data of the left ear will be easier to distinguish compared to the right ear. In all three settings, our model scores higher on correctly classifying tinnitus subjects than control subjects. This may be due to the more similar patterns of tinnitus patients across different datasets than those of control subjects. It also shows that our model can effectively learn the data patterns of tinnitus patients across datasets.
\subsection{Visualization of Learned Representations}
\subsubsection{Effectiveness of Side-aware Domain Adaptation} As shown in Figs.~\ref{embedding} (a-b), we visualize the raw signals and the representations in the target dataset after domain adaptation, respectively. Each subject contains two signals, i.e., one signal from the left and the right ear, which is the same color. It is obvious that the adapted representations are closer to each other than the raw features, e.g., dark blue and light green. This indicates that our side-aware domain adaptation effectively aligns signals from the left and right sides to a unified domain. SUDA can eliminate the side difference and make it easier for classifiers to detect whether the signals are from the same subject.
\subsubsection{Effectiveness of Disentangled Auto-encoder} We plot the raw signals and the pure features learned by $f_{pe}$ in DSUDA in Figs.~\ref{embedding} (c-d). The tinnitus subjects are colored pink and the control subjects are blue. We can observe that in the raw features, the blue and pink nodes are entangled and hard to distinguish. After disentangling the noise-free classification information, the pink nodes are more clustered together and further away from the blue node. It shows that our model can effectively exclude noise and purify the data.
\section{Discussion}
We set default experimental settings as 1) updating step per batch for DAE and SUDA: 1; 2) learning rates of DAE and SUDA: 1.0e-3; 3) loss ratios $\alpha$, $\beta$, $\eta$: 1.0. We do ablations studies to discuss how the updating steps, learning rates, and loss ratios can affect the performance. Each time, we fix other settings as the default setting, use a fixed random seed, and vary the factors we want to study.
\subsection{Effect of Updating Steps}
The results of the different updating steps for DAE and SUDA are shown in the tables ~\ref{Ablation_k2} and~\ref{Ablation_k1}, respectively. We can find that the performance of SUDA is stable at different updating steps and shows a slightly increasing trend as the number of updating steps increases. In contrast, the results of varying updating steps of DAE vary widely, and the best results are achieved when the DAE is updated 10 times per batch. The reason may be that the optimization of SUDA can only affect the classification indirectly, while DAE directly extracts features to classify the signal. So minor changes can lead to non-negligible performance differences.
\subsection{Effect of Learning rates}
We show results with varied learning rates of two model components in Tables~\ref{Ablation_lr2} and \ref{Ablation_lr1}. We can find that the model performance fluctuates up and down as the learning rate increases but is overall stable. Note that we set updating steps as the default value, i.e., 1, thus, according to Table~\ref{Ablation_k2}, the both-ear accuracy would be around 0.725, and the best accuracy (0.775) can not be achieved. This explains why results in ablation studies (Tables~\ref{Ablation_k1}-\ref{Ablation_sdp}) are significantly lower than the best results.
\subsection{Effect of loss ratios}
We summarize results with different loss ratios in Table~\ref{Ablation_dratio}-\ref{Ablation_sdp}. We set the varying ranges based on the values and importance of $L_{d}$ and $L_{side}$: 0.01 to 1 for $\alpha$ ($L_{d}$ ); 1.0e-6 to 1 for $\beta$ and $\eta$ ($L_{side}$). Overall, the performance is stable when changing $\alpha$ and $\eta$, but decreases drastically when $\beta$ is larger than 0.1. The reason lies in that during optimizing DAE, we mainly focusing on representation disentangling, and higher ratios of $L_{side}$ may misguide the optimization direction.
\section{Conclusion and Future Direction}
\subsection{Conclusion}
We propose a novel model, dubbed DSUDA, for EEG-based cross-datasets tinnitus diagnosis. DSUDA integrates disentangled auto-encoder and side-aware unsupervised domain adaptation to adapt the classifier trained on a source domain to the unlabeled target domain. The model can overcome domain discrepancy with domain variance adaptation, overcome inherent left-right ear pattern difference by subject-level side adaptation, and relieve the negative influence of EEG's low signal-to-noise ratio by disentangling pure information from noises. Our model consistently outperforms state-of-the-art competitors with a large margin regarding classification accuracy and other criteria, demonstrating that the side-aware unsupervised domain adaptation facilitates the model's generalization and the disentangled auto-encoder improves classifying performance.
\subsection{Future Direction}
Our model can be further extended to new scenarios in the future. There are several promising directions. For example, we can apply the DSUDA in other EEG-based tasks, such as activity or emotion recognition. The problems we addressed with our model, i.e., distribution discrepancy and low signal-to-noise ratio, are universal in EEG-based tasks. Thus, our model can achieve good performance after being modified to fit other tasks. The potential of DSUDA is worth to dig. Or, we can implement DSUDA in real-world clinics or hospitals to assist doctors or therapists in diagnosing and treat tinnitus, which would be of great value to the community and healthcare. Also, we can improve DSUDA with interpretable components to help researchers in medical or bioinformatics reveal latent but essential findings, which may facilitate the cure of tinnitus. Taking other factors into consideration is another direction to improve our model. Although DSUDA already eliminates the influence of ear-side information, many other factors like ages or genders may also change the EEG signal patterns and thus impair classification accuracy.
\section{Acknowledgment}
The authors would like to thank Dr. Hannah Guest and Dr. Roland Schaette for sharing the datasets.
\bibliographystyle{IEEEtran}
|
\section*{Acknowledgements}
\newpage
\section{Introduction}
The intensifying natural disasters and emergence of new threats, such as cyber-attacks and pandemics, have led to a paradigm shift towards infrastructure system resilience. The physical and functional risks posed by such extreme events to our infrastructure systems are compounded by climate change, the drastic modifications to the built environment, and the growing interdependence among urban systems. Therefore, the infrastructure systems are no longer designed for operational efficiency alone; equal emphasis is given to their ability to withstand disasters and minimize the resultant societal and economic impacts from unanticipated service disruptions \citep{Gay2013}. Disaster mitigation and resilience enhancement alternatives have become crucial aspects considered in the design and management of existing infrastructure systems.
Several approaches exist to analyze the resilience capabilities of interdependent infrastructure systems and evaluate the resilience project alternatives. These approaches can be broadly classified into two, namely empirical- and computational approaches \citep{Mitsova2021}. Empirical approaches rely on datasets, records, and reports based on historical events to identify patterns and severity in physical and functional disruptions to infrastructure systems \citep{McDaniels2007, Luiijf2009}. Data collected during historical breakdown events are used to characterize interrelationships among different infrastructure systems. On the other hand, computational approaches attempt to replicate the physical-, cyber-, geographic-, and logical dependencies among infrastructure components using mathematical and logical functions. Several models, such as, network theory-based models \citep{Holden2013,Praks2017}, system dynamics models \citep{Pasqualini2005,Powell2008}, agent-based models \citep{Tesfatsion2003,Thompson2019} and input-output models \citep{Haimes2005,Oliva2010}, have been extensively used for modeling interdependent infrastructure systems and analyzing their resilience. More advanced computational models adopt co-simulation of multiple domain-specific infrastructure simulators instead of a homogeneous method to replicate the collective behavior of interdependent infrastructure systems \citep{Wang2022}.
Recent computational approaches have been increasingly adopting simulation-based machine learning (ML) models to predict the network-wide impacts and use these predictions for the analysis of interdependent infrastructure resilience \citep{Golkhandan2021}. Most of the ML models for infrastructure resilience analysis focus on the optimal allocation of resources for improving the absorptive and recovery capabilities of infrastructure systems. \cite{Alemzadeh2020} developed Artificial Neural Network (ANN) models to approximate post-disaster repair sequences for optimal recovery of interdependent infrastructure systems by training simulated data. \cite{Sun2020} applied a Deep Q-learning (DQN) algorithm on an interdependent power-water-transport model to predict the optimal repair crew allocation to flood-affected bridges that would minimize the cumulative impact on the network. \cite{Dehghani2021} applied Deep Reinforcement Learning (DRL) to identify the optimal long-term preventive maintenance strategy that maximizes the resilience of power systems.
Although ML models are powerful tools to accurately predict infrastructure system resilience (i.e., network-wide impacts of different disruptive events), they require large training datasets to ensure adequate prediction accuracy. Most of the aforementioned ML models learn the vulnerability and resilience attributes at a component level, leading to a large number of features in the model, also known as the `\textit{curse of dimensionality}' \citep{Turati2016}. When the dimensionality of a problem increases, generalizing the trends becomes harder because the training dataset of fixed size can cover only a small fraction of the possible input combinations \citep{Domingos2012}. While using a larger simulation dataset could be a potential solution to the problem of high-dimensionality, several advanced infrastructure simulation models are computationally intensive and may require considerable time for simulating even a small number of disaster scenarios \citep{Liu2019, Zou2021}. The consequences of high dimensionality could be even more severe when the system response of large-scale interdependent infrastructure networks is to be learned by the ML algorithms.
To mitigate the negative effects of high dimensionality, our study exploits clustering methods to identify similar infrastructure components in terms of their topological and functional properties and later incorporates that information to enhance the accuracy of resilience prediction models. Several studies in the literature have demonstrated that topological and functional attributes of components influence the infrastructure network vulnerability and resilience characteristics \citep{Cadini2009,Nicholson2016,Balakrishnan2020}.
The overall objective of this paper is to propose a novel network clustering-based approach for achieving dimensionality reduction in simulation-based machine learning models for infrastructure network analysis. The specific objectives are as follows:
\begin{enumerate}
\item Demonstrate how the relationship between network topology and infrastructure resilience, as established in the literature, can be leveraged to incorporate network structure characteristics in ML prediction models.
\item Present a methodological framework to apply network clustering to reduce dimensionality in infrastructure resilience prediction models.
\item Propose unsupervised and supervised methods to find the optimal number of clusters in each infrastructure system in an interdependent network.
\end{enumerate}
The rest of the paper is organized as follows: Section~\ref{sec:method} presents the methodological framework adopted in this paper; Section~\ref{sec:case_study} demonstrates the application of the methodology on a synthetic interdependent power-water-transport network; and Section~\ref{sec:conclusions} summarizes the findings and discusses the scope for further research.
\section{Methodology}
\label{sec:method}
The methodological framework adopted in the study is outlined in Figure~\ref{fig:methodology}.
\begin{sidewaysfigure}[htbp]
\centering
\includegraphics[width = 0.85\textwidth]{ML_methodology_full.pdf}
\caption{Framework for developing ML-based infrastructure resilience models.}
\label{fig:methodology}
\end{sidewaysfigure}
The methodology is implemented in three steps, namely, (a) generation of simulation dataset, (b) network component clustering, and (c) dimensionality reduction and development of prediction models. In the first step, we employ an interdependent infrastructure model to simulate the infrastructure disruption data required for developing the prediction model. The second step is the application of appropriate clustering algorithms to categorize infrastructure components based on their topological and functional characteristics. Finally, we develop ML models for predicting the network-wide impacts using disaster and recovery-related features. To summarize, we propose algorithms that combine network component clustering with existing ML prediction algorithms to reduce dimensionality and improve model prediction simultaneously. In the rest of the section, we discuss the above steps in detail.
\subsection{Generation of simulation dataset}
This step derives the relevant target and predictor features using an interdependent infrastructure simulation model. The step further constitutes two subtasks: interdependent infrastructure simulation and feature extraction.
\subsubsection{Interdependent infrastructure simulation}
In this study, we employed \textit{InfraRisk}, an open-source Python-based integrated simulation platform, for simulating infrastructure disruptions and subsequent recovery in interdependent power-water-transport networks \citep{Balakrishnan2021}. \textit{InfraRisk} integrates existing domain-specific infrastructure simulators (\textit{pandapower} for power systems, \textit{wntr} for water distribution systems, and a static traffic assignment package for transport network) via an object-oriented interface to perform the interdependent simulations. All the three infrastructure simulators assign flows (power, water, or traffic) in the respective networks by minimizing the total loss subject to network-specific constraints and dependencies with other infrastructure systems.
Consider an interdependent infrastructure network $\mathbb{K}$ with constituent infrastructure systems $K:K \in \mathbb{K}$. An infrastructure system $K$ can be represented as a graph $G(V_{K}, E_{K})$, where $V_{K}$ is the set of nodes and $E_{K}$ the set of links connecting the nodes. In addition, $\mathbb{K}$ also includes the dependencies between infrastructure systems, which are represented as links. We denote the set of consumers who are dependent on $\mathbb{K}$ by $N$, and the resource supply from $K$ to each consumer $n\in N$ at simulation time $t \in T$ under normal operating conditions is represented by $S_{n}^{K}(t)\in \mathbf{S}^{\mathbb{K}\times T}$.
To implement the simulation, we need to define the disaster scenario and the recovery strategy for infrastructure system $K$. We can define $h:h\in H$ as the disaster scenario which results in the failure of infrastructure nodes $V_{K}^{h}\subseteq V_{K}$ and links $E_{K}^{h}\subseteq E_{K}$.
On the other hand, the recovery strategies $P_{K}^{h}$ defines the schedule of repair actions (repair sequence and start time). Normally, each system has its own set of repair crews and restoration strategies to expedite recovery actions.
The consequences of initial infrastructure disruptions due to the hazard, the network-wide functional disruptions due to infrastructure interdependencies, and the subsequent recovery efforts are reflected in the resource supplied to the consumers during the disrupted conditions $\left (s_{n}^{K}(t)\in \mathbf{s}^{\mathbb{K}\times T}\right )$ and is captured by \textit{InfraRisk}. To summarize, the interdependent infrastructure model can be represented using Equation~\ref{eq:black_box}.
\begin{equation}
\mathbf{s}^{\mathbb{K}\times T} = \mathbf{\Gamma}\left(\mathbb{K}, h,\mathbb{P}^h\right),
\label{eq:black_box}
\end{equation}
where $\mathbf{\Gamma}(\cdot)$ is the simulation model and $\mathbb{P}^h = \{P_K^{h}: K\in \mathbb{K}\}$.
\subsubsection{Feature extraction}
The simulation model provides three types of data required to develop the prediction models.
\begin{enumerate}[i.]
\item Timeline of consumer-level and system-level functional performance under normal and disrupted network conditions.
\item Topological characteristics of the infrastructure systems.
\item System disruption and recovery characteristics, such as the list of initially failed components and the repair sequencing strategy used.
\end{enumerate}
Since the objective of the ML model is to predict the network-level resilience, we converted the consumer-level resource supply values generated by the simulation model to network-level resilience metrics. In this study, we used Prioritized Consumer Serviceability (PCS) as the measure of performance (MOP) for tracking the performance of the networks \citep{Balakrishnan2021}, defined as
\begin{equation}
\text{PCS}_{K}(t) = \left( \frac{\sum_{\forall i: S_{n}^{K}(t) > 0} s_{n}^{K}(t)}{\sum_{\forall i: S_{n}^{K}(t)> 0}S_{n}^{K}(t)} \right), \text{where~} 0 \leq s_{n}^{K}(t) \leq S_{n}^{K}(t).
\label{eq:pcs}
\end{equation}
The resource supply under normal operating conditions $S_{n}^{K}(t)\in \mathbf{S}^{\mathbb{K}\times T}$ is computed by performing the interdependent infrastructure simulation without failing any infrastructure components ($\mathbf{S}^{\mathbb{K}\times T} = \mathbf{\Gamma}(\mathbb{K})$).
The resilience of infrastructure systems is quantified using the concept of equivalent outage hours (EOH). The EOH corresponding to each infrastructure network for a given disaster scenario $\gamma_{K}$ is given by Equation~\ref{eq:system_eoh}.
\begin{equation}
\gamma_{K} = \frac{1}{3600}\int_{t_{0}}^{t_{max}} \left [ 1 - PCS_{K}(t)\right ]dt,
\label{eq:system_eoh}
\end{equation}
where $t_{0}$ is the time of occurrence of the disaster event in the simulation and $t_{max}$ is the maximum simulation time (both in seconds). Mathematically, $[1-PCS_{K}(t)]$ is the mean unmet demand (slack demand) in $K$ at time $t$ and $\gamma_{K}$ is the area of the resilience triangle \citep{Bruneau2003} formed by the PCS curve. The unit of $\gamma_{K}$ is system performance-hours.
The target variable for the ML model is the resilience metric of the interdependent infrastructure network, computed as the weighted equivalent outage hours ($\overline{\gamma}$) of individual infrastructure systems (Equation~\ref{eq:wEOH}).
\begin{equation}
\overline{\gamma} = \sum_{K\in \mathbb{K}}w_{K}\gamma_{K}
\label{eq:wEOH}
\end{equation}
where $w_{K}$ is the weight assigned to system $K$.
In addition to the model inputs ($V_{K}^{h}$, $E_{K}^{h}$, and $P_K^{h}$) and outputs ($\mathbf{s}^{\mathbb{K}\times T}$, $\mathbf{S}^{\mathbb{K}\times T}$, and $\overline{\gamma}$), we also extracted the topological features of infrastructure components, such as, centrality values, from the infrastructure network to aid network clustering.
\subsection{Infrastructure component clustering}
Clustering is an unsupervised learning method that attempts to identify the most natural way of partitioning a dataset based on similarity and dissimilarity among observations \citep{Xu2015}. Clustering could reveal the underlying structure of the dataset, which could be used to build supervised learning models with simple features and a better prediction accuracy \citep{Trivedi2015}. In this study, clustering is proposed to identify infrastructure components that are similar in their vulnerability and resilience characteristics. For this purpose, we investigate the use of topological and functional features of infrastructure components that are identified as indicators of infrastructure vulnerability, criticality, and resilience in the literature.
Figure~\ref{fig:cluster_concept} illustrates the clustering concept introduced in the study for partitioning infrastructure components. Any infrastructure network can be treated as a set of nodes and links, where nodes represent producers, consumers, or intermediate transfer points, whereas the links denote the connections and interdependencies among the infrastructure nodes \citep{Svendsen2007, Svendsen2008}. Therefore, each component in the system can be assigned a group (or cluster) based on its topological and functional properties (color-coded in Figure~\ref{fig:cluster_concept}).
\begin{figure}[htbp]
\centering
\begin{subfigure}[h]{0.475\textwidth}
\centering
\includegraphics[width=\textwidth, trim = {0.05cm 0.05cm 0cm 0cm}, clip]{original_network2.pdf}
\caption{Original network with no clustering}
\end{subfigure}%
\begin{subfigure}[h]{0.475\textwidth}
\centering
\includegraphics[width = \textwidth, trim = 0.05cm 0.05cm 0cm 0cm, clip]{clustered_network2.pdf}
\caption{Clustered network}
\end{subfigure}
\caption{Sample illustration of clustering infrastructure network components}
\label{fig:cluster_concept}
\end{figure}
A number of algorithms are available for performing clustering of datasets, such as, K-Means, K-Medoids, agglomerative propogation, and DBSCAN, depending upon the types of the datasets. For systematic reviews of clustering algorithms and their applications, see \cite{Celebi2015} and \cite{Rodriguez2019}.
Consider the infrastructure system $K\in\mathbb{K}$ which can be represented as a graph $G(V_{K},E_{K})$. Without clustering, we can incorporate a characteristic associated with the infrastructure components (for example, post-disaster functional status) in two different ways as follows:
\begin{enumerate}
\item One feature for each component so that there will be $|V_{K}| + |E_{K}|$ additional features corresponding to each infrastructure system $K$ in the ML prediction model.
\item A single feature which aggregates the component-level values using mathematical operations, such as summation, count, average, maximum, and minimum.
\end{enumerate}
The above two approaches have their own advantages and limitations. While the first approach may effectively learn from the spatial and structural aspects of the characteristic to be considered and ensure a high level of prediction accuracy, it leads to the issue of high-dimensionality. For a fully connected graph with $m$ nodes, $m + m(m-1)/2$ features for every component-level characteristic need to be constructed in this approach if both node- and link-level information are to be used in the ML model. In the second approach, the number of features could be considerably lower than in the first approach; however, aggregation would lead to the loss of useful information related to components, leading to a low prediction accuracy. The clustering approach aims at finding a middle ground between the above two extreme approaches. By employing appropriate clustering algorithms, we can identify $\ell_{K}^{V}:1\leq \ell_{K}^{V}\leq |V_{K}|$ node clusters and $\ell_{K}^{E}:1\leq \ell_{K}^{E}\leq |E_{K}|$ link clusters with similar functional and topological properties in the infrastructure network $K$.
Several studies have demonstrated that centrality measures \citep{Cadini2009,Dunn2013,Balakrishnan2020}, such as degree centrality, betweenness centrality, and eigenvector centrality, could capture the vulnerability and resilience of infrastructure network components. In addition, many studies have shown that specific functional properties, such as flow rates under normal operating conditions, could serve as indicators of the vulnerability and importance of infrastructure components in a system \citep{Nicholson2016}. To incorporate node-specific and link-specific characteristics, clustering of nodes and links are done separately. Table~\ref{tab:topo_funcs} enlists the potential topological and functional features that are considered in the current study for network component clustering.
\begin{sidewaystable}
\begin{threeparttable}[htbp]\small
\caption{Topological and functional features used for infrastructure component clustering}
\begin{tabular}{p{4cm}p{5.5cm}p{8cm}}
\toprule
Feature & Node component $(i)$ & Link component $(i,j)$ \\
\midrule
Degree centrality\tnote{1} &
$\begin{aligned}
& C_{d}(i)=\sum_{v \in V}a_{i,v}
\end{aligned}$ &
$\begin{aligned}
& C_{d}(i)=\sum_{v \in V}a_{i,v}; &
& C_{d}(j)=\sum_{v \in V}a_{j,v}
\end{aligned}$\\
Betweenness centrality\tnote{2} &
$\begin{aligned}
C_{b}(i) = \frac{\sigma_{st}(i)}{\sigma_{st}}
\end{aligned}$ &
$\begin{aligned}
& C_{b}(ij) = \frac{\sigma_{st}(i,j)}{\sigma_{st}}
\end{aligned}$\\
Eigenvector centrality\tnote{3} &
$\begin{aligned}
& C_{e}(i) = \frac{1}{\lambda}\sum_{v \in V}a_{i,v}C_{e}(v)
\end{aligned}$ &
$\begin{aligned}
& C_{e}(i) = \frac{1}{\lambda}\sum_{v \in V}a_{i,v}C_{e}(v);
& C_{e}(j) = \frac{1}{\lambda}\sum_{v \in V}a_{j,v}C_{e}(v)
\end{aligned}$\\
Closeness centrality\tnote{4} &
$\begin{aligned}
& C_{c}(i) = \frac{|V|-1}{\sum_{v \in V}d(i,v)}
\end{aligned}$ &
$\begin{aligned}
& C_{c}(i) = \frac{|V|-1}{\sum_{v \in V}d(i,v)}; &
& C_{c}(j) = \frac{|V|-1}{\sum_{v \in V}d(j,v)}
\end{aligned}$\\
\midrule
Flow-rate\tnote{5} &
$\begin{aligned}
& Q_{i} = \sum_{j\in M(i)}\text{abs}(q_{ji})
\end{aligned}$ &
$\begin{aligned}
& Q_{ij} = \text{abs}(q_{ij})
\end{aligned}$\\
Weighted flow-rate &
$\begin{aligned}
& \bar{Q}_{i} = \sum_{j\in M(i)}C_{b}(ij)\text{abs}(q_{ji})
\end{aligned}$ &
$\begin{aligned}
& \bar{Q}_{ij} = C_{b}(ij)\text{abs}(q_{ij})
\end{aligned}$\\
\bottomrule
\end{tabular}
\label{tab:topo_funcs}
\begin{tablenotes}\footnotesize
\item[1] $i,j,v \in V$ are nodes in network $G$. $(i,j)\in E$ are links. $a_{i,n}=1~\text{if}~(i,n)\in E~\text{else}~0$.
\item[2] $\sigma_{st}$ is the number of all shortest paths in $G$. $\sigma_{st}(i)$ and $\sigma_{st}(i,j)$ are number of shortest paths passing through $i$ and $(i,j)$, respectively.
\item[3] $\lambda$ is a constant equal to the largest positive element in the eigenvector.
\item[4] $d(i,v)$ is the shortest distance between $i$ and $v$ nodes if a path exists between them.
\item[5] $q_{ji}$ is the maximum daily flow-rate from $j$ to $i$ during normal operation. $M(i)$ is the set of neighbor nodes of $i$.
\end{tablenotes}
\end{threeparttable}
\end{sidewaystable}
\subsection{Dimensionality reduction and development of prediction models}
We implement dimensionality reduction using an iterative clustering algorithm introduced in this study. The iterative clustering algorithm combines clustering methods with regression algorithms to produce concise infrastructure resilience prediction models. This step consists of two subtasks: construction of cluster-level features and development of ML models.
\subsubsection{Construction of cluster-level features from component-level features}
Once the infrastructure components are categorized into different clusters, the next step is to derive the cluster-level features from the simulation dataset. In the dataset generated using \textit{InfraRisk}, the component-level information to be incorporated in the ML model is their initial functional states (disrupted or operational) after the occurrence of a disaster event.
If a node cluster and a link cluster for infrastructure system $K$ generated by the clustering algorithm are denoted by $L_{K}^{v}:L_{K}^{v}\subseteq V_{K}$ and $L_{K}^{e}:L_{K}^{e} \subseteq E_{K}$, respectively, then the cluster-level features corresponding to the initial functional states of infrastructure components are derived as in Equation~\ref{eq:cluster_status}.
\begin{equation}
\delta_{i}^{L} =
\begin{cases}
1 & \text{if}~i\in L, \\
0 & \text{otherwise.}
\end{cases}\\
\label{eq:cluster_status}
\end{equation}
where $\delta_{i}^{L}$ is the indicator of whether a component $i$ belongs to the cluster $L$ or not.
Let $F^h$ denotes the set of cluster-level feature values corresponding to a topological or functional characteristic of the interdependent infrastructure network $\mathbb{K}$ and hazard $h$. The cluster-level feature $f_{L}^{h}:f_{L}^{h} \in F^{h}$ representing the initial disaster impact in the cluster $L$ is computed as the total number of failed components that belong to it (Equation~\ref{eq:cluster_feature}).
\begin{equation}
f_{L}^{h} = \sum_{\forall i\in \left\{V_{K}^{h},E_{K}^{h}: K\in \mathbb{K}\right\}}\delta_{i}^{L}.
\label{eq:cluster_feature}
\end{equation}
\subsubsection{Development of ML prediction models}
The last step of our method is to develop the prediction models by employing state-of-the-art ML algorithms. Since the target variable (weighted EOH) in this study is continuous, we consider only regression algorithms. Commonly used regression algorithms include multiple linear regression (MLR), support vector regression (SVR), decision tree (DT), Random Forest (RF), and neural network regression (NN). For detailed discussions and applications of the major ML regression algorithms, we refer the readers to \cite{Berk2008, James2013}.
Finally, we build the ML models using cluster-level features and recovery strategy as the predictors (Equation~\ref{eq:ml_model}).
\begin{equation}
\overline{\gamma} = \Phi\left ( \mathbb{F}, \mathbb{P} \right )
\label{eq:ml_model}
\end{equation}
where $\mathbb{F} = \{F^{h}:h\in H\}$, $\mathbb{P} = \{\mathbb{P}_{h}:h\in H\}$, and $\Phi(\cdot)$ is the ML algorithm. To obtain robust and accurate models, cross-validation and hyper-parameter tuning are performed.
Two performance metrics are used to evaluate the goodness-of-fit of the ML models. The first metric is the coefficient of determination ($R^{2}$), which is defined as the proportion of the variation in the target variable captured by the prediction model (Equation~\ref{eq:R-squared}).
\begin{equation}
R^{2} = 1 - \frac{\sum_{i}(y_{i} - \hat{y}_{i})^{2}}{\sum_{i}(y_{i} - \bar{y})^{2}}
\label{eq:R-squared}
\end{equation}
where $y_{i}$ is the $i$th observed value, $\hat{y_{i}}$ is the $i$th predicted value, and $\bar{y}$ is the mean observed value.
The second performance metric is the root mean square error ($RMSE$), which is the standard deviation of the prediction errors in the model (Equation~\ref{eq:RMSE}).
\begin{equation}
RMSE = \sqrt{\sum_{i}(y_{i} - \hat{y}_{i})^{2}}
\label{eq:RMSE}
\end{equation}
The goal is to have a high $R^2$ and low $RMSE$ values. The unit of $RMSE$ in the this study is system performance-hours.
A major aspect that is not yet resolved in the method is determining the optimal number of clusters in each infrastructure system for building the final ML model. Since an increase in the number of clusters would enhance the capability of the model to capture the spatial and network structure characteristics, an improvement in the quality of model prediction is expected. The optimal cluster count in each infrastructure system is determined using the {\em elbow} method \citep{Yuan2019} and our proposed iterative clustering method.
The elbow method is an unsupervised method in which the sum of squared distances between observations and the centroids of the clusters they belong to is used as the performance measure for evaluating the consistency of clusters. The `elbow' of the curve connecting the sum of squared distances and total cluster count is determined, and the corresponding number of clusters in each infrastructure system is identified.
The second method proposed in this study to identify the optimal number of clusters in each infrastructure network is an iterative clustering algorithm (Figure~\ref{fig:iter_clust_algo}). It is a supervised method in which the performance of the ML model on test dataset is used as the performance measure. In this method, setting the ML model corresponding to the elbow method as the base model, we subsequently increased or decreased an equal number of clusters in all infrastructure systems to develop additional ML models. Once we had an adequate number of cluster combinations and corresponding ML models, we evaluated the improvement to the model goodness-of-fit ($R^{2}$) due to an increase in the number of clusters, and then adopt the most efficient model, i.e., ML model with highest $R^2$.
\begin{figure}[t!]
\centering
\includegraphics[width = .8\textwidth, trim = {0 0cm 0 0cm}, clip]{Iter_clust_algo2.pdf}
\caption{Iterative clustering method}
\label{fig:iter_clust_algo}
\end{figure}
Consider $\{l_{K}: K\in\mathbb{K}\}$ are the cluster counts corresponding to the infrastructure systems in the interdependent network. If the cluster counts determined by the elbow method are $\{l_{K}^{e}: K\in\mathbb{K}\}$, the maximum number of clusters that can be removed simultaneously from each network is $b^{(-)} =min\{l_{K}^{e}-1: K\in\mathbb{K}\}$, and the maximum number of clusters that can be added is $b^{(+)} = min\{|V_{K}| + |E_{K}| - l_{K}^{e}:K\in\mathbb{K}\}$. In the iterative clustering algorithm, we start by building the initial model with cluster counts $\{l_{K}^{e} - b^{(-)}: K\in\mathbb{K}\}$ and evaluating the model train and test performance metrics. In the subsequent iterations, we repeat the procedure after updating the cluster counts $\{l_{K}^{e} + b: K\in\mathbb{K}\}$, where $b \in \{-b^{(-)},\dots,b^{(+)}\}$. Finally, the optimal cluster counts in infrastructure networks are obtained by finding the `knee' of the curve between the test dataset $R^{2}$ and the total cluster count. We employ the {\em kneedle} algorithm \citep{Satopaa2011} for this purpose.
\section{Case Study}
\label{sec:case_study}
The proposed methodology is implemented to develop a resilience prediction model for the Micropolis interdependent infrastructure network. Micropolis is a virtual city designed for 5000 inhabitants with water, power, and road networks \citep{Brumbelow2007}. We used hazard module in \textit{InfraRisk} to generate synthetic flood events and fail infrastructure components randomly based on the disaster intensity. A total of 325 flood scenarios are generated, assuming infrastructure components closer to the Micropolis stream are more likely to fail from a hazard (Figure~\ref{fig:micropolis_floods}). Each disaster scenario results in the failure of a specific set of infrastructure components in the interdependent infrastructure network. For this case study, we considered only water mains, power lines, and road links for failure as they are the most critical to the functioning of the respective infrastructure systems. It is found that 52 water mains, 22 power lines, and 17 road links along the water stream are either located or traversing through the regions exposed to the simulated floods. We limited the maximum number of failures in each flood scenario to 35 components to reduce the computational effort required for the case study. The disruptions to water links are modeled as leaks/pipe breaks, whereas that of power lines and road links are modeled by isolating them from the network.
\begin{figure}[htbp]
\centering
\includegraphics[width = 0.85\textwidth, trim = {0 0 0cm 0cm}, clip]{direct_impact.png}
\caption{Micropolis network exposed to floods}
\label{fig:micropolis_floods}
\end{figure}
We assumed that each failed pipeline is remotely isolated by a predefined set of shutoff valves 10 minutes after the disaster occurred (considering sensing and actuation times). By isolating the leaking pipelines, the loss of water is minimized; however, isolating some segments of the water system would cutoff consumers located within the isolated regions. Once a water pipe is repaired, the corresponding isolating valves are opened, conditional upon whether that would interfere with the remaining repair actions. Similarly, when a power line is fully repaired, the corresponding circuit breakers are closed to allow electric power to flow through the line. In the case of damaged road links, each link is added back to the network after repair, and then the traffic assignment model recomputes the traffic flows.
In this case study, each infrastructure system is assigned a repair crew for performing the post-disaster recovery. For implementing network recovery by component repair, three repair strategies are considered as follows:
\begin{itemize}
\item Betweenness centrality-based: Those components with a higher value of betweenness centrality are repaired first.
\item Maximum flow-based: Those components that handle larger resource flow rates are repaired first.
\item Zone-based: Components are repaired based on the zone in which they are located. The zones are prioritized in the order of central business district, industrial, and residential areas.
\end{itemize}
The recovery model in \textit{InfraRisk} also takes the accessibility to disrupted components into consideration and dynamically modifies repair sequences during the simulation.
Figure~\ref{fig:scenario_data} presents the simulation results corresponding to one of the 325 simulated flood events. The flood event resulted in the failure of 14 water mains, six power lines, and four road links (Figure~\ref{fig:disrupted_compons}).
\begin{figure}[b!]
\centering
\begin{subfigure}[h]{0.65\textwidth}
\centering
\fbox{\includegraphics[width = \textwidth]{disrupted_compons.png}}
\caption{Physically disrupted components and crew locations}
\label{fig:disrupted_compons}
\end{subfigure}
\begin{subfigure}[h]{0.65\textwidth}
\centering
\includegraphics[width=\textwidth]{pcs_plot.pdf}
\caption{System recovery curves}
\label{fig:pcs_plot}
\end{subfigure}
\caption{Simulation results from a simulated flood event}
\label{fig:scenario_data}
\end{figure}
Figure~\ref{fig:pcs_plot} shows the performance of the Micropolis water and power network during the flood event when the capacity-based strategy is chosen for the network recovery. The infrastructure system performance is measured using the MOP in Equation~\ref{eq:pcs}. Both performance curves follow the typical resilience triangle used to characterize system resilience. The power system is restored to the pre-disaster state in approximately 33 hours, whereas the water crew takes approximately 68 hours to complete all the repair actions. The road network is fully restored in approximately 51 hours. Using Equation~\ref{eq:system_eoh}, the equivalent outage duration (in hours) corresponding to the disaster scenario in the water network is estimated to be 17.55 system performance-hours, and that in the power network is 2.92 system performance-hours.
The consumer-level outages in power- and water utility services during the flood event are shown in Figure~\ref{fig:consumer_data}.
\begin{figure}[b!]
\centering
\begin{subfigure}[h]{0.65\textwidth}
\centering
\fbox{\includegraphics[width=\textwidth]{water_eoh.png}}
\caption{Consumer-level equivalent water outage hours}
\label{fig:water_eoh}
\end{subfigure}
\begin{subfigure}[h]{0.65\textwidth}
\centering
\fbox{\includegraphics[width = \textwidth]{power_eoh.png}}
\caption{Consumer-level equivalent power outage hours}
\label{fig:power_eoh}
\end{subfigure}
\caption{Consumer-level impacts due to the simulated flood event}
\label{fig:consumer_data}
\end{figure}
Even though the consumers in the adjacent areas along the flooded stream are affected by water outages the most, consumers in other parts of the network (especially those in the western and central Micropolis) are also affected by water outages (Figure~\ref{fig:water_eoh}). The leakage through failed pipes resulted in a reduced water head in the tank. The closure of shutoff valves isolated many consumers in other regions even though they were not directly affected by the flood event. On the other hand, the power outage is less severe compared to the water outage and is limited to the western side of the flooded stream (Figure~\ref{fig:power_eoh}). The drop in resilience values is mainly attributed to the consumers in downstream of failed power lines or opened circuit breakers who are disconnected from the rest of the power network.
Next, the network resilience metrics (weighted equivalent outage hours) corresponding to all the disaster scenarios are calculated by assigning equal weights of 0.5 in Equation~\ref{eq:wEOH}. Only water and power systems are considered for evaluating the network resilience. Figure~\ref{fig:simulation_data} presents the distribution of the resilience metrics and their relationship with the recovery strategy adopted and the number of physically disrupted components. The results show a positive correlation between the weighted equivalent outage hours and the number of initially failed components.
\begin{figure}[htbp]
\centering
\includegraphics[width = 0.65\textwidth, trim = {0 0 0cm 0cm}, clip]{auc_scatter_plot.pdf}
\caption{Weighted equivalent outage hours versus failure count}
\label{fig:simulation_data}
\end{figure}
Next, ML models to predict the resilience (in terms of weighted equivalent outage hours) are developed as in Equation~\ref{eq:ml_model}. We used the Random Forest algorithm for this case study because of its simplicity and robustness. The ML models are built using 75\% of the data (training dataset), and the rest 25\% of the data (test dataset) is used for validation. The hyperparameters corresponding to the maximum depth and the total number of trees in the Random Forest algorithm are tuned for each model using 3-fold cross-validation.
To identify the optimal clusters, we used the elbow method and the iterative clustering method combined with the {\em kneedle} algorithm. Figure~\ref{fig:elbow_knee} shows the results from the machine learning models developed using these methods.
\begin{sidewaysfigure}[htbp]
\centering
\includegraphics[width = 0.85\textwidth, trim = {0 .25cm 0 0.25cm}, clip]{prog_clust2.pdf}
\caption{Relationship between model performance and number of infrastructure clusters.}
\label{fig:elbow_knee}
\end{sidewaysfigure}
In the case of the elbow method, the optimal number of clusters in water, power, and transport systems are found to be six, seven, and seven, respectively. The train- and test $R^{2}$ corresponding to the optimal cluster counts are 0.98 and 0.87, respectively. The corresponding $RMSE$ values are 1.71 and 4.59 system performance-hours.
The machine learning models developed based on the iterative clustering algorithm reveal that the increase in the number of clusters in infrastructure systems initially leads to a noteworthy improvement in the model prediction. However, subsequent increases in the number of clusters only lead to marginal improvements in the same metrics. When the kneedle algorithm is used, the optimal number of clusters based on the iterative clustering method is found to be eight (two clusters in the power system and three each in the water system and the transport system). The corresponding Random Forest model has a train $R^{2}$ of 0.97 and a test $R^{2}$ of 0.85. The train $RMSE$ is 2.26 system performance-hours, and the train $RMSE$ is 4.97 system performance-hours.
Both elbow and the iterative clustering method resulted in improved prediction models compared to that of the model with the single cluster-level feature for each infrastructure system. In the case of the elbow method, the relative improvement observed with the elbow method in test $R^{2}$ is 16.13\%, whereas that using the iterative clustering method is 13.52\%. At the same time, the models identified using the elbow method and the iterative clustering method resulted in significant reductions of 27.67\% and 21.65\% in test $RMSE$, respectively. The results show that the iterative clustering method resulted in a model with considerably fewer cluster features than the elbow method (eight cluster features compared to 20 cluster features) without compromising too much on the prediction accuracy. The summary of the models developed in this study is presented in Table~\ref{tab:model_details}.
\begin{sidewaystable}[htbp]\small
\centering
\caption{Comparison of performance metrics of prediction models}
\begin{tabular}{lccccccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{2.25cm}{Clustering method}} & \multicolumn{1}{c}{\multirow{2}{1.25cm}{Total clusters}} & \multicolumn{1}{c}{\multirow{2}{1.5cm}{Train $R^{2}$}} & \multicolumn{1}{c}{\multirow{2}{1.5cm}{Test $R^{2}$}} & \multicolumn{1}{c}{\multirow{2}{1.25cm}{Train $RMSE$}} & \multicolumn{1}{c}{\multirow{2}{1.5cm}{Test $RMSE$}} & \multicolumn{2}{c}{Relative change$^{\dagger}$} \\
\cmidrule{7-8}
& & & & & & Test $R^{2}$ & Test $RMSE$ \\
\midrule
Single cluster & 3 & 0.8121 & 0.7485 & 5.54 & 6.34 & -- & --\\
Elbow method & 20 & 0.9821 & 0.8693 & 1.71 & 4.59 & +16.13\% & -27.67\%\\
Kneedle method & 8 & 0.9680 & 0.8497 & 2.26 & 4.96 & +13.52\% & -21.65\% \\
\bottomrule
\end{tabular}
\label{tab:model_details}
\begin{tablenotes}\footnotesize
\item[1] $^{\dagger}$All percentages are relative to the test dataset metrics obtained in the single cluster method.
\end{tablenotes}
\end{sidewaystable}
\section{Conclusions}
\label{sec:conclusions}
In this study, we introduced infrastructure component clustering methods to generate concise infrastructure resilience prediction models with fewer features than the traditional models, thereby resolving the problem of high dimensionality. Disaster scenarios and resultant impacts on interdependent infrastructure networks are simulated using an interdependent infrastructure model. The disaster impacts on infrastructure are quantified using well-established resilience metrics. Prediction models are developed by applying machine learning algorithms. We clustered the component-level features into cluster-level features to reduce the number of features in the models (dimensionality reduction). The clusters are identified by partitioning infrastructure components with similar topological and functional characteristics as indicators of component vulnerability and importance. Finally, we proposed algorithms for determining the optimal number of infrastructure component clusters based on elbow- and iterative clustering methods.
The clustering approach is a simple transformation technique that reduces the number of features (dimensionality reduction) and improves the model performance simultaneously. Since the clustering technique reduces the number of features in the model, improved prediction accuracy could be achieved with smaller simulation datasets. Therefore, the methodology can be adopted when simulation models' data generation is computationally expensive and time-consuming.
The methodology could be further improved by considering the following aspects to produce more accurate prediction models.
\begin{itemize}
\item Along with topological and functional characteristics, simulation data may also be used to improve the quality of network partitioning.
\item Clustering algorithms that implicitly learn the infrastructure network structure could produce more relevant clusters for resilience prediction.
\item Interdependencies are currently not considered for clustering of infrastructure components.
\item Additional component-level features relevant to resilience (for example, repair times) could be used in the clustering process to enhance the quality of clustering.
\end{itemize}
Though this paper only focused on interdependent infrastructure systems, our methodology can be used to design efficient ML models for any network problem where the characteristics of vertices or edges are treated as features. The proposed methodology could find applications to solve similar network problems in various fields, such as chemistry, medicine, finance, and social science.
\section*{Acknowledgements}
This research is supported by the National Research Foundation, Prime Minister’s Office, Singapore under its Campus for Research Excellence and Technological Enterprise (CREATE) programme.
\section*{CRediT Author Statement}
\noindent \textbf{Srijith Balakrishnan}: Conceptualization, Methodology, Data curation, Software, Visualization, Formal Analysis, Writing-Original Draft, Project administration, Supervision. \textbf{Beatrice Cassottana}: Conceptualization, Methodology, Validation, Resources, Writing-Review and Editing. \textbf{Arun Verma}: Conceptualization, Methodology, Validation, Writing-Review and Editing.
\section*{Data Availability}
The \textit{InfraRisk} package can be downloaded from \href{https://github.com/srijithbalakrishnan/dreaminsg-integrated-model}{GitHub}\footnote{\href{https://github.com/srijithbalakrishnan/dreaminsg-integrated-model}{https://github.com/srijithbalakrishnan/dreaminsg-integrated-model}}. Documentation and codes for sample simulations are available in the \textit{InfraRisk} package. The simulation data and codes pertaining to this study can be shared upon request.
\setlength{\bibsep}{0pt plus 0.3ex}
{\small |
\section{Introduction}
\label{sec:introduction}
\subsection{Modular Production}
Modular production systems based on Automated Guided Vehicles (AGVs) allow a drastically increased flexibility of the production plant or setup. Instead of a long line of machines connected by conveyor belts, production islands (consisting of one or several machines) are spread over the production plant and connected via AGVs. The inputs and outputs of these production islands usually also contain buffers.
The increased flexibility allows advancements like multiple product variants in the same plant, alternative stations (two stations of the same type as performance improvement or for failure safety) or the reuse of expensive stations at different stages of the production.
\subsection{The Job Shop Scheduling Problem}
\label{ssec:JSSP}
In modular production environments, the optimized scheduling of the AGVs (the decision which AGV is assigned to which transport task) is crucial for the system's overall performance. This task leads to the Job Shop Scheduling Problem (JSSP), which is -- like the related Travelling Salesman Problem -- a computationally ``hard'' (NP-complete) problem of enormous practical importance, \cite{Garey1976}.
Due to the typically factorial growth of possible arrangement with the size of the problem, exact solutions can only be found for very small problems. Since the AGV to Job scheduling has to be done in real-time, heuristic strategies are used. This is usually done with algorithms like nearest neighbor \cite{10.1007/BFb0032050} or cost tables \cite{https://doi.org/10.1002/nav.3800020109}. All of these methods have various advantages and drawbacks, but they are all reaching a limit with increasing complexity and system size. Varying driving times (because of human machine interaction or other vehicles in the systems) or processing times highly affect the systems performance and stability. Wrong or suboptimal scheduling decisions can lead to deadlocks in the system.
We show that Deep-Q based deep reinforcement learning agents are capable to overcome the aforementioned problems and also allow the calculation of optimal solutions in real-time.
\subsection{Deep Q-Learning}
\label{ssec:deep_Q}
Reinforcement learning aims to train an agent to interact with the environment in a way that maximizes the expected cumulative reward. The policy of the agent $\pi(\theta)$ is defined as a function that maps from a given state $s$ to a suitable action $a$.
Q-learning is a value-based approach, which means that the agent will try to approximate the expected cumulative reward (the “value”) for each possible pair of (state/action). This value is formalized as
\begin{equation}
Q^{*}(s, a)=\mathbb{E}_{s^{\prime}}\left[r+\gamma \max _{a^{\prime}} Q^{*}\left(s^{\prime}, a^{\prime}\right) \mid s, a\right]
\end{equation}
The expected cumulative reward $Q(s,\,a)$ in state $s$ for an action $a$ is defined as the sum of the immediate reward $r(s,\,a)$ and the highest Q-value for the next states among all possible actions. A discount factor $\gamma\in[0,\,1]$ is applied to the future reward in order to control how to rate future rewards compared to the immediate ones.
Due to the usually large number of state-action pairs, filling the whole table of Q values is possible only for very small problems. An approach to overcome this obstacle is \emph{Deep Q-learning}. This is a type of Q-learning where the value function is approximated by a deep neural network, that is, by a network with several layers. This is justified by the universal function approximation property of sufficiently deep neural networks, \cite{Hornik1989}.
Each time the Agent makes a decision, an experience (initial state, action, reward, next state) is saved into a buffer (the Replay Memory). The weights of the network are trained using backpropagation on mini-batches, sampled from the Replay Memory, \cite{Sutton2018}.
\subsection{Structure of the article}
\label{ssec:structure}
This paper is structured as follows: In section \ref{sec:environment}, the environment used for training the DRL agents is discussed in detail, including the particular configurations of the environments, as well as technical aspects, such as using OpenAI's gym interface, benchmark settings and heuristic approaches that have been used for comparison. The design of the DRL agent is then discussed in section \ref{sec:design_DRLagent}, where, based on Sec.~\ref{ssec:deep_Q}, a more detailed explanation of the reward signal used for training is given. The main findings of this paper are summarized in section \ref{sec:results}, where all chosen production setups are presented, together with the respective performance of the DRL agents and heuristic approaches. Finally, in section \ref{sec:conclusions}, we conclude with a short overall summary and point out possible paths for future investigations.
\section{Construction of the environment}
\label{sec:environment}
\subsection{Production plant modeling}
A production plant is modelled by a set of $n_{M}$ modular workstations. Each workstation consists of an Input Buffer ($IB$), a Production Unit ($PU$) and an Output Buffer ($OF$). Both buffers have a capacity of $n_{\rm buf}$ and are running as FIFO queues (First In-First Out). In addition, we defined a Start of Line station ($Source$) and an End of Line station ($Sink$). The $Source$ can be seen as a station with only an $OB$ and the $Sink$ as a station with only an $IB$.
Parts are defined by their part type PT. For each part type $\text{PT}_j$, a sequence of $n_{O}$ operations is defined. This sequence represents the workstations that the parts of type $\text{PT}_j$ must visit in the right order. The sequence always starts at the Source and ends at the Sink. The same workstation can be used multiple times in a sequence. Different part types do not necessarily use the same workstations. Parts are carried by $n_V$ AGVs $V$ with speed $v_V$. A vehicle can only carry 1 part at a time. The transfer of a part between a vehicle and a station, or between 2 units of a station, has a duration $T_{\rm transfer}$.
A weighted multi-directional graph $G$, called “waypoint graph", is defined to represent the distances between the stations. The nodes are the waypoints in the production plant through which the vehicles can pass. Each node $N_{j}$ is described by Cartesian coordinates {$x_{j}$, $y_{j}$}. The edges are the paths connecting these locations. Each edge is either unidirectional or bidirectional and has a weight corresponding to the euclidean distance between its nodes. Each buffer of each workstation is linked to a node of the waypoint graph. A node can only be assigned to a maximum of one buffer.
In our model, the release order of parts at the source is modelled by a uniform distribution over the part types. A source clock $C_{\rm source}$ defines when a new part is released to the system. Our goal is to maximize the production throughput by assigning jobs to the AGVs. A job consists of driving to a station $S_a$, picking a part $P_j$, driving to another station $S_b$, dropping the part $P_j$. Jobs must be assigned in such a way as to ensure that there are no deadlocks. This problem is defined as the Vehicle Management problem (VM), which is solved with deep learning (more precisely with proximal policy optimization) in \cite{Mayer2021}.
\subsection{Implementation details}
\subsubsection{Discrete-event simulation}
The modular production plant is simulated in Python using the Simpy library \cite{matloff2008introduction}, a framework to build Discrete-Event Simulations (DES) \cite{osti_6893405}. In a DES, events are triggered by the environment, each of them at a predefined instant in time. Between two consecutive events, the environment is static. We chose this approach because of its efficiency as compared to continuous simulations. Having a simulation running in Python has also been beneficial for the communication between the Agent and the Environment.
We performed a benchmark of our simulation tool in various cases to estimate the speed-up factor of our approach over simulations not relying on discrete event simulation. We defined the complexity of a configuration as the number of AGVs $N_{\rm agvs}$ times the number of Machines $N_{\rm machines}$. Each production plant configuration has been simulated for 12 hours using the FIFO agent. Since our simulation is not computing each intermediate steps when an AGV moves, but instantly jumps to the next relevant event (e.g. AGV reaches destination), the speed-up factor shows a strong dependence on the parameters of the environment, such as distances, AGV speed and processing times. For typical environments used in training, it takes between 360~ms and 2.6~s to simulate 1h of production, depending on the complexity of the environment. The calculation for the benchmark has been conducted on a laptop computer using a single thread (Dell XPS 15, Intel core i7 2.60Hz). Lacking an obvious definition of complexity of the environment renders quantitative speed-up factors somewhat arbitrary. We thus refrain from providing quantitative data at this point, but emphasize, that our DES based approach shows a very clear advantage in simulation time over discretized time simulations, which is a very desirable property simulation environments used for Deep Reinforcement Learning should possess.
\subsubsection{OpenAI Gym interface}
We use OpenAI’s library Gym \cite{brockman2016openai} to implement the interface between the Simpy simulation and the decision making process. In particular, the library provides utilities to handle observations, actions and rewards. The main concepts handled by the library are:
\begin{itemize}
\item \textbf{Observation:} object that describes the state of the Simpy simulation at a given instant. In our case, the Observation is a human-readable dictionary of values. Gym library takes care of serializing this data into a vector that can be sent to the neural network.
\item \textbf{Action:} object that describes the action taken by an Agent.
\item \textbf{Reward:} float value describing the amount of reward for a given action. The goal of the training is always to maximize the expected, cumulative reward.
\item \textbf{Environment:} the minimal acceptable environment is an object implementing a “step” method. This “step” method takes as input an action and outputs a tuple of four elements:
\begin{itemize}
\item \textbf{Observation:} the state of the environment once the action has been performed.
\item \textbf{Reward:} amount of reward achieved by the previous action
\item \textbf{Done:} a boolean value set to true if the environment must be reset.
\item \textbf{Info:} any information useful for debugging. Not used in our case.
\end{itemize}
\end{itemize}
Each time the "step" method of the environment is called, a new Experience -- a tuple (initial state, action taken, next state, reward, done) -- is stored in the memory for the learning process of the agent.
We had two major benefits of using this framework:
\begin{itemize}
\item Gym contains utilities in order to serialize and deserialize the Observation and Action concepts from a human-readable object to a vector suitable for a neural network.
\item Gym implements several well known problems \cite{openaigymenvs}. These tasks proved useful for testing and debugging the implementation of our agent against environments already solved by Deep-Q agents.
\end{itemize}
\subsubsection{Communication protocol}
The “Agent” is the algorithm that makes the decisions. We defined a common interface for both, deterministic and parameterized agents, which allowed us to seamlessly plug them into the simulation for comparison. Deterministic agents are algorithms that are entirely hard-coded, while parameterized agents require a training phase before evaluation.
The Agent Interface uses two methods:
\begin{itemize}
\item \textbf{Act:} takes as input an Observation state describing the environment and returns the action to perform.
\item \textbf{Step:} takes as input an Experience and returns nothing. This method is used for training the parameterized agents in order to improve their policy.
\end{itemize}
The Controller is an element of the Simpy environment that makes the connection between the AGVs and the Agent. In a production environment, this would be the Fleet Management System.
The communication process is the following:
\begin{enumerate}
\item The environment is initialized with all its elements.
\item The simulation starts.
\item When an AGV is available (it has finished its current task), it triggers a custom Simpy Event.
\item The Controller stops the simulation and computes an Observation state from the environment.
\item The Observation state is sent to the Agent.
\item The Agent takes an action that is communicated to the Controller.
\item The Controller deserializes the action and assigns the corresponding job to the waiting AGV. A job is a station from which an AGV must pick up a part.
\begin{enumerate}
\item If multiple AGVs are waiting at the same time, the Controller sends the job to the waiting AGV that will perform the action the fastest (i.e. the closest AGV for that job).
\item The Agent is then called immediately after to define another job for the other waiting AGVs until all AGVs have a job.
\end{enumerate}
\item Once all the AGVs have a job, the simulation starts over at 2.
\end{enumerate}
The Agent also has the possibility to choose not to assign any task. In that case, the waiting AGV will stay inactive until a new event is triggered.
\subsection{Benchmark settings}
\subsubsection{Production plant settings}
Several production plants have been configured in order to compare the Agents on different types of problems.
\paragraph{Mayer-Classen-Endisch}
This configuration is defined in \cite{Mayer2021}. It has 2 machines, 1 AGV and 1 part type. We re-implemented the simulation, taking the same parameters to verify that our approach reproduces the results of \cite{Mayer2021}.
\paragraph{1-machine-big}
This environment has 1 machine, 2 AGVs and 1 part type. The particularity lies in the distances between the nodes of the waypoint graph (see figure \ref{fig:plan_1_machine_big}). The pairs “source/machines input” and “machines output/sink” are very close, while the pairs “source/sink” and “machines input/machines output” are very far apart. Since there are 2 AGVs, the optimal solution is to have one vehicle doing circles between the source and the machine's input, while another vehicle does the same between the machine's output and the sink. We introduced this environment, expecting that the FIFO Agent would struggle with this setup.
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{plan_1_machine_big.png}
\caption{1-machine-big configuration: optimizing the driving time is essential here.}
\label{fig:plan_1_machine_big}
\end{figure}
\paragraph{3-machines-loop}
We set up a configuration with 3 machines and a unique part type (figure \ref{fig:plan_3_machines}). The particularity here is, that the sequence of operations to perform on a part has a loop, passing twice through the same machine. Each part $P_{j}$ has to follow the sequence $\text{Source} \rightarrow M1 \rightarrow M2 \rightarrow M1 \rightarrow M3 \rightarrow \text{Sink}$. This is a well known type of problem in modular production that often leads to deadlocks when using static algorithms.
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{plan_3_machines.png}
\caption{3 machines configuration: the sequence of operations in this setup includes a loop.}
\label{fig:plan_3_machines}
\end{figure}
\paragraph{6-machines-grid}
This is a family of problems starting with the same production plant configuration. The goal is to incrementally increase the difficulty of the scenarios and monitor how it impacts the performance of the agent at each step. For all sub-problems, the production plant layout, shown in figure \ref{fig:plan_grid_6_machines}, is shared. It has 6 machines distributed as a grid (2 rows of 3 machines), but depending on the scenario, not all of them are in use.
Complexity is increased in 2 ways:
\begin{itemize}
\item By increasing the number of machines to visit. Each sub-problem has 2 part types that follow similar paths in the production plant. 3 scenarios are considered, using either 2 machines ($M1 \rightarrow M2$ and $M2 \rightarrow M1$), 4 machines ($M1 \rightarrow M3 \rightarrow M2 \rightarrow M4$ and $M2 \rightarrow M4 \rightarrow M1 \rightarrow M3$) or 6 machines ($M1 \rightarrow M3 \rightarrow M5 \rightarrow M2 \rightarrow M4 \rightarrow M6$ and $M2 \rightarrow M4 \rightarrow M6 \rightarrow M1 \rightarrow M3 \rightarrow M5$).
\item By increasing the number of AGVs (from 1 to 4). With only 1 vehicle, the limiting factor is usually the speed of the vehicle, whereas with a larger number of vehicles it is possible to reach a 100\% workload on the machines.
\end{itemize}
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{plan_grid_6_machines.png}
\caption{6-machines grid configuration: depending on the scenario, not all machines are in use.}
\label{fig:plan_grid_6_machines}
\end{figure}
\subsubsection{Deterministic Agents}
Deterministic agents follow a hard-coded algorithm and are used as a baseline for result comparison. In this section, we call $\text{AGV}_{\rm longest}$ the AGV that has been waiting the longest for a new assignment and $P_{\rm longest}$ the part that has been waiting the longest in the output buffer of a station (named $S_{\rm longest}$).
An AGV is considered as active, if it has a task assigned to it and inactive otherwise.
\paragraph{FIFO}
The First-In First-Out (FIFO) algorithm consists of assigning $S_{\rm longest}$ to $\text{AGV}_{\rm longest}$.
\paragraph{Nearest Neighbor}
The Nearest Neighbor (NN) algorithm consists of selecting $S_{\rm longest}$ and compute for all the AGVs the time it will take to reach the station exit containing this part:
\begin{itemize}
\item If an AGV is inactive, it estimates the time it will take to drive to $S_{\rm longest}$.
\item If an AGV is active, it estimates the time it will take to finish the current task and then reach $S_{\rm longest}$.
\end{itemize}
The Agent assigns the job to the closest AGV in terms of duration. If this AGV is already active, the Agent doesn’t assign any task.
\paragraph{Cost Table}
The Cost Table algorithm is an extension of the Nearest Neighbor approach. Instead of doing an estimation only for the station $S_{\rm longest}$, it builds an array of size \texttt{[Number of inactive AGVs, Number of waiting stations]}, containing the time estimation for every AGV/Station pair (i.e., “costs”). The Agent solves the linear sum assignment problem on this Cost Table using the linear sum assignment algorithm \cite{linearsumassignment} implemented in SciPy \cite{2020SciPy-NMeth}. The solution of the problem is a list of AGV/Station pairs that minimizes the duration to start all the tasks. Since the Agent can only make one assignment at a time, it arbitrarily selects the task with the AGV that has been waiting the longest among the AGVs of the solution. The Agent is immediately called again to re-compute the Cost Table and re-solve the optimization problem until all AGVs are active, or no parts are waiting in a station output.
\subsubsection{Methodology}
Our goal is to be able to compare a trained Agent against static algorithms in order to prove the ability of DRL to tackle the Vehicle Management problem in a modular production environment.
We found out that static algorithms are highly vulnerable to the input clock $C_{\rm source}$. If $C_{\rm source}$ is too low, the AGVs will almost immediately overload the first machine with parts and run into a deadlock. On the contrary, if $C_{\rm source}$ is too high, the optimization problem doesn’t make sense anymore because it is too easy to get an output clock $C_{\rm sink}$ equal to $C_{\rm source}$.
In order to still be able to have baseline results to compare with, we first find the optimal $C_{\rm source}^*$ for each agent and each environment configuration. $C_{\rm source}^*$ is defined as the lowest input clock for which the simulation doesn’t run into a deadlock situation. We use a dichotomic search algorithm to find this value, rounded to 1 second. By construction, $C_{\rm sink}$ is necessarily equal or greater than $C_{\rm source}^*$. Once $C_{\rm source}^*$ is found, we run the simulation for 12 hours of production, gathering the metrics throuhout the run.
For the DRL Agent, $C_{\rm source}$ is set to $0$ during both, train and test phases. It is assumed that it should learn how to avoid deadlocks by not overloading the first machine.
\section{Design of the DRL Agent}
\label{sec:design_DRLagent}
In this section, we specify the deep-Q learning approach, briefly discussed in Sec.~\ref{ssec:deep_Q}, to our specific situation and discuss some details of its implementation.
\subsection{Observation, action, and reward design}
\label{ssec:observation_action_reward}
\subsubsection{Observation state}
The observation state is the vector describing the Environment that is sent to the Agent. It contains all the information needed to choose the next action. We defined a state $S_v$ for each AGV and a state $S_{\rm unit}$ for each station unit. The environment state $S$ is the concatenation of all the $S_v$ and $S_{\rm unit}$ states.
For each vehicle $V_i$, the state $\text{Sv}_i$ consist of:
\begin{itemize}
\item \texttt{Action\_state}: current action performed by the vehicle as a 1-hot encoded vector: \texttt{DRIVING}, \texttt{TRANSFERRING\_IN}, \texttt{TRANSFERRING\_OUT}, \texttt{WAITING\_FOR\_ORDER}, \texttt{WAITING\_TO\_DROPDOWN}, \texttt{WAITING\_TO\_PICKUP}
\item \texttt{Carried part}: a representation of the part that is carried, if any (see below).
\item \texttt{Current order target}: 1-hot encoded vector representing the destination of $V_i$, if driving. If $V_i$ is not currently driving, it is a zero vector of appropriate dimension.
\item \texttt{Last visited node}: 1-hot encoded vector representing the last node which $V_i$ has visited.
\end{itemize}
\noindent For each unit $U_i$, the state $\text{Sunit}_i$ consists of:
\begin{itemize}
\item \texttt{Action state}: current action performed by the unit as a 1-hot encoded vector: \texttt{PROCESSING}, \texttt{TRANSFERRING\_IN}, \texttt{TRANSFERRING\_OUT}, \texttt{WAITING\_TO\_DROPDOWN}, \texttt{WAITING\_TO\_PICKUP}
\item \texttt{Carried parts}: a representation of each part that is in the unit (see below).
\end{itemize}
For each part is the production plant, either in a station unit or an AGV, a state representation is defined as consisting of:
\begin{itemize}
\item \texttt{Part\_type}: 1-hot encoded vector representing the part type. The information is useful to determine what are the next operations to be performed on the parts.
\item \texttt{Part\_completion}: float value between 0 and 1. When a part is at the Source, the completion value is 0 while at the Sink it is set to 1. In between, the completion value is increased linearly each time an action is performed on the part (driving or processing).
\item \texttt{Part\_next\_station}: 1-hot encoded vector representing the next station the part needs to visit (or Sink if processing is done). In principle,
if only \texttt{part\_type} and \texttt{part\_completion} is provided, the Agent should be able to learn the next steps for a part by itself. However, we discovered by experience that providing redundant information helped to increase the overall performance.
\end{itemize}
\subsubsection{Action}
The set of all possible actions is equivalent to the set of stations where it is possible to pick-up a part (i.e., the Source and all the workstations). The Sink is not included in the set of actions, since it’s only possible to drop off a part there. In contrast to \cite{Mayer2021}, the set of all dropdown actions \texttt{A\_drop} is not defined in our case, since the dropdown station is always implicit and only depends on the carried part. This is a limitation in the case of a production plant where the same operation can be performed by 2 different workstations, but it is beyond the scope of this paper.
In addition to this set of actions, we also allow the Agent to take a “do no\-thing” action, for which no task is assigned. The inactive vehicles stay inactive until another event triggers a request to the Agent.
\subsubsection{Reward}
The reward function has been one of the most difficult components to configure in this setup. We realized that the performance of the Agent and its capability to learn are highly dependent on the reward signal. In the end, we built a score function for each environment state $S_t$. From a state $S_t$ and for an action $A_t$ that resulted in the state $S_{t+1}$, we defined the reward $R(S_t, A_t)$ as $\text{Score}(S_{t+1}) - \text{Score}(S_t)$.
We built the score of a state as the sum of 4 components:
\begin{itemize}
\item Per-part reward $S_{\rm pp}$: a positive reward for each completed part that reaches the Sink.
\item Per-part completion reward $S_{\rm pp\%}$: a positive reward each time a part is being processed. We define the completion of a part (between 0 and 1) as the percentage of processing steps that have been completed over all steps. When the part is at the Source, the completion is 0, while it is 1 at the Sink. This reward helps the Agent learn about the intermediate steps by providing shorter-term feedback.
\item Per-assigned decision reward $S_{\rm decisions}$: a positive reward for each decision taken by the Agent that is actually assigned to an AGV. In particular, the “do-nothing” action is not rewarded. We noticed a slight increase in performance when forcing the Agent into being less lazy.
\item Per-second reward $S_{\rm time}$: a positive reward for each second that has been simulated without encountering a deadlock.
\end{itemize}
We denote by $N_{\rm decisions}(t)$ the number of assigned decisions since the beginning of the simulation, by $N_{\rm pp}(t)$ the total number of completed parts, and by $N_{\rm pp\%}(t)$ the sum of each part completion. In practice $N_{\rm pp\%}(t) \ge N_{\rm pp}(t)$ at all time.
A difficulty we had has been to balance the 4 components of the score function. Depending on the environment configuration, the expected number of assigned decisions and parts processed were highly variable, whereas the number of simulated seconds was always the same (12-hours when no deadlocks). To balance those differences, we defined 3 configurable values: $E_{\rm pp}$ (expected processed parts), $E_{\rm decisions}$ (expected assigned decisions) and $E_{\rm seconds}$ (expected duration). The final score function is defined as a weighted average of the 4 components:
\begin{equation}
S(t) =\begin{cases}
K \, \left[ \frac{N_{\rm pp}(t) + N_{pp\%}(t)}{E_{\rm pp}} + \frac{N_{\rm decisions}(t)}{ E_{\rm decisions}}\right], & \text{if deadlocked},\\[6pt]
K \, \left[\frac{N_{\rm pp}(t) + N_{\rm pp\%}(t)}{E_{\rm pp}} + \frac{N_{\rm decisions}(t)}{E_{\rm decisions}} + \frac{t}{E_{\rm seconds}} \right], & \text{otherwise}.
\end{cases}
\end{equation}
In practice, we found that $K=4000$ worked best for us. $E_{\rm pp}$, $E_{\rm decisions}$ and $E_{\rm seconds}$ are estimated based on the Cost Table Agent performances.
\subsection{DQN implementation}
In this research, we tested the vanilla DQN agent described in \cite{Mnih} and implemented some additional promising extensions to improve the agent's performance. This section briefly describes the implemented extensions. The exact hyperparameters used for the performance can be found in \ref{tab:ddqn_hyperparameters}. In \cite{Mnih}, the authors showed that the vanilla DQN-agent performed better in 43 of 49 games as the best previously known reinforcement learning baselines, setting a new standard in the industry. However, further investigations of the DQN showed that it is prone to be overly optimistic when estimating Q-values when using Q-learning. Hence, it overestimates the Q-values. This over-optimism could hinder the convergence to the agent's optimum, or the agent might not even converge at all. In \cite{HasseltGS15}, the authors showed that the over-optimism could be counteracted by applying Double Q-learning rather than Q-learning. The resulting agent is called DDQN and requires almost no changes to the existing implementation.
Moreover, in DQN, the baseline approach to handle the exploration-exploitation dilemma is the straightforward epsilon-greedy strategy. This strategy relies on random permutations of the sequence of actions to explore the environment. However, in some environments, where the reward is very sparse, it can be challenging to explore the environment sufficiently using this strategy. Hence, we also tried to drive exploration using noisy networks as described in \cite{APMOGM17} by inducing noise when estimating the Q-values. Moreover, in this research, we also tested agents using a dueling network architecture, splitting the data stream in the neural network into two separate data streams as described in \cite{WangFL15}. One data stream estimates the value of the state, while the other data stream estimates the advantage for picking an action in that state. Finally, the most significant extension that showed the most robust performance during all tests is the prioritized experience replay (PER) as described in \cite{schaul2015prioritized}. The PER samples the experiences prioritized based on the temporal difference error rather than sampling uniformly.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c c|}
\hline
Parameter & Value \\
\hline
Replay memory size & 1e5 \\
Batch size & 64 \\
Gamma & 0.99 \\
Learning rate & 0.001 \\
Target update rate & 24 \\
Update rate & 4 \\
& \\
Epsilon & 1.0 \\
Epsilon decay & 0.9995 \\
Epsilon min & 0.01 \\
& \\
NN nb layers & 2 \\
FC1 units & 64 \\
FC2 units & 32 \\
\hline
\end{tabular}
\end{center}
\caption{DDQN training hyperparameters.}
\label{tab:ddqn_hyperparameters}
\end{table}
\section{Experimental results}
\label{sec:results}
\begin{table}[ht]
\centering
\bgroup
\def\arraystretch{1.1}
\begin{tabular}{|l|l||r|r|r|r|}
\hline
Environment & Agent & 1 AGV & 2 AGVs & 3 AGVs & 4 AGVs \\
\hline \texttt{Mayer-Classen-Endisch} & \texttt{FIFO}& & 50 \quad & & \\
& \texttt{Nearest Neighbor} & & 50 \quad & & \\
& \texttt{Cost Table} & & 50 \quad & & \\
& \texttt{DDQN} & & 0 \quad & & \\
\hline \texttt{1-machine-big} & \texttt{FIFO}& & 83 \quad & & \\
& \texttt{Nearest Neighbor} & & 48 \quad & & \\
& \texttt{Cost Table} & & 11 \quad & & \\
& \texttt{DDQN} & & 0 \quad & & \\
\hline \texttt{3-machines-loop} & \texttt{FIFO}& 153 \quad & 81 \quad & 56 \quad & \\
& \texttt{Nearest Neighbor} & 153 \quad & 80 \quad & 57 \quad & \\
& \texttt{Cost Table} & 120 \quad & 82 \quad & 55 \quad & \\
& \texttt{DDQN} & 0 \quad & 0 \quad & 0 \quad & \\
\hline \texttt{2-machines-grid} & \texttt{FIFO}& 31 \quad & 25 \quad & 25 \quad & 25 \quad \\
& \texttt{Nearest Neighbor} & 31 \quad & 25 \quad & 24 \quad & 22 \quad \\
& \texttt{Cost Table} & 20 \quad & 25 \quad & 24 \quad & 22 \quad \\
& \texttt{DDQN} & 0 \quad & 0 \quad & 0 \quad & 0 \quad \\
\hline \texttt{4-machines-grid} & \texttt{FIFO}& 41 \quad & 25 \quad & 25 \quad & 25 \quad \\
& \texttt{Nearest Neighbor} & 41 \quad & 25 \quad & 25 \quad & 25 \quad \\
& \texttt{Cost Table} & 37 \quad & 25 \quad & 25 \quad & 24 \quad \\
& \texttt{DDQN} & 0 \quad & 0 \quad & 0 \quad & 0 \quad \\
\hline \texttt{6-machines-grid} & \texttt{FIFO}& 53 \quad & 29 \quad & 25 \quad & 25 \quad \\
& \texttt{Nearest Neighbor} & 53 \quad & 29 \quad & 25 \quad & 25 \quad \\
& \texttt{Cost Table} & 48 \quad & 28 \quad & 25 \quad & 25 \quad \\
& \texttt{DDQN} & 0 \quad & 0 \quad & 0 \quad & 0 \quad \\
\hline
\end{tabular}
\egroup
\caption{Optimal source clock}
\label{tab:results_optimal_clock}
\end{table}
\begin{table}[ht]
\centering
\bgroup
\def\arraystretch{1.1}
\begin{tabular}{|l|l||r|r|r|r|}
\hline
Environment & Agent & 1 AGV & 2 AGVs & 3 AGVs & 4 AGVs \\
\hline \texttt{Mayer-Classen-Endisch} & \texttt{FIFO}& & \textbf{71.8} \quad & & \\
& \texttt{Nearest Neighbor} & & \textbf{71.8} \quad & & \\
& \texttt{Cost Table} & & \textbf{71.8} \quad & & \\
& \texttt{DDQN} & & \textbf{71.8} \quad & & \\
\hline \texttt{1-machine-big} & \texttt{FIFO}& & 23.6 \quad & & \\
& \texttt{Nearest Neighbor} & & 74.9 \quad & & \\
& \texttt{Cost Table} & & 143.3 \quad & & \\
& \texttt{DDQN} & & \textbf{143.4} \quad & & \\
\hline \texttt{3-machines-loop} & \texttt{FIFO}& 23.4 \quad & 44.2 \quad & 63.8 \quad & \\
& \texttt{Nearest Neighbor} & 23.4 \quad & 44.8 \quad & 62.8 \quad & \\
& \texttt{Cost Table} & \textbf{26.9} \quad & 43.7 \quad & \textbf{65.0} \quad & \\
& \texttt{DDQN} & 26.3 \quad & \textbf{45.4} \quad & 60.2 \quad & \\
\hline \texttt{2-machines-grid} & \texttt{FIFO}& 115.8 \quad & 143.7 \quad & \textbf{143.8} \quad & 143.7 \quad \\
& \texttt{Nearest Neighbor} & 115.8 \quad & \textbf{143.8} \quad & \textbf{143.8} \quad & 143.7 \quad \\
& \texttt{Cost Table} & \textbf{123.1} \quad & \textbf{143.8} \quad & \textbf{143.8} \quad & 143.7 \quad \\
& \texttt{DDQN} & 115.2 \quad & 139.2 \quad & \textbf{143.8} \quad & \textbf{143.8} \quad \\
\hline \texttt{4-machines-grid} & \texttt{FIFO}& 87.4 \quad & \textbf{143.4} \quad & 143.4 \quad & \textbf{143.5} \quad \\
& \texttt{Nearest Neighbor} & 87.4 \quad & \textbf{143.4} \quad & \textbf{143.5} \quad & \textbf{143.5} \quad \\
& \texttt{Cost Table} & \textbf{95.7} \quad & \textbf{143.4} \quad & \textbf{143.5} \quad & 143.3 \quad \\
& \texttt{DDQN} & 88.8 \quad & 128.9 \quad & 138.9 \quad & 134.1 \quad \\
\hline \texttt{6-machines-grid} & \texttt{FIFO}& 67.5 \quad & 123.5 \quad & 143.1 \quad & \textbf{143.2} \quad \\
& \texttt{Nearest Neighbor} & 67.5 \quad & 123.5 \quad & \textbf{143.2} \quad & \textbf{143.2} \quad \\
& \texttt{Cost Table} & \textbf{74.1} \quad & \textbf{127.8} \quad & \textbf{143.2} \quad & \textbf{143.2} \quad \\
& \texttt{DDQN} & 68.6 \quad & 79.0 \quad & 117.4 \quad & 124.0 \quad \\
\hline
\end{tabular}
\egroup
\caption{Throughputs in parts per hour (pph)}
\label{tab:results_throughput}
\end{table}
\begin{table}[ht]
\centering
\bgroup
\def\arraystretch{1.1}
\begin{tabular}{|l|l||r|r|r|r|}
\hline
Environment & Agent & 1 AGV & 2 AGVs & 3 AGVs & 4 AGVs \\
\hline \texttt{Mayer-Classen-Endisch} & \texttt{FIFO}& & 862 \quad & & \\
& \texttt{Nearest Neighbor} & & 862 \quad & & \\
& \texttt{Cost Table} & & 862 \quad & & \\
& \texttt{DDQN} & & \textbf{863} \quad & & \\
\hline \texttt{1-machine-big} & \texttt{FIFO}& & 284 \quad & & \\
& \texttt{Nearest Neighbor} & & 899 \quad & & \\
& \texttt{Cost Table} & & 1720 \quad & & \\
& \texttt{DDQN} & & \textbf{1721} \quad & & \\
\hline \texttt{3-machines-loop} & \texttt{FIFO}& 281 \quad & 531 \quad & 768 \quad & \\
& \texttt{Nearest Neighbor} & 281 \quad & 538 \quad & 755 \quad & \\
& \texttt{Cost Table} & \textbf{326} \quad & 525 \quad & \textbf{782} \quad & \\
& \texttt{DDQN} & 317 \quad & \textbf{547} \quad & 725 \quad & \\
\hline \texttt{2-machines-grid} & \texttt{FIFO}& 1392 \quad & \textbf{1726} \quad & 1726 \quad & 1726 \quad \\
& \texttt{Nearest Neighbor} & 1392 \quad & \textbf{1726} \quad & 1726 \quad & \textbf{1727} \quad \\
& \texttt{Cost Table} & \textbf{1479} \quad & \textbf{1726} \quad & 1726 \quad & \textbf{1727} \quad \\
& \texttt{DDQN} & 1383 \quad & 1672 \quad & \textbf{1727} \quad & \textbf{1727} \quad \\
\hline \texttt{4-machines-grid} & \texttt{FIFO}& 1051 \quad & \textbf{1724} \quad & \textbf{1724} \quad & 1724 \quad \\
& \texttt{Nearest Neighbor} & 1051 \quad & \textbf{1724} \quad & \textbf{1724} \quad & 1724 \quad \\
& \texttt{Cost Table} & \textbf{1151} \quad & \textbf{1724} \quad & \textbf{1724} \quad & \textbf{1725} \quad \\
& \texttt{DDQN} & 1068 \quad & 1551 \quad & 1670 \quad & 1611 \quad \\
\hline \texttt{6-machines-grid} & \texttt{FIFO}& 812 \quad & 1485 \quad & \textbf{1722} \quad & \textbf{1723} \quad \\
& \texttt{Nearest Neighbor} & 812 \quad & 1485 \quad & \textbf{1722} \quad & \textbf{1723} \quad \\
& \texttt{Cost Table} & \textbf{892} \quad & \textbf{1537} \quad & \textbf{1722} \quad & \textbf{1723} \quad \\
& \texttt{DDQN} & 828 \quad & 951 \quad & 1413 \quad & 1491 \quad \\
\hline
\end{tabular}
\egroup
\caption{Total parts produced over 12h}
\label{tab:results_total_parts}
\end{table}
\subsection{Per-scenario results}
The benchmark has been run on each scenario for 12 hours with 4 different agents: 3 static (FIFO, Nearest Neighbor, Cost Table) and 1 trained (DDQN). Before running the simulation, we determined the optimal source clock $C_{\rm source}*$ for each static agent. The main metric that we monitor is the overall throughput, defined in parts per hour (pph). We also measured the total amount of parts produced in 12 hours.
Optimal source clocks can be found in table \ref{tab:results_optimal_clock}, throughput results in table \ref{tab:results_throughput}, and total parts produced results in table \ref{tab:results_total_parts}.
\subsubsection{Mayer-Classen-Endisch}
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{results_simple_cases.png}
\caption{Results on Mayer-Classen-Endisch and $1$-machine-big environments.}
\label{fig:simple_cases_results}
\end{figure}
The first environment we trained on has been the one described by \cite{Mayer2021}. Results are shown in figure \ref{fig:simple_cases_results}. The trained agent reached a throughput of 71.8 parts/hour over 12 hours (863 parts). The maximum throughput in this case is 72 parts/hours (864 parts). This environment has been a good first example, since we have been able to reproduce 2 results described by \cite{Mayer2021}:
\begin{itemize}
\item The agent starts with a ramp-up process before reaching the optimal throughput. This explains the difference of 1 produced part observed over the 12 hours.
\item The final policy learned by the agent is the same, namely, the AGV loops through the production plant in a certain order: $\text{Source} \rightarrow M1_{\rm input} \rightarrow M1_{\rm output} \rightarrow M2_{\rm input} \rightarrow M2_{\rm output} \rightarrow \text{Sink} \rightarrow \text{Source} \rightarrow \ldots$
\end{itemize}
Finally, it is also worth mentioning that the baseline agents (FIFO, Nearest Neighbor and Cost Table) have also been able to achieve the same performance (862 parts produced). The main benefit of the DDQN agent is its robustness to the source clock.
\subsubsection{1-machine-big}
With this configuration, we expected the FIFO agent to be tricked, while a Cost Table approach was supposed to tackle the problem raised by the long distances in the production plant. We confirmed this intuition, since they reached 23 parts/hour and 143.4 parts/hour respectively, the optimal throughput in the equilibrium state being 144 parts/hour. The same performance has been achieved by the trained agent. Results can be compared in figure \ref{fig:simple_cases_results}.
For both Cost Table and DDQN approaches, the final policy in the equilibrium state is to have an AGV looping between the Source and the Machine entry and the other AGV looping between the Machine exit and the Sink, thus avoiding the long drive between Machine entry and exit.
It is also worth mentioning that the Nearest Neighbor approach gets intermediate results (74.9 parts/hour) which confirm that all static approaches are not equivalent even in small scenarios.
\subsubsection{3-machines-loop}
The difficulty in this scenario is to manage the AGV(s) in order not to overload the first machine. Results show that in the end, the different agents get on average similar results (see figure \ref{fig:3_machines_results}). In the details, Cost Table is slightly better than DDQN with 1 AGV (26.9 vs 26.3 pph), slightly worse with 2 AGVs (43.7 vs 45.4 pph) and significantly better with 3 AGVs (65.0 vs 60.2 pph). This shows the difficulties the DDQN agent has to manage multiple AGVs with our framework. We will discuss this aspect further in \ref{sssec:several_agvs}.
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{results_3_machines.png}
\caption{Results on the $3$-machines-loop environment.}
\label{fig:3_machines_results}
\end{figure}
\subsubsection{6-machines-grid}
\paragraph{2-machines scenario}
The 2-machines scenario seems quite easy to tackle for the static agents (see figure \ref{fig:grid_2_machines_results}). The Cost Table approach performs better with 1 AGV but starting from 2 AGVs, the processing time is more limiting than transportation, which means that even a sub-optimal policy is able to obtain the best theoretical throughput (144 pph).
The DDQN agent is able to reach FIFO/Nearest Neighbor performances with 1 AGV (115 pph) but is not optimal since Cost Table performs significantly better (123 pph). With 2 AGVs it is not optimal either, performing at 139 pph even though all static agents are already reaching 144 pph. Above that (3 and 4 AGVs), comparisons don't make sense, since all agents reach the limit.
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{results_grid_2_machines.png}
\caption{Results on the grid-$2$-machines environment.}
\label{fig:grid_2_machines_results}
\end{figure}
\paragraph{4-machines scenario}
The 4-machines scenario is similar to the previous one (see figure \ref{fig:grid_4_machines_results}). Cost Table performs better than other approaches with 1 AGV (96 pph vs 89) and then all static agents reach the limit (144 pph).
However, this scenario starts to show the limitation of the DDQN agent. It is able to learn how to manage the AGVs to transport parts and still avoid deadlocks, but in a sub-optimal way. The best results are achieved with 3 AGVs (139 pph), a better performance than when a 4th AGV was available (134 pph).
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{results_grid_4_machines.png}
\caption{Results on the \texttt{grid-4-machines} environment.}
\label{fig:grid_4_machines_results}
\end{figure}
\paragraph{6-machines scenario}
In the 6 machines scenario, it is even more difficult for the DDQN to reach optimal performances -or at least similar to static agents- (see figure \ref{fig:grid_6_machines_results}). It is worth mentioning that even if the DDQN performs worse than Cost Table in all cases (69 vs 74 pph with 1 AGV, 79 vs 128 pph with 2, 117 vs 143 with 3 and 124 vs 143 with 4), it is still able to run the 12 hours without deadlocks.
\begin{figure}[ht]
\includegraphics[width=1.0\textwidth]{results_grid_6_machines.png}
\caption{Results on the \texttt{grid-6-machines} environment.}
\label{fig:grid_6_machines_results}
\end{figure}
\subsection{Improvement perspectives}
\subsubsection{Improve the observation state}
Our final observation state does not provide all the information needed by the agent to take the optimal decision. In particular, with our approach, the DDQN agent can't take advantage of the exact position of the AGVs and the transportation times. Our efforts to include this information in the observation state have not yet been successful. However, this information is available and used by the Cost Table agent, which can explain its better performances.
\subsubsection{Improve reward signal and deadlock punishment}
Another learning of our study is the sensitivity of the DDQN agent to the reward signal, especially to the deadlock punishment signal. In the end, our agent learned how to avoid running into a deadlock but at the cost of been too protective in some cases (i.e., it is more risky and not enough rewarded to start processing a new part as compared to wait for the production plant to be less full). This balance has been tough to configure and is still not entirely satisfying.
\subsubsection{Improve management of several AGVs}
\label{sssec:several_agvs}
The study has shown the limitation of our approach when it comes to managing several AGVs at the same time. We believe that there is room for improvement in the communication protocol between the Agent and the AGVs. Our intuition is that the current set of actions the Agent can take is too limited. We had difficulties into making the Agent choose both the Station and the AGV that must perform the action.
\section{Conclusions and Outlook}
\label{sec:conclusions}
In this paper, we investigated the deployment of Deep-Q based DRL agents to job-shop scheduling problems in modular production facilities using discrete event simulations for the environment. Here are the key findings of our study:
\begin{itemize}
\item We have developed a simulation that runs fast enough to train DRL agents. We have established a framework to configure environments of different complexity, on which we can additionally run a baseline of three static agents.
\item For each environment, we trained a DDQN agent that is able to run the production plant deadlock-free. The trained agent has learned some complex strategies (beating FIFO in $1$-machine-big and handling a loop in $3$-machines-loop environment).
\item We proved that it is possible to apply DRL to the vehicle management problem. From this, there is room for improvement, especially for managing complex scenarios with more AGVs and stations.
\item Beyond the scope of this paper, there exist various avenues for further exploration. For instance, we hypothesize that a DRL approach increases stability in the system. This is suggested by the fact the DDQN agent is robust against the source clock, while static agents need additional precautions. We also performed simulations in which we introduced noise (in the source clock, the transports, the parts transfers and the processing). The trained agents tend to be more robust in those cases, avoiding deadlocks.
\end{itemize}
With these findings, we have laid out a foundation for future studies that can build upon these findings. Both, the agents, and the environment, can be improved to facilitate environments of a richer variety, and particular focus can be dedicated to further investigating the robustness of the DRL agents in comparison to the heuristic approaches. Also, the issue of dead-lock avoidance could be investigated more thoroughly. As for the environment and apart from increasing the complexity by means of adding stations, a more diverse work-flow, or more AGVs, one could furthermore envision an environment in which the AGVs themselves are controlled by DRL agents, considering additional factors such as battery status, nearest charging station, position of the other AGVs, etc. However, all these avenues are left unaddressed in this study, as the purpose of this paper is to establish a first, preliminary exploration of the feasibility of Deep-Q based DRL approaches in modular production environments.
\section{Acknowledgements and references}
\label{sec:acknowledgements}
Partial funding through the Austrian Research Promotion Agency FFG, Kleinprojekt Nr. 883243, is kindly acknowledged.
Parts of the findings of this article are also given in \cite{Hasenbichler2021}.
|
\section{Introduction}
Traditional deep neural networks are becoming larger, deeper, and more complex in order to keep pushing and defining the competitive limits for machine learning. Modern deep learning is mostly based on traditional supervised learning, leveraging massive amounts of labeled data to train models. As such, there is often an abundance of unlabeled data that can not be used in a traditional supervised context. The incentive to research and develop alternative methods has been growing in recent years. Self-supervised learning models and frameworks like CPC\citep{oord2019cpc}, SimCLR \citep{chen2020simclr}, BYOL \citep{grill2020byol}, and SimSIAM \citep{chen2020simsiam} show promising results in low-labeled data regimes, with representations generalizing well in downstream tasks like visual image classification. The mentioned models usually leverage contrastive loss functions to learn co-occuring modalities present in the unlabeled data. These loss functions have already been explored, studied, and evaluated rather extensively, but there does not seem to exist a defined upper bound for the NT-Xent loss used by the SimCLR framework. Therefore, this short papers aims at defining one such boundary.
\section{The SimCLR framework}
\label{sec:SimCLR}
The simple framework for contrastive learning on visual representations (SimCLR) proposed by \citep{chen2020simclr} builds upon properties of noise contrastive estimation (NCE) and is inspired by previous work in the related field. It aims to simplify the process of contrastive representation learning, by not requiring the use of memory banks for sampling nor maximization of mutual information between latent space representations \citep{wu2018memory, oord2019cpc}. Instead, positive pairs are generated by leveraging data augmentations and SimCLR learns representations by maximizing an agreement between the positive pairs. The authors demonstrate that leveraging SimCLR in a low-labeled data regime produces competitive performance and even succeeds in outperforming the AlexNet when fine-tuned on 100$\times$ fewer labels. The framework can be formalized by four main components:
\begin{itemize}
\item A stochastic data augmentation module $\mathcal{T}$ is used to sample two augmentations $t, t' \in \mathcal{T}$ which generates two augmented views $\tilde{x_i} = t(x), \tilde{x_j} = t'(x)$ of an original datapoint $x$. The generated views are referred to as a positive pair, and are assumed to be correlated since they originate from the same datapoint $x$.
\item An artificial neural network encoder $f$ which is trained to extract features from the augmented views $\tilde{x_i}$ and $\tilde{x_j}$. Thus, the encoder maps $f(\tilde{x_i}) = h_i$ where $h_i \in {\rm I\!R}^d$ is the feature of interest in the embedding space, which is to be used for the domain-specific downstream tasks.
\item An artificial neural network projection head $g$ that performs a non-linear mapping of the embedding feature to the latent space, $g(h_i) = z_i \in {\rm I\!R}^m$. The authors suggest to use a simple MLP for projection head, with $ReLU(x) = max(0, x)$ activation functions between layers.
\item A contrastive loss function $\mathcal{L}$ that is applied on the samples in the latent space. Here, the authors propose the temperature scaled loss NT-Xent which is based on noise contrastive estimation and cross entropy \citep{chen2020simclr}.
\end{itemize}
\section{The NT-Xent loss}
\label{sec:NT-Xent}
The normalized temperature-scaled cross entropy loss (NT-Xent) is designed to maximize an agreement between positive pairs in a mini-batch $\mathcal{MB}$. No explicit negative sampling is performed, instead, $\mathcal{MB}$ is constructed such that $N$ datapoints are sampled and applied to $\mathcal{T}$ to generate positive pairs $\tilde{x_i}$ and $\tilde{x_j}$. This yields $\mathcal{MB} = \{\tilde{x_{1i}}, \tilde{x_{1j}}, ..., \tilde{x_{Ni}}, \tilde{x_{Nj}} \}$, and as such, $|\mathcal{MB}| = 2N$. Thus, given a positive pair, the other $2(N-1)$ datapoints are to be treated as negative samples. The agreement to maximize is the cosine distance of similarity,
\begin{equation}
\label{eq:cosine-similarity}
sim(z_i, z_j) = \frac{z_i \cdot z_j}{\lVert z_i \rVert \lVert z_j \rVert}
\end{equation}
where $z_i$ and $z_j$ are latent space representations of a sampled positive pair. The total loss for all positive pairs is formalized as
\begin{equation}
\label{eq:NT-Xent}
\mathcal{L}^{\mathrm{NT-Xent}}= -\frac{1}{N}\sum_{i,j\in \mathcal{MB}} \mathrm{log}\frac{\mathrm{exp}(sim(z_i, z_j)/\tau)}{\sum_{k=1}^{2N} \bbone_{[k \neq i]} \mathrm{exp}(sim(z_i, z_k)/\tau)}
\end{equation}
where $\tau$ is the temperature scalar, and $\bbone_{[k\neq i]} \in \{0, 1\}$ is an indicator function to indicate that the similarity between the same point, $sim(z_i, z_k)_{[k=i]}$, is not contributing to the normalization. Equation (\ref{eq:NT-Xent}) can be rewritten accordingly by applying the logarithmic rules
\begin{equation}
\begin{aligned}
\label{eq:NT-Xent_alt}
\mathcal{L}^{\mathrm{NT-Xent}} &= -\frac{1}{N}\sum_{i,j\in \mathcal{MB}} \Big( sim(z_i, z_j)/\tau- \mathrm{log}\sum_{k=1}^{2N}\bbone_{[k\neq i]}\mathrm{exp}(sim(z_i, z_k)/\tau)\Big)\\
&= \underbrace{-\frac{1}{N}\sum_{i,j\in \mathcal{MB}}sim(z_i, z_j)/\tau}_{\mathcal{L}_{alignment}} + \underbrace{\frac{1}{N}\sum_i^N\mathrm{log}\Big(\sum_{k=1}^{2N}\bbone_{[k \neq i]}\mathrm{exp}(sim(z_i, z_k)/\tau)}_{\mathcal{L}_{distribution}}\Big)
\end{aligned}
\end{equation}
and as such there are two parts of the loss, namely, \textit{alignment} and \textit{distribution}, as proposed by \citep{wang2020loss} and further studied by \citep{chen2020loss}. Furthermore, a similar way to analyze the NT-Xent loss was proposed by Grill et al. in the bootstrap your own latent (BYOL) paper where direct comparisons were made to the proposed SimCLR loss \citep{grill2020byol}. The second term includes summation over the well known \textit{LogSumExponent} (LSE) term which is defined as
\begin{equation}
\label{eq:LSE}
\mathrm{LSE}(x_1, ..., x_n) = \mathrm{log}\Big(\mathrm{exp}(x_1) + \cdots + \mathrm{exp}(x_n)\Big)
\end{equation}
where $x_{i,k}$ will denote $sim(z_i, z_k)/\tau$, $k\in \{1, ..., 2N\}$, and $n=2N$ to be consistent with the NT-Xent loss notations. It can be proven that LSE is bound accordingly (strictly greater for $n>1$ and strictly lower unless all arguments are equal)
\begin{equation}
\label{eq:LSE_bound}
max(x_{i,1}, ..., x_{i,n}) \leq \mathrm{LSE}(x_{i,1}, ..., x_{i,n}) \leq max(x_{i,1}, ..., x_{i,n}) + \mathrm{log}(n).
\end{equation}
Thus, using equations (\ref{eq:LSE}) and (\ref{eq:LSE_bound}) one can rewrite equation (\ref{eq:NT-Xent_alt} to produce an upper bound for the average similarity between the positive pairs in $\mathcal{MB}$, accordingly
\begin{equation}
\begin{aligned}
\mathcal{L}^{\mathrm{NT-Xent}} &= -\frac{1}{N}\sum_{i,j\in\mathcal{MB}}x_{i,j} + \frac{1}{N}\sum_i^N\mathrm{LSE}(x_{i,1}, ..., x_{i,n}) \\
&\leq -\frac{1}{N}\sum_{i,j\in\mathcal{MB}}x_{i,j} + \frac{1}{N}\sum_i^N\Big( max(x_{i,1}, ..., x_{i,n}) + \mathrm{log}(n)\Big) \\
&= -\frac{1}{N}\sum_{i,j\in\mathcal{MB}}x_{i,j} + \mathrm{log}(n) + \frac{1}{N}\sum_i^N max(x_{i,1}, ..., x_{i,n})
\end{aligned}
\end{equation}
and by rearranging the expression, multiplying by $\tau$, and returning to the original complete notations, one can get the final expression as
\begin{equation}
\label{eq:avg_sim}
\frac{1}{N}\sum_{i,j\in\mathcal{MB}}sim(z_i, z_j)\leq \tau\mathrm{log}(2N) - \tau\mathcal{L}^{\mathrm{NT-Xent}} + \frac{\tau}{N}\sum_i^N max(sim(z_i, z_1)/\tau, ..., sim(z_i, z_{2N})/\tau).
\end{equation}
\section{Final remarks}
From equation (\ref{eq:avg_sim}) one can clearly see the three crucial terms that put constraints on and alter how well a SimCLR trained model can learn to position positive pairs closely and aligned in the latent space. As always, the objective is to minimize the loss $\mathcal{L}$ which is done by means of gradient descent. However, since the loss expression includes terms directly related to properties of the learned representation, there is an interplay between updating model parameters $\theta$ and altering where the samples are positioned in the latent space. Hence, analyzing the behavior of the loss by only considering one of the terms is highly theoretical, and might not represent the true behavior. Therefore, this paper only aims at presenting the upper bound for average similarity, but hopefully enabling future work to practically analyze and evaluate what has been proposed.
\section{Acknowledgments}
I would like to present a sincere thanks to Arvind Kumar and Erik Fransén for granting me the opportunity to write a master's thesis on self-supervised learning and magnetoencephalography, which inspired me to write this brief paper. Furthermore, I am very grateful for the supervision that Erik Fransén has provided, which motivated me to keep being creative and passionate about machine learning.
\bibliographystyle{unsrtnat}
|
\section{Introduction}
Cyber attacks against power system grids can have significant social and economic consequences. The threats can be broadly divided into two categories -- (i) attacks that directly target the power grid's supervisory control and data acquisition (SCADA) system, and (ii) attacks that indirectly target the power grid's control loops by manipulating end-user internet-of-things (IoT) enabled electrical appliances. Direct attacks against the SCADA system, such as false data injection attacks and/or coordinated cyber-physical attacks against power grid state estimation have received significant attention \cite{Liu2009, LakshGT2021, LakshRMT2021}. In contrast, indirect attacks that target a large number of demand-side appliances in a Botnet-type attack have been studied only recently \cite{Dabrowski2017, Soltan2018}. Unlike the SCADA assets, these devices cannot be monitored continuously due to their large numbers.
The focus of this work is on
\emph{load-altering attacks} (LAAs), which refer to a sudden and abrupt change in the power grid demand by synchronously turning on/off a large number of IoT-enabled high-wattage appliances \cite{Dabrowski2017, Soltan2018, HuangUSENIX2019, AminiLAA2018, LakshIoT2021}. LAAs pose a major threat to power grid operations since they can potentially disrupt the balance between supply and demand. It has been shown \cite{Dabrowski2017, Soltan2018} that such attacks can lead to unsafe frequency excursions, line outages, and/or increase the grid's operational costs. Moreover, dynamic LAAs, in which the attacker injects a series of load perturbations over time, can also destabilize the power grid's frequency control loop \cite{AminiLAA2018}. Subsequent work has also focused on detecting LAAs using data-driven approaches based on the data gathered from phasor measurement units \cite{AminiIdentification2019, lakshminarayana2021datadriven}.
Understanding the impact of LAAs is an important component in risk analysis. Existing work on quantifying the impact of LAAs can be categorized into two approaches -- (i) simulation-based approach \cite{Dabrowski2017, Soltan2018, HuangUSENIX2019, AminiLAA2018}, and (ii) analytical approach \cite{LakshIoT2021}. Under the former, the attack impact is assessed by simulating the power grid's control loops (e.g., frequency dynamics) \cite{Dabrowski2017, Soltan2018, HuangUSENIX2019} or computing the system's eigenvalues under LAAs \cite{AminiLAA2018}. However, the results presented in these works correspond to only a few specific LAA scenarios (i.e., specific load perturbations injected at the victim nodes). Evaluating the attack impact under all possible spatial distributions of LAAs (over the victim nodes) requires performing extensive simulations considering different combinations of the victim nodes and attack magnitudes, which can be computationally prohibitive. To overcome these issues, an analytical approach based on the theory of second-order dynamical systems was proposed in \cite{LakshIoT2021}. The closed-form analytical functions to evaluate the attack impact (in terms of the dynamic response and eigensolutions of the power grid's frequency control loop) only need to be computed once, thus avoiding the requirement for repeated simulations. However, the analytical approach is restricted to a second-order model with a direct current (DC) power flow model. Extending these results to higher-order models (e.g., one that considers both frequency and voltage dynamics) and
involving the non-linear alternating current (AC) power flow model is non-trivial. This is important to obtain a realistic assessment of the LAA impact.
To overcome the aforementioned limitations, we apply a sampling approach to map LAA magnitudes and their spatial distributions (across the different victim nodes) to their attack impact. In particular, the focus of this work is on LAAs that lead to the activation of power system emergency responses (ERs) which disconnect critical power system components (e.g., generators/load/transmission lines). A common sampling approach is \emph{Monte Carlo} simulations, which would apply randomly generated realisations of LAAs from an underlying distribution to a simulated model of the physical system. However, power grid design philosophies, such as $N-1$ scheduling, make the grid resilient to various contingencies (including cyber attacks). This implies component disconnections induced by LAAs can be extremely rare. Consequently, Monte Carlo sampling can be computationally expensive, as the majority of sampled LAAs will not result in component disconnections, requiring a large number of realisations applied to the power system model to generate a sample of the rare event (i.e., the activation of an ER).
To avoid an exhaustive search to locate potential LAAs that lead to network failures, we instead employ a novel methodology in the context of LAAs based on a Markov chain Monte Carlo (MCMC) approach for rare-event sampling, known as the \emph{skipping sampler}. The proposed approach is designed to reduce the time and computational effort spent on evaluating LAAs that are of little interest (i.e. those that do not result in an emergency response), allowing it to more efficiently construct a sample of a rare event. The skipping sampler has been applied to draw samples of low probability, high impact events in power networks in the literature (see [12] and [13]). Section~\ref{sc:skipping_sampler} provides a detailed discussion on the skipping sampler.
We evaluate the framework by performing extensive simulations using the Kundur two-area system (KTAS) \cite{kundur1994power}. We simulate the power grid's transient dynamics by a third-order model, which accounts for both the frequency as well as the voltage dynamics \cite{schm2014}. The LAAs that
perturb these dynamics are modelled according to the \emph{Log-normal} distribution (since real-world cyber attack magnitudes are well modelled by this distribution \cite{Edwards2016}). If the local frequency metrics at any node exceed pre-set tolerances, appropriate ERs, such as generation/load shedding, are activated. Our results show that in the KTAS network, most instances of ERs are activated for two specific spatial distributions of LAAs, namely (i) when the attacker increases the load throughout the system or (ii) when the attacker decreases the load in the over-provisioned area (with excess generation) and increases the load in the under-provisioned area (with excess load). In particular, attacks that exacerbate the power imbalance in the system (i.e., type (ii)) can trigger inter-connector line disconnections, and lead to other network failures.
The rest of the paper is organised as follows. Section \ref{sec:Prelim} introduces the system model; Section~\ref{sc:statmodel} presents the statistical model and details of the rare event sampling approach.
Section~\ref{sec:Sims} describes the simulation results and Section~\ref{sec:Conc} concludes. The simulation parameters are provided in an online Appendix found in \cite{extended}.
\section{System model}\label{sec:Prelim}
\subsection{Power Grid Model}
Using the third-order model for the generator, the power system model for rare-event sampling simulations will also include a model for governor action, automatic voltage regulation and a model of protection system ERs if the frequency or RoCoF exceeds pre-defined thresholds. We consider a power grid represented by $\mathcal{G} = \{ \mathcal{N}, \mathcal{W} \},$ where $\mathcal{N}$ is the set of buses and $\mathcal{W}$ is the set of transmission lines. The set of buses $\mathcal{N}$ consists of $N$ generation buses and $L$ load buses, with $|\mathcal{N}| = N+L$. At each generation bus $i= 1, \dots N$, the dynamics for the phase angle $\delta_i$, voltage magnitude $E_i$ and governor action $\rho_i$ are given respectively by:
{\small\begin{subequations}
\label{eq:network_model}
\begin{empheq}[left = \empheqlbrace\,]{flalign}
& M(\psi) \ddot{\delta}_{i}+D \dot{\delta}_{i}=\psi_i \chi^G_i- \chi^L_i(\mathcal{R}_i)-\nonumber \\ &\hspace{3cm}E_{i}\sum_{j=1}^{N+L}B_{ij}(\Omega_{ij})E_{j}\sin(\delta_{ij} )\label{eq:freq21} \\
& S_{i}\dot{E}_{i}=\psi_i (E_{f,i}-\text{v}_i) -E_{i}+ X_{i}\sum_{j=1}^{N+L}B_{ij}(\Omega_{ij})E_{j}\cos(\delta_{ij} )\label{eq:volt21}\\
& \dot{\rho}_{i}= -A_{i}\dot{\delta}_{i}(1-1_{\mathcal{\mathcal{\mathcal{W}}}}[\dot{\delta}_{i}]).\label{eq:gov21}
\end{empheq}
\end{subequations}}
\noindent In a similar manner, the dynamics for $\delta_i$ and $E_i$ at each load bus $i=N+1, \dots, N+L$ are given by:
{\small \begin{subequations}
\label{eq:network_model_load}
\begin{empheq}[left = \empheqlbrace\,]{flalign}
& M(\psi) \ddot{\delta}_{i}+D \dot{\delta}_{i}= -\chi^L_i(\mathcal{R}_i)- E_{i}\sum_{j=1}^{N+L}B_{ij}(\Omega_{ij})E_{j}\sin(\delta_{ij})\label{eq:freq_load} \\
& S_{i}\dot{E}_{i}=\psi_i E_{f,i}-E_{i}+X_{i}\sum_{j=1}^{N+L}B_{ij}(\Omega_{ij})E_{j}\cos(\delta_{ij}. )\label{eq:volt_load}
\end{empheq}
\end{subequations}}
In equations~\eqref{eq:network_model} and ~\eqref{eq:network_model_load}, $\psi_i$, $\Omega_{ij}$ and $\mathcal{R}_i$ are indicator variables associated with generator, line and load disconnections respectively, explained in Section~\ref{sec:emer_res}. The system angular momentum, $M(\psi)= \sum_{j=1}^{N} \psi_j H_j$ is given by the sum of each generator's inertia constant, $H_i$ (see online Appendix for parameter values).
\begin{table}[!h]
\centering
\caption{Variables used in ~\eqref{eq:network_model} and ~\eqref{eq:network_model_load}. }
\begin{tabular}{|c|l|c|}
\hline
\textbf{Symbol}&\textbf{Meaning}&\textbf{Units}\\\hline
$A_i$&Governor's droop response &MW/rad\\ $B_{ij}(\Omega_{ij})$&Susceptance matrix&p.u.\\ $\chi_i^G$&Net generation at node $i$&p.u.\\
$\chi_i^L(\mathcal{R})$&Net loads at node $i$&p.u.\\
$D$&System damping& \%\\
$\delta_i$ & Phase angle & p.u\\
$\delta_{ij}$&$\delta_i - \delta_j$&p.u.\\
$\dot{\delta}_i$&Frequency & p.u\\
$\ddot{\delta}_i$&Rate of change of frequency (RoCoF)&p.u.\\
$E_i$ & Voltage&p.u.\\
$E_{f,i}$&Machine $i$ rotor field voltage&p.u.\\ $M(\psi)$&System angular momentum & Ws$^2$\\
$\Omega_{ij}$&Line disconnection indicator&-\\
$\psi_i$ & Generator shed indicator&-\\
$\mathcal{R}_i$&UFLS counter&-\\
$S_i$&Machine $i$ transient time constant&s\\
$X_i$&Machine $i$ equivalent reactance&ohms\\
$\mathcal{W}$&Governor's deadband frequency range&Hz\\
\hline
\end{tabular}
\label{tb:model_variables}
\end{table}
As we assume the network to be lossless, the elements of $B_{ij}(\Omega_{ij})$ correspond to the imaginary part of the elements of the network's admittance matrix \cite{kundur1994power}. The net generation at node $i$ is given by $\chi^G_i = \min\{P_i^{\text{max}},P^G_i + \rho_i\}$, where $P_i^{\text{max}}$ is the nominal maximum power output of generator $i$, $P_i^G$ is the equilibrium power of the generator and $\rho_i$ is the power contributed by a governor unit, whose dynamics are given in \eqref{eq:gov21}. The variable v$_i$ accounts for the action of automatic voltage regulation (see online Appendix). The net load at node $i$, $\chi^L_i$, is inclusive of the LAA and a load disconnection scheme, and is discussed in Sections~\ref{sec: laa_model} and \ref{sec:emer_res}. The remaining parameters are given in Table~\ref{tb:model_variables}.
\subsection{Load-Altering Attack Model}
\label{sec: laa_model}
Several security vulnerabilities have been identified in IoT-enabled high-wattage consumer appliances (see, e.g., \cite{Soltan2018}). These vulnerabilities can be exploited by a strategic attacker to cause security incidents such as information disclosure and privilege escalation, leading to a change in the device's operational settings (e.g., switch ON/OFF or change the mode of operation). Considering a $2~$kW power rating for the ACs, and a Botnet-scale attack that potentially compromises tens of thousands of such devices, LAAs can lead to a sudden load change of several MWs of power \cite{Soltan2018}.
Not all loads are expected to be susceptible to LAAs, thus we decompose $P^L_i$, the equilibrium load at bus~$i$, into a vulnerable part, given by $\nu P^L_i$, where $\nu \in [0,1]$ denotes the proportion of equilibrium loads in the network vulnerable to an LAA; and a secure part (i.e., protected or non-smart loads) $(1-\nu)P^L_i$. LAAs at node $i$, denoted $u_i$, are modelled as $u_i \coloneqq \eta_i\nu P^L_i$, where $\eta_i \in [-1,1]$ is the proportional change to equilibrium vulnerable load. Thus, the load at node $i$, inclusive of the LAA, is given by
{ \begin{align}
\chi^L_i &= (1-\nu)P^L_i + \nu P^L_i + u_i \nonumber\\
& = (1-\nu)P^L_i + (1+\eta_i)\nu P^L_i.
\end{align}}
This constrains the authority of the attacker to alter vulnerable loads at node $i$ to a minimum of $0$~MW (i.e. $\min(u_i) = -P^L_i \nu$), or, at maximum, double vulnerable load demand (i.e. $\max(u_i) = P^L_i \nu$). This restriction to the maximum LAA reflects the finite capacity of inactive loads the attacker can activate during an LAA.
\vspace{-0.2 cm}
\subsection{Emergency Responses}
\label{sec:emer_res}
ERs refer to systems designed to protect sensitive power system components from excessive frequency deviations following a change in the active power balance. In this section we provide a brief discussion of ER employed, and refer the reader to \cite{goodridge2020} and \cite{goodridge2021} for a detailed mathematical description.
\emph{ Generation shedding:} To protect synchronous generators, we model two independent schemes intended to disconnect the generator from the network: (i) RoCoF-induced generation shedding (RIGS) - the generator is disconnected when nodal RoCoF $|\ddot{\delta_i}|$ exceeds an upper threshold; (ii) over frequency generation shedding (OFGS) - generation is shed when nodal frequency $\dot{\delta_i}$ exceeds a pre-set upper limit. The binary variable $\psi_i$ models the activation of generation shedding at node $i$ during the simulation: $\psi_i = 1$ initially under normal operation; however, when either threshold is met and generator $i$ is disconnected, $\psi_i$ is set to 0 until the end of the simulation.
\emph{ Under-frequency load shedding (UFLS):} We model UFLS schemes as a progressive disconnection of loads when the frequency $\dot{\delta}_i$ falls below a strictly decreasing sequence of four frequency thresholds $F^U \coloneqq \{F^U_1,\dots,F^U_4\}$ where $F^U_{j-1} > F^U_j$. In our model, at each frequency threshold, 10\% of equilibrium loads $P^L_i$ is automatically disconnected to arrest the decline in nodal frequency. Letting $\mathcal{R}_i \in \{0,1,2,3,4\}$ count the total number of UFLS activations at node $i$ at each time step $t$ in the simulation, the \textit{net load} is a dynamic variable in the power system model:
{\small \begin{equation}
\chi^L_i(\mathcal{R}_i) = \bigg(1- 0.1\mathcal{R}_i \bigg)\bigg((1-\nu)P^L_i+ (1+\eta_i)\nu P^L_i\bigg)
\end{equation}
\emph{ Line disconnection:} In the KTAS network, we model the disconnection of the line connecting Areas 1 and 2 when the power flow through the line, given by $\phi_{ij} \coloneqq B_{ij}E_i E_j sin(\delta_i - \delta_j)$, exceeds a pre-set power threshold $P^{\phi}$. When excess power flow is detected through the inter-connector line, the indicator $\Omega_{ij}$ switches from 1 to 0 for the remainder of the simulation, setting the $ij^{th}$ element of $B_{ij}$ to 0 \cite{goodridge2021}.
The ER model inspects the continuous time variables $\dot{\delta}_i(t)$, $\ddot{\delta}_i(t)$ and $\phi_{ij}(t)$ from the power system model~\eqref{eq:network_model} at regular time intervals. Once a criteria for activation is observed, the corresponding ER is activated. This is represented in \eqref{eq:network_model} as a discontinuity, where changes to the relevant input variables (power injection, load or network topology) are applied. Subsequently, the simulation is resumed with the new network parameters.
\section{Statistical Model for LAAs}\label{sc:statmodel}
In this section, we present the statistical model for the distribution of LAAs and describe the proposed rare-event sampling approach to identify the impactful LAAs.
\subsection{Modeling the Unconditional Distribution of LAAs}
There are several studies that document the frequency and magnitude of cyber breaches in enterprise networks \cite{Edwards2016}. For instance, \cite{Edwards2016} demonstrates the size of data breaches can be well-modelled by the log-normal family of distributions. Assuming nodal LAAs magnitudes are independent and follow a similar distribution, we model $U\in \mathbb{R}^{N+L} \coloneqq[|u_1|, \dots, |u_{N+L}| ]$ as
\begin{equation}
\label{eq:rho}
U \sim \prod_{i=1}^{N+L} \text{Lognormal} (\mu_i,\sigma_i^2).
\end{equation}
We also note that our analysis is not restricted to the log-normal family of distributions, and can be extended to any underlying distribution in a straightforward manner.
\vspace{-0.2 cm}
\subsection{Rare-Event Sampler for LAAs}
\label{sc:skipping_sampler}
If given an unconditional density $\rho$ over $\mathbb{R}^{N+L}$ and a rare event of interest $C \subset \mathbb{R}^{N+L}$, \textit{rare-event sampling} involves drawing elements from $\rho$ conditioned on the occurrence $C$. {The density of this conditional distribution for the element $U \in \mathbb{R}^{N+L}$ is:
{ \begin{equation}
\pi(U) = \frac{\rho(U)\mathbbm{1}_C(U)}{\rho(C)},
\label{eq:pi}
\end{equation}}
where $\rho(C)$ is the probability of the event $C$ occurring, and
\begin{equation}
\mathbbm{1}_C(U) = \begin{cases} 1 & U \in C\\
0 & U \notin C. \end{cases}
\end{equation}
\noindent In the context of our research, $C$ is the set of LAA magnitudes which result in the activation of at least one ER and $\rho$ is the distribution in \eqref{eq:rho}. As $C$ is expected to be rare, we employ the \textit{skipping sampler} MCMC algorithm to efficiently draw samples of $U \in C $.} The skipping sampler is formalised in Algorithm~\ref{alg:skip}. We provide an intuitive explanation of the algorithm in the following.
{\small \begin{algorithm}[!ht]
\SetAlgoLined
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\textbf{Input:} initial state $U_1$;\\
\textbf{for i = 1 to n do:}
\BlankLine
\DontPrintSemicolon
Generate an initial proposal $\ensuremath{Z_1}$ distributed according to the density $q(\ensuremath{y} - \ensuremath{U_i})d\ensuremath{y}$;\\
Calculate the direction $\ensuremath{\Phi}=\left( \ensuremath{Z_1}-\ensuremath{U_i}\right)/\left\Vert \ensuremath{Z_1}-\ensuremath{U_i}\right\Vert $;\\
Generate a halting index $K \sim K_{\varphi}$;\\
Set $k=1$ and:\\
\While{$\ensuremath{Z}_k \notin C \textbf{\textup{ and }}$ $k< K$}{
Generate a distance increment $R$ distributed according to $q_{r|\ensuremath{\Phi}}\left(r|\ensuremath{\Phi}\right)$;\\
Set $\ensuremath{Z}_{k+1}=\ensuremath{Z}_k+ \ensuremath{\Phi} R$;\\
k=k+1;\\
}
Set $\ensuremath{Z}\coloneqq \ensuremath{Z}_k$;\\
Evaluate the acceptance probability:
\begin{equation}\label{eq:ap2}
\alpha(\ensuremath{U_i},\ensuremath{Z})=
\begin{cases}
\min\left(1, \frac{\pi(\ensuremath{Z})}{\pi(\ensuremath{U_i})} \right) & \text{ if } \pi(\ensuremath{U_i}) \neq 0, \\
1, & \text{ otherwise, }
\end{cases}
\end{equation}
Generate a uniform random variable $V$ on $(0,1)$;\\
\uIf{$V \leq \alpha(\ensuremath{U_i},\ensuremath{Z})$}{$\ensuremath{U}_{i+1}=\ensuremath{Z}$;}
\Else{$\ensuremath{U}_{i+1}=\ensuremath{U_i}$;}
\KwRet{$\ensuremath{U}_{i+1}$}.\;
\textbf{Output}: final sample $[U_1, \dots, U_h]$
\caption{Skipping sampler algorithm}
\label{alg:skip}
\end{algorithm}}
As a Metropolis-class algorithm \cite{Moriarty2019}, the skipping sampler can be understood as a two step procedure: (1) a \textit{proposal step} where, starting from a state $U_n \in \mathbb{R}^{N+L}$, a potential new state $Z$ for the final sample is generated; (2) an \textit{acceptance/rejection} step, which determines whether the proposed state $Z$ is included in the final sample, according to a specified acceptance probability. This ensures the distribution of the sample follows the desired target distribution $\pi$. If it is accepted, the proposal is included in the final sample and becomes the starting state for the next proposal step. This procedure is repeated a desired number of times, after which the final sample is returned.
The skipping sampler improves the sampling of $C$ by using a specialised proposal function which `skips' over $C^c$ - the set of LAAs which do not lead to the activation of an ER (which are not of interest); until the rare event is sampled or the skipping process is terminated. Thus, the proposal of the skipping sampler was designed to enable efficient transitions between connected components of $C$. Denoting the current state $U_n$, if the initial proposal $Z_1 \notin C$, we update the initial proposal (or `skip') by an adding an independent random distance increment $R_2$ in the direction $\Phi = \frac{Z_1 - U_n}{|Z_1 -U_n|}$, where $R_k$ has the conditional distribution of $\left\Vert Z_1-U_n\right\Vert$ conditioned on $\Phi$. The proposal function continues this linear update procedure until either $C$ is entered or the budget for skipping is exhausted \cite{Moriarty2019}.
%
%
\section{Simulations}\label{sec:Sims}
\subsection{Simulation Settings}
\label{sc:ktas}
Our case study is based on the KTAS power grid \cite{kundur1994power}. We take a Kron reduced version consisting of $N=4$ generation buses and $L=2$ load buses as shown in Figure \ref{fg:kundurbat}. At $t=0^-$, the system is modelled in equilibrium, with power flows from Area 1 to Area 2 through the line connecting nodes $5$ and $6$ (tie line). The system parameters are such that the system is $N-1$ secure, in the sense that the loss of a generator (in the absence of any other disturbance) does not trigger an ER. The initial conditions of the above system of equations, denoted $\delta_i\left(0\right)$, $E_i\left(0\right)$, $\rho_i(0)$, $P_i^G$ and $P^L_i$ are set equal to equilibrium states which can be determined numerically, such that $\ddot{\delta}_i \approx 0$. These values along the those of the parameters of ~\eqref{eq:network_model} and \eqref{eq:network_model_load} can be found in the online Appendix \cite{extended}.
%
\begin{figure}[tbph]
\vspace{-0.2 cm}
\centering\includegraphics[width=2.5in]{kundur4.pdf}
\caption{{Schematic drawing of the Kundur two-area 4 node network after Kron reduction. Generator buses (green circles) correspond to nodes $i=[1,\dots,4]$ and load buses (brown circles) correspond to nodes $i=5,6$. Line lengths are indicated.}}
\label{fg:kundurbat}
\vspace{-0.1 cm}
\end{figure}
To generate LAA instances, we implement the skipping sampler proposed in Section~\ref{sc:skipping_sampler}. During each proposal step, we sample an $N+L-$dimension LAA vector $U$ from Lognormal distribution as in \eqref{eq:rho}. We investigated various values for $\sigma_i \in [1, 8]$ which controls the rareness of large LAAs. For this study, we present the results for $\mu_i = 0$ and $\sigma_i =4$ for $i = 1,\dots, N+L$, and reserve a detailed sensitivity analysis for a dedicated manuscript. We apply $U$ as an input to the power system model~\eqref{eq:network_model} at $t = 0$, with frequency dynamics simulated for $15$ seconds following the LAA using MATLAB. We conduct $n=60,000$ proposals, which generated a final sample of $h \approx 10,500$ LAAs conditioned on the activation of at least one ER. This is an acceptance rate of 17.5\%, within the 15 - 48\% rate considered optimal for exploring of a sample \cite{Gamerman2006}. Following \cite{Moriarty2019}, no burn-in nor thinning was required, thus all samples collected were available for analysis. Since these responses are an undesirable event (from a system operator's point of view), we label such instances as a ``network failure".
\vspace{-0.2 cm}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{Nodal_LAA_Distribution2
\caption{{Distribution of the absolute value LAAs at each node of the KTAS network conditioned on a network failure event, for different levels of network vulnerability $\nu$ and the unconditional distribution ($\rho$). Note - bins for successive values of $\nu$ are offset slightly to the right to improve readability.}}
\label{fig:LAA_Distribution}
\vspace{-0.1 cm}
\end{figure}
\subsection{Simulation Results}
We evaluate the susceptibility of the network to LAAs at both a local and global scope for three regimes of network vulnerability to LAAs- a `secure network' ($\nu \le 45\%)$, a `moderately vulnerable network' ($45\% < \nu \le 65\%)$ and a `highly vulnerable network' ($\nu > 65\%)$.
{\bf Local Analysis -- Identifying the vulnerable nodes:}
Figure~\ref{fig:LAA_Distribution} plots the distribution of LAA magnitudes at each node, conditioned on the occurrence of a network failure, for different degrees of network vulnerability $\nu$. We observe the following : (i)
only at nodes $5$ and $6$ does the conditional distribution of LAA magnitudes differ significantly from $\rho$, with network failures associated with larger magnitude LAAs, located in the low density region of $\rho$. This is driven by the design of the KTAS network, where most loads are concentrated at nodes $5$ and $6$, giving the attacker sufficient leverage over system frequency to trigger a network failure. Thus, network failures are primarily driven by LAAs at these nodes, mostly independent of the LAA magnitude at nodes $1-4$; (ii) as network vulnerability $\nu$ increases, the distribution of LAA magnitudes at node 6 shifts rightwards, implying larger magnitude LAAs at node $6$ become more prevalent in the sample; however (iii) we note that large magnitude LAAs are rare under their assumed Log-normal distribution. Thus, Fig.~\ref{fig:LAA_Distribution} reveals network failures, driven by large LAA magnitudes at node $6$, are indeed low probability events. Summarily, these results suggest nodes~$5$ and $6$ are the critical nodes in the KTAS network from which the attacker can launch the most impactful attacks, regardless of the degree of network vulnerability.
\begin{figure}[!h]
\centering
\includegraphics[width=0.5\textwidth]{nodal_failures_vul.png}
\caption{{Average number of each ER per sample $\epsilon_i, i = 1,\dots,4$ (defined in \eqref{eqn:epsilon_def}) in the KTAS network for different degrees of network vulnerability to LAAs.}}
\label{fig:nodal_Failure}
\vspace{-0.1 cm}
\end{figure}
Figure~\ref{fig:nodal_Failure} illustrates the average number of activations of each ER per sample for different levels of network vulnerability, defined as
{\small\begin{align}
\epsilon_{f} = \frac{\text{Total ER$_f$ activations in the sample}}{\text{Sample size} }. \label{eqn:epsilon_def}
\end{align}}
Herein, the indices $f = 1,\dots,4$ correspond to
RoCoF-induced generation shedding (RIGS), over-frequency generator shedding (OFGS), UFLS activations, and inter-connector line trips respectively. Note that $\epsilon_{f} \in [0,1]$ for $f = 1,2,4$. However, $\epsilon_3 \in [0,4]$, as each node can experience a maximum of 4 UFLS events ($f=3$) during the simulation (see Section~\ref{sec:emer_res}).
For secure networks, i.e, when $\nu \le 45\%$, the average number of RIGS is comparatively small, as low-magnitude LAAs are unable to induce sufficiently large RoCoF deviations to trigger generator disconnection. Additionally, we observe all generator nodes experience OFGS and UFLS events at similar rates of $\epsilon_2 \approx 0.1$ and $\epsilon_3 \approx 1$ respectively. Together, these imply for a secure KTAS network, each class of ER is similarly likely at any node in the network. Thus, to counter the potential threat of LAAs in a secure KTAS network will require a system-wide solution, e.g. - a coordinated automatic generation control (AGC) system.
As $\nu$ increases, it grants the attacker greater authority over network loads and the ability to significantly disrupt the active power balance of the network. This is associated with increased rates of RIGS, UFLS and line disconnections, as these are triggered by large power deviations. However, when $\nu > 65\%$, the average number of OFGS responses declines, as large changes in loads result in RIGS dominating generation shedding events.
In contrast to secure networks, where nodes are similarly vulnerable to UFLS events under LAAs, Figure~\ref{fig:nodal_Failure} reveals a differential in the nodal risk of an UFLS event in moderately and highly vulnerable networks. For example, when $\nu > 65\%$, nodes $3, 4$ and $6$ experience approximately twice as many UFLS events as nodes $1, 2$ and $5$. This can be best understood through a global analysis of the KTAS network, which is discussed next.
{\bf Global Analysis - Spatial Distribution of LAAs:}
Recall that the KTAS network is comprised of two areas- Area 1 (nodes 1, 2 and 5) with excess generation, and Area 2 (nodes 3, 4 and 6) with excess demand. With sufficient authority over network loads, an attacker can exploit the global pre-LAA power imbalances of each area to trigger an ER by decreasing loads in Area 1 to exacerbate the excess generation, and increasing loads in Area 2 to exacerbate the generation deficit. Thus, as $\nu$ increases, we observe fewer UFLS responses in Area 1, more UFLS responses in Area 2, and a substantial increase in power transferred across the inter-connector, increasing the rates of line disconnections (Figure~\ref{fig:nodal_Failure}).
The global susceptibility of the KTAS network to LAAs is also illustrated by Figure~\ref{fig:Area_Failure}, which plots the probability of the relative changes in loads in each Area of the KTAS network, conditioned on the occurrence of a network failure. First, we note the KTAS network is secure against global reduction of loads in both Areas, as it is likely governors are able to adjust generator output to handle the loss of loads in all vulnerability regimes. Instead, we observe that failure events occur primarily for large increases in loads, dominated by two scenarios: (i) when the LAA increases the demand in both Areas- this scenario is the primary driver of failure events when KTAS is secure, as the attacker increases loads at both nodes 5 and 6 beyond the maximum power capabilities of generators, triggering a failure anywhere in the network; (ii) the attacker decreases loads in Area~1 and increases them in Area~2- this scenario is rare when the KTAS is secure, as the attacker generally lacks sufficient authority to exploit the imbalances of the KTAS network. However, as $\nu$ (and the attacker's authority) increases, so too does the conditional probability of scenario (ii) (see Figure~\ref{fig:nodal_Failure}).
\begin{figure}[!t]
\hspace{-1cm}
\includegraphics[width=0.55\textwidth]{area_load_attacks.png}
\caption{{Probability of network failure for increase/decrease of system load in each area for the KTAS network.}}
\label{fig:Area_Failure}
\vspace{-0.1 cm}
\end{figure}
\section{Conclusions and Future Research}
\label{sec:Conc}
In this work, we present a framework to evaluate the impact of LAAs on power grid operations using a rare-event sampling approach.
The proposed approach provides a comprehensive framework to examine the impact of LAAs under different potential spatial distributions of LAAs across the power network. Our results identify the nodes from which the attacker can launch the most impactful LAA and further illustrate how the attacker can exploit the inter-area power imbalances in the network to trigger ER events.
Future work includes (i) an extension of the analysis to dynamic LAAs against power grids \cite{AminiLAA2018}, (ii) considering correlations in LAA injections, (iii) showing scalability of the proposed method to large-dimensional systems.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\begin{figure*}
\centering \includegraphics[scale=1]{fig1.pdf}
\caption{\textbf{Mechanisms of hydrodynamic Coulomb drag}. (a) Schematic representation of the phenomenon under study: a liquid flow induces an electronic current in a solid along which it flows. (b)-(c) Schematic of the two limiting mechanisms of current generation. In (b), electrons (full lines) are driven directly by Coulomb interactions (dashed line) with liquid charge fluctuations (hydrons, wavy line). In (c) electrons are driven by phonons (wavy line), that are excited through hydrodynamic friction with the liquid. }
\label{fig1}
\end{figure*}
New functionalities in nanoscale fluid transport have been achieved by exploiting analogies with condensed matter phenomena. The analogy between surface charge in a nanochannel and doping in a semiconductor~\cite{Bocquet2010} has led to the development of nanofluidic diodes~\cite{Vlassiouk2007,Picallo2013}, and transistors~\cite{Karnik2006}; the similarity between ionic and electronic Coulomb interactions~\cite{Kavokine2021} allows for ionic Coulomb blockade~\cite{Kavokine2019,Tanaka2017,Feng2016a,Chen2021}. More recently, it has been suggested that -- beyond mere analogies --, nanofluidic transport can directly couple to electronic effects within the channel wall. For instance, it has been predicted that a quantum contribution to hydrodynamic friction results from the interaction of charge fluctuations in the liquid with electronic excitations in the solid~\cite{Kavokine2022}.
A more straightforward example of such a liquid-electron coupling is apparently provided by the numerous observations of a liquid-flow-induced current (or voltage drop) within a solid wall~ \cite{Ghosh2003, Newaz2012, HoLee2013, Yin2014, Yang2018, Rabinowitz2020, Marcotte2022}. The mechanisms proposed to explain the current generation include the buildup of a streaming potential~\cite{Cohen2003,Newaz2012}, charging/discharging of a pseudo capacitance~\cite{Yin2014,Yang2018,Park2017}, or adsorbed/desorbed ion hopping~\cite{Persson2004, Dhiman2011}. In all these cases, the liquid is in fact simply acting on the solid as an average external potential. Yet, some of the most recent experimental results~\cite{HoLee2013,Rabinowitz2020,Marcotte2022} cannot be explained by the above-mentioned mechanisms, and therefore suggest to examine the possibility of "intrinsic" current generation through a molecular-scale liquid-electron interaction, which would be analogous to the Coulomb drag phenomenon in condensed matter physics (Fig. \ref{fig1}a).
In Coulomb drag, an electric current driven through a conductor induces a current in a nearby -- yet electrically insulated -- conductor. This is due to charge fluctuations in the driven conductor creating particle-hole excitations with non-zero momentum in the passive conductor, which results in a current, provided that the conductors are not particle-hole symmetric~\cite{Narozhny2016}. A similar process could occur at the interface between a solid and a flowing liquid: excitations in the solid would then be generated through Coulomb interactions with (collective) charge fluctuations in the liquid, that we shall in the following call \emph{hydrons} (Fig. \ref{fig1}b). To our knowledge, this qualitative mechanism still lacks a theoretical description.
However, an alternative Coulomb drag mechanism, where the liquid-electron interaction is mediated by the solid's acoustic phonons (in short, \emph{phonon drag}, see Fig. \ref{fig1}c), has been first proposed by Kr\`al and Shapiro, and formalized in a Boltzmann equation framework~\cite{Kral2001}. Phonon drag has been invoked, for instance, to account for the experimental results of ref.~\cite{HoLee2013}.
In this paper, we develop a non-equilibrium theory of electronic current generation at a solid-liquid interface, that includes both types of solid-liquid interactions, allowing us to compare their relative importance and unveil their interplay. In order to do so, we make use of the Keldysh framework of perturbation theory to study the solid-liquid system under the non-equilibrium conditions of flow~\cite{Kavokine2022}. Our approach is thus free of the approximations involved in the more traditional Boltzmann transport formalism, allowing us, if needed, to account for electron-electron interactions. Strikingly, we find that the current generation triggers a \emph{quantum feedback} mechanism at the solid-liquid interface, that reduces the total hydrodynamic friction.
The paper is organized as follows. In Sec. II we present our model and state the main results. In Sec. III, the formal derivation is carried out. The reader interested only in the physical outcomes may skip directly to Sec. IV, where we evaluate explicitly the Coulomb drag current. In Sec. V, we derive the hydrodynamic friction renormalization resulting from the current generation; finally Sec. VI establishes our conclusions.
\paragraph*{Units and conventions.} We set the Boltzmann constant $k_{\rm B}=1$ (that is, we express the temperature in energy units), but otherwise use SI units throughout the text. Matrices are denoted with bold capital letters. We use the following convention for the $d$-dimensional Fourier transform:
\beqa
\hat{F}(\tb{q},\omega)&=&\int \dd^d\tb{r}\dd t\, F(\tb{r},t)e^{-i\tb{q}\cdot\tb{r}+i\omega t},\nonumber\\
F(\tb{r},t)&=&\int \frac{\dd^d\tb{q}\dd\omega}{(2\pi)^{d+1}}\, \hat{F}(\tb{q},\omega)e^{i\tb{q}\cdot\tb{r}-i\omega t}.\nonumber
\eeqa
\section{Model and main results}
We consider a two-dimensional solid occupying the plane $z=0$, in contact with a semi-infinite liquid occupying the half-space $z>0$, and flowing along the $x$ direction with a velocity $\tb{v}_{\rm \ell}$, as depicted in Fig. \ref{fig1}a. The system is at temperature $T = 300~\rm K$ (or 26 meV). The flow field $\tb{v}_{\rm \ell}$ is assumed uniform in the interfacial liquid layer~\cite{Kavokine2022}. The liquid interacts with the solid through Coulomb forces. The corresponding \emph{electron-hydron} Hamiltonian is
\beq\label{Heh}
\mc{H}_{\rm h/e}(t) = \int\dd\tb{r}\dd\tb{r}_{\rm e}\, n_{\ell}(\tb{r}-\tb{v}_{\ell} t,t)V^{\rm C}(\tb{r}-\tb{r}_{\rm e})n_{\rm e}(\tb{r}_{\rm e},t),
\eeq
where $n_{\ell}$ and $n_{\rm e}$ are the liquid and solid charge density, respectively, and $V^{\rm C}(\r) = e^2/(4\pi \epsilon_0 r)$ is the Coulomb potential. Following~\cite{Kavokine2022}, we treat $n_{\ell}$ as a free bosonic field, whose correlation functions are related to the liquid's dielectric response. An additional liquid-solid interaction originates from short-range repulsion forces which result in "classical" hydrodynamic friction~\cite{Kavokine2022,Bocquet2007}. We will assume that this hydrodynamic friction transfers momentum to the solid's acoustic phonons~\cite{Kral2001}: those phonons then acquire a non-zero average momentum. This effect can be modeled by adding a shift $\v_{\rm ph}$ to the phonon (sound) velocity. The electron-phonon interaction Hamiltonian is then of the form
\beq\label{Heph}
\mc{H}_{\rm ph/e}(t) = \int\dd\tb{r}\dd\tb{r}_{\rm e}\, \phi_{\rm ph}(\r- \r_{\rm e} - \v_{\rm ph} t) n_{\rm e}(\tb{r}_{\rm e},t).
\eeq
Here, $\phi_{\rm ph}$ is proportional to the local lattice displacement; its exact expression depends on the particular solid under consideration. The "phonon wind" velocity $\v_{\rm ph}$ is not known \emph{a priori} and it will be determined self-consistently by establishing the system's momentum balance (see Sec. \ref{wind}).
The treatment of this model within non-equilibrium perturbation theory allows us to obtain two key analytical results. First, we obtain an explicit expression for the electronic current density $\mathbf{j}$ induced in the solid by the liquid flow for each electronic band:
\begin{equation}\label{current}
\langle\tb{j} \rangle = \frac{e\hbar}{T} \int_{-\infty}^{+\infty}\frac{ \d\omega}{2 \pi}\int_0^{+\infty} \frac{\d q }{2 \pi} \, \frac{q^2 \left(\nabla_{q} \xi_q \right)\v_{\rm e}(q)}{\tn{cosh}^2\left[\frac{\hbar(\omega + \xi_q)}{2T}\right]} \frac{\tau_q }{1+ \tau_q^2 \omega^2}.
\end{equation}
Here $\tau_q$ is the lifetime of the quasiparticle at energy $\hbar\xi_\mathbf{q}=u_\mathbf{q}-\mu$ where $u_\mathbf{q}$ is the band dispersion and $\mu$ is the chemical potential. $\v_e$ is the electron "wind velocity", which is a linear combination of the phonon wind velocity $\v_{\rm ph}$ and the hydron wind velocity (or simply, liquid flow velocity) $\v_{\ell}$:
\beq\label{ve0}
\v_{\rm e}(q)=\frac{\tau_\mathbf{q}}{\tau^{\rm ph/e}_\mathbf{q}}\tb{v}_{\rm ph}+\frac{\tau_\mathbf{q}}{\tau^{\rm h/e}_\mathbf{q}}\tb{v}_{\rm \ell}.
\eeq
$1/\tau^{\rm ph/e}_\mathbf{q}$ and $1/\tau^{\rm h/e}_\mathbf{q}$ are the phonon and hydron contributions to the total quasiparticle scattering rate $1/\tau_{\mathbf{q}}$. This result is valid under a few reasonable assumptions on the electronic self-energy (see Sec. III), and as long as the electronic structure has no band crossings close to the Fermi level.
Second, we predict a reduction of the hydrodynamic friction $\lambda$ coefficient due to the current generation. We recall that the solid-liquid friction force is given by $\mathbf{F} = - \lambda \mathcal{A} \v_{\ell}$, where $\cal A$ is the surface area. Accounting for the current generation, $\lambda$ is modified from its "bare" value $\lambda_0$ according to
\beq\label{lambda}
\lambda = \frac{\lambda_0}{1+(\lambda_{\rm h/ph} + \frac{\tau}{\tau^{\rm ph/e} }\lambda_{\rm h/e}) / \lambda_{\rm um} }.
\eeq
Here, $\lambda_{\rm h/ph}$ and $\lambda_{\rm h/e}$ are the phononic and electronic contributions to the fluctuation-induced solid-liquid friction~\cite{Kavokine2022}; $\lambda_{\rm um} = 3 \zeta(3) T^3/(2\pi \hbar^2 c^4 \tau_{\rm um})$ has the dimension of a friction coefficient, and is expressed in terms of the sound velocity $c$ in the solid, and the typical phonon lifetime $\tau_{\rm um}$.
We demonstrate that this friction reduction is a quantum effect, that takes its roots in the solid's electronic excitations.
These results are derived in detail in the following sections. In Sec. IV, we evaluate the flow-induced current for different material systems, and successfully compare our predictions with experimental data~\cite{Marcotte2022}. In Sec. V, we show that the correction to hydrodynamic friction in Eq.~\eqref{lambda} can be non-negligible, and leads to significant hydrodynamic slippage in systems where it would not typically be expected.
\section{Non-equilibrium perturbation theory}
\subsection{Description in the Keldysh framework}
We describe the system's dynamics in terms of three types of real-time Green's functions: the Retarded, Advanced and Keldysh Green's functions, defined, for both bosons and fermions, according to
\beqa\label{definition_Green}
\left\{ \begin{array}{l}
G^{\rm R}(\tb{r},t,\tb{r}',t')= -i\theta(t-t')\langle[\psi(\tb{r},t),\psi^\dagger(\tb{r}',t')]_s\rangle,\\
G^{\rm A}(\tb{r},t,\tb{r}',t')= i\theta(t'-t)\langle[\psi(\tb{r},t),\psi^\dagger(\tb{r}',t')]_s\rangle,\\
G^{\rm K}(\tb{r},t,\tb{r}',t')=-i\langle[\psi(\tb{r},t),\psi^\dagger(\tb{r}',t')]_{-s}\rangle,
\end{array}\right.
\eeqa
where $\psi^{\dagger}$ and $\psi$ are the particles' creation and annihilation operators, and $[A,B]_\pm=AB\pm BA$, $s$ being $+$ for fermions and $-$ for bosons. The Retarded and Advanced Green's functions contain information on the system's elementary excitations. For non-interacting electronic quasiparticles in a translationally-invariant system at equilibrium, the Fourier-transformed Green's functions are
\beq
G^{\rm R,A}_0(\tb{q},\omega)=\frac{1}{\omega-\xi_\tb{q} \pm i0^+}
\label{GReq}
\eeq
where $\hbar\xi_\tb{q}=u_\tb{q}-\mu$ is the quasiparticle energy: $u_\tb{q}$ is the band dispersion and $\mu$ is the chemical potential. The Keldysh Green's function contains information on the quasiparticle distribution. At equilibrium, it satisfies the fluctuation-dissipation theorem:
\beq
G^{\rm K}(\tb{q},\omega)=\frac{2 i}{f(\omega)}\tn{Im}[G^{\rm R}(\tb{q},\omega)]
\label{GKeq}
\eeq
where $f(\omega)=\tn{cotanh}\left(\frac{\hbar\omega}{2T}\right)$. Given its importance for the subsequent discussion, we recall the derivation of this result in Appendix \ref{appendix_FDT}. For non-interacting electrons, Eqs.~\eqref{GReq} and \eqref{GKeq} yield
\beq
G^{\rm K}(\tb{q},\omega)=(2n_{\rm F}(\omega) -1) \times 2 i \pi \delta(\omega- \xi_{\mathbf{q}}),
\eeq
where we recover indeed the Fermi-Dirac distribution $n_{\rm F}(\omega) = 1/(e^{\hbar\omega/T} + 1)$.
We use the letter $D$ to denote bosonic Green's functions. For free bosons (such as phonons) with dispersion $\omega_{\mathbf{q}}$ at equilibrium,
\beq
D^{\rm R,A} = \frac{\omega_\mathbf{q}}{(\omega\pm i0^+)^2-\omega_\mathbf{q}^2}.
\eeq
The bosonic fluctuation-dissipation theorem reads (see Appendix \ref{appendix_FDT})
\beq\label{DFD}
D^{\rm K}(\tb{q},\omega)=2 if(\omega)\tn{Im}[D^{\rm R}(\tb{q},\omega)].
\eeq
so that, at equilibrium
\beq
D^{\rm K} = (2n_{\rm B}(\omega) +1) \times i \pi \left[\delta(\omega- \omega_\mathbf{q})-\delta(\omega+ \omega_\mathbf{q})\right],
\eeq
where we recover the Bose-Einstein distribution $n_{\rm B}(\omega) = 1/(e^{\hbar\omega/T} - 1)$.
The Keldysh Green's functions are therefore the analogues of the occupation distribution functions in the approximate Boltzmann formalism. They will be key in determining the non-equilibrium state of the system. Indeed, as shown in Appendix \ref{appendix_current}, the current density (within one electronic band) is given by
\beq
\langle \mathbf{j} \rangle = 2i e \int \frac{\d \mathbf{q} \d \omega}{(2\pi)^3} (\nabla_{\mathbf{q}} \xi_{\mathbf{q}}) G^{\rm K} (\mathbf{q},\omega).
\label{current0}
\eeq
In addition, the non-equilibrium density-density response function, which will be required for obtaining the correction to the hydrodynamic friction coefficient, can be computed starting from the non-equilibrium Green's functions (see Appendix \ref{appendix_susceptibility}).
\begin{figure*}
\centering \includegraphics[width=0.8\textwidth]{fig2.pdf}
\caption{\textbf{Non-equilibrium diagrammatic expansion.} (a) Dyson equation for the matrix Green's function $\tb{G}$ (thick line). The thin line is the bare Green's function $\tb{G}^0$. (b) First order self-energy diagram, which is computed for each of the electron-boson interactions. (c) Dyson equation for the random phase approximation (RPA) screening of the boson propagator (dashed line). (d) Dyson equation for the boson-electron vertex, within RPA. (e) Diagrammatic expansion for the electric current, which is related to the Keldysh Green's function at equal point in space and time. The square represents a gradient operator (multiplication by $\mathbf{q}$ in momentum space). (f) Notation for the electron and boson Green's functions. We have absorbed the RPA renormalisation of the vertices into the definition of the bosonic propagator. }
\label{fig2}
\end{figure*}
\subsection{Dyson equation}
Our task is now to compute the non-equilibrium Green's functions in the presence of the perturbations $\mc{H}_{\rm h/e}(t)$ and $\mc{H}_{\rm ph/e}(t)$. In the Keldysh formalism, we consider the matrix Green's function
\beqa\label{definition_Green0}
\tb{G}=\left( \begin{array}{cc}
G^{\rm R} & G^{\rm K}\\
0 & G^{\rm A}
\end{array}\right).
\eeqa
The perturbation series may be partially resummed by introducing a (matrix) self-energy $\bf \Sigma$. The Green's function then satisfies the non-equilibrium Dyson equation
\beq
\bf G = G_0 + G_0 \otimes \boldsymbol{\Sigma} \otimes G,
\label{dyson}
\eeq
which is represented diagrammatically in Fig. \ref{fig2}a. Here, $\otimes$ represents convolution in space and time, as well as matrix multiplication. We assume that the system is translationally invariant parallel to the interface, and that it has reached a steady state: we may then Fourier-transform Eq.~\eqref{dyson}. With the convolutions becoming products in Fourier space, and using that $G_0^{\rm A}(\mathbf{q},\omega) = G_0^{\rm R} (\mathbf{q},\omega)^*$ and $\Sigma^{\rm A}(\mathbf{q},\omega) = \Sigma^{\rm R} (\mathbf{q},\omega)^*$, we obtain
\beqa\label{G2R}
G^{\rm R,A}(\mathbf{q},\omega) &=& \dfrac{G_0^{\rm R,A} - |G_0^{\rm R}|^2 \Sigma^{\rm A,R}}{\left|1- G_0^{\rm R} \Sigma^{\rm R}\right|^2},\\
G^{\rm K}(\mathbf{q},\omega) &=& \dfrac{G_0^{\rm K} + |G_0^{\rm R}|^2 \Sigma^{\rm K}}{\left|1- G_0^{\rm R} \Sigma^{\rm R}\right|^2}. \label{G2K}
\label{dysonGK}
\eeqa
Using Eqs. \eqref{GReq} and \eqref{GKeq} for the equilibrium Green's functions, we find that the first term in Eq.~\eqref{dysonGK} vanishes if the self-energy is non-zero. Then, recalling Eq.~\eqref{current0}, we obtain a first very general expression for the flow-induced electric current (within a given electronic band):
\beq
\langle \mathbf{j} \rangle = 2i e \int \frac{\d \mathbf{q} \d \omega}{(2\pi)^3} (\nabla_{\mathbf{q}} \xi_{\mathbf{q}}) \dfrac{|G_0^{\rm R}|^2 \Sigma^{\rm K}}{\left|1- G_0^{\rm R} \Sigma^{\rm R}\right|^2}.
\label{current2}
\eeq
We note that this expression is valid far from equilibrium, and that it allows for systematic inclusion of electron-electron interactions.
\subsection{Non-equilibrium self-energy}\label{SE}
In order to proceed, we need to evaluate the non-equilibrium self-energy $\boldsymbol{\Sigma}$. It contains contributions from both the electron-phonon and electron-hydron interaction. Neglecting diagrams where the phonon and hydron propagators cross, they may be computed separately: $\boldsymbol{\Sigma} = \boldsymbol{\Sigma}_{\rm h/e} + \boldsymbol{\Sigma}_{\rm ph/e}$. Furthermore, the two contributions are in fact formally identical, since both the phonons and the hydrons are free bosons coupled to the electrons. Therefore, we only need to compute a generic electron-boson self-energy, resulting from a perturbation of the form given in Eq.~\eqref{Heh}.
We will consider a single diagram for this self-energy, as shown in Fig. \ref{fig2}b; we verify that higher order diagrams are indeed negligible under most conditions (see Appendix \ref{appendix_higherorder}). We account for the electronic screening of the bosonic propagator and boson-electron interaction vertices within the random phase approximation (Fig. \ref{fig2}c-d). We will absorb the screened vertices into the definition of the full bosonic propagator, that we denote as $\tb{D}$. The electric current, which is expressed in terms of the Keldysh Green's function evaluated at equal points in space and time, can then be represented as a sum of "ice cone" diagrams (Fig. \ref{fig2}e). These diagrams are reminiscent of the Aslamazov-Larkin diagrams~\cite{Aslamasov1968} that typically represent Coulomb drag in condensed matter systems~\cite{Kamenev1995, Narozhny2012,Kamenev2009}. Their evaluation typically involves the computation of a non-linear susceptibility (triangle diagram). In our case, this complication can be avoided, as the non-equilibrium self-energy can be readily evaluated.
Using the Keldysh formalism Feynman rules for the boson-fermion interaction~\cite{Rammer2007} (see Appendix \ref{appendix_selfenergy}) we obtain the components of the self-energy diagram in Fig. \ref{fig2}b as:
\begin{widetext}
\beqa\label{SR}
\Sigma^{\rm R,A}(\mathbf{q},\omega) &=& - \int \frac{\d \mathbf{q}' \d \omega'}{(2\pi)^3} \mathcal{M}(\mathbf{q}-\mathbf{q}',\mathbf{q}) \left( f(\omega' -\mathbf{q}'\!\cdot\! \v_{\rm b}) + \frac{1}{f(\xi_{\mathbf{q}-\mathbf{q}'})} \right)\frac{ \mathrm{Im} \, [ D^R(\mathbf{q}',\omega'-\mathbf{q}'\!\cdot\!\v_{\rm b})]}{\omega-\omega'- \xi_{\mathbf{q}-\mathbf{q}'} \pm i 0^+} , \\
\Sigma^{\rm K}(\mathbf{q},\omega) &=& 2i\pi \int \frac{\d \mathbf{q}' }{(2\pi)^3} \mathcal{M}(\mathbf{q}-\mathbf{q}',\mathbf{q})\left( 1+ \frac{f(\omega-\xi_{\mathbf{q}-\mathbf{q}'} - \mathbf{q}' \!\cdot\!\v_{\rm b}) }{f(\xi_{\mathbf{q}-\mathbf{q}'})} \right) \Im{D^R(\mathbf{q}',\omega-\xi_{\mathbf{q}-\mathbf{q}'}-\mathbf{q}'\!\cdot\!\v_{\rm b})},\label{SK}
\eeqa
and using $\mathrm{Im} \,[1/(\epsilon + i 0^+)] = - \pi \delta(\epsilon)$ yields
\begin{equation}\label{SR2}
\mathrm{Im} \left[\Sigma^{\rm R} (\mathbf{q},\omega)\right] = \pi \int \frac{\d \mathbf{q}' }{(2\pi)^3} \mathcal{M}(\mathbf{q}-\mathbf{q}',\mathbf{q})\left( f(\omega -\xi_{\mathbf{q}-\mathbf{q}'}-\mathbf{q}'\!\cdot\! \v_{\rm b}) + \frac{1}{f(\xi_{\mathbf{q}-\mathbf{q}'})} \right) \Im{D^R(\mathbf{q}',\omega-\xi_{\mathbf{q}-\mathbf{q}'}-\mathbf{q}'\!\cdot\!\v_{\rm b})} .
\end{equation}
\end{widetext}
Here $\v_{\rm b}$ is the boson wind velocity, and $\mathcal{M}(\mathbf{q}-\mathbf{q}',\mathbf{q}) \equiv | \langle \mathbf{q}-\mathbf{q}' | e^{-i\mathbf{q}' \r} | \mathbf{q} \rangle |^2$ are matrix elements computed between the electronic states $|\mathbf{q}\rangle$. The fluctuation-dissipation theorem in Eq. \eqref{GKeq} is therefore not satisfied for the non-equilibrium self-energy. Nevertheless, we may always express the Keldysh component in the form
\beq\label{SFD}
\Sigma^{\rm K}(\tb{q},\omega)=\frac{2 i}{f(\omega - \mathbf{q} \!\cdot\!\v_{\rm eff}(\mathbf{q},\omega))}\tn{Im}[\Sigma^{\rm R}(\tb{q},\omega)],
\eeq
which defines $\v_{\rm eff}$ as a frequency and momentum dependent \emph{effective wind velocity}.
\begin{figure*}
\centering \includegraphics[scale=1]{fig3.pdf}
\caption{\textbf{Non-equilibrium self energy.} (a) Imaginary part of the electron-boson self-energy (at the Fermi level) for the four electron-boson couples discussed in the text, as a function of the charge carrier (electron or hole) density. Electron-phonon and electron-hydron scattering occur at a similar rate in the 2DEG model, whereas in graphene electron-hydron scattering is much more efficient than electron-phonon scattering. (b). Same as in (a), but plotted as a function of frequency: $\omega = 0$ corresponds to the Fermi-level. The relatively weak variations within the thermal window (grey rectangle) justify the use of the impurity approximation. (c) Deviation of the effective electronic wind velocity (defined in the text) from the bosonic wind velocity, as a function of the latter's magnitude.
}\label{fig3}
\end{figure*}
One can now evaluate explicitly the self-energy. The calculations steps are detailed in Appendix B. In a few words, we assume {for simplicity} that the liquid is water, whose interfacial charge fluctuations are described in terms of a \emph{surface response function}. As demonstrated by the extensive molecular dynamics simulations of ref.~\cite{Kavokine2022}, this response function {for water} can be modelled as a sum of two Debye peaks, so that
\begin{equation}\label{Dwater}
D_{\rm w}^{\rm R}(\mathbf{q},\omega) = -\frac{1}{\hbar}\frac{V_q^C}{\epsilon(q)} \sum_{k = 1,2} \frac{f_k}{1- i \epsilon(q) \omega/\omega_{\mathrm{D},k} } ,
\end{equation}
where $V_q^C= e^2/(2\epsilon_0 q)$ is the Fourier-transformed Coulomb potential, $\omega_{\rm D, 1} \approx 1.5~\rm meV$ and $\omega_{\rm D,2} \approx 20~\rm meV$ are the Debye frequencies and $f_{1,2}$ are the corresponding oscillator strengths;
$\epsilon (q)$ is the RPA dielectric function of the electronic system, that accounts for the screening of the interaction vertices.
{These results can be extended to any other liquid using a relevant description of their response function}.
The acoustic phonon propagator can be written as \cite{Bruus2004}
\begin{equation}\label{Dphonon}
D_{\rm ph}^{\rm R}(\mathbf{q},\omega) = \frac{1}{\hbar}V_q^{\rm ph/e} \frac{\omega_{\mathbf{q}}^2}{(\omega+i0^+)^2-\omega_{\mathbf{q}}^2},
\end{equation}
with $\omega_{\mathbf{q}} = cq$, $c$ being the phonon velocity, and $V_{\mathbf{q}}^{\rm ph/e}$ is the material-dependent screened electron-phonon interaction.
For the description of the electronic system, we consider two different models: a two-dimensional electron gas (2DEG) with an effective mass $m$; and graphene, treated within the Dirac cone approximation, characterized by the constant Fermi velocity $v_{\rm F}$. We refer to Appendix B for the details on the associated electronic structure and dielectric properties.
The results of these calculations are summarized in Fig. \ref{fig3}.
Fig. \ref{fig3}a shows the imaginary part of the retarded self energy at the Fermi level, $\mathrm{Im} \left[\Sigma^{\rm R}(q = k_{\rm F},\omega = 0)\right]$, computed by numerical integration according to Eq. \eqref{SR2}, where we have separated the hydron and phonon contributions. This quantity represents the scattering rate of the low-energy electronic quasiparticles. The electron-hydron interaction, as well as the electron-phonon interaction in the 2DEG are essentially screened Coulomb interactions and they yield a similar order of magnitude for the associated scattering rate. On the other hand, the electron-phonon interaction in graphene has a peculiar form (see Appendix \ref{appendix_selfenergy2}), so that the corresponding scattering rate is 2 to 3 orders of magnitudes lower. This will have an importance for the global momentum balance discussed in Sec. \ref{wind}.
For the purpose of computing the Coulomb drag current, we need to determine how the electron-boson scattering affects the electronic distribution function, that varies typically over a scale $T$ around the Fermi level. If the boson energy is much smaller than $T$, we may approximate it as 0: the electrons then see a random static impurity potential. Within this \emph{impurity approximation}, we further assume
\beq
\left\{
\begin{array}{ll}
&\mathrm{Im} \, [\Sigma^{\rm R} (\mathbf{q},\omega)] \approx \mathrm{Im} \, [ \Sigma^{\rm R} (\mathbf{q},\xi_{\mathbf{q}})] \equiv - 1/ \tau_{\mathbf{q}}, \\
&\mathrm{Re} \, [\Sigma^{\rm R} (\mathbf{q},\omega)] \approx 0.
\end{array}
\right.
\eeq
Then, as detailed in Appendix \ref{appendix_selfenergy2}, the electron-boson scattering rate can be computed as
\begin{equation}\label{selfenergy2}
\frac{1}{\tau_{\mathbf{q}}} = \pi \frac{T}{\hbar^2} \int \frac{\d \mathbf{q}'}{(2\pi)^2} V_{\mathbf{q}-\mathbf{q}'} \delta(\xi_{\mathbf{q}} - \xi_{\mathbf{q}'}),
\end{equation}
where $V$ is the electron-boson interaction. Fig. \ref{fig3}b shows the frequency dependence of the scattering rate (imaginary part of the self-energy) at $q = k_{\rm F}$, at a fixed chemical potential $\mu = 100~\rm meV$ for both 2DEG and graphene.
One observes that the variation of the self-energy in a window of width $2T$ around zero frequency (grey rectangle in Fig. \ref{fig3}b) is relatively weak, justifying the use of the impurity approximation (dashed lines in Fig. \ref{fig3}b) in the following computations.
If one neglects the angular dependence the integrand in Eq. \eqref{selfenergy2}, the scattering rate assumes an intuitive Fermi golden rule form:
\begin{equation}\label{SEapprox}
\frac{1}{\tau_{\mathbf{q}}} \approx \pi \frac{T}{\hbar} V_{\mathbf{q}}N(u_\mathbf{q}),
\end{equation}
where $N(u_\mathbf{q})$ is the density of states at energy $u_\mathbf{q}$. Here, the quantity $ \hbar\omega_{\rm b}V_{\mathbf{q}}/\mc{A}$ plays the role of the squared matrix element, and $T/\hbar\omega_{\rm b}$ is the number of bosonic modes on which the electrons can scatter, $\hbar\omega_{\rm b}$ being the typical bosonic energy (see Appendix \ref{appendix_FGR} for a detailed derivation).
The simplified expression in Eq.~\eqref{SEapprox} allows us to understand the scalings observed in Figs.~\ref{fig3} a-b. The frequency dependence of the self-energy (Fig.~\ref{fig3}b) is roughly consistent with it being proportional to the density of states, which is independent of energy in the 2DEG ($N(u)= \theta(u)\times m/\hbar^2$) and proportional to the energy in graphene ($N(u)=\theta(u) \times 2u/(\pi\hbar)$, for the upper Dirac cone). In a 2DEG with reasonable electronic density, the screening length is much shorter than the Fermi wavelength, so that for $q\sim k_{\rm F}$, $V_{\mathbf{q}} \approx 1/(2 N(u_{\mathbf{q}})) $: we then obtain a "Planckian" scattering time $\tau_{\mathbf{q}} \approx 2 \hbar / \pi T$~\cite{Zaanen2004}. This result could be expected on dimensional grounds. Indeed, once the boson energy has been neglected, the temperature is the only energy scale in the problem. The electron-boson scattering rate evaluated at the Fermi momentum is then expected to be nearly independent of electronic density, as observed in Fig.~\ref{fig3}a. The situation is different in graphene, where the screened Coulomb potential scales like $1/q$ for $q\sim k_{\rm F}$ and the density of states $N(u_\mathbf{q}) \propto q$ since $u_{\mathbf{q}} = \hbar v_{\rm F} q$. Therefore, we expect again the self-energy to weakly depend on the momentum (hence, the electronic density) for the electron-hydron scattering. For the electron-phonon scattering, where the effective potential does not depend on $q$, we expect $\tau_{\mathbf{q}} \propto 1/q$, so that $\tau_{k_{\rm F}} \propto n^{-1/2}$, consistently with Fig.~\ref{fig3}a.
We now come back to the Keldysh component of the self-energy, which we evaluate numerically according to Eq. \eqref{SK}. We then compute the effective electronic velocity $\v_{\rm eff}$ defined in Eq. \eqref{SFD}, at $q = k_{\rm F}$ and $\omega = 0$. The deviation of $v_{\rm eff}$ from {the boson wind velocity} $v_{\rm b}$ is plotted in Fig. \ref{fig3}c as a function of $v_{\rm b}$,
for the different electron-boson couples. While the liquid flow velocity can be as low as $1~\rm \mu m/s$, we will find that phonon wind velocities can reach thousands of m/s. For this whole range of boson velocities $v_{\rm b}$ we find that $v_{\rm eff}$ remains within 5\% of $v_{\rm b}$, with a stronger deviation appearing only for $\v_{\rm b}$ in excess of $1~\rm km\cdot s^{-1}$. We may therefore safely assume $\v_{\rm eff} \approx \v_{\rm b}$, and evaluate $\Sigma^{\rm K}$ according to a \emph{quasi-equilibrium} fluctuation-dissipation theorem:
\beq
\Sigma^{\rm K}(\tb{q},\omega)=\frac{2 i}{f(\omega - \mathbf{q} \!\cdot\!\v_{\rm b} )}\tn{Im}[\Sigma^{\rm R}(\tb{q},\omega)],
\label{quasiFDT}
\eeq
which differs from the equilibrium version only by the frequency shift $\mathbf{q} \!\cdot\!\v_{\rm b}$. We note here the power of the Keldysh framework, which allows us to control the approximation leading to Eq.~\eqref{quasiFDT}, and potentially explore conditions where it no longer holds.
\subsection{Quasi-equilibrium state}
When Eq.~\eqref{quasiFDT} is satisfied for all electron-boson self-energies $\Sigma_j$, we will say that the system is in a \emph{quasi-equilibrium state}. Within the impurity approximation, we denote
$\tn{Im}[\Sigma_j]=-1/\tau^j_{\mathbf{q}}$; then, $\tau_{\mathbf{q}}^{-1}=\sum_j(\tau^j_{\mathbf{q}})^{-1}$ is the total electron scattering rate at the energy $\xi_{\mathbf{q}}$. Eqs. \eqref{G2R} and \eqref{G2K} for the non-equilibrium Green's functions now become:
\beqa\label{G3R}
G^{\rm R}(\mathbf{q},\omega) &=& \dfrac{\sum_j\tn{Re}[\Sigma_j] +i \tau^{-1}_\mathbf{q}}{(\omega-\xi_{\tb{q}})^2+ \tau^{-2}_\mathbf{q}}\\
G^{\rm K}(\mathbf{q},\omega) &=& \dfrac{2i}{f(\omega-\tb{q}\!\cdot\!\tb{v}_{\rm e}(q))}\dfrac{\tau^{-1}_\mathbf{q}}{(\omega-\xi_{\tb{q}})^2+ \tau^{-2}_\mathbf{q}} \label{G3K}
\eeqa
where the \emph{electron wind velocity} {defined as},
\beq\label{ve}
\v_{\rm e}(q)=\sum_j\frac{\tau_\mathbf{q}}{\tau^j_\mathbf{q}}\tb{v}_j,
\eeq
is a convex combination of the different boson wind velocities $\tb{v}_j$; Eq.~\eqref{G3K} is valid as long as these are small compared to the Fermi velocity. We note that we may include static impurities as an additional scatterer with zero velocity. Physically, each bosonic wind blows on the electrons through its electron-boson interaction. Each bosonic velocity contributes to the total wind velocity with a weight that is given by the corresponding electron-boson scattering rate.
We therefore find that the non-equilibrium Green's function satisfies the same quasi-equilibrium fluctuation-dissipation theorem as the individual self-energies:
\beq \label{GFD}
G^{\rm K}(\tb{q},\omega)=\frac{2i}{f(\omega-\tb{q}\!\cdot\!\tb{v}_{\rm e}(q))}\tn{Im}[G^{\rm R}(\tb{q},\omega)]
\eeq
Eqs.~\eqref{G3R} and \eqref{GFD} provide a complete picture of the non-equilibrium electronic state. Due to the scattering on the different bosons, the spectral function is broadened, and acquires a width $\tau_{\mathbf{q}}^{-1}$. In addition, the occupation of the broadened states undergoes a Doppler shift $\mathbf{q}\!\cdot\! \v_e$ with respect to the equilibrium Fermi-Dirac occupation. This becomes apparent if one evaluates the actual electronic density in energy-momentum space (see Appendix \ref{appendix_current}):
\beq\label{ne}
n_{\rm e}(\tb{q},\omega)=\frac{G^{\rm K}(\mathbf{q},\omega)}{2i}+\,\tn{Im}[G^{\rm R}(\tb{q},\omega)],
\eeq
which reduces to
\beq
n_{\rm e}(\tb{q},\omega)=n_F(\omega-\tb{q}\cdot\tb{v}_{\rm e}(q))\frac{2\tau}{1+\tau^2(\omega-\xi_{\tb{q}})^2}
\eeq
where $n_F(\omega)= 1/ (e^{\hbar\omega/T}+1)$ is the Fermi-Dirac distribution. The electrons appear to acquire an average velocity equal to the electron's wind velocity, which corresponds precisely to the electric current that we evaluate in the next section.
We make one last remark concerning the electron density-density response function. Starting from its expression in terms of the Green's functions we demonstrate in Appendix \ref{appendix_susceptibility} that it satisfies a quasi-equilibrium fluctuation-dissipation theorem as long as the Green's functions satisfy one:
\beq \label{chiFD}
\chi_{\rm e}^{\rm K}(\tb{q},\omega)=2if(\omega-\tb{q}\!\cdot\!\tb{v}_{\rm e}(q))\tn{Im}[\chi_{\rm e}^{\rm R}(\tb{q},\omega)].
\eeq
This result will be important for the evaluation of fluctuation-induced friction forces in the next section.
\section{Flow-induced electric current}
\subsection{General expression}
Starting from the formal expression in Eq. \eqref{current0}, and using the quasi-equilibrium Green's function in Eq. \eqref{G3K}, we immediately obtain, after angular integration and to first order in the wind velocity, an explicit expression for the flow-induced electronic current:
\begin{equation}\label{currentgeneral}
\langle\tb{j} \rangle = \frac{e\hbar}{T} \int_{-\infty}^{+\infty}\frac{ \d\omega}{2 \pi}\int_0^{+\infty} \frac{\d q }{2 \pi} \, \frac{q^2 \left(\nabla_{q} \xi_q \right)\v_{\rm e}(q)}{\tn{cosh}^2\left[\frac{\hbar(\omega + \xi_q)}{2T}\right]} \frac{\tau_q }{1+ \tau_q^2 \omega^2}.
\end{equation}
The wind velocity $\v_e(q)$ is given by Eq. \eqref{ve}. In the presence of electron-hydron, electron-phonon and electron-impurity scattering, it explicitly writes
\beq\label{ve2}
\v_{\rm e}(q)=\frac{\tau_\mathbf{q}}{\tau^{\rm ph/e}_\mathbf{q}}\tb{v}_{\rm ph}+\frac{\tau_\mathbf{q}}{\tau^{\rm h/e}_\mathbf{q}}\tb{v}_{\rm \ell}
\eeq
where $1/\tau_\mathbf{q}=1/\tau^{\rm ph/e}_\mathbf{q}+1/\tau^{\rm h/e}_\mathbf{q}+1/\tau^{\rm im}_\mathbf{q}$, the last term being the impurity scattering rate.
We thus obtain our first main result, as anticipated in Sec. II: we predict current generation in the solid due to the flow of a neutral liquid.
\subsection{Wind velocity : global momentum balance} \label{wind}
\begin{figure*}
\centering \includegraphics[scale=1]{fig4.pdf}
\caption{\textbf{Momentum balance.} (a) Diagram representing the momentum fluxes in the solid-liquid system, separated into four subsystems, represented by the rectangular boxes. The liquid (flowing at velocity $\v_{\ell}$) is a momentum source, and the "crystal" is a momentum sink. The momentum fluxes in and out of the phonon and electron subsystems need to be balanced in the steady state. The phonons (electrons) accumulate a momentum $\Delta \tb{P}_{\rm ph}$ ($\Delta \tb{P}_{\rm e}$), corresponding to a Doppler shift $\hbar q v_{\rm ph}$ ($\hbar q v_e$) of their momentum distribution. $\tau_{\rm um}$ is the phonon umklapp scattering time and $\tau_{\rm im}$ is the electron impurity scattering time.
(b) Phonon and electron wind velocities (normalized by the liquid velocity) as a function of the electronic density, for $\lambda_{\rm um} \approx 2~\rm N \cdot s \cdot m^{-3}$ and $\lambda_0 = 2 \times 10^6 ~\rm N \cdot s \cdot m^{-3}$.}\label{fig4}
\end{figure*}
\begin{table*}
\begin{tabular}{ l rcl rcl c }
\hline \hline
Interaction & \multicolumn{3}{l}{\quad\quad Momentum transfer} & \multicolumn{3}{l}{\quad\quad Force} &Scattering time \\
\hline
Classical friction & \quad Liquid &$\rightarrow$& Phonons \quad & $\tb{F}_0 $&=&$\lambda_0\mc{A}\tb{v}_{\ell}$ & \\
Phonon-hydron interaction &\quad Liquid &$\leftrightarrow$ &Phonons \quad & $\tb{F}_{\rm h/ph} $&=&$\lambda_{\rm h/ph}\mc{A}(\tb{v}_{\ell}-\tb{v}_{\rm ph})$ & \\
Electron-hydron interaction &\quad Electrons &$\leftrightarrow$& Liquid \quad & $\tb{F}_{\rm h/e} $&=&$\lambda_{\rm h/e}\mc{A}(\tb{v}_{\ell}-\tb{v}_{\rm e})$ & $\tau_{\rm h/e}$ \\
Electron-phonon interaction &\quad Phonons &$\leftrightarrow$& Electrons \quad & $\tb{F}_{\rm ph/e} $&=&$\lambda_{\rm ph/e}\mc{A}(\tb{v}_{\rm ph}-\tb{v}_{\rm e})$ & $\tau_{\rm ph/e}$ \\
Umklapp processes &\quad Phonons &$\rightarrow$& Crystal \quad & $\tb{F}_{\rm um} $&=&$\Delta\tb{P}_{\rm ph}/\tau_{\rm um}=\lambda_{\rm um}\mc{A}\tb{v}_{\rm ph}$ & $\tau_{\rm um}$ \\
Impurities &\quad Electrons &$\rightarrow$& Crystal \quad & $\tb{F}_{\rm im} $&=&$\Delta\tb{P}_{\rm e}/\tau_{\rm im}$ & $\tau_{\rm im}$ \\ \hline
Total friction & \quad Liquid &$\rightarrow$& Phonons+Electrons \quad & $\tb{F} $&=&$\lambda\mc{A}\tb{v}_{\ell}=(\lambda_0+\delta\lambda)\mc{A}\tb{v}_{\ell}$ & \\ \hline \hline
\end{tabular}
\caption{\textbf{Forces in the solid-liquid system.} List of interactions that may transfer momentum between the different components of the solid-liquid system, and notations for the associated forces, friction coefficients and scattering times.}\label{able}
\end{table*}
In order to evaluate the current in Eq.~\eqref{currentgeneral}, we require one last ingredient, which is the velocity $\v_{\rm ph}$ of the phonon wind. As mentioned in Sec. II, we evaluate it self-consistently, by enforcing momentum conservation in the solid-liquid system.
In the stationary state, the phonons accumulate a momentum that we denote $\Delta \tb{P}_{\rm ph}$. We model this momentum accumulation by giving all the phonons an average velocity $\v_{\rm ph}$: this means shifting the phonon distribution according to
\beq
n_{\rm ph}(\tb{q},\omega)=n_{\rm B}(\omega-\tb{q}\cdot\tb{v}_{\rm ph})\times2\pi\delta(\omega-cq)
\eeq
where $n_{\rm B}(\omega)= 1 / (e^{\hbar\omega/T}-1)$ is the Bose-Einstein distribution and $c$ is the sound velocity.
Then,
\beq
\frac{\Delta \tb{P}_{\rm ph}}{\mc{A} }= \int\frac{\dd^2\tb{q}}{(2\pi)^2}\hbar\tb{q}\,n_{\rm B}\left( qc-\tb{q}\cdot\tb{v}_{\rm ph}\right),
\eeq
which becomes, to first order in $\v_{\rm ph}$,
\beq
\frac{\Delta \tb{P}_{\rm ph}}{\mc{A} }= \frac{3\zeta(3)}{2\pi}\frac{T^3}{ \hbar^2c^4}\tb{v}_{\rm ph},
\eeq
$\zeta$ being the Riemann function. The contributions to the momentum flux in and out of the phonon system are summarized in Fig. \ref{fig4}a, and the associated notations are explicited in Table 1. The phonons receive momentum from the flowing liquid through the classical, roughness-induced contribution to the hydrodynamic friction force, $\tb{F}_0 = \lambda_0 \mc{A} \v_{\ell}$, and through the phononic contribution $\tb{F}_{\rm h/ph}$ to fluctuation-induced friction~\cite{Kavokine2022}. In Appendix \ref{sec_force}, we extend the framework of ref.~\cite{Kavokine2022} to account for the non-equilibrium state of the solid, and show that as long as the quasi-equilibrium fluctuation-dissipation theorem (Eq. \eqref{chiFD}) holds for the system's density response functions, the fluctuation-induced friction force is proportional to the differential velocity:
\begin{equation}
\tb{F}_{\rm h/ph} = \lambda_{\rm h/ph}\mc{A} (\v_{\ell} - \v_{\rm ph}).
\end{equation}
We note that, conversely, $\tb{F}_0$ {-- the classical, roughness induced friction --} does not depend on the phonon velocity: indeed, it originates largely in defects on the solid's surface, which do not move even if the phonons accumulate momentum.
The phonons lose momentum mainly through umklapp processes; we denote $\tau_{\rm um}$ the corresponding relaxation time. Formally, in an umklapp process, the interference of multiple phonons converts their momentum into a global translation of the crystal lattice. In practice, however, the 2D material layer remains fixed, and the momentum is transferred to the underlying substrate, which we do not describe explicitly. The momentum lost by the phonons per unit time and unit area through umklapp processes is therefore
\begin{equation}
\frac{\tb{F}_{\rm um}}{\mc{A} } = \frac{\Delta \tb{P}_{\rm ph}}{\tau_{\rm um}\mc{A} } = \frac{3\zeta(3)}{2\pi}\frac{T^3}{ \hbar^2c^4\tau_{\rm um}}\tb{v}_{\rm ph} \equiv \lambda_{\rm um} \v_{\rm ph},
\end{equation}
where we have defined the umklapp friction coefficient $\lambda_{\rm um}$. In graphene, $\tau_{\rm um}\approx 10 ~\rm ps$ \cite{Klemens1994} and $c\approx 2\e{4}$ m/s \cite{Nika2009,Ochoa2011,Cong2019}, so that $\lambda_{\rm um}\approx 2 ~\rm N \cdot s \cdot m^{-3}$. In addition, the phonons loose momentum through quantum friction with the conduction electrons, which is analogous to the water-electron quantum friction studied in ref.~\cite{Kavokine2022}, with the Coulomb interaction being replaced by the electron-phonon interaction. This type of friction is also known as a \emph{current-induced force} in the context of nanoscale electron transport~\cite{DiVentra2008}. The corresponding momentum flux is $\tb{F}_{\rm ph/e} = \lambda_{\rm ph/e}\mc{A} (\v_{\rm ph} - \v_{\rm e})$. In a steady state, the incoming and outgoing momentum fluxes (or forces) must compensate:
\beq\label{balance1}
\tb{F}_{0}+\tb{F}_{\rm h/ph}= \tb{F}_{\rm um}+\tb{F}_{\rm ph/e}.
\eeq
Using Eq. \eqref{ve2} for the electronic velocity $\v_{\rm e}$, we then obtain explicitly the phonon wind velocity as
\beq \label{vph}
\tb{v}_{\rm ph}=\frac{\lambda_0+\lambda_{\rm h/ph}+\frac{\tau}{\tau_{\rm h/e}}\lambda_{\rm ph/e}}{\lambda_{\rm um}+\lambda_{\rm h/ph}+\left(1-\frac{\tau}{\tau_{\rm ph/e}}\right)\lambda_{\rm ph/e}}\v_{\rm \ell}.
\eeq
This formula is consistent with the roughness-induced friction (with coefficient $\lambda_0$) being a momentum source, and the umklapp processes (with equivalent friction coefficient $\lambda_{\rm um}$) being a momentum sink. We do not have a practical way of evaluating the acoustic phonon contribution $\lambda_{\rm h/ph}$ to the fluctuation-induced friction. Nevertheless, guided by the qualitative ideas of ref.~\cite{Kavokine2022}, we expect it to be very small, since it is associated with in-plane lattice displacements at wavelengths much larger than the atomic spacing, that have short-range contact interactions with the liquid. Henceforth, we will assume it to be negligible compared to $\lambda_{\rm um}$ and $\lambda_0$. The electron-phonon friction coefficient, on the other hand, is evaluated explicitly in Sec. V, and we find that $\lambda_{\rm ph/e} \ll \lambda_{0}$ under all practical conditions; it is, however, comparable to $\lambda_{\rm um}$ at large electronic density.
Therefore, in the absence of impurities, the phonon wind velocity may be simplified, after some rearrangements, to
\beq \label{vph2}
\v_{\rm ph} \approx \frac{\lambda_0}{\lambda_{\rm um}+\lambda_{\rm ph/e}\tau/\tau_{\rm h/e}} \v_{\ell}.
\eeq
Typically, {on a molecularly rough surface, the roughness-induced friction coefficient is of order} $\lambda_0 \sim 10^5 - 10^6~\rm N \cdot s \cdot m^{-3}$~\cite{Bocquet2010}, so that $\v_{\rm ph} \gg \v_{\ell}$: the phonon velocity is orders of magnitude larger than the flow velocity. Fig. \ref{fig4}b shows the phonon and electron wind velocities (normalized by $\v_{\ell}$) as a function of the electronic density. Under all practical conditions the phonon wind is faster than the electron wind, whose velocity is reduced by the electron-hydron interaction. This results in the electron-hydron interaction actually making a \emph{negative} contribution to the Coulomb drag, as explained in the next section.
One may draw an analogy between current generation by hydrodynamic Coulomb drag and current generation through the photoelectric effect. The former benefits from strong electron-phonon interactions, as those help transfer momentum from the liquid to the electrons. Conversely, the latter is suppressed by electron-phonon scattering, as phonon drag slows down the photo-generated charge carriers: reducing phonon drag is a key challenge in the engineering of Perovskite materials, which are the state-of-the-art for photoelectric panels \cite{Yang2015,Kentsch2018,Zhang2021}. Such a complementarity calls for further studies of the potential interplay between these two phenomena.
\subsection{Quantitative estimates: Coulomb drag vs. phonon drag}
\begin{figure*}
\centering \includegraphics[scale=1]{fig5.pdf}
\caption{\tb{Coulomb drag current}. Electro-fluidic conductivity $\sigma_{\rm el}=\langle\tb{j}\rangle/\v_{\rm \ell}$ as computed with Eq.~\eqref{currentgeneral}, for different models of the solid's band structure. (a) Two-dimensional electron gas, with different values of the effective mass $m$ (expressed in units of electron mass $m_{\rm e}$). (b) Semi-conductor with electron and hole masses $m=0.1m_{\rm e}$, and band gap $2u_{\rm g}$. (c) Graphene, with constant Fermi velocity in the Dirac cone approximation. In (a)-(c) we assumed $\frac{\lambda_0}{\lambda_{\rm um}}= 10^6$. The dashed lines correspond to the prediction of Eq.~\eqref{currentapprox}, with no chemical potential dependence of the electronic wind velocity.}\label{fig5}
\end{figure*}
We are now in position to evaluate the electronic current according to eqs. \eqref{currentgeneral} and \eqref{ve2}. In order to make quantitatives estimates, we will use $\lambda_{\rm um} \approx 2~\rm N \cdot s \cdot m^{-3}$ as computed for graphene in the previous section, and $\lambda_0 = 2 \times 10^6 ~\rm N \cdot s \cdot m^{-3}$, which is reasonable for water on a large area
{carbon} sample {exhibiting} some {suface} ripples and defects. Furthermore, we make use of the values for $\lambda_{\rm ph/e}$ computed in Sec. V B.
Then, at not too large electronic density, $\v_{\rm ph} \approx (\lambda_0/\lambda_{\rm um}) \v_{\ell} \sim 10^6 \v_{\ell}$.
As in Sec. III, we will consider two models for the electronic structure and electron-phonon interaction: a two-dimensional electron gas (2DEG) with an effective mass $m$, and graphene, treated within the Dirac cone approximation, characterized by the constant Fermi velocity $v_{\rm F}$. In addition we will consider a model of a direct band gap semiconductor as the combination of an electron gas and a hole gas. For the two-band systems we will evaluate the total current as the sum of the currents in the two bands, which amounts to neglecting interband scattering. Fig. \ref{fig5} shows schematics of the three band structures under consideration
We define the \emph{electro-fluidic conductivity} as
\begin{equation}
\sigma_{\rm ef} \equiv {\langle j \rangle \over v_{\ell}}.
\end{equation}
We plot in Fig. \ref{fig5}a-c the electro-fluidic conductivity as a function of the chemical potential $\mu$. One first notes the difference in scaling of $\sigma_{\rm ef}$ with $\mu$ between the three model systems. These scalings are most conveniently understood in the limit of weak interactions and low temperature $\hbar/\tau_{k_{\rm F}} \ll T \ll \mu$. The first inequality means that the broadening of the electronic distribution due to electron-boson scattering is negligible compared to the thermal broadening. The second inequality means that the Fermi-Dirac distribution is well approximated by a step function. Eq.~\eqref{currentgeneral} accordingly simplifies to
\beq \label{currentapprox}
\langle\tb{j}\rangle\approx 2e v_{\rm F} N(\mu) \times \hbar k_{\rm F} \v_{\rm e}(k_{\rm F}),
\eeq
where $N(\mu)$ is the density of states at the Fermi level. The effect of the temperature appears to cancel out, leaving us with a {transparent} expression that is intuitive in a zero-temperature picture of Coulomb drag. The current is the electronic charge times the electronic velocity (which is the Fermi velocity), times the charge carrier density contributing to the current. The latter is the density of states at the Fermi level, times the energy range around the Fermi level in which the charge carriers can contribute to the current: this is given by the "Doppler shift" $\hbar k_F v_e$.
In the 2DEG, the density of states $N(\mu) = m/\hbar^2$ is independent of chemical potential, and $v_{\rm F} k_{\rm F} \propto \mu$: one expects a linear scaling of $\sigma_{\rm ef}$ with $\mu$ when $\mu \gg T$. A correction to this scaling comes from the chemical potential dependence of the electron-phonon friction coefficient: $\lambda_{\rm ph/e}\propto \mu$ (see Appendix \ref{sec_lambda}), which contributes to reducing the phonon wind velocity at high $\mu$ (see Fig. \ref{fig4}b). As a consequence, the current is expected to saturate at high chemical potential. Our model for the semi-conductor is a combination of two-dimensional electron and hole gases, therefore a linear scaling of $\sigma_{\rm ef}$ is obtained for both positive and negative chemical potential (Fig \ref{fig5}b). The electro-fluidic conductivity is suppressed when the chemical potential is within the band gap because of the lack of charge carriers, similarly to the 2DEG at negative chemical potential.
In graphene, $N(\mu) \propto \mu$, $k_{\rm F} \propto \mu$ and $v_{\rm F}$ is independent of $\mu$: we thus expect $\sigma_{\rm ef} \propto \mu |\mu|$ (the sign of $\sigma_{\rm ef}$ reflects the nature -- electron or hole -- of the charge carriers). Fig. \ref{fig5}c shows a slight deviation from this quadratic scaling, which is due to the dependence on chemical potential of the electron-phonon scattering time that contributes to the wind velocity $\v_e$ (see Fig. \ref{fig3}c and Fig. \ref{fig4}b).
In the electron gas model, $\sigma_{\rm ef}$ is tunable by the charge carrier effective mass. As the effective mass increases, the Fermi velocity is reduced, but the density of states and the Fermi momentum increase: the latter dominate, and overall, at low enough chemical potential ($\mu \lesssim 0.2~\rm eV$) $\sigma_{\rm ef} \propto m$. However, at larger chemical potential, the electron-phonon friction coefficient, that scales as $m^2$ (see Appendix~\ref{sec_lambda}), reduces the phonon wind velocity (see Fig. \ref{fig4}b), so that $\sigma_{\rm ef} \propto 1/m$. These scalings, illustrated in Fig. \ref{fig5}b, suggests that flat-band materials are likely to exhibit a significant hydrodynamic Coulomb drag effect.
At similar chemical potential, the electro-fluidic conductivity is found to be about two orders of magnitude larger in the 2DEG than in graphene. For instance at $\mu\sim 100$ meV, $\sigma_{\rm ef} \approx 20~\rm nA \cdot s \cdot \mu m^{-2}$ for the 2DEG and $\sigma_{\rm ef} \approx 0.1~\rm nA \cdot s \cdot \mu m^{-2}$ for graphene. This difference is mainly due to the wind velocity $\v_e$, which is determined according to Eq. \eqref{ve2} from the phonon wind velocity $\v_{\rm ph}$ and the flow velocity $\v_{\ell}$, with $\v_{\rm ph} \gg \v_{\ell}$ (see Fig. \ref{fig4}b). In the 2DEG the electron-phonon and electron-hydron scattering times are similar, so that $\v_e \approx \v_{\rm ph}/ 2$. In graphene, the electron-hydron scattering is much faster than the electron-phonon scattering, so that $\v_e \sim 10^{-2} \v_{\rm ph} \gg \v_{\ell}$. Thus, despite the different orders of magnitude, the phonon drag is the main driving force for the electronic current in both model systems. However, this does not imply that the electron-hydron interactions are negligible: in fact, they reduce the "bare" phonon drag current by a factor of 2 in the 2DEG and by a factor of $10^2$ in graphene, by providing a supplementary momentum relaxation pathway for the electrons. In other words, the electron-hydron interaction makes a negative contribution to the Coulomb drag. This effect is at the root of the quantum feedback phenomenon discussed in Sec. V.
\subsection{Comparison with experiments}
These results are in good qualitative agreement with the various experimental reports of electronic currents induced by liquid flows \cite{Ghosh2003, Newaz2012, HoLee2013, Yin2014, Yang2018, Rabinowitz2020, Marcotte2022}. Being more specific and quantitative, the recent nanoscale experiments in ref.~\cite{Marcotte2022} can be systematically compared to our theoretical predictions. In ref.~\cite{Marcotte2022}, an atomic force microscope is used to deposit a liquid droplet (ionic liquid or silicon oil) on the surface of a multilayer graphene sample, connected to two metallic electrodes. A current proportional to the droplet velocity is measured. The experimental electro-fluidic conductivities are in the range $2-8~\rm nA \cdot s \cdot \mu \rm m^{-2}$. The electronic structure of the experimental sample can be described as a zero gap semi-conductor.
This measured values are then fully reproduced by our theoretical prediction (Eq.~\eqref{currentgeneral}), considering effective masses in the range $0.2-0.8~ m_{\rm e}$ where $m_{\rm e}$ is the electron mass. Furthermore, the zero gap semi-conductor model reproduces the experimentally observed linear scaling of $\sigma_{\rm ef}$ with chemical potential \cite{Marcotte2022}. Therefore, our theory is quantitatively consistent with the experiments of ref.~\cite{Marcotte2022} and is able to account for the particularly strong Coulomb drag currents (in the 10 nA range) that were generated by the motion of a micrometer-sized droplet at a few $\mu \rm m/s$. The key factor that is responsible for a strong current is the hydrodynamic friction force, which is large {in the experiments of ref.~\cite{Marcotte2022}} due to the high viscosity of the liquids used and the wrinkling of the sample surface: {this results in an efficient transfer of} a significant amount of momentum to the sample's phonon modes. The crucial role of phonons is further supported by the fact that the effect could be observed with a non-ionic silicon oil, where liquid-electron Coulomb interactions are negligible.
\section{Quantum feedback and negative friction}
\begin{figure*}
\centering \includegraphics[scale=1]{fig7.pdf}
\caption{\textbf{Quantum feedback and negative friction.} (a) Schematic representation of the quantum feedback phenomenon: a fraction of the momentum transferred to the solid by classical friction is returned to the fluid by the solid's internal excitations. (b) Electron-boson friction coefficients as a function of the electron density for a 2DEG with $m = 0.1~\rm m_{\rm e}$ and graphene, evaluated according to~\eqref{lambda_boson}. The calculations has been carried out using the electronic susceptibilities in the zero-temperature limit and in the RPA approximation. (c) Hydrodynamic friction coefficient $\lambda$ (in units of the classical contribution $\lambda_0$) for a 2DEG with $m=0.1 m_{\rm e}$ and graphene.}
\label{fig7}
\end{figure*}
\subsection{Derivation}
Our analysis in Sec.~IV revealed that the hydrodynamic Coulomb drag current is determined by a subtle combination of electron-phonon and electron-hydron interactions. Indeed, the electron-hydron scattering provides a supplementary momentum relaxation pathway for the electrons, which prevents them from aligning to the phonon wind velocity. This immediately implies that the electrons actually transfer momentum \emph{to} the flowing liquid: they make a \emph{negative} contribution to the hydrodynamic friction force, as shown schematically in Fig. \ref{fig7}. In this section, we explicitly evaluate this negative contribution and assess its practical consequences.
The liquid interacts with the solid through the classical roughness-induced friction force $\tb{F}_0$, and through the phononic and electronic contributions to the fluctuation-induced (quantum) friction $\tb{F}_{\rm h/ph}$ and $\tb{F}_{\rm h/e}$, respectively. The total hydrodynamic friction coefficient $\lambda$ is then defined according to
\beq
\lambda\mc{A} \v_{\rm \ell} = \tb{F}_{0}+\tb{F}_{\rm h/ph}+\tb{F}_{\rm h/e}.
\eeq
Introducing the individual friction coefficients as in Sec. IV B (see Table 1 for notations),
\beq
\lambda\v_{\rm \ell} = \lambda_{0} \v_{\ell} +\lambda_{\rm h/ph} (\v_{\ell} - \v_{\rm ph}) +\lambda_{\rm h/e}(\v_{\ell} - \v_e).
\eeq
In the limit where $\lambda_0$ dominates all other friction coefficients, and using eqs. \eqref{ve2} and \eqref{vph2} for the velocities $\v_{\rm e}$ and $\v_{\rm ph}$, a rearrangement yields:
\beq
\lambda = \lambda_0 - \delta \lambda\quad\tn{with}\quad\delta\lambda=\frac{\lambda_0}{1+\frac{\lambda_{\rm um}+\frac{\tau}{\tau_{\rm im}}\lambda_{\rm ph/e}}{\lambda_{\rm h/ph}+\frac{\tau}{\tau_{\rm ph/e}}\lambda_{\rm h/e}}},
\label{delta_lambda}
\eeq
where we have included the possibility for the electrons to lose momentum through impurity scattering (at a rate $\tau_{\rm im}^{-1}$). This is our second key result, anticipated in Sec. II. Strikingly, $\delta \lambda$ represents a negative contribution to hydrodynamic friction. It is, however, always smaller than $\lambda_0$, so that the total friction coefficient remains positive, and there is no violation of the laws of thermodynamics. But a key observation is that it becomes equal to $\lambda_0$ in the absence of impurity or umklapp scattering. This amounts to formally considering a solid that is unable to relax momentum: then, even if the solid has a rough surface, all the momentum the liquid loses through the classical friction $\tb{F}_0$ is sent back by the solid's electronic and phononic fluctuations, so that the total friction vanishes. In practice, however, there is always some amount of momentum relaxation that keeps the friction from vanishing. The amount of negative friction is ultimately obtained by balancing the relaxation processes (umklapp and impurity scattering), and the processes that allow the solid to return momentum to the liquid: the phonon-hydron and electron-hydron interactions, the latter corresponding to quantum friction~\cite{Kavokine2022}.
In order to go beyond a qualitative discussion, we evaluate the electron-hydron and electron-phonon friction coefficients in the framework of ref.~\cite{Kavokine2022}. A general electron-boson friction coefficient is given by
\beq
\lambda_{\rm b/e} =\frac{\hbar^2}{8\pi^2 T}\int_0^\infty\frac{\dd\omega\dd q}{\tn{sinh}^2\left(\frac{\hbar\omega}{2T}\right)} \,q^3\,\frac{\tn{Im}[\chi_{\rm e}^{\rm R}]\tn{Im}[D^{\rm R}]}{|1-\chi_{\rm e}^{\rm R}D^{\rm R}|^2},
\label{lambda_boson}
\eeq
where $D^{\rm R}$ is the retarded bosonic propagator in which the electron-boson interaction has been absorbed (see Sec. III).
We evaluate Eq.~\eqref{lambda_boson} by numerical integration for both the graphene and 2DEG model systems, using zero temperature RPA expressions for the electronic susceptibility~\cite{Wunsch2006,Mihaila2011}. The results are plotted in Fig. \ref{fig7}b as a function of the electronic density. The relative values of the friction coefficients are consistent with the respective scattering times: indeed, one roughly expects $\lambda_{\rm b/e} \propto \tau^{-1}_{\rm b/e}$ (see Appendix C3). In particular, in the graphene case $\lambda_{\rm ph/e} \ll \lambda_{\rm um} \approx 2 ~\rm N \cdot s \cdot m^{-3}$ for all reasonable electronic densities, while in the 2DEG model this is true only for low enough electronic densities.
In any case the term proportional to $\lambda_{\rm ph/e}$ in Eq.~\eqref{delta_lambda} is negligible in the limit of low impurity scattering: $1/\tau_{\rm im} \to 0$. If we further assume $\lambda_{\rm h/ph} \ll \frac{\tau}{\tau_{\rm ph/e}}\lambda_{\rm h/e}$ following the argument of Sec. IV B, Eq.~\eqref{delta_lambda} simplifies to
\beq
\lambda = \lambda_0 - \delta \lambda\quad\tn{with}\quad\delta\lambda=\frac{\lambda_0}{1+ \frac{\tau_{\rm ph/e}}{\tau} \frac{\lambda_{\rm um}}{\lambda_{\rm h/e}}}.
\label{delta_lambda2}
\eeq
The resulting estimate for $\delta \lambda$ is plotted in Fig. \ref{fig7}c. In doped graphene, with an electronic density $10^{14}~\rm cm^{-2}$, we find that the negative quantum contribution reduces the hydrodynamic friction by about 12\%. In a 2DEG with parabolic dispersion, the quantum contribution to friction is stronger, so that at the same electronic density (and taking the graphene value for $\lambda_{\rm um}$) the friction reduction is nearly 80\%.
\subsection{Discussion}
According to Eq.~\eqref{delta_lambda2}, the negative contribution to friction represents a \emph{quantum feedback} of the solid on the liquid flow: it originates from electronic excitations returning momentum that the liquid has lost due to the surface roughness. It is significant if the quantum friction coefficient $\lambda_{\rm h/e}$ is non-negligible compared to the umklapp friction coefficient. This is not a very stringent requirement: it is in fact satisfied for our two model systems, at least in part of the considered electron density range.
The negative quantum friction effect is thus predicted to be quantitatively important.
It complements the picture of quantum hydrodynamic friction developed in ref.~\cite{Kavokine2022}, where the solid was assumed to remain in an equilibrium state. This was justified in particular for a truly semi-infinite solid, where the surface can quickly transmit momentum to the bulk. In a 2D material that is weakly coupled to the underlying substrate, we find that momentum accumulation and the resulting Coulomb drag current are non-negligible. As electrons receive momentum from the liquid through the phonon-mediated interaction, they begin to move faster than the liquid, thus helping it flow along the surface and reducing the total hydrodynamic friction force. In this way, quantum friction can play a significant role even on non-atomically-flat surfaces.
Negative quantum friction may provide a clue as to why the experimentally measured water friction on graphene~\cite{Xie2018} appears to be much lower than what is predicted by essentially all molecular simulations~\cite{Kannam2013,Tocci2014,Tocci2020,Thiemann2022}. Even in ab initio simulations at the Born-Oppenheimer level, it is not possible to account for Coulomb drag or fluctuation-induced quantum friction. Nevertheless, our quantitative estimates cannot account for the full extent of the discrepancy, and the specific case of water on pristine graphene will be the subject of further investigation.
Ultimately, negative quantum friction provides a previously unexplored pathway for designing surfaces with low hydrodynamic friction. Friction reduction is first achieved by minimizing momentum relaxation in the solid: the best case scenario is a suspended 2D material, or a 2D material weakly coupled to its substrate. Then, the water-solid quantum friction coefficient needs to be maximized, so as to allow for efficient momentum transfer back to the liquid. Conditions for high quantum friction have been detailed in ref.~\cite{Kavokine2022}; typically, quantum friction benefits from high electronic densities and large effective masses. As outlined above, for the momentum transfer to be efficient, the quantum friction coefficient needs to be large compared to $\lambda_{\rm um}$, which is several orders of magnitude smaller than a typical hydrodynamic friction coefficient: many materials may satisfy this requirement. Potential candidates include magic angle twisted bilayer graphene~\cite{Stauber2016}, and metallic transition metal dichalcogenides such as $\rm VS_2$ and $\rm TaS_2$~\cite{DaJornada2020}.
\section{Conclusion: the importance of Quantum Interfacial Effects}
In this work, we introduce a novel perspective on hydrodynamics at solid-liquid interfaces. The usual description of such interfaces relies on the continuum Navier-Stokes equation, or -- going down to the molecular scale -- on the interactions between fluid molecules and the surface corrugation. This approach becomes insufficient to account for more advanced quantum couplings that arise between the liquid and the solid.
Specifically, we start here by addressing the question of electronic current generation by interfacial liquid flows. Using the non-equilibrium Keldysh framework, we systematically investigate the mechanisms that couple fluid motion to the electronic degrees of freedom within the solid material, which include Coulomb and phonon-mediated interactions. The theory provides quantitative estimates for the electro-fluidic conductivity -- defined as the electronic current response to the fluid motion --, which are fully corroborated by experimental reports.
Going further, our theoretical framework reveals a quantum feedback mechanism, that provides a negative contribution to hydrodynamic friction: in a very counter-intuitive way, quantum effects may reduce hydrodynamic friction at the solid-liquid interface. The Coulomb drag current can in fact be "faster" than the liquid flow, and return momentum to the liquid through electron-hydron scattering. The resulting \emph{negative quantum friction} provides a unique opportunity to tune hydrodynamic friction by choosing specific electronic properties of the confining wall. This broadens the scope of solid-liquid quantum friction beyond the water-carbon interfaces discussed in ref.~\cite{Kavokine2022}, as we find that it can play a role even for materials with non-negligible surface roughness.
More generally, our results provide a new way of thinking about the interaction of liquids with solids, and in particular water-solid interfaces. By bridging fluid dynamics and condensed matter theory, we picture the interface dynamics in terms of the collective excitations of both the liquid and the solid, instead of real-space molecular interactions. Water charge fluctuations -- which we dub \emph{hydrons} -- couple to the quantum excitations inside the confining solid as first proposed in ref.~\cite{Kavokine2022}. Here, this approach bears fruit by accounting for existing experimental results, and predicting a novel quantum feedback effect. It opens the way to quantum engineering of fluid transport: quantum effects can become valuable assets for future water technologies.
The emerging interface between hydrodynamics, electrodynamics, condensed matter physics and quantum mechanics, is an uncharted territory that begs for further exploration.
\section*{Acknowledgements}
The authors acknowledge fruitful discussions with A. Marcotte, M. Lizee and A. Siria. The Flatiron Institute is a division of the Simons Foundation. B.C. acknowledges funding from a J.-P. Aguilar grant of the CFM Foundation, and thanks Olivier Parcollet for hosting him at the Flatiron Institute. L.B. acknowledges funding from the EU H2020 Framework Programme/ERC Advanced Grant agreement number 785911-Shadoks.
\newpage
|
\section{Introduction}
Many astrophysical processes, including the gravitational collapse of molecular
clouds, are complex and multi-scaled. Residing in the Galactic disk, the clouds
are the nurseries of stars. They are open systems that interact with the
environment constantly. Their collapses involve an interplay between turbulence
\citep{2004RvMP...76..125M}, gravity, magnetic field \citep{2014prpl.conf..101L},
ionization radiation, and Galactic shear, resulting in highly complex density
distributions.
\begin{figure}
\includegraphics[width = 0.45 \textwidth]{illus.pdf}
\caption{\label{fig:illus} {\bf Evaluation the Level-Set Density Exponent.}
We first divide a region using a set of isosurfaces. A typical region
$R_1$ would be surrounded by an isosurface at $\rho = \rho_1$. Inside this
region, there exists one or a few subregions ($R_{2, i}$) surrounded by isosurfaces
at $\rho = \rho_2$. The equations on the right hand side describe how the
density exponent $k_{\rho}$ at voxels included in $R_1$ yet not included in
$R_{2, i}$) is computed. The spacing between the adjacent
isosurfaces are exaggerated for a clearer view. }
\end{figure}
A clear picture of how the collapse occurs is yet to be achieved. Modern, high
spatial dynamical range observations provide maps that contain an unprecedented
amount of information \citep{2016A&A...587A.106Z}. Measures like density
(Probability Distribution Function) PDF \citep{2009A&A...508L..35K},
correlation function \citep{2003ApJ...588..881P} and fractal dimension
\citep{1993A&A...270..451H} and spectral correlation function \citep{1999ApJ...524..887R}
have been proposed to quantify the structure of the
star-forming regions and these are summarized in \citet{2021PASP..133j2001B}
. Despite their successes, these measures are degenerate,
where original data from regions are often represented using curves where the spatial information is lost. Although the density PDF has enjoyed good success
in quantifying the cloud structure and connecting simulations with observations
\citep{2009A&A...508L..35K,2011ApJ...727L..20K,2012ApJ...750...13C,2016MNRAS.461.3027L,2021MNRAS.507.4335K},
the loss of information during the compression means that the complexity of the
underlying structures is often overlooked. Besides, to derive these measures,
one needs to specify the boundary of a region in advance, which can be a
a challenging task and efforts have been made to address the boundary issue
\citep{2015A&A...578A..97L,2015IAUGA..2253648A} The criteria for collapse
remains an open question. Addressing the transition from clouds to clumps which
hosts clustered star formation, the virial parameter
\citep{1992ApJ...395..140B} is often used to study whether a region should
collapse or not, and gravitationally unbound clouds like the Perseus can be divided into regions which gravitationally bound by themselves
\citep{2015A&A...578A..97L}. Along this direction, \citep{Li2017} provides a
criteria describing the transition to gravitationally-bound clumps. Analytical
models have been established to describe the evolution of density structure and
the star formation efficiencies in a turbulent medium
\citep{2005ApJ...630..250K,2007ApJ...661..972P,2015MNRAS.450.4035F,2017ApJ...834L...1B,2019MNRAS.488.1501G}.
Power-law density structures such as $\rho \propto r^{k_{\rho}}$ ($k_\rho <
0$) are common in systems that have reached intermediate asymptotic states,
such that the behaviors are independent of the details of the initial and/or
boundary conditions \citep{1996sssi.book.....B,Goldenfeld}. One such example is
the gravitational collapse of a molecular cloud, where stationary collapse leads
to $\rho\sim r^{-2}$ \citep{1969MNRAS.144..425P,
2009ARep...53.1127P,2011ApJ...727L..20K,Girichidis2014,2015ApJ...814...48N,2015ApJ...804...44M,2018MNRAS.474.5588D,2018MNRAS.477.4951L,2019MNRAS.485.3224D},
and accretion flow around dense objects have $\rho\sim r^{-1.5}$
\citep{1941MNRAS.101..227H,1952MNRAS.112..195B}. For these systems, measuring
the power-law density exponent $k_\rho$ would enable us to distinguish
different structures, and the value of $k_{\rho}$ can be directly compared
against models to achieve understanding. Various attempts have been made to
measure the density exponent. The most obvious approach is to fit spherical
models to observational data. However, as the majority of regions we study are
non-spherical, this approach is limited in practice. Another way is to derive
the power-law density exponent using the density probability distribution
function (PDF) \citep{Girichidis2014,2016MNRAS.461.3027L,2020ApJ...903...56P}.
Although the procedure is straightforward, this statistical approach only
allows for the derivation of an ``effective'' density exponent, which contains
no information on how gas organizes spatially.
To fully exploit the diagnostics power of the density exponent $k_{\rho}$, importing the Level-Set Method (LSM),
we propose a new formalism to measure its value for non-spherical yet centrally condensed regions. The Level-Sets are contourlines, and the Level-Set Method is a conceptual framework where analyses of
surfaces and shapes can be performed with the help of Level-Sets.
By applying the LSM to state-of-the-art high dynamical range observations of star-forming regions, we obtain spatially-resolved maps of the density exponent, and reveal, for the first time, the complexity and regularity of molecular cloud structures.
\begin{figure*}
\centering
\includegraphics[width = 0.9 \textwidth]{figure_0.pdf}
\caption{\label{fig:illus} {\bf Methods for analyzing cloud structure.} We consider an example made of two clumps (clump 1: $\rho \sim (r + 0.01 )^{-1}$, clump 2: $\rho \sim (r + 0.015)^{-3}$). Because of the truncation at small radii, the density exponent of the clump 2 is larger than -3.
{\bf Upper Left:} Density distribution. {\bf Upper Right:} Density PDF. {\bf Lower Left:} Density Exponent Map. Color represents the density exponent, and brightness represents surface density. The colormap is the same as the one in Fig. \ref{fig:overview}. {\bf Lower Right:} mass-weighed density exponent distribution. The normalization of the $y$-axis is arbitrary. \label{fig:0}}
\end{figure*}
\section{Method}
For a spherical system, the density structure can be described as $\rho(r)$ where $r$ is the radius. To measure the steepness of the density profile in a given location, we adopt a local model where $\rho \sim r ^{k_{\rho}}$, and in the vicinity of $r$, the value of $k_{\rho}$ can be derived as
$$
k_{\rho}(r) = \frac{{\rm log}(\rho(r + \delta_r))- {\rm log}(\rho(r))}{ {\rm log}(r + \delta_r) - {\rm log}(r) }\;,
$$
where $\delta_r << r$.
The goal is to measure the density exponent for clouds of arbitrary geometries. Assuming a 3D density structure $\rho(x, y, z)$,
we divide the region using a
set of densely-spaced iso-density contours, after which each subregion should be surrounded by
a contour at $\rho = \rho_1$, one or a few contours at $\rho = \rho_2$, and it
should contain values ranging from $\rho_{1}$ to $\rho_{2}$. Assuming
that the region $R_1$ has a volume of $V_1$, inside this region, there
can be a few ($n$) subregions $R_{2, i}$ surrounded by isosurfaces with $\rho
= \rho_{2}$ ($\rho_2 > \rho_1$), and these subregions have volumes of $V_{2,
i}=V_{2,1}, \ldots V_{2,n}$. The size of the region can be approximated as $r_1
\propto (V_1)^{1/3}$, and the effective size of all subregions altogether can be
approximated as $r_2 \propto ({\sum_i V_{2, i}})^{1/3}$. The \textit{Level-Set Density Exponent} is
\begin{equation}
k_{\rho} = \frac{{\rm log}(\rho_2) - {\rm log}(\rho_1)}{{\rm log}(r_2) - {\rm log}(r_1)} \;.
\end{equation}
The procedure
is illustrated in Fig. \ref{fig:illus}, and the resulting map is called the \emph{Density Exponent Map}.
The advantage of the Density Exponent Analysis lies in its robustness and resolving power:
the method is directly applicable to maps that contain heterogeneous structures and can be used to distinguish these structures.
As an example (Fig. \ref{fig:0}), we construct a model which contains two spherical clumps of different density profiles. We derive its density PDF and produce a density exponent map. The density PDF contains limited information since the spatial information is lost completely. In contrast, the Density Exponent Map indicates that the map should be separated into regions characterized by different density exponents and the map contain values of the density exponent at every location. The additional spatial information retrieved by the method makes it a powerful tool to analyze complex, spatially inhomogeneous datasets, such as the density structures of nearby molecular clouds. We should note that our resolved map of the density exponent is not completely identical to the density exponent derived in analytical models \citep{1969MNRAS.144..425P,
2009ARep...53.1127P,2011ApJ...727L..20K,Girichidis2014,2015ApJ...814...48N,2015ApJ...804...44M,2018MNRAS.474.5588D,2018MNRAS.477.4951L,2019MNRAS.485.3224D}, as in those cases, the density fluctuations caused by e.g. turbulence are not accounted for.
A final remark: one can be temped to evaluate the density exponent in 2D, where
\begin{equation}
k_{\Sigma} = \frac{{\rm log}(\Sigma_2) - {\rm log}(\Sigma_1)}{{\rm log}(r_2) - {\rm log}(r_1)} \;,
\end{equation}
where $\Sigma$ represents the surface density, and derive $k_\rho$ using $k_{\rho} = k_{\Sigma} - 1$. However, this approach is limited, and the above-mentioned equation holds only if the structures we study are completely self-similar \footnote{\url{https://en.wikipedia.org/wiki/Abel_transform}.}.
\section{Results \& Discussions}
Located at a distance of $\sim$ 290 pc
\citep{2018ApJ...869...83Z}, the Perseus star-forming region is nearby and well-resolved.
We use the surface density map derived using data from the
Herschel and the Planck telescope \citep{2016A&A...587A.106Z}. The original map is the opacity of dust at 870 $\mu$m. We converted the map into a map of the surface density using a calibrated conversion factor provided in their Appendix.
The
cloud has a size of $\sim$ 30 pc, and the map has a resolution of 36 arcsec ($\sim
0.05\;\rm pc$). The enormous ($\sim 10^3$) spatial dynamical range allows detailed studies.
Derivation of the level-set density exponent $k_{\rho}$ requires 3D density
distributions. As observations are done mostly in 2D, we develop a method (see Method \ref{sec:appen:3d})
to construct 3D density distributions using 2D maps. This is achieved by first decomposing a 2D surface density map into component maps that contain structures of different sizes and by assigning thicknesses to these component maps
and combining them. For cloud-like density structures, the
reconstruction allows us to measure the mass-weighted mean density exponent to an
accuracy of $\lesssim 0.1$.
In our calculations, we focus on gas with
$\rho_{\rm H_2} > 1000 \;\rm cm^{-3}$.
This corresponds to 40\% of the gas contained in the Perseus region and the region is surrounded by a diffuse envelope that contains gas that does not contribute directly to the star formation. We also excluded unresolved regions -- patches surround by contours whose sizes are smaller than 3 pixels (0.08 pc), from our analyses.
To derive the density exponent map, the data is divided using 100 contour levels equally spaced in log($\rho$).
\begin{figure*}
\includegraphics[width = 1.0 \textwidth]{region_evolution.pdf}
\caption{ {\bf Results from the Perseus molecular cloud.}
{\bf Left:}
Density Exponent Map of the Perseus region. We plot the results at the midplane of the reconstructed 3D density distribution. The brightness represents the density, and the colors represent the density exponent $k_{\rho}$. The color channels have response functions of Gaussian shapes (see the colormap on the upper right).
{\bf Right:} Density exponent plotted against the mean density. Pie charts show the dense gas mass fractions of the corresponding regions, where the yellow areas stand for the diffuse gas and the orange areas stand for the dense gas. A relation between dense gas fraction and density exponent can be found online. For a single region, the mass-weighted mean density exponents has an estimated uncertainty of 0.1, caused by our density reconstruction (See our online material).
\label{fig:overview} }
\end{figure*}
In Fig. \ref{fig:overview} we plot the density exponent distribution at the cloud center plane, which contains all the line-of-sight density maximums. $k_{\rho}$ ranges from -3.5 to -0.5, forming a highly inhomogeneous pattern that contains variations on different scales.
We note that when the structures we study can be characterized by strict power laws, the slope of density PDF is $3/k_\rho $, and the slope of the surface density PDF is $2/(k_\rho + 1)$. In our case, the range of $k_{\rho}$ translates to density PDF slopes that range from -1 to -6, and a surface density PDF slopes as shallow as -1. We note that these are very rough translations, as the derivations hold only if the underlying structures are strictly self-similar.
\subsection{Inter-regional variations: Density-driven collapse}
The Perseus clouds can be separated into a few pc-sized subregions. Each region has a corresponding gravitational potential dip \citep{2015A&A...578A..97L}, and these regions can collapse to form star clusters or associations. We first divide the cloud into these regions and evaluate parameters including the mass-weighed mean density exponent and the mean densities.
We also characterize these regions by deriving a quantify called the dense gas fraction (see Methods \ref{sec:appendix:dense}). Since all the dense gas would collapse to form individual or multiple stars, the dense gas mass fraction is a direct indicator of the star formation activity.
Correlations between these quantities are summarized in the right panel of Fig. \ref{fig:overview}.
In general, regions of steeper density profiles have higher dense gas mass fractions. Thus, the formation of dense gas, which occurs on tiny scales, is deeply linked to a global steepening of the density profile. Although the link between dense gas mass fractions and the steepness of the density profile inferred from the shape of the surface density PDF is somehow established \citep{2015A&A...577L...6S,2016MNRAS.461.3027L}, our maps provide a more detailed view, where steepening occurs at the high-density regions found around the dense cores.
By plotting the mass-weighted density exponent
\begin{equation}
k_{\rho, \rm mean} = \int \frac{k_{\rho} \rho {\rm d} v}{\rho {\rm d} v} \;,
\end{equation}
against the mean density (Fig. \ref{fig:overview}), we find that regions of higher densities tend to have steeper density profiles, which suggests a \emph{density-driven collapse} scenario.
Since all regions belong to the Perseus molecular cloud, we assume that they have almost the same age $t_{\rm Perseus}$. Provided that they evolve at paces set by their free-fall time $t_{\rm ff}$, which is related to the mean densities by $t_{\rm ff}\approx
\sqrt{1 / (G \rho_{\rm mean, init})}$, the evolutionary stage can be parameterized using $f = t_{\rm Perseus}/t_{\rm ff} \sim \rho_{\rm mean, init}^{1/2}$.
Regions with larger $\rho_{\rm mean, init}$ are more evolved, have higher densities when observed (higher $\rho_{\rm mean, present}$), with steeper density profiles and higher dense gas fraction. In reality, due to the presence of turbulence, the actual collapse time can be a few times the collapse \citep{2018MNRAS.477.4951L,2022ApJ...927...75A}. But this additional factor will not change our conclusion.
The picture of a steepening density profile is consistent with results from a recent paper \citep{2020arXiv200914151G}, and the correlations point to a simple picture where the mean gas density sets the pace at which a region evolves, leading to the steepening of the density profile \footnote{One might be worried about the effect of protostellar outflows \citep{2016ARA&A..54..491B}. Indeed, regions like the NGC1333 \citep{2000A&A...361..671K} contain a number of these. However, the effects should not be significant, as recent simulations found that outflows lead to an excess of low-density gas which we do not analyze \citep{2022ApJ...927...75A}. }.
The map also allows us to identify a new region called ``Perseus Cirrous"
(Fig. \ref{fig:overview}) for the first time. Although overlooked by previous
studies, the region stands out in our analyses due to its shallow density
profile ($k_{\rho } \approx -1$). The region occupies the shallow end of the
density exponent parameter space and its structure should be representative of
the structure gas at the early stages of gravitational collapse. On the other hand,
since this region has a density profile that is significantly shallower than
those the others, it is possible that this region belongs to the gravitationally
unbound phase, whereas others belong to the bound phase, as suggested by
previous studies \citep{2007ApJ...665..416K,2012ApJ...750...13C,Li2017}.
\subsection{Summary \& Future Extensions}
The evolution of molecular clouds is exemplary of complex, multi-scale processes. Regions in molecular clouds appear to be gravitationally bound at
$\sim $ pc scale \citep{2015A&A...578A..97L}, and the collapse of dense cores, which is directly related to star formation, occurs at $\lesssim $ 0.05 pc. Importing the Level-Set Method, we develop a new, robust formalism to compute spatially resolved maps of the density exponent. On the pc scale, the mass-weighted mean density exponent correlates with the star formation activity. On smaller scales ($\lesssim$ 1 pc), the density exponent $k_{\rho}$ still exhibits significant variations. This complex pattern results from a continued steepening of the density profile driven by gravitational collapse.
The spatial information our method provides is valuable for large, inhomogeneous datasets. The Level Set-based formalism can be modified to suit different models, for example, to derive the scale length for exponential-like structures, which we will explore in the future.
We expect our method to be effective for other structures including the density structure of granular materials and
the Large-Scale Structure of the
Universe.
One limitation of our method is that it still requires 3D density. In the current approach, we are using 3D density distribution constructed from 2D surface density maps. For non-self-similar structures, the relation between exponent evaluated in 2D and 3D are not trivial \footnote{This has been well-understood under the Abel Transform \url{https://en.wikipedia.org/wiki/Abel_transform}.}, and our ongoing studies (Li \& Zhou in prep) suggest that compared to the current approach, evaluating the density exponent from 2D is less accurate but significantly easier to implement, making widespread applications feasible. These will be explored in our future papers.
\section*{Acknowledgements}
We thank the referee for a careful reading of the paper and for the comments.
GXL acknowledges supports from
NSFC grant W820301904 and 12033005.
\section*{Data availability}
The data this publication makes use of is publicly available, the link can be found in \citet{2016A&A...587A.106Z}, and we use code is from \citet{2022arXiv220105484L}, which is available at \url{https://gxli.github.io/Constrained-Diffusion-Decomposition/}.
\bibliographystyle{mnras}
|
\section{Introduction}
\input{introduction-ikematsu-20220412.tex}
\section{Preliminaries}
\paragraph*{Notations and Consensus. }
Unless noted otherwise, let any algorithm in this paper be a probabilistic
polynomial time Turing Machine.
We also define as follows: let $\mathbb{F}$ be a finite field used throughout this paper,
$\mathrm{M}_n(\mathbb{F})$ be the set of $n$-square matrices over $\mathbb{F}$ and
$\mathrm{GL}_n(\mathbb{F})$ be the set of $n$-square invertible matrices over $\mathbb{F}$.
\begin{definition}
[Search MinRank Problem]\label{def:Search-Minrank-Problem} The
search MinRank problem is defined as follows. Given a positive integer
$r\in\mathbb{N}$ and $m$ matrices
$M_{0},M_{1,}\ldots,M_{m-1}\in\mathrm{M}_n(\mathbb{F})$,
find $\alpha=(\alpha_{1},\ldots,\alpha_{m-1})\in\mathbb{F}^{m-1}$
such that $\mathrm{rank}(M)=r$, where $M=\sum_{i=1}^{m-1}\alpha_{i}M_{i}-M_{0}$.
\end{definition}
\subsection*{Decisional MinRank Problem}
In this paper, we use the hardness of the decisional version of the
MinRank problem as the basic assumption of the security since it is
much simpler to prove the security based on the decisional version
compared to the search version above.
\begin{comment}
Informally, the decisional MinRank problem is as follows: given a
positive integer $r\in\mathbb{N}$ and $m$ $n$-square matrices over
a finite field $\mathbb{F}$: $M_{0},M_{1,}\ldots,M_{m-1}$, decide
whether there exists $\alpha=(\alpha_{1},\ldots,\alpha_{m-1})\in\mathbb{F}^{m-1}$
such that $\mathrm{rank}(M)=r$, where $M=\sum_{i=1}^{m-1}\alpha_{i}M_{i}-M_{0}$.
The formal definition of decisional MinRank problem is as follows.
\end{comment}
\begin{definition}
[Decisional MinRank Problem] An algorithm $\mathcal{D}$ is said
to $(t,\varepsilon)$-solve the decisional MinRank problem associated
with the finite field $\mathbb{F}$ and $r,m,n\in\mathbb{N}$ if $\mathcal{D}$
runs in $t$ units of time and the following holds.
\begin{align*}
& \left| %
\Pr\left[\mathcal{D}^{\mathsf{IGen}}(\mathbb{F},r,m,n)=1\right]
-
\Pr\left[\mathcal{D}^{\mathsf{LossyGen}}(\mathbb{F},r,m,n)=1\right]
\right|
\geqq\varepsilon,
\end{align*}
where:
\begin{itemize}
\item $\mathcal{D}^{\mathsf{IGen}}$ denotes that $\mathcal{D}$ receives
the input from the oracle $\mathsf{IGen}$ which generates an instance
of MinRank problem that has at least one solution, i.e., $m$
matrices: $M_{0},M_{1,}\ldots,M_{m-1}\in\mathrm{M}_n(\mathbb{F})$,
such that there exists $\alpha=(\alpha_{1},\ldots,\alpha_{m-1})\in\mathbb{F}^{m-1}$
satisfying the following:
\begin{align}
\mathrm{rank}\left(\sum_{i=1}^{m-1}\alpha_{i}M_{i}-M_{0}\right) & =r,\label{eq:Igen}
\end{align}
\item $\mathcal{D}^{\mathsf{LossyGen}}$ denotes that $\mathcal{D}$ receives
the input from the oracle $\mathsf{LossyGen}$ who generates $m$
random matrices $M_{0},M_{1,}\ldots,M_{m-1}\in\mathrm{M}_n(\mathbb{F})$,
which \uline{do not} necessarily
have $\alpha=(\alpha_{1},\ldots,\alpha_{m-1})\in\mathbb{F}^{m-1}$
satisfying Eq. (\ref{eq:Igen}).
\end{itemize}
The decisional MinRank problem associated with the finite field $\mathbb{F}$
and $r,m,n\in\mathbb{N}$ is said to be $(t,\varepsilon)$-hard if
there is no algorithm $\mathcal{D}$ which $(t,\varepsilon)$-solves
the problem.
\end{definition}
\begin{remark}
Buss et al. \cite{DBLP:journals/jcss/BussFS99} and
Courtois\cite{DBLP:conf/asiacrypt/Courtois01}
have proven that the decisional MinRank problem is NP-Complete.
\end{remark}
\section{Proposed Scheme }
In this section, first we describe our proposed identification scheme.
Then we show that our proposed scheme satisfies the standard properties
such as completeness, soundness, and zero-knowledgeness.
\subsection{Construction}
\subsubsection*{Key Generation}
Given the security parameter as input, the key generator generates
the public $pk$ and the secret key $sk$ which satisfy the following properties.
The public key $pk$ consists of a positive integer $r\in\mathbb{N}$
and $m$ matrices $M_{0},M_{1,}\ldots,M_{m-1}\in\mathrm{M}_n(\mathbb{F})$.
The secret key $sk$ consists of $\alpha=(\alpha_{1},\ldots,\alpha_{m-1})\in\mathbb{F}^{m-1}$
such that $\mathrm{rank}(M)=r$, where $M=\sum_{i=1}^{m-1}\alpha_{i}M_{i}-M_{0}$.
\begin{comment}
\begin{remark}
Here, we do not describe the concrete implementation of the key generator.
We will describe it in the later section when we discuss about the
concrete implementation of the scheme.
\end{remark}
\end{comment}
\subsubsection*{Interactive Protocol}
A single elementary round of interactive protocol between a prover
$P(pk,sk)$ and a verivier $V(pk)$ is described as follows.
Similar to Courtois' ID scheme \cite{DBLP:conf/asiacrypt/Courtois01},
we also employ the hash function $H$ which acts as
a commitment with \emph{perfectly hiding} and
\emph{computational binding} properties.
\begin{itemize}[leftmargin=15mm]
\item [{Step{ }1:}] $P$ randomly generates $S_{0},S_{1},T_{0},T_{1}\in\mathrm{GL}_n(\mathbb{F})$
and $X_{0},X_{1}\in\mathrm{M}_n(\mathbb{F})$.
Next, $P$ randomly generates $\beta_{0}=(\beta_{0,1},\ldots,\beta_{0,m-1})\in\mathbb{F}^{m-1}$ and
$\beta_{1}=(\beta_{1,1},\ldots,\beta_{1,m-1})\in\mathbb{F}^{m-1}$, then
computes the following
\hspace{-10mm}
\begin{tabular}{l@{\hspace{-1mm}}|@{\hspace{-2mm}}l}%
\begin{minipage}[t]{43mm}\vspace{-5mm}
\begin{gather*}
N_{0} =\sum_{i=1}^{m-1}\beta_{0,i}M_{i} \\
U_{0,0} =T_{0}N_{0}S_{0}+X_{0} \\
U_{0,1} = T_{0}MS_{0}+U_{0,0} \\
R_{0} = (S_{0},T_{0},X_{0})
\end{gather*}
\end{minipage}
&
\begin{minipage}[t]{43mm}\vspace{-5mm}
\begin{gather*}
N_{1}=\sum_{i=1}^{m-1}\beta_{1,i}M_{i}\\
U_{1,0}=T_{1}N_1S_{1}+X_1 \\
U_{1,1}=T_{1}MS_{1}+U_{1,0} \\
R_{1}=(S_{1},T_{1},X_{1})
\end{gather*}
\end{minipage}
\end{tabular} \\
Finally, $P$ sends $Y=(Y_{0},Y_{1})$ to $V$ where the followings hold.
%
\begin{align}
Y_{0} & =(H(U_{0,0}),H(U_{0,1}),H(R_{0})) \\
Y_{1} & =(H(U_{1,0}),H(U_{1,1}),H(R_{1})).
\end{align}
\item [{Step{ }2:}] $V$ parses $Y_{0}$ and $Y_{1}$ as $Y_{0}=(Y_{0,0},Y_{0,1},Y_{0,2})$
and $Y_{1}=(Y_{1,0},Y_{1,1},Y_{1,2})$. Then, $V$ chooses randomly
$c\in\left\{ 0,1,2,3\right\} $ and sends $c$ to $P$.
\item [{Step{ }3:}] $P$ computes $Z_{0,0},Z_{0,1},Z_{1,0},Z_{1,1}$ according
to the value of $c$ as follows. \\
%
\ \\
\begin{tabular}{l|cc}
\hline
\textbf{Case} $c=0$:
& $Z_{0,0}=U_{0,0}$, & $Z_{1,0}=R_{1}$, \\
& $Z_{0,1}=U_{0,1}$, & $Z_{1,1}=\beta_{1}$. \\
\hline
%
\textbf{Case} $c=1$:
& $Z_{0,0}=R_{0}$, & $Z_{1,0}=R_{1}$, \\
& $Z_{0,1}=\beta_{0}$, & $Z_{1,1}=\beta_{1}+\alpha$. \\
\hline
%
\textbf{Case} $c=2$:
& $Z_{0,0}=R_{0}$, & $Z_{1,0}=R_{1}$, \\
& $Z_{0,1}=\beta_{0}+\alpha$ & $Z_{1,1}=\beta_{1}$ \\
\hline
\textbf{Case} $c=3$:
& $Z_{0,0}=R_{0}$, & $Z_{1,0}=U_{1,0}$, \\
& $Z_{0,1}=\beta_{0}$, & $Z_{1,1}=U_{1,1}$. \\
\hline
\end{tabular}
\\
\item [{Step{ }4:}] %
$V$ parses $Z=(Z_{0},Z_{1}$) into $Z_{0,0},Z_{0,1},Z_{1,0},Z_{1,1}$.
And then $V$ performs verification procedure according to the value
of $c$ as shown in Fig. \ref{fig:checking}.
If all corresponding checking equations hold, $V$ outputs $1$ (accept),
otherwise $V$ outputs $0$ (reject).
%
%
\end{itemize}
\begin{figure*}[htbp]
\fbox{
\scalebox{0.75}{%
\begin{tabular}{l|l}
\begin{minipage}[t]{1.3\columnwidth}%
\begin{description}
\item [{Case{ }$c=0$:}]\ \\
$Z_{1,0}$ is parsed as $Z_{1,0}=(\widetilde{S},\widetilde{T},\widetilde{X})$
and $Z_{1,1}$ is parsed as $Z_{1,1}=(\widetilde{\gamma}{}_{1},\ldots,\widetilde{\gamma}_{m-1})$.
\begin{itemize}
\item []$H(Z_{0,0})\stackrel{?}{=}Y_{0,0},\quad H(Z_{0,1})\stackrel{?}{=}Y_{0,1}$,
$\mathrm{rank}(Z_{0,1}-Z_{0,0})\stackrel{?}{=}r$,
\item []$\widetilde{S}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$, $\widetilde{T}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$H(Z_{1,0})\stackrel{?}{=}Y_{1,2},\quad H(\widetilde{U})\stackrel{?}{=}Y_{1,0}$,
where\vspace{-2mm}
\end{itemize} \vspace{1mm}
\hspace{10mm}
$
\begin{aligned}
\widetilde{U}
=\widetilde{T}\left(\sum_{i=1}^{m-1}\widetilde{\gamma}_{i}M_{i}\right)\widetilde{S}+\widetilde{X}.
\end{aligned}
$
\end{description}
\begin{description}
\item [{Case{ }$c=1$:}]\ \\
$Z_{0,0}$ is parsed as $Z_{0,0}=(\widehat{S},\widehat{T},\widehat{X})$
and $Z_{0,1}$ is parsed as $Z_{0,1}=(\widehat{\gamma}_1,\ldots\widehat{\gamma}_{m-1})$.
$Z_{1,0}$ is parsed as $Z_{1,0}=(\widetilde{S},\widetilde{T},\widetilde{X})$
and $Z_{1,1}$ is parsed as $Z_{1,1}=(\widetilde{\mu}_{1},\ldots\widetilde{\mu}_{m-1})$.
\begin{itemize}
\item []
$\widehat{S}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$, $\widehat{T}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$H(Z_{0,0})\stackrel{?}{=}Y_{0,2}, H(\widehat{U})\stackrel{?}{=}Y_{0,0}$,
where\vspace{-2mm}
\end{itemize}
\begin{gather}
\widehat{U} =\widehat{T}\left(\sum_{i=1}^{m-1}\widehat{\gamma}_{i}M_{i}\right)\widehat{S}+\widehat{X}.
\label{eq:widehat-U}
\end{gather}
\begin{itemize}
\item[] $\widetilde{S}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$\widetilde{T}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$H(Z_{1,0})\stackrel{?}{=}Y_{1,2}$,
$H(\widetilde{W}-\widetilde{T}M_{0}\widetilde{S})\stackrel{?}{=}Y_{1,1}$,
where\vspace{-2mm}
\end{itemize}
\begin{gather}
\widetilde{W} =\widetilde{T}\left(\sum_{i=1}^{m-1}\widetilde{\mu}_{i}M_{i}\right)\widetilde{S}+\widetilde{X}.
\label{eq:widetilde-W}
\end{gather}
\end{description}
\end{minipage}
%
%
&
\begin{minipage}[t]{1.3\columnwidth}%
\begin{description}
\item [{Case{ }$c=2$:}]\ \\
$Z_{0,0}$ is parsed as $Z_{0,0}=(\widehat{S},\widehat{T},\widehat{X})$
and $Z_{0,1}$ is parsed as $Z_{0,1}=(\widehat{\mu}_{1},\ldots\widehat{\mu}_{m-1})$.
$Z_{1,0}$ is parsed as $Z_{1,0}=(\widetilde{S},\widetilde{T},\widetilde{X})$
and $Z_{1,1}$ is parsed as $Z_{1,1}=(\widetilde{\gamma}_{1},\ldots\widetilde{\gamma}_{m-1})$.
\begin{itemize}
\item []
$\widehat{S}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$, $\widehat{T}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$H(Z_{0,0})\stackrel{?}{=}Y_{0,2}$,
$H(\widehat{W}-\widehat{T}M_{0}\widehat{S})\stackrel{?}{=}Y_{0,1}$,
where
\end{itemize}
\begin{gather}
\widehat{W} =\widehat{T}\left(\sum_{i=1}^{m-1}\widehat{\mu}_{i}M_{i}\right)\widehat{S}+
\widehat{X}.
\label{eq:widehat-W}
\end{gather}
\begin{itemize}
\item []
$\widetilde{S}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$\widetilde{T}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$.
$H(Z_{1,0})\stackrel{?}{=}Y_{1,2},$
$H(\widetilde{U})\stackrel{?}{=}Y_{1,0}$,
where
\end{itemize}
\begin{gather}
\widetilde{U} =\widetilde{T}\left(\sum_{i=1}^{m-1}\widetilde{\gamma}_{i}M_{i}\right)\widetilde{S}+\widetilde{X}.
\label{eq:widetilde-U}
\end{gather}
\end{description}
%
%
\begin{description}
\item [{Case{ }$c=3$:}]\ \\
$Z_{0,0}$ is parsed as $Z_{0,0}=(\widehat{S},\widehat{T},\widehat{X})$
and $Z_{0,1}$ is parsed as $Z_{0,1}=(\widehat{\gamma},\ldots\widehat{\gamma}_{m-1})$.
\begin{itemize}
\item []
$\widehat{S}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$\widehat{T}\stackrel{?}{\in}\mathrm{GL}_n(\mathbb{F})$,
$H(Z_{0,0})\stackrel{?}{=}Y_{0,2}$,
$H(\widehat{U})\stackrel{?}{=}Y_{0,0}$,
where
\end{itemize}
$\qquad\qquad
\begin{aligned}
\widehat{U} =\widehat{T}\left(\sum_{i=1}^{m-1}\widehat{\gamma}_{i}M_{i}\right)\widehat{S}+\widehat{X}.
\end{aligned}
$
\begin{itemize}
\item []$H(Z_{1,0})\stackrel{?}{=}Y_{1,0},\quad H(Z_{1,1})\stackrel{?}{=}Y_{1,1}$,
$\mathrm{rank}(Z_{1,1}-Z_{1,0})\stackrel{?}{=}r$,
\end{itemize}
\end{description}
%
\end{minipage} \\
\end{tabular}
}
}
\caption{Checking equations performed by the verifier ($V$) in Step 4 of the elementary round of interactive protocol.}
\label{fig:checking}
\end{figure*}
\begin{remark}
The response $Z$ is said to be a \emph{valid response with respect to challenge}
$c$ if all checking equations in the verifier side corresponding
to the value of $c$ hold.
\end{remark}
\
\begin{remark}
A full identification scheme consists of $\ell$ repetitions of the
single elementary round of interactive protocol and the verifier will
accept the prover if and only if $V$ outputs $1$ in all $\ell$
rounds.
\end{remark}
\begin{remark}
Here we assume that the length of the input into the hash function $H$
is larger than that of the output, that is why we can assume that
$H$ acts as a perfectly hiding commitment.
We also assume that $H$ is collision resistant, i.e., for any
polynomial algorithm, it is hard to find two distinct inputs
with the same output.
That is why we can assume that $H$ acts a computational binding
commitment.
Any common standard hash functions such as SHA-128, SHA-256, SHA-512
is assumed to have these properties.
\end{remark}
\subsection{Completeness}
Here we show that any prover who possesses the secret key and follows
the procedure of the honest prover will always be accepted by the
verifier.
\begin{theorem}[Completeness]
\label{thm:completeness}Let $P$ be a prover who possesses the secret
key $sk$ corresponding to the public key $pk$ of our proposed identification
scheme. Let $P$ generate $Y$ in Step 1 according to the described
procedure and send it to the verifier. Then for any received challenge
$c\in\{0,1,2,3\}$ from the verifier, if $P$ computes $Z$ according
to described procedure, $Z$ is a valid response with respect to challenge
$c$.
\end{theorem}
In order to prove the above theorem, it is sufficient to show that
for each challenge $c\in\{0,1,2,3\}$, $Z$ which is generated accordingly
in the procedure of the prover will satisfy all the corresponding
checking equations on the verifier side.
See Section \ref{subsec:Proof-of-Completeness}
for the detailed proof.
\subsection{Soundness }
In order to prove the soundness of our proposed scheme, we will use
the following proposition.
\begin{proposition}%
\label{prop:three-combinations}
Let $Y$ denote the value sent by
the prover in the Step 1 to the verifier and let $Z^{(c)}$ denote
the valid response with respect to the challenge $c\in\{0,1,2,3\}$.
Then, from $Y$ and any three combinations of elements from the set
$\{Z^{(0)},Z^{(1)},Z^{(2)},Z^{(3)}\}$ we can efficiently compute
the solution of the search MinRank problem represented by the public key.
\end{proposition}
We describe the detailed proof of above proposition in Section \ref{subsec:Proof-of-Three-Combinations}.
Based on above proposition, we can easily see that the following
corollary holds.
\begin{corollary}%
\label{cor:invalid}
If the public key has no corresponding secret key,
the success probability of any prover
to be accepted by the verifier in all $\ell$ rounds
of a full identification at most $1/2^\ell$.
\end{corollary}
The security of our scheme against key-only impersonation attack,
i.e., soundness, is based on the hardness of decisional MinRank problem,
as stated by the following theorem.
\begin{theorem}\label{thm:soundness}
Let $\mathcal{A}$ be an algorithm such that given the public key
$pk$, it is accepted in all $\ell$ rounds of the full identification
protocol with probability $\varepsilon_{\mathcal{A}}\geqq\frac{1}{2^{\ell}}$,
where the probability is taken over the random coins of $\mathcal{A}$,
the key generator, and the verifier. Then, we can construct an algorithm
which $(t,\varepsilon)$-solves the decisional MinRank problem associated
with the finite field $\mathbb{F}$ and $r,m,n\in\mathbb{N}$
such that the following holds. \vspace{-5mm}
\begin{align*}
\varepsilon =\varepsilon_{\mathcal{A}}-\frac{1}{2^{\ell}},
\quad
t = t_{\mathcal{A}},
\end{align*}
where $t_{\mathcal{A}}$ is the maximum total time of $\mathcal{A}$
interacting in one full identification protocol.
\end{theorem}
\begin{corollary}
If the decisional MinRank problem is $(t,\varepsilon)$-hard, then
the success probability of
any adversary attempting to impersonate a prover
without secret key within $t$ time units
is upper-bounded by $\varepsilon+1/2^\ell$.
\end{corollary}
\subsection{Zero-Knowledgeness}
The following theorem is to guarantee that no knowledge on the secret leaked by
communication with the prover.
\begin{theorem}[Zero-Knowledgeness]%
\label{thm:zk}
For any verifier $V$, there exists an algorithm $M$ which given input
the public key $pk$, perfectly simulates the view of verifier with the
same distribution as the view of $V$ engaging with the prover possessing
$pk$ and the secret key $sk$.
\end{theorem}
\section{Proofs of Main Theorems }
\subsection{Proof of Theorem \ref{thm:completeness}\label{subsec:Proof-of-Completeness}}
It is sufficient to show that for each challenge $c\in\{0,1,2,3\}$,
$Z$ which is generated accordingly in the procedure of the prover
will satisfy all the corresponding checking equations on the verifier
side.
Let us check for each case of challenge.
\begin{description
\item[Case{ }$c=0$:]
Since $Z_{0,0}=U_{0,0}$ and $Y_{0,0}=H(U_{0,0})$,
it is obvious that $H(Z_{0,0})=H(U_{0,0})=Y_{0,0}$ holds. Similarly,
since $Z_{0,1}=U_{0,1}$ and $Y_{0,1}=H(U_{0,1})$ holds, it is obvious
that $H(Z_{0,1})=H(U_{0,1})=Y_{0,1}$. Since $Z_{0,0}=U_{0,0}$ and
$Z_{0,1}=U_{0,1}$, the followings hold.
\begin{align*}
\mathrm{rank}\left(Z_{0,1}-Z_{0,0}\right)
& =\mathrm{rank}\left(U_{0,1}-U_{0,0}\right) \\
& =\mathrm{rank}\left(T_{0}MS_{0}\right) \\
& \stackrel{(a)}{=}\mathrm{rank}(M) = r.
\end{align*}
Eq. $(a)$ holds since $T_{0}$ and $S_{0}$ are invertible matrices.
Since $Z_{1,0}=R_1$ and $Y_{1,2}=H(R_1)$, it is obvious
that $H(Z_{1,0})=H(R_1)=Y_{1,2}$ holds.
Also, we can easily see that $(\widetilde{S},
\widetilde{T},\widetilde{X})=(S_1,T_1,X_1)$.
Since $S_1, T_1$ are invertible matrices,
so are $\widetilde{S}, \widetilde{T}$.
Since $Z_{1,1}=\beta_1$, it is obvious that
$(
\widetilde{\gamma}_1,\ldots,\widetilde{\gamma}_{m-1}
)=
(
\beta_{1,1},\ldots,\beta_{1,m-1}.
)
$
Thus, the followings hold.
\begin{align*}
\widetilde{U}
& =\widetilde{T}
\left(
\sum_{i=1}^{m-1}\widetilde{\gamma}_{i}M_{i}
\right)
\widetilde{S}+\widetilde{X} \\
& = T_1\left(\sum_{i=1}^{m-1}\beta_{1,i}M_{i}\right)S_1+X_1 \\
& = T_1 N_1 S_1+X_1
= U_{1,0}.
\end{align*}%
Hence, since $Y_{1,0}=H(U_{1,0})$,
$H(\widetilde{U})=H(U_{1,0})=Y_{1,0}$ holds.
\item [{Case{ }$c=1$:}]%
Since $Z_{0,0}=R_0$ and $Y_{0,2}=H(R_0)$, it is obvious that
$H(Z_{0,0})=Y_{0,2}$ holds.
Hence, one can see that
$(\widehat{S},\widehat{T},\widehat{X})=
({S}_0,{T}_0,{X}_0)$
holds.
Since $Z_{0,1}=\beta_0$, it is obvious that
$(\widehat{\gamma}_{1},\ldots,\widehat{\gamma}_{m-1})
=(\beta_{0,1},\ldots,\beta_{0,m-1})$ holds.
Thus, the following holds.
\begin{align*}
\widehat{U}
& =\widehat{T}
\left(
\sum_{i=1}^{m-1}\widehat{\gamma}_{i}M_{i}
\right)
\widehat{S}+\widehat{X} \\
& = T_0\left(\sum_{i=1}^{m-1}\beta_{0,i}M_{i}\right)S_0+X_0 \\
& = T_0 N_0 S_0+X_0
= U_{0,0}.
\end{align*}%
Hence, since $Y_{0,0}=H(U_{0,0})$, automatically
$H(\widehat{U})=H(U_{0,0})=Y_{0,0}$ holds.
Next, since $Z_{1,0}=R_1$ and $Y_{1,2}=H(R_1)$, it is obvious that
$H(Z_{1,0})=Y_{1,2}$ holds.
Hence, one can see that
$(\widetilde{S},\widetilde{T},\widetilde{X})=
({S}_1,{T}_1,{X}_1)$
holds.
Since $Z_{1,1}=\beta_1+\alpha$, it is obvious that
$(\widetilde{\mu}_{1},\ldots,\widetilde{\mu}_{m-1})
=(\beta_{1,1}+\alpha_1,\ldots,\beta_{1,m-1}+\alpha_{m-1})$ holds.
Thus, the following holds.
\begin{align*}
\widetilde{W}-\widetilde{T}M_0\widetilde{S}
& =\widetilde{T}
\left(
\sum_{i=1}^{m-1}\widetilde{\mu}_{i}M_{i}
\right)
\widetilde{S}+\widetilde{X}
-\widetilde{T}M_0\widetilde{S}
\\
& = T_1 \left(
\sum_{i=1}^{m-1}(\beta_{1,i}+\alpha_i)M_{i}
\right)S_1+X_1
\\
& \quad
- T_1 M_0 S_1 \\
& = T_1 N_1 S_1 + T_1 M S_1 + T_1 M_0 S_1 +X_1 \\
& \quad - T_1 M_0 S_1 \\
& = T_1 N_1 S_1 + X_1
= U_{1,0}
\end{align*}%
Hence, since $Y_{1,0}=H(U_{1,0})$, automatically
$H(\widetilde{W}-\widetilde{T}M_0\widetilde{S})=H(U_{1,0})=Y_{1,0}$
holds.
\item [{Case{ }$c=2$:}]%
This case is similar to the case $c=1$ with additional notes as follows:
\begin{itemize}
\item any variable in the form of $\widehat{*}$ notation
switches with the resembling variable
in the form of $\widetilde{*}$ notation,
\item any variable in the form of ${*}_0$ notation
switches with the resembling variable
in the form of ${*}_1$ notation,
\item for any numeric $j$, any variable in the form of $*_{0,j}$ notation
switches with the resembling variable in the form of
$*_{1,j}$ notation.
\end{itemize}
\item [{Case{ }$c=3$:}]%
This case is similar to the case $c=0$ with the same additional
notes as in the case $c=2$.
\end{description}
\subsection{Proof of Proposition \ref{prop:three-combinations}\label{subsec:Proof-of-Three-Combinations}}
It is sufficient to show that from $Y$ and any combination of three
elements from the set of the valid responses $\{Z^{(0)},Z^{(1)},Z^{(2)},Z^{(3)}\}$,
we can compute $\alpha=(\alpha_{1},\ldots,\alpha_{m-1})\in\mathbb{F}^{m-1}$
such that $\mathrm{rank}\left(\sum_{i=1}^{m-1}\alpha_{i}M_{i}-M_{0}\right)=r$
holds, where $r$ and $M_{0},\ldots,M_{m-1}$ are generated by the
key generation algorithm as elements of the public key.
\begin{remark}
Note that
in our proposed scheme, we assume that $H$ has computational
binding property.
Hence, we can assume that for any polynomial time algorithm,
if $H(a)=H(b)$, then $a=b$ must hold
except with negligible probability.
\end{remark}
\paragraph*{Case 1: $Y$ and $(Z^{(0)},Z^{(1)},Z^{(2)})$. }
Let $Z_{0,0}^{(1)}$ be parsed as $Z_{0,0}^{(1)}=(\widehat{S}^{(1)},\widehat{T}^{(1)},\widehat{X}^{(1)})$
and $Z_{0,1}^{(1)}$ be parsed as $Z_{0,1}^{(1)}=(\widehat{\gamma}_{1},\ldots\widehat{\gamma}_{m-1})$.
Also let $Z_{0,0}^{(2)}$ be parsed as $Z_{0,0}^{(2)}=(\widehat{S}^{(2)},\widehat{T}^{(2)},\widehat{X}^{(2)})$
and $Z_{0,1}^{(2)}$ be parsed as $Z_{0,1}^{(2)}=(\widehat{\mu}_{1},\ldots\widehat{\mu}_{m-1})$.
Since the following holds:
\begin{align*}
H\left(\widehat{S}^{(1)},\widehat{T}^{(1)},\widehat{X}^{(1)}\right) & =H\left(\widehat{S}^{(2)},\widehat{T}^{(2)},\widehat{X}^{(2)}\right)=Y_{0,2},
\end{align*}
we can define as follows: $(\widehat{S},\widehat{T},\widehat{X}):=(\widehat{S}^{(1)},\widehat{T}^{(1)},\widehat{X}^{(1)})=(\widehat{S}^{(2)},\widehat{T}^{(2)},\widehat{X}^{(2)})$.
From $H(Z_{0,0}^{(0)})=Y_{0,0}$ and Eq. \eqref{eq:widehat-U}, we obtain as follows.
\begin{align}
Y_{0,0} & =H(Z_{0,0}^{(0)})
= H\left(\widehat{T}\left(\sum_{i=1}^{m-1}\widehat{\gamma}_{i}M_{i} \right)
\widehat{S}+\widehat{X}
\right)
\label{eq:y00-z00(0)-u(1)} \\
& \quad \Rightarrow Z_{0,0}^{(0)}=\widehat{T}\left(\sum_{i=1}^{m-1}
\widehat{\gamma}_{i}M_{i}\right)\widehat{S}+\widehat{X},
\nonumber
\end{align}
Similarly, from $H(Z_{0,1}^{(0)})=Y_{0,1}$ and Eq. \eqref{eq:widehat-W},
we also have the followings hold.
\begin{align}
%
Y_{0,1}
& =H(Z_{0,1}^{(0)})
\nonumber \\
&
=
H\left(\widehat{T}\left(\sum_{i=1}^{m-1}\widehat{\mu}_{i}M_{i}
\right)\widehat{S}+\widehat{X}-\widehat{T}M_{0}\widehat{S}\right) \\
& \Rightarrow Z_{0,1}^{(0)}=\widehat{T}\left(\sum_{i=1}^{m-1}\widehat{\mu}_{i}M_{i}\right)\widehat{S}+\widehat{X}-\widehat{T}M_{0}\widehat{S} \nonumber
\end{align}
Finally, we have the followings hold.
\begin{align*}
\mathrm{rank}(Z_{0,1}^{(0)}-Z_{0,0}^{(0)}) & =\mathrm{rank}\left(\widehat{T}\left(\sum_{i=1}^{m-1}(\widehat{\mu}_{i}-\widehat{\gamma}_{i})M_{i}-M_{0}\right)\widehat{S}\right) \\
& \stackrel{(a)}{=}\mathrm{rank}\left(\sum_{i=1}^{m-1}(\widehat{\mu}_{i}-\widehat{\gamma}_{i})M_{i}-M_{0}\right),
\end{align*}
where
Eq. $(a)$ holds since $\widehat{S},\widehat{T}$ are non-singular.
Therefore, we can set $\alpha_{i}=\widehat{\mu}_{i}-\widehat{\gamma}_{i}$
for $i\in[1,m-1]$, since
$\mathrm{rank}(Z_{0,1}^{(0)}-Z_{0,0}^{(0)})=r$ holds.
\paragraph*{Case 2: $Y$ and $(Z^{(0)},Z^{(2)},Z^{(3)})$. }
Similar to Case 1. The only diffference is that all relations
and components of $Z^{(1)}$ in Case
1 are substituted by those of $Z^{(3)}$.
\paragraph*{Case 3: $Y$ and $(Z^{(1)},Z^{(2)},Z^{(3)})$. }
Let $Z_{1,0}^{(2)}$ be parsed as $Z_{1,0}^{(2)}=(\widetilde{S}^{(2)},\widetilde{T}^{(2)},\widetilde{X}^{(2)})$
and $Z_{1,1}^{(2)}$ be parsed as $Z_{1,1}^{(2)}=(\widetilde{\gamma}_{1},\ldots\widetilde{\gamma}_{m-1})$.
Also let $Z_{1,0}^{(1)}$ be parsed as $Z_{1,0}^{(1)}=(\widetilde{S}^{(1)},\widetilde{T}^{(1)},\widetilde{X}^{(1)})$
and $Z_{1,1}^{(1)}$ be parsed as $Z_{1,1}^{(1)}=(\widetilde{\mu}_{1},\ldots\widetilde{\mu}_{m-1})$.
Since the following holds:
\begin{align*}
H\left(\widetilde{S}^{(1)},\widetilde{T}^{(1)},\widetilde{X}^{(1)}\right) & =H\left(\widetilde{S}^{(2)},\widetilde{T}^{(2)},\widetilde{X}^{(2)}\right)=Y_{1,2},
\end{align*}
we can define as follows: $(\widetilde{S},\widetilde{T},\widetilde{X}):=(\widetilde{S}^{(1)},\widetilde{T}^{(1)},\widetilde{X}^{(1)})=(\widetilde{S}^{(2)},\widetilde{T}^{(2)},\widetilde{X}^{(2)})$.
From $H(Z_{1,0}^{(3)})=Y_{1,0}$ and Eq. \eqref{eq:widetilde-U}, we obtain as follows.
\begin{align}
Y_{1,0} & =H(Z_{1,0}^{(3)})
= H\left(\widetilde{T}\left(\sum_{i=1}^{m-1}\widetilde{\gamma}_{i}M_{i} \right)
\widetilde{S}+\widetilde{X}
\right)
\\
& \quad \Rightarrow Z_{1,0}^{(3)}=\widetilde{T}\left(\sum_{i=1}^{m-1}
\widetilde{\gamma}_{i}M_{i}\right)\widetilde{S}+\widetilde{X},
\nonumber
\end{align}
Similarly, from $H(Z_{1,1}^{(3)})=Y_{1,1}$ and Eq. \eqref{eq:widetilde-W},
we also have the followings hold.
\begin{align}
%
Y_{1,1}
& =H(Z_{1,1}^{(3)}) \nonumber \\
&
=
H\left(\widetilde{T}\left(\sum_{i=1}^{m-1}\widetilde{\mu}_{i}M_{i}
\right)\widetilde{S}+\widetilde{X}-\widetilde{T}M_{0}\widetilde{S}\right) \\
& \Rightarrow Z_{1,1}^{(3)}=\widetilde{T}\left(\sum_{i=1}^{m-1}\widetilde{\mu}_{i}M_{i}\right)\widetilde{S}+\widetilde{X}-\widetilde{T}M_{0}\widetilde{S} \nonumber
\end{align}
Finally, we have the followings hold.
\begin{align*}
\mathrm{rank}(Z_{1,1}^{(3)}-Z_{1,0}^{(3)}) & =\mathrm{rank}\left(\widetilde{T}\left(\sum_{i=1}^{m-1}(\widetilde{\mu}_{i}-\widetilde{\gamma}_{i})M_{i}-M_{0}\right)\widetilde{S}\right) \\
& \stackrel{(a)}{=}\mathrm{rank}\left(\sum_{i=1}^{m-1}(\widetilde{\mu}_{i}-\widetilde{\gamma}_{i})M_{i}-M_{0}\right),
\end{align*}
where
Eq. $(a)$ holds since $\widetilde{S},\widetilde{T}$ are non-singular.
Therefore, we can set $\alpha_{i}=\widetilde{\mu}_{i}-\widetilde{\gamma}_{i}$
for $i\in[1,m-1]$, since
$\mathrm{rank}(Z_{1,1}^{(3)}-Z_{1,0}^{(3)})=r$ holds.
\paragraph*{Case 4: $Y$ and $(Z^{(0)},Z^{(1)},Z^{(3)})$. }
Similar to Case 3. The only diffference is that all relations
and components of $Z^{(2)}$ in Case
1 are substituted by those of $Z^{(0)}$.
\subsection{Proof Sketch of Corollary \ref{cor:invalid}}
Recall that based on Proposition \ref{prop:three-combinations},
we know that in any single round, if the prover can answer correctly
three out of four possible challenges from the verifier,
it means that the prover knows the secret key corresponding
public key. Thus, in the case that the public key has no
corresponding valid secret key, even a prover with unbounded
resources must not be able to answer correctly more than
two out of four possible challenges in any single round.
Otherwise, it will contradict with the assumption that
the public key
that the public key has no corresponding secret key.
\subsection{Proof Sketch of Theorem \ref{thm:soundness}}
Let define algorithm $\mathcal{D^{\mathsf{InputGen}}}(\mathbb{F},r,m,n)$
as follows. First, $\mathcal{D}$ retrieves inputs from the oracle
$\mathsf{InputGen}$ in the form of $m$ $n$-square matrices over
the finite field $\mathbb{F}$: $M_{0},\ldots,M_{m-1}$. Then, $\mathcal{D}$
simulates the key generation algorithm of the identification scheme
by setting the public key $pk$ as $r$ and $M_{0},\ldots,M_{m-1}$.
Next, $\mathcal{D}$ inputs $pk$ to $\mathcal{A}$ and runs $\mathcal{A}$
as the prover and $\mathcal{D}$ acts as the honest verifier. If $\mathcal{A}$
successfully gives valid responses in all $\ell$ rounds of the full
identification protocol, $\mathcal{D}$ outputs $1$, otherwise, $\mathcal{D}$
outputs $0$.
Note that if $\mathsf{InputGen}$ is $\mathsf{IGen}$, the probability of $\mathcal{D}$
outputs $1$ is exactly $\varepsilon_\mathcal{A}$.
Meanwhile, when $\mathsf{InputGen}$ is $\mathsf{LossyGen}$,
based on Corollary \ref{cor:invalid}, the probability of
$\mathsf{D}$ outputs $1$ is at most $1/2^\ell$.
Thus, denoting the system parameters $(\mathbb{F},r,m,n)$ as $\mathrm{par}$,
we obtain as follows.
\begin{align*}
\left|\Pr[\mathcal{D}^{\mathsf{IGen}}(\mathrm{par})=1]
-\Pr[\mathcal{D}^{\mathsf{LossyGen}}(\mathrm{par})=1] \right |
& \geqq \varepsilon_\mathcal{A}- \frac{1}{2^\ell}.
\end{align*}
This proves Theorem \ref{thm:soundness}.
\subsection{Proof Idea of Theorem \ref{thm:zk}}
It is sufficient to prove that
given any $c\in\{0,1,2,3\}$, we can
create valid response $Z_{0,0},Z_{0,1},Z_{1,0},Z_{1,1}$
and the commitment $Y_0,Y_1$ without using secret key
such that their distribution is the same as
the distribution of the response and commitment
generated by a honest prover who possesses
valid secret key. Note that we can put the
responses and commitment
into two independent groups:
$(Y_0,Z_{0,0},Z_{0,1})$ and
$(Y_1,Z_{1,0},Z_{1,1})$,
such that
each group is corresponding to the
set of responses and commitment
in Courtois' ID scheme~\cite{DBLP:conf/asiacrypt/Courtois01}.
Hence, it is easy to see that
we can apply the proof of zero-knowledge
for Courtois' ID scheme into our proposed
scheme.
\section{Parameter Selections}
\subsection{Complexity of MinRank Problem}
\input{complexity.tex}
\input{complexity-nakamura-20220412.tex}
\subsection{Communication Costs}
We will estimate the communication costs based on the assumption that
we use random seed and pseudorandom generator to generates $S_{0},S_{1},T_{0},T_{1},X_{0},X_{1},\beta_{0},\beta_{1}$.
Let $Z^{(c)}$ denote the valid response of the prover with respect to challenge
$c$ for any $c\in\{0,1,2,3\}$. For simplicity, here we assume that
all matrices are $n$-square matrices and $\mathbb{F=\mathbb{F}}_{q}$,
where $q$ is a power of some prime. Thus, we have as follows.
\begin{align*}
|Z^{(0)}|=|Z^{(3)}| & \approx2n^{2}\log_{2}q+|\mathrm{seed}_{\overline{STX}}|+|\mathrm{seed}_{\beta}|, \\
|Z^{(1)}|=|Z^{(2)}| & \approx2|\mathrm{seed}_{\overline{STX}}|+|\mathrm{seed}_{\beta}|+(m-1)\log_{2}q,
\end{align*}
where $\mathrm{seed}_{\overline{STX}}$ is the seed for generating
$(S_{0},T_{0},X_{0})$ or $(S_{1},T_{1},X_{1})$ and $\mathrm{seed}_{\beta}$
is the seed for generating $\beta_{0}$ or $\beta_{1}$.
\begin{comment}
\subsubsection*{Estimation for $128$-bit security}
For achieving
$128$-bit security, we need to have $128$ repetitions of the elementary
round. Let $\#Z_{1/2}$ denote the average of the total cost of sending
all responses in our proposed identification scheme. Thus, we can estimate $\#Z_{1/2}$ as follows.
\begin{align*}
\#Z_{1/2} & =\frac{128}{4}(|Z^{(0)}|+|Z^{(1)}|+|Z^{(2)}|+|Z^{(3)}|) \\
& =\frac{128}{4}(2|Z^{(0)}|+2|Z^{(1)}|)=64(|Z^{(0)}|+|Z^{(1)}|) \\
& \approx64(2n^{2}\log_{2}p+|\mathrm{seed}_{\overline{STX}}|+|\mathrm{seed}_{\beta}| \\
& \quad +2|\mathrm{seed}_{\overline{STX}}|+|\mathrm{seed}_{\beta}|+(m-1)\log_{2}p)
\end{align*}
In Courtois' identification scheme \cite{DBLP:conf/asiacrypt/Courtois01},
for achieving $128$-bit security, we need to have $219$ repetitions
of the elementary round since the cheating probability for each elementary
round is $2/3$.\footnote{This number $219$ is based on the fact that $\ell=219$ is the least
integer such that $(2/3)^{\ell}\leqq2^{-128}$ holds.} In Courtois' identification scheme, we have $c\in\{0,1,2\}$. Based
on the same assumption as above on the matrices, the field, and the
random seeds, we have as follows.
\begin{align*}
|Z^{(0)}| & \approx2n^{2},
|Z^{(1)}| \approx|\mathrm{seed}_{\overline{STX}}|+|\mathrm{seed}_{\beta}|, \\
|Z^{(2)}| & \approx|\mathrm{seed}_{\overline{STX}}|+(m-1)\log_{2}p.
\end{align*}
Let $\#Z_{2/3}$ denote the average of the total cost of sending
all responses in Courtois' identification scheme. Thus, we can estimate
$\#Z_{2/3}$ as follows.
\begin{align*}
\#Z_{2/3} & =\frac{219}{3}(|Z^{(0)}|+|Z^{(1)}|+|Z^{(2)}|) \\
& =73(|Z^{(0)}|+|Z^{(1)}|+|Z^{(2)}|) \\
& \approx73(2n^{2}\log_{2}p+|\mathrm{seed}_{\overline{STX}}|+|\mathrm{seed}_{\beta}| \\
& \quad
+ |\mathrm{seed}_{\overline{STX}}|+(m-1)\log_{2}p)
\end{align*}
For $128$-bit security, it is common to assume that all random seeds
have $128$ bit length. Thus, we can finalize our estimation as follows.
\begin{align*}
\#Z_{1/2} & \approx128\left(n^{2}+\frac{1}{2}(m-1)\right)\log_{2}p+64\times5\times128 \\
& =128\left(n^{2}+\frac{1}{2}(m-1)\right)\log_{2}p+40960, \\
\#Z_{2/3} & \approx146\left(n^{2}+\frac{1}{2}(m-1)\right)\log_{2}p+73\times3\times128 \\
& =146\left(n^{2}+\frac{1}{2}(m-1)\right)\log_{2}p+28032.
\end{align*}
\end{comment}
Let $\#R_{1/2}$ and $\#R_{2/3}$ denote
the necessary number of rounds to achieve $\ell$-bit
security for our scheme and Curtois' respectively.
It is easy to see that
$\#R_{1/2}=\ell$ and $\#R_{2/3}=2/3\lceil\ell/(\log_2{3}-1)\rceil$.
Assuming that the seeds for $\ell$-bit security are $\ell$ bits,
we have the following general equations for estimating the average
total
communication costs for $\ell$-bit security.
\begin{align*}
\#Z_{1/2} & \approx
\ell
\left(
\left(
n^2 + (m-1)
\right)\log_2 q
+ \frac{5\ell}{2}
\right), \\
\#Z_{2/3} & \approx
\frac{2}{3}
\left\lceil
\frac{\ell}{\log_2 3-1}
\right\rceil
\left(
\left(
n^2 + (m-1)
\right)\log_2 q
+ \frac{3\ell}{2}
\right),
\end{align*}
where $\#Z_{1/2}$ and $\#Z_{2/3}$
denote the average total communication costs of our proposed scheme
and that of Curtois' respectively.
\subsection{Security Parameters}
\input{securityparameters-arxiv.tex}
\section{Conclusion}
\input{conclusion.tex}
\vspace{1mm}
\noindent{\bf{Acknowledgements}}
This work was supported by JST CREST Grant Number
JPMJCR2113, JSPS KAKENHI Grant Number JP19K20266, JP20K19802,
JP20K03741, JP18H01438, and JP18K11292.
\bibliographystyle{plain}
|
\section*{Reduction of the coupled dynamic equations}
\label{app:equations-reduction}
The system of two equations for the surface membrane and the reservoir
\begin{equation}
\systeme[][:]{\partial_t n = D \Delta n+\delta(z)[k_D m - k_An(\vec{r},z=0,t)]:
\partial_t m= \mu\nabla^2\frac{\delta \mathcal{F}}{\delta m}+k_A n(\vec{r},z=0,t) - k_Dm},\\
\label{eq:coupledBath-membrane}
\end{equation}
may be reduced to a single equation. By defining the Fourier transform of a function $A_x$ as $A_{x}=\int \frac{d{q}}{\sqrt{2\pi}}e^{i{x}{q}}\tilde A_{q}$, we write
\begin{equation}
\tilde n(\vec{q},z=0,w) = \frac{k_D \tilde m(\vec{q},w)\tilde I(\vec{q},w)}{\sqrt{2\pi}+k_A\tilde I(\vec{q},w)}, \\
\label{eq:nq0w}
\end{equation}
where
$\tilde I(\vec{q},w)=\int dk_z\frac{1}{iw +D(k_z^2+\vec{q}^2)}$. Replacing (\ref{eq:nq0w}) in (\ref{eq:coupledBath-membrane}) for $\tilde I\neq 0$ we obtain
\[
\partial_t m= \mu\nabla^2\frac{\delta \mathcal{F}}{\delta m}+\mathcal{F}^{-1}\Big[\Big(\tilde K(\vec{q},w) - 1\Big)k_D\tilde m(\vec{q},w))\Big],
\] with $
\tilde K(\vec{q}, w)=\frac{1}{\frac{\sqrt{2\pi}}{\tilde I(\vec{q},w)}+k_A }.$
The space-Fourier transform of this quantity is
\begin{equation}
\hat K(\vec{q},t)=\kappa e^{-q^2Dt}\Big(\frac{1}{\sqrt{\pi \kappa t}}-\text{Erfcx}(\sqrt{\kappa t})\Big),\\
\label{eq:K(q,t)}
\end{equation}
where ${\kappa}=\pi\frac{k^2_{A}}{2D}$, and $\text{Erfcx}(t)$ is the standard error function $\text{Erfcx}(t)=e^{t^2}\frac{2}{\sqrt{\pi}}\int_t^\infty e^{-z^2}dz$.
\section*{Numerical solution}
We discretize Eq.~\ref{eq:DynamicEvolution} so that the surface particles' density $m$ is a matrix of elements $(i,j)$ at each simulation step $p$, with a given initial condition $m(i,j,p=0)$. To obtain the evolution of $m$ we use the standard Euler semi-implicit integration method for the time variable and a five-point discretization of the Laplacian terms in Fourier space for the spatial coordinates.
Compared to a standard Cahn-Hilliard equation we have to consider an extra term involving the integral of our kernel. We approximate the integral in Eq.~\ref{eq:DynamicEvolution} at a simulation step $p$ by the Riemann sum
$\int_{0}^{t}dt'\hat K(\vec{q},t-t')m(\vec{q},t')\simeq \sum_{p'=M}^{p-1}\hat K(q_i,q_j,p-p')m(q_i,q_j,p')\Delta t$, where $\hat K$ is given by Eq.~\ref{eq:K(q,t)}, and $M$ is the number of previous configurations that will be taken into account in the calculation.
\begin{figure}
\begin{center}
{\includegraphics[width=1\linewidth]{PD_ic.png}}
\end{center}
\caption{Snapshot of systems that evolved during $10^4$ simulation steps for different values of the absorption and desorption rates, $k_A$ and $k_D$, respectively, at fixed diffusion constant $D = 0.1$. We consider two different initial conditions (shown on the insets). The effect of matter exchange between the membrane and the bath is observed
at relatively short times, especially for larger values of $k_D$. The images highlighted with dashed gray lines correspond to systems that evolved under the instantaneous kernel (Eq. 5).}
\label{fig:initial_conditions}
\end{figure}
\section*{Short-time effects}
To further verify the multi-domain state of the system when matter exchange is considered, we studied
the cases where the initial condition is a single domain, either a bubble or a stripe, as shown in Fig.~\ref{fig:initial_conditions}. The effect of the kernel is immediately observed in the simulations for both considered kernels.
\end{document}
|
\section{Introduction}
\label{sec:introduction}
{Security} and privacy are vital to emerging cyber-physical systems due to the increasing reliance of many applications on shared cyber infrastructures \cite{CPS1}, \cite{CPS2}.
In the framework of discrete event systems (DESs) \cite{2010Lafortunebook}, \cite{2019Wonhambook}, \cite{2020Chrisbook}, where the underlying system is usually modeled as a finite automaton or a Petri net \cite{2017Ma}, \cite{2017Zhu}, \cite{2018Zhu}, \cite{2017Zhang},
the problems of security and privacy have been widely investigated in terms of an {\em information flow} property called opacity \cite{Lafortune2018}. Opacity characterizes the ability of a system to hide a secret from an external observer, i.e., a curious eavesdropper (or simply eavesdropper in the sequel) who is assumed to have full knowledge of the system's structure and partial observation capability.
In a typical opacity setting, secrets are represented as subsets of states or strings (sequences of events), leading respectively to stated-based and language-based opacity.
Specifically, Saboori and Hadjicostis introduced in \cite{2007Chris} a state-based verification approach for the notion of current-state opacity in a finite automaton.
This approach aims to deduce if the eavesdropper can infer whether the current state of the system is a secret state or not, based on its observations. Later on, the notion of current-state opacity was extended to initial-state opacity \cite{2008Chris}, initial-and-final-state opacity \cite{2013Wu}, $K$-step opacity \cite{2011Chris}, and infinite-step opacity \cite{2012Chris}. Lin proposed a notion of language-based opacity in \cite{2011Lin}.
Regardless of the definition of opacity that is proposed, an opacity property holds if, for any string that reveals a secret,
there exists at least another string that does not reveal the secret and appears identical to the eavesdropper (i.e., it generates the same sequence of observations); this implies that the eavesdropper's estimate never reveals the system's secrets.
Opacity enforcement becomes necessary when a system is not opaque, and has been extensively studied mainly following two approaches: supervisory control strategies that restrict the system's behavior in order to avoid opacity violations \cite{2012Chrisa}, or enforcement via output obfuscation \cite{2014Wu}, \cite{2016Wu}, \cite{2018Wu}, \cite{2018Ji}, \cite{2019Ji}.
More specifically, Wu and Lafortune \cite{2014Wu,2016Wu} first proposed a novel opacity enforcement mechanism based on the use of insertion functions, where the insertion function takes as input the sequence of observable events of the system and outputs a modified sequence to the eavesdropper via the insertion of additional strings of observable events before each observable event.
Subsequently, they proposed edit functions \cite{2018Wu}, which are similar to insertion functions, but more powerful as they can alter the output behavior of the system by inserting, deleting, or replacing events.
Ji $\emph{et al.}$ \cite{2018Ji}, \cite{2019Ji} extended insertion and edit functions, respectively, into a more general setting where the insertion and edit functions may become known to the intruder. The developments for both of the above works are based on a game theoretical approach that has also been used in \cite{2015Yin,2019Yin}.
The notion of opacity is inversely related to that of event diagnosis, which tries to determine if certain (not directly observable) events have occurred or not \cite{2010Lafortunebook}, \cite{2020Chrisbook}.
This decision (event diagnosis) is based on observable events (occurring before and after the event in question) and on knowledge of the system model.
More specifically, researchers have modeled faults of a system as unobservable events and considered how to determine whether one of these events has occurred \cite{Lafortune2018} (more generally, how to classify or identify them).
For example, Sampath $et$ $al.$ \cite{1995Sampath} proposed a diagnostic model called a diagnoser that can be used to track the behavior of the underlying system modeled as a deterministic finite automaton under partial observation; the diagnoser can also be used to verify diagnosability, i.e., the ability of the system to deduce all occurrences of faults after a bounded number of observable events.
Note that the complexity of constructing the diagnoser is exponential with respect to the number of states of a system.
Another construction called a verifier was proposed in \cite{2001Kumar}, \cite{2002Lafortune}, which can be used to verify diagnosability with polynomial complexity.
Inspired by opacity enforcement and event diagnosis, we focus in this paper on the following two problems: (1) If certain events of a system are deemed as secrets, under what circumstances do their occurrences get revealed to an external eavesdropper? (2) Is it possible to conceal (as long as needed) the secret events using an obfuscation mechanism?
To answer these questions, this paper first formulates the problem of event concealment in a DES modeled as a non-deterministic finite automaton under partial observation.
We consider a scenario where particular unobservable events of a system are assumed to be secret, and their inference (diagnosis) by the external eavesdropper needs to be prevented.
To this end, the secret events are classified into two categories: {\em concealable} and {\em unconcealable}.
A secret event is said to be concealable if for any sequence containing the secret event, there exists at least another sequence that does not contain the secret event and produces the same observation. In other words, the eavesdropper cannot infer the occurrence of the secret event with finite observations.
It is also possible that a secret event is unconcealable, which means that the eavesdropper is able to eventually detect the occurrence of the secret event following some observation sequence.
If a given system is unconcealable (i.e., one or more secret events are unconcealable), the problem of concealability enforcement naturally arises.
We consider the case where a defensive function can be placed at the output of the system to obfuscate the observations to the eavesdropper by appropriately modifying the original events generated by the system (via replacement, insertion, or deletion).
In order to characterize the defensive function's ability to conceal the occurrence of secret events, the notion of enforceablility is proposed.
The defensive function is said to be $C$-enforceable if there exists at least one safe sequence that does not reveal the occurrence of the secret events, implying that the defender has a chance to manipulate observations so as to make the eavesdropper observe the safe sequence regardless of what event sequence has been generated by the system.
In this way, our goal of concealing the secret events can be achieved.
Hence, the concealability of the system is said to be $C$-enforceable if there exists an $C$-enforceable defender.
In addition, two scenarios about the use of the defensive functions are considered: one with constraints (in terms of how events are inserted/substituted) and the other one without constraints.
Note that the obfuscation technique (defensive functions) that we use to enforce concealability is related to prior works on state-based opacity \cite{2007Chris} and opacity enforcement based on insertion and edit functions.
More specifically, by reducing event concealabity to a current-state opacity problem (by introducing additional absorbing secret states after the system executes a secret event) and using the obfuscation techniques in \cite{2014Wu}, \cite{2016Wu}, \cite{2018Wu}, \cite{2018Ji}, one can verify whether event concealability is $C$-enforceable or nor via a necessary and sufficient condition that can be obtained with exponential complexity.
Unlike these approaches, our dual aim is to obtain, with polynomial complexity, one necessary condition and one sufficient condition respectively for enforcing event concealability in the system.
If the necessary condition is not satisfied, then we reach the same conclusion as by using the above more complex approaches (i.e., that concealability is not $C$-enforceable); alternatively, we are certain that event concealability is $C$-enforceable if the sufficient condition is satisfied.
Overall, the main contributions of this paper are summarized as follows:
\begin{enumerate}
\item The privacy of a system is considered in terms of concealing secret events and the concept of event concealment is proposed.
A formal description of secret events and concealability is provided, and a diagnoser-based approach is developed to verify concealability of the system.
The relationship between event concealment and event diagnosis is also characterized precisely.
\item The problem of concealability enforcement is formulated if a system is unconcealable.
We briefly discuss how enforceablility of defensive functions can be verified with exponential complexity using the approaches inspired by state estimation and diagnosis.
Moreover, we propose a polynomial complexity construction that can be used to obtain one necessary and one sufficient condition for event concealability, depending on whether the defender operates under constraints or not.
\end{enumerate}
\section{Preliminaries}
\subsection{System model}
Let $E$ be a finite set of events and $E^*$ be the set of all finite length strings (finite sequences of events) over $E$ including the empty string~$\varepsilon$. For a string $\lambda\in E^*$, we denote its length by $|\lambda|$ (the length of the empty string is defined as $|\varepsilon|=0$). A language $L\subseteq E^*$ is a subset of finite-length strings.
A prefix of a string $\lambda\in E^*$ is a string $u\in E^*$ for which there exists a string $v\in E^*$ such that $uv=\lambda$, where $uv$ denotes the concatenation of strings $u$ and $v$.
The prefix closure of $\lambda$ is the set of all of its prefixes denoted by $\overline{\lambda}=\lbrace u\in E^*|\exists v\in E^*, uv=\lambda \rbrace$.
We denote by $L\backslash\lambda$ the postlanguage of $L$ after $\lambda$, i.e., $L\backslash \lambda=\lbrace u\in E^*| \lambda u\in L\rbrace$.
Given $\alpha\in E$ and $\lambda\in E^*$, we say that $\lambda$ contains $\alpha$, written as $\alpha\in\lambda$, if there exist $u,v\in E^*$ such that $\lambda=u\alpha v$.
In this paper, we model a DES as a non-deterministic finite automaton (NFA) $G =(X, E, f, X_0)$, where $X$ is the finite set of states, $E$ is the finite set of events, $f : X\times E \rightarrow 2^X$ is the state transition function, and $X_o\subseteq X$ is the set of possible initial states.
The transition function $f$ can be extended to $X\times E^*\rightarrow 2^X$ in the usual manner: $f(x,\alpha \lambda)=\bigcup_{x'\in f(x,\alpha)}f(x',\lambda)$ for $x\in X$, $\alpha\in E$, and $\lambda\in E^*$, with $f(x,\varepsilon)=\lbrace x\rbrace$.
The generated language of NFA $G$, denoted by $L(G)$, is defined as $L(G) =\lbrace \lambda\in E^*\mid\exists x\in X_0, f(x,\lambda)\ne\emptyset\rbrace$.
The language $L(G)$ is said to be $\emph{live}$ if for all $\lambda\in L(G)$, there exists at least one event $\alpha\in E$ such that $\lambda \alpha\in L(G)$.
Note that without loss of generality, we assume that the set $X_0$ is a singleton.
A deterministic finite automaton (DFA) is a special case of an NFA, denoted by $G_d =(X, E, f, x_0)$, where
$X$ and $E$ have the same interpretations as in an NFA, $x_0\in X$ is the initial state, and the state transition function is captured by a possibly partially defined deterministic function $f : X\times E \rightarrow X$. Transition function $f$ can be extended to $X\times E^*\rightarrow X$ in the usual manner: $f(x,\alpha \lambda)=f(f(x,\alpha),\lambda)$ for $x\in X$, $\alpha\in E$, and $\lambda\in E^*$, with $f(x,\varepsilon)=x$ (note that $f(x,\alpha \lambda)$ is undefined if $f(x,\alpha)$ is undefined).
The generated language of DFA $G_d$, denoted by $L(G_d)$, is defined as $L(G_d) =\lbrace \lambda\in E^*\mid f(x_0,\lambda)!\rbrace$, where $f(x_0,\lambda)!$ means that $f(x_0,\lambda)$ is defined.
The event set $E$ in an NFA or DFA can be partitioned into a set of observable events $E_o$ and a set of unobservable events $E_{uo}=E- E_o$ due to partial observation of a system.
For a sequence $\lambda\in L(G)$, the natural projection (with respect to the set of observable events $E_o$) is defined as $P: E^*\rightarrow E_o^*$, where $P(\alpha)=\alpha$, if $\alpha\in E_o$; $P(\alpha)=\varepsilon$, if $\alpha\in E_{uo}\cup\lbrace\varepsilon\rbrace$; and $P(\lambda \alpha)=P(\lambda)P(\alpha)$, where $\alpha\in E$, and $\lambda\in E^*$. Essentially, the natural projection strips out the unobservable events from a given sequence of events (and presents the sequence of observable events).
The inverse projection of $P$ is defined as $P^{-1}: E_o^*\rightarrow E^*$ with $P^{-1}(\omega)=\lbrace \lambda\in E^*|P(\lambda)=\omega\in E_o^*\rbrace$.
For state $x\in X$, we denote by $E_o(x)=\lbrace \sigma\in E_o\mid f(x,\sigma)\neq\emptyset\rbrace$ the set of observable events defined at state $x$.
\subsection{Event Diagnosis}
Consider a system to be diagnosed that is modeled as an NFA $G =(X, E, f, X_0)$ with $E=E_o\dot{\cup}E_{uo}$.
Let $E_f\subseteq E_{uo}$ denote the set of fault events to be diagnosed.
For the sake of simplicity, we consider a single type of faults, i.e., $E_f=\lbrace f\rbrace$.
In this perspective, the notion of diagnosability (defined below) is the ability to detect every fault occurrence in the system, where a necessary and sufficient condition for diagnosability is provided in \cite{1995Sampath}.
\begin{definition}
Given a system $G$ and its generated language $L(G)$, $G$ is diagnosable w.r.t. a natural projection $P$ and a set of fault events $E_f$ if for all $f\in E_f$, and all $\lambda=\lambda_1f\in L(G)$, where $\lambda_1\in E^*$, there exists a finite number $n\in \mathbb{N}$ such that for all $\lambda_2\in L(G)\backslash \lambda$ with $|\lambda_2|\geq n$, and all $\lambda'\in P^{-1}(P(\lambda\lambda_2))\cap L(G)$, $f\in\lambda'$ holds.
\end{definition}
In plain words, the notion of diagnosability considers a fault to be diagnosable if it can be detected with certainty within a finite number of observable events after its occurrence.
Naturally, a system is diagnosable if it is able to detect any fault occurrence with a finite delay using the record of observed events.
\section{Event concealment Formulation}
Since online services and network communications are increasingly significant to our daily life,
the privacy of users may be exposed to potentially malicious eavesdroppers through the process of information transmission.
For this reason, secrets have been well defined as a subset of states or sequences of events in various opacity formulation in DESs \cite{2007Chris}, \cite{2008Chris}, \cite{2011Chris}, \cite{2012Chris}, \cite{2011Lin}.
In this work, we are interested in a scenario where a subset of events is deemed as the secret of a given system, and define the corresponding problem of event concealment for secret events of the system under partial observation.
More specifically, we consider a system modeled as an NFA $G = (X, E, f, X_0)$ with event set $E=E_o\dot{\cup} E_{uo}$, where $E_o$ and $E_{uo}$ are the sets of observable events and unobservable events, respectively.
We assume, without loss of generality, that the set of secret events $E_S$ is a subset of $E_{uo}$ (i.e., $E_S\subseteq E_{uo}$) since observable secret events will be trivially revealed.
For the sake of simplicity, we consider a single type of secrets, i.e., $E_S=\lbrace s\rbrace$.
We make the following assumptions (similar to \cite{1995Sampath}) for the NFA $G$:\\
(1) The language of $G$ is live, which means that there is at least one event defined at each state $x\in X$;\\
(2) There exists no cycle of unobservable events in $G$;\\
(3) Every secret event is unobservable ($E_S\subseteq E_{uo}$).
\begin{definition}
Given a system $G$ and its generated language $L(G)$, the occurrence of secret event $s\in E_S$ is said to be concealable in $L(G)$ w.r.t. a natural projection $P$ if for each $\lambda=\lambda_1s\lambda_2\in L(G)$, where $\lambda_1,\lambda_2\in E^*$, there exists $\lambda'\in P^{-1}(P(\lambda))\cap L(G)$ such that $s\notin \lambda'$ holds.
The system $G$ is concealable w.r.t. $P$ and $E_S$ if for all $s\in E_S$, $s$ is concealable.
\end{definition}
Suppose that a secret event $s\in E_S$ occurs after a sequence $\lambda_1$. Event $s$ is said to be concealable if after any indefinitely extendible continuation $\lambda_2\in E^*$, there exists a sequence $\lambda'$ that has the same observation as that of $\lambda=\lambda_1s\lambda_2$ but contains no secret event $s$; this implies that the secret event $s$ cannot be detected.
\begin{definition}
Given a system $G$ and its generated language $L(G)$, the occurrence of secret event $s\in E_S$ is said to be unconcealable in $L(G)$ w.r.t. a natural projection $P$ if there exists $\lambda_s=\lambda_1s\lambda_2\in L(G)$, where $\lambda_1,\lambda_2\in E^*$, and for all $\lambda'\in P^{-1}(P(\lambda_s))\cap L(G)$, $s\in \lambda'$ holds. Such a sequence $\lambda_s$ is called an $s$-revealing sequence.
The system $G$ is unconcealable w.r.t. $P$ and $E_S$ if there exists at least one unoncealable secret event $s\in E_S$.
\end{definition}
Note that, even if the secret event $s\in E_S$ is assumed to be unobservable, the secret can be indirectly detected following some observations after its occurrence, i.e., $P(\lambda_s)$ may lead to the conclusion that $s$ has occurred.
Apparently, a system is concealable if all secret events are concealable, whereas the system is unoncealable if some (at least one) secret events are unconcealable.
\subsection{Comparison between Event Diagnosis and Event Concealment}
From Definition 1, a fault is said to be diagnosable if it can be detected within a finite number of observable events after its occurrence. In order to detect the fault accurately, one should ensure that the fault can be detected for {\em any} execution after the occurrence of the fault.
When considering a secret event, once it gets revealed under {\em some} execution after its occurrence, it means that the privacy of the system has been compromised (since there is a possibility for the secrecy of the event to be compromised).
Thus, there is clearly an inverse relationship between event diagnosis and event concealment.
However, the two notions are not exactly the opposite of each other as clarified by the following proposition.
For the sake of comparison, we assume that the set of secret events is exactly the same as the set of fault events, i.e., the set of particular events $E_p=E_S=E_f$.
\begin{proposition}
Given a system $G=(X,E,f,X_0)$ and its generated language $L(G)$, $G$ is not diagnosable w.r.t. a natural projection $P$, and a set of particular events $E_{p}=E_S=E_f\subseteq E_{uo}$ if $G$ is concealable w.r.t. $P$ and $E_p$.
\end{proposition}
The proof of Proposition 1 is omitted since it follows directly from Definitions 1, 2, and~3.
Consider a system $G$ with its generated language $L(G)$.
From Definition 1, one can deduce that $G$ is not diagnosable if for some $\lambda_1f\in L(G)$, there exists a subsequent finite sequence $\lambda_2$ of arbitrary length such that $\lambda=\lambda_1f\lambda_2\in L(G)$ and we can find $\lambda'\in P^{-1}(P(\lambda))$ with $f\notin \lambda'$.
Hence, we can conclude that if $G$ is concealable, then $G$ is not diagnosable with respect to $P$ and $f$ by following Definition 2; however, the reverse is not necessarily true. The following example is provided to illustrate the differences between event diagnosis and event concealment.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.3\textwidth]{egmotivation.pdf}
\caption{System $G$.}
\label{egmotivation}
\end{figure}
\begin{eg}
Consider the system modeled as the NFA $G=(X,E,f,X_0)$ shown in Fig.~\ref{egmotivation}, where $X=\lbrace1,2,3,4\rbrace$, $X_0=\lbrace1\rbrace$, and $E=E_o\dot{\cup} E_{uo}$ with set of observable events $E_o=\lbrace a,b\rbrace$ and set of unobservable events $E_{uo}=\lbrace \sigma\rbrace$.
Event $\sigma$ is a special event (fault or secret event).
If the dotted cycle of event $a$ at state 4 is not present, event $\sigma$ is not diagnosable since the occurrence of event $\sigma$ is indistinguishable regardless of the executions of event $a$ or $b$.
In this case, event $\sigma$ is concealable according to Definition 2 while it is not diagnosable according to Definition 1.
On the other hand, if the dotted cycle of event $a$ at state 4 is considered, event $\sigma$ is still not diagnosable since the system may execute the sequence $ab$ repeatedly any number of times.
However, it is unconcealable since the occurrence of event $a$ at state 4 may lead to disclosure of event $\sigma$ (i.e., following the occurrence of $(ab)^n$ for some $n$, we subsequently observe event $a$).
\hfill$\diamond$
\end{eg}
\subsection{Problem formulation}
In the remainder of this paper, we consider a system modeled as an NFA $G = (X, E, f, X_0)$ with a set of initial states $X_0=\lbrace x_0\rbrace$ and a set of secret events $E_S=\lbrace s\rbrace\subseteq E_{uo}$, where $E=E_o\dot{\cup} E_{uo}$ (with $E_o$ being the set of observable events and $E_{uo}$ being the set of unobservable events).
We are interested in hiding from an external observer (a curious eavesdropper) confidential information of the system that is represented as the occurrence of events from a subset of $E_{uo}$, i.e., $E_S\subseteq E_{uo}$, which is called the secret set.
Accordingly, the privacy of the system is the concealment of the occurrence of the secret events.
Although the secret events are unobservable, they may be revealed following a finite number of observations after their occurrences.
The problem of event concealment is stated as follows, where the goal is to determine whether the secret events of the system are concealed.
\begin{problem}
Given an NFA $G=(X,E,f,X_0)$ with $E=E_o\dot{\cup} E_{uo}$ ($E_o$ being the sets of observable events and $E_{uo}$ being the set of unobservable events), determine whether or not $G$ is concealable w.r.t. a natural projection $P$ and a set of secret events $E_{S}\subseteq E_{uo}$.
\end{problem}
\section{Diagnoser and verifier constructions}
There are two ways to systematically verify diagnosability (i.e., the property that a fault will be diagnosed under any system behavior), namely the construction of a diagnoser as proposed in \cite{1995Sampath}, or the construction of a verifier as presented in \cite{2001Kumar}, \cite{2002Lafortune}.
Note that in order to detect/identify secret events, in this section we also present diagnoser and verifier constructions with some modifications.
We first define a secret label $l\in\lbrace N,S\rbrace$ to capture the diagnostic information at each state of the system, where label $S$ is associated with a state that can be reached via a sequence of events that contains the secret event $s\in E_S$, whereas label $N$ is associated with a state that can be reached via a sequence of events that does not contain the secret event.
\subsection{Diagnoser Construction}
The diagnoser $G_d$ of a given system $G=(X,E,f,X_0)$ is a DFA $G_d = (X_{d}, E_o, f_d, x_{d_0})$ determined from $G$ \cite{1995Sampath}, where $X_d\subseteq 2^{X\times \lbrace N, S\rbrace}$ is the set of states, $x_{d_0}=\lbrace (x,S)|\exists u=u'su''\in E_{uo}^*,s\in E_S, x\in f(x_0,u)\rbrace\cup\lbrace (x,N)|\exists v\in(E_{uo}-E_S)^*,x\in f(x_0,v)\rbrace$
is the initial state, and $f_d$ is the transition function constructed in a manner similar to the transition function of an observer \cite{1990ozeveren}, which includes attaching secret labels $l$ to the states and propagating these labels from state to state. Note that the resulting state space $X_d$ is a subset of $2^{X\times \lbrace N, S\rbrace}$ composed of the states of the diagnoser that are reachable from $x_{d_0}$ under $f_d$.
Given a system $G=(X,E,f,X_0)$ and its diagnoser $G_d=(X_d,E_o,f_d,x_{d_0})$, a state $x_d$ in $X_d$ is of the form $x_d=\lbrace (x_1,l_1),(x_2,l_2),...,(x_n,l_n)\rbrace$, where $x_i\in X$ and $l_i\in \lbrace N,S\rbrace$ for all $i\in\lbrace 1,2,\dots,n\rbrace$.
A state $x_d=\lbrace (x_1,l_1),(x_2,l_2),...,(x_n,l_n)\rbrace$ is called a normal state if $l_i = N$ for all $(x_i,l_i)\in x_d$; $x_d$ is called a secret state if $l_i = S$ for all $(x_i,l_i)\in x_d$; and $x_d$ is called an uncertain state if there exist $(x_i,l_i),(x_j,l_j)\in x_d$ such that $l_i = N$ and $l_j = S$, where $i, j\in\lbrace 1,2,\dots,n\rbrace$.
The state space $X_d$ can be accordingly partitioned into three disjoint sets $X_d=X_{N_d}\dot{\cup} X_{S_d}\dot{\cup} X_{NS_d}$, where $X_{N_d}$ is the set of normal states, $X_{S_d}$ is the set of secret states, and $X_{NS_d}$ is the set of uncertain states.
\begin{figure*}[htbp]
\centering
\subfigure[System $G$.]{
\centering
\includegraphics[width=0.4\textwidth]{Example1.pdf}
\label{G}
}
\subfigure[Diagnoser $G_d$.]{
\centering
\includegraphics[width=0.2\textwidth]{Gdfortheexample1.pdf}
\label{Gd}
}
\subfigure[$G_o$.]{
\centering
\includegraphics[width=0.2\textwidth]{Gofortheexample1.pdf}
\label{Go}
}
\subfigure[Verifier $G_v$.]{
\centering
\includegraphics[width=0.4\textwidth]{Gvfortheexample1.pdf}
\label{Gv}
}
\caption{Illustration of diagnoser and verifier constructions of a given system.}
\label{example1}
\end{figure*}
\subsection{Verifier Construction}
Given a system $G = (X, E, f, X_0)$, its verifier is proposed in \cite{2001Kumar} and its construction can be obtained via the following two steps:
\begin{enumerate}
\item First, we construct an NFA $G_o=(X_o,E_o,f_o,x_0^o)$ from $G$ with $L(G_o)=P(L(G))$, where $X_o=\lbrace (x_i,l_i)|x_i\in X, l_i\in \lbrace N,S\rbrace\rbrace$ is the finite set of states, $E_o$ is the set of observable events, $x_0^o=(x_0,N)$ is the initial state, and $f_o: X_o\times E_o\rightarrow 2^{X_o}$ is the transition function for $x\in X_o$, $e\in E_o$, defined as follows:
\begin{equation*}
f_o((x,l),e)=\left\{
\begin{aligned}
&\lbrace(x',N)|\exists t\in (E-E_S)^*, P(t)=e,\\
&x'\in f(x,t)\rbrace, \qquad\quad \text{if}\quad l=N, \\
&\lbrace(x',S)|\exists u=u'su''\in E^*, P(u)=e,\\
&x'\in f(x,u)\rbrace, \qquad \text{ if}\quad l=N, \\
&\lbrace(x',S)|\exists v\in E^*, P(v)=e,\\
&x'\in f(x,v)\rbrace, \qquad \text{ if}\quad l=S. \\
\end{aligned}
\right.
\end{equation*}
\item Then, a verifier, captured by $G_{v}$, can be obtained by composing $G_o$ with itself. It is defined as $G_{v}=(X_{v},E_o,f_{v}, x_{v_0})$, where $X_{v}=X_o\times X_o$ is the set of states, $E_o$ is the set of observable events, $f_{v}: X_v\times E_o\rightarrow 2^{X_v}$ is the transition function defined by $f_v((x_i,l_i,x_j,l_j),e)=(f_o((x_i,l_i),e),f_o((x_j,l_j),e))$ for all $e\in E_o$ if and only if both $f_o((x_i,l_i),e)$ and $f_o((x_j,l_j),e)$ are defined, and $x_{v_0}=x_0^o\times x_0^o=\lbrace(x_0,N),(x_0,N)\rbrace\in X_v$ is the initial state.
\end{enumerate}
Since the secret label $l$ is attached to each state in the verifier, we can partition $X_v$ into three disjoint sets in a similar manner as done for the diagnoser.
Each state $x_v\in X_v$ can be seen as a pair $\lbrace(x_i,l_i),(x_j,l_j)\rbrace$, where $x_i,x_j\in X$ and $l_i,l_j\in\lbrace N,S\rbrace$.
State $x_v=\lbrace \lbrace(x_i,l_i),(x_j,l_j)\rbrace$ is called a normal state if $l_i=l_j=N$; it is called a secret state if $l_i=l_j=S$; and it is called an uncertain state if $l_i\neq l_j$.
The state space $X_v$ can be partitioned into three disjoint sets $X_v=X_{N_v}\dot{\cup} X_{S_v}\dot{\cup} X_{NS_v}$, where $X_{N_v}$ is the set of normal states, $X_{S_v}$ is that of secret states, and $X_{NS_v}$ is that of uncertain states.
Since the diagnoser $G_d$ and $G_o$ are driven by the observable events of system $G$, we have $P(L(G))=L(G_d)=L(G_o)$.
Also, since the verifier $G_v$ is obtained from composing $G_o$ with itself, we have $L(G_v)=L(G_o)\cap L(G_o)=L(G_o)$.
Therefore, one can capture the fact that $P(L(G))=L(G_d)=L(G_v)$, which means that both $G_d$ and $G_v$ can track the observable sequences of the system $G$ to infer the specific unobservable events, i.e., the secret events.
\begin{eg}
Consider the system $G=(X,E,f,X_0)$ as shown in Fig. \ref{G}, where $X=\lbrace1,2,3,4,5,6,7\rbrace$, $X_0=\lbrace1\rbrace$, and $E=E_o\dot{\cup} E_{uo}$ with $E_o=\lbrace a,b,c,d\rbrace$ and $E_{uo}=\lbrace s\rbrace$.
Assume that the set of secret events $E_S$ equals $E_{uo}$, i.e., $E_S=E_{uo}=\lbrace s\rbrace$.
If we observe event $c$, we are not sure if $s$ has happened since there are two paths resulting in the observation of event $c$ but only the first one ($sc$) contains $s$.
If we observe sequence $cb$, then we know for sure that event $s$ has occurred since the string $cb$ enables the identification of the occurrence of event $s$.
Two diagnostic models are proposed next to capture diagnostic information in accordance to system activity.
The diagnoser $G_d$ (shown in Fig. \ref{Gd}) and $G_o$ (shown in Fig. \ref{Go}) are driven by the observable events in $G$. The verifier $G_v$ shown in Fig. \ref{Gv} is constructed by the composition $G_o$ with itself. For the sake of simplicity, each state $(x,l)$ is written as $xl$.
As mentioned above, if any observable event is generated by $G$, the diagnoser assembles all possible states that the system might reach whereas the verifier tracks pairs.
For instance, if event $c$ occurs from the initial state, $G_d$ will reach state $\lbrace3S,5N,6S\rbrace$, whereas $G_v$ will reach one of four possible states ($\lbrace 3S,3S\rbrace$, $\lbrace 5N,5N\rbrace$, $\lbrace 3S,5N\rbrace$, and $\lbrace 5N,3S\rbrace$).
Note that since states $\lbrace5N,3S\rbrace$ and $\lbrace3S,5N\rbrace$ are the same in terms of labels, we only reserve state $\lbrace3S,5N\rbrace$ in the remainder of the paper.
An external observer can detect the occurrence of the secret event $s$ via the diagnoser $G_d$ or the verifier $G_v$, i.e., as long as an observable sequence $da^*$ is captured.
A formal approach to verify concealability of the system is presented in the next section.
\hfill$\diamond$
\end{eg}
\begin{remark}(Computational complexity of a diagnoser and a verifier)
Given a system $G$, its diagnoser $G_d$ can be constructed via the method in \cite{1995Sampath} with exponential complexity of $O(2^{|X|\times 2^{|E_S|}}\times|E_o|)=O(2^{2|X|}\times|E_o|)$; and its verifier $G_v$ can be constructed via the method in \cite{2001Kumar} with polynomial complexity of $O((|X|\times 2^{|E_S|})^2\times |E_o|)=O(4|X|^2\times|E_o|)$.
\end{remark}
\section{Verification of Concealability}
This section presents a necessary and sufficient condition for concealability based on a diagnoser construction.
The idea is that a secret event is concealable if and only if for each arbitrarily long sequence in which the secret event occurs at least once, we can find at least one other arbitrarily long sequence in which the secret event does not appear and the two sequences have the same observable projection.
To proceed, several definitions are provided below.
\begin{definition}
Given a system $G=(X,E,f,X_0)$ and its diagnoser $G_d=(X_d,E_o,f_d,x_{d_0})$, a set of states $\lbrace x_{d_1},x_{d_2},\dots,x_{d_{n+1}}\rbrace\subseteq X_d$ is said to form a cycle $\mathcal{C}_d$ w.r.t. an observable sequence $\lambda=\lambda_1\lambda_2\dots\lambda_n\in E_o^*$ if $f_d(x_{d_i},\lambda_i)=x_{d_{i+1}}$ for $i\in\lbrace1,2,\dots,n\rbrace$, and $x_{d_{n+1}}=x_{d_1}$.
\end{definition}
\begin{definition}
A cycle $\mathcal{C}_d$ formed by a set of states $\lbrace x_{d_1},x_{d_2},\dots,x_{d_{n+1}}\rbrace\subseteq X_d$ is said to be a normal cycle, denoted by $\mathcal{C}_{N_d}$, if for all $x_{d_ i}\in\lbrace x_{d_1},x_{d_2},\dots,x_{d_{n+1}}\rbrace$, $x_{d_i}\in X_{N_d}$, where $i\in\lbrace1,2,\dots,{n}\rbrace$; it is said to be a secret cycle, denoted by $\mathcal{C}_{S_d}$, if for all $x_{d_ i}$, $x_{d_i}\in X_{S_d}$; and it is said to be an uncertain cycle, denoted by $\mathcal{C}_{NS_d}$, if for all $x_{d_i}$, $x_{d_i}\in X_{NS_d}$.
\end{definition}
\begin{lemma}
Given a system $G=(X,E,f,X_0)$ and its diagnoser $G_d=(X_d,E_o,f_d,x_{d_0})$, there exists a secret cycle $\mathcal{C}_{S_d}$ in $G_d$ if there exists one $s$-revealing sequence $\lambda_s=\lambda_1s\lambda_2\in L(G)$, where $s\in E_S$.
\end{lemma}
\begin{proof}
Assume that there exists an $s$-revealing sequence $\lambda_s$; this means that the occurrence of the secret event $s$ will be revealed anyhow.
Once the occurrence of the secret event is detected, both its diagnostic state and subsequent states are labeled as $S$.
Since the considered system is live, it means that eventually (after a finite number of events) the system will enter into a cycle.
Thus, it will inevitably reach a secret cycle $\mathcal{C}_{S_d}$.
\end{proof}
\begin{theorem}
Consider a system $G=(X,E,f,X_0)$ and its diagnoser $G_d=(X_d,E_o,f_d,x_{d_0})$. $G$ is concealable w.r.t. a natural projection $P$ and a set of secret events $E_S\subseteq E_{uo}$ if and only if there exists no secret cycle $\mathcal{C}_{S_d}$ in $G_d$.
\end{theorem}
\begin{proof}
Necessity:
We first prove that if $G$ is concealable, then there exists no secret cycle $\mathcal{C}_{S_d}$ in $G_d$.
By following Definition 2, $G$ is concealable if for all $s\in E_S$, and all $\lambda=\lambda_1s\lambda_2\in L(G)$, there exists $\lambda'\in P^{-1}(P(\lambda))\cap L(G)$ such that $s\notin \lambda'$ holds. This means that there exists no $s$-revealing sequence, which implies the absence of a secret cycle $\mathcal{C}_{S_d}$ in $G_d$ due to Lemma 1.
Sufficiency:
We now prove that if there exists no secret cycle $\mathcal{C}_{S_d}$ in $G_d$, then $G$ is concealable.
By contrapositive, assume that $G$ is unconcealable. Then there exists a corresponding (arbitrarily long) $s$-revealing sequence $\lambda_s=\lambda_1s\lambda_2$ with $s\in E_S$ according to Definition 3.
In conclusion, there exists a secret cycle $\mathcal{C}_{S_d}$ in $G_d$ by following Lemma~1.
\end{proof}
Fig. \ref{example1} provides an example of a system that is not concealable.
Specifically, one can notice that there are two secret cycles represented as $\mathcal{C}_{S_{d_1}}$ formed by $\lbrace 4S\rbrace$ and $\mathcal{C}_{S_{d_2}}$ formed by $\lbrace 7S\rbrace$ in $G_d$ as shown in Fig. \ref{Gd}.
Hence, system $G$ is unconcealable since an eavesdropper is able to detect the occurrence of the secret event if cycles $\mathcal{C}_{S_{d_1}}$ and $\mathcal{C}_{S_{d_2}}$ are reached.
\begin{remark}(A verifier cannot be used to verify concealability)
The reason why a verifier cannot have the same functionality as the diagnoser is because it tracks the behavior of system $G$ in a pairwise manner which does not allow us to be conclusive about concealability.
One can consider that the diagnoser $G_d$ tracks all possible diagnostic information when it observes the behavior of $G$, while the verifier $G_v$ tracks only two strings in $L(G)$ that look identical under projection $P$.
To verify concealability, one needs to check whether the occurrence of secret events has been detected in all diagnostic states of a diagnoser construction.
Intuitively, the presence of a secret cycle $C_{S_d}$ in $G_d$ leads to the conclusion that the occurrence of secret events has been detected; that is, the system $G$ is unconcealable.
However, the presence of a secret cycle in a verifier $G_v$ does not necessarily imply the violation of concealability.
For example, if the occurrence of observable sequence $\lambda\in P(L(G))$ can reach an uncertain cycle $C_{NS_d}$ in $G_d$, then there exist at least one secret cycle, one normal cycle, and one uncertain cycle in $G_v$ via the same execution $\lambda$.
Thus, a verifier cannot be used to verify the concealability of a system.
Nevertheless, as we argue later in the paper, the verifier can still be used to obtain necessary and/or sufficient conditions for $C$-Enforceability.
\end{remark}
\section{Manipulation of Observations under Defense}
Suppose that an eavesdropper aims to determine whether the secret events have occurred or not by tracking observable
sequences of a given system $G=(X,E,f,X_0)$.
From the previous section, concealability of a system is violated if, under some system behavior, the occurrence of any one secret event can be detected, i.e., there exists a secret cycle $\mathcal{C}_{S_d}$ in $G_d$.
To enforce concealability of a system, we allow an external defender at the interface of the system to manipulate actual observations generated by the system such that a different diagnostic state is perceived at the eavesdropper.
We assume that each output event of the system can be manipulated via the operations of replacement, insertion, or deletion.
Note that only one output event of the system can be manipulated at a time.
In other words, a given event is not allowed to be replaced and simultaneously has an event inserted before it.
More specifically, given any single observation $t\in E_o$, we assume that the defender has the capability to replace $t$ with any other observable event in the set of replaced events $E_R\subseteq E_o\cup \lbrace\varepsilon\rbrace$, or insert before it any single event in the set of inserted events $E_I\subseteq E_o$.
Such an event in $E_R$ or $E_I$ appears to the intruder to be identical to the events generated by the system.
We introduce the notation ``$/$'' used to represent defensive operations, so that defensive actions can be distinguished from events generated by a system.
Given an observable event $t\in E_o$ generated by the system, a defender is implemented by a defensive function $D: E_o\rightarrow 2^{E_o/E_o'}$, where $E_o/E_o'=\lbrace t/o\mid t\in E_o, o\in E_R\cup \lbrace et\mid e\in E_I\rbrace\rbrace$. The defensive function is defined as $D(t)=\lbrace t/o\mid t/o\in E_o/E_o' \rbrace$ and captures the set of possible defensive actions for the given $t$.
Note that $t/o$ indicates event $t$ is replaced with event $o$ (with $o=\varepsilon$ implying a deletion operation and $o=t$ implying no operation), and $o=et$ is used to indicate that event $e$ is inserted before event $t$.
For an observable sequence $u= t_1t_2\dots t_n\in E_o^*$, $D(t_i)$ can be treated differently according to each observable event $t_i$, depending on the time instant it is observed ($i$ is related to the position of the observable event in the observable sequence of length $|u|=n$, i.e., $i\in \lbrace1,2,\dots,|u|\rbrace$).
Moreover, we require that every defensive function $D(t_i)$ should be causal: when dealing with the observable sequence $u$ of the system, $D(t_i)$ should manipulate observations step by step, i.e., $D(t_i)$ defined at observation $t_i$ is determined by the current observation $t_i$, the previous observations $t_1t_2\dots t_{i-1}$ and the previous alterations $D(t_1)D(t_2)\dots D(t_{i-1})$.
In order to record the defensive actions easily for the observation sequence $u= t_1t_2\dots t_n\in E_o^*$,
a string-based defensive function that can capture the set of possible sequences of defensive actions is introduced, denoted by $D^*:E_o^*\rightarrow 2^{(E_o/E_o')^*}$ and defined as $D^*(u)=D^*(t_1t_2\dots t_{n-1}) D(t_n)$ for the observation sequence $u$.
Recall that notation $/$ is added into the defensive actions, which should be stripped out from $D^*$ such that a defensive observable sequence can be recognized and transmitted to the eavesdropper.
Therefore, we introduce a defensive projection $P_D: (E_o/E_o')^*\rightarrow E_o^*$, which is defined recursively as
\begin{equation*}
P_D(\sigma)=\left\{
\begin{aligned}
o, &\quad \sigma=t/o\\
et, &\quad \sigma=t/et\\
\end{aligned}
\right.
\end{equation*}
where $t\in E_o$, $o\in E_R$, $e\in E_I$, and $P_D(\omega\sigma)=P_D(\omega)P_D(\sigma)$ for $\omega\in (E_o/E_o')^*$ and $\sigma\in E_o/E_o'$.
\begin{remark}
The defensive function can be viewed as an interface at the system output that is used to confuse the eavesdropper.
In order to consider how the defensive function works under some constraints, the following two scenarios are presented in the developments of this paper:
\begin{enumerate}
\item $\forall t\in E_o$, $D(t)$ is defined with $E_R= E_o\cup\lbrace\varepsilon\rbrace$ and $E_I=E_oe$;
\item $\forall t\in E_o$, $D(t)$ is defined with $E_R= E_{or}\cup\lbrace\varepsilon\rbrace$ and/or $E_I= E_{oi}$, where $E_{or}\subset E_o$ and $E_{oi}\subset E_o$.
\end{enumerate}
\end{remark}
Scenario (1) is called the unconstrained case, i.e., each event $t$ in $E_o$ that is produced by the system can be replaced with any other event in $E_o$, or any event in $E_o$ can be inserted before the output of event $t$.
Scenario (2) is called the constrained case, i.e., each event $t$ in $E_o$ that is generated by the system can be replaced with some of the other events in $E_{or}$ but not necessarily all events, and/or only some events in $E_{oi}$ (but not all events) can be inserted before the output of event $t$.
One problem of interest is to identify whether the constructed defender has the capability under all possible system activity to conceal secret events before they are revealed, namely, enforce the concealment of secret events.
Such a problem can be addressed using the enforcement framework of event concealment proposed in the next section.
\section{Concealability Enforcement Formulation}
In this section, the notion of $C$-Enforceability is introduced, which characterizes the ability of a defender to manipulate the observations of output events such that the occurrence of secret events can be concealed to the eavesdropper regardless of system activity.
To deal with the enforcement of concealability, we assume that a considered system is unconcealable in the remainder of this paper.
\begin{definition}\label{enforceable}
Given a system $G$ and its generated language $L(G)$, a string-based defensive function $D^*$ is said to be $C$-enforceable w.r.t. a natural projection $P$ and a set of secret events $E_{S}$ if the following conditions hold:
\begin{enumerate}
\item for all $\lambda\in L(G)$, $D^*(P(\lambda))$ is defined;
\item for each $s$-revealing sequence $\lambda_s=\lambda_1s\lambda_2\in L(G)$, where $s\in E_S$, $\lambda_1, \lambda_2\in E^*$, there exists at least one sequence $u=t_1t_2\dots t_n\in L(G)$ with $t_i \notin E_S$ for $i\in\lbrace1,2,\dots,n\rbrace$, such that $P(u)\in P_D(D^*(P(\lambda_s)))$; furthermore, for each continuation $\lambda'\in E^*$ such that $\lambda_s\lambda'\in L(G)$, we can find $u'=t_1't_2'\dots t'_{n'}\in L(G)$ with $t_i'\notin E_S$, such that $P(u')\in P_D(D^*(P(\lambda_s\lambda')))$ and $P(u)\in \overline{P(u')}$.
\end{enumerate}
\end{definition}
Intuitively, concealability requires that for each sequence containing secret events (including an $s$-revealing sequence $\lambda_s$), there should exist a sequence that contains no secret event in a system, such that they can be observationally indistinguishable following defensive actions. In this case, a defender can manipulate actual observations and make the sequence without secret event appear feasible to the eavesdropper.
The first condition requires that,
as the interface of the system, the string-based defensive function should be able to react to every observable sequence of events that can be generated by the system, such that defensive actions (replacements, or insertions, or deletions) can be utilized.
In other words, in order to retain the privacy of a system, for each $s$-revealing sequence $\lambda_s$, the corresponding defensive actions should include a feasible one that does not disclose the occurrence of the secret event as described in the second condition of Definition 6.
Furthermore, for the sake of maintaining accuracy and completeness of defensive actions, the last part of the second condition guarantees that for each subsequent activity, defensive actions can retain the concealment of the occurrence of secret events indefinitely.
If both conditions are satisfied, one says that concealability of the system can be enforced by a $C$-enforceable string-based defensive function.
\begin{definition}
Given a system $G$ and its generated language $L(G)$, $G$ is said to be $C$-enforcing w.r.t. a natural projection $P$ and a set of secret events $E_{S}$ if there exists a $C$-enforceable string-based defensive function $D^*$.
\end{definition}
\begin{eg}
Recall that system $G$ as shown in Fig. \ref{G} is unconcealable by Theorem 1 since there exist two secret cycles $\mathcal{C}_{S_{d_1}}$ formed by $\lbrace 4S\rbrace$ and $\mathcal{C}_{S_{d_2}}$ formed by $\lbrace 7S\rbrace$ in $G_d$ as shown in Fig. \ref{Gd}.
Accordingly, one can capture two $s$-revealing sequences $\lambda_{s_1}=sda^*$ and $\lambda_{s_2}=cd^*sbd^*$ and a sequence containing no secret event $\lambda=cd^*$, where their natural projections are $P(\lambda_{s_1})=da^*$, $P(\lambda_{s_2})=cd^*bd^*$, and $P(\lambda)=cd^*$, respectively.
Since $cd^*\in P_D(D^*(da^*))$ (by replacing event $d$ with event $c$ and by replacing event $a$ with $d$ indefinitely) and $cd^*\in P_D(D^*(cd^*bd^*))$ (by replacing event $b$ with $d$ once), we conclude that the string-based defensive function $D^*$ is $C$-enforceable following Definition 6 and $G$ is $C$-enforcing following Definition 7.
\hfill$\diamond$
\end{eg}
Note here that it is important to be able to determine what action to take at each time an observable event is generated (in this case we can distinguish between the two strings, and take corresponding defensive actions, based on the very first observable event).
Consider a system modeled as an NFA $G=(X,E,f,X_0)$, where $E=E_o\dot{\cup} E_{uo}$ with $E_o$ and $E_{uo}$ being the sets of observable and unobservable events, respectively, and $E_S\subseteq E_{uo}$ being the set of secret events.
We consider the case that the system is unconcealable.
Then, under some system behavior, the eavesdropper can indirectly detect the occurrence of the secret events by utilizing its knowledge of the system model and by analyzing the observable behavior of the system.
The defender proposed in this section can alter observable output events of the system $G$ by replacement, insertion, or deletion.
The problem of enforcing concealability of the system aims to determine whether concealability is $C$-enforceable if it is violated, which equivalently identifies whether the defender can conceal the occurrence of the secret events before they are revealed.
It can be formalized as follows.
\begin{problem}
Given an NFA $G=(X,E,f,X_0)$ with $E=E_o\dot{\cup} E_{uo}$ ($E_o$ being the set of observable events and $E_{uo}$ being the set of unobservable events), determine whether or not $G$ is $C$-enforcing w.r.t. a natural projection $P$ and a set of secret events $E_{S}\subseteq E_{uo}$.
\end{problem}
\section{Verification of $C$-Enforceability}
Recall that two scenarios about defensive functions are considered: one involves constraints on how outputs are modified, and the other does not involve such constraints.
In this section, considering defensive functions without constraints, we first propose a necessary and sufficient condition for a string-based defensive function to be $C$-enforceable; this condition can be checked in polynomial time based on a verifier construction, an approach that has lower complexity compared to the one based on the diagnoser.
Then, a defensive verifier is constructed to capture all desirable defensive actions following system activity.
Finally, an $E$-verifier is built by a special synchronization mechanism between a verifier and a defensive verifier to verify the $C$-Enforceability of the defender in the scenario, where the defensive function has to operate under constraints.
Specifically, the E-verifier can be used to obtain, with polynomial complexity, one necessary and one sufficient condition for $C$-Enforceability; in case that the sufficient condition is satisfied, the E-verifier allows us to obtain a strategy to enforce concealability, also with polynomial complexity;
these will be contrasted against the approach using constructions with exponential complexity \cite{2014Wu} (the latter however, provides a necessary and sufficient condition).
\subsection{Necessary and Sufficient Condition for $C$-Enforceability of Defensive Functions without Constraints}
A malicious entity (eavesdropper) that aims to detect the occurrence of secret events can do so by observing and analyzing the behavior of a given system (via a diagnoser or a verifier).
Since a string-based defensive function can create a perturbed output sequence that causes a different interpretation from the eavesdropper’s point of view, we aim to find an observable sequence including no secret event in a system.
\begin{definition}
Given an NFA $G=(X,E,f,X_0)$ and its verifier $G_v=(X_v,E_o,f_v,x_{v_0})$, a set of states $\lbrace x_{v_1},x_{v_2},\dots,x_{v_{n+1}}\rbrace\subseteq X_v$ is said to form a cycle $\mathcal{C}_v$ w.r.t. an observable sequence $\lambda=\lambda_1\lambda_2\dots\lambda_n\in E_o^*$ if $x_{v_{i+1}}\in f_v(x_{v_i},\lambda_i)$ for $i\in\lbrace1,2,\dots,n\rbrace$ and $ x_{v_{n+1}}= x_{v_1}$.
\end{definition}
\begin{definition}
A set of states $\lbrace x_{v_1},x_{v_2},\dots,x_{v_{n+1}}\rbrace\subseteq X_v$ is said to be a normal cycle, denoted by $\mathcal{C}_{N_v}$, w.r.t. an observable sequence $\lambda=\lambda_1\lambda_2\dots\lambda_n\in E_o^*$ if for all $x_{v_i}\in \lbrace x_{v_1},x_{v_2},\dots,x_{v_{n+1}}\rbrace$, $x_{v_i}\in X_{N_v}$; it is said to be a secret cycle, denoted by $\mathcal{C}_{S_v}$, if for all $x_{v_i}$, $x_{v_i}\in X_{S_v}$; and it is said to be an uncertain cycle, denoted by $\mathcal{C}_{NS_v}$, if for all $x_{v_i}$, $x_{v_i}\in X_{NS_v}$.
\end{definition}
\begin{definition}
Consider an indefinitely extendible sequence $\lambda=\lambda_1\lambda_2\dots\lambda_k\dots\in L(G_v)$.
Sequence $\lambda$ is said to be safe (denoted as $\lambda_{safe}$) if there exists a set of states $\lbrace x_{v_0},x_{v_1},\dots,x_{v_k},\dots\rbrace$ such that $x_{v_i}\in f_v(x_{v_{i-1}},\lambda_i)$ with $x_{v_i}\in (X_{N_v}\cup X_{{NS}_v})$ for all $i\in\lbrace 1,2,\dots, k,\dots\rbrace$.
\end{definition}
Note that for large enough $k$, there exists $n\leq k$ such that $\lbrace x_{v_n},\dots,x_{v_k}\rbrace$ forms a normal cycle $\mathcal{C}_{N_v}$ or an uncertain cycle $\mathcal{C}_{NS_v}$, where $x_{v_n}=x_{v_k}$.
\begin{theorem}
Consider an NFA $G=(X,E,f,X_0)$ and its verifier $G_v=(X_v,E_o,f_v,x_{v_0})$.
Under the defensive functions without constraints, a string-based defensive function $D^*$ is $C$-enforceable if and only if there exists an indefinitely extendible safe sequence $\lambda_{safe}\in L(G_v)$.
\end{theorem}
\begin{proof}
Necessity: Obviously, if a string-based defensive function $D^*$ is $C$-enforceable, then there exists at least one indefinitely extendible safe sequence $\lambda_{safe}\in L(G_v)$ according to Definition 6.
Sufficiency: If there exists a safe sequence $\lambda_{safe}\in L(G_v)$, then it holds that for all $\lambda_{safe}'\in\overline{P(\lambda_{safe})}$, $\lambda_{safe}'$ is safe, i.e., we have $f_v(x_{v_0},\lambda_{safe}')\cap (X_N\cup X_{NS})\neq\emptyset$.
For each $s$-revealing sequence $\lambda_s=\lambda_1s\lambda_2\in L(G_v)$, there always exists a defensive observable sequence $\lambda\in P_D(D^*(P(\lambda_s)))$ such that $\lambda=\lambda_{safe}'$ with the length of $\lambda$ being equal to that of $\lambda_{safe}'$ (e.g., by replacing the first observable event in $P(\lambda_s)$ with the first observable event in $\lambda_{safe}$, the second observable event in $P(\lambda_s)$ with the second observable event in $\lambda_{safe}$, etc.).
This can be retained for any continuation $\lambda'\in E_o^*$ (i.e., $\lambda\lambda'\in L(G_v)$) by considering appropriate continuations of $\lambda_{safe}'$, such that the eavesdropper observes a prefix of the sequence $\lambda_{safe}$. Therefore, $D^*$ is $C$-enforceable.
\end{proof}
\begin{eg}
As shown in Fig. \ref{Gv}, there exists a safe sequence $\lambda_{safe}=cd^*$ in the verifier $G_v$. An uncertain cycle $C_{NS_v}$ formed by $\lbrace 3S,5N\rbrace$, a normal cycle $C_{N_v}$ formed by $\lbrace 5N,5N\rbrace$, or a secret cycle $C_{S_v}$ formed by $\lbrace 3S,3S\rbrace$ will be eventually reached from the initial state via the occurrence of $\lambda_{safe}$, which leads to the fact that the occurrence of the secret event can be concealed perpetually.
Once $\lambda_{safe}$ is identified, the defender could trick the eavesdropper into thinking that the system is executing the safe sequence $\lambda_{safe}$ regardless of the system activity.
Thus, one can infer that the string-based defensive function $D^*$ is $C$-enforceable following Theorem 2 and $G$ is $C$-enforcing.
\hfill$\diamond$
\end{eg}
\begin{remark}(A verifier can be used to verify $C$-Enforceability in the unconstrained case)
Recall that a string-based defensive function is said to be $C$-enforceable if, regardless of what event is generated by a given system, it can manipulate observations and output a sequence that does not reveal the occurrence of secret events.
Thus, in the unconstrained case, one can use a verifier construction to verify $C$-Enforceability with low complexity.
Note that existence of the safe sequence $\lambda_{safe}$ can be found via Breadth-First Search \cite{DFS} on the verifier with complexity $O(|X_v|\times|E_o|)$.
Hence, concealability enforcement can be verified with complexity $O(|X_v|\times|E_o|)=O(4{|X|}^2\times|E_o|)$.
\end{remark}
\subsection{Defensive Verifier}
The eavesdropper can utilize its knowledge of the system model to determine the authenticity of the received observation according to the transition structure of a system.
It is essential for a defender to generate a modified observation that matches a valid observation, i.e., one that can be generated by the system.
When the actions of the defender have some constraints, a defensive verifier that captures all admissible defensive actions is proposed.
Suppose that the defensive functions have some constraints, i.e., observable events generated by the system can only be replaced with some events in the set of replaced events $E_R$ (with $E_R\cap E_o\neq E_o$) or inserted events need to belong to the set of inserted events $E_I$ (with $E_I\cap E_o\neq E_o$).
More precisely, we can impose constraints differently on the defensive functions for each observable event generated by the system, as illustrated in the following example.
\begin{eg}
Consider the system $G$ as shown in Fig. \ref{G} when the defensive function is under constraints.
We use the verifier $G_v$ to capture observable behavior of the system $G$.
For the set of observable events $E_o=\lbrace a,b,c,d\rbrace$, we impose constraints on corresponding defensive actions as follows: $D(a)=\lbrace a/a,a/d\rbrace$, $D(b)=\lbrace b/b\rbrace$, $D(c)=\lbrace c/c\rbrace$, and $D(d)=\lbrace d/d,d/cd,d/\varepsilon\rbrace$.
Note that, apart from these constraints, the defender should guarantee that each defensive action is consistent with the system structure and previous defensive actions (so that the eavesdropper cannot realize that the system is under defense).
For instance, even though $d/ad\in D(d)$ holds, if we insert event $a$ before event $d$ at the initial state $\lbrace1N,1N\rbrace$ in $G_v$, then the eavesdropper will be able to infer that something has been altered since sequence $ad$ is not defined at the initial state.
\end{eg}
Here we denote the set of possible defensive actions of the system $G$ by $E_D$, which is defined as $E_D=\bigcup_{t\in E_o}D(t)$.
A defensive verifier ${G_D}$ is defined below to capture all admissible defensive actions, i.e., the set of defensive actions with given constraints $E_D$.
\begin{definition}\label{G_{ad}}
Given an NFA $G=(X,E,f,X_0)$ and its verifier $G_v=(X_v,E_o,f_v,x_{v_0})$, a defensive verifier is an NFA $G_D=(X_D,E_D,f_D,x_{D_0})$, where $X_D=X_v-X_{S_v}$ is the state space, $x_{D_0}=x_{v_0}$ is the initial state (assumed to be outside $X_{S_v}$), $E_D$ is the set of defensive actions with constraints, and $f_D$ is the transition function that implements defensive actions for $t/o\in E_D$, and $x_D\in X_D$:
\begin{equation*}
f_D(x_D,t/o)=\left\{
\begin{array}{rll}
&f_v(x_D,\varepsilon) \\
& \quad\text{if}\quad P_D(t/o)=\varepsilon \quad\text{(deletion)},\\
&f_v(x_D,t')\cap X_D \\
& \quad \text{if}\quad P_D(t/o)=t' \quad\text{(replacement)},\\
&(\bigcup_{x_D'\in f_v(x_D,t')}f_v(x_D',t))\cap X_D \\
& \quad \text{if}\quad P_D(t/o)=t't \quad\text{(insertion)}.\\
\end{array}
\right.
\end{equation*}
\end{definition}
Without loss of generality, we can consider $X_D$ to be the set of states in $G_D$ that are accessible from the initial state $X_{D_0}$.
Note that $f_D$ is defined such that $f_D(x,t/o)=f_v(x,P_D(t/o))$ if $f_v(x,P_D(t/o))\cap X_{S_v}=\emptyset$ (otherwise, the states in $X_{S_v}$ are excluded).
Informally, the defensive verifier $G_D$ can be constructed by cloning $G_v$ and pruning away all secret states and all original events while adding defensive events, and then taking the accessible part of the resulting automaton.
For each state $x_D\in X_D$, we denote by $E_D(x_D)=\lbrace t/o\in E_D| f_D(x_D,t/o)\neq\emptyset\rbrace$ the set of possible defensive actions at state $x_D$.
An example is provided below to demonstrate the defensive verifier.
\begin{figure}[htbp]
\centering
{
\includegraphics[width=0.3\textwidth]{NDGvfortheexample.pdf}
}
\caption{Defensive Verifier $G_D$.}
\label{defensiveverifier}
\end{figure}
\begin{eg}
Consider the system $G$ and its verifier $G_v$ as shown in Fig. \ref{example1} with $E_D$ as described in Example 5. Its defensive verifier $G_D$ is built from the verifier $G_v$ in Fig. \ref{Gv} via Definition \ref{G_{ad}}. The defensive verifier demonstrates all possible defensive actions (including replacements, insertions, and deletions) with constraints.
For instance, the defensive actions that can be taken at the initial state $(1N,1N)$ in $G_D$ are $E_D{\lbrace(1N,1N)\rbrace}=\lbrace c/c,d/\varepsilon,d/cd\rbrace$.
More specifically, defensive action $d/\varepsilon$ or $d/cd$ can be implemented if observable event $d$ is generated by the system, while defensive action $c/c$ can be implemented if observable event $c$ is generated by the system.
\end{eg}
\begin{remark}
The defensive verifer can be regarded as a defender that tracks the defensive actions in an effort to conceal the secret events and confuse the eavesdropper.
Since the set of states of $G_D$ is identical to (or a subset) of the set of states of $G_v$, we can also classify the states of $G_D$ in two disjoint sets: the set of normal states $X_{N_v}$, and the set of uncertain states $X_{NS_v}$ (recall that the set of secret states has been pruned away).
\end{remark}
\subsection{Verification of $C$-Enforceability under Defensive Functions with Constraints}
A $C$-enforceable string-based defensive function should ensure that all possible defensive actions keep the eavesdropper confused regardless of system activity.
Thus, when the defensive function is subject to constraints, we propose a new construction by composing a verifier and a defensive verifier of a given system to capture all admissible defensive actions following system activity.
To capture the observable behavior of the verifier, we use $E_o(x_v)=\lbrace \sigma\in E_o\mid f_v(x_v,\sigma)\neq\emptyset\rbrace$ to denote the set of observable events defined at state $x$ for $x\in X_v$.
\begin{definition}\label{RdefinitionKverifier}
Given an NFA $G=(X,E,f,X_0)$ with its verifier $G_v=(X_v,E_o,f_v,x_{v_0})$ and its defensive verifier $G_{D}=(X_D,E_D,f_D,x_{D_0})$, the $E$-verifier is an NFA that captures the composition of $G_v$ and $G_D$, denoted by $V_E=(X_E, E_D, f_E, x_{E_0})$, satisfying the following:
\begin{enumerate}
\item $X_E=\lbrace (x_v,x_D)\mid x_v\in X_v, x_D\in X_D\rbrace$ is the state space;
\item $x_{E_0}= (x_{v_0},x_{D_0})$ is the initial state;
\item $E_D$ is the set of defensive actions;
\item $f_E$ is the transition function defined as follows:
$$f_E((x_v,x_D),t/o)=f_v(x_v,t)\times f_D(x_D,t/o)$$ $\noindent$ where $t\in E_o(x_v)$, $t/o\in D(t)$.
\end{enumerate}
\end{definition}
Note that given a state $(x_v,x_D)\in X_E$ and an event $t\in E_o(x_v)$, if we have only one defensive action, e.g., $D(t)=\lbrace t/o\rbrace$, it may evolve in multiple ways via $f_v(x,t)$ and/or $f_D(x_D,t/o)$ (i.e., the $E$-verifier is non-deterministic).
\begin{definition}
Consider an $E$-verifier $V_E=($$X_E$, $E_D,$ $f_E,$ $x_{E_0})$. Defensive action $t/o\in E_D$ is said to be feasible if and only if $f_E((x_v,x_D),t/o)\ne\emptyset$.
\end{definition}
\begin{lemma}
Given an NFA $G=(X,E,f,X_0)$ with its verifier $G_v=(X_v,E_o,f_v,x_{v_0})$ and its defensive verifier $G_{D}=(X_D,E_D,f_D,x_{D_0})$, the $E$-verifier $V_E=(X_E, E_D, f_E, x_{E_0})$ enumerates feasible defensive actions following system activity.
\end{lemma}
The proof of the above lemma follows directly from Definition $\ref{RdefinitionKverifier}$ and is thus omitted.
The main idea underlying Definition $\ref{RdefinitionKverifier}$ is to compute a parallel composition of $G_v$ and $G_D$, where the event synchronization is performed with respect to $E_o$ from $G_v$ and $E_D$ from $G_D$.
\begin{figure*}[htbp]
\centering
{
\includegraphics[width=0.7\textwidth]{VEfortheexample1.pdf}
}
\caption{$E$-verifier $V_E$.}
\label{Everifier}
\end{figure*}
\begin{eg}
Consider the NFA $G$ in Fig. \ref{G} with verifier $G_v$ and defensive verifier ${G_D}$ as shown in Fig. \ref{defensiveverifier}.
The $E$-verifier of $G$ is constructed following Definition \ref{RdefinitionKverifier} as depicted in Fig. \ref{Everifier} (for convenience, states have been renamed as $A,B,C,D,E$ and $F$ as shown in the figure).
Since the $E$-verifier is non-deterministic, it may reach different states under the same defensive action for a given event, e.g., event $c$ is inserted before event $d$ at state $(A,A)$ such that we can reach states $(B,D)$ and $(B,C)$ via the defensive action $d/cd$.
Note that we draw dotted arrows corresponding to infeasible defensive actions in response to event $a$ and $b$, denoted by $a/$ and $b/$, on states $(B,A)$, $(C,C)$, and $(C,D)$.
This is done in order to visualize infeasible defensive actions in response to particular events:
if we look at states $\lbrace 4S,4S\rbrace$ (represented as $B$ in Fig. \ref{Everifier}) and $\lbrace 5N,5N\rbrace$ (represented as $C$ in Fig. \ref{Everifier}) in the verifier $G_v$, event $a$ should be defined at state $B$ and event $b$ should be defined at state $C$; however, there is no feasible defensive action that can be taken at states $\lbrace1N,1N\rbrace$, $\lbrace5N,5N\rbrace$, and $\lbrace3S,5N\rbrace$ (represented as $A$, $C$, and $D$ respectively in Fig. \ref{Everifier}) in the defensive verifier $G_D$.
Thus, the dotted arrow labeled $a/$ at state $(B,A)$ is used to indicate that there is no corresponding defensive action for event $a$, and the dotted arrow labeled $b/$ at states $(C,C)$ and $(C,D)$ is used to indicate that there is no corresponding defensive action for event $b$.
Next, we consider the problem of verifying $C$-enforceability of a string-based defensive function in terms of the construction of $V_E$, where a necessary condition is provided below.
\hfill$\diamond$
\end{eg}
\begin{definition}
Consider an NFA $G=(X,E,f,X_0)$ and its $E$-verifier $V_E=(X_E, E_D, f_E, x_{E_0})$. Let $X_{C_{x_v}}\subseteq X_E$ denote the set of co-relative states for state $x_v\in X_v$, defined as $X_{C_{x_v}}=\lbrace (x_v,x_{D_i})\in X_E| x_{D_i}\in X_D \rbrace$. The set of states $X_{C_{x_v}}$ is said to be legal if for all $t\in E_o(x_v)$, there exists $(x_v,x_{D_i})\in X_{C_{x_v}}$ and defensive action $t/o\in E_D$ such that $f_E((x_v,x_{D_i}), t/o)$ is defined; $X_{C_{x_v}}$ is said to be illegal otherwise.
\end{definition}
Recall that for state $(x_v,x_{D})$ in the $E$-verifier, the first component represents a pair of possible actual states that the system can be in at current time, whereas the second one represents a pair of possible fake states perceived by the eavesdropper after defensive actions.
Since the $E$-verifier is nondeterministic, multiple state of the form $(x_v,x_D)$ might be allowable at current time following a sequence of observations.
Thereby, the set of states $X_{C_{x_v}}$ for state $x_v$ can be regarded as a set that contains all possible fake states (perhaps not all of them simultaneously reachable following a particular sequence of observations) if the verifier is at state $x_v$.
Therefore, for the defensive function to be able to react to system activity, we need that for each $t\in E_o(x_v)$, we can find among all states in $X_{C_{x_v}}$ at least one defensive action $t/o$ that is feasible.
If we cannot find any defensive action, the correlative state set $X_{C_{x_v}}$ is problematic since it means that at state $x_v$ in the verifier (which is certainly reachable under some sequence of observations) we will have no way of maintaining secrecy following observable events generated by the system.
\begin{theorem}
Consider an NFA $G=(X,E,f,X_0)$ with its verifier $G_v=(X_v,E_o,f_v,x_{v_0})$, its defensive verifier $G_{D}=(X_D,E_D,f_D,x_{D_0})$, and its $E$-verifier $V_E=(X_{E}, E_D, f_E, x_{E_0})$.
If a string-based defensive function $D^*$ is $C$-enforceable, then for all $x_v\in X_v$, the corresponding set of co-relative states $X_{C_{x_v}}$ is legal.
\end{theorem}
\begin{proof}
By contradiction, if for some $x_v\in X_v$, the set of co-relative states $X_{C_{x_v}}$ is illegal, then it means that
for a system action $t\in E_o$ that is defined at state $x_v$, there exists state $(x_v,x_{D_i})\in X_{C_{x_v}}$ such that $f_E((x_v,x_{D_i}), t/o)=\emptyset$.
In this case, there is no defensive action that can be taken corresponding to the system action $t$, which violates the first condition of Definition 6.
In other words, for any sequence that takes us to $x_v$ in the verifier, there is at least one continuation of this sequence that cannot be protected, which violates the second condition of Definition 6.
Therefore, if a string-based defensive function $D^*$ is $C$-enforceable, then for all $x_v\in X_v$, its set of co-relative states $X_{C_{x_v}}$ is legal.
\end{proof}
\begin{algorithm}[htbp]
\caption{Necessary condition for $C$-~Enforceability of a string-based defensive function}
\label{AlgorithmE}
\LinesNumbered
\KwIn{$E$-verifier $V_E=(X_{E}, E_D, f_E, x_{E_0})$, where $X_{E}=\lbrace (x_v,x_D)\mid x_v\in X_v, x_D\in X_D\rbrace$.}
\For{$x_v\in X_v$}{
Let $X_{C_{x_v}}$ be the set of co-relative states for state $x_v$, i.e., $X_{C_{x_v}}=\lbrace(x_v,x_{D_i})|(x_v,x_{D_i})\in X_E\rbrace$\;
\For{$t\in E_o(x_v)$}{
\If{$\nexists (x_v,x_{D_i})\in X_{C_{x_v}}, t/o\in D(t)$, such that $f_E((x_v,x_{D_i}),t/o)\ne\emptyset$}{
\Return Not $C$-enforceable.
}
}
}
\Return $C$-enforceable.
\end{algorithm}
Given an $E$-verifier, we can check the necessary condition for the string-based defensive function to be $C$-enforceable by following Algorithm~1.
For each state $x_v\in X_v$, the set of co-relative states $X_{C_{x_v}}$ is computed in Line 2.
Then, in Lines 3--5, we check whether there exist no states in $X_{C_{x_v}}$ that can respond to event $t\in E_o(x_v)$. If there exists at least one state, then the set of co-relative states $X_{C_{x_v}}$ is legal and has been examined; otherwise, we can conclude that the string-based defensive function is not $C$-enforceable.
In the end, the necessary condition is satisfied if co-relative states for each state in the verifier are legal.
Note that Theorem 3 provides a necessary condition for $C$-Enforceability.
Even though the condition might be satisfied, it is possible that a string-based defensive function may not be $C$-enforceable.
Next, a notion of $E$-forbidden state is defined below and such states are removed from the $E$-verifier.
Based on the reduced version of $E$-verifier, a sufficient condition for $C$-enforceability is obtained.
\begin{definition}
Given a system $G=(X,E,f,X_0)$ and its $E$-verifier $V_E=(X_E, E_D, f_E, x_{E_0})$, a state $(x_v,x_D)\in X_E$ is said to be $E$-forbidden if for some $t\in E_o(x_v)$, we have $f_E((x_v,x_D),t/o)=\emptyset$ for all $t/o\in D(t)$.
\end{definition}
A state $(x_v,x_D)\in X_E$ is said to be $E$-forbidden if there is no subsequent state that can be reached for some $t$ generated by the system, i.e., for some $t\in E_o(x_v)$, there is no defensive action that can be taken in response to system execution $t$.
Note that when the same defensive action evolves from some originating $E$-verifier state to multiple destination states (on account of the nondeterminacy of the $E$-verifier), if one of the multiple destination states is $E$-forbidden, then only that state needs to be avoided.
Assume that for a given state $(x_v,x_D)\in X_E$ and an event $t\in E_o(x_v)$, there is a defensive action $t/o\in D(t)$ such that $f_E((x_v,x_D),t/o)=\lbrace (x_{v_1},x_{D_1}), (x_{v_2},x_{D_2})\rbrace$, where state $(x_{v_1},x_{D_1})$ is $E$-forbidden.
In this case, the removal of state $(x_{v_1},x_{D_1})$ will not make state $(x_v,x_D)$ $E$-forbidden: since state $(x_{v_2},x_{D_2})$ is retained, it can be reached from state $(x_v,x_D)$ via the defensive action $t/o$.
In order to ensure that the $E$-verifier captures only feasible defensive actions following system activity, a formal procedure for constructing the reduced $E$-verifier by removing $E$-forbidden states, denoted by $RV_E$, is presented in Algorithm 2 below.
A sufficient condition for $C$-enoforceability is obtained subsequently.
\begin{algorithm}[htbp]
\caption{Construction of reduced $E$-verifier}
\label{AlgorithmVE}
\LinesNumbered
\KwIn{$E$-verifier $V_E=(X_{E}, E_D, f_E, x_{E_0})$, where $X_{E}=\lbrace (x_v,x_D)\mid x_v\in X_v, x_D\in X_D\rbrace$.}
\KwOut{Reduced $E$-verifier $RV_E=(X_{ER}, E_D, f_E, x_{E_0})$.}
$X_{EF}=\emptyset$, where $X_{EF}$ denotes the set of $E$-forbidden states\;
\For{$(x_v,x_D)\in X_E-X_{EF}$}{
\For{$t\in E_o(x_v)$}{
\If{$\forall t/o\in D(t)$, $f_E((x_v,x_D),t/o)\cap (X_E-X_{EF})=\emptyset$}{
$X_{EF}:=X_{EF}\cup \lbrace(x_v,x_D)\rbrace$\;
}
}
}
Go back to Step 2 until no more new $E$-forbidden states are produced\;
$X_{ER}=X_E-X_{EF}$\;
\Return $RV_E$.
\end{algorithm}
Next we briefly explain how Algorithm 2 works. Given an $E$-verifier, the states are marked as $E$-forbidden if they have no subsequent states for some event $t$ that is defined in the verifier, as shown in Lines 3--5. Nevertheless, once certain states are pruned away, new forbidden states may be created. Then, the reduced $E$-verifier $RV_E$ is constructed by pruning away all $E$-forbidden states in an iterative manner until no more such states are generated.
\begin{theorem}
Given an NFA $G=(X,E,f,X_0)$ with its verifier $G_v=(X_v,E_o,f_v,x_{v_0})$, its defensive verifier $G_{D}=(X_D,E_D,f_D,x_{D_0})$, and its reduced $E$-verifier $RV_E=(X_{ER}, E_D, f_E, x_{E_0})$, a string-based defensive function $D^*$ is $C$-enforceable if for all $x_v\in X_v$, there exists $(x_v,x_D)\in X_{ER}$, where $X_{ER}$ is the state space of the reduced $E$-verifier.
\end{theorem}
\begin{proof}
State $(x_v,x_D)\in X_{ER}$ contains two components, where the first one satisfies $x_v\in X_v$ and the second one satisfies $x_D\in X_D$ (i.e., $x_v$ and $x_D$ are states from the state spaces of the verifier $G_v$ and the defensive verifier $G_D$, respectively.)
If for all $x_v\in X_v$, there exists $(x_v,x_D)\in X_{ER}$. This means that each observable event generated by the verifier has at least one defensive action such that the first condition for Definition~\ref{enforceable} holds.
Moreover, each subsequently visited state in $x_{ER}$ has the same property (otherwise, it would have been pruned away).
For each state $(x_v,x_D)\in X_{ER}$, the case that $x_v\in X_{S_v}$ while $x_D\in X_{S_v}$ is not possible since secret states are pruned away in $G_D$ (in fact $x_D\notin X_{S_v}$ regardless of $x_v$).
Thus, the second condition for Definition \ref{enforceable} holds.
Moreover, for long enough sequences a cycle is reached regardless of the execution in $RV_E$ by virtue of Algorithm 2 (there remains no $E$-forbidden state at the end of Algorithm 2).
This implies that the diagnostic information in $x_v$ can be concealed indefinitely, which satisfies the second part of the second condition of Definition \ref{enforceable}.
Therefore, one can conclude that the string-based defensive function $D^*$ is $C$-enforceable.
\end{proof}
\begin{remark}
Note that Theorem 4 provides a sufficient condition for $C$-Enforceability.
If the condition holds, the reduced $E$-verifier can be used to take defensive action in response to system activity (any action that is feasible at the present state —or one of the present states— of the $E$-verifier can be taken since there will always be a feasible action in all of the future states).
\end{remark}
\begin{eg}
Consider the $E$-verifier $V_E$ shown in Fig. \ref{Everifier}.
First, we check whether the necessary condition in Theorem 3 is violated via Algorithm 1.
For instance, the set of correlative states for state $B$ denoted by $X_{C_B}=\lbrace (B,A), (B,D), (B,C)\rbrace$. By $E_o(B)=\lbrace a\rbrace$, we only need to check if a defensive action for $a$ is possible among $(B,A), (B,D)$, and $(B,C)$.
Since there are subsequent defensive actions from $(B,D)$ and $(B,C)$, the set of correlative states for state $B$ is legal and retained.
Similarly, we check the set of correlative states for state $C$ denoted by $X_{C_C}=\lbrace (C,C), (C,D)\rbrace$.
Note that $E_o(C)=\lbrace b,d\rbrace$, but there exists no defensive actions for $b$.
Thus, the set of correlative states for state $C$ is illegal and Algorithm 1 stops.
At this point, we know that the string-based defensive function is not $C$-enforceable.
One can also check whether the sufficient condition in Theorem 4 is violated via the reduced $E$-verifier $RV_E$ constructed by following Algorithm 2 (of course, in this case we already know that the sufficient condition will be violated since the necessary condition is not satisfied; nevertheless, we provide the discussion below for illustration purposes).
Since state $(B,A)$ has no continuation (there is no defensive action for event $a$), it is marked as $E$-forbidden.
For states $(C,C)$ and $(C,D)$, we have that the set of observable events is $E_o(C)=\lbrace b,d\rbrace$.
They are marked as $E$-forbidden since there is no defensive action for event $b$.
For state $(A,A)$, there is an observable event $d\in E_o(A)$ with two defensive actions $D(d)=\lbrace d/\varepsilon, d/cd\rbrace$.
Even though $f_E((A,A),d/\varepsilon)$ leads to an $E$-forbidden state $(B,A)$, there exists a feasible defensive action $d/cd$; therefore, state $(A,A)$ is not marked as $E$-forbidden.
The reduced $E$-verifier $RV_E$ can be directly obtained by removing states $(B,A)$, $(C,C)$, and $(C,D)$.
Thus, one concludes that the string-based defensive function is not $C$-enforceable since the sufficient condition in Theorem 4 is violated (there is no state composed with $C$ and $F$ in $RV_E$), which confirms that concealability of the system cannot be enforced by the string-based defensive functions under the given constraints.
The important aspect of the above discussions is that both conditions (sufficient and necessary) can be checked with polynomial complexity.
\hfill$\diamond$
\end{eg}
\begin{remark}
Consider a system $G=(X,E,f,X_0)$, where the number of states is $|X|$ and the number of secret events is $|E_S|=1$.
Recall that the state space of the verifier $G_v$ has $|X_v|=(|X|\times 2^{|E_s|})^2=4 |X|^2$ states; identical complexity is exhibited by the defensive verifier, i.e., $|X_D|=4 |X|^2$.
Then, the number of states of $RV_E$ is bounded by $|X_{ER}|=|X_v|\times|X_D|=16 |X|^4$.
Overall, the structural complexity of $RV_E$ is $O(|X|^4)$.
\end{remark}
\section{Discussion: Connections with opacity enforcement problems}
\begin{figure*}[htbp]
\centering
\subfigure[Defensive diagnoser $G_{D_d}$.]{
\centering
\includegraphics[width=0.2\textwidth]{DGdfortheexample1.pdf}
}
\subfigure[$E$-diagnoser $D_E$.]{
\includegraphics[width=0.5\textwidth]{DEfortheexample1.pdf}
}
\label{finaleg}
\caption{Concealability enforcement using $E$-diagnoser.}
\end{figure*}
In prior works such as \cite{2014Wu}, \cite{2016Wu}, \cite{2018Wu}, \cite{2018Ji}, the current state opacity enforcement problem is considered based on insertion and edit functions.
Since the enforcement technique using insertion functions is similar to that of using edit functions, we only recall the one using insertion functions here.
More specifically, a desired estimator is proposed to generate safe system actions and a feasible estimator is proposed to enumerate all possible insertion operations in \cite{2014Wu}.
Then, the construction of the so-called All Insertion Structure (AIS) is proposed to depict a game scenario between a given system and a defender; the AIS is constructed by synchronizing the desired estimator and the feasible estimator.
The notion of \emph{$i$-enforceability} is used to verify whether the opacity of the system is enforceable based on insertion functions.
Note that the AIS allows us to obtain a necessary and sufficient condition for $i$-enforceability but it is constructed in exponential time w.r.t. the number of states of the system.
In this work, we utilize similar obfuscation techniques (named defensive functions) to solve the problem of event concealment.
Clearly, concealability enforcement can be converted to a current state opacity problem (e.g., special states reached after the occurrence of secret events can be treated as secret states) and one can subsequently use an $E$-diagnoser (an AIS-like structure) to verify concealability enforcement.
Nonetheless, our focus is not on constructions with exponential complexity (which can provide one necessary and sufficient condition) but on constructions with polynomial complexity, which can provide one necessary condition and one sufficient condition.
The following example is proposed to explain the gap between the two different methods.
\begin{eg}
Recall that Theorem 3 provides a necessary condition, whereas Theorem 4 provides a sufficient condition for $C$-Enforceability.
Based on the sufficient condition and the necessary condition proposed in the previous section, we can try to verify whether a string-based defensive function may be $C$-enforceable via Theorem 3 at first (if the necessary condition is not satisfied then $C$-Enforceability is not possible).
If the necessary condition is satisfied, we can construct a reduced $E$-verifier via Algorithm~2 to check if the sufficient condition holds: if it holds, then $C$-Enforceability is possible (and the construction can also be used to enumerate possible defensive actions).
The issue is that there might exist systems where the necessary condition in Theorem 3 is satisfied whereas the sufficient condition in Theorem 4 is not.
To reach a definite conclusion, for the same example (the same system and the same constraints imposed in the defensive functions) as we described in Example 8, we can use a diagnoser-based approach to track observable behavior of the system and verify concealability.
Specifically, we can construct a defensive diagnoser and an $E$-diagnoser (composition of the diagnoser and the defensive diagnoser) in a similar manner as done using the verifiers, as shown in Fig. 5.
Note that there are also dotted arrows corresponding to $a/$ and $b/$ at states $(\lbrace3S,5N,6S\rbrace,\lbrace3S,5N,6S\rbrace)$ and $(\lbrace4S\rbrace,\lbrace1N,2S\rbrace)$, which means that there is no defensive action from such states.
After removing the above problematic states iteratively, to facilitate presentation, we denote the reduced construction by $RD_E$.
One can obtain a necessary and sufficient condition for $C$-Enforceability in terms of $RD_E$: a string-based defensive function is $C$-enforceable if and only if the $RD_E$ is not empty.
Apparently, in this case, state $(\lbrace4S\rbrace,\lbrace1N,2S\rbrace)$ has no subsequent defensive action and should be removed in $D_E$.
Also state $(\lbrace3S,5N,6S\rbrace,\lbrace3S,5N,6S\rbrace)$ has no subsequent defensive action for event $b$ and should be removed in $D_E$.
Then, the initial state $(\lbrace1N,2S\rbrace,\lbrace1N,2S\rbrace)$ also turns out to be problematic since there is only one defensive action ($c/c$) for event $c$ leading to a problematic state (that has been removed).
Thus, state $(\lbrace1N,2S\rbrace,\lbrace1N,2S\rbrace)$ should also be removed.
In other words, the $RD_E$ for this system is empty and concealability cannot be enforced.
\hfill$\diamond$
\end{eg}
\section{Conclusions}
This paper deals with the problem of event concealment for concealing secret events in a system modeled as an NFA under partial observation.
We formally define concealable and unconcealable secret events.
A necessary and sufficient condition is given to verify whether the system is concealable or not.
If concealability of the system does not hold, then we deal with the problem of concealability enforcement.
The notion of $C$-enforceability characterizes whether an external defender has the capability to use an obfuscation strategy that manipulates the outputs generated by the system such that the occurrence of unconcealable events cannot be revealed.
Note that two scenarios about defensive functions are considered, one without constraints and one with constraints.
It is worth mentioning that our focus of this paper is on the use of reduced complexity constructions (with polynomial complexity) to provide one necessary condition and one sufficient condition for $C$-enforceability.
Future work will touch upon the scenario where each type of defensive operations is assigned with a positive cost and will address the synthesis of a concealing mechanism under a given upper bound on the total cost.
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:intro}
It has been now well established that Galactic globular clusters (GCs) host multiple populations (MPs) of stars, at odds with the traditional paradigm of star clusters harbouring objects all with the same age and uniform initial chemical composition.
A large number of spectroscopic observations have in fact shown that individual GCs are
characterised by anti-cor\-re\-la\-ted star-to-star variations among C, N, O, Na (in some cases also Mg and Al) and He \citep[see, e.g.,][for reviews]{gcb:12, bl18, gratton:19, csreview20}, while
their colour magnitude diagrams (CMDs) confirm negligible age spreads despite the chemical inhomogeneity.
The more popular scenarios for the formation of MPs invoke subsequent episodes of star formation \citep[see, e.g.][]{dercole08, decressin08, r22} whereby stars with CNONa (and He) abundance patterns
similar to those observed in the
field are the first stars to form (we will denote them as P1), while stars
showing a range of N and Na (and He) enrichment and C and O depletion
(we will denote them as P2) were formed several $10^6$ up to $\sim10^8$ years
later --depending on the adopted scenario-- from chemically processed material, ejected by some class of more massive objects born in the first epoch of star formation.
All the proposed scenarios have however difficulties in quantitatively matching the
observed abundance patterns, and no consensus has been reached yet on the mechanism responsible for the formation of MPs in a cluster \citep[see, e.g.][]{bastian15,r15,bl18}.
Complementary to high-resolution spectroscopy, photometry has been crucial to enlarge the sample of clusters investigated for the presence of MPs, the sample of stars surveyed
in each cluster, and the range of evolutionary phases (including the main sequence, generally too faint to be investigated spectroscopically in GCs) where chemical abundance variations have been detected \citep[see, e.g.,][]{sumo, p15, m17, nieder17, dondoglio}. Photometry has also been instrumental for the discovery and characterisation of MPs in massive extragalactic old- and intermediate-age clusters \citep[see, e.g.,][and references therein]{larsen14, dale16, gilligan19, h19,lagioia19,lagioiab, mart19, nardiello19, sara19, sara20, cadelano22}, demonstrating that the MP phenomenon is not restricted to the Milky Way oldest massive clusters.
The ability of photometry to detect MPs rests on the fact that abundance
variations of the elements involved in the anticorrelations affect stellar effective temperatures,
luminosities, and spectral energy distributions
\citep[see, e.g.,][]{salaris:06, M4UBV, yong:08, sswc:11, cmpsf:13, dale16, mucc16, m17, dale18, s19}. For example, filters covering wavelengths shorter than $\lesssim$4500~\AA~can be especially sensitive to star-to-star differences in C, N, and O abundances.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{chmap.pdf}
\caption{Chromosome map of RGB stars in NGC~2808. The left and middle panel display, respectively, the $F814W$-$(F275W-F814W)$ CMD
and the $F814W$-$C_{F275W, F336W, F438W}$ diagram, with overimposed the corresponding red and blue
fiducial lines (see text for details). The right panel displays the resulting chromosome
map. P1 stars are displayed as orange open circles, P2 stars are shown as black crosses.
\label{fig:chmap}}
\end{figure*}
Taking advantage of photometry in the {\em Hubble Space Telescope} ($HST$)
Wide Field Camera 3 filters $F275W$, $F336W$, and $F438W$
from the {\sl UV legacy survey of Galactic GCs} \citep[see, e.g.,][]{p15}, and data in the $F814W$ filter from the Wide Field Channel of the $HST$ Advanced Camera for Survey \citep[][]{sarajedini}, \citet{milone:15, m17} have introduced the pseudo two-colour diagram $\Delta_{F275W, F814W}$-$\Delta_{C\ F275W, F336W, F438W}$ named \lq{chromosome map\rq}. In this diagram, different populations can be easily identified, especially when considering red giant branch (RGB) stars.
The RGB stars belonging to the P1 population of a cluster are expected to be generally distributed around the origin
of the chromosome map coordinates ($\Delta_{F275W, F814W} \sim$ 0, $\Delta_{C\ F275W, F336W, F438W} \sim$ 0),
covering a
narrow range of $\Delta_{F275W, F814W}$ and $\Delta_{C\ F275W, F336W, F438W}$ values, whilst P2 stars (with a range of abundances of
C, N, O, Na and He) span a wide range of both coordinates \citep{milone:15, m17, carretta18}.
On the other hand, as reported by \citet{m17}, the large majority of their sample of 57
Galactic GCs display $\Delta_{F275W, F814W}$ values (and $\Delta_{C\ F275W, F336W, F438W}$ to a much lesser extent) for P1 stars that cover a range much larger than what
expected from photometric errors only.
Unresolved binaries do certainly contribute to produce an extended P1 sequence in
$\Delta_{F275W, F814W}$ , but their fraction in GCs is too small to fully explain
the phenomenon, as shown by \citet{marino:19} and \citet{martins} -- see
also the discussion in \citet{kamann}.
The reason for these extended P1 sequences must be therefore some unexpected level of chemical non-uniformity, that does not alter the relative abundances of the metals. Identifying the culprit of this inhomogeneity has therefore a major impact on our understanding of the mechanism of formation of globular clusters.
In the assumption that P1 stars in a cluster have a uniform metal content, \citet{milone:15}, \citet{He18} and \citet{lsb18} have shown that variations of the initial abundance of He can explain the extended sequences, because of their effect on the stellar effective temperatures: An increased initial He abundance produces hotter, hence bluer RGB stars at a given luminosity, and
lower $\Delta_{F275W, F814W}$ values in the chromosome maps.
The analysis of GC chromosome maps by \citet{marino:19b} concluded more generally that the extended P1 sequences
can be due either to a range of
initial helium abundances, or to a range of metallicity (of about 0.1~dex), the more metal poor component populating the lower $\Delta_{F275W, F814W}$ values (hotter and bluer RGB stars).
The assumption of uniform metallicity was corroborated by the spectroscopy
of six RGB stars distributed along the extended P1 of the GC NGC~2808 by \citet{cabrera19}, which did not reveal any significant spread in metallicity.
Also \citet{latour} did not find any metallicity variation among P1 stars in NGC~2808, by employing MUSE spectra for 1115 RGB stars distributed along
the various sequences of the cluster chromosome map.
On the other hand, the spectroscopic study by
\citet{marino:19} of 18 RGB stars belonging to the extended P1 of NGC~3201 has disclosed a
[Fe/H] range (hence a range of total metallicity if the abundance ratios among the more
abundant metals are uniform, as typically measured in P1 stars) on the order of 0.1-0.15~dex.
Very recently, \citet{legnardi} have studied in detail two metal rich GCs with extended P1,
namely NGC~6362 and NGC~6838 ([Fe/H]$\sim -1.10$ and $\sim -$0.8, respectively), devising appropriate combinations of magnitudes in the
$F275W$, $F336W$, $F438W$ and $F814W$ filters for the clusters' P1 subgiant branch stars,
to disentangle the effect of metallicity and helium variations. By comparisons with theoretical subgiant branch isochrones, they found that a range of metallicity and not $Y$ is present among P1 stars in these two clusters and, by extrapolation, in all other GCs with extended P1 in their chromosome maps. By employing the width of the P1 sequences in the $(F275W-F814W)$ colour
at a reference
$F814W$ magnitude determined by \citet{m17} for 55 GCs, \citet{legnardi} estimated from theoretical isochrones the presence of metallicity ranges --denoted in terms of [Fe/H] ranges-- ranging from a
few 0.01~dex up to much higher values of 0.15-0.30~dex for metal poor GCs, at odds with inferences from spectroscopic analyses, that point to uniform
[Fe/H] abundances in the majority of Galactic GCs \citep[see, e.g.,][]{carrettametal}.
This result has therefore major implications not only for the models of massive clusters' formation, but also for the spectroscopic investigations of chemical abundances in GCs.
For this reasons we believe that a further investigation of the origin of the extended P1 sequences is warranted, to corroborate independently the results by \citet{legnardi}. To this purpose, we present here an alternative method to assess whether a
metallicity or helium spread are the cause of the extended P1 in Galactic GCs. It makes use of information from photometry like \citet{legnardi} study but, instead of subgiants, it uses the same RGB stars along the extended P1 sequence of the chromosome maps to disentangle the effect of metallicity and $Y$, and we show here our results for the GCs NGC~6341 (M~92), NGC~2808, and NGC~6362 \citep[one of the two clusters investigated by ][with their method to disentangle the effect of metallicity and helium]{legnardi}, which cover a large metallicity range, between $\sim -$2.15 and $\sim-$1.10.
Section~\ref{method} introduces the method, based on the properties of theoretical isochrones tested
empirically on the GC M~2, and Sect.~\ref{example} presents an application to our sample of three clusters. Section~\ref{discussion} closes the paper by discussing the results.
\section{RGB stars in the $F814W$-$(F275W-F814W)$ colour magnitude diagram}
\label{method}
The photometric identification of P1 RGB stars takes advantage of
the chromosome maps described by \citet{m17}, that make use of the
combination of a $F814W$-$(F275W-F814W)$ CMD
and a $F814W$-$C_{F275W, F336W, F438W}$ diagram, where the pseudocolor
$C_{F275W, F336W, F438W}$ is defined as $C_{F275W, F336W, F438W}=(F275W-F336W)-(F336W-F438W)$.
They are produced as follows:
First, red and blue fiducial lines in each of these two diagrams are calculated by determining the values of the 4th and the 96th percentile of the $(F275W-F814W)$ and $C_{F275W, F336W, F438W}$ distributions
in various magnitude bins across the RGB, as shown in the left and middle panel of Fig.~\ref{fig:chmap} for the cluster NGC~2808 \citep[the photometry for this cluster and all other clusters discussed in this paper is from][]{ndatabase}.
In these diagrams P2 stars are
distributed between the red and blue fiducials, whilst the supposedly chemically homogeneous P1 stars are expected to lie around the red fiducials.
As a second step, the width of the RGB in the $(F275W-F814W)$ colour
(denoted here as $W_{F275W,F814W}$)
and in the $C_{F275W, F33W, F438W}$ pseudocolour
(denoted as $W_{C\ F275W,F336W,F438W}$)
is calculated at a reference luminosity,
set to two $F814W$ magnitudes above the cluster turn off, taking the colour and pseudocolor differences between the red and blue fiducial.
Finally, for each RGB star the differences in $(F275W-F814W)$ and
$C_{F275W,F336W,F438W}$ with respect to the red fiducials taken at the star $F814W$ magnitude are calculated, and normalised to the values of $W_{F275W,F814W}$ and $W_{C\ F275W,F336W,F438W}$.
These quantities are denoted as $\Delta_{F275W, F814W}$ and
$\Delta_{C\ F275W, F336W, F438W}$ respectively, and are defined as follows:
\begin{equation}
\Delta _{F275W,F814W}= W_{F275W,F814W} \frac{X-X_{\rm R}}{X_{\rm R}-X_{\rm B}}
\label{eqx}
\end{equation}
\begin{equation}
\Delta _{C\ F275W,F336W,F438W} = W_{C\ F275W,F336W,F438W} \frac{Y_{\rm R}-Y}{Y_{\rm R}-Y_{\rm B}},
\label{eqy}
\end{equation}
where X=$(F275W-F814W)$, Y=$C_{F275W,F336W,F438W}$, and R and B denote the red and blue fiducial line, respectively.
The chromosome map is the plot of the stars' position in a $\Delta_{C\ F275W, F336W, F438W}$ vs $\Delta_{F275W, F814W}$ diagram, shown in the right panel
of Fig.~\ref{fig:chmap} for the GC NGC~2808.
Following the definitions of these two quantities, the
values $\Delta_{F275W, F814W}$=0 and
$\Delta_{C\ F275W, F336W, F438W}$=0
correspond to objects lying on the red fiducial
lines while $\Delta$ values different from zero denote colour and
pseudocolour distances (defined as positive for $C_{F275W,F336W,F438W}$, and negative for $(F275W-F814W)$) from such lines.
The P2 stars typically follow a sequence extending from the origin of both coordinates towards increasing
$\Delta_{C\ F275W, F336W, F438W}$ and decreasing $\Delta_{F275W, F814W}$, as shown in
Fig.~\ref{fig:chmap}, whilst
P1 stars should be clustered around the origin because of their expected chemical homogeneity,
with a small spread just due to photometric errors \citep[see, e.g.][for more details]{m17, csreview20}.
However, as already mentioned, a large number of GCs display extended P1
sequences in the chromosome map of their RGB stars, like the case of NGC~2808 shown in Fig.~\ref{fig:chmap}, where
the sequence of P1 stars is clearly elongated mainly
along the $\Delta_{F275W, F814W}$ axis \citep[see, e.g.][for more details about the identification
of this elongated sequence with P1 stars]{m17, csreview20}, meaning that P1 objects describe multiple RGB sequences between the red and blue fiducials in the $F814W$-$(F275W-F814W)$ CMD.
In the chromosome maps these multiple RGB sequences are compressed essentially along one segment, because the way the maps are calculated. In the following we show how we can employ the $F814W$-$(F275W-F814W)$ CMD of P1 stars to help discriminate between inhomogeneities in He and metallicity as the origin of the extended P1 phenomenon.
Figure~\ref{fig:deltaiso} displays the $F814W$-$(F275W-F814W)$ CMD
of two $\alpha$-enhanced 12~Gyr isochrones
with [Fe/H]=$-$1.3 and $-$1.6 respectively, from
\citet{pietrinferni:06}. The relative abundance distribution of the metals
is typical of P1 stars and is the same for the two isochrones; it is just the total metallicity (hence [Fe/H]) that varies.
Due to the shape of the RGBs in this CMD, the colour difference at fixed magnitude
between the two isochrones increases with
decreasing $F814W$. This is more clearly seen in the lower panel of the figure, where we display the more metal rich RGB as a reference straight line at $(F275W-F814W)$=0, and plot
on the horizontal axis the colour differences $\Delta(F275W-F814W)$
between the metal poorer RGB and this reference line at varying $F814W$,
in the range between about 2 and 4 magnitudes above the main sequence turn off, which is approximately the magnitude range considered in the analysis presented in the next section.
\begin{figure}
\centering
\includegraphics[width=0.90\columnwidth]{figDeltacol-crop.pdf}
\caption{{\sl Upper panel:} The $F814W$-$(F275W-F814W)$ CMD of
two RGBs belonging to 12~Gyr old isochrones with [Fe/H]=$-$1.6
(dashed line) and $-$1.3 (solid line), respectively. {\sl Lower panel:} Difference
of the $(F275W-F814W)$ colour at varying $F814W$ between
the [Fe/H]=$-$1.6 (dashed line) and the \lq{verticalised\rq} metal richer RGBs (solid line),
taken as reference (see text for details).
\label{fig:deltaiso}}
\end{figure}
The colour separation $\Delta(F275W-F814W)$ increases when moving to brighter magnitudes, while
the effect of a spread of initial helium abundance (at fixed metallicity) is different,
as shown by Fig.~\ref{fig:deltaisoHe}. Here we display the RGB of the
[Fe/H]=$-$1.3 isochrone of the previous figure, together with an isochrone for
the same metallicity and age, but calculated with the helium mass fraction $Y$ increased by 0.05.
In this case, the colour separation $\Delta(F275W-F814W)$ between the two RGBs
tends to decrease when moving to brighter magnitudes, the opposite behaviour compared to the case
of metallicity variations.
\begin{figure}
\centering
\includegraphics[width=0.90\columnwidth]{figDeltacolHe-crop.pdf}
\caption{As Fig.~\ref{fig:deltaiso} but this time the dashed line represents a 12~Gyr old isochrone with [Fe/H]=$-$1.3 and the helium mass fraction $Y$ increased by
0.05.
\label{fig:deltaisoHe}}
\end{figure}
In addition to these predictions from theory,
the GC M~2 gives us the opportunity to verify empirically the behaviour of RGBs with different metallicity predicted by the models.
This cluster has been found to host stars characterised by large iron variations, with three main metallicity groups at [Fe/H] = --1.7, --1.5, and --1.0, respectively \citep{yong14}. While the presence of an intrinsic metallicity spread between the two metal poorer groups (we denote these groups as the main cluster component) is still controversial \citep{lardo16}, several independent studies have confirmed the existence of a small component (accounting for $\sim$1\% of the cluster mass)
with [Fe/H]=--1 \citep{yong14,miloneM2,lardo16}. Such metal rich stars do not exhibit star-to-star variations in light elements or $s$-process enhancement (e.g. they have P1-like chemical composition; \citealp{yong14}). They are located on a well defined, narrow red sequence which runs parallel to the main RGB body and that can be followed down to the SGB and main sequence, supporting the case for cluster membership \citep{miloneM2}.
\begin{figure}
\centering
\includegraphics[width=.72\columnwidth]{m2_P1_ALL.pdf}
\caption{{\sl Top panel:} Chromosome map of M~2. The P1 and P2 stars in the main cluster component are displayed as large orange circles and black crosses, respectively, while the P1 metal rich component is shown as green squares.
{\sl Middle panel:} $F814W$-$(F275W-F814W)$ CMD of the three groups of stars shown
in the chromosome map.
{\sl Bottom panel:} Similar to the right panel of Fig.~\ref{fig:deltaiso}. The orange circles display the difference
of the $(F275W-F814W)$ colour between the P1 stars of the main cluster component, and the \lq{verticalised\rq} fit to the CMD of the metal rich component taken as reference (dashed line). The individual metal rich P1 stars are also plotted around the verticalised fit.
\label{fig:m2_new}}
\end{figure}
The top panel of Fig.~\ref{fig:m2_new} displays the chromosome map of the cluster,
including the metal rich P1 component --that
comprises a very small fraction of the cluster population, and therefore lies to the red
of the red fiducials in the diagrams used to calculated the chromosome map-- located at values
$\Delta_{F275W, F814W}$ larger than 0.5.
The $F814W$-$(F275W-F814W)$ CMD used to determine the chromosome map is
shown in the middle panel of
the same figure, whilst the bottom panel displays the colour
differences $\Delta(F275W-F814W)$ between stars belonging to the P1 of the main component and the \lq{verticalised\rq} cubic fit to the position of the metal rich P1 stars in the $F814W$-$(F275W-F814W)$ CMD, as a function of $F814W$.
We can see clearly that the $\Delta(F275W-F814W)$ values increase with
increasing brightness along the RGB, as predicted by theory.
As a conclusion, the opposite trends of the colour differences $\Delta(F275W-F814W)$
with $F814W$ for the case
of RGBs with different metallicity or different initial helium, provides us with a diagnostic that can be applied to the GCs with extended P1.
To assess how this diagnostic performs in case of populations with a continuous
distribution of
helium or metal abundances, we performed a few numerical experiments.
Figure~\ref{fig:sinObs} shows the equivalent of Figs.~\ref{fig:deltaiso} and \ref{fig:deltaisoHe}, but for the $F814W$-$(F275W-F814W)$ CMDs of
three pairs of synthetic samples with 200 RGB stars each, calculated from 12~Gyr $\alpha$-enhanced isochrones \citep{pietrinferni:06}. This number of stars is a typical average value for the clusters
in \citet{m17} sample, that we will use for the analysis described below.
The left panels
of each row display samples with a spread of metallicity (0.15~dex and uniform distribution) while the right panels are samples with a spread of
$Y$ ($\Delta Y$=0.05 and uniform distribution)\footnote{These ranges of metallicity and $Y$ produce
widths of the RGB in the $F814W$-$(F275W-F814W)$ CMD, that are consistent with the ranges measured
by \citet{m17}.}. From top to bottom, the three rows display simulations in order of increasing metallicity regimes. Random Gaussian photometric errors with 1$\sigma$ dispersion typical of
the cluster photometries used in the next section (average values on the order of $\sim$0.01~mag) are also included in these simulations.
The straight red lines located at $\Delta(F275W-F814W)$=0.0 are the verticalised red fiducials of the $F814W$-$(F275W-F814W)$ CMDs of the samples. Each red fiducial corresponds to a fit to the 95th percentile (to account for potential outliers due to the photometric error) of the colour distribution as a function of $F814W$.
In the simulations with a metallicity spread, the red fiducials represent the position in the CMD of the populations with the highest metallicity, whilst in case of the simulations with a range of $Y$ the red fiducials denote the position of the objects with the lowest $Y$.
The horizontal coordinate of each panel displays the colour difference $\Delta(F275W-F814W)$ between the individual synthetic stars and the corresponding red fiducial at the star $F814W$ magnitude.
The various panels show that, in the case of a metallicity spread, the range of $(F275W-F814W)$ colours spanned by the population increases with increasing $F814W$ luminosity, and this trend steepens
with increasing mean metallicity. This is evident when considering the
second straight line displayed in each panel, which is the blue fiducial of the distribution of points in the diagram; it has been derived as a
linear fit to the 5th percentile of the distribution of the colour differences as
a function of $F814W$.
In the metal poor regime ([Fe/H] around $-$2) this line is almost vertical,
and the slope (always in the sense of increasing $\Delta(F275W-F814W)$ with
increasing $F814W$ luminosity) increases with increasing mean metallicity.
Moreover, for a fixed spread of [Fe/H], the range of $(F275W-F814W)$ colours increases with
increasing mean metallicity.
The opposite effect can be seen in the case of populations with a range of $Y$.
The range of $\Delta(F275W-F814W)$ values decreases with increasing luminosity, at all metallicities.
The exact value of the slope --not the sign-- of the blue edge of the $\Delta(F275W-F814W)$ distribution as a function of $F814W$, and whether
the value of the slope is significantly different from zero depends in general on the size of the stellar sample, the photometric error (that in our simulations are set to the typical values of the GC sample we are going to consider in our analysis), the value of the mean metallicity,
the metallicity (or $Y$) range of the RGB sample, and, very importantly, on the probability distribution of the various metallicities (or $Y$ values) within the adopted range.
In the case of the choice of parameters of Fig.~\ref{fig:sinObs}, multiple realizations of the simulations show that in both scenarios we can always recover a slope that is significantly different from zero.
We have then varied the probability distribution of metallicity and $Y$ by performing the same simulations described before, but this time
assuming Gaussian distributions for both metallicity and $Y$. The mean values and 1$\sigma$ spreads were chosen in order to cover, within $\pm$3$\sigma$ from the mean, the same total ranges
of [Fe/H] and $Y$ as in Fig.~\ref{fig:sinObs}.
Also in this case we formally derive slopes of the blue fiducials of the $\Delta(F275W-F814W)$ distribution as a function of $F814W$
with different sign in case of metallicity or $Y$ variations, but in many realizations they were often not significantly different from zero.
The implication of these numerical experiments is that it is possible to use the distribution of P1 RGB stars
in the $F814W$-$(F275W-F814W)$ CMDs to determine whether a range of metallicity or $Y$ is present, but
with the available observational data we can find statistically significant results only for certain distributions of metallicity (or $Y$) among the cluster P1 stars. This is similar to the situation with the method employed by \citet{legnardi} based on photometry of subgiants, that can be applied to clusters with P1 and P2 stars well separated along the subgiant branch.
\begin{figure}
\centering
\includegraphics[width=.98\columnwidth]{sint_poor.pdf}
\includegraphics[width=.98\columnwidth]{sint_int.pdf}
\includegraphics[width=.98\columnwidth]{sint_rich.pdf}
\caption{As Fig.~\ref{fig:deltaiso}, but for synthetic samples of P1 stars
(200 objects) including a 0.15~dex uniform spread of metallicity (left panels), or
a uniform spread $\Delta Y$=0.05 (right panels).
The top row displays the results for populations with
[Fe/H] distributed between $-$2.15 and $-$2.0~dex, and for a spread of Helium at
[Fe/H]=$-$2.15. The middle row shows the case for [Fe/H] between $-$1.6 and $-$1.45~dex,
and for a spread of helium at [Fe/H]=$-$1.3. The bottom row displays the case
for [Fe/H] between $-$0.85 and $-1.0$~dex, and for a range of helium at [Fe/H]=$-0.7$.
In each panel the solid vertical lines at $\Delta(F275W-F814W)$=0 are the verticalised red fiducials of the $F814W$-$(F275W-F814W)$ CMDs of the samples. The other line in each panel correspond to the blue fiducial of the $\Delta(F275W-F814W)$ distribution as a function of $F814W$ (see text for details).
\label{fig:sinObs}}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=.92\textwidth]{m92_P1_ALL.pdf}
\includegraphics[width=.92\textwidth]{n2808_P1_ALL.pdf}
\includegraphics[width=.92\textwidth]{n6362_P1_ALL.pdf}
\caption{From top to bottom, the three rows of panels refer
to the clusters M~92, NGC~2808, and NGC~6362 respectively. In each row
the left panel displays the cluster RGB chromosome map, with small black crosses
denoting P2 stars, and orange circles the P1 population.
The middle panel displays the
corresponding $F814W$-$(F275W-F814W)$ CMD, with a
cubic polynomial fit to the red edge of the P1 sequence (red solid line -- see text).
The faint limit of the CMDs corresponds approximately to 1.5-2.0~mag above the turn off.
The right panel shows the
difference of the $(F275W-F814W)$ colour of P1 stars
with respect the \lq{verticalized\rq} fit shown
in the middle panel (red line).
The solid black line is a linear best fit to the 5th percentile of the colour differences as a function of $F814W$ (see text for details), corresponding to the blue fiducial of the $\Delta(F275W-F814W)$ distribution as a function of $F814W$,
while the dotted line denotes a comparison vertical sequence, to emphasise the sign of the slope the blue fiducial in this diagram (see text for details). The dashed black line is the best fit to the 50th percentile of the colour differences as a function of $F814W$.
A colour-coded number density 2D map is also displayed.
The average 1-$\sigma$ $(F275W-F814W)$ photometric error over the $F814W$ covered in the figure is show in the bottom left corner of each panel.
\label{fig:n2808_new}}
\end{figure*}
\section{The extended P1 of three Galactic globular clusters}
\label{example}
Based on the results of the previous section, we investigate here the behaviour in the $F814W$-$(F275W-F814W)$ CMD
of RGB stars belonging to three GCs in \citet{m17} sample,
with well populated RGB photometry in the appropriate filters, and a well extended P1 in their chromosome maps, namely NGC~6341 (M~92), NGC~2808,and NGC~6362 \citep[see][]{m17}, with spectroscopic
[Fe/H]= --2.35, --1.18, and --1.07, respectively \citep{carrettametal}\footnote{The number of RGB stars in these clusters' photometries range between $\sim$150 (for NGC6362) and $\sim$400 (for NGC~2808).}.
Unfortunately we cannot employ in our analysis the GC NGC~3201 --whose P1 RGB stars display a range of metallicities according to the spectroscopic analysis by \citet{marino:19}-- because the available photometry does not have enough stars along the RGB to perform our analysis (see the previous section).
For each cluster we have determined first the chromosome map, to identify their P1 stars.
Given that we are interested in high-precision photometry of cluster members, we retain for our analysis only stars with membership probability greater than $>95\%$. Additionally, only stars with relatively small photometric uncertainties ($<$0.05, 0.04, 0.03, 0.02, 0.01~mag in F275W, F336W, F438W, F606W and F814W respectively) have been selected to produce the chromosome maps.
Cluster members along the RGB have been identified by performing a multivariate analysis on the photometric data through a Gaussian mixture model algorithm (GMM; \citealp{gmm}). For each giant star in a given cluster photometry, we have computed colour (and colour-indices) from all possible combination of the magnitudes available in the \citet{nardiello19} catalogues --including $(F275W-F814W)$ and $C_{F275W, F336W, F438W}$. Member stars have then been selected using the constraint that similar objects (e.g. stars belonging to the cluster) lie on a well-defined RGB sequence in the multi-dimensional space defined by all the possible CMD combinations. In such a way, stars which are likely contaminants (field stars and/or binaries) are easily spotted by the algorithm because they are scattered around when using different colour (or colour index) combinations to create the CMD.
Once the cluster RGB stars have been selected, we have followed the procedure described in Sect.~\ref{method} to derive the chromosome maps shown in Fig.~\ref{fig:n2808_new}.
The P1 stars have been then identified according to their position in the chromosome map \citep[following][]{m17} and they are plotted as orange circles in figure~\ref{fig:n2808_new}.
The left hand panels of Fig~\ref{fig:n2808_new} show the chromosome maps of (from top to bottom) M~92, NGC~2808, and NGC~6362 respectively.
From a visual comparison they are very similar to the ones published by \citet{m17}. Also, the number ratios of P1 stars to the total number of RGB stars we obtained for these clusters are consistent with the values listed in Table~2 of \citet{m17}.
The middle panels of Fig.~\ref{fig:n2808_new} show the position of the RGB stars of the chromosome maps in the $F814W$-$(F275W-F814W)$ CMDs. The redder envelope of the P1 population in these CMDs has been derived by fitting the 95th percentile (to account for potential outliers) of the
colour distribution as a function of $F814W$ with a third order polynomial.
This red fiducial line is shown in red in the middle panels of the same figures, and corresponds
to the location of P1 stars with either the highest metallicity --if metallicity varies
along the extended P1 sequence in the chromosome maps-- or the lowest helium abundance --if helium changes along the P1 sequence.
The $\Delta(F275W-F814W)$ colour differences between the individual RGB stars and the fiducial lines in the middle panels are plotted in the right hand panels of Fig.~\ref{fig:n2808_new}. The red solid lines denote the verticalised red fiducials described before, whilst the black solid lines are linear fits to the 5th percentile of the $\Delta(F275W-F814W)$ distribution as a function of $F814W$, and
represent the blue edge, or blue fiducial, of the distribution of $\Delta(F275W-F814W)$ values.
We have verified that for each cluster the sign of the slope of this blue fiducial does not change if we fit alternatively the 10th or even the 15th percentile of the colour distribution, to be more conservative regarding the presence of unresolved binaries and blue
stragglers progeny
(that have been already largely filtered out by our procedure to determine the chromosome maps) which are predicted to be
located mainly at the bluer end of the extended P1 sequence in the chromosome maps
\citep[see][]{martins, marino:19}.
We find in all clusters that the range of $\Delta(F275W-F814W)$ values increase with increasing $F814W$ brightness; this is a clear signature for the presence of a range of metal abundances among their P1 stars (see Fig.~\ref{fig:sinObs}). Errors on the magnitudes show a decreasing trend with increasing luminosity, thus the photometric error does not contribute to the broadening of the RGB with luminosity.
The lines denoting the median of the colour differences as a function of $F814W$, and the number density of points in these diagrams also show
in all three clusters a general trend consistent with the sign of the slope of the blue fiducial. The absolute values of the slope of the blue fiducial in all three clusters are significantly different from zero at more than 95\% confidence level.
A range of helium as the
main driver of the observed range of $(F275W-F814W)$ colours among P1 stars seems to be ruled out.
\section{Summary and discussion}
\label{discussion}
We have exploited $HST$ near-UV and optical photometry of RGB stars in the Galactic GCs
M~92, NGC~2808, and NGC~6362, to determine whether metal or helium abundance variations are the cause of the extended P1 sequences in their chromosome maps.
The cluster NGC~6362 is in common with \citet{legnardi} study, and we find for this cluster and all others in our sample that a range of metal abundance does exist among their P1 stars.
This confirms the results of \citet{legnardi} based on the analysis of subgiant branch stars in NGC~6362 and NGC~6838, and solidifies the case
for the existence of unexpected variations of metal abundances in most globular clusters.
\begin{figure}
\centering
\includegraphics[width=.479\textwidth]{sinFe.pdf}
\caption{{\em From left to right:} the $F814W$-$(F275W-F814W)$ and $F814W$-C$_{(F275W,F336W,F438W)}$
diagrams for a synthetic sample with --1 $<$ [Fe/H] $<$ --0.85) like the one shown in Fig.~\ref{fig:sinObs}.
The blue and red fiducial lines shown in each panel as solid lines (determined as described in Sect.~\ref{method})
highlight the different widths of the RGB in the two diagrams.
The width (in magnitudes) of the horizontal axis of the two figures is the same.
\label{fig:p1_map}}
\end{figure}
A range of metallicity does not only explain the P1 extension in the $(F275W-F814W)$ colour, but also the associated smaller range of the $C_{F275W, F336W, F438W}$ pseudocolour covered by P1 stars in the
chromosome maps. This is shown in
Fig.~\ref{fig:p1_map}, which displays the $F814W$-$(F275W-F814W)$ and the $F814W$-$C_{F275W, F336W, F438W}$ diagrams of a synthetic population with --1 $<$ [Fe/H] $<$ --0.85, like the one shown in Fig.~\ref{fig:sinObs}.
Assuming that a metallicity spread --as found for NGC~6362 and NGC~6838-- is the cause of the P1 extension in the chromosome maps of all clusters in \citet{m17} sample, \citet{legnardi} determined the correspondent range of [Fe/H]
--a proxy for a range of total metallicity in P1 stars-- for the sample studied by \citet{m17}, comparing the width of the RGB ($W_{F275W,F814W}$) taken two $F814W$ magnitudes above the main sequence turn off tabulated by \citet{m17} with theoretical isochrones.
They derived internal variations ranging from less than 0.05 to 0.30~dex, that mildly correlate with cluster mass and metallicity.
Our study corroborates these conclusions, by directly
establishing the existence of a metallicity spread in a slightly larger sample of clusters distributed over a wider [Fe/H] range compared to \citet{legnardi} study, using an independent method.
Figure~\ref{fig:n2808_new} also shows that
the P2 distribution in the $F814W$-$(F275W-F814W)$ CMDs of NGC~6362 appears clustered around the blue edge of the P1 distribution, whilst for both M~92 and NGC~2808 the red
edge of the P2 distribution is close to the red edge of the P1 one.
Also, in case of M~92 the width of the P2 sequence is
comparable to the width of the P1 one, whilst it is larger in NGC~2808.
This has implications for the metallicity distribution of P2 stars in these clusters,
if we consider the analysis by \citet{lsb18} \citep[see also][]{He18} who have shown that at fixed metallicity the $(F275W-F814W)$ colour of RGB P2 stars becomes bluer when the extension of the light element anticorrelations (due mainly to the decrease of oxygen) and $Y$ increase.
The red edge of the P2 CMD corresponds to the P2 stars
with the smaller values of the light element abundance anticorrelations and
helium enhancement, hence its position with respect to the P1 colour distribution
is likely a fair indicator of the metallicity out of which these P2 stars have formed.
In case of NGC~6362 these P2 stars have formed with the same [Fe/H] of the metal poorer P1 population, while in the other clusters they formed with a [Fe/H] closer to the metal richer P1 component.
Also, given that at fixed [Fe/H] the $(F275W-F814W)$ colour of RGB P2 stars increases
with increasing $Y$ and decreasing oxygen, a comparable or narrower width of the P2
sequence compared to the P1 counterpart suggest a smaller
[Fe/H] range in the P2 cluster population.
These conclusions echo those by \citet{legnardi}, based on the distribution of the P1 and P2 stars in the chromosome maps, and highlight the emerging ever increasing complexity of the chemical makeup of GCs.
The existence of a metallicity spread amongst P1 stars and a different distribution of the total metallicity in P1 and P2 populations might have implications for the determination of the relative He abundances, based on the comparison of theoretical and observed
colours of P1 and P2 RGB stars \citep{He18}, that need to be assessed. We also need to investigate the
impact of both helium and total metallicity variations on the interpretation of the horizontal branch morphologies of individual clusters, using synthetic horizontal branch modelling
\citep[see, e.g.,][]{dale13, csp14, scp, tailo19}.
Finally, it is also important to further validate these results about metallicity spreads with dedicated high-resolution spectroscopic analyses.
So far, as already mentioned, spectroscopy of NGC~2808 P1 stars distributed along the extended sequence in the chromosome map has not disclosed any metallicity spread, and
for example the spectroscopic analysis of a sizable sample of P1 RGB stars in NGC~6362 by \citet{muc16} did not reveal any statistically significant spread.
This lack of detection might be due to measurements not sufficiently precise compared to the range of metal abundances in these two clusters, that are
equal to $\sim$0.1~dex and $\sim$0.05~dex for NGC~2808 and NGC~6362 respectively \citep[see][]{legnardi}. Moreover, in case of NGC~6362 it is possible that P1 stars on
the blue side of the chromosome map have not been sampled by the observations.
In the case of M~92 \citet{langer} found a spread of 0.18~dex (consistent with
$\sim$~0.15~dex estimated by \citealp{legnardi} for P1 stars) in iron peak elements among three bright red giants, but a recent analysis by \citet{mesz} does not find any significant metallicity spread in this cluster.
\begin{acknowledgements}
The authors would like to thank Ivan Cabrera-Ziri for
useful comments and discussions.
M. Salaris acknowledges support from The Science and Technology Facilities Council Consolidated Grant ST/V00087X/1.
C. Lardo acknowledges funding from Ministero dell'Università e della Ricerca through the Programme {\em Rita Levi Montalcini} (grantPGR18YRML1).
Isochrones and photometry used in this study are available at \url{http://basti.oa-teramo.inaf.it/index.html} and
\url{https://archive.stsci.edu/prepds/hugs/}, respectively.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
The physical manifestations of vacuum electromagnetic nonlinearities have been a fascinating topic of research since the discovery by Euler and Heisenberg \cite{EH} of a striking prediction of quantum electrodynamics (QED), that is, the light-by-light scattering arising from the interaction of photons with virtual electron-positron pairs. As is well known, the physical consequences of this crucial finding, such as vacuum birefringence and vacuum dichroism, have been largely considered from different points of view \cite{Adler, Costantini,Ruffini,Dunne}. However, despite remarkable progress \cite{Bamber,Burke,Pike,Tommasini1,Tommasini2,PVLAS}, this prediction has not yet been confirmed.
It is appropriate to remark, in this context, that recently the ATLAS and CMS collaborations at the Large Hadron Collider (LHC) have reported on the high energy gamma-gamma pair emission from virtual gamma-gamma scattering in ultraperipheral Pb-Pb collisions \cite{Atlas,Cms}. However, as emphasized in \cite{Robertson2}, in these results there is no modification of the optical properties of the vacuum. In addition, the coming of laser facilities has given rise to various proposals to probe quantum vacuum nonlinearities \cite{Battesti, Ataman}.
An interesting example is provided by the experiment (DeLLight project) \cite{Robertson2}, which exploits the change in the index of refraction due to nonlinear electrodynamics.
In this connection, it may be recalled that different nonlinear electrodynamics of the vacuum may have significant contributions to photon-photon scattering such as Born-Infeld \cite{Born} and Lee-Wick \cite{Lee-Wick} theories. As is well known, these electrodynamics were introduced in order to avoid the divergences inherent in the Maxwell theory at short distances.
With these ideas in mind, in previous works \cite{Nonlinear,Logarithmic,Nonlinear2,Nonlinear3}, we have considered the physical effects presented by different models of $(3+1)$-D nonlinear electrodynamics in vacuum. Evidently, this has helped us to gain insights into the peculiarities of quantum vacuum nonlinearities in different contexts. For example, the Generalized Born-Infeld, and Logarithmic Electrodynamics the field energy of a point-like charge is finite, which also exhibit the vacuum birefringence phenomenon. As well as we have studied the lowest-order modifications of the static potential within the framework of the gauge-invariant but path-dependent variables formalism, which is an alternative to the Wilson loop approach.
We further note that recently an interesting study on vacuum Cherenkov radiation in Euler-Heisenberg-like nonlinear electrodynamics has been considered in \cite{Macleod}. As is well known, a charged particle moving in a medium under an external electromagnetic field emits Cherenkov radiation when its velocity of light exceeds the phase velocity in that medium.
It is worth recalling, at this stage, that any variation of the velocity of light with respect to $c = {1 \mathord{\left/
{\vphantom {1 {\sqrt {{\varepsilon _0}{\mu _0}} }}} \right.
\kern-\nulldelimiterspace} {\sqrt {{\varepsilon _0}{\mu _0}} }}$, where ${\varepsilon _0}$ and ${\mu _0}$ are the vacuum permittivity and the vacuum permeability respectively, is due to that light propagates in a medium. In this manner, we have to introduce the constants $\varepsilon$ and $\mu$ which characterize the medium. As is well known, the velocity of light in a medium is less than the velocity of light in vacuum by a factor (index of refraction) $n = {{\sqrt {\varepsilon \mu } } \mathord{\left/
{\vphantom {{\sqrt {\varepsilon \mu } } {\sqrt {{\varepsilon _0}{\mu _0}} }}} \right.
\kern-\nulldelimiterspace} {\sqrt {{\varepsilon _0}\,{\mu _0}} }}$.
From the previous remark it follows that both $\varepsilon$ and $\mu$ are positives. Nevertheless, as was first hypothesized in \cite{Veselago}, considerable attention has been paid recently to the $\varepsilon<0$ and $\mu<0$ case. The interest in studying this case is mainly due to the laboratory construction of an exotic form of dielectric,
{\it metamaterial}, with both $\varepsilon$ and $\mu$ negatives. As it offers a valuably observational window on the constitutive parameters $\varepsilon$ and $\mu$, electrodynamics with
metamaterial features has stimulated lots of experimental works \cite{Caloz}.
In this context it is particularly important to notice that, as emphasized in \cite{Caloz}, the four possible sign combinations in the pair $(\varepsilon, \mu)$ are $(+,+)$, $(+,-)$, $(+,-)$ and $(-,-)$. Evidently, this last combination corresponds to Veselago's materials. Thus, in this work we focus on the interesting possibility if non-linearities can induce any of the above combinations, excepting the first one. In other words, one of our goals is to understand what might be another observational signature of nonlinearities.
Another of our goals is devoted to study the stability of the above radiation scenario in the presence of other nonlinear electrodynamics.
Of special interest will be to check the effects of the ''medium'' on the production of this radiation. To do this, we will work out the radiated energy for logarithmic electrodynamics following the conventional path of calculating the Poynting vector. As we will see, our analysis renders manifest the vital role played by vacuum electromagnetic nonlinearities in triggering the radiated energy for logarithmic electrodynamics.
Our work is organized as follows. In Sec. 2, we describe the vital role played by nonlinearities in triggering a material with $(\varepsilon > 0, \mu <0)$, which corresponds to a ferrimagnetic material. Subsequently, the permittivity and susceptibility induced by nonlinearities are considered in order to obtain the refractive index via the dispersion relation for logarithmic electrodynamics.
This would not only provide the theoretical setup for our subsequent work, but also fix the notation. In Sec. 3, we consider the calculation of the electromagnetic radiation. A summary of our work is the subject of Sec. 4.
In our conventions the signature of the metric is ($+1,-1,-1,-1$).
\section{General aspects}
\subsection{On ferrimagnetic materials}
As already mentioned, we now explore the interesting possibility if non-linearities can induce any of the above combinations, excepting the first one.
For this purpose we begin by considering a generic Lagrangian density:
\begin{equation}
{\cal L} = {\cal L}({\cal F},{\cal G}), \label{Meta05}
\end{equation}
where the arguments of ${\cal L}$ are the usual electromagnetic field invariants, that is, ${\cal F} \equiv - \frac{1}{4}{F_{\mu \nu }}{F^{\mu \nu }} = \frac{1}{2}\left( {\frac{{{{\bf E}^2}}}{{{c^2}}} - {{\bf B}^2}} \right)$ and
${\cal G} \equiv - \frac{1}{4}{F_{\mu \nu }}{\tilde F^{\mu \nu }} = \frac{{\bf E}}{c} \cdot {\bf B}$.\\
Next, after splitting ${F^{\mu \nu }}$ in the sum of a classical background, $F_B^{\mu \nu }$, and a small fluctuation, ${f^{\mu \nu }}$, the corresponding linearized field equations read
\begin{eqnarray}
{\partial _\mu }\left( {{C_1}{f^{\mu \nu }} + {C_2}{{\tilde f}^{\mu \nu }}} \right) - \frac{1}{2}{\partial _\mu }\left( {k_B^{\mu \nu \kappa \lambda }{f_{\kappa \lambda }} + t_B^{\mu \nu \kappa \lambda }{{\tilde f}_{\kappa \lambda }}} \right) \nonumber\\
- \frac{1}{4}{\partial _\mu }\left( {{\varepsilon ^{\mu \nu \kappa \lambda }}{t_{B\kappa \lambda \rho \sigma }}{f^{\rho \sigma }}} \right)
= - {\partial _\mu }\left( {{C_1}F_B^{\mu \nu } + {C_2}{{\tilde F}^{\mu \nu }}} \right) + {j^\nu }, \nonumber\\
\label{Meta10}
\end{eqnarray}
where $k_B^{\mu \nu \kappa \lambda } = {D_1}F_B^{\mu \nu }F_B^{\kappa \lambda } + {D_2}\tilde F_B^{\mu \nu }\tilde F_B^{\kappa \lambda }$ and $t_B^{\mu \nu \kappa \lambda } = {D_3}F_B^{\mu \nu }F_B^{\kappa \lambda }$.
Whereas ${C_1} = {\left. {\frac{{\partial {\cal L}}}{{\partial {\cal F}}}} \right|_B}$, ${C_2} = {\left. {\frac{{\partial L}}{{\partial {\cal G}}}} \right|_B}$, ${D_1} = {\left. {\frac{{{\partial ^2}{\cal L}}}{{\partial {{\cal F}^2}}}} \right|_B}$, ${D_2} = {\left. {\frac{{{\partial ^2}{\cal L}}}{{\partial {{\cal G}^2}}}} \right|_B}$ and ${D_3} = {\left. {\frac{{{\partial ^2}{\cal L}}}{{\partial {\cal F}\partial {\cal G}}}} \right|_B}$.
At this stage, we are not bound to consider a constant and uniform electromagnetic background, so that the coefficients $C_1$, $C_2$, $D_1$, $D_2$ and $D_3$ are, in principle, space-time-dependent. This is why the
field equations take the form of eq. (\ref{Meta10}). However, in what follows, we adopt space-time constancy of the background, so that the coefficients above are not acted upon by the space-time derivatives,
giving rise to the constitutive tensors to be present soon below.
However, in what follows we will write the equations of motion in the case $j^{\nu}=0$, in the presence of a constant background with both electric and magnetic fields $({\bf E},{\bf B})$. We thus find
\begin{equation}
\nabla \cdot {\bf d} = 0, \label{Meta15}
\end{equation}
where ${d_i} = {\varepsilon _{ij}}{e_j} + {\xi _{ij}}{b_j}$.
Whereas, ${\varepsilon _{ij}}$ and ${\xi _{ij}}$ are given by
\begin{equation}
{\varepsilon _{ij}} = {\delta _{ij}} + {\alpha _i}{E_j} + {\beta _i}{B_j}, \label{Meta20a}
\end{equation}
and
\begin{equation}
{\xi _{ij}} = - {c^2}{\alpha _i}{B_j} + {\beta _i}{E_j}, \label{Meta20b}
\end{equation}
here we have used the notation $ \boldsymbol{\alpha} \equiv \frac{1}{{{C_1}}}\left( {\frac{{{D_1}}}{{{c^2}}}\,{\bf E} + \frac{{{D_3}}}{c}\,{\bf B}} \right)$ and $\boldsymbol{\beta} \equiv \frac{1}{{{C_1}}}\left( {{D_2} \,{\bf B} + \frac{{{D_3}}}{c}\, {\bf E}} \right)$. Note that the tensors $\varepsilon$ and $\xi$ are completely determined by the electromagnetic background. Throughout, ${\bf e}$ and ${\bf b}$ are the electric and magnetic fields arising from the fluctuation ${f^{\mu \nu }}$.
On can now further observe that
\begin{equation}
\nabla \times {\bf h} = \frac{1}{{{c^2}}}\frac{\partial }{{\partial t}}{\bf e}, \label{Meta25}
\end{equation}
where ${h_i} \equiv \mu _{ij}^{ - 1}{b_j} + {\eta _{ij}}{e_j}$.
In this case, $\mu _{ij}^{ - 1}$ and ${\eta _{ij}}$ are given by
\begin{equation}
\mu _{ij}^{ - 1} \equiv {\delta _{ij}} - {B_i}{\gamma _j} - {E_i}{\Delta _j}, \label{Meta30a}
\end{equation}
and
\begin{equation}
{\eta _{ij}} \equiv {B_i}{\alpha _j} - \frac{1}{{{c^2}}}{E_i}{\beta _j} = - \frac{1}{{{c^2}}}{\xi _{ji}}, \label{Meta30b}
\end{equation}
where we have defined $\boldsymbol{\gamma} \equiv \frac{1}{{{C_1}}}\left( {{D_1}\,{\bf B} - {D_3} \,{\bf E}} \right)$ and
$\boldsymbol{\Delta} \equiv \frac{1}{{{C_1}}}\left( { - \frac{{{D_3}}}{c} \,{\bf B} + \frac{{{D_2}}}{{{c^2}}}\, {\bf E}} \right)$.
Incidentally, we would like to point out that, in a recent paper \cite{Casana}, the authors carry out a detailed study of bi-isotropic and bi-anisotropic material media in terms of a general class of constitutive tensors. In our case, we stress that it is the vacuum - subject to strong external electromagnetic fields - that acts as the material medium, with
permittivity and permeability tensors completely determined by
the electromagnetic background, according to the equations cast previously. In the particular cases of a purely electric or a purely magnetic external field, the constitutive tensors naturally arise as symmetric $3\times3$-matrices.
One can easily verify that for an external field ${\bf B}$ (${\bf E}=0$) we have ${\alpha _i} = \frac{{{D_3}}}{{c\,{C_1}}}{B_i}$, ${\beta _i} = \frac{{{D_2}}}{{{C_1}}}{B_i}$, ${\gamma _i} = \frac{{{D_1}}}{{{C_1}}}{B_i}$ and ${\Delta _i} = - \frac{{{D_3}}}{{c\,{C_1}}}{B_i} = - {\alpha _i}$. We thus find that,
\begin{equation}
{\varepsilon _{ij}} = {\delta _{ij}} + \frac{{{D_2}}}{{{C_1}}}{B_i}{B_j}, \label{Meta35a}\end{equation}
\begin{equation}
{\xi _{ij}} = - c\,\frac{{{D_3}}}{{{C_1}}}{B_i}{B_j}, \label{Meta35b}
\end{equation}
\begin{equation}
\mu _{ij}^{ - 1} = {\delta _{ij}} - \frac{{{D_1}}}{{{C_1}}}{B_i}{B_j}, \label{Meta35c}\end{equation}
and
\begin{equation}
{\eta _{ij}} = \frac{{{D_3}}}{{c\,{C_1}}}{B_i}{B_j} = - \frac{1}{{{c^2}}}{\xi _{ij}}.\label{Meta35d}
\end{equation}
By defining the matrix ${{\cal B}_{ij}} \equiv {B_i}{B_j}$, which is a symmetric one with eigenvalues $0,0$ and ${{\bf B}^2}$, the expressions for $d$ and $h$, in matrix notation, become
\begin{equation}
d = e + \frac{{{D_2}}}{{{C_1}}}{\cal B}\,e - c\,\frac{{{D_3}}}{{{C_1}}}{\cal B}\,b, \label{Meta40a}
\end{equation}
\begin{equation}
h = b - \frac{{{D_1}}}{{{C_1}}}{\cal B}\,b + \frac{{{D_3}}}{{c\,{C_1}}}{\cal B}\,e. \label{Meta40b}
\end{equation}
Whereas
\begin{equation}
\nabla \cdot {\bf d} = 0, \label{Meta40c}
\end{equation}
\begin{equation}
\nabla \times {\bf h} = \frac{1}{{{c^2}}}\frac{\partial }{{\partial t}}{\bf d}. \label{Meta40d}
\end{equation}
After the matrix ${\cal B}$ is diagonalized, we then have
\begin{equation}
{\partial _i}{d_i} = 0, \label{Meta45}
\end{equation}
\begin{equation}
{\varepsilon _{ijk}}{\partial _j}{h_k} = \frac{1}{{{c^2}}}\frac{\partial }{{\partial t}}{d_i}, \label{Meta50}
\end{equation}
where the matrix ${\cal B} = \left( {\begin{array}{*{20}{c}}
{{{\bf B}^2}}&0&0\\
0&0&0\\
0&0&0
\end{array}} \right)$. \\
An immediate consequence of this result is that
${d_x} = \left( {1 + \frac{{{D_2}}}{{{C_1}}}{{\bf B}^2}} \right){e_x} - c\,\frac{{{D_3}}}{{{C_1}}}{{\bf B}^2}\,{b_x}$, ${d_y} = {e_y}$ and ${d_z} = {e_z}$. We also find that ${h_x} = \left( {1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}} \right){b_x} + \frac{{{D_3}}}{{c\,{C_1}}}{{\bf B}^2}\, {e_x}$, ${h_y} = {b_y}$ and ${h_z} = {b_z}$.\\
This leads to the following expressions for ${\varepsilon _{xx}}$ and ${\mu _{xx}}$, that is,
\begin{equation}
{\varepsilon _{xx}} = 1 + \frac{{{D_2}}}{{{C_1}}}{{\bf B}^2}, \label{Meta55}
\end{equation}
and
\begin{equation}
\mu _{xx}^{ - 1} = 1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}. \label{Meta60}
\end{equation}
It should be further noted that, in principle, the positivity of the above expressions is not assured.
However, from equation (\ref{Meta60}) it is evident that
\begin{equation}
{\mu _{xx}} = \frac{1}{{1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}}},
\end{equation}
this then implies that ${\mu _{xx}}<0$ if $\frac{{{D_1}}}{{{C_1}}}{{\bf B}^2} > 1$. We see, therefore, a remarkable feature of nonlinearities of the electromagnetic field.
The preceding considerations clearly show that the nonlinearities induce the second combination mentioned above $(\varepsilon > 0, \mu <0)$, which corresponds to a ferrimagnetic material in the classification given in \cite{Caloz}.
So far our treatment is completely general. However, for the specific case of logarithmic electrodynamics, the permittivity and susceptibility induced by nonlinearities will be considered in the next Subsection.
\subsection{Some features of logarithmic electrodynamics}
We now proceed to explore other relevant aspects on nonlinearities. Let us commence our undertaking by considering logarithmic electrodynamics \cite{Logarithmic}:
\begin{equation}
{\cal L} = - \beta ^2 \ln \left[ {1 - \frac{{{\cal F}}}{{\beta ^2 }} - \frac{{{\cal G}^2 }}{{2\beta ^4 }}} \right],
\label{log05}
\end{equation}
recalling again that ${\cal F} = \frac{1}{2}\left( {\frac{{{{\bf E}^2}}}{{{c^2}}} - {{\bf B}^2}} \right)$ and ${\cal G} = \frac{{\bf E}}{c} \cdot {\bf B}$. We further note that the parameter $\beta$ measures the nonlinearity of the theory and in the limit $\beta \to \infty$ the Lagrangian (\ref{log05}) reduces to the Maxwell theory.
As already stated, in this Subsection we will be mainly interested in the dispersion relations for the electrodynamics under consideration. The first step in this direction is to consider a generic Lagrangian density ${\cal L}={\cal L}({\cal F},{\cal G})$, in the presence of a constant background with both $({\bf E},{\bf B})$. As before, after splitting $F^{\mu\nu}$ in the sum of a classical background $F^{\mu\nu}_{B}$, and a small fluctuation, $f^{\mu\nu}$, the corresponding linearized equations of motion read
\begin{equation}
\nabla \cdot {\bf d} = {\partial _i}{d_i} = {\varepsilon _{ij}}{\partial _i}{e_j} + {\xi _{ij}}{\partial _i}{b_j} = 0, \nonumber\\
\end{equation}
\begin{equation}
\nabla \times {\bf e} = - \frac{\partial }{{\partial t}}{\bf b}\,\,\, \Rightarrow {\varepsilon _{ijk}}{\partial _j}{e_k} = - \frac{\partial }{{\partial t}}{b_i}, \nonumber\\
\end{equation}
\begin{equation}
\nabla \cdot {\bf b} = 0,\nonumber\\
\end{equation}
\begin{eqnarray}
\nabla \times {\bf h} = \frac{1}{{{c^2}}}\frac{\partial }{{\partial t}}{\bf e} \nonumber\\
\Rightarrow{\varepsilon _{ijk}}\mu _{kl}^{ - 1}{\partial _j}{b_l} + {\varepsilon _{ijk}}{\eta _{kl}}{\partial _j}{e_l} &=& \frac{1}{{{c^2}}}{\varepsilon _{ij}}{\partial _t}{e_j} + \frac{1}{{{c^2}}}{\xi _{ij}}{\partial _t}{b_j}. \nonumber\\
\label{log10}
\end{eqnarray}
Throughout, ${\bf e}$ and ${\bf b}$ are the electric and magnetic fields arising from the fluctuation $f^{\mu\nu}$.
By considering the plane waves
\begin{equation}
e_{i} = e_{0i}\,{e^{i\left( {{\bf k} \cdot {\bf x} - wt} \right)}}, \,\,\,\,\, b_{i} = b_{0i}\,{e^{i\left( {{\bf k} \cdot {\bf x} - wt} \right)}}, \label{log15}
\end{equation}
from the Eqs. (\ref{log10}) it follows that
\begin{equation}
{M_{in}}\left( {w,{\bf k};{{\bf E}_B},{{\bf B}_B}} \right){e_{0n}} = 0, \label{log20}
\end{equation}
where
\begin{eqnarray}
{M_{in}} &=& \frac{{{w^2}}}{{{c^2}}}{\varepsilon _{in}} + \frac{w}{{{c^2}}}{\xi _{ij}}{\varepsilon _{jmn}}{k_m} + {\varepsilon _{ijk}}{\varepsilon _{lmn}}\mu _{kl}^{ - 1}{k_j}{k_m} \nonumber\\
&+& w{\varepsilon _{ijk}}{\eta _{kn}}{k_j},\label{log25}
\end{eqnarray}
in which we have used
\begin{equation}
{\varepsilon _{ij}} = {\delta _{ij}} + {\alpha _i}{E_j} + {\beta _i}{B_j}, \label{log30a}
\end{equation}
\begin{equation}
{\xi _{ij}} = - {c^2}{\alpha _i}{B_j} + {\beta _i}{E_j}, \label{log30b}
\end{equation}
\begin{equation}
\mu _{ij}^{ - 1} = {\delta _{ij}} - {B_i}{\gamma _j} - {E_i}{\Delta _j}, \label{log30c}
\end{equation}
\begin{equation}
{\eta _{ij}} = - \frac{1}{{{c^2}}}{\xi _{ji}} = {\alpha _j}{B_i} - \frac{1}{{{c^2}}}{\beta _j}{E_i}. \label{log30d}
\end{equation}
In passing we recall that ${\alpha _i}$, ${\beta _i}$, ${\gamma _i}$ and ${\Delta _i}$ are given in terms of $C_{1}$, $D_{1}$, $D_{2}$ and $D_{3}$.
It is of interest to note that in the particular case of an external electric field ${\bf E}=0$, Eq. (\ref{log20}) takes the particularly simple form
\begin{equation}
\left( {\frac{{{w^2}}}{{{c^2}}}{\varepsilon _{in}} + {\varepsilon _{ijk}}{\varepsilon _{lmn}}\mu _{kl}^{ - 1}{k_j}{k_m}} \right){e_{0n}} = 0. \label{log35}
\end{equation}
Here we have used that $D_{3}=0$ (${{\cal G}_{Background}} = 0$), whereas ${\xi _{ij}} = 0$ and ${\eta _{ij}} = 0$. We accordingly express Eq. (\ref{log35}) in the form
\begin{eqnarray}
&\biggl[&{{\!\!\!\! \frac{{{w^2}}}{{{c^2}}}{\varepsilon _{in}} + \mu _{in}^{ - 1}{{\bf k}^2} + \left( {tr{\mu ^{ - 1}}} \right){k_i}{k_n} + {\delta _{in}}\left( {\mu _{jk}^{ - 1}{k_j}{k_k}} \right)}} \nonumber\\
&-&\!\! {{k_i}\mu _{nj}^{ - 1}{k_j} - \mu _{ij}^{ - 1}{k_j}{k_n} - {\delta _{in}}\left( {tr{\mu ^{ - 1}}} \right){{\bf k}^2}} \biggr]{e_{0n}} = 0, \label{log40}
\end{eqnarray}
where
\begin{equation}
{\varepsilon _{in}} = {\delta _{in}} + \frac{{{D_2}}}{{{C_1}}}{B_i}{B_n} \equiv {\varepsilon _{ni}}, \label{log45a}
\end{equation}
\begin{equation}
\mu _{jk}^{ - 1} = {\delta _{jk}} - \frac{{{D_1}}}{{{C_1}}}{B_j}{B_k} \equiv \mu _{kj}^{ - 1}, \label{log45b}
\end{equation}
\begin{equation}
tr{\mu ^{ - 1}} = 3 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}. \label{log45c}
\end{equation}
Making use of these relations, we find that Eq. (\ref{log40}) reduces to
\begin{eqnarray}
\left[ {\frac{{{w^2}}}{{{c^2}}}{\mu _{mi}}{\varepsilon _{in}} + {{\bf k}^2}{\delta _{mn}} - {k_m}{k_n} + \left( {tr{\mu ^{ - 1}}} \right){k_i}{\mu _{im}}k_{n}} \right. \nonumber\\
\left. { - \left( {tr{\mu ^{ - 1}}} \right){{\bf k}^2}{\mu _{mn}} + \left( {\mu _{ij}^{ - 1}{k_i}{k_j}} \right){\mu _{mn}} - {k_i}{k_j}\mu _{jn}^{ - 1}} \right]{e_{0n}} = 0, \nonumber\\
\label{log50}
\end{eqnarray}
where ${\mu _{ij}} = {{\delta _{ij}} + \frac{{{{{D_1}} \mathord{\left/
{\vphantom {{{D_1}} {{C_1}}}} \right.
\kern-\nulldelimiterspace} {{C_1}}}}}{{1 - {{{D_1}} \mathord{\left/
{\vphantom {{{D_1}} {{C_1}{B^2}}}} \right.
\kern-\nulldelimiterspace} {{C_1}{{\bf B}^2}}}}}{B_i}{B_j}}$. \\
It is also important to observe that in the configuration space we have $w = i{\partial _t}$, ${k_i} = - i{\partial _i}$ and ${{\bf k}^2} = - {\nabla ^2}$. Hence, we readily verify that Eq. (\ref{log50}) can be brought to the form
\begin{eqnarray}
\left[ {-\frac{1}{{{c^2}}}\partial _t^2{\mu _{mi}}{\varepsilon _{in}} - {\delta _{mn}}{\nabla ^2} + {\partial _m}{\partial _n} - \left( {tr{\mu ^{ - 1}}} \right){\mu _{mi}}{\partial _i}{\partial _n}} \right. + \nonumber\\
\left. { \left( {tr{\mu ^{ - 1}}} \right){\mu _{mn}}{\nabla ^2} - {\mu _{mn}}\mu _{ij}^{ - 1}{\partial _i}{\partial _j} + {\mu _{mi}}\mu _{nj}^{ - 1}{\partial _i}{\partial _j}} \right]{e_{0n}} = 0. \nonumber\\
\label{log55}
\end{eqnarray}
Thus, finally we end up with
\begin{eqnarray}
\left[ {{\mu _{mi}}{\varepsilon _{in}}\frac{1}{{{c^2}}}\partial _t^2 - \left( {{\mu _{mn}}tr{\mu ^{ - 1}} - {\delta _{mn}}} \right){\nabla ^2} - {\partial _m}{\partial _n}} \right. + \nonumber\\
\left. { \left( {tr{\mu ^{ - 1}}} \right){\mu _{mi}}{\partial _i}{\partial _n} + \left( {{\mu _{mn}}\mu _{ij}^{ - 1} - {\mu _{mi}}\mu _{nj}^{ - 1}} \right){\partial _i}{\partial _j}} \right]{e_{0n}} = 0. \nonumber\\
\label{log60}
\end{eqnarray}
Before proceeding our analysis of the dispersion relation, we call attention to the fact that
\begin{eqnarray}
{C_1} = \frac{{\partial {\cal L}}}{{\partial {\cal F}}} = \frac{1}{{1 - \frac{{\cal F}}{{{\beta ^2}}} - \frac{{{{\cal G}^2}}}{{2{\beta ^4}}}}}, \nonumber\\
{C_2} = \frac{{\partial {\cal L}}}{{\partial {\cal G}}} = \frac{1}{{{\beta ^2}}}{\cal G}{C_1}, \nonumber\\
{D_1} = \frac{{{\partial ^2}{\cal L}}}{{\partial {{\cal F}^2}}} = \frac{1}{{{\beta ^2}}}C_1^2, \nonumber\\
{D_2} = \frac{{{\partial ^2}{\cal L}}}{{\partial {{\cal G}^2}}} = \frac{1}{{{\beta ^2}}}{C_1} + \frac{1}{{{\beta ^6}}}{{\cal G}^2}C_1^2, \nonumber\\
{D_3} = \frac{{{\partial ^2}{\cal L}}}{{\partial {\cal F}\partial {\cal G}}} = \frac{1}{{{\beta ^4}}}{\cal G}C_1^2. \label{log65}
\end{eqnarray}
Restricting our considerations to the ${\bf E}=0$ case, we have ${\cal F} = - \frac{1}{2}{{\bf B}^2}$ and ${\cal G}=0$. We thus find ${C_1} = \frac{1}{{1 + \frac{{{{\bf B}^2}}}{{2{\beta ^2}}}}}$, $C_{2}=0$, $\frac{{{D_1}}}{{{C_1}}} = \frac{1}{{{\beta ^2} + \frac{{{{\bf B}^2}}}{2}}}$, $\frac{{{D_2}}}{{{C_1}}} = \frac{1}{{{\beta ^2}}}$ and $D_{3}=0$. Making use of the foregoing results one encounters that, ${\varepsilon _{ij}} = {\delta _{ij}} + \frac{{{D_2}}}{{{C_1}}}{B_i}{B_j}$, have two eigenvalues 1 and $1 + \frac{{{D_2}}}{{{C_1}}}{{\bf B}^2}$. In fact, for logarithmic electrodynamics this eigenvalue reduces to $1 + \frac{{{{\bf B}^2}}}{{{\beta ^2}}}$. Similarly, from ${\mu _{ij}} = {\delta _{ij}} + \frac{{\frac{{{D_1}}}{{{C_1}}}}}{{1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}}}{B_i}{B_j}$, we again have two eigenvalues $1$ and $\frac{1}{{1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}}}$. For logarithmic electrodynamics the previous eigenvalue becomes $\frac{{{\beta ^2} + \frac{{{{\bf B}^2}}}{2}}}{{{\beta ^2} - \frac{{{{\bf B}^2}}}{2}}}$.
Next, by making use of ${\varepsilon _{in}}$, $\mu _{in}^{ - 1}$ and $tr{\mu ^{ - 1}}$ in the dispersion matrix ($M_{in}$), we can write the corresponding matrix as:
\begin{eqnarray}
{M_{in}} &=& \left( {\frac{{{w^2}}}{{{c^2}}} - {{\bf k}^2}} \right){\delta _{in}} + \frac{{{D_1}}}{{{C_1}}}\left( {{{\bf B}^2}{{\bf k}^2} - {{\left( {{\bf B} \cdot {\bf k}} \right)}^2}} \right){\delta _{in}} \nonumber\\
&+& \left( {\frac{{{D_2}}}{{{C_1}}}\frac{{{w^2}}}{{{c^2}}} - \frac{{{D_1}}}{{{C_1}}}{{\bf k}^2}} \right){B_i}{B_n} + \left( {1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}} \right){k_i}{k_n} \nonumber\\
&+& \frac{{{D_1}}}{{{C_1}}}\left( {{\bf B} \cdot {\bf k}} \right)\left( {{k_i}{B_n} + {k_n}{B_i}} \right). \label{log70}
\end{eqnarray}
We are now in position to examine the condition, $\det M = 0$, in order to obtain the dispersion relations. We also recall that the index of refraction is given by $n \equiv \frac{{|{\bf k}|c}}{w}$, and after some manipulations, it follows that
\begin{eqnarray}
M_{in}&=&\left[ {\frac{1}{{{n^2}}} - 1 + \frac{{{D_1}}}{{{C_1}}}\left( {{{\bf B}^2} - {{\left( {{\bf B} \cdot \hat {\bf k}} \right)}^2}} \right)} \right]{\delta _{in}} \nonumber\\
&+& \left( {\frac{{{D_2}}}{{{C_1}}}\frac{1}{{{n^2}}} - \frac{{{D_1}}}{{{C_1}}}} \right){B_i}{B_n} \nonumber\\
&+&\left( {1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}} \right){{\hat {\bf k}}_i}{{\hat {\bf k}}_n} + \frac{{{D_1}}}{{{C_1}}}\left( {{\bf B} \cdot \hat {\bf k}} \right)\left( {{{\hat k}_i}{B_n} + {{\hat k}_n}{B_i}} \right), \nonumber\\
\label{log75}
\end{eqnarray}
where $\hat {\bf k} \equiv \frac{{\bf k}}{{|{\bf k}|}}$.
This last expression clearly shows that the $M$-matrix does not depend on $|{\bf k}| = {{2\pi } \mathord{\left/
{\vphantom {{2\pi } \lambda }} \right.\kern-\nulldelimiterspace} \lambda }$. Then the refractive index, $n$, arising from the condition, $det M=0$, does not depend on $\lambda$ but of the relative direction between the propagation vector $\hat {\bf k}$ and the external field. In this manner, we obtain an effective refractive index $n = n\left( {{\bf B},\hat {\bf k}} \right)$.
To further elaborate on the comparative features of the index of refraction, we shall examine two different situations. First, we consider ${n_ \bot }$ if $\hat {\bf k} \,\bot \,{\bf B}$. In this case, the condition, $det M=0$, reads
\begin{eqnarray}
\det \left[ {\left( {\frac{1}{{{n^2}}} - 1 + \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}} \right)} \right.{\delta _{in}} + \left( {1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}} \right){{\hat k}_i}{{\hat k}_n}
\nonumber\\
\left. {\left( {\frac{{{D_2}}}{{{C_1}}}\frac{1}{{{n^2}}} - \frac{{{D_1}}}{{{C_1}}}} \right){B_i}{B_n}} \right] = 0. \label{log80}
\end{eqnarray}
From this equation it is clear that the determinant has the form $\det \, \left( {a{\delta _{ij}} + b{u_i}{u_j} + c{v_i}{v_j}} \right)$, whose solution is given by $a\left[ {\left( {a + b{u^2}} \right)\left( {a + c{v^2}} \right) - bc{{\left( {u \cdot v} \right)}^2}} \right]$. It is a simple matter to verify that the condition, $det M=0$, becomes
\begin{equation}
\frac{1}{{{n^2}}}\left( {\frac{1}{{{n^2}}} - 1 + \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}} \right)\left[ {\frac{1}{{{n^2}}}\left( {1 + \frac{{{D_2}}}{{{C_1}}}{{\bf B}^2}} \right) - 1} \right] = 0. \label{log85}
\end{equation}
Thus, we finally obtain two modes associated to the direction of propagation $\hat {\bf k}$, that is,
\begin{equation}
n_ \bot ^2 = \frac{1}{{1 - \frac{{{D_1}}}{{{C_1}}}{{\bf B}^2}}}, \label{log90}
\end{equation}
and
\begin{equation}
n_ \bot ^2 = 1 + \frac{{{D_2}}}{{{C_1}}}{{\bf B}^2}. \label{log90}
\end{equation}
For logarithmic electrodynamics we have $\frac{{{D_1}}}{{{C_1}}} = \frac{{{C_1}}}{{{\beta ^2}}} = \frac{1}{{{\beta ^2} + \frac{{{{\bf B}^2}}}{2}}}$ and $\frac{{{D_2}}}{{{C_1}}} = - \frac{1}{{{\beta ^2}}}$. Hence we see that the two modes take the form
\begin{equation}
n_ \bot ^2 = \frac{{{\beta ^2} + \frac{{{{\bf B}^2}}}{2}}}{{{\beta ^2} - \frac{{{{\bf B}^2}}}{2}}},\label{log95}
\end{equation}
and
\begin{equation}
n_ \bot ^2 = 1 + \frac{{{{\bf B}^2}}}{{{\beta ^2}}}. \label{log100}
\end{equation}
Second, we consider ${n_\parallel }$ if ${\hat {\bf k}\,\parallel \, {\bf B}}$. By using ${\bf B} = \xi |{\bf B}|\hat {\bf k}$ and ${\bf B} \cdot \hat {\bf k} = \xi |{\bf B}|$, where $\xi = \pm 1$ stands parallel or anti-parallel to the propagation direction. After some manipulations the condition $det M=0$ becomes
\begin{equation}
\frac{1}{{{n^2}}}{\left( {\frac{1}{{{n^2}}} - 1} \right)^2}\left( {1 + \frac{{{D_2}}}{{{C_1}}}{{\bf B}^2}} \right) = 0.\label{log105}
\end{equation}
In this case, the corresponding mode associated to the direction of propagation $\hat {\bf k}$ becomes ${n_\parallel } = 1$.
In summary then, we easily verify that the previous electromagnetic vacuum acts like a birefringent medium with two indices of refraction determined by the relative direction between the propagation vector $\hat {\bf k}$ and the external field. More recently, this has also helped us to gain insights into the peculiarities about vacuum nonlinearities such as calculating the bending of light \cite{inverse}.
\section{Electromagnetic radiation}
As already mentioned, our immediate objective is to compute the electromagnetic radiation produced by a moving charged particle interacting with a medium characterized by nonlinearities of the electromagnetic field.
With this in view, the starting point are the Maxwell equations for a moving charged particle in a medium characterized by logarithmic electrodynamics:
\begin{eqnarray}
\nabla \cdot {\bf e} = \frac{{4\pi }}{\varepsilon }{\rho _{ext}}, \nonumber\\
\nabla \cdot {\bf b} = 0, \nonumber\\
\nabla \times {\bf e} = - \frac{1}{c} \frac{{\partial {\bf b}}}{{\partial t}}, \nonumber\\
\nabla \times {\bf b} = \frac{{\varepsilon \mu }}{c}\frac{{\partial {\bf e}}}{{\partial t}} + \frac{{4\pi \mu }}{c}{{\bf j}_{ext}}. \label{Red65}
\end{eqnarray}
where ${\rho _{ext}}$ and ${{\bf j}_{ext}}$ denote the external charge and current densities. Whereas ${\bf d} = \varepsilon {\bf e}$ and ${\bf b} = \mu {\bf h}$. Here we have simplified our notation by setting ${\bf E}_{p}={\bf e}$ and ${\bf B}_{p}={\bf b}$.
It is straightforward to see that the foregoing equations can be written alternatively in the form
\begin{equation}
{\nabla ^2}{\bf b} - \frac{{\varepsilon \mu }}{{{c^2}}}\frac{{{\partial ^2}{\bf b}}}{{\partial {t^2}}} = - \frac{{4\pi \mu }}{c}\nabla \times {{\bf j}_{ext}}, \label{Red70}
\end{equation}
and
\begin{equation}
{\nabla ^2}{\bf e} - \frac{{\varepsilon \mu }}{{{c^2}}}\frac{{{\partial ^2}{\bf e}}}{{\partial {t^2}}} = \frac{{4\pi \mu }}{{{c^2}}}\frac{{\partial {{\bf j}_{ext}}}}{{\partial t}} + \frac{{4\pi }}{\varepsilon }\nabla {\rho _{ext}}, \label{Red75}
\end{equation}
where the external charge and current densities are given by: ${\rho _{ext}}\left( {t,{\bf x}} \right) = Q\delta \left( x \right)\delta \left( y \right)\delta \left( {z - vt} \right)$ and ${\bf j}\left( {t,{\bf x}} \right) = Qv\delta \left( x \right)\delta \left( y \right)\delta \left( {z - vt} \right){\hat {\bf e}_z}$.
In passing we note that, for simplicity, we are considering the $z$ axis as the direction of the moving charged particle.
Next, in order to solve equations (\ref{Red70}) and (\ref{Red75}), we shall begin by performing a Fourier transform to momentum space via
\begin{equation}
f(t,{\bf x}) = \int {\frac{{dw{d^3}{\bf k}}}{{{{\left( {2\pi } \right)}^4}}}} {e^{ - iwt + {\bf k} \cdot {\bf x}}}f\left( {w,{\bf k}} \right), \label{Red80}
\end{equation}
where $f$ stands for the electric and magnetic fields. Then, the corresponding electric and magnetic fields read:
\begin{equation}
{\bf b}\left( {w,{\bf k}} \right) = - \,i\,\frac{{4\pi \mu }}{c}\frac{{{\bf k} \times {{\bf j}_{ext}}\left( {w,{\bf k}} \right)}}{\cal O}, \label{Red85}
\end{equation}
and
\begin{equation}
{\bf e}\left( {w,{\bf k}} \right) = i\,\frac{{4\pi }}{\varepsilon }\frac{{{\bf k}\,{\rho _{ext}}\left( {w,{\bf k}} \right)}}{\cal O} - i\frac{{4\pi \mu w}}{{{c^2}}}\frac{{{{\bf j}_{ext}}\left( {w,{\bf k}} \right)}}{\cal O},\label{Red90}
\end{equation}
where
\begin{equation}
{\cal O} = \frac{{{w^2}}}{{{c^{\prime 2}}}} - {{\bf k}^2}, \ \ \
\frac{1}{{{c^{\prime 2}}}} \equiv \frac{{\varepsilon \mu }}{{{c^2}}}. \label{Red91}
\end{equation}
In the same way, the external charge and current densities, in the Fourier space, take the form: $\rho_{ext} \left( {w,{\bf k}} \right) = 2\pi Q\delta \left( {w - {k_z}v} \right)$ and ${{\bf j}_{ext}}\left( {w,{\bf k}} \right) = 2\pi Qv\delta \left( {w - {k_z}v} \right){\hat {\bf e}_z}$.
From the above we can proceed to obtain ${\bf b}\left( {w,{\bf x}} \right)$ and ${\bf e}\left( {w,{\bf x}} \right)$. It is clear now that ${\bf b}\left( {w,{\bf x}} \right)$ is given by
\begin{equation}
{\bf b}\left( {w,{\bf x}} \right) = \int {\frac{{{d^3}{\bf k}}}{{{{\left( {2\pi } \right)}^3}}}} \; {e^{i{\bf k} \cdot {\bf x}}}\;{\bf b}\left( {w,{\bf k}} \right). \label{Red95}
\end{equation}
We may now take advantage of the axial symmetry of the problem under consideration. If we take cylindrical coordinates, equation (\ref{Red95}) becomes
\begin{eqnarray}
{\bf b}\left( {w,{\bf x}} \right) &=& - \frac{{i\mu Qv}}{{\pi c}}{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}}\int_0^\infty {d{k_T}{k_T}} \nonumber\\
&\times&\int_0^{2\pi } {d\alpha } \frac{{{e^{i{k_T}{x_T}\cos \alpha }}}}{\left. {\cal O} \right|_{{k_z} = {w \mathord{\left/
{\vphantom {w v}} \right.
\kern-\nulldelimiterspace} v}}} \left( {{k_T}\sin \alpha\, \pmb{\hat \rho} - {k_T}\cos \alpha\, \pmb{\hat \phi} } \right). \nonumber\\
\label{Red100}
\end{eqnarray}
In passing we recall that $\int_0^{2\pi } {d\theta } {e^{ix\cos \theta }}\sin \theta = 0$ and $\int_0^{2\pi } {d\theta } {e^{ix\cos \theta }}\cos \theta = 2\pi i{J_1}\left( x \right)$, which implies
\begin{equation}
{\bf b}\left( {w,{\bf x}} \right) = - \frac{{2\mu Qv}}{c}{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}} \int_0^\infty {d{k_T}\;k_T^2} \;\frac{{{J_1}\left( {{k_T}{x_T}} \right)}}{{{{\left. {\cal O} \right|}_{{k_z} = {w \mathord{\left/
{\vphantom {w v}} \right.
\kern-\nulldelimiterspace} v}}}}}\;\pmb{\hat \phi}, \label{Red105}
\end{equation}
where ${{J_1}\left( {{k_T}{x_T}} \right)}$ is a Bessel function of the first kind.
In this case, ${\left. {\cal O} \right|_{{k_z} = {w \mathord{\left/
{\vphantom {w v}} \right.
\kern-\nulldelimiterspace} v}}} = {w^2}\left( {\frac{1}{{{c^{\prime 2}}}} - \frac{1}{{{v^2}}}} \right) - {\bf k}_T^2$. From this last expression it follows that
\begin{equation}
{\bf b}\left( {w,{\bf x}} \right) = \frac{{2\mu Qv}}{c}{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}} \int_0^\infty {d{k_T}}\; k_T^2\;\frac{{{J_1}\left( {{k_T}{x_T}} \right)}}{{\left( {k_T^2 + {\sigma}^{2}} \right)}}\; \pmb {\hat \phi}, \label{Red110}
\end{equation}
where ${\sigma ^2} = {w^2}\left( {\frac{1}{{{v^2}}} - \frac{1}{{{c^{\prime 2}}}}} \right)$.
We next observe that the previous expression can be brought to the form
\begin{eqnarray}
{\bf b}\left( {w,{\bf x}} \right) &=& \frac{{2\mu Qv}}{c}{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}}\int_0^\infty {dy\,{e^{-y{\sigma}^{2}}}} \nonumber\\
&\times&\int_0^\infty {d{k_T}}\; k_T^2\;{e^{ - yk_T^2}}{J_1}\left( {{k_T}{x_T}} \right)\pmb {\hat \phi}.\label{Red115}
\end{eqnarray}
From this last expression it follows that
\begin{equation}
{\bf b}\left( {w,{\bf x}} \right) = \! \frac{{2\mu Qv}}{c}{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}}\, \frac{{{x_T}}}{4}\int_0^\infty {dy\frac{1}{{{y^2}}}{e^{-y {\sigma}^{2} - {{x_T^2} \mathord{\left/
{\vphantom {{x_T^2} {4y}}} \right.
\kern-\nulldelimiterspace} {4y}}}}} \,\pmb {\hat \phi}, \label{Red120}
\end{equation}
or, in terms of the modified Bessel function,
equation (\ref{Red120}), becomes
\begin{equation}
{\bf b}\left( {w,{\bf x}} \right) = \frac{{2\mu Qv}}{c}\,{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}}\,\sigma {K_1}\left( {\sigma {x_T}} \right)\,\pmb {\hat \phi}, \label{Red125}
\end{equation}
where, in cylindrical coordinates, ${x_T} = \rho$.
Now we come to the calculation of the electric field. From the expression (\ref{Red90}), we find that the electric field may be written in the form
\begin{eqnarray}
{\bf e}\left( {w,{\bf x}} \right) &=& i\,8 {\pi}^{2} Q\int {\frac{{{d^3}k}}{{{{\left( {2\pi } \right)}^3}}}} \frac{w}{{v\;{\cal O}}}\;\delta \left( {{k_z} - {\raise0.7ex\hbox{$w$} \!\mathord{\left/
{\vphantom {w v}}\right.\kern-\nulldelimiterspace}
\!\lower0.7ex\hbox{$v$}}} \right){e^{i{\bf k} \cdot {\bf x}}} \nonumber\\
&\times&\left\{ {\frac{{{k_x}}}{{\epsilon }}\,\pmb{{\hat e}_x} + \frac{{{k_y}}}{{\epsilon }}\,\pmb{{\hat e}_y} + \left( {\frac{{{k_z}}}{{\epsilon }} - \frac{\mu\,w\,v}{c^{2}}} \right)\pmb{{\hat e}_z}} \right\}. \nonumber\\
\label{Red130}
\end{eqnarray}
In the same way as was done for the magnetic field, we then get
\begin{eqnarray}
{e_\rho }\left( {w,{\bf x}} \right) &=& \frac{{i\,Q}}{{v{{{\pi }}}}} \int_0^\infty {d{k_T}\,{k_T}}\,{e^{iw{z \mathord{\left/
{\vphantom {z v}} \right.
\kern-\nulldelimiterspace} v}}}\frac{1}{{{{\left. {\cal O} \right|}_{{k_z} = {w \mathord{\left/
{\vphantom {w v}} \right.
\kern-\nulldelimiterspace} v}}}}} \nonumber\\
&\times&\left\{ {\frac{{2\pi i{k_T}}}{{\epsilon }}{J_1}\left( {{k_T}{x_T}} \right)} \right\}, \label{Red135}
\end{eqnarray}
and
\begin{eqnarray}
{e_z}\left( {w,{\bf x}} \right) &=& \frac{{i\,Q}}{{v{{{\pi }}}}} \int_0^\infty {d{k_T}\,{k_T}}\,{e^{iw{z \mathord{\left/
{\vphantom {z v}} \right.
\kern-\nulldelimiterspace} v}}}\frac{1}{{{{\left. {\cal O} \right|}_{{k_z} = {w \mathord{\left/
{\vphantom {w v}} \right.
\kern-\nulldelimiterspace} v}}}}} \nonumber\\
&\times&\left\{ {2\pi\frac{w}{\epsilon}\left( {\frac{1}{{v}} - \frac{v}{c^{\prime2}}} \right){J_0}\left( {{k_T}{x_T}} \right)} \right\}.
\label{Red140}
\end{eqnarray}
The integral occurring on the right-hand side of the previous expressions can be as before. We get accordingly
\begin{equation}
{e_\rho }\left( {w,{\bf x}} \right) = \frac{{2Q}}{{v\varepsilon }}{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}}\sigma {K_1}\left( {\sigma {x_T}} \right), \label{Red145}
\end{equation}
and
\begin{equation}
{e_z}\left( {w,{\bf x}} \right) = -\,i\,\frac{{2Q}}{{v\varepsilon }}{e^{i{{wz} \mathord{\left/
{\vphantom {{wz} v}} \right.
\kern-\nulldelimiterspace} v}}}w\left( {\frac{1}{v} - \frac{v}{{{c^{\prime 2}}}}} \right){K_0}\left( {\sigma {x_T}} \right). \label{Red150}
\end{equation}
We are now equipped to compute the corresponding radiated energy in the case
under consideration.
In order to accomplish this purpose, let us start by observing that the density of power carried out by the radiation fields across the surface bounding the volume $V$ is given by the real part of the Poynting vector (time averaged value)
\begin{equation}
{\bf S} =\frac{c}{{2\pi }}{\mathop{\rm Re}\nolimits} \left( {{\bf e} \times {\bf {h^ * }}} \right). \label{Red155}
\end{equation}
We further recall that we will calculate the power radiated through the surface \cite{Das}, that is,
\begin{equation}
{\cal E} = \int_{ - \infty }^\infty {dt} \int\limits_S {d{\bf a} \cdot {\bf S}}. \label{Red160}
\end{equation}
It is worth emphasizing that in our case we shall consider a cylinder as the integration surface. Also, it may be mentioned that in order to get a meaningful expression we shall use a cylinder infinitesimally small \cite{PRA}.
Consequently, the power radiated per unit length through the surface
then reads
\begin{eqnarray}
{\cal E} &=& \frac{c}{{2\pi }}{\mathop{\rm Re}\nolimits} \int_0^\infty {dw} \left\{ {2\pi {\rho _0}{{\left. {{S_\rho }} \right|}_{\rho = {\rho _0}}}} \right\} \nonumber\\
&+& \frac{c}{{2\pi }}{\mathop{\rm Re}\nolimits} \int_0^\infty {dw} \left\{ {\frac{\partial }{{{\partial _z}}}\int_0^{{\rho _0}} {\int_0^{2\pi } {{S_z}} \,\rho\, d\rho\, d\phi } } \right\}, \label{Red165}
\end{eqnarray}
where ${S_\rho } = - {e_z}h_\phi ^ *$, ${S_z} = {e_\rho }h_\phi ^ *$ and ${\rho _0} \to 0$. Let us also recall here that, in our case, the $\phi$-component of the Poynting vector $\left( {{S_\phi }} \right)$ vanishes.
According to equations (\ref{Red125}), (\ref{Red145}) and (\ref{Red150}), the expression for the power radiated per unit length (\ref{Red165}) takes the form
\begin{equation}
{\cal E} = - \pi\, \frac{{{Q^2}v}}{{{c^2}}}\,\frac{{{n^2}}}{\varepsilon }\int_0^\infty {dww\left( {1 - \frac{{{c^2}}}{{{n^2}{v^2}}}} \right)}. \label{Red170}
\end{equation}
One immediately sees that this expression is similar to that encountered in the Cherenkov radiation theory \cite{Das}. This last expression clearly shows the role played by vacuum electromagnetic nonlinearities in triggering the radiated energy. We also point out that in equation (\ref{Red170}) we have used the asymptotic behavior of the Bessel (${K_\nu }\left( x \right) \to \frac{\pi }{{\sqrt {2x} }}{e^{ - x}}$), since we are describing outgoing radiation.
In connection with this last expression (\ref{Red170}) a few comments are in order. First, it should be recalled that ${\cal E}$ represents the rate of energy lost due to radiation along the trajectory of the charged particle, $-\frac{{dE}}{{dt}}$, where $E$ is the energy of the charged particle. Second, we also recall that the $w$ integration has physical meaning only over the range where $n > {c \mathord{\left/
{\vphantom {c v}} \right.\kern-\nulldelimiterspace} v}$. Third, we further note that the vacuum of the electrodynamics studied in this work describes a non-dispersive ''medium'', which is verified because $\varepsilon$ and $\mu$ are constant. In other words, the velocity of electromagnetic waves in this ''medium'' does not depend on the frequency of the waves.\\
\section{Final remarks}
In summary, we have studied some observational signatures of nonlinearities of the electromagnetic field. First, we shown the vital role played by nonlinearities in triggering a material with $(\varepsilon > 0, \mu <0)$, which corresponds to a ferrimagnetic material. Secondly, the permittivity and susceptibility induced by nonlinearities have been studied in order to obtain the refractive index via the dispersion relation for logarithmic electrodynamics. Finally, we have considered
the radiation produced by a moving charged particle (with uniform velocity) interacting in nonlinear medium. Let us also recall here that particles moving with a uniform velocity in vacuum do not lead to radiation. As already mentioned, one is lead to the interesting conclusion that the above radiation is driven by the medium through which the particle travels like the one that happens in the Cherenkov effect. Lastly, we will be focusing efforts to understand in more detail the physical consequences of electrodynamics with metamaterial features, including Cherenkov radiation, in the near future.
\section*{Acknowledgments}
One of us (P. G.) was partially supported by ANID PIA / APOYO AFB180002.
|
\section{Introduction}
Nonlinear effects introduce new beam dynamics phenomena that might open up the possibility of devising novel beam manipulation techniques. This is the case, for instance, when shaping the transverse beam distribution by means of adiabatic crossing of a stable nonlinear resonance. Such a process is at the heart of the so-called beam splitting that is used for the CERN Multiturn Extraction (MTE)~\cite{PhysRevLett.88.104801,PhysRevSTAB.7.024001,PhysRevSTAB.12.014001} and has been successfully implemented as a routine part of operation of the CERN Proton Synchrotron since several years~\cite{Borburgh:2137954,PhysRevAccelBeams.20.014001,PhysRevAccelBeams.20.061001}.
However, this is not the only nonlinear manipulation that can be devised. Indeed, under the inspiration of~\cite{PhysRevLett.110.094801}, it has been found that a controlled redistribution of the invariants can be achieved between the two transverse degrees of freedom~\cite{our_paper7}, provided that an appropriate two-dimensional nonlinear resonance is crossed. This opens novel options in terms of manipulation of the transverse beam emittances.
It is therefore natural to study whether nonlinear effects can be used efficiently to reduce the linear invariants of a transverse beam distribution, thus generating a cooling of the transverse beam emittance. The basis of this approach to beam cooling is the observation that nonlinear effects do not preserve the linear invariant, i.e.\ the linear action, or the so-called Courant-Snyder invariant. In this sense, they can be used to reduce the value of the linear invariant without violating the symplectic character of the Hamiltonian dynamics. Therefore, the comparison of the value of the linear invariant before and after the action of the nonlinear forces, i.e.\ when the dynamics is linear and expressed as a rotation around the origin of phase space, is a correct indicator of the reduction of the invariant for each individual particle, and hence of the whole beam distribution and of the corresponding emittance.
In this paper, the initial step towards the development of a nonlinear cooling of a particle distribution is discussed. We present a framework to cool an annular beam distribution, i.e.\ a distribution with nonzero density in an interval of radii $r_1 < r < r_2, \; r_1 > 0$ in the normalized phase space. It is well-known that annular beam distributions are generated as the result of applying a single transverse kick to a centered beam in the presence of decoherence. Hence, a potential application of annular beam cooling could be the restoration of the initial centered distribution after a transverse kick.
A general discussion of the systems that can be used to devise a cooling method for an annular beam distribution is presented in Section~\ref{sec:gen}, while the considered models are presented in Section~\ref{sec:theory} together with some results of the theory of adiabatic trapping applied to the models. In the same section, several cooling protocols are defined, and their performance analyzed in detail by means of extensive numerical simulations, whose results are presented and discussed in Section~\ref{sec:numres}. Finally, conclusions are drawn in Section~\ref{sec:conc}, with some mathematical details reported in the Appendices.
\section{General considerations on the model chosen} \label{sec:gen}
The general idea underlying the approach developed to achieve cooling of the emittance of an annular beam distribution is based on creating stable islands in phase space. This can be done by slowly modulating the parameters to vary the area of the islands to cause the particles to cross the separatrices. By then moving the resonance islands in phase space their action can be changed and eventually reduced.
To create stable phase-space islands, a resonance needs to be excited. The MTE experience suggests using a H\'enon-like map as a model, close to stable low-order resonances, e.g.\ $1/4$, $1/5$. If the initial annulus lies outside the chain of islands, then by changing the linear frequency one can act on the area of the central region and of the islands to trap particles in the center. This reduces the action by a quantity equivalent to the area of the islands divided by $2\pi$, according to the separatrix crossing theory.
A simple analysis of the scaling laws of the parameters of the islands, found in~\cite{Bazzani:262179}, suggests that this approach is feasible only for resonances of order $n=4$. However, to get the best cooling results one needs two parameters to control the position and the area of the resonance islands. Acting on the sextupolar coefficient is not efficient since this acts as a global-scale parameter~\cite{Bazzani:262179} and hence changes the dynamic aperture of the map. Therefore, an octupolar kick should be added to the sextupolar one to provide an additional free parameter. The estimates for the area of the islands and the central region can be derived using the results of~\cite{Bazzani:262179} and~\cite{PhysRevSTAB.12.024003}. However, the main drawback of this approach is the thick stochastic layer generated by the octupolar kick around the outer part of the separatrix of the four stable islands. This has the effect of inducing the loss of particles, which makes the method unreliable. These observations make the approach based on H\'enon-like maps unsuitable for the application under consideration.
Ongoing studies suggest that trapping into islands and transport from within the islands can also be efficiently achieved using AC-modulated magnets~\cite{our_paper4}. The most straightforward option consists of creating one island using an AC dipole in a $1:1$ resonance condition, i.e.\ with the oscillation frequency close to the linear tune of the system. It is worth recalling that AC dipoles have been widely studied in the field of accelerator physics, with essential applications to beam diagnostics (see e.g.\ ~\cite{peggstang,bei1999beam,PhysRevSTAB.5.054001,PhysRevSTAB.8.024401,PhysRevSTAB.11.084002,PhysRevSTAB.16.071002,PhysRevAccelBeams.19.054001,PhysRevAccelBeams.22.031002}, for an overview of AC dipole studies and applications). A cooling method for annular beams will therefore be devised based on Hamiltonian system modeling of the stable islands used to perform the adiabatic trapping, and subsequent transport, under the influence of an AC dipole.
\section{Theory} \label{sec:theory}
\subsection{The Hamiltonian model}
Horizontal betatronic motion in the presence of an AC dipole can be described by the Hamiltonian of a generic oscillator with a sextupolar nonlinearity and a dipolar time-dependent excitation~\cite{PhysRevSTAB.5.054001, peggstang, bei1999beam}, namely
\begin{equation}
\mathcal{H}(x,p_x,t) = \omega_0 \frac{x^2+p_x^2}{2} + \frac{k_3}{3}x^3 + \varepsilon x \cos\omega t \, ,
\label{eq:ham_xp}
\end{equation}
where
\begin{equation}
k_3 = \frac{1}{B_0 \rho} \, \frac{\partial^2 B_y}{\partial x^2} \, \ell \, ,
\end{equation}
and $B_0 \rho$ stands for the magnetic rigidity of the reference particle, $B_y$ is the transverse component of the magnetic field, and $\ell$ is the physical length of the magnetic element. We remark that the choice of the sextupolar nonlinearity is rather arbitrary, as other types of nonlinearity might be used, as long as they generate an amplitude-detuning term. On the other hand, from the standpoint of applications, the use of a sextupolar nonlinearity is very convenient as it is present in all magnetic lattices of circular accelerators.
Using the action angle coordinates $(\phi,J)$ of the unperturbed ($\varepsilon=0$) system and averaging on the fast Fourier components, the Hamiltonian reads
\begin{equation}
\mathcal{H}(\phi,J,t) = \omega_0 \, J + \frac{\Omega_2}{2} J^2 + \varepsilon\sqrt{2J}\cos\phi\cos\omega t \, ,
\end{equation}
where $\Omega_2 = g(\omega_0) \,k_3^2$ and $g(\omega_0)$ is a function of the linear frequency~\cite{Bazzani:262179}, representing an amplitude tuning term that can be derived using normal forms applied to the Hamiltonian~\eqref{eq:ham_xp}. We recall that $J(x,p_x)$ is an adiabatic invariant of the unperturbed system if the frequency $\omega_0$ is slowly modulated.
If we change the coordinates to refer the system to a rotating reference frame with slow angle $\gamma = \phi-\omega t$, taking into account the generating function $F=J(\phi-\omega t)$ and its time derivative $\pdv{F}{t}=-\omega J$, the transformation gives
\begin{equation}
\mathcal{H}(\gamma,J,\psi) = (\omega_0-\omega) J + \frac{\Omega_2}{2} J^2 + \varepsilon\sqrt{2J}\cos(\gamma+\psi)\cos\psi \, ,
\end{equation}
where $\psi=\omega t$.
One can average the fast variable $\psi$, using
\begin{equation}
\frac{1}{2\pi}\int_0^{2\pi}\dd\psi\,\cos(\gamma+\psi)\cos\psi = \frac{1}{2}\cos\gamma \, ,
\end{equation}
yielding the new averaged Hamiltonian
\begin{equation}
\mathcal{H}(\gamma,J) = (\omega_0-\omega) J + \frac{\Omega_2}{2} J^2 + \frac{\varepsilon}{2}\sqrt{2J}\cos\gamma \, ,
\end{equation}
which, after a rescaling, can be written in the following form
\begin{equation}
\mathcal{H}(\gamma,J) = 4J^2 - 2\lambda J + \mu\sqrt{2J}\cos\gamma \, ,
\label{eq:hamavJ}
\end{equation}
where
\begin{equation}
\lambda = \frac{4}{\Omega_2}(\omega-\omega_0)\,, \qquad \mu = \frac{4\varepsilon}{\Omega_2} \, .
\label{eq:params}
\end{equation}
Equation~\eqref{eq:hamavJ} represents a well-known Hamiltonian~\cite{neish1975,Neishtadt2013} that can be conveniently written in the form
\begin{equation}
\mathcal{H}(X,Y) = (X^2+Y^2)^2 - \lambda(X^2+Y^2) + \mu X
\label{eq:hamavxy}
\end{equation}
using the Cartesian coordinates $X=\sqrt{2J}\cos\gamma$, $Y=\sqrt{2J}\sin\gamma$.
When $\lambda > (3/2)\mu^{2/3}$, a hyperbolic fixed point exists only for $Y=0$ and
\begin{equation}
X = x_\mathrm{c} = \frac{\sqrt{6\lambda}}{3} \cos(\frac{\pi}{6}+\alpha) \, ,
\label{eq:xc}
\end{equation}
where
\begin{equation}
\alpha = \frac{1}{3}\asin(\frac{3\sqrt{6}}{4}\frac{\mu}{\lambda^{3/2}}) \, .
\label{eq:alpha}
\end{equation}
\begin{figure}[htp]
\centering
\includegraphics[width=0.7\columnwidth]{phsp.pdf}
\caption{Phase-space portrait of the Hamiltonian~\eqref{eq:hamavxy} with parameters $\lambda=0.1$, $\mu=0.01$. The red line represents the separatrix.}
\label{fig:ham_phsp}
\end{figure}
The phase space portrait of the Hamiltonian~\eqref{eq:hamavxy} is shown in Fig.~\ref{fig:ham_phsp}, and it can be divided into three regions: the inner regions $G_1$ and $G_2$ (and $G_3=G_1 \cup G_2$) and the region outside them.
Let us compute the area $A_i$ of any region $G_i$. If $\mathcal{H}_c$ is the value of the Hamiltonian in $(X= x_\mathrm{c}, Y=0)$, the equation $\mathcal{H}(\gamma, J) = \mathcal{H}_c$ has the solution
\begin{equation}
J(\gamma) = \frac{\lambda-2x_\mathrm{c}^2}{2} - 2x_\mathrm{c}\sqrt{\lambda-2x_\mathrm{c}^2}\sin\gamma + 2x_\mathrm{c}^2\sin^2\gamma \, ,
\end{equation}
and $J(\gamma)=0$ for $\gamma=\gamma_0$ with
\begin{equation}
\gamma_0 = \asin\frac{\sqrt{\lambda-2x_\mathrm{c}^2}}{2x_\mathrm{c}} \, .
\end{equation}
The area of $G_1$ in polar coordinates is thus given by
\begin{equation}
A_1 = \int_{-\gamma_0}^{\pi - \gamma_0}\dd\gamma\, J(\gamma) = \frac{\pi\lambda}{2}-K_1 - K_2 \, ,
\end{equation}
while the area of $G_3$ is given by
\begin{equation}
A_3 = \int_{-\pi-\gamma_0}^{\gamma_0}\dd\gamma\, J(\gamma) = \frac{\pi\lambda}{2}+K_1 + K_2
\label{eq:A3}
\end{equation}
so that
\begin{equation}
A_2 = A_3 - A_1 = 2(K_1+K_2) \, ,
\end{equation}
where
\begin{align}
K_1 &= \lambda\asin(\frac{\sqrt{\lambda-2x_\mathrm{c}^2}}{2x_\mathrm{c}}) \, , \label{eq:K1} \\
K_2 &= \frac{3}{2}\sqrt{\lambda-2x_\mathrm{c}^2}\sqrt{6x_\mathrm{c}^2-\lambda} \, . \label{eq:K2}
\end{align}
Let us now consider a particle which lies in the outer region with an action $J_0 > A_3/(2\pi)$. The area enclosed by its orbit will be $A_0 = 2\pi J_0$. If we start a slow modulation of the parameters $\lambda=\lambda(t)$, $\mu=\mu(t)$, according to the theory of adiabatic separatrix crossing~\cite{neish1975,Arnold:937549}, at $t=t^*$, when the condition $A_3=A_0$ is met for $\lambda=\lambda^*$, $\mu=\mu^*$, the particle is captured into $G_1$ or $G_2$ as a random event. Defining
\begin{equation}
\xi_i = \frac{\dv*{A_i}{t}(\lambda^*,\mu^*)}{\dv*{A_3}{t}(\lambda^*,\mu^*)} \qquad i=1,\,2 \, ,
\end{equation}
the probability $P_i$ of trapping in $G_i, \, i=1,\,2 $ is given by
\begin{equation}
P_i =\begin{dcases}
\quad 1 &\quad\text{ if } \xi_i > 1\\
\quad\xi_i &\quad\text{ if } 0<\xi_i < 1\\
\quad 0 &\quad\text{ if } \xi_i < 0
\end{dcases}\, .
\end{equation}
After trapping, the resulting action $J$ is given by $A_i/(2\pi)$, where $A_i$ is computed when trapping occurs, namely for $\lambda=\lambda^*$ and $\mu=\mu^*$.
Given an initial distribution of particles, all of which have an initial action in the close neighborhood of $J_0$, the expectation value of their final action is
\begin{equation}
\av{J} = {\frac{P_1 A_1 + P_2 A_2}{2\pi}}\eval_{\lambda^*,\mu^*} \, ,
\end{equation}
and we have $\left \langle J \right \rangle \le J_0$, since $P_1+P_2=1$, $A_1+A_2=A_3=2\pi J_0$, and $A_i>0$, $P_i>0$. Hence, the final expected action is smaller than the initial one, i.e.\ the Courant-Snyder invariant of the particle has been reduced. For a distribution of particles with action $J_0$, this results in a cooling of the beam.
Furthermore, when trapping occurs at ($\lambda^*$, $\mu^*$) we have $A_3 = 2\pi J_0$ and using $A_3 = \pi\lambda^*/2 + K_1 + K_2 = 2\pi J_0$, we obtain the expression
\begin{equation}
K_1 + K_2 = \pi\qty(2J_0 - \frac{\lambda^*}{2}) \, .
\end{equation}
Substituting $K_1+K_2$ into the expressions for $A_1$ and $A_2$, one obtains
\begin{equation}
A_1(\lambda^*,\mu^*) = \pi(\lambda^*-2J_0) \qquad A_2(\lambda^*,\mu^*) = \pi(4J_0 - \lambda^*) \, .
\end{equation}
We note that the values of $A_1$ and $A_2$ at the crossing time do not depend on $\mu^*$.
We can then rewrite $\av{J}$ using $P_2=1-P_1$, which gives
\begin{equation}
\av{J} = 2J_0-\frac{\lambda^*}{2} + P_1(\lambda^* - 3 J_0)
\label{eq:finJ}
\end{equation}
having calculated $P_1$ at $\lambda=\lambda^*$, $\mu=\mu^*$.
\subsection{Cooling protocols}
We envisage three possible protocols to achieve beam cooling, since we can trap particles by varying only $\lambda(t)$, only $\mu(t)$, or both parameters. We will present the three possible processes in this order, referring to them as Protocol~A,~B and~C, respectively.
\subsubsection{Variation of \texorpdfstring{$\lambda$}{l} (Protocol~A)} \label{sec:theorA}
If we keep $\mu$ constant, $\dv*{A_i}{t} = \pdv*{A_i}{\lambda} \cdot \dv*{\lambda}{t}$, and the probabilities are thus given by
\begin{equation}
\xi_i = {\frac{\dv*{A_i}{\lambda}(\lambda^*)}{\dv*{A_3}{\lambda}(\lambda^*)}}\eval_{\lambda=\lambda^*} \qquad i=1,\,2 \, .
\end{equation}
Their expressions have been computed in~\cite{neish1975,Neishtadt2013} and read
\begin{equation}
\begin{aligned}
\pdv{A_1}{\lambda} &= \frac{\Theta}{2}\,,\qquad &\pdv{A_2}{\lambda} &= \pi - \Theta\,, \\
P_1 &= \frac{\Theta/2}{\pi-\Theta/2}\,, \qquad
&P_2 &= \frac{\pi-\Theta}{\pi-\Theta/2} \, ,
\end{aligned}
\label{eq:probB}
\end{equation}
where
\begin{equation}
\Theta = \acos(\frac{\lambda}{2x_\mathrm{c}^2}-2) \, .
\label{eq:theta}
\end{equation}
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\columnwidth]{theor_lambda.pdf}
\caption{Cooling ratio $\av{J} /J_0$ for trapping in $G_1$ and $G_2$ with the variation of $\lambda$ according to Eq.~\eqref{eq:finJ}, for three values of $\mu^*$.}
\label{fig:mugraph}
\end{figure}
Figure~\ref{fig:mugraph} shows $\av{J} /J_0$ as a function of $J_0$ for different values of $\mu^*$. We find that the minimum value of $\av{J} /J_0$ is independent of $\mu^*$ (the proof is given in Appendix~\ref{sec:app}).
A numerical computation of this minimum value gives $\av{J} /J_0$ = $\num{0.3957}$. Given $J_0$, we can always find a value $\mu$ that optimizes the cooling, with the final action reduced to $\approx 40\%$ of the initial value.
Strictly speaking, when $\varepsilon\neq 0$, as in the final state of this protocol, the emittance is not equal to the average value of the adiabatic invariant. The reason for this is that the emittance is computed assuming that the dynamics induces a rotation around the origin, whereas the adiabatic invariant is computed with respect to the fixed point around which the initial conditions actually evolve. In fact, when $\varepsilon\neq 0$, and especially when particles are trapped both in $G_1$ and in $G_2$, as in the final state of this protocol, they are not rotating around the origin. We observe also that if such a cooled beam were transferred to another accelerator, then its emittance would indeed be equal to the average action of the particle distribution. In this sense, the cooling ratio $\av{J} /J_0$ calculated from Eq.~\eqref{eq:finJ} is the lower bound to the actual ratio between the final and initial emittance values.
This situation could be solved or at least mitigated if it were possible to develop a protocol of adiabatic transport that, after the trapping phase, would preserve the actions of the particles while reducing $\mu$ to zero. However, one should consider that when trapping is achieved by means of a variation of $\lambda$ only, the cooling is not particularly efficient, since at best the cooling ratio is $\approx 60\%$. The methods that we are going to present in the following sections are, in theory, capable of achieving total cooling.
\subsubsection{Variation of \texorpdfstring{$\mu$}{m} and complete trapping in \texorpdfstring{$G_2$}{G2} (Protocol~B)} \label{sec:theorB}
For the protocol based on the variation of $\mu$, the area derivatives $(i=1,\,2)$ are given by
\begin{equation}
\dv{A_i}{\mu} = \dv{\alpha}{\mu}\dv{x_\mathrm{c}}{\alpha}\dv{A_i}{x_\mathrm{c}} \, ,
\end{equation}
where
\begin{equation}
\begin{aligned}
\dv{\alpha}{\mu} &= \frac{1}{2}\sqrt\frac{12}{8\lambda^3-27\mu^2}\, , \quad &\dv{x_\mathrm{c}}{\alpha} &= -\frac{\sqrt{6\lambda}}{3}\sin(\frac{\pi}{6}+\alpha)\, , \\
\dv{A_1}{x_\mathrm{c}} &= -2\frac{(6x_\mathrm{c}^2-\lambda)^{3/2}}{x_\mathrm{c}\sqrt{2x_\mathrm{c}^2-\lambda}}\, , \quad
&\dv{A_2}{x_\mathrm{c}} &= 4\frac{(6x_\mathrm{c}^2-\lambda)^{3/2}}{x_\mathrm{c}\sqrt{2x_\mathrm{c}^2-\lambda}} \, .
\end{aligned}
\end{equation}
Thus, we have $\xi_1 = -1$ and $\xi_2 = 2$, which means that $P_1=0$ and $P_2=1$. All particles are therefore trapped in $G_2$, with an action value
\begin{equation}
J = \frac{A_2}{2\pi} = 2J_0 - \frac{\lambda^*}{2}
\label{eq:theor_m1}
\end{equation}
Cooling is possible in the interval $\lambda^*/4\le J_0 \le \lambda^*/2$, i.e.\ $2J_0\le \lambda^*\le 4J_0$, which corresponds to the existence of the square roots $\sqrt{\lambda^*-2x_\mathrm{c}^2}$ and $\sqrt{6x_\mathrm{c}^2-\lambda^*}$.
On the other hand, for $\lambda^*>4J_0$, the initial condition does not belong to the outer region but to the inner region, $G_1$. In that case, the separatrix crossing occurs when $A_1=2\pi J_0$ and the particle is trapped into $G_2$ at an action $A_2(\lambda^*,\mu^*)/(2\pi)$. Using the expressions of $A_1$ and $A_2$, we find that the resulting expected final action is
\begin{equation}
J=\frac{\lambda^*}{2}-2J_0 \, ,
\label{eq:theor_m2}
\end{equation}
which means that cooling is also possible for $4J_0 \le \lambda^* \le 6J_0$, i.e.\ $\lambda^*/6\le J_0 \le \lambda^*/2$.
After being trapped in $G_2$, the particle distribution has a smaller action than the initial one, but, as before, the definition of the adiabatic invariant, being $\mu\neq 0$, is not related to $(x^2+p_x^2)/2$. Therefore, a transport process must be designed to reduce $\mu$ to zero without losing particles from $G_2$. Since the particles are trapped in region $G_2$, we need to keep its area constant, i.e.\ $\dv*{A_2}{t}=0$, or
\begin{equation}
\dv{A_2}{t} = \dv{\lambda}{t}\qty( \pdv{A_2}{\lambda} + \dv{x_\mathrm{c}}{\lambda} \pdv{A_2}{x_\mathrm{c}} ) = 0 \, .
\label{eq:mutransp}
\end{equation}
This can be used to derive a differential equation for $\mu(\lambda)$
\begin{equation}
\dv{\mu}{\lambda} = -2x_\mathrm{c}\sqrt\frac{\lambda-2x_\mathrm{c}^2}{6x_\mathrm{c}^2-\lambda}\asin\frac{\sqrt{\lambda-2x_\mathrm{c}^2}}{2x_\mathrm{c}} \, .
\label{eq:odemu}
\end{equation}
Following this equation, as $\lambda$ is reduced $\mu$ increases, and while $A_2$ remains constant $A_1$ is reduced to zero, which occurs when $\mu=(2\lambda/3)^{3/2}$. We can then safely reduce both $\mu$ and $\lambda$ to zero, stopping the perturbation: in fact, as $\mu$ is kept below $(2\lambda/3)^{3/2}$ no island is present in the phase space.
\subsubsection{Coupled variation of \texorpdfstring{$\lambda$}{l} and \texorpdfstring{$\mu$}{m} and complete trapping in \texorpdfstring{$G_1$}{G1} (Protocol~C)} \label{sec:theorC}
One could also devise a protocol in which both $\lambda$ and $\mu$ are modulated. We can express $\mu$ as a function of $\lambda$, and the expression of the capture probabilities becomes
\begin{equation}
P_i = {\frac{\pdv*{A_i}{\lambda} + \mu' \pdv*{A_i}{\mu}}{\pdv*{A_3}{\lambda} + \mu' \pdv*{A_3}{\mu}}}\eval_{\lambda=\lambda^*,\,\mu=\mu^*} \;\; i=1,\,2 \, ,
\end{equation}
where the prime symbol denotes the derivative w.r.t.\ $\lambda$.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{theor_lm.pdf}
\caption{Left: implicit solution $\lambda^*(\mu^*)$ of the equation $A_3(\lambda^*,\mu^*)=2\pi J_0$. Center: expected cooling ratio $J/J_0$ for trapping particles via the coupled variation of $\lambda$ and $\mu$, as a function of $\mu^*$. Right: required value of $\mu'$ to achieve the cooling efficiency shown in the center plot, as a function of $\mu^*$. Thanks to the ratios of variables reported on the axes, the plotted functions are unique and independent from the value of $J_0$.}
\label{fig:theor_lm}
\end{figure*}
The trapping probability is calculated at the jumping point $(\lambda^*,\,\mu^*)$. Therefore, we can define an implicit function $\lambda^*(\mu)$ that resolves the equation $A_3 = A_0$ (see Fig.~\ref{fig:theor_lm}, left). Then, we optimize the probability by imposing that: (a) all particles are trapped in region $G_1$; (b) the area $A_1$ is minimized at the trapping point. For the first condition, the equation $P_1=1$, $P_2=0$, gives the following condition on $\mu'$
\begin{equation}
\mu' = - {\frac{\pdv*{A_2}{\lambda}}{\pdv*{A_2}{\mu}}}\eval_{\lambda^*,\,\mu^*} \, .
\label{eq:muprime}
\end{equation}
Note that the signs of the partial derivatives of $A_2$ w.r.t.\ $\lambda$ and $\mu$ ensure that $\mu'<0$.
When $P_1=1$, $P_2=0$ and $2\pi \av{J} = A_1 = \lambda^* - 2J_0$, we can minimize $\av{J}$ choosing the minimum $\lambda^*$ for which trapping is possible. This corresponds to $A_1=0$, from which $\lambda^*=2J_0$, and the equation $A_3=2\pi J_0$ becomes
\begin{equation}
K_1 + K_2 = \pi J_0 \, ,
\end{equation}
that can be solved by setting $K_1=\pi J_0$ and $K_2=0$. From $K_1=\pi J_0$ we have the equation
\begin{equation}
\asin(\frac{\sqrt{2J_0-2x_\mathrm{c}^2}}{2x_\mathrm{c}}) = \frac{\pi}{2} \, ,
\end{equation}
which is solved when the argument of the arc-sine is $1$, so
\begin{equation}
\sqrt{2J_0-2x_\mathrm{c}^2} = 2 x_\mathrm{c} \,\implies\, 6x_\mathrm{c}^2 - 2J_0 = 6x_\mathrm{c}^2-\lambda = 0 \, .
\end{equation}
It is straightforward to verify that this implies $K_2=0$. Additionally, this condition induces $\pdv*{A_2}{\mu}=0$, and $\mu'$ diverges. Thus, a perfect cooling, i.e.\ in which the final value of the action is zero, would require to change $\mu$ infinitely fast, which contradicts the adiabatic condition we made to apply the theoretical results.
Although it is not possible to provide an analytical expression for the implicit solution $\lambda^*(\mu^*)$ of equation $A_3=2\pi J_0$, we can prove that the graphs shown in Fig.~\ref{fig:theor_lm} represent the unique solution after having properly scaled the axes. In particular, we find (the details are reported in Appendix~\ref{sec:app}), that the graph of the implicit solution of equation $A_3=2\pi J_0$ is independent of $J_0$ if we rescale $\lambda^*\to \lambda^*/J_0$ and $\mu^* \to \mu^*/J_0^{3/2}$ (see Fig.~\ref{fig:theor_lm}, left). Similar laws hold for the expected cooling $J/J_0$, which is a function of the only variable $\mu^*/J_0^{3/2}$ (see Fig.~\ref{fig:theor_lm}, center), and for the required $\mu'$, which fulfills the functional relation $\mu'/\sqrt{J_0} = f(\mu^*/J_0^{3/2})$ (see Fig.~\ref{fig:theor_lm}, right).
\section{Simulation results} \label{sec:numres}
We perform numerical simulations of the dynamics generated by the Hamiltonian of Eq.~\eqref{eq:ham_xp} varying $\lambda$ and $\mu$ according to the protocols previously described. In these simulations, we set $\omega_0/(2\pi)=0.414$, $k_3=1$, and invert the relations of Eq.~\eqref{eq:params} to obtain the values of $\omega$ and $\varepsilon$ as a function of $\lambda$ and $\mu$ at each time step. The amplitude-detuning parameter $\Omega_2$ has been evaluated for the unperturbed Hamiltonian at $\varepsilon=0$ by using the algorithm to evaluate the tune described in~\cite{Bartolini:316949}, to give $\Omega_2=-0.3196$.
The initial distributions used in the simulations are an infinitely thin annular distribution with initial action $J_0 = (x_0^2 + p_{x,0}^2)/2$, while uniformly distributed according to the angle variable $\phi_0 = \atan(p_{x,0}/x_0)$, i.e.\ with the p.d.f.\
\begin{equation}
\rho_{J_0}(\phi, J) = \frac{\delta(J-J_0)}{2\pi} \, .
\end{equation}
\subsection{Protocol~A: Cooling by varying \texorpdfstring{$\lambda$}{l}}
This protocol is divided in two phases. The first one is a matching phase, to slowly adapt the initial distribution to the phase space topology, as when $\mu\neq 0$ the elliptic fixed point is shifted. We will increase $\mu$ until the chosen value $\mu^*$ while keeping $\lambda=0$.
In the first phase, for time $t\in [0,t_1]$, we set
\begin{equation}
\begin{dcases}
\lambda(t) &= 0\\
\mu(t) &= \mu^* \frac{t}{t_1} \, .
\end{dcases}
\end{equation}
The actual trapping occurs in the second phase. The parameter $\lambda$ increases linearly from $0$ to a value $\Delta\lambda$. In order to trap particles at $J_0$, one needs $\Delta\lambda >\lambda^*$, where $\lambda^*=\lambda^*(\mu^*,J_0)$. We then set, for time $t\in [t_1,2t_1]$
\begin{equation}
\begin{dcases}
\lambda(t) &= \Delta\lambda \left (\frac{t-t_1}{t_1} \right ) \\
\mu(t) &= \mu^* \, .
\end{dcases}
\end{equation}
We remark that although the proposed protocol, for the sake of simplicity, envisages two phases of the same duration, it is certainly possible to remove this constraint to adapt the duration of each phase to make it as adiabatic as possible.
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{varj_lambda.pdf}
\caption{Simulated cooling ratio obtained by applying Protocol~A for different values of $\Delta\lambda$, as function of the initial annular distribution action $J_0$. A comparison with the theoretical bound on the cooling efficiency given by Eq.~\eqref{eq:finJ} is presented. The Hamiltonian~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $\mu^*=\num{7.5e-3}$, $t_1 = \num{5e4}$.}
\label{fig:result_l}
\end{figure}
Figure~\ref{fig:result_l} shows the simulated $\av{J} /J_0$ for different annular distributions $\rho_{J_0}$ as a function of the initial action $J_0$ using three values of $\Delta\lambda$ (with $\mu^*=\num{7.5e-3}$), and compares it with the theoretical estimate given by Eq.~\eqref{eq:finJ}. We remark that the scale of $\lambda$ and $\mu$ are related with that of $J_0$ and hence the selected values of $\mu^*$ do not have any specific meaning, as any change would simply rescale the $J_0$ axis in Fig.~\ref{fig:result_l}.
We observe two effects that are the root of the difference between the theoretical reduction of $\av{J} /J_0$ and the observed behavior. For larger values of $\Delta\lambda$, the cooling range is increased at the expense of the minimum cooling ratio. Given $\Delta\lambda$ and $\mu^*$, for large values of $J_0$, $\lambda$ is never big enough to achieve trapping, since the $\lambda^*$ value that solves $A_3(\lambda^*,\mu^*)$ is larger than $\Delta\lambda$. Furthermore, increasing $\Delta\lambda$ to trap more particles moves the center of $G_2$ far from the origin of the phase space (all fixed points of Eq.~\eqref{eq:hamavxy}, from the solution of the resulting cubic equation, are $O(\sqrt{\lambda})$ for large values of $\lambda$), thus decreasing the effective cooling ratio.
\subsection{Protocol~B: Cooling by varying \texorpdfstring{$\mu$}{m}}
This protocol consists of three phases: the first phase is used to perform particle trapping, with the second and the third needed to transport the particles back to the center of the phase space by progressively reducing the strength of the AC dipole.
In the first phase, for times $t\in[0,t_1]$, we have the following.
\begin{equation}
\begin{dcases}
\lambda(t) & =\lambda^*\\
\mu(t) &= \mu_1 \frac{t}{t_1} \, ,
\end{dcases}
\end{equation}
and the condition $\mu_1 > \mu^*$, where $\mu^*$ solves the equation $A_3(\lambda^*,\mu^*)=2\pi J_0$.
In the second phase, the differential equation~\eqref{eq:mutransp}
is solved. For $t\in[t_1,t_2]$, we set $\lambda(t) =\lambda^* - \dot\lambda \left (t-t_1 \right )$ and obtain $\mu(t)$ by numerically integrating the Cauchy problem
\begin{equation}
\begin{dcases}
\dv{\mu}{t} &= \dv{\lambda}{t}\dv{\mu}{\lambda} = -\dot\lambda\dv{\mu}{\lambda}\\
\mu(t_1) &= \mu_1 \, ,
\end{dcases}
\end{equation}
where $\dv*{\mu}{\lambda}$ is given by Eq.~\eqref{eq:mutransp}. The second phase is stopped at time $t_2$ once the condition $\mu(t_2)=\mu_2 = (2\lambda(t_2)/3)^{3/2}$ is met. The third phase follows for times $t\in [t_2,t_2+t_1]$, with
\begin{equation}
\begin{dcases}
\lambda(t) & =\lambda(t_2) \left [1- \left (\frac{t-t_2}{t_1} \right ) \right ]\\
\mu(t)&= \left ( \frac{2}{3}\lambda(t) \right )^{3/2} \, .
\end{dcases}
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=0.7\columnwidth]{mu_protocol.pdf}
\caption{Evolution of $\lambda(t)$ and $\mu(t)$ during the three phases of Protocol~B.}
\label{fig:protocol_mu}
\end{figure}
The plots of the time evolution of $\lambda$ and $\mu$ are shown in Fig.~\ref{fig:protocol_mu}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\columnwidth]{result_m.pdf}
\caption{Expected and simulated cooling ratio for trapping in $G_2$ using Protocol~B as a function of $\lambda^*$. The initial distribution is $\rho_{0.05}$. The Hamiltonian~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $\mu_1=0.02$, $t_1 = 1/\dot\lambda = \num{5e4}$.}
\label{fig:result_m}
\end{figure}
Figure~\ref{fig:result_m} shows the simulated cooling ratio $\av{J}/J_0$, as a function of $\lambda^*$, for an initial annular distribution $\rho_{J_0}(\phi,J)$ with $J_0=\num{0.05}$, together with the theoretical expected value given by Eqs.~\eqref{eq:theor_m1} and~\eqref{eq:theor_m2}.
\begin{figure*}[h]
\centering
\includegraphics[trim=5truemm 30truemm 5truemm 5truemm,width=\textwidth,clip=]{distr_mu.pdf}
\caption{Distributions at the end of the first (left), second (center), and third (right) phase for an initial distribution $\rho_{0.05}$ for Protocol~B for three values of $\lambda^*$. The color scale represents the initial angle $\phi_0$ and the initial distribution is the same as that shown in the left plots of Fig.~\ref{fig:angle_lm}. Note that for $\lambda^* > 4J_0=0.2$ the angular dependence of the final action is reversed w.r.t.\ $\lambda^*<0.2$. The Hamiltonian of Eq.~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $\mu_1=0.02$, $t_1 = 1/\dot\lambda = \num{5e4}$}
\label{fig:angle_m}
\end{figure*}
We note that the theory presented earlier accurately describes the simulated cooling ratio unless it is in the vicinity of $\lambda^*=4J_0=\num{0.2}$, where the theory predicts total cooling, while in simulation, $\av{J}/J_0\approx 10\%$. This is due to the angular dependence we averaged upon in our analysis, as can be inferred from Fig.~\ref{fig:angle_m}. This figure shows the distributions at the end of each of the three phases of Protocol~B for the same initial annular distribution for three values of $\lambda^*$. We observe that at the end of each phase the action of the particles, which were all the same at the beginning, are spread according to their initial phase. For example red particles, which correspond to the initial phase $\pi$, result in the innermost position when $\lambda^*=0.15$ and in the outermost position when $\lambda=0.25$. This behavior reverses for cyan particles, which have $\phi_0=0$. This means that particles with different initial angles are trapped at slightly different values of $J$. Some particles are trapped earlier or later than expected, with a larger or smaller value of $J$ than that given by theory. In the graphs, it is also visible that the inner and outer particles are reversed, depending on whether $\lambda^*< 4J_0$ or $\lambda^*>4J_0$. When $\lambda\approx 4J_0$, however, all particles are trapped at a higher value than expected no matter when they cross the separatrix, thus increasing $\av{J} $. In our simulations, we were able to reach $\av{J} /J_0 = 0.078$, for a cooling efficiency of $92\%$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\columnwidth]{result_circ_m.pdf}
\caption{Cooling ratio for trapping in $G_2$ using Protocol~B, at different values of $\lambda^*$, as a function of the initial action of the annular distribution $J_0$. The Hamiltonian~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $\mu_1=0.02$, $t_1 = 1/\dot\lambda = \num{5e4}$.}
\label{fig:result_circ_m}
\end{figure}
In Fig.~\ref{fig:result_circ_m} we show the dependence of the cooling ratio on the value of the initial action $J_0$ for three values of $\lambda^*$. The range in which $\av{J} /J_0<1$ represents the possible interval of actions of a thick annular distribution that could be cooled using Protocol~B. Note that according to the theoretical predictions cooling is possible in the range $\lambda^*/6 \leq J_0 \leq \lambda^*/2$ and the optimal cooling ratio is found at $J_0 = \lambda^*/4$.
An animation of the trapping process for a thick annular distribution is available as Supplemental Material~\footnote{See \url{https://gitlab.cern.ch/fcapoani/nonlinear-cooling-animations} for an animation showing the evolution of an initial thick annular distribution under Protocol~B.}
\subsection{Protocol C: Cooling by varying \texorpdfstring{$\lambda$}{l} and \texorpdfstring{$\mu$}{m}}
This protocol requires two phases: the first to adapt the phase space; the second for trapping and transport. Our goal, besides trapping the particles inside $G_1$, is to ensure that both at the beginning and at the end of the process the adiabatic invariant is as close as possible to the linear action variable $J=(x^2+p_x^2)/2$, which is true if the AC dipole is switched off, i.e.\ when $\mu=0$. Thus, in the first phase, $\mu$ is gradually increased, while keeping $\lambda=0$ (i.e.\ $\omega=\omega_0$), until it reaches the value needed to initiate the trapping process. In the second phase, the derivative of $\mu(\lambda)$ is kept at a constant value $\mu'$ while increasing $\lambda$, and taking advantage of the fact that as $\mu'<0$, we can slowly reduce $\mu$ until it reaches zero to recover the equivalence between the adiabatic invariant and $J$.
In the first phase, for times $t\in [0,t_1]$, we set
\begin{equation}
\begin{dcases}
\lambda(t) &= 0\\
\mu(t) &= \mu_\text{max} \frac{t}{t_1} \, ,
\end{dcases}
\end{equation}
where $\mu_\text{max}=\mu^*+\lambda^*|\mu'|$. This ensures that during the second phase when $\lambda=\lambda^*$, $\mu$ is exactly $\mu^*$ and its derivative $\mu'$ has the appropriate value. The values of $\mu^*$ and $\lambda^*$ are obtained by choosing a solution of the implicit equation $A_3=A_0$ for the selected value of $J_0$ that corresponds to the desired cooling. From Eq.~\eqref{eq:muprime}, the desired value of $\mu'$ is also calculated.
In the second phase, where $t \in [t_1,2t_1]$, we have
\begin{equation}
\begin{dcases}
\lambda(t) &= \frac{\mu_\text{max}}{|\mu'|} \left (\frac{t-t_1}{t_1} \right ) \\
\mu(t) &= \mu_\text{max}-|\mu'|\lambda(t) \, .
\end{dcases}
\end{equation}
\begin{figure}[htp]
\centering
\includegraphics[width=0.7\columnwidth]{protocol_lm.pdf}
\caption{Evolution of $\lambda(t)$ and $\mu(t)$ during the two phases of Protocol~C. The two values $\lambda_\text{max}$ and $\mu_\text{max}$ have expressions in function of the computed $\lambda^*$, $\mu^*$, and $\mu'$, i.e.\ $\mu_\text{max}=\mu^* +\lambda |\mu'|$, $\lambda_\text{max} = \lambda^* + \mu^*/|\mu'|$.}
\label{fig:protocol_lm}
\end{figure}
When the process ends and $\mu=0$ is reached, $G_2$ disappears as the perturbation provided by the AC dipole has been switched off, and the particles trapped in $G_1$ have been transported to the center of the phase space. The values of $\lambda$ and $\mu$ during the whole procedure are plotted in Fig.~\ref{fig:protocol_lm}.
We remark that although the proposed protocol envisages two phases of the same duration, it is possible to remove this constraint to adapt the duration of each phase to make them as adiabatic as possible.
\begin{figure}[htb]
\centering
\includegraphics[width=0.7\columnwidth]{result_lm.pdf}
\caption{Expected and simulated cooling ratio for trapping in $G_1$ using Protocol~C as a function of $\mu^*$ for an initial distribution $\rho_{0.05}$. The Hamiltonian~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $t_1 = \num{1e5}$.}
\label{fig:varlm}
\end{figure}
\begin{figure*}[htp]
\centering
\includegraphics[trim=10truemm 15truemm 10truemm 0truemm,width=\textwidth,clip=]{angle_lm.pdf}
\caption{Particle distributions when applying Protocol~C at the beginning (left), after the first phase (middle), and at the end of the second phase (right), for two values of $\mu^*$. The hue encodes the initial angle of the action distribution. The Hamiltonian~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $t_1 = \num{1e5}$.}
\label{fig:angle_lm}
\end{figure*}
In Fig.~\ref{fig:varlm} we show the simulated cooling ratio $\langle J\rangle/J_0$ for an initial annular distribution $\rho_{0.05}(\phi,J)$, as a function of $\mu^*$, and a comparison with the theoretically expected value $\av{J} = (\lambda^*(\mu^*) - 2J_0)/(2\pi)$. It can be seen that the agreement between theory and simulation is remarkable up to a certain breakdown value of $\mu^*$. This breakdown is due to the angular dynamics that has been neglected in the averaging process of the theory. In Fig.~\ref{fig:angle_lm} we show the initial distribution, the situation at the end of the first phase and the final distribution of particles for two different values of $\mu^*$, using the hue to represent the initial angle $\phi_0$. For both values of $\mu^*$, we observe that the distribution after the first phase is no longer infinitely thin, and that the action of each particle depends on the initial angle. As a result each particle crosses the separatrix at a different time at the end of the second phase resulting in different values of the final action. For $\mu^*$ smaller than the breakdown threshold, all particles are still trapped in $G_1$, and this angular dependence is averaged out. On the other hand, for higher values of $\mu^*$, particles that at the end of the first phase are in the outer part of the distribution can also be trapped in $G_2$ at high amplitude, thus dramatically increasing the value of the final action. We again stress that we cannot expect to reach $100\%$ cooling as $|\mu'|$ and $\mu_\text{max}$ would need to reach unlimited values. The best cooling that we could achieve in our numerical simulations is $92\%$, at $\av{J} /J_0=0.08$.
\begin{figure}[htp]
\centering
\includegraphics[width=0.7\columnwidth]{circlesratio_lm.pdf}
\caption{Simulated cooling ratio $\av{J}/J_0$ for initial distributions $\rho_{J_0}$ using Protocol~C at different values of $\mu^*$, having computed $\mu'$ and $\mu_\text{max}$ for $J_0 = \hat J_0 = 0.05$ (indicated by a vertical dotted line in the plot). The Hamiltonian~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $t_1 = \num{1e5}$}
\label{fig:circles_lm}
\end{figure}
\begin{figure*}[htp]
\centering
\includegraphics[trim=5truemm 5truemm 5truemm 0truemm,width=\textwidth,clip=]{final_distr.pdf}
\caption{Final particle distributions after applying Protocol~C, for initial distributions $\rho_{0.045}$ (left) and $\rho_{0.05}$ (right), having computed $\mu'$ for $\hat J_0 = 0.05$. The hue indicates the initial angle $\phi_0$. We see that, although the process has been optimized for $\hat J_0=0.05$, for $J_0<\hat J_0$ (and close to the minimum shown in Fig.~\ref{fig:circles_lm}) the resulting cooling is better. The Hamiltonian~\eqref{eq:ham_xp} has been used, with $k_3=1$, $\omega_0/(2\pi)=0.414$, $\Omega_2=-0.3196$, $t_1 = \num{1e5}$, $\mu^*=0.01$.}
\label{fig:final_example_lm}
\end{figure*}
To study the applicability of the cooling protocol to a more realistic particle distribution, we have looked at an ensemble of infinitely thin annular distributions $\rho_{J_0}$ covering a certain interval in $J_0$. The values of $\mu'$ and $\mu_\text{max}$ have been chosen to optimize the trapping for a particular value of $J_0$, $\hat J_0 = 0.05$. The results are shown in Fig.~\ref{fig:circles_lm}. It is clearly visible that for different values of $\mu^*$, which translates into different cooling targets for particles at $\hat J_0$, a significant range of action values is actually cooled. The width of this \emph{cooling well}, i.e.\ the range of $J_0$ where $\av{J} /J_0<1$, is the thickness of the annular distribution that the protocol can handle successfully.
We note that, contrary to theoretical expectations, the minimum value of $\av{J} /J_0$ does not occur at $\hat J_0$, although this difference tends to decrease as $\mu^*$ increases. This is due once more to the angular dynamics. Using the same parameters as the plots shown in Fig.~\ref{fig:circles_lm}, two final distributions are shown in Fig.~\ref{fig:final_example_lm} using the hue of the color to identify the initial phase. The right plot shows the case where the initial distribution is $\rho_{0.05}$, i.e.\ the initial conditions are selected at $\hat J_0$, while the left plot shows the case where the initial distribution is $\rho_{0.045}$, where the initial actions have a value $J_0<\hat J_0$, but close to the minimum. In the left plot, a gap in the final distribution is clearly visible. This can be explained by the fact that in this case some particles are trapped earlier (the red dots in the plots) due to the spread of the action after the first phase. These can end up either in $G_2$ or in $G_1$, according to the probability law, but when their areas are smaller. The average final action is therefore reduced more by this effect than by the increase induced by the particles in $G_2$.
An animation of the trapping process for a thick annular distribution is available as Supplemental Material~\footnote{See \url{https://gitlab.cern.ch/fcapoani/nonlinear-cooling-animations} for an animation showing the evolution of an initial thick annular distribution under Protocol~C.}
\section{Conclusions} \label{sec:conc}
In this paper, beam manipulations based on nonlinear beam dynamics have been devised with the goal of achieving cooling for annular transverse beam distributions. Such a beam distribution can be generated after a beam is kicked in the transverse direction. The possibility of achieving cooling by means of crossing stable resonances generated by static magnetic elements has been ruled out, however the use of an AC dipole for such manipulations has proven to be very successful.
A Hamiltonian model describing the transverse dynamics in the presence of an AC dipole has been studied using concepts from the adiabatic theory for Hamiltonian systems. This has allowed the design of three cooling protocols, two of which proved to be extremely effective with a simulated best performance of $\approx 90\%$ cooling. In physical terms this observed cooling is achieved by controlling the strength and frequency of the AC dipole according to the specifications of the proposed protocols.
Detailed numerical simulations carried out on the considered Hamiltonian systems have revealed a rich phenomenology that could be explained in detail by using adiabatic theory for Hamiltonian systems. Although an infinitely thin annular distribution was initially used, the two best protocols have been shown to have a significant cooling range. It therefore seems possible to be able to use them to cool a transverse annular beam distribution of finite thickness. Numerical studies on more realistic accelerator models will be considered in the future in view of experimental tests on a real machine.
Such annular beam distributions are also representative of the beam halo, which opens up the study of future applications to halo manipulation that could result in experimental tests at the LHC.
\section*{Acknowledgments}
We would like to express our warm thanks to Xavier Buffat for motivating discussions on the topic of this paper.
|
\section{Introduction}
Manipulating coherent superpositions of quantum states and nonclassical pathways has been a central question to many areas of science, such as coherent control \cite{Shapiro2000,Bauerle2018,Hikosaka2019} and quantum information \cite{Reiserer2015}. Important applications include controlling chemical reactions \cite{Peller2020} or electron dynamics in ultrafast molecular dissociation \cite{Kling2006,Sansone2010,Singh2010,Xu2017}, creating quantum switches \cite{Cirac1997,Kimble2008,Reiserer2015} and enhancing nonclassicality in extended systems such as light-harvesting compounds \cite{Wilde2009,O’Reilly2014}. In this wide range of scenarios, decoherence must be kept at bay, so that the timescales of interest are much shorter than the decoherence times. This brings one's attention to how coherent superpositions of quantum states and nonclassical pathways may be controlled in attoscience.
Attoscience has emerged from the interaction of matter with very intense laser fields, typically of the order of $10^{13}\mathrm{W}/\mathrm{cm}^2$ and deals with some of the shortest time scales in nature, of the order of $10^{-18}s$. These extremely short timescales bring about the possibility of controlling real-time electron dynamics \cite{Krausz2009,Lepine2014}, and have triggered many applications. Examples are subfemtosecond imaging of matter (for reviews see, e.g., \cite{Lein2007,Salieres2012R}), high-harmonic spectroscopy \cite{Marangos2016}, ultrafast photoelectron holography \cite{Huismans2011,Faria2020}, attosecond electron or hole migration \cite{Smirnova2009,Mairesse2010,Calegari2014,Kuleff2016,Calegari2016Migration}, and, more recently, strong-field phenomena in solids \cite{Ghimire2014,Vampa2018,Yu2019,Ortmann2021} and nanostructures \cite{Ciappina2017,Ciappina2019}. It has even been speculated that, in the future, laser-induced electron dynamics may lead to optoelectronic computers with switching rates 100,000 times higher than existing digital electronic systems \cite{Schiffrin2013,Schultze2013,sommer2016attosecond}.
Recently, it has been shown that quantum interference and nonclassical pathways play a vital role in strong-field enhanced ionisation of diatomic molecules \cite{chomet2019quantum}. Enhanced ionisation is known for over three decades \cite{zuo1995charge,Seideman1995}, and consists in an increase of at least one order of magnitude in the ionisation rate of a stretched molecule, in comparison with that of an atom with a similar ionisation potential. It has been attributed to the narrowing of the upfield potential barrier due to the presence of an adjacent, downfield potential well, and to coupled charge-enhanced resonant states. Although the phenomenon is widely known, phase-space tools, such as the Wigner quasi-probability distributions \cite{wigner1932on} have shed new light on its behaviour \cite{he2008strong,takemoto2011time,chomet2019quantum}. Phase-space quasi-probability distributions have found enormous success in many research areas such as quantum optics \cite{schleich2011quantum,Barnett2005,leonhardt2010essential}, quantum information \cite{Braunstein2005,Serafini2017}, chemical physics \cite{Miller2001,Miller2005} and cold gases \cite{Blakie2008}, but are hitherto underused in attoscience (for a recent review on the overall landscape within this research field see our article \cite{chomet2021attoscience}). Nevertheless, they provide valuable insight into the dynamics of these systems of interest and have been used to study phenomena such as strong-field ionisation \cite{czirjak2000wigner,he2008strong,takemoto2011time,zagoya2014quantum}, rescattering \cite{Graefe2012,Baumann2015} and entanglement \cite{Czirjak2013}. For recent work investigating different pathways in enhanced ionisation see \cite{Xu2015,Liu2021a}.
In 2008, \textit{momentum gates} occurring in phase space under strong oscillating laser fields were identified \cite{he2008strong}. These gates facilitate the flow of quasi-probability from one molecular centre to the other causing many ionisation bursts within a field cycle. These bursts were later attributed to a non-adiabatic response to the time-dependent field gradients \cite{Takemoto2010,takemoto2011time}. However, our previous publication \cite{chomet2019quantum} dispelled this notion, by showing that these momentum gates are present when there is no time dependence on the field \cite{chomet2019quantum}, or even no field at all \cite{kufel2020alternative}, and are primarily caused by quantum interference. The term \textit{quantum bridge} was coined to describe this mechanism.
In the present article, we aim at controlling the pathways behind enhanced ionisation, which have been identified in \cite{chomet2019quantum}. Some pathways may be understood within a quasi-static, semiclassical picture, whereby the quasiprobability flows along the field gradient, with tails following equienergy curves \cite{balazs1990wigner, czirjak2000wigner, zagoya2014quantum}. Other pathways are enabled by quantum interference, which creates a bridge and provides a passage for direct intra-molecular quasi-probability flow. For a time-dependent field, one must bear in mind that the phase-space configuration also changes as time progresses. Therefore, the field can act both in favour of the cyclic motion of the quantum bridges (and the enhanced ionisation) or against it. These different temporal behaviours depend on a wealth of parameters and may be explored in order to control strong-field ionisation. Recently, it has been shown that quantum interference plays an important role even in the semiclassical ionisation pathway \cite{hack2021quantum_wigner}.
Previous analyses focused on understanding the physical mechanisms behind the quantum bridges, and were restricted to a homonuclear molecular potential in the field-free setting, or under a strong static or monochromatic laser field. The initial wavepacket was taken to be an upfield or downfield localised Gaussian, or a symmetric coherent superposition thereof known as a cat state \cite{schleich2011quantum}. Here, we focus on how the quantum bridges can be manipulated by using different relative phases and wavepacket localisations in the initial superposition state, molecular potentials of heteronuclear type, where each well is weighted asymmetrically, and electric field configurations.
This is a complex task: With a wealth of tunable parameters, such as internuclear separation, relative phase and wavepacket localisation in the initial wave packet, differing nuclear charges, laser intensity and frequency as well as the pulse shape, it necessitates optimisation techniques. Machine learning methods have been explored in a wide range of scenarios in quantum physics (see the reviews \cite{Dunjko2018,Carleo2019,Carrasquilla2020} and the recent tutorial \cite{Dawid2022}) and ultrafast photonics \cite{Genty2021}. Recently, they have been employed in the attosecond science context to control attosecond pulses \cite{BelmiroChu2021}, retrieve the structure of a large molecule from laser-induced electron diffraction patterns \cite{Liu2021b} and to track quantum pathways within a spatio-temporal Feynman path integral framework \cite{Liu2020}.
In this work, we employ machine-learning techniques for dimensionality reduction, such as the T-Distributed Stochastic Neighbour Embedding (t-SNE) \cite{vandermaaten08a:tsne_original_paper} and Principal-Component Analysis (PCA) \cite{Pearson1901} to find overall trends and patterns that allow us to establish a hierarchy of parameters and classify different regimes for strong-field ionisation. Throughout, we are aided by a qualitative phase-space analysis, which allows us to understand the physics behind the patterns encountered. We first employ a model molecule in a static field as a proof of concept, and, subsequently, use time-dependent two-colour driving fields. The autocorrelation function exhibit several distinct behaviours in time, including a stepwise profile in which ionisation is switched on or off. We also provide a physical interpretation for these features.
This article is organised as follows. In section \ref{sec:model} we discuss our model, including its general features (section \ref{sec:features}) and the physical quantities of interest (section \ref{sec:tools}). Section \ref{sec:dimReduction} is devoted to the dimensionality reduction methods and their implementation. Subsequently, in section \ref{sec:static} we provide a qualitative analysis (\ref{sec:qualitative}) and a proof of concept (\ref{sec:staticTSNE}) for static fields, and in section\ref{sec:TD} we show how strong-field ionisation in two-colour fields can be optimised to produce a stepwise temporal behaviour. We commence by examining the different clusters that result from the t-SNE (section \ref{sec:TDtsneResults}) and analyse them using physical arguments (section \ref{sec:TDfullAnalysis}). Finally, in section \ref{sec:conclusions} we summarize our results and provide possible outlooks for the present studies.
\section{Model}
\label{sec:model}
\subsection{General features}
\label{sec:features}
Our system is a one-dimensional single-electron molecule whose wave function dynamics $\Psi(x,t)$ are obtained by solving the time-dependent Schr\"odinger equation (TDSE) in atomic units using the split operator method \cite{feit1982solution}
\begin{equation}\label{eq:tdse}
i \frac{\partial}{\partial t} \Psi(x,t) =
\left(
- \frac{1}{2} \frac{\partial^2}{\partial x^2} + V(x) + x E(t)
\right)
\Psi(x, t).
\end{equation}
The quantity $E(t)$ is the external laser field and $V(x)$ is the molecular binding potential given by
\begin{equation}
\label{eq:potential}
V(x) = Z_{\mathrm{r}} V_0(x-R/2) + Z_{\mathrm{l}} V_0(x+R/2),
\end{equation}
where $R$ is the internuclear distance. The first and second term appearing on the right-hand side of equation (\ref{eq:potential}) shall be called the left and right potential wells, respectively. Each potential well $V_0(x)$ is chosen to have soft-core form
\begin{equation}\label{eq:softcore}
V_0(x) = -\frac{1}{\sqrt{x^2 + a}},
\end{equation}
where $a$ is known as the softening parameter. This removes the singularity at $x=0$ of the true Coulomb potential, yet
remains long-range \cite{javanainen1988numerical,eberly1990scale, su1991model}. This parameter is typically chosen such that the ionisation energy coincides with the ground state of the potential \cite{rae1994saturation}; throughout the following analysis $a=1$ is chosen. Using a Gaussian wavepacket (Eq.~(\ref{eq:gaussian_i})) with $p_0 = q_0 = 0.0$ and $\gamma = 0.5$, this ground state energy is ${{E}_{{\rm sc}}}=-0.67$ a.u..
The homonuclear case is recovered by setting $Z_{\mathrm{l}}=Z_{\mathrm{r}}=1$. The symbol $Z$ is chosen in reference to nuclear charge; however, in this toy model, $Z_i$ is allowed to be a continuous variable. Eq.~(\ref{eq:softcore}) provides qualitative insight into the behaviour of Coulomb potentials, which suffices for the objectives of the present work. Recently, however, the soft-core potential has been modified to exhibit quantitative agreement with realistic three dimensional models \cite{Majorosi2018,Majorosi2020}.
The external laser field is either taken to be static, that is $E(t)=E_0$, or a time-dependent linearly polarised polychromatic field such that
\begin{equation}\label{eq:polychrom}
E(t)=E_0 \left[ \mathrm{cos}(\omega t) + r_t \mathrm{cos}(b \omega t+\phi) \right],
\end{equation}
where $b$ is the frequency ratio between the first and the second driving wave, $r_t$ is the field-strength ratio, $\phi$ is the relative phase between the two driving waves, and $\omega$ is the frequency of the fundamental.
This molecular model is analysed in phase space. There, the bound and continuum regions may be identified, and an example of is shown in Figure~\ref{fig:phaseSpace}. Fixed points (points that satisfy $\dot{x}=\dot{p}=0$) here are either centres or saddles \cite{Arrowsmith1992}. Centres correspond to minima of the effective potential, and are surrounded by closed orbits. Saddles correspond to maxima of the effective potential, and delimit different dynamical regions of phase space. For instance, the central saddle separates the two molecular wells with energy $E_C$, and the Stark saddle is located between the downfield well and the continuum with energy $E_S$. The energy curves of those saddles in phase-space are called separatrices. We define the energy difference between these two saddles as
\begin{equation}
\label{eq:sepEnergyDiff}
\Delta E = E_C - E_S.
\end{equation}
$\Delta E$ is either negative (Fig.~\ref{fig:phaseSpace}(a)) or positive (Fig.~\ref{fig:phaseSpace}(c)) and characterises two different phase space configurations: closed (Fig.~\ref{fig:phaseSpace}(b)) and open (Fig.~\ref{fig:phaseSpace}(d)) separatrices.
\begin{figure}
\centering
\includegraphics[width=0.6\linewidth]{sepTest4.png}
\caption{Phase portraits for the one-dimensional homonuclear molecular models described by the binding potentials (\ref{eq:potential}), using internuclear separations of $R = 4$ and $R = 6$ and a static field $E=0.07.$ a.u. [left panels], together with the corresponding effective potentials [right panels]. The Stark and the central saddles are indicated by the labels S and C in the figure, and their energy by the labels $\mathrm{E_S}$ and $\mathrm{E_C}$. The field-free separatrices and potentials are given by the dashed red lines.
}
\label{fig:phaseSpace}
\end{figure}
The initial state of the electron $\Psi(x, 0)$ is approximated as a Gaussian wavepacket,
\begin{equation}\label{eq:gaussian_i}
\Psi(x, 0)=\langle x \mid \Psi(0)\rangle=\left(\frac{\gamma}{\pi}\right)^{\frac{1}{4}} \exp \left\{-\frac{\gamma}{2}\left(x-q_{0}\right)^{2}+\mathrm{i} p_{0}\left(x-q_{0}\right)\right\},
\end{equation}
of width $\gamma$ and initial momentum $p_0 = 0$. The state shall be called localised in the left well when $q_0 = - R/2$, denoted $\Psi_{\mathrm{l}}(x, 0)$, or in the right well when $q_0 = R/2$, denoted $\Psi_{\mathrm{r}}(x, 0)$. Alternatively, we may form delocalised coherent superpositions.
We introduce some asymmetry between the initial left and right wavepackets in a delocalised setting. As such, the initial state in \cite{chomet2019quantum} is generalised by including a relative phase $\theta \in [-\pi, \pi]$ and a localisation parameter $\alpha \in [0, 1]$, giving
\begin{equation}\label{eq:genwp}
\Psi_{\alpha, \theta}(x, 0) \coloneqq \frac{\sqrt{\alpha} \Psi_{\mathrm{l}}(x, 0)+\sqrt{1-\alpha} e^{i \theta} \Psi_{\mathrm{r}}(x, 0)}{\sqrt{1+2 \sqrt{\alpha(1-\alpha)} \cos (\theta) \mathcal{I}_{o}}}.
\end{equation}
Under some circumstances, we also consider different widths $\gamma_l$ and $\gamma_r$ for the wavefunctions $\Psi_{\mathrm{l}}$ and $\Psi_{\mathrm{r}}$, respectively.
The constant of normalisation in the denominator depends on the initial state overlap $\mathcal{I}_0$:
\begin{equation}
\mathcal{I}_{0}=\int \Psi_{\text {l}}^{*}(x, 0) \Psi_{\text {r}}(x, 0) \mathrm{d} x.
\end{equation}
Note that $\Psi_{1, \theta}(x, 0) = \Psi_{\mathrm{l}}(x, 0)$ and $\Psi_{0, \theta}(x, 0) = \Psi_{\mathrm{r}}(x, 0)$ (up to a global phase), whilst intermediate values of $\alpha$ produce coherent superpositions of varying weight.
If one considers $\Psi_{\frac{1}{2}, 0}(x, 0)$ and $\Psi_{\frac{1}{2}, \pi}(x, 0)$, this will lead to even and odd cat states, for which the wavepacket localisation in each centre are equally weighted.
\subsection{Relevant quantities}
\label{sec:tools}
Of particular interest is the phenomenon of \textit{enhanced ionisation}, characterised by anomalous peaks in the ionisation rate for relatively large internuclear separations. The effect is understood to arise from the presence of strongly coupled charge resonant states in combination with a narrowing of the effective potential for the upfield charge centre \cite{zuo1995charge,Seideman1995}. The latter is a result of the neighbouring downfield centre enabling efficient population transfer from upfield to the continuum. The ionisation rate $\Gamma$ from time $t=0$ to $t=T$ is quantified using
\begin{equation}
\label{eq:rate}
\Gamma=-\ln \left(\frac{|\mathcal{P}(T)|^{2}}{|\mathcal{P}(0)|^{2}}\right) \frac{1}{T}
,\end{equation}
where $\mathcal{P}(t)$ is the probability function at $t$, given by
\begin{equation}
\mathcal{P}(t)=\int^{+\infty}_{-\infty}\Psi^*(x,t)\Psi(x,t)dx.
\label{eq:prob}
\end{equation}
This definition of ionisation rate was used in the seminal paper \cite{zuo1995charge} in the context of enhanced ionisation of molecules. Numerically, the integral limits of Eq.~(\ref{eq:prob}) are finite, taken throughout the box of size \(x=-100\) to \(x=100\) a.u. Because of irreversible ionisation Eq.~(\ref{eq:prob}) will decrease with time and be less than unity. No absorber was used, and reflections are minimised by using a total grid size twice as large as the box size. The final time is chosen to be $T = 150~\mathrm{ a.u.}$ as it has been used in previous publications \cite{chomet2021attoscience,chomet2019quantum,kufel2020alternative}. Throughout, $\Gamma$ will be written in arbitrary units.
Eq.~(\ref{eq:rate}) will be employed throughout in order to test our optimisation methods for static fields. The static field in this work is chosen in such a way that the left well is located downfield and the right well upfield.
Another relevant quantity, which provides insight into the resulting wave function dynamics, is the autocorrelation function
\begin{equation}\label{eq:auto}
a(t)=\int \Psi^{*}(x, t) \Psi(x, 0) \mathrm{d} x,
\end{equation}
which is given by the overlap integral between the initial and time propagated wave function. The autocorrelation function will be employed to assess the behaviour of the system in time-dependent electric fields.
The Wigner quasi-probability distribution \cite{Wigner1932} is a useful tool in the analysis of phase-space dynamics (for a review see \cite{Weinbub2018}). It is defined as
\begin{equation}\label{eq:wigner}
W(x, p, t)=\frac{1}{\pi} \int_{\infty}^{-\infty} \mathrm{d} \xi \Psi^{*}(x+\xi, t) \Psi(x-\xi, t) \mathrm{e}^{2 \mathrm{i} p \xi}.
\end{equation}
The Wigner function is real and normalised, with the property that its marginals correspond to physical probability distributions for each conjugate variable, respectively. W(x,p,t) may assume negative values, and this can provide an indication of nonclassicality \cite{benedict1999wigner}.
\section{Dimensionality Reduction}
\label{sec:dimReduction}
The present problem depends on many parameters: The initial wavepacket localisation $\alpha$, its widths $\gamma_l$ and $\gamma_r$ and relative phase $\theta$; molecular parameters like the weights $Z_l$ and $Z_r$ and the inter-nuclear distance $R$; and many external electric-field parameters such as $\phi$, $r_t$, $b$ and $E_0$. The main challenge to overcome when studying the effect of a large number of parameters simultaneously is how to visualise the results. Indeed, each data point will exist in a high-dimensional space (the number of dimensions equal to the number of parameter used), and ideally we would like to project our results down to a two-dimensional space.
To this end we will use T-Distributed Stochastic Neighbour Embedding (t-SNE), an unsupervised dimensionality reduction technique, meaning it will find overall trends and patterns without any prior knowledge on the origin of the data. It can be briefly summarised in two steps:
\begin{itemize}
\item Define $G_{xy}$ and $C_{xy}$ as the probability of picking a pair of points (x,y) in high dimensional and two dimensional space respectively. Both probabilities are characterised such that they are larger if the neighbouring points are close. $G_{xy}$ is a Gaussian distribution and $C_{xy}$ a Cauchy distribution.
\item Minimise the Kullback–Leibler divergence \cite{Kullback1951} between the two distributions with respect to their projection in two-dimensional space using the gradient descent method \cite{Cauchy1847}.
\end{itemize}
A more detailed explanation of the workings of the t-SNE method can be found in \cite{vandermaaten08a:tsne_original_paper}. As a reality check, we will compare all t-SNE projections to Principal Component Analysis (PCA) \cite{Pearson1901} dimensional reduction. Indeed, t-SNE is stochastic, so each run can yield different results. It focuses on preserving the local structure of the data and is a non-linear technique. On the other hand, PCA is linear, deterministic, and preserves global properties while potentially losing low-variance deviations between neighbours. Therefore by comparing both projections we can guarantee the accuracy of the t-SNE results. For clarity, some key results obtained with the PCA are given in the Appendix A.
To make use of the t-SNE capabilities we need to include information about the wavefunction in the static field case or its evolution in the time-dependent case.
\subsection{Time evolution characterisation}
\label{sec:characterisation}
In order to understand the system dynamics through the use the dimensionality reduction methods, we wish to characterise the electronic dynamics by a single value. In the static-field case we will employ the ionisation rate $\Gamma$ (\ref{eq:rate}). For the time-dependent case we choose to quantify the shape of the autocorrelation function (\ref{eq:auto}), as it carries more temporal information. A preliminary investigation showed that the autocorrelation function can take three distinct shapes (Figure \ref{fig:functSorting}): `steps', `constant', and `other', which represent different behaviours.
The classification of autocorrelation functions into the defined categories is performed by a simple deterministic decision tree algorithm. First, we check whether a function is `constant', that is, if it does not fall below a given threshold value. Next, we look for the steps. Steps are characterised by the intervals where the autocorrelation function stays approximately constant, followed by a steep decline. Hence, if we find a constant interval longer than some `critical length', we categorise it as a `step'. We define a constant interval as an interval where the difference of any two values is smaller than a critical threshold, `$\epsilon$'. If the function is not caught by these checks, it is classified as `other'.
In the results presented in section \ref{sec:TD} the threshold value is set to 0.8, the critical length is set to 100, while the critical $\epsilon$ is 0.03. All of the autocorrelation function arrays used have 3000 elements.
\begin{figure}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/ACFsorterExamples.png}
\caption{Examples of the three potential outputs of the autocorrelation function sorting algorithm. The parameters used vary greatly but are all within the range shown in Table.~\ref{table:paramTable}.}
\label{fig:functSorting}
\end{figure}
\subsection{Method}
\label{sec:dataRed}
To obtain a data point, $N-1$ parameters are randomly uniformly distributed within the range shown in Table.~\ref{table:paramTable}.
The electronic wavefunction is then evolved using the time-dependent Schr\"odinger equation and we characterise the time evolution information by a single value as described in the above section In case of a static external field studied in section \ref{sec:static} this is done by computing the ionisation rate $\Gamma$ (Eq.~(\ref{eq:rate})). For the time-dependent field in section \ref{sec:TD} we quantify the shape of the electron autocorrelation function (see section \ref{sec:characterisation}). The $N-1$ parameters chosen randomly as well as the single value characterising the time evolution form a data point of size $N$.
In certain situations specific data points are removed from the final data set in order to focus the visualisation on the matter at hand. For example in section \ref{sec:TDtsneResults}, only 1332 data points are used from the 1 million originally computed using random parameters. Indeed, in order to focus on the `step' output of the autocorrelation function, approximately 99.9\% of data points with outputs `constant' and 99.97\% of data points with output `other' are removed.
We then use the t-SNE to reduce our N-dimensional data points to 2 dimensions, projecting the data set onto the two axis $\mathrm{y}_1$ and $\mathrm{y}_2$. The data points will have clustered into different groups, the nature of which will be understood by plotting the projection as a function of one of the input parameters. This will allow us to determine which parameter or which combination of parameters influence the ionisation rate or the autocorrelation function step, and to what degree. The overall method described here and used throughout this paper is summarised schematically in Fig.~\ref{fig:schematic}. Finally, the PCA method is used to ensure that the observed behaviour is neither a remnant of the stochastic nature of the algorithm, nor due to the removing of data points from the data set. Examples of PCA results are shown in Appendix A.
\begin{table}
\begin{center}
\begin{tabular}{||c| c c c c c c c c ||}
\hline
Static Field &$\alpha$ & $\theta$ & $\gamma_r$ & $\gamma_l$ & R & $Z_r$ & $Z_l$ & $E_0$\\ [0.5ex]
\hline\hline
min & 0.0 & 0.0 & 0.2 & 0.2 & 3.0 & 1.0 & 1.0 & 0.04\\
\hline
max & 1.0 & $\pi$ & 1.0 & 1.0 & 12.0 & 2.0 & 2.0 & 0.08\\
\hline
\end{tabular}
\begin{tabular}{||c| c c c c c c c c c c c||}
\hline
Time-dependent Field &$\alpha$ & $\theta$ & $\gamma_r$ & $\gamma_l$ & R & $Z_r$ & $Z_l$ & $E_0$ & $b$ & $r_t$ & $\phi$\\ [0.5ex]
\hline\hline
min & 0.0 & 0.0 & 0.2 & 0.2 & 4.0 & 1.0 & 1.0 & 0.04 & 0.5 & 0.0 & -$\pi$\\
\hline
max & 1.0 & $\pi$ & 1.0 & 1.0 & 10.0 & 1.5 & 1.5 & 0.08 & 2.0 & 1.0 & $\pi$ \\
\hline
\end{tabular}
\end{center}
\caption{Table of all parameters used throughout this paper with the minimum and maximum value of the range they are randomly chosen from.\label{table:paramTable}
}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{flowchart.png}
\caption{Schematic outline of the methodology used throughout the paper.}
\label{fig:schematic}
\end{figure}
\section{Static fields}
\label{sec:static}
We will start our study with an analysis of enhanced ionisation in static fields. While the ultimate goal is to understand and control enhanced ionisation in time-dependent fields, the purpose of this static field section is two-fold:
\begin{itemize}
\item Obtain an in depth understanding of how certain parameters influence the system, and expand on the conclusions drawn in \cite{chomet2021attoscience,chomet2019quantum}.
\item Establish the effectiveness of the t-SNE dimensionality reduction technique as a method of simultaneous multi-parameter analysis. This will allow us to use it as a primary tool in section \ref{sec:TD}.
\end{itemize}
\subsection{Qualitative analysis}
\label{sec:qualitative}
As a starting point, we will perform a qualitative investigation of how specific parameters influence the quantum bridges and the ionisation rate in a static field using the generalised initial wavepacket given by Eq.~\eqref{eq:gaussian_i}. We employ phase-space arguments and Wigner functions to facilitate the interpretation.
\subsubsection{Initial electron wavepacket.}
\label{sec:initialwp}
Our earlier work on the subject \cite{chomet2019quantum} led to the conclusion that a major predictor of enhanced ionisation was the localisation of the initial electron wavepacket. However, only three configurations were studied: localised around the right well, localised around the left well, and delocalised equally around both centres. Here we start with a more thorough analysis of this parameter by considering it a continuous variable, $\alpha$, see Eq.~(\ref{eq:genwp}). As shown in Fig.~\ref{fig:relAmpl}, the ionisation rate is linearly proportional to the electron localisation, with $\alpha = 0.0$ (initially localised upfield) leading to the highest ionisation rate. This can be understood as the downfield centre narrowing the effective-potential barrier for the upfield centre, hence enhancing upfield ionisation.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{CDTPlots/amplitude.png}
\caption{Ionisation rate $\Gamma$ for an initial wave packet $\Psi_{\alpha, \theta}(x, 0)$ with respect to the localisation parameter $\alpha$, comparing maximum $\theta = 2.4$ and minimum $\theta = -0.8$. The external static field has strength $E_0 = 0.06$, and the molecular parameter used are internuclear distance $R = 7$, $Z_r = 1.0$ and $Z_l = 1.0$.}
\label{fig:relAmpl}
\end{figure}
In addition to a continuous localisation parameter, the state of the initial electron wavepacket is varied by a relative phase factor $\theta$, see Eq.~(\ref{eq:genwp}), between the superposition of $\Psi_r$ and $\Psi_l$. When using an equally delocalised wavepacket, $\alpha = 0.5$, the ionisation rate as a function of $\theta$ (in rad) peaks around $\theta = 2.4$ and is minimal for $\theta=-0.8$. This has little overall influence on the ionisation rate as seen in Fig.~\ref{fig:relAmpl}. Finally, the initial wavepacket widths $\gamma_r$ and $\gamma_l$, see Eq.~(\ref{eq:gaussian_i}), are set to differ from $\gamma = 0.5$, which is the minimal ground-state energy of a field free single-centre soft-core potential with softening parameter $a = 1.0$. This has no major influence on the ionisation rate and is not shown here.
\subsubsection{Heteronuclear molecule $Z_{l,r}$.}
Finally, we expand our study to include heteronuclear molecules. We analyse the ionisation rate while varying the value of molecular weights $Z_r$ and $Z_l$, of Eq.~\ref{eq:potential}. The study of heteronuclear molecules is motivated by enabling a change in the critical internuclear distance where the phase-space separatrices change configuration and in the dynamics of the quantum bridges. From Fig.~\ref{fig:rate-contour} studying the ionisation rate as a function of both charges $Z_r$ and $Z_l$ simultaneously shows a large region of suppressed ionisation as $Z_r > 1.3$. The effect of $Z_l$ is a lot less pronounced, leading to high ionisation rates even at $Z_l = 2.0$. While open separatrices do not guarantee a high ionisation rate, closed separatrices always suppress ionisation. To have a better understanding of the sharp decrease in ionisation rate with respect to $Z_r$, we look at the Wigner quasi-probability distribution in Fig.~\ref{fig:WignerZrZl}. Comparing the bound region around the right molecular centre in (a) and (b), as $Z_r$ increases, the right molecular well deepens greatly. This leads the upfield population to stay trapped and greatly suppresses ionisation. On the other hand, the increase of the left bound region in (c) due to an increase in $Z_l$ does not greatly influence the ionisation rate. This is because in enhanced ionisation the majority of the ionised population comes from the upfield centre, as shown in \cite{chomet2019quantum}. There is very little change to the upfield (right-hand-side here) centre when increasing $Z_l$.
\begin{figure}
\centering
\includegraphics[width=0.6\linewidth]{CDTPlots/ZrZl.png}
\caption{Contour plot of the ionisation rate $\Gamma$ as a function of $Z_l$ and $Z_r$ using an external static field $E_0 = 0.06$ with internuclear distance $R = 7$. The white lines indicate the separatrix energy difference $\Delta E$, with the 0.0 line indicating the transition from open separatrices ($\Delta E$ > 0) to closed separatrices ($\Delta E$ < 0).}
\label{fig:rate-contour}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{CDTPlots/ZrZlstatic2.png}
\caption{Wigner quasi-probability distribution of the ionisation from a static field $E_0 = 0.06$ with internuclear distance $R = 7$ for (a) $Z_r = 1.0$ and $Z_l = 1.0$, (b) $Z_r = 1.5$ and $Z_l = 1.0$ and (c) $Z_r = 1.0$ and $Z_l = 1.5$. The snapshots are taken at $t = 25~\mathrm{a.u.}$.}
\label{fig:WignerZrZl}
\end{figure}
Now, a more complete overview would go through the effect of $Z_l$ and $Z_r$ with respect to various parameters, which would be slow and exhaustive. Instead we propose in the following section to use dimensionality reduction techniques to analyse the effect of multiple parameters at once.
\subsection{Proof of concept - machine learning techniques}
\label{sec:staticTSNE}
We now aim to study the effect of multiple parameters on the ionisation rate, starting with a limited number in order to fully understand the results and to focus on the effect of the chosen parameters. Those are the internuclear distance R, the width of the initial superposed gaussian upfield and downfield wavepackets $\gamma_l$ and $\gamma_r$, the phase difference $\theta$ between them, and the localisation parameter of the wavepacket $\alpha$. The value of these parameters is chosen from a predetermined range described in section \ref{sec:dataRed}.
Data points comprised of these five parameters as well as the ionisation rate are projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE algorithm, shown in Figure~\ref{fig:tsne1}. This method is described in section \ref{sec:dataRed}
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{static/total1.png}
\caption{Plot of the 3000 6-dimensional data points that have been projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE dimensional reduction technique described in section \ref{sec:tools}. Each subplot has the projected data plotted as a function an initial input parameter. The initial input parameters include (a) the ionisation rate $\Gamma$ (given in arbitrary units), (b) the internuclear distance $R$, (c) the electron localisation $\alpha$. Also included but not shown here are the left initial wavepacket width $\gamma_l$, the right wavepacket width $\gamma_r$ and the phase difference between the left and right initial wavepackets $\theta$. PCA distributions using the same data points have been analysed for accuracy (not shown). The external field is taken to be static with $E_0 = 0.0534$ and the molecular weights $Z_r = 1.0$ and $Z_l = 1.0$, matching the results from \cite{chomet2019quantum}. }
\label{fig:tsne1}
\end{figure*}
From Figure~\ref{fig:tsne1}, we see that the projected data bring into focus the connection between the ionisation rate, the internuclear distance and the electron localisation. These match the results from \cite{chomet2019quantum} and section \ref{sec:qualitative}. That is, peak ionisation yield is present at and around the peak internuclear distance of $R = 6.8$ a.u., shown in Figure~\ref{fig:tsne1}(b), and when the initial wavepacket is localised upfield, shown in Figure~\ref{fig:tsne1}(c).
The low ionisation rate regions, seen in Figure~\ref{fig:tsne1}(a), split into two distinct parts seen in Fig.~\ref{fig:tsne1}(b) which correspond respectively to regions of too low and too high internuclear distance, studied in depth in \cite{chomet2019quantum}. The ionisation rate falls linearly as the electron localisation changes from localised upfield to localised downfield, as seen in Fig.~\ref{fig:relAmpl}. The effect of the gaussian widths $\gamma$ and the phase difference $\theta$ is a lot weaker than those discussed above, and are not shown here.
We will now look at the full range of parameters. In addition to the previous six, these following results will include the effects of the nuclear weight number $Z_r$ and $Z_l$, and the electric field strength $E_0$. Because of these additional parameters, we posit that there is no longer a clear internuclear distance peak, but instead a larger range of optimal internuclear distances, dependent on $E_0$, $Z_r$ and $Z_l$. For that reason the difference in energy between the central and Stark separatrix $\Delta E$ will also be analysed, but will not be an input parameter.
As seen in Figure~\ref{fig:StaticTsne2}(a), the algorithm successfully separates two clusters of low ($\Gamma < 0.002$) and high ($\Gamma > 0.05$) ionisation yield. Within the high ionisation cluster, the data points are further distinguished by a very high ($\Gamma > 0.3 $) ionisation yield region. From the distribution of other parameters in these two clusters, we can determine which lead to a high ionisation yield.
\begin{figure*}[tbp]
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{static/2ionisationRate.png}
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{static/2zR.png}
\end{minipage}\\
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{static/2ampl.png}
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{static/2E0.png}
\end{minipage}\\
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{static/2R.png}
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{static/2sepEnergy.png}
\end{minipage}\\
\caption{Plot of the 16 995 9-dimensional data points that have been projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE dimensional reduction technique described in section \ref{sec:tools}. Each subplot has the projected data plotted as a function an initial input parameter. The initial input parameters include (a) the ionisation rate $\Gamma$, (b) the right molecular well depth $Z_r$, (c) the electron localisation $\alpha$, (d) the field strength $E_0$, (e) the internuclear distance $R$. Also included but not shown here are the left molecular well depth $Z_l$, the left initial wavepacket width $\gamma_l$, the right wavepacket width $\gamma_r$ and the phase difference between the left and right initial wavepackets $\theta$. In subplot (f) the separatrix energy difference $\Delta E$ is shown for each data point, however it is not one of the input parameters. The input data set has been reduced from 100 000 data points to only have either $\Gamma < 0.002$ or $\Gamma > 0.05 $. PCA distributions using the original 100 000 data points are shown in Appendix A. }
\label{fig:StaticTsne2}
\end{figure*}
The strongest predictor of high ionisation rate is the nuclear depth factor $Z_r$, [see Fig.~\ref{fig:StaticTsne2}(b)], which in this configuration is associated to the upfield molecular well. As shown in Figure~\ref{fig:rate-contour}, the ionisation rate falls of sharply as soon as $Z_{r} > 1.0$. While the value of $Z_r$ is originally chosen randomly in the range [1:2], the average value of $Z_r$ in the high ionisation yield cluster is 1.16, and 1.12 in the $\Gamma > 0.3 $ region. Indeed, 90\% of the $\Gamma > 0.3 $ ionisation yield data points have $Z_r < 1.25$.
In Fig.~\ref{fig:StaticTsne2}(c), echoing the conclusions from Fig.~\ref{fig:tsne1} and Fig.~\ref{fig:relAmpl}, only 4\% of the high ionisation yield data points have the initial wavepacket of the electron localised downfield, with $\alpha > 0.8$. Similarly, 30\% of the high ionisation yield cluster data points are for an initial wavepacket localised upfield, with $\alpha < 0.2$. This proportion goes up to 72\% for $\Gamma > 0.3 $.
From Fig.~\ref{fig:StaticTsne2}(d), the median initial field strength $E_0$ of the high ionisation yield cluster is 0.07, despite the parameter being originally randomly chosen from a [0.04:0.08] range. This means the majority of high ionisation yield data points have a high field strength. Indeed, 82\% of data points have $E_0 > 0.06~ \mathrm{a.u.}$ in that cluster and 88\% in the $\Gamma > 0.3 $ region.
As hypothesised earlier, the clear internuclear distance peak from Fig.~\ref{fig:tsne1} is now less pronounced in Fig.~\ref{fig:StaticTsne2} (e) as we are now also varying the molecular weights $Z$ and the field amplitude $E_0$. Still, 72\% of data points within the very high, $\Gamma > 0.3 $, ionisation yield region have internuclear distances $R$ in the range $6 < R < 9$. This is surprising as the initial range for $R$ is randomly taken from [3:12]. Therefore, one would expect that one third of the data points would be in this range.
Finally, in Figure~\ref{fig:StaticTsne2}(f), we can deduce the effect of the separatrices on the ionisation yield. Indeed if the separatrices are nested (see Fig.~\ref{fig:phaseSpace}), meaning that $E_S > E_C$, or $\Delta E <0$, then the ionisation rate is suppressed. In contrast, 99.5\% data points in the high ionisation yield cluster have $\Delta E > -0.5$, and 86\% of them have $\Delta E > 0$. This separation is present despite $\Delta E$ not being used as an input parameter.
In conclusion, the t-SNE enables us to visualise the effect of multiple parameters simultaneously. All the conclusions drawn in section \ref{sec:qualitative} are visualised in Fig.~\ref{fig:StaticTsne2}. Parameters with little to no effect are swiftly identified and separated from relevant parameters. Not only that, but a hierarchy of the effect of different parameters is also established much more rapidly. For example the effect of $Z_l$ on the ionisation rate being greater than that of $\alpha$. Because of this we will be able to use the t-SNE in the following section to rapidly assert what parameters are relevant and to what degree.
\section{Time dependent fields - control of enhanced ionisation - optimising autocorrelation step function}
\label{sec:TD}
When using a polychromatic field (see Eq.~\ref{eq:polychrom}) and specific parameters, we can obtain a `controlled' ionisation release, which translates to a step function in the autocorrelation function plot.
In the following section we will be looking at what parameters lead to step functions and why specific configurations lead to a controlled ionisation release.
\subsection{t-SNE application to step functions}
\label{sec:TDtsneResults}
\subsubsection{The 3 clusters and electron localisation.}
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/allParam1/cluster5.png}
\caption{Plot of the 1332 12-dimensional data points that have been projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE dimensionality reduction technique described in section \ref{sec:tools}. Each subplot has the projected data plotted as a function of an initial input parameter (shown by the colour bar). All units are in [a.u.]. unless stated otherwise. The input parameters shown in this figure are (a) The step function shape obtained by the sorting algorithm in section \ref{sec:characterisation} that outputs `constant', `other', or the length of the step in the autocorrelation function and (b) The initial electron wavepacket localisation represented by the parameter $\alpha$ in Eq.~(\ref{eq:genwp}). The full list of input parameters as well as details on the initial data set used can be found in section \ref{sec:tools}. PCA distributions using 500 000 data points are shown in Appendix A.}
\label{fig:TDTsneCluster}
\end{figure*}
The t-SNE algorithm separates the data into three different clusters, whose nature can be seen in the two plots of Fig.~\ref{fig:TDTsneCluster}. Therein, the data points are coloured with respect to (a) their autocorrelation function type and (b) their initial electron localisation $\alpha$. Indeed, one cluster groups all autocorrelation functions classified as `constant' or `other', while the other two contain the autocorrelation functions classified as `steps'. Those two clusters are understood by looking at the distribution with respect to the initial electron localisation. All autocorrelation functions that are classified as `steps' have their initial electron wavepacket localised completely at one centre. The clusters are separated between having the initial electron wavepacket localised at the left centre ($\alpha > 0.95$, or dark red) and localised at the right centre ($\alpha < 0.05$, or dark blue). From this it is clear that the initial wavepacket must be localised either to the left or to the right for there to be steps. This is explained when looking at Fig.~\ref{fig:localisation2}, in section \ref{sec:TDeLoc}.
\subsubsection{Difference between the two `step' clusters and the `other' and `constant' cluster. }
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/allParam1/stepParam2.jpg}
\caption{Plot of the 1332 12-dimensional data points that have been projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE dimensional reduction technique described in section \ref{sec:tools}. Each subplot has the projected data plotted as a function an initial input parameter (shown by the colour bar). All units are in [a.u.]. unless stated otherwise. The input parameter shown in this figure are (a) The ratio $r_t$ between the two colour field amplitudes, (b) the overall field amplitude $E_0$ in Eq.~(\ref{eq:polychrom}) and (c) the internuclear distance $R$. The full list of input parameters as well as details on the initial data points used can be found in section \ref{sec:tools}. PCA distributions using 500 000 data points have been checked for accuracy (not shown).}
\label{fig:TDTsneStepParam}
\end{figure*}
From looking at the different clusters, we can find other correlations, shown in Fig.~\ref{fig:TDTsneStepParam}, between certain parameters and the presence of step functions. most notably, the field ratio $r_t$ between the two driving waves, in subplot (a). Indeed, while originally the value of $r_t$ is chosen randomly from a range [0:1], the two step clusters have a median $r_t$ value of 0.81 and an average of 0.75. Moreover, longer step lengths (more than 150 a.u., or orange/red) have a $r_t$ median value of 0.85 and an average of 0.8.
Next, the overall electric field strength $E_0$, chosen at random from a range of [0.04:0.08], has an median and average value of $E_0 = 0.07~\mathrm{a.u.}$. These can be understood as by seeing that both a high field ratio $r_t$ and a high field strength $E_0$ lead to a higher field peak during which more of the population will escape, causing a deeper drop in the autocorrelation function.
Finally, the internuclear distance is clearly closer to the peak internuclear distance of $R = 7~\mathrm{a.u.}$, which is also confirmed by the standard deviation being only 1.05 instead of 1.3. $R$ is also for $92.6 \%$ of all data in the step clusters above a minimum internuclear distance of $R = 6~\mathrm{a.u.}$. The cause of this effect is more complex, and is discussed in detail in section \ref{sec:TDsep}.
\subsubsection{Symmetry between the `localised to the left' and `localised to the right' step clusters. }
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/allParam1/leftRightSym3.png}
\caption{Plot of the 1332 12-dimensional data points that have been projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE dimensionality reduction technique described in section \ref{sec:tools}. Each subplot has the projected data plotted as a function an initial input parameter (shown by the colour bar). All units are in [a.u.]. unless stated otherwise. The input parameters shown in this figure are (a) the phase difference between the left and right initial wavepackets $\theta$, (b) the left initial wavepacket width $\gamma_l$, (c) the right wavepacket width $\gamma_r$, (d) the left molecular well depth $Z_l$ and (e) the right molecular well depth $Z_r$. The full list of input parameters as well as details on the initial data points used can be found in section \ref{sec:tools}. PCA distributions using 500 000 data points have been checked for accuracy (not shown).}
\label{fig:TDTsneLeftRightSym}
\end{figure*}
The value of $\alpha$ is not the only difference between the two step clusters. The next series of parameters, shown in Fig.~\ref{fig:TDTsneLeftRightSym}, are connected to the symmetry between the left and right molecular wells. First, there is quite naturally the distribution of $\gamma_l$ and $\gamma_r$. From their definition in Eq.~(\ref{eq:gaussian_i}), $\gamma_l$ only influences the initial wavepacket localised to the left. Therefore, if the initial wavepacket was localised to the left, only the initial left wavepacket width $\gamma_l$ is a relevant parameter, and vice versa. Since, as shown in Fig.~\ref{fig:TDTsneCluster}, both step clusters use initially localised wavepackets exclusively, only one $\gamma$ parameter influences each cluster and the $\theta$ parameter (subplot (a)) has no effect on them. As seen in Fig.~\ref{fig:TDTsneLeftRightSym} (b) and (c), 22.8\% of $\gamma_l$ in the localised right step cluster are above 0.8, but only 6.8\% are above 0.8 in the localised left step cluster. Similarly, 27.5\% of $\gamma_r$ in the localised left step cluster are above 0.8, while 5.3\% are above 0.8 in the localised right step cluster. The width $\gamma = 0.5$ corresponds to the minimal ground-state energy of a field free single-centre soft-core potential with our current parameters, and the vast majority of step functions have a $\gamma$ value that is within the range [0.2 : 0.8]. As $\gamma$ deviates from 0.5, part of the initial wavepacket bleeds out of the bound region despite being initially localised. This leads to oscillations during the flat `step' portion of the autocorrelation function and explains the absence of initial wavepacket widths $\gamma$ above 0.8 in the step clusters.
The other set of parameters symmetric with respect to the left/right direction are the molecular weights $Z_l$ and $Z_r$. Their effect on one cluster should be mirrored on the other. From Figs.~\ref{fig:TDTsneLeftRightSym}(d) and (e), while there is no clear connection between the localised left cluster and $Z_r$, $Z_l$ is more likely to be close to 1.0 (and vice-versa). Indeed there are 42\% of data points that have $Z_l < 1.2$ in the localised right cluster, but 70\% in the localised left cluster. Similarly, 39\% of data points have $Z_r < 1.2$ in the localised left cluster, but 67\% in the localised right cluster. This is further investigated in section \ref{sec:TDsep}.
\subsubsection{Link between different parameters and the step length.}
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/allParam1/stepLength.png}
\caption{Plot of the 1332 12-dimensional data points that have been projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE dimensionality reduction technique described in section \ref{sec:tools}. Each subplot has the projected data plotted as a function of an initial input parameter (shown by the colour bar). All units are in [a.u.] unless stated otherwise. The input parameters shown in this figure are (a) The field-frequency ratio $b$ and (b) the field offset $\phi$ in Eq.~(\ref{eq:polychrom}). The full list of input parameters as well as details on the initial data points used can be found in section \ref{sec:tools}. PCA distributions using 500 000 data points have been checked for accuracy (not shown).}
\label{fig:TDTsneStepLength}
\end{figure*}
Going back to the distribution with respect to the autocorrelation function type in Fig.~\ref{fig:TDTsneCluster}, those two `step' clusters are then also distinguished into regions of `short' (less than 150 a.u., or lime yellow) and `long' (more than 150 a.u., or orange/red) step length.
From Fig.~\ref{fig:TDTsneStepLength}, we can already glean that the field-frequency ratio $b$ affects the step length. However, there is no clear link with the field offset $\phi$. This is further expanded upon in section \ref{sec:TDfreq}.
\subsection{Analysis of the step functions}
\label{sec:TDfullAnalysis}
In order to obtain a controlled burst of ionisation, two pre-requisites are needed: First, a time interval in which the population stays bounded, and, second, a short burst of enhanced ionisation. Following the conclusions drawn from section \ref{sec:TDtsneResults}, we already know what parameter range is required or preferred. In this section, we will provide an intuitive, physics-based analysis for the controlled-ionisation burst conditions.
\subsubsection{Electron localisation.}
\label{sec:TDeLoc}
From Fig.~\ref{fig:TDTsneCluster}, it is clear that the initial electron wavepacket must be localised, either around the left or right molecular well. Fig.~\ref{fig:localisation1} illustrates the difference between using an initial localised wavepacket and a delocalised wavepacket. This, along with the Wigner function shown in Fig.~\ref{fig:localisation2}, allows us to understand the mechanism behind the controlled ionisation release.
To obtain a burst of enhanced ionisation, the system must be in an optimal configuration. This was initially found in a study using static fields in \cite{chomet2019quantum}, and further expanded upon in section \ref{sec:static}. A key point is that the initial wavepacket must be localised upfield. This allows the population to escape directly through the quantum bridge to the semiclassical escape pathway and ultimately the continuum. It also stops the quantum bridge from cycling through the momentum space and bringing the population back to the upfield centre. Similarly, the population must stay bounded for an interval of time, meaning that the system must be in the configuration with lowest ionisation rate. As seen in Fig.~\ref{fig:relAmpl}, for static fields, this happens when the initial wavepacket is localised downfield. For the time dependent polychromatic field, the terms `upfield' and `downfield' become relative as the configuration changes with the direction of the field. However in either case the initial wavepacket must be localised on one of the molecular centres. As seen in the behaviour of the Wigner function in Fig.~\ref{fig:localisation2} [right column] these two conditions are not possible if the initial wavepacket is delocalised. Because of the quantum interference in between the two molecular centres, there is always the cyclic motion of the quantum bridge.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{TD/ACs_and_field_stars.png}
\caption{Comparison of auto-correlation functions at R = 6.8 a.u. with initial wavepacket localised upfield or delocalised. Direction of the field is denoted by the colour of the background, blue regions are positive values, pink are negative. The field is shown as the orange dashed lines with values shown on the rightmost y-axis and has form $E \propto \cos{\omega t} + \cos{2 \omega t}$ where $ \omega = 0.057$ a.u. The stars denote autocorrelation function values at times $t = 10~ \mathrm{a.u.}$, $t = 25~ \mathrm{a.u.}$, $t = 50~ \mathrm{a.u.}$, and $t = 85~ \mathrm{a.u.}$.}
\label{fig:localisation1}
\end{figure}
\begin{figure}
\begin{minipage}{0.80\textwidth}
\centering
\includegraphics[width=\linewidth]{TD/3UpDelocWigner.png}
\end{minipage}\\
\caption{ Wigner quasiprobability distributions computed for internuclear distances R = 6.8 a.u. with initial wavepacket localised at the right potential well ($\alpha = 0.0$) [left column] or delocalised between the two molecular wells ($\alpha = 0.5$)[right column]. The external field has frequency $ \omega = 0.057$ a.u. and strength $E_0 = 0.0534 ~\mathrm{a.u.}$. The snap shots are taken at times $t = 10~ \mathrm{a.u.}$ for (a) and (a'), $t = 25~ \mathrm{a.u.}$ for (b) and (b'), $t = 50~ \mathrm{a.u.}$ for (c) and (c'), and $t = 85~ \mathrm{a.u.}$ for (d) and (d'). These times are indicated by stars in Fig.~\ref{fig:localisation1}.}
\label{fig:localisation2}
\end{figure}
The localisation of the wavepacket must work in conjunction with the sign of the external field for the stepwise behaviour to be present in the autocorrelation function. If the wavepacket is localised to the left and the external field $E(t)$ is positive, or if the wavepacket is localised to the right and $E(t)$ is negative, this will result in downfield localisation. Alternatively, upfield localisation occurs for a wavepacket localised to the left and negative $E(t)$, or a wavepacket localised to the right and positive $E(t)$.
In order to obtain a step function, the localisation must be upfield during the drop and downfield during the flat part of the step. In Fig.~\ref{fig:localisation1}, the different field signs are shown by the different coloured shaded areas. Indeed, during the initial drop, the external field is positive and the population localised to the right, and therefore upfield. It escapes rapidly through the quantum bridge as seen by the Wigner function in Fig.~\ref{fig:localisation2}(a) and the separatrices are open. As soon as the field changes sign, the approximately constant behaviour in the autocorrelation function begins. As seen in Figs.~\ref{fig:localisation2}(b), (c) and (d), during that time the wavepacket localisation is now downfield. Because of that, despite the separatrices being sometimes open, like in subplots (b) and (d), ionisation is suppressed. When the field is no longer strongly negative around $t = 60$, the separatrices are closed, as shown in subplot (c). This also stops the population from escaping. The autocorrelation function therefore has a step until the field becomes positive again, with the separatrices strongly open. From this analysis and the clustering obtained by the t-SNE distribution, it is clear the electron localisation is the most important parameter in obtaining a controlled ionisation release.
\subsubsection{Separatrices.}
\label{sec:TDsep}
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/ERtotal3.png}
\caption{Different plots of autocorrelation functions [Top row] and $\Delta E$ [Bottom row] for $R = 4~\mathrm{a.u.}$, $R = 7~\mathrm{a.u.}$ and $R = 10~\mathrm{a.u.}$ with [Left column] $E_0 = 0.04~\mathrm{a.u.}$, [Middle column] $E_0 = 0.06~\mathrm{a.u.}$ and [Right column] $E_0 = 0.08~\mathrm{a.u.}$. The external laser field is represented by the red dotted line (not to scale). We consider an initial wavepacket localised to the right, with $\gamma_r$ = 0.5, and equal
molecular weights $Z_r = Z_l = 1.0$. The external field amplitude and frequency ratios are $r_t = 1.0$ and $b = 0.5$, respectively. }
\label{fig:TdSepER}
\end{figure*}
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/WigR4a.png}
\caption{Evolution of the Wigner function with internuclear distance $R = 4~\mathrm{a.u.}$ and external field parameters $r_t = 1.0$, $b = 0.5$ and $E_0 = 0.04~\mathrm{a.u.}$. The snap shots are taken at times (a) $t = 65~ \mathrm{a.u.}$, (b) $t = 85~ \mathrm{a.u.}$ and (c) $t = 100~ \mathrm{a.u.}$. Initial wavepacket localised to the right, with $\gamma_r$ = 0.5. Molecular weights $Z_r = Z_l = 1.0$.}
\label{fig:TdSepWig4}
\end{figure*}
\begin{figure*}[tbp]
\centering
\includegraphics[width=0.7\textwidth]{TD/totalZrZl.png}
\caption{Different plots of autocorrelation functions and $\Delta E$ for [Left column] $Z_r = 1.0~\mathrm{a.u.}$, $Z_r = 1.25~\mathrm{a.u.}$ and $Z_r = 1.5~\mathrm{a.u.}$ with $Z_l = 1.0~\mathrm{a.u.}$. and [Right column] $Z_l = 1.0~\mathrm{a.u.}$, $Z_l = 1.25~\mathrm{a.u.}$ and $Z_l = 1.5~\mathrm{a.u.}$ with $Z_r = 1.0~\mathrm{a.u.}$. The initial wavepacket is localised to the right and other parameters are set to $R = 7~\mathrm{a.u.}$, $E = 0.06~\mathrm{a.u.}$, $r_t = 1.0$ and $b = 0.5$. The external laser field is represented by the red dotted line (not to scale).}
\label{fig:TdSepZ}
\end{figure*}
On top of the electron localisation, the presence of step functions is strongly dependent on the value of $E_0$, $r_t$, $Z_l$, $Z_r$ and $R$, as seen in Fig.~\ref{fig:TDTsneStepParam} and Fig.~\ref{fig:TDTsneLeftRightSym}. Both $E_0$ and $r_t$ have a similar effect on the external laser field: when $r_t$ is close to 1.0 and when $E_0$ is large, the field peak increases in absolute value. As shown in Fig.~\ref{fig:StaticTsne2}, when using a static field, increasing $E_0$ leads to higher ionisation yield. In our time-dependent case, when $E_0$ or $r_t$ is high, the value of $\Delta E$ at the field peak is much higher, meaning the separatrices are wide open. This phenomenon is best seen in Fig.~\ref{fig:TdSepER} for a fixed value of $R$. In the second row it is clear that the increase in $E_0$ leads to higher maximum values of $\Delta E$. For example, at internuclear distance $R = 10$, $E_0 = 0.04$ has $\Delta E = 0.7$ at the field maximum, while $E_0 = 0.08$ has $\Delta E = 1.6$. Compared to the corresponding autocorrelation functions in the first row, we see that when matched by the correct electron localisation, this will lead too a much steeper drop during the decreasing part of the autocorrelation function step. Using our previous example, at $E_0 = 0.04$ the first step plateaus at 0.9, while $E_0 = 0.08$ at 0.5.
The effect of the internuclear distance $R$ is at a glance not clear. When focusing on the initial drop, the connection between $\Delta E$ and the initial drop of the correlation seem contradictory. Intuitively one would expect a that a higher maximum $\Delta E$ lead to a steeper drop of the autocorrelation function, as was the case for varying $r_t$ and $E_0$. But the opposite appears true with varying $R$.
This is understood by results in \cite{chomet2019quantum}, and illustrated in Fig.~\ref{fig:TdSepER}. As $R$ is becomes too large, the quantum bridge between the two molecular wells weakens and the ionisation rate falls. As consequence the initial drop of the autocorrelation function is low. With $E_0$ being low as well, this leads to a `constant' autocorrelation function as seen in Fig.~\ref{fig:TdSepER} for $R = 10$ and $E = 0.04$.
When $R$ is too small, the quantum bridges between the molecular wells are at their strongest, and initially facilitate population transfer, leading to a steeper drop. However, as seen in section \ref{sec:static}, a small $R$ does not lead to a high ionisation rate. That is understood in results from \cite{chomet2019quantum}, and illustrated by the Wigner functions in Fig.\ref{fig:TdSepWig4}: the quantum bridge brings population back to the right-side core. This cyclic motion does not follow the direction of the external field. As a consequence, the quasiprobability distribution does not stay within one centre during the `flat' portion of the step autocorrelation function, as illustrated in Fig.\ref{fig:TdSepER}.
While the effect of $Z_l$ and $Z_r$ differs greatly depending on the step cluster analysed, it is mirrored, so conclusions on one can be expanded to both. Focusing on the initially localised to the right cluster from Fig.\ref{fig:TDTsneLeftRightSym}, we see that while $Z_l$ has little influence, the vast majority of step autocorrelation functions has $Z_r < 1.2$. This is illustrated in Fig.\ref{fig:TdSepZ}: The initial drop of the step autocorrelation function greatly reduces as $Z_r$ increases. From the corresponding $\Delta E$ plot, the peak separatrix energy difference goes from 0.7 for $Z_r$ = 1.0 to 0.6 for $Z_r$ = 1.5. Compared to the effect of the separatrix energy difference of $E_0$ and its effect on the initial autocorrelation function drop, it appears something else must be at play. A better understanding is obtained when looking at the ionisation rate in static fields as a function of $Z_r$ and $Z_l$. As explained in section \ref{sec:qualitative}, the biggest difference between Fig.~\ref{fig:WignerZrZl} is not the separatrix energy difference, but the range of the bound region, especially in the momentum space. Back to the time dependent situation, as $Z_r$ increases the population (here initially localised to the right) stays trapped at the right molecular well, and the initial step drop will be much shorter.
\subsubsection{Step length and two colour field frequencies.}
\label{sec:TDfreq}
From Fig.~\ref{fig:TDTsneStepLength} we can already observe that the step length is connected to the value of the two colour field frequency ratio $b$. Indeed, while step functions of step length less than $110~\mathrm{[a.u.]}$ have on average $b = 1.5$, step function of step length greater than $150~\mathrm{[a.u.]}$ have on average $b = 0.9$.
\begin{figure*}[tbp]
\centering
\includegraphics[width=\textwidth]{TD/plotA1.png}
\caption{Plot of the 27 194 8-dimensional data points that have been projected onto two axis $\mathrm{y}_1$ and $\mathrm{y}_2$ using the t-SNE dimensional reduction technique described in section~\ref{sec:tools}. All units are in [a.u.]. unless stated otherwise. Each subplot has the projected data plotted as a function a initial input parameter. The initial input parameters include (a) The output of the step function algorithm from section~\ref{sec:characterisation} that is the length of the step in the autocorrelation function and (b) the frequency-ratio of the second colour field $b$ as well as not shown here the offset of the second field $\phi$, the field ratio $r_t$, the internuclear distance $R$, the left molecular well depth $Z_l$, the right molecular well depth $Z_r$, the right initial wavepacket width $\gamma_r$. All data points use an initially localised to the right wavepacket with $E_0 = 0.07$. The input data set has been reduced from 1 000 000 data points to only look at step autocorrelation functions. PCA distributions using the original 1 000 000 data points have been analysed for accuracy (not shown).}
\label{fig:TdFieldTsne}
\end{figure*}
\begin{figure*}[tbp]
\label{fig:TDfrequency}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{TD/mediumStepB05.png}
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{TD/longStep.png}
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{TD/mediumStepB13.png}
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{TD/shortStep.png}
\end{minipage}
\caption{Plot of the corresponding autocorrelation function and external field in each of the four different clusters in Fig.~\ref{fig:TdFieldTsne}, in order of increasing field frequency ration $b$. Constant parameters are $E_0 = 0.07$, $R = 7$, $\alpha = 0.0$, $\gamma_r = 0.5$ . The parameters used in each subplot are (a) from C1; $b = 0.5$, $r_t = 1.0$, $\phi = -0.2$, $Z_r = 1.2$ and $Z_l = 1.2$, (b) from C2; $b = 0.7$, $r_t = 1.0$, $\phi = -0.2$, $Z_r = 1.2$ and $Z_l = 1.2$, (c) from C3; $b = 1.3$, $r_t = 1.0$, $\phi = 1.0$, $Z_r = 1.1$ and $Z_l = 1.1$ and (d) from C4; $b = 2.0$, $r_t = 1.0$, $\phi = -0.13$, $Z_r = 1.05$ and $Z_l = 1.5$.}
\label{fig:TdFieldStep}
\end{figure*}
In order to view the relationship between the step length and other parameters in greater detail, we will now fix certain parameters in order to get a clearer picture of the effect the others, as well as only conserve data points with `step' autocorrelation functions. First, we will start with the initial wavepacket localised at the right centre. This is equivalent to focusing only on the `localised to the right' cluster of Fig.~\ref{fig:TDTsneCluster}. Because of this, we also remove the initial wavepacket width $\gamma_l$ and $\theta$ parameters. Finally, we will fix the field strength to $E_0 = 0.07~\mathrm{a.u.}$. These results are shown in Fig.~\ref{fig:TdFieldTsne}.
The t-SNE algorithm separates the data into four different clusters. The C1 and C3 clusters represent step lengths between $110$ and $260$. They are distinguished by the value of the frequency ratio $b$ of the two-color fields: the largest C1 having an average of $b = 0.65$ and the smallest C3 $b = 1.32$. Next, the C2 cluster is the smallest and corresponds to all autocorrelation functions with a very long (greater than $260$) step length. Finally, the C4 cluster is the largest cluster and corresponds to all autocorrelation functions with a short step length. Here the average step length is $1.09$. An example of autocorrelation function and external field within each cluster is shown in Fig.~\ref{fig:TdFieldStep}.
As seen in the previous sections, the temporal location of the field peak, when matched with the population localisation, determines the step drop. From this, we can assume that the specific times for which the field peaks occur match the step length. This can be seen in Fig.~\ref{fig:TDfrequency} (a) and (b), when increasing the second field frequency-ratio from $b=0.5$ to $b = 0.7$, the distance between two field peaks increases and the step length goes from 187 to 285. This is also confirmed when looking at an example step function from clusters (c) and (d): The distance between field peaks matches the step length.
\section{Conclusions}
\label{sec:conclusions}
In this work we use dimensionality reduction techniques, namely the the t-distributed stochastic neighbor embedding (t-SNE) method and Principal component analysis (PCA) to investigate the effect of multiple parameters at once in enhanced ionisation. We show how quantum effects in strong-field ionisation of stretched diatomic molecules may be understood, classified and synchronised with the external laser field in order to create tailored ionisation bursts.
Thereby, the t-SNE was crucial to establish a hierarchy of parameters, and manipulate the relevant time scales. These time scales are either dictated by the field, or by the molecular system. The latter are associated to quantum-interference structures that provide a direct, intra-molecular population transfer via a quantum pathway in phase space, the quantum bridges \cite{chomet2019quantum,kufel2020alternative}. The results presented here have been mainly obtained with the t-SNE, while the PCA was used to rule out possible algorithm-dependent artefacts.
The effectiveness and accuracy of the t-SNE is first showcased on the ionisation rate using a static field. The expected results are found using phase-space arguments and Wigner quasiprobability distributions. They completely match the conclusions drawn from the t-SNE distributions. Furthermore, a hierarchy in the importance of different parameters is quickly and effectively demonstrated. Because of this, we confidently used the t-SNE to project data of autocorrelation functions and their parameters for time-dependent fields.
Using a two-colour field, we obtained a controlled ionisation release, which translates to a stepwise behaviour in the autocorrelation function. By using the t-SNE to project multi-dimensional data sets into 2 dimensions, ionisation profiles obtained while varying multiple parameters at once are organised and understood.
A conclusion from previous work \cite{chomet2019quantum} was that a key to understand enhanced ionisation is the interplay between the intra-molecular quantum bridges and the external field. This also holds for controlled ionisation release (represented by a step function in the autocorrelation function), where the system requires a short burst of enhanced ionisation followed by a time interval for which the population stays bounded.
Quantum bridges are tied to many predictors for a step function in the autocorrelation function, including the strongest one: electron localisation. A delocalised initial wavepacket creates an interference pattern in both positive and negative momentum space. This leads to ionisation bursts that do not follow the external field. Therefore, we cannot obtain a time interval for which the population stays bounded. Moreover, the quantum bridges are highly tied to the internuclear distance. For that reason, configurations that suppress momentum gates do not lead to the burst of ionisation needed. This is the case when the internuclear distance is too large. Alternatively, a too small internuclear distance leads to cyclic motion in the momentum space, regardless of the initial state \cite{chomet2019quantum}. This also stops the population from staying bounded at one centre.
On the other hand, a higher field peak maximum leads to stronger ionisation bursts. A stronger field peak is achieved by increasing the field amplitude $E_0$ or adding a second field. This is neatly quantifiable by the separatrix energy difference $\Delta E$. From this we determine bounded regions with closed separatrices and ionisation burst regions when the localisation is upfield with open separatrices. The timing of these field peaks matching with the upfield population lead to the start of the autocorrelation function drops. The field frequency determines the distance between the field peaks and is therefore the key to controlling the length of the steps.
Finally, changing the molecular weights $Z_{r,l}$ affects the presence of step functions. As the upfield molecular weight increases, the upfield bound region increases and this greatly suppresses ionisation.
In summary, we can understand the physical cause and requirements for controlled ionisation release by separating the phenomenon into two steps, a short burst of enhanced ionisation and a time interval where the population stays bounded. For static fields, the optimal configuration for enhanced ionisation is pinpointed to an initial wavepacket localised upfield at optimal internuclear distance with a high intensity external field and a low upfield molecular weight. The separatrices are open with a high separatrix energy difference. This also holds during the time-dependent ionisation burst time interval, and will be periodic as the upfield and downfield configurations will change along with the sign of the external field, matching the field frequency. For the population to stay bounded, either the field strength must be low, separatrices closed or barely open, or the population must be localised in the downfield centre and therefore not follow the changes in the field strength and direction. In addition, the population must not cycle around through momentum gates, meaning the internuclear distance cannot be too short. The t-SNE was greatly effective at separating results into clusters and therefore visualising results due to specific combinations of parameters. For example, an overall analysis of the effect of $Z_r$ on the step function autocorrelation function reveals only negligible effects. However, the step function results were separated into two clusters $\alpha < 0.2$ and $\alpha > 0.8$. The effect of the increased bound region and ionisation suppression from $Z_r$ only affecting the $\alpha < 0.2$ cluster is brought into focus. Moreover, the t-SNE presented a hierarchy of the correlation between different parameters and the quantity of study. From that, the role of electron localisation was pinpointed as the key parameter.
The importance of electron localisation is in agreement with the findings by several groups, reported over many years. In fact, resolving and ultimately controlling electron localisation in extended or dissociating molecules holds the promise of steering chemical reactions. This has led to many studies, both theoretical and experimental, employing, for instance, pump-probe schemes \cite{Yudin2005,He2007,Sansone2010,Singh2010,Wang2014,Xu2017}, CEP stabilised few-cycle pulses \cite{Roudnev2004,Kling2006}, long-wavelength fields \cite{Liu2011,Liu2013} or synthesised wave forms \cite{Lan2012}. Moreover, orthogonally polarised fields have also been applied to trace or control the ionisation site \cite{Wu2012,Wu2012b,Liu2017}. Regarding enhanced ionisation, electron localisation in the upfield well, together with coupled charge-resonant states, were widely mentioned in the literature as major contributing factors. This holds both for seminal enhanced ionisation papers \cite{zuo1995charge,Seideman1995} and more recent work in which models for time-dependent ionisation bursts have been developed \cite{Kelkensberg2011,Takemoto2010,takemoto2011time,Takemoto2011Bohmian}.
Since potential wells occur in a wide range of physical systems, the present techniques can be applied not only to more realistic molecular models, but also to solids and nanostructures. Thereby, a crucial issue would be to incorporate other degrees of freedom, multielectron dynamics and assess the role of decoherence. Loss of coherence can be caused by many physical mechanisms, such as coupling with additional degrees of freedom, intensity fluctuations, and incoherent emission from across the focal volume (for a brief discussion of some decoherence mechanisms in the context of strong-field quantum sensing see \cite{Maxwell2021}).
Recent studies using pump-probe schemes in $\mathrm{H}_2$ have shown that coupling of electronic and vibrational degrees of freedom affects coherence and may hinder electron localisation \cite{Vrakking2021}. Still, studies in large molecules indicate that phase relations may be preserved even in systems with many degrees of freedom \cite{Delgado2021}, and could even be controlled leading to tailored ionisation enhancement or suppression \cite{li2020control}. Our simplified model is meant as a proof of concept and a first step towards more realistic scenarios.
\section*{Acknowledgements} We would like to thank Eduard Oravkin, Pavol Drotar and Samo Hromadka for useful discussions. This work was funded by grant No.\ EP/J019143/1, from the UK Engineering and Physical Sciences Research Council (EPSRC).
|
\section{Introduction}
With the world-wide commercial deployment of the fifth-generation (5G) mobile communication systems starting in the mid-2019, research efforts on future wireless networks, e.g., the sixth-generation (6G) networks, are gaining momentum. Especially in 2020, China, the United States, Europe, Japan and South Korea all announced their national research plans for future wireless networks. Even though it still takes time to reach a broad consensus on key technologies, Internet-of-Everything (IoE) has been recognized as one major vision of the future network in both academia and industry \cite{1}.
As forecasted by multiple bodies, applications like smart industry, smart transportation, and smart health are most likely to be fast-growing areas in future network. With these emerging vertical applications, the number of intelligent devices worldwide is expected to experience an explosive growth, which is predicted to reach up to 125 billion by 2030 \cite{2}. Coming with the ultra-massive connectivity of intelligent devices, stringent requirements, e.g., transmission capacity, coverage, latency, reliability, and energy consumption, emerge to realize ultra-massive machine-type communication (umMTC) in IoE. Besides the realization of umMTC, IoE also faces challenges in terms of sensing. Intelligent devices can be distributed over a large area covering indoors, urbans, factories and remote rural regions. It is challenging to obtain and share accurate sensing information of the surrounding environment in these areas, especially under power constraints. Furthermore, sensing information obtained by these intelligent devices is expected to assist the wireless communications via beamforming and channel estimation, referred to as integrated sensing and communication (ISAC), which makes the issue more pressing. A third feature of the IoE is that information security needs careful consideration because the devices are applied in various fields, such as medical care, social life, and even national power grids. In these fields, the acceptance and widespread of future IoE mainly relies on the protection of privacy and security. However, these requirements cannot be completely fulfilled by using existing technologies in the 5G network. It calls for new wireless technologies to support these requirements in future IoE systems beyond 5G.
\begin{figure*}[!t]
\centering
\includegraphics[width = 18cm,height= 18cm]{Figure1.eps}
\caption{The organization of this paper and the interplay of IRC technologies and other technologies.}
\label{fig1}
\end{figure*}
Recently, wireless technologies enabled by intelligent reflection communication (IRC), consisting of reconfigurable intelligent surface (RIS) and ambient backscatter communication (AmBC), have received much attention due to their ability to realize smart propagation with low costs and reduced power consumption. Unlike traditional technologies, IRC can actively provide supplementary wireless links, and even transmit additional information, by exploiting environmental signals, which exhibits great potential in IoE applications.
Specifically, RIS is a planar surface composed of a large number of passive reconfigurable reflecting elements \cite{3}, each being able to control the amplitude and phase shift of incident signals independently, thereby collaboratively realizing accurate three-dimensional (3D) beamforming. In this way, the RIS is able to adjust the transmission environment around the IoE devices and strengthens the received signals at the intended receiver for ubiquitous communication. It successfully extends the coverage of devices at cell edges or blocked by obstacles.
As another important component of IRC, AmBC allows intelligent devices to communicate with each other utilizing signals from environmental radio-frequency (RF) sources \cite{4}. Different from changing propagation environments by RIS, AmBC sends data to a backscatter receiver by modulating and reflecting surrounding environmental signals. Thus, the AmBC can be used to connect IoE devices without batteries, build smart sensors, and allow devices to operate independently with minimal manual intervention. Therefore, AmBC serves as a key enabler for achieving high energy efficiency in future IoE with marginal maintenance costs.
Furthermore, the performance of AmBC depends on the RF sources and strength of electromagnetic (EM) waves in the environment. As a potential way to change the EM properties, RIS can enhance AmBC from various aspects through adjusting propagation environments, such as performing interference management, expanding communication range, and providing additional high-quality links.
Although a number of surveys on this topic of IRC have recently appeared as \cite{3}\cite{4}, and references therein, it is still incompletely clear what roles can IRC technologies play in future IoE systems. Therefore, the main contribution of this article is to provide an in-depth understanding and discuss potential applications of IRC in future IoE systems as Fig. \ref{fig1}.
The rest of this paper is organized as follows. Section \uppercase\expandafter{\romannumeral2} explores two types of IRC-empowered communication scenarios in IoE. Then, Section \uppercase\expandafter{\romannumeral3} provides the mutual benefits of IRC-empowered IoE communication and sensing. Section \uppercase\expandafter{\romannumeral4} exposes the impacts of IRC on IoE security. Furthermore, Section \uppercase\expandafter{\romannumeral5} presents a summary of open issues and challenges in IRC-empowered IoE. Finally, Section \uppercase\expandafter{\romannumeral6} draws a conclusion.
\begin{figure*}[!t]
\centering
\includegraphics[width = 18cm,height= 7.5cm]{Figure2.eps}
\caption{Architecture for RIS-empowered SAGIN in future IoE. (a) Deployment of RISs in satellite communications, (b) Deployment of RISs in aerial networks, and (c) Deployment of RISs on ground networks.}
\label{fig2}
\end{figure*}
\section{Communication in IRC-Empowered IoE}
In this section, as shown in Fig. \ref{fig1}, we elaborate on the IRC-empowered IoE communication from two aspects, i.e., the space-air-ground integrated network (SAGIN) and green communications.
\vspace{-10.pt}
\subsection{Motivation of SAGIN and Green Communications}
The SAGIN is considered as one of the essential features for future wireless networks, which makes full use of low, medium and high frequency spectrum resources to achieve seamless global coverage \cite{5}. As an effective method to improve network coverage, the IRC technologies are naturally applied, together with satellite and UAV communications, to realize the vision of SAGIN.
On the other hand, with widespread applications of IoE, a growing number of IoE devices are going to be connected. Hence, in practice, an important concern is to supply sufficient and sustainable power for all these devices. Although low-power wireless technologies enable battery life of current devices to last up one to three years, the labour cost of replacing batteries for an enormous number of devices is still unacceptable, especially for devices located in wilderness, culverts, tunnels, underwater, indoors, etc. Therefore, developing wireless technologies for green IoE communication becomes important, where IRC is envisioned to be a potential solution owing to its low power consumption.
\vspace{-10.pt}
\subsection{IRC-Empowered SAGIN in Future IoE}
\emph{1) Deployment of RISs in Satellite Networks:} For uninhabited areas where IoE devices distributed sparsely, e.g., deserts, oceans, and forests, low earth orbit (LEO) satellite communication has been an effective way of achieving seamless coverage, which is not susceptible to geographical factors. However, the LEO satellite brings the requirements of steerable antennas for tracking due to its motion with a relatively high velocity. Obviously, these requirements are unlikely to be met for IoE devices with low-complexity hardware.
Specifically, the application of RIS in satellite networks is expected to deliver the benefits from two aspects. On one hand, RISs can reduce the hardware complexity of IoE devices by allowing over-the-air signal processing rather than requiring computations on the IoE devices. On the other hand, RISs are able to realize efficient beam tracking and beam directing by configuring the phase shift of each reflecting element, thus the coverage area is adaptively adjusted with respect to the stochastic geometry of IoE devices. A direct application of this case is to deploy RISs on the solar panel of LEO satellites for broadcasting and beamforming~\cite{6}, see Fig. 2(a), where the RISs' phase shifts are adjusted according to the IoE devices, as well as their propagation channels, to enhance the quality of service (QoS). Moreover, it is exemplified that an RIS-assisted satellite can provide up to an order-of-magnitude improvement in terms of the achievable rates \cite{6}.
Furthermore, high velocity of LEO satellites also brings challenges for localization especially when multiple interconnected LEO satellites cooperate to communicate with the IoE devices. To solve this, a state-of-the-art RIS based on the holographic principle becomes an attractive way of achieving high spatial resolutions in satellite networks. Specifically for the holographic RIS, a large number of tiny and inexpensive reconfigurable elements are integrated into a compact space in order to achieve holographic beamforming with a spatially continuous aperture \cite{7}. Compared to a conventional RIS, the holographic RIS generates desired beams through holographic recording and reconstruction so that higher spatial resolution is achieved. Meanwhile, it is easy to mount a holographic RIS on mobile terminals due to its highly integrated and ultra-thin structure. We envision that holographic RIS applied in LEO satellites can promote global connectivity and high-rate communications for IoE systems.
\emph{2) Deployment of RISs in Aerial Networks:} For complex terrains lacking infrastructure, e.g., mountain villages, temporary hotspots, and disaster areas, where the air-to-ground channels easily suffer from blockage, the combinations of unmanned aerial vehicles (UAVs) and RISs provide more flexibility and improve the signal quality of IoE devices in the absence of line-of-sight (LoS) connectivity \cite{8}.
Fig. 2(b) depicts two typical cases of combining UAVs and RISs in aerial networks. In the first case, the RIS is installed on the facade of a building to reflect the signals from IoE devices to the UAV, which are further relayed to a nearby base station (BS). Virtual LoS links can be established between the UAV and IoE devices via RISs when direct LoS links are blocked by obstacles. Thus, the QoS of IoE devices can be stably met, which is beneficial to coverage extension. However, this deployment poses fundamental performance limitations. It can be difficult to find a suitable place for RIS installation and only the users at the same side of the RIS are served.
In a second case, RIS is mounted on an UAV (named as a flying RIS) to address the above issue \cite{8}. For example, an RIS is installed as a separate horizontal surface at the bottom of the UAV, which can be realized through lumped elements embedded in the reflector units. Compared to terrestrial RIS, the placement of the aerial RIS can be more flexibly optimized because of the mobility of the UAV and the ability of $360^{\circ}$ panoramic full-range coverage to serve a larger number of terminals. In addition, owing to the high altitude of the UAV, the flying RIS has a higher probability to establish strong LoS links with the IoE devices.
\emph{3) Deployment of RISs on Ground Networks:} For urban areas where the IoE devices are distributed densely, the communication is vulnerable because the devices are frequently placed at cell edges or blocked by obstacles, where the transmitted signals suffer from deep fading. One solution to this problem is deploying RISs on outdoor buildings for providing supplementary links \cite{3}, as shown in Fig. 2(c). The RIS is composed of passive scattering elements with a designed physical structure. Each scattering element can be controlled in a software-defined manner to change the EM properties of the incident signals. By a joint control of all these artificial scattering elements, the reflecting phase shifts and angles of the incident signals can be properly tuned to create a desirable multi-path effect. Thus, the reflected signals are added constructively to improve the received signal power of the IoE devices. In addition, RIS can also be mounted in indoor ceilings to help reflect signals for extending the coverage when indoor users are located in dead zones.
Most current researches on terrestrial RISs focus on two-dimensional (2D) planar array. However, due to the limited coverage of terrestrial RIS compared to aerial RIS, terrestrial RIS can be extended to a 3D spherical surface \cite{9}. The spherical RIS exhibits additional advantages of higher received signal strength and wider coverage compared to flat RIS. It can thus incorporate with future networks comprising of devices from the ground vertically up to the space and bridge the terrestrial cellular networks, UAV networks, and satellite networks. Meanwhile, its influence on performance needs to be further investigated with beamforming optimization.
\begin{figure}[!t]
\centering
\includegraphics[width = 9cm,height= 7.5cm]{Figure3.eps}
\caption{RIS-assisted AmBC communication systems.}
\label{fig3}
\end{figure}
\renewcommand\arraystretch{2}
\begin{table*}[t]
\caption{The potential communication scenarios for IRC-empowered IoE.}
\begin{center}
\begin{tabular}{>{\columncolor{pink!25}}l|lll}
\hline
\rowcolor{blue!15}\multicolumn{2}{l}{IRC-empowered IoE Schemes}&\multicolumn{1}{l}{Features}&\multicolumn{1}{l}{Benefits}\\
\hline
&\cellcolor{gray!20} &\cellcolor{gray!20}&$\bigcdot$ \cellcolor{gray!20}Enhance the QoS level\\
&\cellcolor{gray!20}\multirow{-2}{*}{RIS-aided satellite networks}&\cellcolor{gray!20}\multirow{-2}{*}{$\bigcdot$ Mounted on the solar panel}&\cellcolor{gray!20}$\bigcdot$ Reduce hardware complexity of IoE devices\\
&\multirow{2}{*}{RIS-aided UAV networks}&$\bigcdot$ Mounted on a fixed building&$\bigcdot$ Establish strong LoS links\\
&&$\bigcdot$ Mounted on a mobile UAV&$\bigcdot$ Achieve $360^{\circ}$ panoramic full-range coverage\\
\cline{2-4}
&\cellcolor{gray!20}&\cellcolor{gray!20}$\bigcdot$ Mounted on outdoor buildings&\cellcolor{gray!20}$\bigcdot$ Create a desirable multi-path effect\\
&\cellcolor{gray!20}\multirow{-2}{*}{RIS-aided ground networks}&\cellcolor{gray!20}$\bigcdot$ Mounted in indoor ceilings&\cellcolor{gray!20}$\bigcdot$ Extend the coverage of users in the dead zones\\
\multirow{-7}{*}{Global Coverage in IoE}&AmBC&$\bigcdot$ Near various types of RF sources&$\bigcdot$ Energize passive IoE devices in SAGIN\\
\hline
&\cellcolor{gray!20}&\cellcolor{gray!20}$\bigcdot$ A combination of WIT and WPT&\cellcolor{gray!20}\\
&\cellcolor{gray!20}\multirow{-2}{*}{RIS-aided SWIPT}&\cellcolor{gray!20}$\bigcdot$ Suitable for the downlink transmission&\cellcolor{gray!20}\multirow{-2}{*}{$\bigcdot$ Enhance energy harvesting performance}\\
&RIS-aided AirComp&$\bigcdot$ Suitable for the uplink transmission&$\bigcdot$ Enhance wireless-powered data aggregation\\
\cline{2-4}
\multirow{-4}{*}{Energy efficiency in IoE}&\cellcolor{gray!20}AmBC&$\cellcolor{gray!20}\bigcdot$ Modulate information on reflected signals&\cellcolor{gray!20}$\bigcdot$ Enable passive devices for signal transmission\\
\hline
\end{tabular}
\end{center}
\label{tabel1}
\end{table*}
\emph{4) AmBC for SAGIN:} Different from RIS, AmBC allows passive IoE devices to harvest energy and to simultaneously transmit their information using the reflection of incident RF signals \cite{4}. As shown in Fig. \ref{fig3}, various types of signals in SAGIN can be utilized by AmBC, such as WiFi, cellular BSs, UAVs, satellites, etc. The WiFi signals are suitable for IoE deployment in indoor applications. The cellular BSs can be a preferred energy source in densely populated urban areas. Whereas the UAVs can be a supplement in sub-urban areas. Since the wide coverage characteristic, satellite signals can be a preferred solution for AmBC in uninhabited areas. In general, AmBC is an indeed promising communication technology to energize passive IoE devices in SAGIN.
\vspace{-10.pt}
\subsection{IRC-Empowered Green Communications in Future IoE}
\emph{1) RIS-Assisted Simultaneous Wireless Information and Power Transfer (SWIPT):} As a combination of wireless information transfer (WIT) and wireless power transfer (WPT), downlink SWIPT has been a viable technology to provide green communications with energy harvesting for ultra-massive low-power devices in IoE systems. However, energy receiver (ER) usually requires much higher power than an information decoder (ID). Therefore, how to improve the efficiency of ER is a critical issue for practical SWIPT systems.
As a cost-effective technology, RIS has the capability of achieving higher spectral and energy efficiencies. Motivated by the success of RIS in wireless communications, it is expected that RIS can play a promising role in SWIPT. In particular, by adjusting the phase shifts of each RIS element, the received signal at the ER can be strengthened, thereby improving the performance of SWIPT and empowering the envisioned battery-free IoE system. In addition, RIS-assisted mobile edge computing (MEC) is also a promising technique to assist SWIPT for IoE devices, by offloading computational tasks to the edge servers with reduced energy consumption \cite{10}.
\emph{2) RIS-Assisted Over-the-Air Computation (AirComp):} Different from downlink SWIPT, uplink AirComp has been developed to enable an efficient data-fusion of sensing data from many concurrent sensor transmissions by utilizing the inherent broadcast nature of wireless communications. AirComp achieves fast wireless data aggregation by reducing both computation and resource consumption, and efficiently power the IoE devices which are densely distributed.
However, in some harsh channel propagation conditions, the harvested energy of IoE devices can be too weak to support reliable uplink transmission, leading to high signal distortion in the AirComp. As a result, it is critical to mitigate these detrimental effects of channel fading. RIS is a promising solution to reconstruct the channel environment, which motivates researchers to design wireless-powered AirComp in RIS-aided IoE systems. Specifically, in \cite{11}, it reveals that deploying an RIS in AirComp can significantly reduce the mean-squared error, which validates the effectiveness of an RIS-aided AirComp.
\emph{3) AmBC for Green IoE:} AmBC technology plays an important role in realizing a green IoE communication system. Traditionally, IoE transceivers contain active RF components to generate and convert RF signals, which makes tiny IoE devices power hungry. Considering AmBC, it has the potential to supply sufficient energy to power tiny IoE devices, e.g., LEDs, pressure sensors, and accelerometers. By exploiting AmBC, the IoE devices do not need to generate their own RF signals, which avoids the requirement for power-consuming active components. It has been reported in \cite{12} that this helps reduce their energy consumption from milli-watts to micro-watts. Therefore, there is no need for dedicating power sources to these tiny IoE devices, since the micro-watts power requirement can be satisfactorily fulfilled by ambient RF signals.
Moreover, we stress that RIS and AmBC can be exploited cooperatively to enhance the communication in a future IoE. Fig.~\ref{fig3} depicts a novel system setup where AmBC devices operate in the presence of an RIS to minimize the consumption of transmit power. This combination further increases the coverage of IoE devices compared with non-RIS-assisted ones.
As a summary of this section, we list potential communication scenarios in IRC-empowered IoE in Table \ref{tabel1}, and illustrate the interplay of IRC and other technologies in Fig. \ref{fig1}.
\vspace{-10.pt}
\section{Sensing in IRC-Empowered IoE}
The use of high frequency bands from millimeter-wave (mmWave) to terahertz (THz), wide bandwidth, and massive antenna arrays in the future enables the integration of sensing and communication. The concept of ISAC is believed to be a key feature of future networks \cite{13}. In this section, as shown in Fig. \ref{fig1}, we analyze the IRC-assisted sensing and sensing-assisted IRC in future IoE systems, respectively, based on the architecture of ISAC.
\vspace{-10.pt}
\subsection{Enhancing Sensing by IRC for Future IoE}
Sensing enhancement is not easy under wireless connections, especially in future IoE systems. Specifically, it is impossible to consume extra resources to improve the quality of sensing since most of IoE devices are constrained nodes with limited power supply and computational capability. Hence, it is expected to use extra degrees of freedom to improve the performance of sensing and IRC is a promising candidate.
\begin{figure*}[!t]
\centering
\includegraphics[width = 18cm,height= 7cm]{Figure4.eps}
\caption{Potential scenarios with the integration of IRC and sensing for future IoE systems. (a) IRC enhances sensing in a smart home, and (b) Sensing enhances IRC in a smart farm.}
\label{fig4}
\end{figure*}
Through the development of AmBC technology, smart sensors could be built, placed permanently inside a structure and set to communicate with each other to notify about this structure’s condition. The AmBC collects enough electricity to power sensors like LEDs, pressure sensors, accelerometers and so on. On the other hand, the communication links generated by RIS can be regarded as a large sensor \cite{13}. It explores the radio wave transmissions, reflections, and scatterings to obtain a better understanding of physical world. For instance, all IoE appliances such as televisions, doorbells, lights, and air conditioners are connected to the gateway in a smart home. However, indoor signals are easily blocked at high frequency bands, while an outdoor RIS installed on the buildings can help improve the communication links. In addition, indoor RISs can be installed on walls or ceilings, to increase the probability that the wireless channels are influenced by the sensing objectives. For example, Fig. 4(a) shows that posture recognition based on IRC-assisted sensing and machine learning (ML) in home IoE scenarios promotes camera-free supervision of, e.g., patients or elderly people, thus emergencies are identified immediately while individual privacy is protected. Even when the users are located in the dead zones where the direct receiving links are blocked, indoor RISs mounted on the ceilings helps reflect the signals for extending the range of sensing. Moreover, indoor WiFi signals can be further designed to power the passive IoE devices such as temperature/humidity sensors by AmBC.
\vspace{-8.pt}
\subsection{Enhancing IRC by Sensing for Future IoE}
For IRC-assisted future IoE systems, most applications mentioned in articles are based on the assumption of perfect channel state information (CSI). However, it is always a challenge to obtain the CSI accurately in practice, especially when the channel dynamics are reconfigured by a passive device as RIS. Fortunately, sensing is able to help the IRC better understand the surrounding environment, thus facilitating efficient channel estimation and beam alignment. For example, in a system with obstacles between the BS and the IoE devices, the blockages in the direct LoS links can be sensed. After a success sensing, the beam is directly pointed to the RISs and then reflected to the IoE devices. It reduces unnecessary overhead and latency for beam recovery. Meanwhile, when the BS senses IoE devices at cell edges, it enhances the signal quality through transmitting sensing-assisted beams to nearby RISs.
Specifically, as shown in Fig. 4(b), sensing is also used to detect humidity levels for agricultural IoE applications, namely a smart farm. The hygrometer communicates with the sprinkler by using RF signals reflected by a flying RIS from the BS, where the AmBC is utilized to energize the passive IoE devices, i.e., a hygrometer and sprinkler. When the hygrometer detects that the humidity of the farmland is lower than a certain threshold, the sprinkler automatically discharges water. This linkage process relies on the quality of energy harvested from ambient wireless signals. A surveillance camera can upload information to the control center according to the perceived water discharge state of the sprinkler, and enhance the reflection of the ambient wireless signals by adjusting both trajectory and orientation of the flying RIS in real time. In summary, sensing can help provide a reliable and controllable communication link for IoE systems.
\vspace{-5.pt}
\section{Security in IRC-Empowered IoE}
In addition to the communication and sensing, security is of paramount importance, especially in the developing 5G networks as many IoE applications carry private, sensitive or confidential data. In recent years, incidents of IoE security occurred frequently, and important infrastructure, such as smart homes, traffic cameras, and even the national power grid, were attacked \cite{14}. Traditionally, security issues are addressed by complicated cryptographic solutions of high complexity, which are difficult to be applied in resource-constrained sensor devices in IoE. In recent years, physical layer security (PLS), by exploiting intrinsic characteristics of the communication medium as a complement to cryptographic methods, attracts increasing attentions \cite{15}.
Due to the function of intelligent reflection, adoption of IRC can further make the PLS more efficient by involving dynamic channel environments into the design loop. In this section, we illustrate some typical potential applications of the IRC-enhanced PLS in IoE systems.
\vspace{-10.pt}
\subsection{Enhancing PLS in Unfavorable Secrecy Environments}
In general, RISs can induce the required phase shifts on the reflected signals to maximize the signal-to-noise ratio (SNR) at the legitimate user, thus enhancing the performance of security, especially in the scenarios that are favorable for eavesdroppers.
As a case study, an RIS-assisted secure IoE communication system is considered, which is equipped with an RIS with sixteen elements. In this case, the link distance from the BS to the eavesdropper (Eve) is smaller than that from the BS to the legitimate IoE device (Bob). Specifically, the average SNRs of the BS-Eve and BS-RIS-Eve links are set to $-3~{\rm dB}$ and $0~{\rm dB}$, respectively. The average SNR of the BS-Bob link is equal to that of the BS-RISs-Bob link, and the target secrecy rate is $0.05$. To compare the secrecy performance of this efficient RIS-assisted IoE scheme, two baselines are additionally tested: the communication scheme without RIS and the communication scheme of RISs with random phase. From Fig. \ref{fig5}, the results show that the RISs with the optimal phase outperform the baseline schemes in terms of secrecy outage. The reason is that the RISs with the optimal phase realize accurate beamforming and maximally improve the received SNR at Bob. Moreover, Fig. \ref{fig5} also shows that the secrecy performance of the scheme with two RISs outperforms the scheme with a single RIS. This is because in the scenario of multiple RISs, more degrees of freedom are provided through the choices of the suitable RIS to maximize the received signal at the destination. In this way, the secrecy performance of Bob is further improved from the perspective of PLS.
\begin{figure}[!t]
\centering
\includegraphics[width = 9cm,height= 7cm]{Figure5.eps}
\caption{RIS-assisted IoE communication in unfavorable secrecy environments.}
\label{fig5}
\end{figure}
\vspace{-10.pt}
\subsection{Enhancing PLS with Flying RISs}
Flying RIS-assisted IoE communication has received much attention in recent years. Compared to a terrestrial RIS with a fixed location, the position of a flying RIS can be adjusted more flexibly in the 3D space with more refined beam resolutions for PLS enhancement. This feature improves the overall security by exploiting the high mobility of the flying RIS.
Besides, it is envisioned that the flying RIS can serve as a mobile cooperative jammer jointly with the ground BS to improve the security. Through transmitting the artificial noise (AN) signals to the malicious nodes, the drawbacks of low hardware complexity for IoE devices to secure communication can be overcame by the flying RIS.
Furthermore, in a scenario where multiple Bobs and Eves exist, a single flying RIS has to fly over some Eves due to its mission requirement, where the possibility of information leakage bursts. This motivates the deployment of multiple collaborative flying RISs to achieve stronger secure communications. For example, based on the locations of Eves, the ground Bobs can be grouped into different clusters with each cluster served by a single flying RIS. In specific, a part of the flying RISs can act as aerial jammers above the Eves, while the others are deployed near Bobs. As such, it may not be necessary for the flying RIS to fly over the Eves, which thus helps reduce the information leakage.
\vspace{-1.pt}
\section{Open Issues and Challenges}
The IRC-empowered future IoE is a growing research area. Addressing the associated design issues is the key to unlock its potential. Despite fruitful research in this area, there are a variety of challenges waiting to be tackled. In this section, we highlight open issues and challenges for future research.
\vspace{-10.pt}
\subsection{Joint Design of Flying RIS for SAGIN}
The flying RIS faces multiple technical challenges for the joint design. The setups of the UAV and RIS should be taken into account together with the transmission optimization. First for UAV, higher requirements of endurance and controllability should be satisfied when carrying RIS. Besides, its trajectory needs to be jointly optimized with the transmit beamforming at the BS and the passive beamforming at RIS, to maximize the performance metrics (e.g., achievable rate, spectral efficiency, ergodic capacity, etc.) of terrestrial nodes. Secondly for RIS, we need to consider multiple factors such as size, orientation and placement when deploying it on the UAV. It is noteworthy that the orientation of RIS mounted on the UAV can significantly affect the air-ground channels.
\vspace{-10.pt}
\subsection{Dynamic CSI Acquisition}
It is observed that a majority of IRC applications are based on the assumption of perfect CSI at the transmitter or the IRC devices. In practice, both RISs and AmBC are passive, and the channel estimation by low-power IoE devices is a challenging task. Meanwhile, the computational complexity of channel estimation grows rapidly as the number of RIS reflection elements increases. In addition, Eves may attack the channel estimation by sending the same pilot sequence, which further deteriorates the channel estimation.
\vspace{-10.pt}
\subsection{Deployment of RISs and AmBC}
Generally speaking, the deployment of RISs at different locations is an intractable problem compared to the deployment of BSs or relays because of the passive nature of RISs. It is still an open challenge to adjust the physical design, deployment, collaboration, and association to enhance RISs-assisted IoE. ML and stochastic geometry-based solutions may be good options for efficient deployment of RISs.
AmBC typically covers an area ranging from several to tens of meters with low data rates. There is an immense need for improvement in range extension, to enhance the coverage of AmBC-assisted IoE. In addition, simultaneous transmission of many passive devices can create uncontrollable interference. Interference management by using AmBC among ultra-massive IoE devices becomes of increasing interest.
\vspace{-10.pt}
\section{Conclusion}
This article provides a comprehensive survey on the applications of IRC in future IoE systems. Specifically, RIS and AmBC are two different forms of IRC. First, the potential benefits in the context of SAGIN and green communications are revealed. RISs and AmBC can be exploited cooperatively to enhance the communication performance of future IoE in these two scenarios. Then, we also analyze the integration of RF sensing and IRC technologies in IoE. IRC technologies can improve sensing performance, and accurate sensing information can provide more reliable communication links. Moreover, we propose to apply IRC to enhance PLS in unfavorable secrecy environments for IoE. Research indicates that flying RISs have great potential in PLS enhancement. Open issues and challenges for realizing IRC-assisted future IoE systems are identified in terms of the joint design of flying RIS for SAGIN, dynamic CSI acquisition, and the deployment of IRC devices.
\vspace{-15.pt}
\section*{Acknowledgment}
This work was supported in part by the National Key Research and Development Program of China under Grant 2020YFB1806600; the National Natural Science Foundation of China (NSFC) under Grants 62022026 and 6211101429; and the Fundamental Research Funds for the Central Universities under Grant 2242022k30002.
\vspace{-10.pt}
|
\section{Introduction}
Modern and prospective wireless communications systems face with a growing number of simultaneously operating user's terminals and strong interference in urban outdoor and indoor environments. To keep reliable coverage and high channel capacity, conventional fixed antenna systems become insufficient. One of the proposed approaches to improve such systems consists in using passive structures for controlling electromagnetic field distributions. Tunable reflective metasurfaces (MSs - electrically dense periodic structures of subwavelength scatterers) and reflectarrays (RAs - periodic arrays of passive individually tuned antennas with typically a half-wavelength spacing) used for that purpose are commonly called Reconfigurable Intelligent Surfaces (RISs). These structures are considered in the literature as powerful and relatively cheap solutions for improving microwave and millimeter-wave wireless channels \cite{Kaina2014,Akyildiz2018}. In outdoor environments, an RIS can assist a fixed base station (BS) dynamically modifying the channel by creating one or several steerable reflected beams as a response to the wave impinging the RIS from the BS. In indoor environments, an RIS can optimize the link by dynamically modifying the spatial distribution of partially standing waves in a given room created by the radiation of a wireless access point, as proposed in \cite{Kaina2014}. In both cases, one or several RISs are to be strategically placed in the environment with a predefined location of the signal source. Applications of RISs, their mathematical models used in wireless communications, as well as their practical realizations have been recently reviewed by several groups of authors \cite{DiRenzo_Access_2019,TreryakovIEEE2020,Rui2020,alexandropoulos2020reconfigurable,Tapio2021}.
In the microwave range, RISs can be implemented in two ways \cite{Tapio2021}: using MSs or RAs. MSs consist of deeply subwavelength unit cells and have small periods compared to the wavelength, while RAs typically have the period of around $\lambda/2$, where $\lambda$ is the operational wavelength in free space. In contrast to RAs, MSs operate in the regime of strongly coupled unit cells, so they are modeled as effectively homogeneous impedance boundaries rather than arrangements of discrete elements \cite{glybovski2016metasurfaces,Eleftheriades22}. Despite MSs potentially offer more advanced reflection control such as achieving perfect reflection to angles larger than $70-80^{\circ}$ from the normal \cite{diaz2017generalized}, the features of their unit cells are much finer than of RAs. For that reason, RAs better fit to the relevant technological constraints that are especially important in the millimeter-wave range \cite{Greenerwave2021} and, therefore, serve as RISs in most papers.
Usually RAs consist of passive metal patch antennas placed over a common ground plane. The patches are excited by the incident wave and scatter (reflect) the signal with a predefined phase delay or advance. To form a desirable reflected field pattern, the corresponding distribution of the reflection phase over the aperture is to be computed and encoded. This can be achieved by adjusting the resonant frequency of each patch according to the local value of the required reflection phase. The most important function of RISs is to reflect the impinging beam into a desirable direction.
For reflection of a normally incident wave to angles of less than $70-80^{\circ}$ from the normal, it is sufficient to approximate a linear reflection phase profile within the aperture. Dynamical steering requires the resonant frequency of each patch is to be gradually tunable. This tunability can be realized e.g. by biasing a varactor diode connected between the patch and the ground plane \cite{garg2001microstrip} or between two halves of the patch. However, gradual and dynamic variation of phase for a large amount of individually controlled patches requires complex and expensive electronic circuits. To simplify the design and make it cheaper, 1-bit (also called digital \cite{Cui2014}) RISs have been proposed approximating the phase profile with only two reflection phase states at the expense of high parasitic diffraction lobes in the scattering pattern \cite{7480359}. 1-bit RISs grant a compromising directivity for the main reflection direction provided that the number of array elements is high enough (practically, several hundreds and more) \cite{8990007}.
In order to electronically switch reflection phase states, several approaches have been proposed. Most of them consist in loading patches or other resonant printed metal shapes backed by a grounded dielectric substrate with biased varactor diodes or positive-intrinsic-negative (PIN) diodes. Varactor diodes allow either uniform phase tuning \cite{4286010,Sievenpiper2003} or discrete phase switching \cite{app7090882,Zhang2018,Li_2022} depending on the capabilities of the control circuits. PIN diodes are useful only for RISs with discrete phase states. 1-bit RISs switched using PIN diodes have been proposed in \cite{Cui2014,Li2017,Zhang2018NatCom,8993760,Chen:20}. A 2-bit version of an RIS controlled with PIN diodes was presented in \cite{9020088}. While at frequencies below 10 GHz both varactor and PIN diode can be employed, at higher frequencies \cite{9036087}, especially in the 5G FR2 band \cite{8826264,Greenerwave2021,Popov2021} and above \cite{5765450}, PIN diodes were mostly used due to lower losses.
\begin{figure}[!h]
\center
\begin{minipage}{1\linewidth}
\center{\includegraphics[width=1.0\linewidth]{Fig1_v2.pdf}}
\end{minipage}
\caption{Operational principle of the proposed RIS remotely controlled from a wireless access point via infrared digital code. Inset shows the organisation of one building block composed of four individually-controlled patches loaded with varactor diodes, a photodiode, a controller (C) and DC drive (D).}
\label{fig:Work_princ}
\end{figure}
Special DC circuits are to be embedded into the design of the RIS to individually bias the diodes for setting phases of individual patches (or groups of patches). In order to decouple the DC and microwave parts of the RIS, the DC circuits including biasing lines and controllers were usually placed on the back side of a multi-layer printed-circuit board (PCB). On the contrary, the patches were placed on the front side. For isolation of the two sides, resonant traps (RF chokes) can be used at the corresponding vias \cite{5765450,7509611,Li_2022}. Embedded biasing circuits can contain distributed controllers placed in every unit cell or usually a group of unit cells. Alternatively, a centralized control unit \cite{Li2017,Zhang2018NatCom,SUN2020883} can be employed. However, in both cases RISs are to be re-designed each time when the aperture dimensions and a number of patches needs to be changed. For instance, scaling the aperture is necessary to obtain a desirable width of a reflected beam or certain spatial resolution of the reflected field distribution. The problem of scaling can be solved by using light to deliver bias voltages from a centralized control unit to diodes, or to deliver control signals from a remote control to every element of the distributed controllers. This approach allows one to encode the given reflection phase distribution without any physical wire connection for transferring biasing voltages or control signals providing a higher degree of freedom and helping strategically arrange RISs. Among other tunability methods for RISs, optical illumination appears the most attractive because it allows photonic functions to be reconfigured in a noncontact manner and with high switching speed \cite{Ren2020}.
Controlling resonant unit cells in the microwave range with light was first proposed in \cite{Light1} and further demonstrated in electromagnetic structures with various functions e.g. in \cite{Light2,Light2018,Chen:20,Dobrykh2020}. Light-controllable RISs have been implemented in several ways as follows. In \cite{Zhang2018} an RA operating around 4 GHz was composed of split patches loaded with varactor diodes in their gaps. The varactors grouped in 1D rows were commonly biased with voltages generated by PIN photodiodes operating in the photovoltaic mode in response to applied visible light from an edge row with 50 white light-emitting diodes (LEDs). By tuning the intensity of LEDs, the RIS switched between states with one or two reflected beams. In \cite{Zhang2020} an RIS was composed of blocks each one containing $4\times 4$ patches loaded with varactor diodes. The reflection phase of each block as a whole was determined by the intensity of light illuminating a meandered chain of 22 PIN photodiodes placed on the back panel producing the bias voltage for all 16 patches of the block simultaneously. Light from white LEDs was focused to illuminate individually $6\times6$ blocks to encode various 2D stepwise phase patterns at frequencies from 5.8 to 7.0 GHz. In \cite{SUN2020883} the authors proposed an infrared-controlled RIS capable of 1-bit phase encoding for reconfigurable 1D beam steering and splitting at frequencies from 4.1 to 4.35 GHz. Compared with the visible-light control mechanism, the infrared-controlled approach can work in longer remote control distances and with higher energy efficiency. The RIS had a centralized control unit with several bias voltage combinations pre-coded in FPGA enabled when receiving coded signals from a remote infrared transmitting circuit. Recently, a light-controlled RIS operating at 6.375 GHz integrated with photoresistors and varactors has been proposed, which can independently obtain a controllable range of reflection phase modulation on each unit \cite{Li_2022}. A full 2D phase encoding capability was first achieved in \cite{Li_2022} by integrating a photoresistors-based circuit on the back side of each unit cell and using the dynamically reconfigurable light field illumination created using a liquid-crystal display placed behind the RIS. This approach simplifies scaling the reflective aperture of the RIS due to independent analog-type biasing of identical unit cells all powered by the same DC voltage. However, it does not allow any distant light control that would be useful for realistic indoor applications.
In this work, we propose and demonstrate a 1-bit RIS with full 2D independent phase control of individual patches via IR light coming from a distant source. In this approach IR light is modulated with digital series control sequences and illuminates the whole aperture of the RIS, while the RIS is composed of identical building blocks powered by the same DC source. Each block contains four independently phase-switchable patches and equipped with one common microcontroller and one photodiode receiving the control sequences. Each block responds only to its own pre-coded address from a sequence and then extracts the information about the required phase states for its four patches and applies the corresponding combination of bias voltages. The operation of the proposed RIS is schematically shown in Figure~\ref{fig:Work_princ}. With this approach one can freely scale the RIS by changing the number of blocks without any complex wire control circuitry connecting them together and with no analog light-field sources. As a result, a distant and robust control via IR makes the proposed RIS advantageous for indoor communications.
\section{Design and Manufacturing of RIS}
In this Section, we describe the design procedure for one building block of the proposed scalable RIS and of an entire reflective aperture with $20\times20$ individually controllable patches with 2D phase encoding capability operating in the Wi-Fi 5-GHz range. Also, the design of the remote control unit is discussed.
In order to experimentally show the proposed principle the following technical requirements were set: frequency range covering Wi-Fi channels 40, 42 (5.17 - 5.25 GHz), vertical polarization, reflecting aperture dimensions of 600 mm $\times$ 600 mm (approximately $10\lambda\times10\lambda$).
Four unit cells were grouped in a building block with one microcontroller separately switching the varactors of four patches between two states. The entire 1-bit RIS was composed of such building blocks all being powered from the same DC voltage. Surface-mount-device (SMD) varactors SMV2019-040LF by Skyworks Solutions \cite{SMV2019-040LF} were chosen.
\subsection{RIS building block and unit cell}
The building block with four identical rectangular patches loaded with two varactor diodes each is schematically shown in inset of Figure~\ref{fig:Work_princ}. All four copper patches were printed on the top of a multilayer PCB with a common ground plane for the microwave and DC parts of the board. The substrate of the patches was made of Rogers 4003C (RO4003) material with a thickness of 1.524 mm, relative permittivity of 3.38 and dielectric loss tangent of 0.0027.
Each patch had a rectangular shape with dimensions $L$ and $W$ and a through via going through a hole in the ground plane and connecting the center of the patch to the DC circuit as shown in Figure~\ref{fig:Unit-cell_geometry}. The vias were made to apply biasing voltages to the patches relative to the ground (GND).
The resonant frequency of the patch was determined by its length $L$ and a capacitance of varactor diodes connecting the patch with the ground plane at a distance $s/2$ from the center of the patch. To connect two diodes between the patch and the ground plane, two rectangular windows with dimensions $a \times b$ were made in the patch and two small contact plates connected to the ground with two additional vias were added. Each diode was soldered between the edge of its window and the corresponding contact plate. In this design, the ground plane and the contact plates were kept at a zero DC potential, while the patch was under a bias voltage formed by the DC circuit.
The latter was positioned at two layers of FR4 substrate behind the ground plane. The FR4 part of the multilayer board had a total thickness of 1.0 mm, substrate permittivity of 4.3 and dielectric loss tangent of 0.025. The RO4003 and FR4 parts were connected by 0.44-mm-thick prerpeg KB-6065 with a dielectric permittivity of 4.2 and dielectric loss tangent of 0.013.
In the literature, a DC biasing circuit is typically isolated from the microwave part of RISs by means of linear quarter-wavelength stubs, radial stubs, inductors or other RF chokes. Unlike in the previous works where a patch was loaded with only one diode and had an RF choke, here we used a symmetric combination of two diodes per patch. This approach removed the current induced on the central via for the normal incidence and oblique incidence of a TE-polarized wave isolating the DC circuit without any choke. This allowed us to save space on the board and reduce losses.
\begin{figure}[t]
\center
\begin{minipage}{1\linewidth}
\center{\includegraphics[width=1.0\linewidth]{Unit-cell_geometry.pdf}}
\end{minipage}
\caption{Unit cell of the proposed RIS consisting of a rectangular patch on the top layer of a multilayer PCB.}
\label{fig:Unit-cell_geometry}
\end{figure}
By varying distance $s$ between the windows with diodes one can adjust the effect of the tunable capacitance to the resonant frequency of the patch. This parameter helped us to obtain 1-bit switching between two states ("0" with a reflection phase of $-90^{\circ}$ and "1" with a reflection phase of $+90^{\circ}$). These two values of phase are convenient for operation of a 1-bit RIS as they correspond to equal detuning of the patch resonance from the central operational frequency and typically correspond to equal magnitudes of the reflection coefficient. With proper $s$ it was possible to provide the bias voltages corresponding to both phases within an available range of the DC supply for the selected diode. Preferably, the difference between the voltages corresponding to both states should be as large as possible. This condition ensures that the difference in capacitance of the diode switched between two states is considerably larger than any parasitic capacitance due to soldering and mounting on a board. Moreover, that difference is to be larger than a typical capacitance deviation given in the datasheet of the diode. In our case, for stable operation of the unit cell, the difference in capacitance should exceed 0.4--0.5 pF, which was considered when choosing separation $s$.
In order to numerically predict the bias voltages providing reflection phases of $\pm90^{\circ}$ and correctly estimate the losses introduced by the RIS upon reflection, equivalent-circuit parameters of the diode were extracted from waveguide measurements as described below.
\subsection{Extraction of varactor diode's parameters}
To characterize the varactor diode as a series RLC equivalent circuit we made a set of PCBs serving as resonant waveguide terminations. Like the unit cell of the RIS, each PCB contained a resonant patch with windows and additional contact plates. However, the patch dimensions were changed in order to fit to the cross-section of WR229 waveguide with cross-section dimensions of 58.17 mm $\times$ 29.08 mm. With patch dimensions $L=30.5$ mm and $W=19.6$ mm the resonance in the reflection coefficient spectrum was around 4.2 GHz with a zero voltage applied to the diodes (initial estimation was made based on the data from the manufacturer's datasheet). The other parameters according to Figure \ref{fig:Unit-cell_geometry} were: $b=5$ mm, $a=6$ mm, $w=0.5$ mm, $l_1=1.7$ mm, $d=0.2$ mm, $s=13.9$ mm, $2r=0.4$ mm.
Three three-layers PCBs were manufactured with one 1.524-mm-thick and one 0.508-mm-thick Rogers 4003C substrate.
The first PCB had two varactors (shown in Figure~\ref{fig:waveguide_meas}(a)) connected in the windows of the patch in the same way as discussed above for the unit cell. Two other PCBs had no varactors. In one of them the patch was disconnected from the contact plates in the windows, while in the other one short-circuiting metal strips connected the patch to the contact plates instead of the varactors.
The diode's parameters were extracted based on the comparison between the simulated and measured spectra of the reflection coefficient at the input of the waveguide section with a length of 500 mm. Numerical simulations were made in CST Microwave Studio 2020. In simulations both varactors were represented by lumped elements with sizes 0.3 mm $\times$ 0.475 mm. The impedance of the lumped elements was described by the series-type equivalent circuit with capacitance $C_{d}(U)$ and resistance $R_{d}(U)$ depending on bias voltage $U$ as well as constant parasitic lead inductance $L_{d}$ and parasitic capacitance $C_{\text{par}}$ as shown in the inset in Figure~\ref{fig:waveguide_meas}(a).
The measurements were carried out in the range of 3--5 GHz with a step of 0.5 MHz. In the measurements the waveguide section was connected to vector network analyzer (VNA) Agilent E8362C calibrated using a standard coaxial kit via coaxial-to-waveguide junction (see Figure~\ref{fig:waveguide_meas}(a)). The PCBs with short-circuiting metal strips and splits at the places of the varactor diodes were used to ensure the correspondence between the calculated and measured reflection coefficient spectra in the absence of the varactors and to avoid possible numerical errors.
Parasitic ripples in reflection coefficient spectra caused by multiple reflections in the waveguide section between the sample and coaxial-to waveguide transition were suppressed using the time gating technique. For the PCB with two diodes bias voltage $U$ was varied from 0 to 5 V with a step of 0.2 V.
\begin{figure*}[t]
\center
\begin{minipage}{1\linewidth}
\center{\includegraphics[width=1.0\linewidth]{FIg_waveguide_3.pdf}}
\end{minipage}
\caption{To extraction of varactor diode's parameters: manufactured waveguide termination with a patch and two varactor diodes (top) and measurement setup (bottom) with a waveguide section and VNA (a); measured and numerically calculated reflection coefficient magnitude (b) and phase (c); measured and numerically calculated normalized real (d) and imaginary (e) part of the normalized impedance of the waveguide load. Inset in (a) shown an equivalent circuit used for varactor diodes.}
\label{fig:waveguide_meas}
\end{figure*}
Two examples of the measured reflection coefficient spectra are given in Figure~\ref{fig:waveguide_meas}(b,c) for $U=0$ V and 3.2 V (the bias voltages eventually taken for the experimental prototype of the RIS). The parameters were obtained by their parametric variation in the numerical model to fit the simulated spectra to the measured ones. With this aim, impedance curves of the PCB loads normalized to a characteristic waveguide impedance $Z_0$ were used. Particularly, the following properties were to made the same: (1) the frequency at which the real part reaches its maximum (the resonance); (2) the maximum level for the real part of the impedance; (3) the slope of the imaginary part at the resonance. Note that for different $U$, all three criteria are to be achieved only by varying the capacitance $C_{d}$ and resistance $R_{d}$ of the diode. It was assumed that the parasitic inductance and capacitance of the diode were independent on $U$. The resulting normalized impedance curves are given in Figures~\ref{fig:waveguide_meas}(d,e). As can be seen from the results, good agreement between measured and simulated curves was achieved. The extracted parameters were the following: $L_{d}=0.2$ nH, $C_{\text{par}}=30$ fF. For $U=0$ V the main parameters were $C_{d}=2.1$ pF and $R_{d}=7.5$ Ohm, while for $U=3.2$ V, $C_{d}=0.87$ pF and $R_{d}=7.1$ Ohm. Some difference from the equivalent series resistance of 4.5 Ohm given in the datasheet can be explained by the fact that the manufacturer's data was provided for low frequencies.
Next, the extracted equivalent-circuit parameters of the varactor diode were used in optimization of the unit cell of the RIS at 5.2 GHz.
\subsection{Design and organization of RIS}
Based on the extracted parameters of the diode, the unit cell was optimized via numerical simulations. Varactors were modeled in the same way as in the previous subsection. The goal was to obtain "0" and "1" phase states at the central frequency of 5.2 GHz for the normal incidence (vertical polarization along $y$-axis). Using Frequency-Domain Solver of CST Microwave Studio, the structure depicted in Figure~\ref{fig:Unit-cell_geometry} was modeled with a Bloch-Floquet port in conjunction with unit-cell boundary conditions. The reflection coefficient was calculated in the range of 4.9--5.5 GHz. The resulting parameters (all in mm) read: $L=16.6$, $W=22.0$, $a=2.0$, $b=2.6$, $w=0.5$, $l_1=0.7$, $d=0.5$, $s=11.6$, $2r=0.5$. The magnitude and phase of the simulated reflection coefficient is shown with solid lines in Figures~\ref{fig:Reflection_sim_meas_finite} (a-b). As can be seen, the chosen parameters indeed provide that "0" and "1" states realized with $U=0$ V and $U=3.2$ V, correspondingly, differ by approximately $180^{\circ}$ in the reflection phase. This is additionally confirmed by the plot of the phase difference for the two voltages vs. frequency (Figure~\ref{fig:Reflection_sim_meas_finite}(c)). The phase difference holds within the range of 150--210$^{\circ}$ (maximum deviation of phase difference of $30^{\circ}$ from $180^{\circ}$) at frequencies from 5.12 to 5.28 GHz. Note that this operational range covers the target bands of Wi-Fi channels 40 and 42 (5.17-- 5.25 GHz). The reflection coefficient magnitude of around 0.5 in the operational range, as shown in the simulations, was due to dissipation losses mostly associated with the varactors.
\begin{figure}
\center
\begin{minipage}{1\linewidth}
\center{\includegraphics[width=0.8\linewidth]{Reflection_sim_meas_2.pdf}}
\end{minipage}
\caption{Simulated (solid curves) and measured (dashed curves) reflection characteristics of the RIS with all patches in "0" state ($U=0$ V) and in "1" state ($U=3.2$ V): reflection coefficient magnitude (a) and phase (b) for a normally incident vertically-polarized plane wave; phase shift between two reflection states (c).}
\label{fig:Reflection_sim_meas_finite}
\end{figure}
Four identical unit cells were combined in one building block of the scalable RIS. The idea behind this combination is that the block has its own control unit that allows it receiving control sequences via infrared light regardless its position within the RIS aperture. In frames of a single block, an array of $2 \times 2$ patches was placed on the top layer of the multilayer PCB (on a grounded Rogers substrate), while the biasing circuit independently forming four control voltages for four pairs of varactor diodes was positioned on FR4 substrates behind the ground plane. The microwave and DC circuits had a common ground plane as an internal layer between Rogers and FR4 substrates kept under a zero potential.
The biasing circuit contained an IR receiver, a microcontroller, and a driver circuit all mounted on the opposite side of the multilayer PCB. The PCB of the block had a circular through hole in the center between corners of four patches, which allowed the IR receiver to be sensitive to IR illuminating the face of the RIS. This solution is advantageous over previous light-controlled RISs illuminated from the back as it simplifies mounting low-profile RISs on walls for indoor applications.
The IR receiver was based on TSOP34338 optoelectronic module that included an optical IR filter, IR photodiode (PD), a band-pass filter (BPF), amplifier with an automatic gain control (AGC), demodulator and a digital output stage for interfacing with the microcontroller.
The biasing circuit used ATTiny441 microcontroller with a small footprint and low power consumption. The board also included a hardware-based serial port making interfacing with a photoreceiver simpler. CodeVision AVR integrated development environment (IDE) was used for microcontroller software development in C programming language.
The bias voltages were applied to the central vias of four patches connecting each patch to a field-effect transistor (FET) based driver circuit. In turn, the DC circuits controlled the reflection state of all four diode pairs by switching between two different bias voltages. The voltage combinations were prescribed by the common controller receiving the IR control signal from the IR remote unit distantly placed in the same room.
Figure~\ref{fig:3}(b) shows the scheme of the biasing circuits of a single building block of the RIS. To test the DC circuit, a prototype of the block was manufactured. Its bottom side (with DC biasing circuit seen) and top side (with four patches and the through hole seen) are shown in Figure~\ref{fig:3} (d) and (e), accordingly. Elements numbering in Figures~\ref{fig:3}(d,e) corresponds to one in Figure~\ref{fig:3}(b).
\begin{figure*}
\center
\begin{minipage}{0.8\linewidth}
\center{\includegraphics[width=1.0\linewidth]{Fig3.pdf}}
\end{minipage}
\caption{Organization of the IR-controllable RIS block and its fabricated prototype: schematic of the IR remote (a) and schematic of the RIS building block with four individually controlled patches (b); prototype of the IR remote circuit with an LED transmitter (c) and prototype of the RIS block shown from the back side (DC biasing circuit elements are shown) (d) and the front side (four patches with a through hole are shown) (e).}
\label{fig:3}
\end{figure*}
In order to create an RIS of the given reflecting aperture dimensions, the corresponding number of identical blocks were manufactured and assembled together. All blocks in the assembly were powered from the same DC source.
Note that adjacent blocks should have their ground planes interconnected. Otherwise gaps between the blocks may affect the operation of the RIS due to resonant diffraction effects. Typically RISs have electrically large aperture dimensions, while RISs with small amount of patches do not make sense. Therefore, it can be recommended to compose the entire RIS from panels each one having at least $5\times 5$ identical proposed blocks. Such large panels can be assembled to a common RIS so that the edges of their PCBs stay close to each other without any electrical connection. The corresponding long gaps are non-resonant and will not affect the RIS operation. In the experiment we used four square $5\lambda\times5\lambda$ panels to compose a RIS with a total amount of $20\times20$ patches.
\subsection{Remote control and IR light delivery}
Each block receives IR light illuminating the entire aperture of the RIS and extracts its own control commands from the digital code with the information about the requested reflection states to be set for its four patches. Despite IR light propagation can be limited due to outdoor conditions, this method in conjunction with digital modulation for coding is beneficial for indoor applications. In the developed RIS prototype, each block had its unique address. An IR transmitter at a remote side sent commands to all blocks simultaneously followed after the corresponding addresses. As a controller of a particular block recognizes its address, it sets phases to four its patches according to the received commands by forming a set of biasing voltages using the DC drivers circuit.
A near-infrared (940 nm) LED was used as a transmitter of the IR remote. The IR communication channel used amplitude modulation and control commands produced a series code sequence. The carrier frequency was 38 kHz. The IR transmitter was controlled from a computer via the USB interface. The schematic organization of the IR remote is shown in Figure~\ref{fig:3}(a), while a photograph of its prototype is shown in Figure~\ref{fig:3}(c).
The phase control protocol used two bytes to form a control packet. First byte was the address field. The most significant bit (MSB) of this byte was always set to zero (an address marker). Thus, this protocol provided 128 unique addresses (from 0 to 127) which was more than enough to control 100 blocks (i.e. 400 patches).
The zero address was a broadcast address used to send phase settings to all blocks simultaneously. This significantly speeds up the setting of the array to all- zeros or all-ones state. The second byte of the packet contained phase settings for four patches of the particular block. The most significant bit of this byte was always set to one (a data marker). The phase states were encoded using bits 0 to 3 of the second byte. The presence of address (MSB set to 0) and data (MSB set to 1) markers allowed recovery in case of possible synchronization errors (due to byte loss or corruption).
Since the two required voltage levels corresponding to "0" and "1" states may require tuning, a special driver circuit was developed to equip each block of the RIS. It allowed changing the two voltage levels simply by changing the voltages of two external power supplies. The third 5V power supply was required for the microcontroller, IR receivers and additional light-emitting diodes (LEDs) used for indication of phase states. The driver consisted of a single field-effect transistor (FET), two Shottky diodes and three resistors. The FET worked as a switch that can be in either open (low resistance) or closed (high resistance) state. Depending on the FET state either one or the other power supply was connected to the driver's output via the corresponding resistor and Shottky diode.
A graphical user interface (GUI) program was developed to encode the phase over the RIS. Open source Lazarus IDE and FreePascal compiler were used. The program allowed to set the phases of individual patches manually or load them all from a file. It also had a test mode that allowed easy testing of all elements of the array.
\subsection{Near- and far-field phase coding}{\label{C}}
The developed RIS belongs to the class of 1-bit reflectarrays with full 2D phase coding capability. To create a desirable spatial distribution of the reflected field, the corresponding phases of individual patches are to be pre-calculated and encoded. The calculation of phases was realized in frames of the same program as one controlled the IR transmitter for sending commands to the RIS.
The following adaptive algorithm was realized to calculate the phases based on the specified distribution of the electric field. Both the far-field distribution (scattering pattern) and intermediate-field distribution (field pattern on a target plane parallel to the aperture) could be set as the input data. Before starting the adaptation, the RIS is set to the initial state when all patches are in the same "1" state. However, the algorithm could start from any arbitrary phase distribution. The use of a preliminary approximation usually did not influence on the final result, but could significantly speed up the adaptation process.
First, the field generated by the RIS in the target area was calculated using a simplified model. The model took into account the excitation field (created by either a plane wave or a point source), the distance from each RIS element to the observation point, the phase of the array element and its individual scattering pattern. All patches were assumed to have a Huygens-like scattering pattern given by the formula: $F(\theta)=\cos^2(\frac{\theta}{2})$ with $\theta$ being a polar angle measured from the normal to the aperture. If the target area is a plane parallel to the aperture, the reflected field $E_k^{\text{RIS}}$ created by the RIS at point $k$ was calculated as
\begin{equation}
E_k^{\text{RIS}} = \sum_{m=1}^{N_xN_y} E_{\text{inc}}^{m}e^{j\Phi_{\text{inc}}^m} \cdot |R_m| e^{j\Phi_{m}}\cdot e^{-jr_{mk}}\frac{1}{2}\left(1+\frac{z_{\text{plane}}}{r_{mk}}\right),
\end{equation}
where $E_{\text{inc}}^{m}$ and $\Phi_{\text{inc}}^m$ are the magnitude and phase of the incident field at the location of patch element $m$, $|R_m|$ -- magnitude of the local reflection coefficient at patch $m$, $\Phi_{m}$ –- reflection phase of patch $m$ ($-\pi/2$ for state "0" and $+\pi/2$ for state "1"), $r_{mk}$ –- distance between patch $m$ and point $k$, $z_{\text{plane}}$ – distance from the target plane to the aperture of the RIS, $N_x, N_y$ - number of patches along $x$- and $y$-axis respectively. In the case when the source is a plane wave, the phase of the incident field was found as:
\begin{equation}
\Phi_{\text{\text{inc}}}^m = k_0\left[x_{m}\cos(\phi_{\text{inc}})\sin(\theta_{\text{inc}})+y_{m}\sin(\phi_{\text{inc}})\sin(\theta_{\text{inc}})\right],
\end{equation}
where, $x_m$, $y_m$ are Cartesian coordinates of the center of patch element $m$, $\phi_{\text{inc}}$ and $\theta_{\text{inc}}$ -- azimut and polar incidence angles in the spherical coordinate system with a polar axis being normal to the RIS, $k_0=2\pi/\lambda$ -- wavenumber in free space.
Second, the difference between the desired and calculated field distributions was determined. The objective function that needed to be minimized was a mean squared error (MSE) between the two magnitude distributions created on the target plane.
Third, the phase state of one of the patches (selected at random) was flipped and the corresponding change in the objective function was calculated. If the objective function decreased the new phase state was kept, otherwise, the element was returned to its original state and the process repeated.
The criteria to stop the iteration process was the absence of changes in the phase distribution for a specified maximum number of iterations. It should be noted that the number of steps depends on the number of patches and desirable field distribution. For the experimental RIS with $N_x=N_y=20$ it was enough to limit the process with 50 iterations.
The above described objective function can be useful for maximizing the reflected power concentration in the chosen region of space. If this region is small enough, and the distance to the target plane is much larger than the dimensions of the RIS (i.e. $z_{\text{plane}} \gg N_x \lambda/2,N_y \lambda/2$), this process is close to simple beam steering. However, the target region may be large and its dimensions may compare to ones of the aperture. In that case, the process can be applied to holographic field synthesis. In the following section, we consider both types of adaptation made with the experimental RIS.
\section{Experimental results}
Measurements of reflected field distributions were made using a 3-axis field scanner installed in an anechoic chamber (7 m $\times$ 5 m $\times$ 3 m) and VNA Rohde$\&$Shwarz ZVB20. The first port of the VNA was connected to the source antenna with a vertical linear polarization ($E_y$) while the second one - to a field probe. The experimental setup for the case of beam steering is shown in Fig.~\ref{IRS_experimental}(a) with the source being a linearly polarized TEM-horn antenna at the distance of 4 meters from the RIS. In this case, for obtaining a scattering pattern, first a complex magnitude of magnetic field was measured over the plane parallel to the RIS at the distance of 10 mm from the patches in presence and absence of RIS. Then, Fourier transform was applied to calculate the far-field scattering pattern from the near-field map using standard methods \cite{Kobayashi}. In the case of holographic field synthesis, the field was measured directly on the target plane.
In both cases, SX-R 3-1 H-field Probe mounted to an arm of the scanner and mechanically manipulated in parallel to $XY$-plane was used. The complex magnitude of the measured field was extracted from $S_{12}$ values stored by the VNA.
\begin{figure*}[t]
\center
\begin{minipage}{1\linewidth}
\center{\includegraphics[width=1\linewidth]{Fig4_v2.pdf}}
\end{minipage}
\caption{Experimental investigation of the proposed RIS at 5.2 GHz: (a) setup for measurements of the reflected field pattern; (b) fabricated RIS composed of $N_x\times N_y=20\times 20$ patch elements shown from the front side (inset shows one building block with $2\times2$ patches); (c) back side of the RIS with electronic circuits and wire interconnections for powering building blocks from a common DC supply (inset shows one building block with $2\times2$ patches). Phase distributions with 1-bit coding and corresponding measured far-field patterns for the normal incidence ($\theta_{\text{inc}}=0$): (d-f) beam steering to scan angle: $\theta_{\text{ref}}=0^{\circ}$; and $\theta_{\text{ref}}=45^{\circ}$ (g-i). Phase distribution (j), desirable magnitude field distribution on the target plane $Z_{\text{plane}}=700$ mm away from the aperture, simulated (l) and measured (m) field distribution on the same plane.}
\label{IRS_experimental}
\end{figure*}
For the experimental characterization, we fabricated a full-sized prototype of the RIS consisting of $N_x\times N_y=20\times 20$ patch elements (i.e. $10\times10$ IR-controlled building blocks with individual controllers inside). For simplicity of fabrication all blocks were grouped into four identical square panels each one being a separately fabricated multilayer PCB. To support the panels and fix their edges close to each other, a special holder was made of 5-mm-thick Plexiglass. The PCBs were mounted to the holder with plastic screws. The prototype with full dimensions of 600 mm $\times$ 600 mm fixed on the holder is shown in Figure~\ref{IRS_experimental}(b) from the front side and in Figure~\ref{IRS_experimental}(c) from the back side. DC supply Rohde$\&$Shwarz HMP2030 was used to power the building blocks providing the required voltage levels.
Experimental investigation of the RIS in different regimes included the following tasks:
\begin{enumerate}
\item Measurements of complex reflection coefficients from the RIS with all patches being set to "0" or "1" state;
\item Indirect determination of far-field scattering patterns in the regime of reflected beam steering by calculating from measured near fields for different phase coding profiles corresponding to different scan angles of the reflected beam;
\item Direct measurements of intermediate-field distributions in the regime of holographic field synthesis on the target plane positioned at distance $Z_{\text{plane}}=700$ mm away from the RIS.
\item Measurements of the signal transmission between two antennas in the presence and absence the RIS.
\end{enumerate}
The results of the corresponding tasks are presented and discussed below in the corresponding subsections.
\subsection{Reflection coefficients}
Magnitudes and phases of the complex reflection coefficient of an incident plane wave from the RIS were measured for various biasing voltage levels applied to the diodes ranging from 2.8 V to 3.6 V with a step of 0.2 V. For each level, the same biasing voltage was applied to all diodes of the RIS (uniform phase distribution). Also the case of a zero biasing voltage was considered in the measurements. The magnitudes and phases were measured indirectly through near-fields \cite{Kobayashi} in the range of 4.9–5.5 GHz with a step of 10 MHz.
The required near-field maps were scanned within the area of 240 mm $\times$ 240 mm with a resolution of 10 mm. The indirectly calculated far-field complex magnitudes were obtained at the post-processing step as discussed in \cite{Kobayashi}. The scan region was chosen with smaller dimensions than ones of the aperture to reduce the effect of edge diffraction and make the obtained reflection coefficients closer to ones numerically calculated for an infinite 2D-periodic structure. At each probe position the magnitude and phase distributions of the $x$-component of the magnetic field were recorded. The same value in the absence of the RIS and in the case where the RIS was replaced with a copper plate of the same dimensions was also measured to determine a complex reflection coefficient as described e.g., in \cite{Johnson,Sayanskiy}. The effect of parasitic signal reflections between the RIS and the horn antenna was reduced using the time gating procedure \cite{Hock}.
The magnitude and phase of the measured reflection coefficient are compared with the simulation results for $U=0$ V and $U=3.2$ V in Figures~\ref{fig:Reflection_sim_meas_finite} (a) and (b) correspondingly. The phase difference between these two states is shown in Figure~\ref{fig:Reflection_sim_meas_finite}(c). As can be seen, at 5.2 GHz the phase difference between "0" and "1" states is precisely equal to $180^{\circ}$. Also, the magnitude levels of the reflection coefficient in both these states are almost equal as desired. The measured results are in good agreement with the simulated ones.
\subsection{Scattering patterns in reflected beam steering}
Far-field scattering patterns were measured at 5.2 GHz for the normal incidence, i.e. the wave from the distant horn impinged the RIS from the direction of the normal to its aperture. Depending on pre-calculated 1-bit phase profiles, different scan angles of an anomalously reflected TE-polarized beam were obtained in the horizontal plane. The uniform phase distribution with all patches set to "0" state and the corresponding measured scattering pattern in the horizontal plane with the main reflected beam in the normal direction are given in Figures~\ref{IRS_experimental}(d-e). The horizontal cut of the measured scattering pattern is shown in Cartesian coordinates versus angle $\theta$ in the horizontal plane with the blue dashed line. A projection of the
scattering pattern onto $XY$ plane (i.e. the 2D map of the scattering pattern in dB vs. coordinates $u=\cos(\phi)\sin(\theta)$, $v=\sin(\phi)\sin(\theta)$) is shown in Figure~\ref{IRS_experimental}(f). The same data for scan angle $\theta_{\text{ref}}=-45^{\circ}$ are given in Figure~\ref{IRS_experimental}(g-i). The measured maximum directivity was $D=31$ dBi for $\theta_{\text{ref}}=0^{\circ}$ and $D=24.1$ dBi for $\theta_{\text{ref}}=-45^{\circ}$.
The scattering pattern corresponding to $\theta_{\text{ref}}=-45^{\circ}$ exhibits a high phase quantization lobe. As a results, the reflected power is equally split between two beams at $\theta_{\text{ref}}=\pm45^{\circ}$, which is an effect of any 1-bit phase-coded RISs \cite{7480359} typically observed when a reflected beam scan angle differs from the mirror-reflected one.
For comparison, numerical simulations of the entire RIS illuminated with a normally incident plane wave were carried out using Time Domain Solver of CST Microwave Studio 2020. The simulated scattering patterns in the horizontal plane shown in Figures~\ref{IRS_experimental}(e,h) with solid red lines are in good agreement with the measured curves.
\subsection{Spatial distributions in holographic field synthesis}
In order to demonstrate the capability of the 1-bit phase coding in the regime of holographic field synthesis we aimed to create a certain magnitude distribution of magnetic field on the target plane $Z_{\text{plane}}=700$ mm away from the RIS. The desirable shape of the distribution was the shape of the capital letter "I" with the sizes of 200 mm $\times$ 200 mm (see Figure~\ref{IRS_experimental}(k)).
Using the iterative adaptation algorithm, described in Section \ref{C}, the phase distribution shown in Figure~\ref{IRS_experimental}(j) was computed. The simulated and measured H-field distributions on the target plane are shown in Figure~\ref{IRS_experimental} (l) and (m) correspondingly. The realized distribution had a blurred shape in comparison with the desirable one due to finite aperture dimensions of the RIS which limited the spatial resolution of the holographic phase synthesis. However, the experimental and simulated distributions resembled shape "I" and looked visually similar to each other.
\subsection{Reflected signal transmission between two antennas}
The aim of an RIS is to enhance the signal propagation between a base station and a user. To estimate the effect of the fabricated RIS in this practical scenario, we experimentally modeled a single-path wave propagation at 5.2 GHz between two horn antennas with a properly phase-encoded RIS acting as an intermediate reflector on the path. The transmission was characterised by comparing the levels of the signal transmitted between the horns in the presence of the RIS and in the absence of it. To satisfy the conditions of the far-field region of the RIS with respect to both horn antennas, in the measurements we used only one panel containing $10\times10$ patches. The experimental setup is schematically shown in Figure~\ref{Trans_improvement}(a). It consisted of two linearly polarized broadband horn antennas placed at distances $R_1=2.7$ and $R_2=4.2$ m from the RIS. As in previous measurements the horns created the main TE polarization with an electric field vector oriented along $y$-axis. Both horns were connected to ports of VNA Rohde$\&$Shwarz ZVB20 with calibrated 50-Ohm coaxial cables. The signal level was proportional to the magnitude of the $S_{12}$-parameter measured in the range of 5.0--5.3 GHz.
\begin{figure*}
\center
\begin{minipage}{1\linewidth}
\center{\includegraphics[width=1\linewidth]{Fig5_v2.pdf}}
\end{minipage}
\caption{To the experimental evaluation of the signal transmission in the presence/absence of the RIS: (a) scheme for measuring the signal transmission between two horn antennas; 1-bit
phase distribution for the normal incidence ($\theta_{\text{inc}}=0^{\circ}$) and reflection to angle $\theta_{\text{ref}}=45^{\circ}$ (b) and the corresponding magnitude of the transmission coefficient between two horn antennas vs. frequency (c); 1-bit phase distribution for oblique incidence ($\theta_{\text{inc}}=-15^{\circ}$) and reflection to angle $\theta_{\text{ref}}=30^{\circ}$ (d) and the corresponding magnitude of the transmission coefficient between two horn antennas vs. frequency (e).}
\label{Trans_improvement}
\end{figure*}
The experimental investigation included several scenarios with different incidence angles and reflection angles in the horizontal ($XZ$) plane realized with different positioning of the horns and RIS in the anechoic chamber. The following scenarios were compared:
\begin{enumerate}
\item RIS encoded to reflect a wave coming from a horn at angle $\theta_{\text{inc}}=0^{\circ}$ to a horn at angle $\theta_{\text{ref}}=45^{\circ}$;
\item RIS encoded to reflect a wave coming from a horn at angle $\theta_{\text{inc}}=-15^{\circ}$ to a horn at angle $\theta_{\text{ref}}=30^{\circ}$;
\item Mirror reflection with all patches set to "0" state;
\item Mirror reflection with all patches set to "1" state.
\end{enumerate}
The calculated 1-bit phase distributions for the first two (optimized transmission) scenarios are shown in Figure~\ref{Trans_improvement}(b,d), while the corresponding measured spectra of $S_{21}$ magnitude are given in Figure~\ref{Trans_improvement}(c,e) with red lines. The obtained transmission coefficient levels were found to be in a good comparison to theoretical predictions made using radar range equation (see Section 2.17 in \cite{Balanis}). The received power $P_{\text{r}}$ at horn 2 can be expressed through the transmit power $P_{\text{t}}$ applied to horn 1 as follows:
\begin{equation}
P_{\text{r}} = P_{\text{t}}\frac{\lambda^2L_xL_y\cos(\theta_{\text{inc}})(1-|S_{11}|^2)(1-|S_{22}|^2)}{(4\pi)^3R_1^2R_2^2})G_{\text{t}}G_{\text{r}}G_{\text{RIS}},
\end{equation}
where, $L_x=L_y=300$ mm are the dimensions of the RIS, $\lambda=57.7$ mm - wavelength in free space, $R_1=2.7$ m and $R_2=4.2$ m - distances from horns 1 and 2 to the RIS, $G_{t}=15$ dB - gain of the transmit horn, $G_{r}=15$ dB - gain of the receive horn, $G_{\text{RIS}}$ - gain of the RIS for the given incidence $\theta_{\text{inc}}$ and reflection $\theta_{\text{ref}}$ angles, $|S_{11}|=-20$ dB and $S_{22}=-20$ dB - return loss levels of the transmit and receive horns, respectively.
It should be noted that this expression is valid only for the case, in which both transmit and receive horns are located in the far-field region of the RIS.
Gain $G_{\text{RIS}}$ was estimated by using the magnitude of the reflection coefficient measurements (0.58 or -4.7 dB at 5.2 GHz) to find the power efficiency and the analytical estimation of the directivity value calculated from the measured scattering pattern. The directivity was estimated as $D=20.1$ dBi in the first scenario and $D=19.5$ dB in the second one. The magnitude of the theoretical transmission coefficient $|S_{21}|=\sqrt{P_{\text{r}}/P_{\text{t}}}$ is given with green dashed lines in Figures~\ref{Trans_improvement}(c,e). As as can be seen, the theoretical levels well coincide with the measured levels in the first two scenarios.
For a comparison the measured signal transmission levels are also shown in Figure~\ref{Trans_improvement}(c,e) for the third and fourth scenarios as well as for the case in which the RIS is absent. As follows from the comparison of the results, the presence of the properly phase-encoded RIS improves the signal transmission between the horns by 28.5 and 25.5 dB in the first and second scenario, respectively. In the third and fourth scenario the RIS improves the signal only by less than 12 dB because of a uniform phase distribution and non-optimal mirror reflection behavior.
\section{Conclusion}
A new approach to design an optically-controlled RIS was proposed and demonstrated. In contrast to previous RISs, identical building blocks containing four identical patches with an embedded individual microcontroller were used. Each block can operate alone or together with multiple similar blocks forming a reflective aperture which is scalable in its dimensions by changing the amount of blocks. Each block can be distantly controlled via IR digital code reacting only to its own commands. The capability of full 2D phase encoding was experimentally demonstrated in the Wi-Fi 5-GHz range in two cases: reflected beam steering in the horizontal plane and holographic field synthesis. Experimental results are in good comparison with analytical predictions and numerical simulations.
Therefore, the possibility of a distant and robust control of a scalable RIS via digital infrared code was shown to be promising for indoor wireless communications.
\section{Acknowledgements}
This work was done in cooperation with Moscow Research Center, Huawei Technologies Co. Ltd., Russia.
The work was supported in part of waveguide measurements by the Russian Science Foundation (Project No. 21-79-30038).
|
\section{Introduction}
This is a special issue of the journal SYMMETRY, devoted to the topic of TACHYONS. I am a theoretical physicist (a rather old one) and I have been invited to write a paper for this volume. So, I feel obliged to start out addressing the question, "What is a Tachyon?" \footnote{At an early stage of this project I asked the editors if they would start the volume off with their own discussion of this question, offering a suggestion on how that might be done. Their response was rather empty; and so I am undertaking this task myself.}
Commonly stated, "Tachyons" are particles that travel at speeds greater than the speed of light.
Commonly believed, Einstein's Theory of Relativity said that particles could never travel at speeds greater than the speed of light.
Physics is a human enterprise composed of people who do actual measurements of physical systems (simple or complex) who are called "experimentalists"; and also people, called "theorists", who draw mathematical frameworks that they can attribute to experimental observations.
The overall consensus of physicists is that there has been no credible observation of anything traveling faster than the speed of light (in vacuum). Nevertheless, some have been enticed to consider that such things may exist. This is not a new field of speculation, it is not a popular field of speculation, but it is a lively topic for some few theorists.
Let me start by clarifying "what Einstein said." Without searching back into all the writings of that great physicist, I will state what I believe to be the strong consensus among today's physicists. According to Einstein's Special Theory of Relativity, any object (huge as a sun or tiny as an atom) that can be found at rest (by us observers) can also be seen as moving with some velocity v but there is a limit: v is always less than c, where c is the speed of light. Light itself always travels at exactly that speed c, never more never less.. (We are looking at motion of objects in otherwise empty space, with no external interactions, and not any gravity.) The mathematical formulation of the Special Theory of Relativity allows the possibility that there might be some other type of particles that would always travel at speeds greater than c; and such things are named "tachyons".
To get some more precise definition of the idea "tachyon", I would start by appealing to the broad concept of SYMMETRY, which is the title of this journal. Among the profession of particle physicists, I would assert that the best definition of the concept "particle" is associated with the mathematical prospect of finding the "irreducible representations of the Poincar\'e group."
What is that? Wikipedia \cite{Wi} says, ``The Poincar\'e group, named after Henri Poincar\'e (1906),was first defined by Hermann Minkowski (1908) as the group of Minkowski spacetime isometries. It is a ten-dimensional non-abelian Lie group that is of importance as a model in our understanding the most basic fundamentals of physics.'' That group is a mathematical formalism that expresses a set of physical symmetries. We have, following Einstein, a world of four dimensions (three dimensions of space plus time) with the following symmetries for the vacuum (assumed empty space in which we may see free particles of all kinds and their interactions). Symmetry of translation, mathematically generated by the four operators $P^\mu = (P^0, P^1, P^2, P^3) = (E, \textbf{p})$ and called "momentum"; symmetry of rotation, mathematically generated by the three operators $J_i $ and called angular momentum; symmetry of transition to a moving reference frame, mathematically generated by the three operators $K_i$ and called boosts. Those six operators, for rotations and boosts, comprise the Lorentz group, a subgroup of the Poincar\'e group.
The mathematical analysis of this Poincar\'e group leads first to the recognition that the four momentum operators $P^\mu$ all commute with one another and this leads to constructing plane waves as the proper basis for any free particle. The angular momentum operators and the boost operators may mix the components of $P^\mu$ but there is one simple combination that is unchanged (invariant) under those operations: it is written as $P^\mu P_\mu$
The idea of "irreducible representations" means that any physical system, viewed as moving freely in otherwise empty space-time, must be characterized by a constant value of $P^\mu P_\mu = E^2 - p^2$. That constant is a number associated with this particular physical system (and the word "mass" comes in here); and we see three distinct categories. These may be characterized by the range of velocities, where we recognize for any free particle $v = p/E$. (Our units have the velocity of light c = 1.)
\begin{eqnarray}
P^\mu P_\mu >0, ordinary\; particles\; with\;v < 1, \\
P^\mu P_\mu = 0, particles\; of\;light \;(photons)\; with \;v = 1, \\
P^\mu P_\mu<0, some\;other \;type\;of\;particle \; with \;v > 1.
\end{eqnarray}
The word "tachyon" is commonly attached to this third category of particles. That is our first step in defining "tachyon". However, when we look at how various theoretical studies have proceeded, it becomes clear that there are more than one set of mathematical rules being taken to define what is a "tachyon." This distinction is particularly vivid in working with a quantum theory of tachyon fields. I would offer two classifications of such theories.
{\bf {Def. A}} Tachyon is a particle (or field) that always travels (propagates) outside of the light cone ($v > 1$).
{\bf {Def. B}} Tachyon is a particle that may be derived from a wave equation (like Klein-Gordon or Dirac) where the usual mass parameter is replaced by an imaginary number.
These two definitions share some domain of overlap; but they can lead to very different physical theories. My own work is within the definition {\bf A}. There is a large body of work by other authors who start with the definition {\bf B}, follow traditional lines of going from classical physics to quantum physics, and then find it necessary to break the symmetry of the Lorentz group. My own work, retains strict adherence to the full Lorentz and Poincar\'e groups; thus there is substantial conflict between these two arenas of tachyon theory. If you want to see some of the clash between these two sets of theories, I offer this transcript of letters related to the peer review of my most recent paper. \url{https://www.ocf.berkeley.edu/~schwrtz/TachyonDebate.pdf}
To illustrate the pivotal question of how one should go from accepted theory of ordinary particles to a theory of tachyons, here is a simple example which is basic to any physicist: How do you count the number of particles? For ordinary particles, especially in their rest frame, the answer is: Count the number of particles $N$ in this particular Volume V at time t.
Mathematically, and working within the Special Theory of Relativity, we assert that there is vector current density $j^\mu (x) = j^\mu(\textbf{x},t)$ that is locally conserved:
\begin{equation}
\partial_\mu j ^\mu (x) = \partial /\partial t\;j^0 +\sum_{i=1}^3\partial/ \partial x^i\; j^i = 0.
\end{equation}
Then we integrate this equation over some chosen volume ${\cal{V}}$ in four-dimensional space-time.
For ordinary matter (that which travels at speeds less than c) we find it convenient to choose that volume as bounded between the two times $t_1 < t_2$, and then also bounded by a space-like surface S. Using Gauss's Theorem, the resulting identity is
\begin{equation}
\int d^3x\; j^0(\textbf{x}, t_2) - \int d^3x\; j^0 (\textbf{x}, t_1) = 0.
\end{equation}
Here is the definition of the number of particles $N(t)$ and it is independent of time. In getting this result we have dropped any terms that sit on that other surface S. This is because we assume that the particles (or the field) is confined to some finite domain of 3-space at time $t_1$; and then, allowed to propagate only at speeds limited by the speed of light, it will be confined to a finite domain of 3-space at all times up to $t_2$. We locate that other surface S far enough out so that there is no matter there at any time of interest.
Now we move on to consider tachyons. These are particles whose velocities are unlimited; so the above construction is false. We still have a locally conserved vector current density; and need to find a different arrangement for the integral in 4-space.\cite{CS7} It turns out that the proper way to count tachyons is to ask: How many particles flow through a fixed area in 3-space, integrating over time. For example,
\begin{equation}
N(z) = \int dt \int dx \int dy\; j^3 (x,y,z,t);
\end{equation}
and this will be independent of the value of z.
I noted earlier that a major difference between my work on tachyons and that of numerous other authors involves how one goes from classical physics to quantum mechanics. The textbook answer, called "canonical quantization", is the path used by those other authors, following {\bf {Def. B}}. They find problems dealing with negative frequency solutions of the wave equations; and their choice is to abandon the symmetry of Lorentz transformations. The interested reader will undoubtedly find that line of work extensively presented in articles expected in this Special Issue of the journal Symmetry.\cite{ST} In my own work (see reference \cite{CS2}) I discussed this matter in detail (it is analogous to the discussion above about N), I rejected "canonical quantization" and found an alternative route to a quantized field theory: simply relying upon {\bf {Def. A}}.
In response to some reviewers' comments, I shall briefly repeat, here, those arguments against using "canonical quantization" for tachyon field theory. The word "canonical" means following the structure of Hamiltonian dynamics in classical physics. The philosophical basis of that theory is to think that we can follow the evolution of any system as time evolves. That way to go from classical to quantum mechanics was first formulated for non-relativistic particle theory and then for relativistic field theory. We are familiar with the Hamiltonian H as the operator that generates time evolution of the physical system. As long as we limit our considerations to fields that propagate within (or on) the light-cone, we can have a well defined system contained in some finite region of space. (This is like the discussion, above, about the number of particles.) However, if we now consider tachyonic fields, this is no longer true. The mathematical question, similar to what I have detailed above, is how to go from a locally conserved energy-momentum tensor, $\partial_\mu T^{\mu \nu} = 0$, to a definition of the time-independent energy (H=E).
I have another way of distinguishing that other line of tachyon theory from my own work: It seems that the objective of those other authors is to take some equation that looks like it implies faster-than-light particles (e.g., start with {\bf {Def. B}}) and then manipulate the mathematics to get rid of troubles that they see and thereby restore the familiar behavior of slower-than-light particles, thus negating {\bf {Def. A}}. A recent paper in the journal Particles \cite{LN} is explicit about that mission. My own mission, I can state clearly, is to start with {\bf{Def. A}}, stay within the strict confines of traditional theories of Relativity in both Classical and Quantum Mechanics, and see where careful exploration may lead. This has been quite successful, as readers will see below. but much remains to be studied.
Ultimately, it is experimental data that will determine the validity or irrelevance of any theoretical offerings. The experiment most currently watched is called KATRIN. This measures the energy spectrum of electrons from the beta-decay of Tritium and may set some limits on neutrino mass; and may possible distinguish among current theories. Otherwise, it is hard to see experiments that will directly say how fast low energy neutrinos move.
There are, however, other types of experimental observations that may be cited as justifying, or contradicting, certain theories about neutrinos as tachyons. My own work has found two such rewarding situations. In my 2017 paper \cite{CS3} I reported that my theory of gravitation produced by a sea of tachyons could explain - qualitatively and quantitatively - the famous phenomenon known as Dark Energy. In my 2022 paper \cite{CS6} I came up with a model of higher spin tachyons that could explain - in rough quantitative agreement - the known mass hierarchy of the three types of neutrinos. My search for a detailed model of how tachyon neutrinos might provide an explanation for the phenomenon known as Dark Matter remains an open question.
What follows is a selection of my writings about tachyons. Section 2, which comes from a talk given at a physics conference in 2019, is meant to briefly summarize my work, over the last decade, to construct a consistent mathematical theory for the idea that tachyons - faster then light particles - might exist in this universe. The physics here focuses on neutrinos and their role in Gravitation and Cosmology. Section 3 is a polemical note entitled "Debunking the Anti-Tachyon Myths"; and it is addressed to a wide arena of physicists and others who believe that tachyons could not exist.
\section{If Neutrinos are Tachyons, we may explain Dark Energy and Dark Matter }
What this section covers is a slightly edited version of a talk given at the Martin B. Halpern Memorial Conference, UC Berkeley, March 30, 2019. This is a quick survey of my previous research publications on a theory of tachyons. The several subsections are titled, OPERA, Classical Tachyons, Negative Energy States, Energy-Momentum Tensor I (DM), Energy-Momentum Tensor II (DE), Quantized Dirac Tachyon Field, Future Work.
\subsection{OPERA}
Start with the OPERA experiment of 2011.
20 GeV neutrinos appeared to travel faster than light by about 1/40,000.
From the formula $E = mc^2/\sqrt{v^2/c^2 -1}$,
\begin{equation}
(v-c)(v+c)/c^2 = (mc^2/E)^2.
\end{equation}
So that would imply a tachyon neutrino with a mass of about 100 MeV. But we are pretty sure that neutrino mass is around 0.1 eV. So the OPERA result going away tells us nothing about possible neutrino tachyons of such a low mass. That is what we are considering.
For a comprehensive review of experimental data linked to the idea of tachyons, I recommend the article by Robert Ehrlich in this Special Issue. \cite{RE}
\subsection{Classical Tachyons}
Classical particles are described by a worldline $\xi^\mu(\tau)$ with a Lorentz invariant form that leads us to note three categories:
\begin{eqnarray}
\eta_{\mu \nu}\;\dot{\xi}^\mu\; \dot{\xi}^\nu = \epsilon \;\;\;\;\;\;\;\;\; \\
ordinary \; particles \; (v < c): &&\epsilon = +1 \nonumber\\
massless \; particles\; (v=c): &&\epsilon = 0 \nonumber\\
tachyons\; (v > c): &&\epsilon = -1 \nonumber
\end{eqnarray}
We usually define the 4-vector $p^\mu = m \dot{\xi}^\mu$. For tachyons this is a space-like vector and so one asks, What about negative energy states?
\subsection{Negative Energy States}
Let's look at a space-time diagram for a general interaction process. Figure 1 shows four particles involved in the reaction, $n \rightarrow p + e + \nu$, where I imagine that the neutrino is a tachyon.
\begin{picture}(100,200)
\thicklines
\put(50,100){\circle*{10}}
\put(55,100){\line(1,1){60}}
\put(55,100){\line(1,-1){60}}
\put(45,100){\line(-1,1){60}}
\put(45,100){\line(-1,-1){60}}
\thinlines
\put(50,105){\line(0,1){40}}
\put(50,95){\line(0,-1){40}}
\put(50,100){\line(-1,2){20}}
\put(55,100){\line(5,1){40}}
\put(50,150){\shortstack{p}}
\put(48,45){\shortstack{n}}
\put(25,145){\shortstack{e}}
\put(97,105){\shortstack{$\nu$}}
\put(0,20){\shortstack{Figure 1. Reaction with an outgoing tachyon.}}
\end{picture}
Now look at Figure 2. Is this a picture of the reaction $ n \rightarrow p + e + \nu$ with the neutrino carrying off negative energy; or is this a picture of the reaction $n + \nu \rightarrow p + e$ with positive energy for all participants?
\begin{picture}(100,200)
\thicklines
\put(50,100){\circle*{10}}
\put(55,100){\line(1,1){60}}
\put(55,100){\line(1,-1){60}}
\put(45,100){\line(-1,1){60}}
\put(45,100){\line(-1,-1){60}}
\thinlines
\put(50,105){\line(0,1){40}}
\put(50,95){\line(0,-1){40}}
\put(50,100){\line(-1,2){20}}
\put(55,100){\line(5,-1){40}}
\put(50,150){\shortstack{p}}
\put(48,45){\shortstack{n}}
\put(25,145){\shortstack{e}}
\put(97,90){\shortstack{$\nu$}}
\put(0,20){\shortstack{Figure 2. Reaction with an incoming tachyon.}}
\end{picture}
This "problem" is the same as noting that the mass-shell equation,
\begin{equation}
p^\mu \; p_\mu = E^2 - p^2 = \pm m^2, \label{b1}
\end{equation}
gives us two separate hyperboloids for ordinary particles (plus sign) but a single hyperboloid for tachyons (minus sign). For ordinary particles, we manage to reinterpret the negative energy solutions as antiparticles and give them positive energy. For tachyons we need to see what it means when we look at a positive energy particle from a different Lorentz frame, where it appears to have negative energy.
Look again at the space-time diagrams above. A "positive energy" tachyon will have $dt/d\tau > 0$ and its trajectory will be seen moving upward - as in Figure 1 - and so we would call that an outward going particle if it sits above the interaction region in time; and we would call it an inward moving particle if the trajectory sits below the interaction region - as in Figure 2. But from another reference frame we may have $dt/d\tau < 0$ for what was formerly an outgoing particle and so this will now look like an incoming particle.
The lesson is that \emph{the labels "in" and "out" are Lorentz invariant for ordinary particles but NOT for tachyons.} Does this matter? No. The physical law which we call the conservation of total energy and momentum is written,
\begin{equation}
\sum_{out} p^\mu _j - \sum_{in} p^\mu_i = 0.\label{b2}
\end{equation}
This is true in any Lorentz frame, even though the individual terms in this equation each transform. In going from Figure 1 to Figure 2 we need only move one $p^\mu$ from the "out" summation to the "in" summation.
\subsection{Energy-Momentum Tensor I (DM)}
Now we look at tachyons in General Relativity.
For ordinary particles we write the Energy-Momentum tensor as,
\begin{equation}
T^{\mu \nu} (x) = m\; \int d\tau\; \dot{\xi}^\mu\; \dot{\xi}^\nu \; \delta^4 (x -\xi(\tau)),
\end{equation}
and, at first, I used this same formula for tachyons.
For a free particle we have the familiar representation,
\begin{equation}
\xi^\mu (\tau)= (\gamma \tau, \gamma \textbf{v}\;\tau), \;\;\;\;\; \gamma = 1/\sqrt{|1-v^2/c^2|}.\end{equation}
Thus, for a very low energy ordinary particle ($v << c$) the dominant term is
\begin{equation}
T^{00} (x)= m \delta^3(\textbf{x} - \textbf{v} t).
\end{equation}
and for a very low energy tachyon ($v >> c$) the dominant terms are
\begin{equation}
T^{i j} (x)= m \gamma v_i\; v_j\; \delta^3 (\textbf{x} - \textbf{v} t).
\end{equation}
These space-components of $T ^{\mu \nu}$ can be very large; and this leads to interesting cosmological modeling.
In my 2011 paper I showed that this leads to attractive forces among co-linear flows of tachyons; and I predicted the possibility of such tachyon "ropes" becoming localized, say, within a galaxy, and creating strong local gravitational fields that could produce the observational effects now ascribed to Dark Matter.
Then, last year, I was led to revise that theory, as follows.
\subsection{Energy-Momentum Tensor II (DE)}
Let's derive that formula for tachyons' energy-momentum tensor from some general principle. General Relativity may be constructed from an action principle with a Lagrangian density that looks like this.
\begin{eqnarray}
{\cal{L}} = \sum m\int d\tau\; \sqrt{\epsilon g_{\mu \nu}(x)\dot{\xi}^\mu\; \dot{\xi}^\nu }\delta^4 (x-\xi(\tau)) + \nonumber \\
\frac{ \sqrt{|det g|}}{8\pi G}\;R\;\;\;\;\;,
\end{eqnarray}
where we sum over all the particles and $R$ is the scalar form of the Riemann curvature tensor, which depends on the metric tensor $g_{\mu \nu}(x)$. Notice that I put an epsilon under the square root to make sure it would always be real, whether we have an ordinary particle or a tachyon.
GR textbooks show how, under variation of the metric $g_{\mu \nu}$, we get exactly Einstein's equation, with the energy momentum tensor, as we wrote it earlier, on the right hand side. Note, however, that the factor epsilon should be sitting there:
\begin{equation}
T^{\mu \nu} (x) = \epsilon m\; \int d\tau\; \dot{\xi}^\mu\; \dot{\xi}^\nu \; \delta^4 (x -\xi(\tau)).
\end{equation}
This new minus sign leads to the unconventional result of repulsive gravitational forces; and now we predict that tachyons, if they exist in the cosmos, would form a pervasive gas and contribute a \emph{negative pressure} when we consider the Robertson-Walker model for the universe. The formula for this pressure is simply,
\begin{equation}
p = -m \gamma \frac{v^2}{3} \rho,
\end{equation}
where $\rho$ is the density of those tachyons, $v$ is their velocity, and $\gamma = 1/\sqrt{v^2/c^2 -1}$.
Using standard numbers for the density and energy of the Cosmic Neutrino Background, assuming that they could be tachyons with a mass around $ 0.1 \;eV/c^2$, gives a numerical value for \textbf{this negative pressure that is within a factor of two of explaining what is commonly called Dark Energy.}
\vskip 1cm
Thus far, considering cosmic neutrinos as tachyons in General Relativity:
ACT I - this may explain Dark Matter
alternatively,
ACT II - this may explain Dark Energy
... waiting for Act III
\subsection{Quantized Dirac Tachyon Field}
Let's look at a quantized field theory for a Dirac tachyon.
\begin{equation}
\psi(x) = \int_{-\infty}^{\infty} d\omega \int d^2 \hat{k}\;k^2\; e^{i(\textbf{k}\cdot\textbf{x} - \omega t)} \sum_{h=\pm 1}\;b_h(\omega,\hat{k})\; v_h(\omega,\hat{k})
\end{equation}
where $k = \sqrt{\omega^2 + m^2}$, $\textbf{k} = k\;\hat{k}$, and $v_h$ is a Dirac spinor for a tachyon (put "im" instead of "m" in the Dirac wave equation), with helicity h. Note that the integral over frequencies extends over all real numbers. I do not separate positive and negative frequency solutions, since a Lorentz transformation of tachyon states readily mixes those two domains. (Please refer back to Section 1 of this paper where I discuss, and criticize, the work of other theorists who do make that separation, and break Lorentz invariance.)
The operator algebra is chosen to be,
\begin{equation}
[b_h(\omega, \hat{k}), b^\dagger _{h' }(\omega ', \hat{k}')]_+ = \delta(\omega - \omega') \delta^2(\hat{k} -\hat{k}')/k^2\; \delta_{h,h'};
\end{equation}
and this gives us the two-point anti-commutator for the fields,
\begin{equation}
[\psi(x),\psi^\dagger(x')]_+ = 0, \;\;\; if\;\;\; |t-t'| > |\textbf{x} - \textbf{x}'|.
\end{equation}
This is the strong condition of causality for tachyon fields: \emph{No signal can travel slower than the speed of light.}
Next, we look at one-particle states and then examine how to calculate the energy-momentum tensor.
What we note is that the helicity h serves to signify what we call the particle and what we call the anti-particle. (For ordinary particles this job was given to the sign of the frequency.) Working from the vacuum state $|0>$ we define:
\begin{equation}
b_{h=+1} (\omega, \hat{k}) |0> = 0, \;\;\; b^\dagger_{h=-1} (\omega, \hat{k}) |0> = 0.
\end{equation}
Then construct one-particle states,
\begin{eqnarray}
|\omega, \hat{k},h=+1> = \sqrt{k} \;b^\dagger_{h=+1}(\omega, \hat{k}) |0> , \\
|\omega, \hat{k},h=-1> = \sqrt{k} \;b_{h=-1}(\omega, \hat{k}) |0> .
\end{eqnarray}
The conserved energy-momentum tensor for the tachyon Dirac field is,
\begin{equation}
T^{\mu \nu} (x) = (i/4)\psi^\dagger (x)\gamma^0 \gamma^5 (\gamma^\mu \stackrel{\leftrightarrow}{\partial}^\nu + \gamma^\nu \stackrel{\leftrightarrow}{\partial}^\mu) \psi(x),
\end{equation}
where $\stackrel{\leftrightarrow}{\partial}=\stackrel{\rightarrow}{\partial} - \stackrel{\leftarrow}{\partial}$.
I want to put this operator between one-particle states; but first there are two important steps. First we embrace this operator with the notation $: ... :$ which means "normal order", to guarantee that this will have zero expectation value in the vacuum state.
The second step is to acknowledge that we need a specific indefinite metric when taking matrix elements between one-particle states of spin one-half tachyons. This comes from investigating the "Little Group" $O(2,1)$ appropriate for tachyons in building a unitary representation of the Lorentz group. This indefinite metric $H$ is simply the helicity operator.
The result of this calculation is,
\begin{equation}
<\omega, \hat{k},h|H :T^{\mu \nu}: |\omega, \hat{k},h> = h\; k^\mu\;k^\nu\;.
\end{equation}
This says that particle and anti-particle contribute to the energy-momentum tensor with \emph{opposite} signs. \textbf{Thus we see the possibility to explain both Dark Matter and Dark Energy with the proposition that Cosmic Background Neutrinos are tachyons, with a mass in the neighborhood of $0.1\; eV/c^2$.}
\subsection{Future Work}
{\bf More Work that Needs to be Done}
\underline{Questions about previous results/claims}
* Should chirality, rather than helicity, define particle vs anti-particle for tachyons?
* Is $k^2 \;d\omega\;d^2\hat{k} $ the correct density of states formula for tachyons?
* Can we show that gravity works to localize tachyon streams?
\underline{New areas to be explored}
* Fitting of tachyon neutrinos into the Standard Model of particles.
* Mass mixing for tachyon neutrinos
* Revise Standard Cosmology Theory for tachyon neutrinos (many sub-topics)
* New experiments to detect low energy tachyon neutrinos
\vskip 1cm
\underline{Summary of Cosmology results - Scaling of Energy and Pressure Components}
Radiation: $\rho\sim\;a^{-4}, \;\;\;\;\; p \sim\;a^{-4}$
Cold Matter: $\rho\sim\;a^{-3}, \;\;\;\;\; p \sim\;0$
Cosmo Const: $\rho\sim\;a^{0}, \;\;\;\;\; p \sim\;a^{0}$
Cold Tachyon Neutrinos: $\rho\sim\;0, \;\;\;\;\; p \sim\;a^{-1.5}\;\;$
or $\;\; p \sim\;a^{-2}$
\vskip 1cm
\underline{Something we can check now?}
If low energy tachyon neutrinos cause Dark Matter effects, those effects should be lessened in the past when their Temperature was near or above 0.1 eV. Is there data on Dark Matter effects as a function of z?
\vskip 1cm
For slow particles under Newton's Gravity,
\begin{equation}
H = \sum_a \frac{1}{2} m_a v_a^2 - \sum_{a < b} \frac{Gm_a m_b}{ |\textbf{x}_a - \textbf{x}_b|}. \nonumber
\end{equation}
And from Einstein's GR we now get,
\begin{eqnarray}
H = -\sum_a \omega_a m_a \gamma_a - \sum_{a , b} \frac{G(\omega_a m_a \gamma_a )(\omega_b m_b \gamma_b)}{ |\textbf{x}_a - \textbf{x}_b|} Z_{ab} ,\;\; \nonumber\\
Z_{ab}= 2 - 4 \textbf{v}_a \cdot \textbf{v}_b + (v_a^2 + v_b^2) - (\epsilon_a \gamma_a^2 + \epsilon_b \gamma_b^2 + 1)\times \; \nonumber \\
\;\times [(1-\textbf{v}_a \cdot \textbf{v}_b)^2 - \frac{1}{2}(1-v_a^2)(1-v_b^2)]. \;\;\nonumber
\end{eqnarray}
For a static system, including tachyons.
\vskip 1cm
"For every complex problem there is a simple solution. And it's always wrong."
--H. L. Mencken
\section {Debunking the Anti-Tachyon Myths}
Some people believe that the Earth is flat; and they can see this is true with their own eyes. Physicists are convinced otherwise; and they can cite abundant evidence on their side.
However, when it comes to Tachyons (faster-than-light particles), a great many physicists believe that they do not, and some will say that they can not, exist; and a number of reasons are recited in support that that prejudice.
In my several papers exploring mathematical frameworks for how Tachyons might fit into physical theory and experiments I have taken the trouble to lay out careful reasoning to debunk those prejudices. Do I have to review all those arguments in every new paper I write? This note is posted to serve that purpose.
Firstly, all my work is done strictly within the established mathematical frameworks of Special and General Relativity. (Some other authors have violated those bounds.)
The argument most commonly heard against the existence of tachyons (anything that travels faster than light) is this: If tachyons existed, one could, in principle, send a signal into the past and this would lead to an unbearable logical paradox. (This so-called paradox is sometimes named the "antitelephone.")
This claim has been debunked some time ago \cite{ER}; and here I shall give my own critique.
\vskip 0.5cm
\textbf{The phrase "send a signal into the past" is utter nonsense. Said more professionally, this phrase is an oxymoron, meaning that it is self-contradictory in terms of the language used.}
\vskip 0.5cm
Let me show this by means of some diagrams that are basic tools in connecting the mathematical principles of the Special Theory of Relativity to the way we physicists talk. The picture below is a space-time diagram that shows two events, labeled A and B, each with a worldline showing the trajectory of a particle (or a signal) ending at that event. The picture at A is described by the word "send" and this means that the particle moves farther and farther away from the location of event A, as time moves forward. The picture at B is described by the word "receive" and this means that the particle moves closer and closer to the location of event B, as time moves forward. Thus the words "send" and "receive" each have within their definition (at least as sensible physicists speak) a specific sense of time evolution. The phrase "send a signal into the past" explicitly violates this definition of the word "send."
\begin{picture}(200,100)
\put(10,10){\vector(0,1){20}}
\put(10,32){t}
\put(10,10){\vector(1,0){20}}
\put(32,10){x}
\put(50,30) {\circle*{5}}
\put(54,26) {A}
\put(50,30){\line(1,3){15}}
\put(130,60){\circle*{5}}
\put(134,58){B}
\put(130,60){\line(1,-3){15}}
\end{picture}
Please note that I have not yet said anything about tachyons. This comes next.
If the particle being sent or received is an ordinary particle (or a photon), then the trajectory will lie within (or on) the lightcone centered at the endpoint A or B. This property will remain the same when either process is viewed from a different Lorentz frame. Thus, for ordinary particles or light we can say that the assignment of the word "send" or "receive" is invariant under (orthochronous) Lorentz transformations. For tachyons, however, the particle trajectory lies outside the lightcone; and this means that a Lorentz transformation can change the appearance of the process. The experiment at A (sending the particle) may be seen by observers in a different Lorentz frame as a version of experiment B (receiving the particle).
Thus, one way to debunk the "antitelephone" paradox is to say that you have misused Special Relativity by tieing together two observations that are actually made in different Lorentz frames and claiming this is physical reality. It is not. The error is to assume that the word "send" always has a Lorentz invariant meaning, when, in fact, that is true only for a select set of particles, excluding tachyons. ( There is some similarity between this discussion about "send" and "receive" and the discussion in Section 2.3 about "in states" and "out states".)
It may seem strange that, as one considers viewing the experiment at A from a sequence of Lorentz frames that move faster and faster away from the original frame, an experiment that was once termed "send" becomes instantaneously transformed into "receive". Next, I note a way to remove this annoyance.
In my 2011 JMP paper, \cite{CS1} Appendix A looks at a scenario of sending tachyon signals between earth and a distant rocket ship, alleging \emph{a causal paradox}. It is argued that an exchange of tachyon signals can lead to a response arriving before the original message was sent out. Simply replacing the point particle by a wave packet shows that, when one carries out the relevant Lorentz transformation, the distinction between sending (emitting) and receiving (absorbing) a tachyon can disappear in a continuous manner.
In my 2016 IJMPA paper,\cite{CS2} I state the appropriate \emph{principle of causality} for tachyons - no propagation slower than the speed of light; and this leads to a consistent mathematical formalism for quantizing such fields. This provides an alternative to the canonical formalism, which is wrong for tachyon fields.
In my 2018 IJMPA paper,\cite{CS4} Section 2 examines the role of tachyons engaged in a general multi-particle interaction. The common idea that \emph{negative energy states} imply physical instability of the system is debunked by recognizing that the naming of $in$ and $out$ states is not Lorentz invariant. The total energy and momentum are still conserved.
In my 2016 paper \cite{CS2} on quantizing tachyon fields, especially for the spin 1/2 (Dirac) case, I deal with the \emph{Little Group O(2,1)} by introducing an indefinite metric (the helicity) into the Fock space.
Then there are experiments, a number of which over the years have claimed to observe neutrinos as tachyons, and then been revised to the opposite conclusion. The 2011 \emph{OPERA experiment} looked at 20 GeV neutrinos and first reported that they travelled faster than light by 1 part in 40,000. That would imply a tachyon mass of about 100 MeV. But we know that neutrino mass is around 0.1 eV; and the excess velocity (v-c) goes as the square of the mass-to-energy ratio. That puts us 14 orders of magnitude below the original (wrong) observation.
Finally, there are theoretical efforts to derive the existence of known particles from some abstract field with complicated self-interactions. The simplest model is a scalar field with a potential that looks like W. If one expands around the central peak, then the resulting particles are found to be tachyons (negative mass-squared). But then one recognizes that those states are unstable; one should instead expand about the minima of W, where one gets ordinary particles. I am not involved in that sort of theorizing.
I start with the question: If tachyons do exist, how would we describe them within our customary mathematical frameworks? The starting point is the relativistically invariant form for any 4-vector (e.g., the energy-momentum of a particle): \newline $p^\mu p_\mu = constant$. That constant may be positive, zero, or negative.
A recent criticism of tachyon-neutrinos is this. Take the Lagrangian density for an ordinary Dirac field,
\begin{equation}
\bar{\psi} [i\gamma^\mu \partial_\mu -m]\psi, \;\;\; \bar{\psi} = \psi^\dagger \gamma^0
\end{equation}
and replace $m$ by $im$. The result is non-Hermitian, so the theory derived from this will violate unitarity - conservation of total probability. But that is not how I do it. I insert $im$ in the Dirac differential equation, derive the equation for the adjoint field $\psi^\dagger$ from this, and then construct a Lagrangian density that will produce both of those equations of motion. The result is,
\begin{equation}
\bar{\psi} \gamma_5 [i\gamma^\mu \partial_\mu -im]\psi.
\end{equation}
With the $\gamma_5$ inserted this is all Hermitian. So there is no problem here.
I have also noted earlier that the conserved current and energy-momentum tensors are,
\begin{eqnarray}
j^\mu = \bar{\psi} \gamma_5 \gamma^\mu \psi, \\
T^{\mu \nu} = (im/4) \bar{\psi}\gamma_5 [\gamma^\mu \stackrel{\leftrightarrow}{\partial^\nu} + \gamma^\nu \stackrel{\leftrightarrow}{\partial^\mu} ]\psi.
\end{eqnarray}
While all these expressions are Hermitian and transform appropriately under proper orthochronous Lorentz transformations, they have the opposite behavior under parity (space inversion) compared to ordinary Dirac theory. Is this a problem? Neutrinos are supposed to interact with other particles in a way that maximizes parity violation; so maybe this is a good sign!
Furthermore, given such forms, we would not want to couple this tachyon-Dirac field to the electromagnetic field. That is good news for our theory for two physical reasons: a charged tachyon would rapidly dissipate its energy via Cherenkov radiation; and neutrinos are uncharged particles.
Let's continue this line of inquiry. Earlier I have said that it is the helicity of Dirac-tachyon states that designates the naming of particle vs anti-particle; and there has been some uncertainly about whether this is a Lorentz invariant rule. Here is a short table of simple calculations of the scalar and the pseudoscalar forms for an ordinary Dirac wavefunction $\psi_o$ and for a tachyon Dirac wavefunction $\psi_t$, both sensibly normalized.
\begin{equation}
\begin{array}{cll}
Scalar & \bar{\psi}_o \psi_o = \omega/|\omega| = \pm 1 \;\;\;\;\;& \bar{\psi}_t \psi_t = 0 \\
Pseudoscalar & \bar{\psi}_o \gamma_5\psi_o = 0 & \bar{\psi}_t \gamma_5\psi_t = h = \pm 1
\end{array}
\end{equation}
That speaks well for using helicity, $h$, for tachyons as we use the sign of the frequency for ordinary Dirac particles.
There is more we might say. In the classical theory for tachyons I have written an additional factor $\zeta = \pm 1$ in front of the Lagrangian density.
\begin{eqnarray}
\zeta m \int d\tau \sqrt{\epsilon \dot{\xi}^\mu \dot{\xi}^\nu g_{\mu \nu}}\;\delta^4(x - \xi(\tau));
\end{eqnarray}
and I have said elsewhere that this sign factor should be the helicity of the particle state. That makes this expression a pseudoscalar, which aligns with the previous discussion.
\vskip 1cm
{\bf {Acknowledgment}}
I am grateful to Professor Robert Ehrlich who encouraged me to write this paper.
\vskip 1cm
|
\section{Introduction}
Diffusion weighted magnetic resonance imaging (dMRI) is the de-facto imaging modality for probing the brain micro-structure in vivo. Biomarkers estimated with dMRI are widely used to study normal and abnormal brain development \cite{alexander2019imaging,huppi2006diffusion}. Estimation of these biomarkers from the dMRI signal entails solving inverse problems that can range from non-linear least squares to complex non-convex optimization problems \cite{novikov2019quantifying,harms2017robust}. Accurate estimation depends on dense high-quality measurements, which may be difficult or impossible to obtain in many applications such as neonatal and pediatric studies.
Machine learning, and in particular deep learning (DL), has emerged as a potent alternative to classical optimization-based estimation methods for dMRI analysis. Rather than solving the inverse problem directly, DL methods aim to learn the complex relation between the dMRI signal and the parameter of interest from a set of training data. Recent studies have used DL models for estimating diffusion tensor, diffusion kurtosis, multi-compartment models, and fiber orientation distribution, to name just a few \cite{golkov2016,gibbons2019simultaneous,karimi2021learning,tian2020deepdti,li2021superdti}. They have shown that DL methods can accurately estimate dMRI biomarkers using far fewer measurements than optimization-based techniques.
Statistical atlases are models of expected anatomy and anatomical variation. They are routinely used in medical image analysis to characterize normal anatomy and for identifying anatomical differences between normal and/or abnormal subjects or populations. In neuroimaging, study-specific atlases offer higher sensitivity and specificity of analysis than off-the-shelf atlases. Several prior works have used dMRI atlases to study brain development and maturation \cite{khan2019fetal,akazawa2016probabilistic}, to assess the impact of neurological diseases \cite{oishi2009atlas,hasan2014serial}, and for various other purposes \cite{hagler2009automated,saghafi2017spatio}.
However, no prior work has attempted to leverage DL and atlases for dMRI biomarker estimation in a unified framework. Such a framework makes much intuitive sense. While advanced DL methods are very effective in learning the complex mapping between the dMRI signal and the biomarker of interest, atlases can supply additional useful information that may be absent from the local diffusion signal. The information contained in the atlas can be particularly useful where the local diffusion signal is not adequate for accurate estimation, such as when the number of measurements is low or the signal is noisy. Some studies have shown that atlases can improve the performance of DL methods for segmentation (e.g., \cite{oguz2018combining,diniz2020esophagus}). There have also been efforts to use atlases or other sources of prior information such as distribution of noise and parameter values with classical estimation methods \cite{veraart2013comprehensive}. For example, Taquet et al. showed that estimation accuracy for diffusion compartment models can be improved by using a population-informed prior in a graphical model \cite{taquet2015improved}. Anderson \cite{andersson2008maximum} used probabilistic priors on model parameters in a maximum likelihood estimation framework for diffusion tensor estimation. Also using a Bayesian estimation approach, Clayden et al. \cite{clayden2016microstructural} found that prior information was very influential for estimating axon radii from dMRI signal. In another Bayesian work for NODDI estimation, Mozumder et al. \cite{mozumder2019population} incorporated the priors learned on 35 subjects. However, to the best of our knowledge, no prior work has used atlases within a DL framework for dMRI biomarker estimation.
On the other hand, designing such a framework is not trivial. An atlas can only represent the average of a population; it lacks the subtle but important variations among the individuals in the population. Furthermore, the correspondence between an atlas and an individual subject's brain is complex and spatially-varying. The atlas may match one subject much better than another subject. Within the brain of a subject, the match between the atlas and the subject depends on the local anatomy. It is not clear how these information can be incorporated into a machine learning framework.
In this work, we propose a framework that brings together the power of DL models and atlases for dMRI biomarker estimation. Our framework nicely addresses the challenges outlined above by providing methods to compute the reliability of the atlas and its degree of correspondence with a subject in a spatially-varying manner. We use our proposed framework to estimate fractional anisotropy and neurite orientation dispersion from down-sampled dMRI data of newborn subjects and show that it is significantly more accurate than standard estimation methods as well as recent DL techniques.
\section{Materials and methods}
\subsection{Data}
We used 300 dMRI scans from the Developing Human Connectome Project (DHCP) dataset \cite{bastiani2019}. Each scan is from a different subject. We used 230 of the scans for model development, including atlas creation and DL model training. We used the same 230 scans to train the competing DL models (described below). We used the remaining (completely independent) 70 scans to test our method and the competing methods. The age of the subjects at the time of scan ranged between 31 and 45 gestational-equivalent weeks. We have focused on this dataset because dMRI analysis for this age range is especially challenging due to higher free water content, incomplete myelination, and overall lower data quality \cite{dubois2014early}.
\subsection{Atlas development}
Because of rapid brain development in the neonatal period, a single atlas cannot represent the entire age range \cite{pietsch2019framework,uus2021multi}. Therefore, we built atlases of the biomarkers of interest separately at one-week intervals. To build atlases for GA of week 35, for example, we used subjects with GA between 34.5 and 35.5. For each GA, we used 10 scans to build the atlas. Our experience and prior works \cite{pietsch2019framework,khan2019fetal,uus2021multi} have shown that approximately 10 subjects are sufficient for each GA.
In this study, we focus on two biomarkers: 1) Fractional anisotropy (FA) from the diffusion tensor model, which is arguably the most widely used dMRI biomarker, and 2) Orientation dispersion (OD) from the NODDI model \cite{zhang2012noddi}, which is a more complex model than diffusion tensor and which has been shown to be a more specific biomarker in some studies \cite{andica2021neurite,palacios2020evolution}.
Given the dMRI signal volumes $\{ s_i \}_{i=1}^n$ for $n$ subjects, we compute the biomarkers of interest separately for each subject. Regardless of which biomarkers are explored, we also always compute the diffusion tensor for each subject because we use the tensors for accurate spatial alignment. This is shown in Figure \ref{fig:atlas_creation}, where we have used $p$ and $q$ to denote the biomarkers considered in this work and $T$ to denote the diffusion tensor. Given the set $\{ T_i \}_{i=1}^n$ of subject tensors, we compute a set of transformations $\{ \Phi_i \}_{i=1}^n$ that align these tensors into a common atlas space and compute a mean tensor $\bar{T}$. This is done using an iterative approach that computes a series of rigid, affine, and diffeomorphic non-rigid registrations and updates $\bar{T}$ at every iteration, rather similar to common practice \cite{pietsch2019framework,khan2019fetal}. Specifically, we perform five iterations of rigid registration, followed by five iterations of affine registration, and finally ten iterations of non-rigid registration. The final registration transform $\Phi_i$ for subject tensor $T_i$ is the composition of the final affine and non-rigid transforms.
\begin{figure}
\includegraphics[width=\textwidth]{atlas_creation.png}
\caption{Left: Biomarkers of interest and diffusion tensor are computed for each subject. Right: tensor-to-tensor registration is used to compute the atlas and atlas confidence maps for each biomarker.} \label{fig:atlas_creation}
\end{figure}
Using the transformations $\{ \Phi_i \}_{i=1}^n$ computed via tensor-to-tensor registration described above, we transform the biomarker maps into the common atlas space. We compute the mean of the transformed biomarker maps as the biomarker atlas. Furthermore, we compute the standard deviation of the transformed biomarker maps as a measure of atlas confidence. Formally, for biomarker $p$:
\begin{equation}
\bar{p}= \text{mean} \Big[ \Phi_i(p_i) \Big]_{i=1}^n , \hspace{5mm} p^{\sigma}= \text{std} \Big[ \Phi_i(p_i) \Big]_{i=1}^n
\end{equation}
\noindent where mean and standard deviation are computed in a voxel-wise manner across subjects. Larger values of $p^{\sigma}$ indicate higher variability/disagreement between the biomarker maps of subjects used to create the atlas and, hence, \emph{lower} confidence. Figure \ref{fig:atlas_creation} shows example atlas and atlas confidence maps for FA and OD.
\subsection{Biomarker estimation for an individual subject}
Given the dMRI signal volume for an individual subject, $s_k$, we compute the desired biomarker(s) for that subject via the following steps, which have been shown for an example subject in Figure \ref{fig:subject_analysis}.
\subsubsection{Step 1: Atlas-to-subject alignement.}
In order to exploit the information encoded in an atlas, we need to accurately register it to the subject space. As in our atlas development described above, we use tensor-to-tensor registration for this alignement. Hence, regardless of the biomarker(s) being estimated, we compute the diffusion tensor, $T_k$, for the subject. We then compute affine+non-linear registration transforms that map the atlas tensor $\bar{T}$ to the subject tensor, $T_k$. We use $\Phi_k$ to denote the composition of these affine and non-linear transforms. Hence, $\Phi_k$ describes the complete spatial alignment from the atlas space to the subject. We denote the template tensor transformed to the subject space with $\bar{T}_k= \Phi_k(\bar{T})$.
The registration between the template tensor and the subject tensor is never perfect. The accuracy of this registration is a potentially important piece of information because it indicates where the prior information is more reliable. In other words, if at voxel $i$ the registration between $\bar{T}$ and $T_k$ is more accurate, then we have a higher incentive to trust the biomarker atlas at the location of that voxel. The accuracy of this registration is spatially varying and depends on at least three factors: 1) Accuracy of computation of $T_k$, which in turn depends on the quality of the diffusion signal, $s_k$, 2) Degree of similarity between the subject and the atlas, and 3) Accuracy of the registration procedure that aligns the atlas to the subject. We propose the following practical formulation to estimate the error of this alignment:
\begin{equation} \label{eq:alignment_error}
\Phi_k^{\text{err}}= \theta \big( \bar{T}_k, T_k \big) . \exp \Big( - \min \big[ \text{FA}(\bar{T}_k) , \text{FA}(T_k) \big] / \tau \Big).
\end{equation}
This formulation has two terms. The first term, $\theta$, measures the angle between the major eigenvectors of $\bar{T}_k$ and $T_k$. Clearly, smaller angles indicate more accurate registration. The second term is introduced to down-weight the registration accuracy for the location of less anisotropic tensors such as gray matter and cerebrospinal fluid (CSF). For CSF, for example, the tensor is spherical and the computed orientation of the major eigenvector is not reliable but the eigenvectors of $T_k$ and $\bar{T}_k$ may be very close to each other by chance, hence artificially making $\theta$ very small. By using the minimum of the FAs, if either $\bar{T}_k$ or $T_k$ has a low anisotropy, the second term will have a larger value. We set $\tau=0.2$, which we found empirically to work well.
\subsubsection{Step 2: Estimation using a DL model.}
To compute a specific biomarker for an individual subject, standard techniques only use the diffusion signal, $s_k$. In our framework, we also utilize the prior information encoded in the biomarker atlas as described above. Specifically, for estimating a biomarker $p$ for subject $k$, we have three additional pieces of information; 1) the biomarker atlas registered to the subject space $\Phi_k(\bar{p})$, 2) the biomarker atlas confidence registered to the subject space $\Phi_k(\bar{p})$, and 3) registration error $\Phi_k^{\text{err}}$. Since these three pieces of information are spatially varying and aligned with the subject space, we simply concatenate them with the diffusion signal to generate the input to our DL model, as shown in Figure \ref{fig:subject_analysis}. For example, for estimating the biomarker $p$ for subject $k$, the input to the network is $\Big[ s_k, \Phi_k(\mathbf{ \bar{p} }), \Phi_k(\mathbf{ p^{\sigma}}) , \Phi_k^{\text{err}} \Big]$. The estimation target is the ground-truth parameter map, $p^{\text{g.t.}}$, computed as described below.
\begin{figure}
\includegraphics[width=0.9\textwidth]{subject_analysis.png}
\caption{Our proposed method for estimating dMRI biomarkers using atlases and DL. (a) We compute the diffusion tensor $T_k$ from the subject data. (b) We compute the registration $\phi_k$ from the atlas to the subject using tensor-to-tensor registration, and then the error of this registration $\Phi_k^{\text{err}}$ using Eq. \ref{eq:alignment_error}. (c) We align the biomarker atlases and atlas confidence maps to the subject using $\Phi_k$. We then feed these information, along with the dMRI signal $s_k$ to the DL model, which aims to predict the ground truth biomarker values.} \label{fig:subject_analysis}
\end{figure}
We used the UNet++ \cite{zhou2018unet} as our DL architecture. As it has been recently demonstrated for image segmentation in \cite{isensee2021nnu}, our experience shows that the exact network architecture is not critical. We used patches of size $48^3$ voxels. The number of network input channels was equal to the number of diffusion measurements plus 3, as described above. We train a separate network for each biomarker, hence the network had only one output channel to compute the scalar biomarker of interest. We set the number of feature maps in the first stage of the network to be 12, which was the largest possible on our GPU memory. For training, we sampled blocks from random locations in the training images. At test time, we used a sliding window with a stride of 16 voxels in each dimension to estimate the biomarker for an input dMRI volume of arbitrary size.
\subsubsection{Compared methods and evaluation strategy.}
For FA estimation, we compared our method with: 1) Constrained weighted linear least-squares (\textbf{CWLLS}) \cite{koay2006unifying}, which is the standard method; 2) \textbf{Deep-DTI} \cite{tian2020deepdti}, which is a recent DL method based on CNNs. This method exploits the anatomical T2 image, in addition to the diffusion signal, for estimation. Hence, for Deep-DTI we also used the T2 image, which we registered to the dMRI volume, and 3) \textbf{Super-DTI} \cite{li2021superdti}, which is a recent DL method based on CNNs. For OD estimation, we compared our method with: 1) \textbf{Dmipy} \cite{fick2019dmipy}, which follows a standard optimization-based estimation approach, 2) Microstructure Estimation using a Deep Network (\textbf{MEDN+}) \cite{ye2017tissue}, which is a DL method that has been inspired by AMICO \cite{daducci2015accelerated} and significantly outperforms AMICO too, and 3) Another recent CNN-based model \cite{gibbons2019simultaneous}, which we refer to as \textbf{CNN-NODDI}.
Each DHCP scan includes 20 $b=0$ measurements and 280 diffusion weighted measurements at $b=400$ (n=64), $b=1000$ (n=88), and $b=2600$ (n=128). For FA it is known that b values close to 1000 are optimal \cite{jones1999optimal}. Hence, we used all 88 measurements in the $b=1000$ shell (along with all b=0 measurements) to compute the ground truth using CWLLS. We then selected subsets of 12 and 6 measurements from this shell for each subject, which represent measurement down-sampling factors of approximately 7 and 15, respectively. To select the 6 measurements, similar to \cite{tian2020deepdti,karimi2022diffusion}, we considered the 6 optimal diffusion gradient directions proposed in \cite{skare2000condition} and chose the measurements that were closest to those directions. To select the 12 measurements, we selected these measurements to be close to uniformly spread on the sphere, as suggested in \cite{jones1999optimal,karimi2021learning}.
For OD (and NODDI), we used all 300 measurements to reconstruct the ground truth using Dmipy \cite{fick2019dmipy}. For DL-based reconstruction, prior works have typically used 20-60 measurements from more than one shell \cite{ye2017tissue,gibbons2019simultaneous}. Here, we choose either 6 and 15 measurements from each of the $b=1000$ and $b=2600$ shells, for a total of 12 and 30 measurements, which represent downsampling factors of approximately 24 and 10, respectively. We selected these measurements to be close to uniformly spread on the sphere, using an approach similar to \cite{karimi2021learning}. For a fair comparison, for both FA and OD we used the same down-sampled datasets for our method and for all competing techniques.
\subsubsection{Implementation and training.}
We used the DTI-TK software \cite{zhang2006deformable} to compute all registrations, including the registrations needed in building the atlases. In creating the biomarker atlases we used MRtrix \cite{tournier2019mrtrix3} to compute the diffusion tensors and FA, and we used Dmipy \cite{fick2019dmipy} to compute OD. We implemented all DL techniques in TensorFlow 1.14 under Python 3.7. We ran all algorithms on a Linux machine with 16 CPU cores and an NVIDIA GeForce GTX 1080 GPU.
We trained our model by minimizing the $\ell_2$ norm between the predicted and ground truth biomarker using Adam \cite{kingma2014}, a batch size of 10, and an initial learning rate of $10^{-4}$ that was reduced by half every time the validation loss did not decrease after a training epoch. For the competing methods, we followed the training procedures recommended in the original papers.
\section{Results and Discussion}
\subsection{Comparison with other techniques}
Tables \ref{table:FA_results} and \ref{table:OD_results} show the reconstruction error for different methods for FA and OD, respectively, computed on the 70 independent test subjects. For both FA and OD, the DL methods were substantially more accurate than the standard optimization-based techniques (i.e., CWLLS and Dmipy). The proposed method achieved lower errors than the other DL methods for both FA and OD at both down-sampling factors. We used paired t-tests to compare our method with the other techniques. For both FA and OD and at both down-sampling factors, the estimation error for our method was significantly lower than the error for any of the compared techniques ($p<0.001$). Figure \ref{fig:sample_results} shows example reconstruction results for different techniques. Our method achieves lower errors across the brain for both FA and OD.
\begin{table*}[!htb]
\footnotesize
\caption{\footnotesize{FA estimation errors for the proposed method and compared techniques.}}
\label{table:FA_results}
\begin{center}
\begin{tabular}{ L{29mm} C{19mm} C{21mm} C{21mm} C{22mm} }
\hline
no. of measurements & CWLLS & Deep-DTI & Super-DTI & Proposed \\ \hline
$n=6$ & $0.111 \pm 0.014$ & $0.048 \pm 0.008$ & $0.048 \pm 0.005$ & $\bm{0.040 \pm 0.005}$ \\
$n=12$ & $0.053 \pm 0.007$ & $0.044 \pm 0.006$ & $0.043 \pm 0.004$ & $\bm{0.039 \pm 0.005}$ \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[!htb]
\footnotesize
\caption{\footnotesize{Comparison of OD estimation errors for the proposed method and competing techniques.}}
\label{table:OD_results}
\begin{center}
\begin{tabular}{ L{29mm} C{19mm} C{21mm} C{21mm} C{22mm} }
\hline
no. of measurements & Dmipy & MEDN+ & CNN-NODDI & Proposed \\ \hline
$n=16$ & $0.138 \pm 0.032$ & $0.064 \pm 0.028$ & $0.058 \pm 0.011$ & $\bm{0.047 \pm 0.004}$ \\
$n=30$ & $0.096 \pm 0.029$ & $0.052 \pm 0.030$ & $0.050 \pm 0.008$ & $\bm{0.044 \pm 0.005}$ \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{figure}
\includegraphics[width=1.0\textwidth]{FAOD.png}
\caption{Example FA (top) and OD (bottom) reconstructions by different methods. In each of the four examples, the bottom row shows maps of absolute estimation error.}
\label{fig:sample_results}
\end{figure}
For both FA and OD, our proposed method showed a smaller increase in error as the down-sampling rate increased. Specifically, the FA estimation error for our method increased by 2.5\% as the number of measurements was reduced from 12 to 6, compared with 9\% and 12\% for Deep-DTI and Super-DTI, respectively. For OD, the estimation error for our method increased by 7\% as the number of measurements was decreased from 30 to 16, compared with 23\% and 16\% for MEDN+ and CNN-NODDI, respectively.
The training time for our model was 10 hours, compared with 10-60 hours for the other DL methods. To estimate FA or OD for a dMRI test volume, our model required $73 \pm 20$ seconds. Approximately 90\% of this time was spent on computing the atlas-to-subject registration. The average computation time for the other DL methods ranged from 10 seconds for SuperDTI to 30 minutes for MEDN+. The average computation times for CWLLS and Dmipy were, respectively, 20 seconds and 3.2 hours. Nonetheless, computation time for dMRI analysis is not a critical factor since fast estimation is typically not a requirement.
\subsection{Ablation experiments}
Table \ref{table:ablation} shows the results of some ablation experiments. We have performed these experiments to show that the superior accuracy of our method compared to the other methods is due to the incorporation of the atlas information as we claim, rather than the differences in network architecture or training. Results show that the error of our method increases significantly when we discard the atlas information. When we only use the dMRI signal (last column in Table \ref{table:ablation}), the FA accuracy of our method is very close to that of Deep-DTI and Super-DTI (Table \ref{table:FA_results}) and the OD accuracy of our method is slightly worse than CNN-NODDI (Table \ref{table:OD_results}). Ablation experiments further show that all three extra pieces of information contribute to the model accuracy, but the contribution of atlas and atlas confidence is larger than the contribution of atlas-to-subject alignment error.
\begin{table*}[!htb]
\footnotesize
\caption{\footnotesize{Results of some ablation experiments. In the column headings, we use $p$ to denote either FA or OD.}}
\label{table:ablation}
\begin{center}
\begin{tabular}{ L{29mm} C{25mm} C{21mm} C{21mm} C{19mm} }
\hline
input to the model & $\big[ s_k, \Phi_k(\mathbf{ \bar{p} }), $ $ \Phi_k(\mathbf{ p^{\sigma}}) , \Phi_k^{\text{err}} \big]$ & $\big[ s_k, \Phi_k(\mathbf{ \bar{p} }), $ $ \Phi_k(\mathbf{ p^{\sigma}}) \big]$ & $\big[ s_k, \Phi_k(\mathbf{ \bar{p} }) \big]$ & $s_k$ \\ \hline
FA, $n=6$ & $\bm{0.040 \pm 0.005}$ & $0.041 \pm 0.005$ & $0.044 \pm 0.006$ & $0.047 \pm 0.006$ \\
OD, $n=16$ & $\bm{0.047 \pm 0.004}$ & $0.050 \pm 0.004$ & $0.055 \pm 0.011$ & $0.060 \pm 0.010$ \\
\hline
\end{tabular}
\end{center}
\end{table*}
\section{Conclusions}
We proposed a novel framework to incorporate atlases and DL for dMRI biomarker estimation. We cannot claim that the design of our framework is ``optimal''. For example, atlas-to-subject alignment may be improved by incorporating anatomical MRI information in addition to the diffusion tensor information. Nonetheless, our work has shown, for the first time, that spatio-temporal atlases can be used within a DL framework to achieve superior biomarker estimation accuracy from down-sampled data.
\section*{Acknowledgment}
This study was supported in part by the National Institute of Biomedical Imaging and Bioengineering and the National Institute of Neurological Disorders and Stroke of the National Institutes of Health (NIH) under award numbers R01EB031849, R01NS106030, and R01EB032366; in part by the Office of the Director of the NIH under award number S10OD0250111; in part by the National Science Foundation (NSF) under award 2123061; and in part by a Technological Innovations in Neuroscience Award from the McKnight Foundation. The content of this paper is solely the responsibility of the authors and does not necessarily represent the official views of the NIH, NSF, or the McKnight Foundation.
The DHCP dataset is provided by the developing Human Connectome Project, KCL-Imperial-Oxford Consortium funded by the European Research Council under the European Union Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement no. [319456]. We are grateful to the families who generously supported this trial.
\clearpage
\bibliographystyle{splncs04}
|
\section{Introduction}
Machine learning (ML) models are increasingly used in safety-critical settings like healthcare~\cite{chen2020ethical, tuggener2019automated}, college admissions~\cite{jamison2017applying}, and law~\cite{alarie2016using}.
Several studies have shown that human decisions can become more accurate when assisted by such ML models \cite{buccinca2020proxy,tschandl2020human,wang2020human}. However, many ML models are ``blackboxes''---they might have too many parameters or be proprietary---and cannot explain their predictions in ways humans understand~\cite{rudin2019stop}.
In such scenarios, users may struggle to understand a model's outputs enough to trust and use its predictions~\cite{dietvorst2015algorithm,doshi2017towards,lipton2018mythos}.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{images/FAccT_intro.pdf}
\caption{An example of an unfair global explanation model. Orange and Blue circles indicate predicted classes, \textit{healthy} or \textit{unhealthy}, respectively. $\square$ and $\triangle$ denote group membership. The red dashed line is a linear explainability model fit to approximate the black blackbox decision boundary. The two figures on the right show that the linear approximate is worse for the $\square$ group.}\label{fig:local_global_expls}
\end{figure}
\emph{Post-hoc explainability} methods have recently begun helping users better understand why blackbox models make certain predictions \cite{ribeiro2016model,ribeiro2018anchors,lakkaraju2019faithful}. A popular post-hoc approach is to train simple, human-interpretable models to \textit{imitate} a blackbox model's behaviour \cite{ribeiro2016should} by maximizing the congruity between simple approximations and blackbox model predictions. Such approximation quality is known as \textit{fidelity}~\cite{craven1995extracting}.
Then, the simpler model can be used either as a new stand-alone model or to explain one prediction at a time~\cite{tan2018distill, ribeiro2016should}. By highlighting important inputs, these explainability methods provide a path towards helping users trust machine learning models in high-impact settings \cite{ribeiro2016model}.
However, it remains unknown when and if these models approximate behavior \textit{fairly}.
If fidelity differs between different pre-defined groups (e.g., between demographics) in a dataset, explainability methods may perpetuate machine bias by encouraging users to trust model predictions for some people but not others.
In this work, we study \textit{to what degree do gaps in fidelity exist between subgroups?}
To answer this question, we measure group-wise fidelity for different post-hoc explanation methods on real tabular datasets that include group membership.
Intuitively, an explanation model is \emph{fair} if it has equally high fidelity for all protected groups. This definition is similar to common group fairness definitions which seek to eliminate gaps in predictive performance across groups~\cite{pleiss2017,hardt2016equalityofopportunity,zhang2022improving}. We introduce two definitions of \textit{fidelity gaps}, or disparities in fidelity across different subgroups. Using these measures, we benchmark two popular families of post-hoc explanation models: local methods, which imitate the boundary of a blackbox around one instance~\cite{ribeiro2016should,lundberg2017unified,ribeiro2018anchors}, and global methods, which imitate the blackbox across all instances \cite{rudin2014algorithms,lakkaraju2019faithful}. We also motivate fidelity gap measurements by showing mathematically that measuring fidelity gaps across subgroups directly connects with prior work on fairness preservation for explainability \cite{dai2021will}.
With a comprehensive audit of explanation fairness, we find that significant fidelity gaps exist between subgroups.
A popular way to train fairer models is through robust optimization \cite{liu2021just,rahimian2019distributionally,sagawa2019distributionally}.
To see how robust training impacts large fidelity gaps, we also study a simple technique for retraining explainability methods to improve their fairness.
We also study potential causes for these fidelity gaps, and highlight mechanisms by which group information can indirectly be used in post-hoc explanations as an important contributing factor. Lastly, we assess the impact of the observed fidelity gaps on real-world decision-making accuracy with a carefully designed simulation study. The major findings of our evaluation are as follows:
\textbf{Explanation fidelity varies significantly between subgroups:} We find that fidelity gaps grow up to 7\% between subgroups in our experiments using four popular datasets. In comparison to average fidelity across all data points, the fidelity of explanations for disadvantaged groups is often significantly lower (up to 21\%).
These findings indicate that judging the quality of explanations by their \textit{average} fidelity alone---a common approach---overestimates explanation quality for some subgroups, potentially leading to worse downstream decision making.
This effect is illustrated in Fig.~\ref{fig:local_global_expls}.
\textbf{Balanced and robust training can reduce but not eliminate fidelity gaps:} We use robust training by adaptively reweighting or balancing groups in training data while training explanation models. This turns out to be a promising direction: fidelity gaps improve across subgroups, though this depends on both the dataset and exact method utilized.
\textbf{Fidelity gaps have an impact on decision-making in the real-world:} Using a simulation study, we find that larger fidelity gaps may lead to disparities in decision making accuracy for different subgroups. This implies that ignoring fidelity gaps between subgroups can have detrimental effects to members of protected groups.
Finally, we categorize and discuss promising directions for evaluating and improving post-hoc explainability methods. In summary, our work is a step towards training fair and reliable explanation models.
\section{Related Work}
\subsection{Explainable Machine Learning}
While ML models achieve outstanding performance, users often find them too complex to trust in practice~\cite{dietvorst2015algorithm}.
To make such \textit{blackbox} models more useful, users require that they be understandable, often
due to laws \cite{bibal2021legal} or preference \cite{holzinger2018machine, roscher2020explainable, bhatt2020explainable}.
To fill this gap, recent approaches ``explain'' a blackbox model's behavior after it is trained \cite{dovsilovic2018explainable, burkart2021survey}. These \textit{post-hoc} explainability methods are now used in safety-critical applications like healthcare \cite{ahmad2018interpretable} and finance \cite{bussmann2021explainable}.
Several explainability methods are increasingly-popular because they make no assumptions about a blackbox model's architecture \cite{ribeiro2016model}, also known as model-agnostic.
In contrast, some methods are designed exclusively for deep learning, requiring their internal structure and gradients \cite{shrikumar2017learning, shrikumar2017learning, bach2015pixel,li2015visualizing, karpathy2015visualizing}.
We consider model-agnostic methods, which can be used for a wider family of blackbox models, including deep learning.
Model-agnostic explainability methods are primarily either \textit{local} or \textit{global} \cite{du2019techniques}.
\textit{Local} methods justify one model prediction at a time, typically by approximating the decision boundary around one data point \cite{plumb2018model, botari2020melime, lundberg2017unified, ribeiro2018anchors, rathi2019generating, ribeiro2016should}. Then, the weights learned by the local models are used to rationalize the blackbox model's prediction. Some of the best-known local methods are LIME \cite{ribeiro2016should}, which learns a sparse linear classifier on a dataset of perturbed samples, and SHAP \cite{lundberg2017unified}, which uses feature-wise Shapley values \cite{roth1988shapley}.
\textit{Global} methods, on the other hand, train interpretable surrogate models of the blackbox model's behavior on an entire dataset, which is then used in lieu of the blackbox.
These methods primarily use tree-based models \cite{lundberg2020local}, rule lists \cite{lakkaraju2019faithful, puri2017magix}, sparse linear models \cite{ustun2016supersparse, zhang2021learning}, and generalized additive models \cite{lou2012intelligible} as surrogates.
\subsubsection{Explainable ML in Safety-critical Settings}
The need for explainability in safety-critical applications is a nebulous and contested topic for several reasons:
\noindent\textbf{Interpretability vs Explainability.} Some prior works advocate for interpretability over explainability~\cite{rudin2019stop,mittelstadt2019explaining,ghassemi2021false}. An explanation model without perfect fidelity is by definition incorrect for some data points~\cite{rudin2019stop}. Our work extends this point; these errors can occur for some groups more than others. Since explanations influence trust~\cite{buccinca2020proxy}, it is important to conduct user studies on the impacts on aversion algorithmic advice~\cite{logg2019algorithm} and over-reliance on algorithmic advice~\cite{dijkstra1998persuasiveness}.
\noindent\textbf{Anchoring Effects.} Explanations can fool people into trusting incorrect models~\cite{bansal2021most,poursabzi2021manipulating}. For example, \citet{poursabzi2021manipulating} find that when people are shown explanations from a bad model, they become more likely to trust the model, even when it is clearly wrong.
In cases like this, people use the explanations while judging the quality of the blackbox models, even though the explanations themselves can be misleading \cite{bansal2021does}.
\noindent\textbf{Mismatched end-user and model-designer goals.}
Many explainability methods aim to assist model debugging, while non-engineer users only choose when to accept a blackbox's decisions \cite{ribeiro2016model,krishnan2017palm,pradhan2021interpretable}.
This mismatch can have downfalls.
For instance, \citet{buccinca2021trust} find that the explanations people find most useful are also the ones they trust \textit{incorrectly}.
Resolving this mismatch requires goal-aware explainability methods along with education to ensure end-users are properly trained in using these methods.
\subsubsection{Desiderata for Post-hoc Explanations}
Most post-hoc explainability methods have three goals:
\indent\textbf{Reliability.} Explanations must be accurate for the right reasons. People often trust explained models \cite{poursabzi2021manipulating,ross2017right}, so ensuring that explanations are faithful to the original model and not simply easy-to-rationalize is essential \cite{ghassemi2021false}.
\indent\textbf{Robustness.} Explanation models should not overfit to spurious patterns in the data~\cite{ghorbani2019interpretation,lakkaraju2020fool} and must be robust in the presence of small distribution shifts at test time~\cite{lakkaraju2020fool}.
\indent\textbf{Simplicity.} Models should be sparse, and leave little room for effects for human cognitive biases such as the anchoring effect~\cite{poursabzi2021manipulating}. Ideal explanations will highlight only the key information needed to understand a model's behavior, encouraging users to engage with explanations in predictable ways~\cite{buccinca2021trust}. However, there is often a trade-off between an explanation's faithfulness and its simplicity~\cite{lakkaraju2017interpretable}. Recent work on cognitive forcing---where users explicitly interact and understand explanations---appears to be a promising direction to address this trade-off~\cite{buccinca2021trust}.
Along with other recent efforts~\cite{ghorbani2019interpretation,lakkaraju2020fool,poursabzi2021manipulating,bansal2021does}, we promote a fourth goal: \textbf{Fairness}.
Explanation quality should not depend on group membership.
We find that this requirement is not yet satisfied by popular explainability methods.
\subsection{Algorithmic Fairness}
\label{sec:algorthmic_fairness_background}
Formalizing fairness is a flourishing research area
~\cite{mehrabi2021survey, chen2018classifier, chouldechova2017fair, berk2017impact, locatello2019challenging, chouldechova2018frontiers, hardt2016equalityofopportunity, zemel2013learning, zafar2017fairness,dwork2011fairnessthroughawareness}.
Recent works define fairness at either the \textit{individual}- or \textit{group}-level.
Individual fairness~\cite{dwork2011fairnessthroughawareness} requires similar predictions for similar individuals; group fairness requires similar predictions for different groups (sex or race, for example).
We consider group-level fairness for binary classification, which we quantify using demographic parity (DP) gap \cite{hardt2016equalityofopportunity,pleiss2017}, a standard group-fairness metric.
We describe this metric probabilistically, allowing calculation of gaps across groups: $\text{DP} = E[\hat{Y}\vert A = a] - E[\hat{Y}\vert A = b] \quad \forall a,b \in A$, where $\hat{Y}$ is a predictor and its DP is measured with respect to attribute $A$.
There are three main strategies for encouraging group fairness
~\citep{caton2020fairness}: pre-processing data to find less-biased representations~\cite{paul2021generalizing};
enforcing fairness while training a model, typically through regularization~\cite{madras2018learning,zhang2018mitigating};
and altering a model's predictions to satisfy fairness constraints after it is trained~\citep{hardt2016equalityofopportunity, agarwal2018reductions, pleiss2017, chen2020}.
In this paper, we utilize the inprocessing method proposed by ~\citet{zhang2018mitigating} for training fair blackbox models. Further, recent work has demonstrated that group-robust training can increase fairness by improving the worst-group accuracy~\cite{sagawa2020groupdro}.
\subsection{Bias in Model Compression and Risks of Fairwashing}
Several recent works study the effects of model and data compression on fairness~\cite{samadi2018price,hooker2020characterising}. For example, ~\citet{samadi2018price} observe that reconstruction error associated with data dimensionality reduction via principal component analysis is higher for some populations.
\citet{hooker2020characterising} show that average accuracy after ML model compression
hides disproportionately high errors on a small subset of examples.
In a similar vein, we study post-hoc explanation models, which are often compressed blackbox models, and assess how they transmit bias. Another related topic is ``fairwashing'': the act of overlooking a model's unfair behavior by rationalizing its predictions via explanations \cite{aivodji2019fairwashing}. Our paper instead considers fairness in how well explanation models imitate blackbox models (rather than the ground-truth), regardless of blackbox model fairness. In concurrent work, ~\citet{dai2022} showed that explanation quality may differ between subgroups, further validating our findings. However, the metrics and methodological focus in their work is on feature-based variations to fidelity. We encourage readers to review their work for an alternate approach to measuring the fairness of explanations.
\section{Measuring the Fairness of Explanations}
\label{sec:background}
Here, we introduce metrics for measuring \emph{fairness of explanation models} or fidelity gaps across subgroups.
\subsection{Notation}
Consider a dataset $\mathcal{D} = \{(\bm{x}_i, y_i)\}_{i=1}^n$ that contains $n$ training data points. $\bm{x}_i \in \mathbb{R}^d$ is the $d$-dimensional feature vector of the $i$-th data point in $\mathcal{D}$ and $y_i \in \{0, 1\}$ is its associated binary label.
We assume binary classification for simplicity.
Let $g_i \in \{1, \dots, G\}$ be a variable defining group membership with respect to the protected attribute for every data point
for a total of $G$ subgroups. In all cases, $g$ serves as auxiliary information and is not used during any model training, unless specified otherwise.
A blackbox classifier $B: \mathbb{R}^d \rightarrow \{0, 1\}$ predicts one binary label per input $x$.
Given classifier $B$, we wish to explain its prediction given some query point $\bm{x}^*$.
To achieve this, an explanation model $E$ is chosen from a set of interpretable models (\textit{e.g.}, linear models or decision trees).
Then, $E$ is trained to \textit{imitate} $B$ either locally (for the feature space near $\bm{x}^*$) or globally (for all data points in $\mathcal{D})$.
\subsection{Fidelity of Explanations}
Given a blackbox $B$ and explainability model $E$, we seek to describe how well $E$ approximates $B$'s behavior. Fidelity, as detailed in Definition \ref{bg:def:fidelity} below, is a powerful measure for this approximation error \cite{aivodji2021characterizing,lakkaraju2017interpretable,lakkaraju2019faithful}, though it disregards group information.
\begin{definition}[Explanation Fidelity~\cite{craven1995extracting}]
\label{bg:def:fidelity}
Given blackbox model $B$ and explanation model $E$, the \textit{explanation fidelity} on data points $(x_i,y_i)_{i=1}^N$ is
$\frac{1}{N}\sum_{i=1}^{N} L(B(x_i), E(x_i))$,
where $L$ is a performance metric.
\end{definition}
For $L$, we use accuracy, AUROC\footnote{AUROC cannot be written directly as a sum but we slightly abuse notation for readability.}, and mean error,
denoted as $Fidelity^{Acc}$, $Fidelity^{AUROC}$, $Fidelity^{Err}$, respectively.
In the following sections, we build up to a definition of explanation fidelity that considers group information.
First, we motivate the need for a metric that measures fidelity across groups (Section~\ref{sec:motivation}), then define two new notions for measuring the fairness of explanations (Section~\ref{sec:metrics}).
\subsection{Fidelity Gaps are Critical to Fairness Preservation}\label{sec:motivation}
\citet{dai2021will} recently introduced \textit{fairness preservation} in surrogate explanation models.
Fairness is \textit{preserved} when the fairness properties of the blackbox model and explanation model are identical.
For example, consider Figure~\ref{fig:local_global_expls}.
A linear explanation $E$ is a high-fidelity approximation of the blackbox $B$'s decisions for one group ($\triangle$), but not the other ($\square$). Here, $B$ seems unfair in predicting the ``unhealthy" class
for the two groups. Meanwhile, $E$ appears fairer.
In this example, $B$'s degree of (un)fairness---the demographic parity gap---is not \emph{preserved} by the explanation model.
For demographic parity, fairness preservation in explanations implies that $B$ and $E$ should have similar DP Gaps (Section~\ref{sec:algorthmic_fairness_background}).
To reliably judge a blackbox's fairness using only its post-hoc explanations, preserving fairness is essential.
If fairness is preserved, then when an explanation seems unfair, we can be confident that the blackbox model is likely similarly unfair as well.
Next, we prove that fairness preservation is directly linked to fidelity gaps across subgroups. While ~\citet{dai2021will} briefly intuit that fairness preservation impacts explanation fidelity via an illustratory example, only a group-conditional blackbox model's decision boundary under imbalanced group sizes is considered. In contrast, we do not make any assumptions about the relative sizes of groups or group-dependence of blackbox model and instead show that fairness preservation is related to fidelity gaps (more so than overall explanation fidelity).
\subsubsection{Fidelity Gaps are related to Fairness Preservation}
\begin{theorem}\label{thm:thm}
Let \(E\) be a post-hoc explanation model trained to imitate predictions of blackbox model \(B\), and mean residual error for a set of $N$ data points in dataset $\mathcal{D}$ is $\frac{1}{N}\sum_{x \in \mathcal{D}}(E(x)-B(x))$. Then, the difference between the Demographic Parity Gaps of \(E\) and \(B\), both with respect to binary valued-protected attribute $g$, is equal to the difference in mean residual error of data points with $g=1$ and $g=0$.
\end{theorem}
The full proof of this theorem is in Appendix~\ref{sec:proof}; the key idea is to expand $E(x_i)=B(x_i)+\epsilon_i$ where $\epsilon_i$ is a residual for each point $x_i$. This is valid when $E$ is trained to imitate $B$ with high fidelity (\textit{e.g.}, minimizing mean squared error or cross-entropy loss). We also empirically validate this theorem on explanation models in Appendix~\ref{sec:empirical_validation}.
From Theorem \ref{thm:thm}, a \emph{sufficient} condition for DP, as computed over instances $x_{i}$ and their local linear classifiers $E_{i}(x_{i})$ or a global model $E$ (where $E_{i}=E \forall i$) is ensuring that the mean residual errors for each group is comparable. This is the same as low fidelity gaps across subgroups where $L$ is the mean error. Note that this does not correspond to mean absolute difference between predictions of $E$ and $B$, but instead their mean difference. Theorems of similar form could be derived for other group fairness definitions, but the $\epsilon$ values and data points considered would depend on the ground truth as well (e.g., for equal-opportunity, the $\epsilon$ difference would only contain terms for data points with positive-class ground truth).
With this motivation in mind, we next introduce two new metrics that measure fidelity gaps across subgroups.
\subsection{Measuring the Fairness of Explainability Methods}\label{sec:metrics}
Building on the definition of average fidelity across groups (Defn.~\ref{bg:def:fidelity}), we introduce two new measurements for the fairness of explanation models by evaluating their fidelity gaps between subgroups. The first metric (Definition~\ref{bg:def:subgroup_avg_gap}) addresses the question: by what degree would relying on the average fidelity alone be detrimental to subgroups of data? The second metric estimates the mean difference in fidelity of explanations between subgroups of data (Definition~\ref{bg:def:subgroup_max_gap}).
Inspired by past work~\cite{liu2021just,craven1995extracting}, the \textit{maximum fidelity gap from average} (Definition \ref{bg:def:subgroup_avg_gap}) computes the difference between the overall, average fidelity and the worst-case subgroup fidelity.
This way, we quantify the maximum degree to which an explanation model's fidelity is lower for disadvantaged groups compared to the average across all subgroups.
\begin{definition}[Maximum Fidelity Gap from Average: $\Delta_{L}$]
\label{bg:def:subgroup_avg_gap}
Let the maximum fidelity gap from average be
\[\Delta_{L}=\max_{j}\left[\frac{1}{N} \sum_{i=1}^{N} L(B(x_i), E(x_i)) - \frac{1}{N_j} \sum_{i : g_i^j=1}L(B(x_i), E(x_i))\right],\]
where $g_i^j=1$ denotes that point $x_i$ belongs to the $j$th subgroup defined by a specific protected attribute $g$ (\textit{e.g.} data points from females), and $N_j$ is the number of data points with $g^j=1$.
\end{definition}
Next, the mean fidelity gap amongst subgroups (Definition \ref{bg:def:subgroup_max_gap}) computes how much an explanation model's fidelity differs over subgroups.
Here, we only consider groups defined by the same sensitive attribute (\textit{e.g.} $g^k$ is male, $g^j$ is female).
\begin{definition}[Mean Fidelity Gap Amongst Subgroups: $\Delta^{group}_{L}$]
\label{bg:def:subgroup_max_gap}
Let the mean fidelity gap amongst subgroups be
\[\Delta_{L}^{group}=\frac{2}{G(G-1)}\sum_{k=1}^{G}\sum_{j=k+1}^{G}\left|\frac{1}{N_k} \sum_{i : g_i^k=1}L(B(x_i), E(x_i)) - \frac{1}{N_j} \sum_{i : g_i^j=1}L(B(x_i), E(x_i))\right|,\]
where $g^j$ denotes the $j^{th}$ subgroup defined by a specific sensitive attribute (\textit{e.g.} datapoints from females), and $N_j$ is the number of datapoints in $g^j$.
\end{definition}
Similar to average fidelity, we choose $L$ to be Accuracy, AUROC, and Mean Error for both fidelity gap measurements (e.g., $\Delta_{AUROC}$ and $\Delta^{group}_{AUROC}$). In all cases, we do not consider intersectional groups due to sample size concerns.
\subsection{Experiments Overview}
Since fidelity gaps across subgroups are closely linked to fairness preservation and risks of fairwashing, we design experiments to audit this quantity. We conduct the following experiments in the sections below\footnote{Code: \url{https://github.com/MLforHealth/ExplanationsSubpopulations}}:
\textbf{Measuring Fidelity Gaps Between Subgroups:} We measure fidelity gaps using metrics defined in ~\ref{bg:def:subgroup_avg_gap} and ~\ref{bg:def:subgroup_max_gap} for four different post-hoc explanation models, and two different blackbox model classes. The aim of this experiment is to study the presence and degree of fidelity gaps in standard explainability methods (Section~\ref{sec:fidelity_gaps_exps}).
\textbf{Assessing the Impact of Robust Training:} We use robust training strategies to train explanation models, and repeat the fidelity gap audits to study if robust training can provide reduced fidelity gaps (Section~\ref{sec:robust_training}).
\textbf{Studying Possible Causes for Fidelity Gaps}: We analyze the impact of blackbox fairness and presence of protected attribute information in feature representations on the fidelity gap (Section~\ref{sec:fair_fidelity_gaps}).
\textbf{Simulation Showing Impact of Fidelity Gaps}: We conduct a simulation and study the quality of decisions made for groups to examine the impacts of unfair explanation models on real-world decision making (Section~\ref{sec:real_world}).
\begin{table}[]
\begin{tabular}{llccccc}
\toprule
\textbf{Dataset} & \textbf{Outcome Variable} & \textbf{n} & $\bm{d}$ & $\bm{d}'$ & \textbf{Protected Attribute ($g$)} \\ \midrule
\texttt{adult} \cite{Dua:2019} & Income $>$ 50K & 48,842 & 9 & 33 & Sex (2 groups) \\
\texttt{lsac} \cite{wightman1998lsac} & Student passes the bar & 20,427 & 8 & 14 & Race (5 groups) \\
\texttt{mimic} \cite{harutyunyan2019multitask} & Patient dies in ICU & 21,139 & 49 & 49 & Sex (2 groups)\\
\texttt{recidivism} \cite{propublica2019} & Defendant re-offends & 6,150 & 6 & 7 & Race (2 groups) \\ \bottomrule
\end{tabular}
\caption{Binary classification datasets used in our experiments. $n$ is the number of samples, $d$ is the number of variables in the original dataset, and $d'$ is the number of features after one-hot encoding categorical variables. \label{tab:datasets}}
\end{table}
\section{Explanation Fidelity Varies Significantly between Subgroups}
\label{sec:fidelity_gaps_exps}
Experimentally, we find that fidelity gaps indeed vary by group in many settings. To show this, we train four post-hoc explainability methods (two local, two global) to explain two different blackbox models trained on the four standard fairness benchmark tabular datasets described in Table 1. Following ~\citet{aivodji2021characterizing}, we randomly split each dataset into four subsets: a training set for blackbox models (50\%), a training set for explanation models (30\%), a validation set for explanation models (10\%), and a held-out test set for evaluating both blackbox and explanation models (10\%). For each dataset, we train both a Neural Network (NN) and a Logistic Regression (LR) model to serve as blackboxes. See Section~\ref{sec:training_regimes} in the Appendix for details on the training regimes, hyperparameter settings, and evaluation metrics for each. In the following sections, we describe the explainability models and fidelity gaps observed.
\subsection{Local Explanation Models}
Local explanation models explain individual predictions from classifiers by learning an interpretable model locally around each prediction. In our experiments, we consider LIME \cite{ribeiro2016should,ribeiro2016model} and SHAP \cite{lundberg2017unified}, which are popular methods that use linear models to elicit each feature's contribution to the blackbox model's prediction. More details are in Appendix \ref{sec:explanation_model_descriptions}.
\textbf{Experiment Setup.} We measure fidelity gaps between subgroups using the two key metrics introduced in Section \ref{sec:metrics} (see Definitions ~\ref{bg:def:subgroup_avg_gap} and ~\ref{bg:def:subgroup_max_gap}).
For each, we select three performance measures: Accuracy ($\Delta_\text{Acc.}^\text{group}$) following prior work \cite{aivodji2019fairwashing}, mean residual error ($\Delta_\text{Err.}^\text{group}$), and also include AUROC ($\Delta_\text{AUROC}^\text{group}$) as a threshold-independent metric. A full table with all metrics can also be found in the Appendix (Section~\ref{sec:full_gaps_tables}).
For accuracy, we use a threshold of 0.5.
Since the four datasets are imbalanced, we use AUROC for model selection while tuning all hyperparameters.
Non-zero fidelity gaps indicate disparities across groups in the explanation models.
\textbf{Results.}
First, we find that LIME disproportionately favors different groups, as shown in Table \ref{tab:local}, where the maximum accuracy gap ranges from 0.1-21.4\%.
This confirms that explanation quality can dramatically differ by subgroup, even without access to group-membership data.
Furthermore, the AUROC/Accuracy between protected groups also ranges significantly (0-6.6\%/0.3-20.6\%), indicating that some members of protected groups are disadvantaged in terms of explanations. Hence, when explanations are judged to be ``high quality" based on average fidelity, it might be misleading and lead to errors in decision-making. Bolded non-zero fidelity gaps are also significantly greater than $0$ with a one-sided Wilcoxon signed-rank test at $p<0.05$.
Second, as expected, SHAP's gaps are consistently zero.
This is because the blackbox and explanation models are trained using identical features, in which case consistency is guaranteed~\cite{lundberg2017unified}.
However, using a subset of features to train the explanation model often leads to more useful explanations~\cite{wong2021leveraging}.
This increases the gaps significantly, as shown in Figure \ref{fig:auroc_gaps}, indicating that SHAP can also suffer from significant gaps in fidelity when used in practice. Since LIME considers sparsity as well, we also run this same experiment for LIME and find that fewer features are indeed associated with larger fidelity gaps (Fig.\ref{fig:auroc_gaps}). Increasing sparsity is a common approach in training explanation models and these experiments indicate that this technique alone may contribute to substantially worse fidelity gaps.
Third, we observe that the fidelity gaps in AUROC tend to be lower for the logistic regression blackbox, possibly because the linearity of the local surrogate models matches logistic regression better than the neural network. Note that the overall fidelity of all models are greater than 85\% (see Table~\ref{tab:full_gaps_local} in the Appendix).
\begin{table*}
\begin{tabular}{lccccc}
\toprule
Dataset & Blackbox Classifier & $\Delta_\text{Acc.}$ & $\Delta^\text{group}_\text{AUROC}$ & $\Delta^\text{group}_\text{Acc.}$ & $\Delta^\text{group}_\text{Err.}$ \\
\midrule
\multirow{2}{*}{\texttt{adult}} & Logistic Regression &
\textbf{0.8\% ± 0.0\%} & 0.1\% ± 0.0\% & \textbf{2.4\% ± 0.1\%} & \textbf{1.9\% ± 0.0\%} \\
& Neural Network &
\textbf{6.9\% ± 0.7\%} & \textbf{3.0\% ± 1.2\%} & \textbf{20.6\% ± 2.0\%} & \textbf{0.8\% ± 0.5\%} \\
\midrule
\multirow{2}{*}{\texttt{lsac}} & Logistic Regression &
\textbf{2.0\% ± 1.0\%} & 0.0\% ± 0.0\% & \textbf{1.5\% ± 0.5\%} & \textbf{1.5\% ± 0.1\%} \\
& Neural Network &
\textbf{21.4\% ± 4.4\%} & \textbf{6.6\% ± 1.2\%} & \textbf{12.2\% ± 2.2\%} & \textbf{3.8\% ± 1.2\%} \\
\midrule
\multirow{2}{*}{\texttt{mimic}} & Logistic Regression &
\textbf{0.4\% ± 0.6\%} & \textbf{3.0\% ± 1.8\%} & \textbf{1.1\% ± 0.3\%} & \textbf{2.0\% ± 0.1\%} \\
& Neural Network &
\textbf{0.8\% ± 0.4\%} & \textbf{ 1.7\% ± 1.5\%} & \textbf{1.4\% ± 0.7\%} & \textbf{1.7\% ± 0.5\%} \\
\midrule
\multirow{2}{*}{\texttt{recidivism}} & Logistic Regression &
0.1\% ± 0.1\% & 0.0\% ± 0.0\% & 0.3\% ± 0.2\% & 0.3\% ± 0.0\% \\
& Neural Network &
\textbf{0.9\% ± 0.3\%} & \textbf{0.7\% ± 0.3\%} & \textbf{2.4\% ± 0.7\%} & \textbf{1.1\% ± 0.1\%} \\
\bottomrule
\end{tabular}
\caption{Performance fidelity gaps across subgroups for \textit{LIME local explanations} using all available features. ± denotes standard deviation computed over 5 replications. Fidelity gaps are significant (one-sided Wilcoxon signed-rank tests at $p<0.05$; marked in \textbf{bold}) between all five groups in the \texttt{lsac} dataset, and between two sensitive groups in other three datasets. $\Delta_{Acc.}$ denotes the maximum fidelity gap of subgroups from average (in terms of accuracy at $0.5$ threshold), and $\Delta^{group}_{m}$ is the mean fidelity gap between subgroups using metric $m$.}\label{tab:local}
\vspace{-10mm}
\end{table*}
\begin{figure}[ht!]
\centering
\begin{subfigure}{.45\textwidth}
\includegraphics[width=1\linewidth]{images/AUROC_nn_local_gap.pdf}
\caption{Local Explanations}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\includegraphics[width=1\linewidth]{images/AUROC_nn_global_gap.pdf}
\caption{Global Explanations}
\end{subfigure}
\caption{The effect of varying the number of features on fidelity gaps using the \texttt{mimic} dataset with a neural network blackbox model. For (a) local explanation models, using fewer features leads to worse fidelity gaps. We observe larger fidelity gaps across subgroups with sparser models, i.e., fewer features in (a) local explanation models. For (b) global explanation models, the gap varies with number of features. We also observe similar trends on other datasets (see Appendix~\ref{sec:varying_features_plots_all}).
}\label{fig:auroc_gaps}
\vspace{-5mm}
\end{figure}
\subsection{Global Explanation Models}
Global explanation methods train one new surrogate model that approximates the behavior of a blackbox model.
This surrogate model should itself be easily understood, and can then be used instead of the blackbox at test time (more background in Appendix \ref{sec:explanation_model_descriptions}).
\textbf{Experiment Setup.} In this experiment, we generate global explanations using two popular choices of interpretable surrogate models: Generalized Additive Model (GAM) \cite{hastie2017generalized} and a sparse decision tree (Tree)~\cite{pedregosa2011scikit}.
GAM combines linear models of different variables during explanation \cite{tan2018distill}, while Tree uses a low-depth, sparse, decision tree.
We evaluate the fidelity of each global method with the original blackbox and compare across subgroups.
As with the local methods, we use both Accuracy and AUROC to evaluate fidelity gaps.
\textbf{Results.}
First, we find that the fidelity gap between subgroups differs substantially from the average for the global explanation models, as shown in Table \ref{tab:global} where the accuracy gap ranges from 0-13.5\%. We again observe that AUROC and accuracy vary substantially between protected subgroups (0-8.1\% and 0.1-7.4\% for protected groups such as sex and race groups in each dataset).
This is especially true for more imbalanced subgroup proportions: having more subgroup categories leads to more disadvantage in protected groups, particularly when the classes are imbalanced themselves (e.g. \texttt{lsac}).
Second, we find that using fewer features (e.g., 15 in Fig.~\ref{fig:auroc_gaps}) may lead to larger gaps in performance between subgroups in sparse decision trees (Trees), bolstering prior findings on training trustworthy models \cite{chang2021interpretable}.
Hence, the gaps shown in Tables \ref{tab:local} and \ref{tab:global} are likely underestimates when using fewer dimensions in explanation models, which is common.
Interestingly, the subgroup with the lowest-quality explanations is not always the minority subgroup---which may be the most disadvantaged---in the datasets for fair ML. We expand this finding in Table \ref{tab:worst_case_groups} in the Appendix.
Additionally, we see that subgroup gaps occur even after training blackbox models with a balanced number of data points from each subgroup for both global and local explanation models (see Table~\ref{tab:balanced_global}).
\begin{table*}[t]
\begin{tabular}{lccccccc}
\toprule
Dataset & Blackbox Classifier & Expl. Model & $\Delta_\text{Acc.}$ & $\Delta^\text{group}_\text{AUROC}$ & $\Delta^\text{group}_\text{Acc.}$ & $\Delta^\text{group}_\text{Err.}$ \\
\midrule
\multirow{4}{*}{\texttt{adult}} & Logistic Regression & GAM & 0.1\% ± 0.0\% & 0.0\% ± 0.0\% & 0.3\% ± 0.0\% & 0.1\% ± 0.0\% \\
& Logistic Regression & Tree & \textbf{1.5\% ± 0.1\%} & \textbf{2.9\% ± 0.4\%} & \textbf{4.5\% ± 0.2\%} & \textbf{1.1\% ± 0.1\%} \\
& Neural Network & GAM &
\textbf{0.8\% ± 0.2\%} & \textbf{0.5\% ± 0.3\%} & \textbf{2.4\% ± 0.5\%} & 0.3\% ± 0.2\% \\
& Neural Network & Tree & \textbf{1.1\% ± 0.1\%} & \textbf{0.6\% ± 0.4\%} & \textbf{3.4\% ± 0.2\%} & \textbf{0.5\% ± 0.4\%} \\
\midrule
\multirow{4}{*}{\texttt{lsac}} & Logistic Regression & GAM &
\textbf{0.9\% ± 0.9\%} & 0.0\% ± 0.0\% & \textbf{0.6\% ± 0.4\%} & \textbf{0.7\% ± 0.3\%} \\
& Logistic Regression & Tree &
\textbf{3.7\% ± 3.1\%} & \textbf{1.1\% ± 0.4\%} & \textbf{2.8\% ± 0.7\%} & \textbf{1.8\% ± 0.5\%} \\
& Neural Network & GAM &
\textbf{13.5\% ± 0.9\%} & \textbf{5.2\% ± 1.2\%} & \textbf{7.3\% ± 1.0\%} & \textbf{3.9\% ± 2.6\%} \\
& Neural Network & Tree &
\textbf{11.5\% ± 2.7\%} & \textbf{5.8\% ± 2.1\%} & \textbf{ 7.4\% ± 1.2\%} & \textbf{4.9\% ± 2.0\%} \\
\midrule
\multirow{4}{*}{\texttt{mimic}} & Logistic Regression & GAM &
0.5\% ± 0.1\% & 0.4\% ± 0.1\% & \textbf{0.9\% ± 0.1\%} & 0.4\% ± 0.2\% \\
& Logistic Regression & Tree &
\textbf{0.6\% ± 0.0\%} & \textbf{8.1\% ± 0.8\%} & \textbf{1.2\% ± 0.1\%} & \textbf{1.9\% ± 0.0\%} \\
& Neural Network & GAM &
\textbf{1.2\% ± 0.3\%} & \textbf{1.8\% ± 1.2\%} & \textbf{2.2\% ± 0.6\%} & \textbf{0.9\% ± 0.3\%} \\
& Neural Network & Tree &
\textbf{1.1\% ± 0.5\%} & \textbf{3.0\% ± 1.5\%} & \textbf{2.0\% ± 0.9\%} & \textbf{1.9\% ± 0.9\%} \\
\midrule
\multirow{4}{*}{\texttt{recidivism}} & Logistic Regression & GAM & 0.1\% ± 0.0\% & 0.1\% ± 0.0\% & 0.3\% ± 0.0\% & 0.5\% ± 0.0\% \\
& Logistic Regression & Tree & 0.0\% ± 0.0\% & 0.4\% ± 0.0\%& 0.1\% ± 0.0\% & \textbf{ 1.2\% ± 0.0\%} \\
& Neural Network & GAM &
0.2\% ± 0.2\% & 0.4\% ± 0.2\% & 0.6\% ± 0.6\% & \textbf{1.1\% ± 0.4\%} \\
& Neural Network & Tree &
\textbf{0.9\% ± 0.3\%} & \textbf{1.0\% ± 0.9\%} & \textbf{2.3\% ± 0.7\%} & \textbf{1.4\% ± 0.3\%} \\
\bottomrule
\end{tabular}
\caption{Fidelity gaps across subgroups for \textit{global} explanation models GAM and Tree. ± denotes standard deviation computed over 5 replications. Fidelity gaps are significant (one-sided Wilcoxon signed-rank tests at $p<0.05$; marked in \textbf{bold}) for all five groups in the \texttt{lsac} dataset, and between two sensitive groups in the other three datasets with both global explanation models. $\Delta_{Acc.}$ denotes the maximum fidelity gap of subgroups from average (in terms of accuracy at $0.5$ threshold), and $\Delta^{group}_{m}$ is the mean fidelity gap between subgroups using metric $m$.}\label{tab:global}
\vspace{-10mm}
\end{table*}
\section{Balanced and Robust Training Reduces Fidelity Gaps}\label{sec:robust_training}
Balanced and robust training methods could provide a path towards improving fidelity gaps, thereby learning fairer explanations~\cite{han2021balancing,adragna2020fairness}.
We showcase two such robust training methods, one for local methods and one for global methods.
For both cases, we choose hyperparameters that maximize the worst-case fidelity across all groups.\footnote{The overall fidelity is not significantly affected by either training approach.}
Ultimately, our experiments indicate that while robust training improves fidelity gaps sometimes, they remain largely pervasive.
\subsection{Robust Local Explanation Models}
\textbf{Experiment Setup.} We train a more-robust version of LIME, using Just Train Twice (JTT)~\cite{liu2021just}, a two-stage training paradigm for training robust ML Models. First, we train an identification model via empirical risk minimization. Then, we extract its set of misclassified training examples. A final model is then trained by upsampling these misclassified examples, scaled by a hyperparameter $\lambda$. This reweighted loss is designed to make the second model more robust.
We use JTT to train LIME's local linear approximations, using linear models for both the identification and final models.
\textbf{Results.}
JTT successfully reduces gaps on three datasets with a NN blackbox model, as shown in Figure~\ref{fig:mod_gaps_lime}.
Interestingly, this is not the case for the \texttt{recidivism} dataset, where JTT does not reduce the gaps and performs the same as standard training. With LR blackboxes (Figure~\ref{fig:mod_gaps_lime_all} in Appendix), the fidelity gaps are already small, so JTT is less impactful. However, non-zero gaps between 1-2\% still persist (e.g., NN blackbox on the \texttt{adult} dataset), indicating that the error-prone regions did not generalize to the test setting. Measuring fidelity gaps is therefore still critical, even if an explanation model is trained to be robust.
\subsection{Robust Global Explanation Models}
\label{sec:robust_exps}
\textbf{Experiment Setup.} We next study balanced training for the global explanation method Tree.
We rebalance the explainability training sets for each dataset by randomly oversampling minority groups, a common approach for improving test error on minority subpopulations~\cite{wei2013role,han2021balancing}. This way, the training set in each case consists of an equal number of examples from each protected group.
Then, we train a Tree model to explain each blackbox model using these balanced datasets.
\textbf{Results.}
As shown in Figure \ref{fig:mod_gaps_lime}, we find that this common rebalancing approach to does not reduce gaps significantly across the board.
Still, some cases look more promising than others.
For example, \texttt{mimic} with NN which indicates this may be a fruitful direction for learning fairer explanations.
This is especially true for cases like \texttt{mimic} with LR, where rebalancing the training set increases the fidelity gap substantially (see Figure~\ref{fig:mod_gaps_lime_all} in Appendix).
\begin{figure*}[t!]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{images/jtt_training_lime_nn_.png}
\caption{Local Explanations (LIME)}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{images/balanced_nn.png}
\caption{Global Explanations (Tree)}
\end{subfigure}
\caption{AUROC Fidelity gaps across subgroups with and without robust training for (a) LIME and (b) Tree-based Models. Improvements are significant with robust training for \texttt{adult} dataset in (a) with a Wilcoxon-signed rank test at $p<0.1$ level. However, balanced training does not help for most datasets in (b). Error bars indicate 95\% confidence intervals.}\label{fig:mod_gaps_lime}
\end{figure*}
\section{On Possible Causes for Fidelity Gaps}
\label{sec:fair_fidelity_gaps}
In our fidelity gap audits in prior sections, we noticed that the fidelity gaps are largest for the least-fair blackboxes (\texttt{adult} and \texttt{lsac} datasets; Tables \ref{tab:global} and \ref{tab:local}). This indicates a potential relationship between the fairness of the blackbox and explainability models. To explore this further, we study the associations between blackbox fairness and fidelity gaps across subgroups in this section. First, we train fair models, and observe that significant non-zero fidelity gaps still persist. Second, we study if protected group information -- e.g. if a data point belongs to a Male or Female individual -- can be predicted from the \emph{feature representations} alone, following prior work in fair representation learning \cite{madras2018learning,zhang2018mitigating}. We find that complex mechanisms by which protected group information can be indirectly predicted could be contributing factors to the fidelity gaps observed.
\subsection{Training Fair Blackbox Models}
Our experiments in previous sections (see Tables~\ref{tab:local} and ~\ref{tab:global}) indicate that fidelity gaps across subgroups occur regardless of the blackbox model's fairness with respect to groundtruth label predictions.
For example, a logistic regression model trained on the \texttt{mimic} dataset is fair with respect to the sex (Table~\ref{tab:blackbox_groundtruth} in Appendix; DP gap of 1\%). However, fidelity gaps are non-zero across sex subgroups with a sparse decision tree global explanation (8.1\%). Strikingly, the gaps in fidelity AUROC often exceed the gaps in AUROC of the blackbox models themselves (e.g., \texttt{mimic} dataset with the Tree model, the difference in classification performance of the blackbox between Male and Female individuals is 3.6\%, while the fidelity gap between subgroups is 8.1\%). However, we do observe larger fidelity gaps in datasets where blackbox models are more accurate but less fair (e.g. \texttt{adult} with an absolute DP gap of 16-17\%).
\textbf{Experiment Setup.} To investigate this further, we train debiased neural network blackbox models for \texttt{lsac} and \texttt{mimic}: both highly-imbalanced by class label, and characterized by the largest and smallest gaps in blackbox model AUROC respectively. Adversarial debiasing following methodology proposed by ~\citet{zhang2018mitigating}\footnote{We use an open-sourced implementation: \url{https://github.com/Trusted-AI/AIF360}} is utilized, wherein an adversary tries to predict the protected group information from classification predictions and labels, while the main classification model (our blackbox model) jointly predicts the primary classification outcome. We use demographic parity as the desired fairness definition.
Our results are shown in Table \ref{tab:fair_blackbox_groundtruth}, which reports the performance of the fair(er) blackbox classifiers.
\textbf{Results.} We debias neural network blackbox models to be fairer, where a model is deemed to be fair if it has an absolute DP gap close to 1\% (9\% and 0.6\% after debiasing for \texttt{lsac} and \texttt{mimic} respectively; improved from 14\% and 2\%). We find that despite fair training, fidelity gaps remain (Table~\ref{fig:Identifiability of groups}, though they are significantly reduced in most cases: the fidelity gap in accuracy decreases from 2\% to 1\% for \texttt{mimic} and 7.4\% to 0.8\% for \texttt{lsac} (Tree)). Note that these results are dependent on our choice of fairness criterion: particularly, for the \texttt{lsac} dataset we find that overall performance is reduced to achieve parity (Table~\ref{tab:fair_blackbox_groundtruth} in Appendix; for absolute DP gap close to 1\%, over 99\% of the blackbox predictions are that the student passes the bar). This indicates that while fair blackboxes could potentially reduce fidelity gaps across subgroups, there might be trade-offs and more potential causes for such gaps. We explore this in Section~\ref{sec:group_pred}.
\subsection{Performance of Predicting Protected Attributes from Feature Representations Alone}
\label{sec:group_pred}
\textbf{Experiment Setup.} One way to achieve group fairness is by removing group information from or debiasing the representations (e.g. with the use of an adversary~\cite{wadsworthachieving,madras2018learning}). Here, we quantify the amount of group information that is present in the data. This is relevant as the absence of group information is a sufficient condition to achieving fairness parity according to standard metrics in fair machine learning. For example, consider equality of opportunity for the positive class, which can be written as $\hat{Y} \indep G | Y = 1$, where G denotes the protected group and Y is the binary groundtruth label. If we have $X \indep G | Y = 1$, then equality of opportunity is achieved for any form of the classifier $\hat{Y} = f(X)$, including $E$ the explanation model and $B$ the blackbox model. Therefore, if no protected group information is present in the positive examples, then the explanation fidelity would not differ between protected groups for positive examples \citep{madras2018learning}.
In this experiment, we first compute the accuracy of detecting the protected group information from all datasets (with a cross-validated model). Then, we select features that have zero mutual information with respect to the protected attribute, and only use these in training the blackbox and explanation model. We expect that the performance of predicting group information from these features will be low. Then, we compute the fidelity gaps -- this allows us to answer the question: do fidelity gaps exist when there is low group information in the data?
\textbf{Results}: First, we observe that in all cases the prediction AUROC is significantly greater than 0.5 (see Table~\ref{fig:Identifiability of groups}) in identifying the minority group. This indicates that the protected group information -- e.g. if the datapoint belongs to a Male/Female individual -- can be predicted with good performance from the feature representations alone\footnote{This was also observed when conditioned on only positive or negative label classes}. Since past work has shown that group information might be indirectly constructed and used by explanations~\cite{lakkaraju2020fool}, this is important to consider. Second, we only use features that have zero mutual information with respect to protected attribute labels in the \texttt{lsac} dataset for training blackbox and explanation models. We see that all models output a single-class prediction which limits our ability to make meaningful conclusions about the impact of group information on this dataset -- note the fidelity gaps are technically zero, though the explanations are trivial. We repeat the same procedure with the \texttt{mimic} dataset by selecting 10 features. The AUROC of predicting protected attribute (sex) from these features is low ($0.54$; also less than $0.57$ for features from positive and negative class). With this representation, we train both NN and LR blackbox models, and GAM/Tree global explanation models. We observe that accuracy-based fidelity gaps ($\Delta_{Acc.}$, $\Delta_{Acc.}^{group}$) decrease to low values not much higher than zero (to 0-0.6\% with GAM and Tree; full table in Appendix~\ref{tab:group_removed_subset} while blackbox model's AUROC is greater than $0.7$). This indicates that fidelity gaps decrease when there is less group information in data representations. However, non-zero fidelity gaps in AUROC still persist for Tree-based models (up to 6.6\%). This is due to low prevalence of positive class predictions with the blackbox model on using the reduced data representation ($\approx$3\% positive class at $0.5$ threshold), which has a large impact on AUROC (since it is a ranking-based metric, and sensitive to degree of imbalance). We highlight that more experiments using interpretable, completely group-independent representations (i.e., an AUROC of 0.5 in predicting protected attribute labels) that still have high groundtruth predictive capability are required to accurately quantify the impact of group information on AUROC-based fidelity gaps. We also note that class imbalance -- and varying degrees of class imbalance for data subgroups -- may be an important factor. Our findings indicate that fidelity gaps persist across a range of class-imbalance ratios, but we leave the estimation of the effect of varying degrees of class imbalance (or positive-class prevalence) across subgroups on explanation fairness for future work.
\begin{figure*}
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=0.98\textwidth]{images/auroc_predicting_sens.pdf}%
\caption{AUROC in detecting minoritized protected group}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\raggedleft
\begin{tabular}{llccc}
\toprule
Dataset & Expl & $Fidelity^{Acc.}$ & $\Delta^\text{group}_{AUROC}$ & $\Delta^\text{group}_{Acc.}$
\\
\midrule
\multirow{2}{*}{\texttt{lsac}} & GAM & 96.6 & 1.4 & 1.5 \\
& Tree & 96.9 & 7.0 & 0.8 \\
\midrule
\multirow{2}{*}{\texttt{mimic}} &GAM & 96.2 & 0.6 & 1.0 \\
& Tree & 94.9 & 4.1 & 1.6 \\
\bottomrule
\end{tabular}
\vspace{8mm}
\caption{Fidelity gaps with fair blackbox models}
\end{subfigure}
\caption{
We find minoritized protected groups can be detected with high AUROC from feature representations alone in (a). As a result, non-zero fidelity gaps persist even when underlying blackboxes are fair as seen in (b).
}\label{fig:Identifiability of groups}
\vspace{-5mm}
\end{figure*}
\section{Simulating the Real-World Impact of Biased Explanations}\label{sec:real_world}
Unfair explanation models can have negative effects on real-world decision making.
To demonstrate this, we conduct a simulation study of ML-assisted law school admissions using the \texttt{lsac} dataset \cite{wightman1998lsac}.
Such systems are already being used in many cases \cite{martinez2021using,pangburn_2019}.
Our results clearly show that worse decisions are made for members of disadvantaged groups when explanations are less fair.
\textbf{Experiment Setup.} To set up our simulation study, we consider an admissions officer that uses a blackbox model that predicts whether a student will pass the bar, though this prediction may be incorrect.
The admissions officer also has an explanation of the model's prediction, which may have low fidelity.
The blackbox model's performance and the explanation fidelity can vary between protected groups---we vary these parameters in this experiment.
We assume that the admissions officer then admits students solely based on their perceived likelihood of the applicant passing the bar, without any knowledge of the applicant's demographics. We assume parameters for the probability that the officer ultimately makes the correct decision. We obtain these parameters from prior user studies assessing the impact of explanations on human+AI decision making accuracy for a different task \cite{bansal2021does}, but believe they serve as reasonable estimates to display the anchoring effect of decisions with explanations observed across a variety of decision-making settings~\cite{poursabzi2021manipulating,bansal2021does}. For further details, please see Appendix \ref{app:sim}.
To simulate the effect of fidelity gaps on decision-making accuracy, we vary the maximum fidelity gaps between the two groups (males and females) and the average from 0\% to 15\%, assuming an average fidelity of 85\% across groups.
We then compute the admissions officer's resulting decision-making accuracy for males and females.
We use sex as the protected attribute of interest for the simulation as both groups pass the bar equally in the dataset, so decision-making accuracy is a fair performance metric.
\textbf{Results.}
As shown in Figure \ref{fig:expl_gap_sim_nn}, we find that larger fidelity gaps lead to larger decision accuracy gap between groups.
So when explanations are less fair, disadvantaged groups may be targeted by worse decisions.
With over 60,000 law school applicants in the U.S. in a typical year \cite{lsac_2018}, over 200 applications would be wrongly admitted/rejected based solely on the fairness of the explanation model, according to this simulation.
Fidelity gaps should therefore be used as fairness metrics for explanation methods: minimizing these gaps leads to fairer decisions.
However, we emphasize that the findings of this simulation are based on some strong assumptions (e.g., reliance on parameters extracted, anchoring effect existence in this admission setting, etc.). Real-world user studies are required to validate these expected findings rigorously across a variety of decision-making setups.
\begin{figure}
\centering
\includegraphics[scale=0.5]{images/expl_gap_sim_nn.pdf}
\caption{Effect of fidelity gap size on a simulated admissions officer's decision accuracy between males and females using a neural network blackbox. Note that larger fidelity gaps lead to larger decision accuracy gaps between males and females; fidelity gaps could disadvantage different groups in practice. Each line is paired with 95\% confidence intervals across 20 simulations.
}\label{fig:expl_gap_sim_nn}
\end{figure}
\section{Discussion}
\subsection{Takeaways for ML Practitioners}
\textbf{Analyze subgroup fidelities.} Our results suggest that ML practitioners using post-hoc explainable models to interpret blackbox models should carefully analyse the fidelity of commonly-used explanations for different groups separately. Especially if there is a target subgroup of interest. If a fidelity gap exists, practitioners should carefully consider its source \cite{wick2019unlocking}, and, where possible, take measures to minimize the impacts on downstream decision-making \cite{rajkomar2018ensuring}. We also highlight the importance of carefully choosing the metric for measuring fidelity (e.g., accuracy, AUROC, precision, etc.): different metrics may be affected by properties of the dataset and hence predictions from a blackbox/explanation model (e.g., class imbalance) differently.
\textbf{Consider the explanation model.} Overall, our findings indicate the existence of fidelity gaps between subgroups is both a \textit{model} and a \textit{data} issue. From Section \ref{sec:fidelity_gaps_exps}, we find that fidelity gaps can vary greatly for the same dataset depending on the explanation model used, and our results in Section \ref{sec:robust_training} show that algorithms that seek to improve worst-case group performance may be a promising direction in reducing fidelity gaps. As such, we recommend careful selection and testing of various explanation models in order to select an equitable model with high overall fidelity.
In addition, model hyperparameters should also be carefully tuned. For example, in models like LIME, there are several hyperparameters that can affect fidelity gaps, such as the sampling variance (Figure~\ref{fig:sigma_var_lime} in Appendix), the number of perturbations, or number of features in the explanation. Exploring the effect of these hyperparameters on explanation quality and fairness is a promising direction of future work. Lastly, extending prior work on methods for fair supervised ML models~\cite{caton2020fairness,paul2021generalizing,madras2018learning,zhang2018mitigating,liu2021just}, we call for similar approaches to training fair and explainable local and global explanation models which have reduced fidelity gap in addition to high overall fidelity.
\textbf{Consider the data.} Our results in Section \ref{sec:fair_fidelity_gaps} indicate that fidelity gaps also depend on data representations. Because some feature representations cause smaller fidelity gaps, practitioners should carefully consider the features used to learn both the blackbox and explanation models \cite{grgic2016case}.
As machine learning models can encode historical biases present in the training corpora \cite{doughman2021gender}, it is crucial to consider the source of such potential biases, and, if possible, take actions to correct them by collecting additional data in a fairness-aware way \cite{anahideh2020fair, jo2020lessons}.
\subsection{Implications of Fidelity Gaps}
\textbf{Algorithmic Modifications to Train Fair Explanation Models with Low Fidelity Gaps.} While we benchmark robust training as an attempt to mediate the fidelity gaps across subgroups, we posit that data distribution-aware methods could lead to lower, less significant fidelity gaps. For example, recent work in causal bootstrapping~\cite{imbens2018causal}
shows the potential to reduce confounding biases in ML model training given causal knowledge~\cite{gowda2021pulling}. Similar strategies relying on partial or complete knowledge of the data generation graph~\cite{geirhos2018imagenet} could prove effective in selecting features and training examples to train fair explanation models.
Post-processing solutions to standard explanation model training could also prove effective, similar to recent work in the space of improving worst-case generalization~\cite{menon2020overparameterisation}. However, such solutions need to be appraised carefully to ensure that the resulting models are both \emph{fair} and remain \emph{interpretable} to users~\cite{wang2020pursuit}.
An interesting follow-up question is whether it is possible to have zero fidelity gaps---perfect worst-case generalization---while retaining good average fidelity under standard training settings. Zero fidelity gaps are possible, of course, when the blackbox and explanation models are identical.
However in more-realistic scenarios, fidelity gaps may simply depend on the data distributions~\cite{menon2020overparameterisation}.
For example, rare subgroups may be more diffidult to approximate, and will naturally have lower fidelity than others~\cite{samadi2018price,hooker2020characterising}.
\textbf{Fidelity Gaps as an Evaluation Metric.} We focus mainly on evaluating the fairness of explanation models using the \textit{fidelity gap} as a metric, assuming that models with smaller fidelity gaps are more desirable. However, recent work in group fairness has found that trying to achieve equal performance for all subgroups tends to worsen welfare for all \cite{hu2020fair, corbett2018measure, zhang2022improving}. Such a fairness/accuracy trade-off is well-documented in the algorithmic fairness literature \cite{kearns2019empirical, zliobaite2015relation}. We posit that there is likely a similar trade-off between the fidelity gap and the overall fidelity of an explanation model. In such cases, motivated by definitions such as minimax Pareto fairness \cite{martinez2020minimax, diana2021minimax}, it may be more appropriate to select explanation models that maximize the fidelity of the worst-case group.
\textbf{Human Implications of Fidelity Gaps.} Explainable ML models form an integral part of sociotechnical systems, given their user-facing nature~\cite{buccinca2020proxy}. Several works have studied the utility of explanations in human--AI joint decision-making~\cite{bansal2021does}. However, the potential failure modes we identify---fidelity gaps leading to worse explanations for some groups---need to be studied further in the context of real human decision-making (in addition to the simulations we conduct). For accurate decision-making in practice, like learning to defer decisions to an expert~\cite{pmlr-v119-mozannar20b}, it is important to communicate clearly and provide end-users with details of performance caveats~\cite{liang2019implicit}.
This requires collaboration between computer scientists and scholars working in the space of computer-mediated communication. A more design-centric approach is required to bridge the gap between researchers and consumers of these models~\cite{chen2022interpretable}.
\section{Conclusion}
In this work, we investigate fairness properties of post-hoc explainability methods.
We ultimately find that significant gaps in performance exist between groups, indicating that some groups receive better explanations than others.
First, we demonstrate experimentally that significant gaps occur in the two main branches of explanation methods using four explainability methods on four common datasets and two blackbox models.
Second, we present a study of robust training methods for improving these gaps.
We find that these methods can improve the fairness of explanation models in some cases.
Third, using a simulation study, we demonstrate that improving explanation fairness could substantially improve decision making accuracy for underserved groups.
Finally, we pose promising directions enhancing post-hoc explainability methods; future work should focus on ensuring explanation quality does not suffer according to group membership while remaining reliable and accurate.
\begin{acks}
Aparna Balagopalan is supported by a grant from the MIT-IBM Watson AI Lab. Haoran Zhang is supported by a grant from the Quanta Research Institute. Dr. Frank Rudzicz is supported by a CIFAR AI Chair. Dr. Marzyeh Ghassemi is funded in part by Microsoft Research, and a Canadian CIFAR
AI Chair held at the Vector Institute. Resources used
in preparing this research were provided, in part, by
the Province of Ontario, the Government of Canada
through CIFAR, and companies sponsoring the Vector Institute. We would like to thank Hammaad Adam, Bret Nestor, Natalie Dullerud, Vinith Suriyakumar, Nathan Ng, and three anonymous reviewers for their valuable feedback.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
|
\subsection{Our results}
\label{sec:intro:setup}
In this work, we focus on the simple yet prevalent case where the potential
outcomes are linear in the collected features (as in the housing market
disequlibrium example), with a residual (random) error term that is uncorrelated
across potential outcomes.
More precisely,
we consider a setting involving $n$ individuals
(observations) and $k$ potential outcomes (models). Each individual $i \in [n]$ has a feature
vector $\vec{x}^{(i)} \in \mathbb{R}^d$ and each model $j \in [k]$ has a vector of regression
parameters $\vec{w}^*_j \in \mathbb{R}^d$. Then, each individual with feature vector
$\vec{x}$ recieves a label for each outcome $j$ equalling
$\smash{y_{j} = \vec{w}_j^{* \top} \vec{x} + \varepsilon_{j}}$,
where $\vec{\varepsilon}$ is assumed to be an standard multivariate Gaussian random
variable. Our setting will involve self-selection bias arising from the fact
that every individual will only choose to reveal one of their labels,
$j_* \in \{1, \ldots, k\}$, as determined by some function of
that individual's full set of labels $y_{1}, \ldots, y_{k}$.
Our goal in this paper is to estimate the $k$ parameter vectors
$\vec{w}^*_1$, $\dots$, $\vec{w}^*_k$ under both the known-index and
unknown-index observational models.
We begin with our results in the known-index setting, which we formally define
in Definition~\ref{defn:index_observed}:
\begin{description}
\item[Known-index Setting:] We observe $n$ samples
$(\vec{x}^{(i)}, y^{(i)}, j_*^{(i)})$, where
$j_*^{(i)} = S(y_{1}^{(i)}, \dots, y_{k}^{(i)})$, and
$y^{(i)} = y_{j_*^{(i)}}^{(i)}$ for
some known self-selection rule
$S : \mathbb{R}^k \rightarrow \{1, \ldots, k\}$.
\end{description}
Here, we can estimate the unknown parameter vectors $\vec{w}^*_1$, $\dots$,
$\vec{w}^*_k$ to arbitrary accuracy and we allow for quite general
self-selection rules.
In particular, we allow the self-selection rule $S$ to be any
\textit{convex-inducing rule} (Definition~\ref{defn:convex-inducing}):
letting $\vec{y} \in \mathbb{R}^k$ be the vector of potential
outcomes, if we fix the $j$-th
coordinate $y_j$ for any $j \in [k]$, then deciding whether $j$ will be the
winner (i.e., whether $j_* = j$) is the same as
deciding whether $\vec{y}_{-j}$ belongs to some convex set (this convex set
can depend on $y_j$).
We formally define the self-selection rule in Definition
\ref{def:selfSelectionRule} and the convex-inducing self-selection rule in
Definition \ref{defn:convex-inducing}. Below we present an informal version of
our estimation theorem for the known-index case. The corresponding formal
version of the theorem can be found in Theorem \ref{thm:knownIndex}.
\begin{inftheorem}[Known-Index Estimation] \label{infthm:knownIndex}
Let $(\vec{x}^{(i)}, y^{(i)}, j_*^{(i)})_{i = 1}^n$ be $n$ observations from
the known-index self-selection model with $k$ linear models $\vec{w}^*_1$,
$\dots$, $\vec{w}^*_k$ as described in Section
\ref{sec:intro:setup}. If the self-selection rule is convex-inducing and that the probability of observing each model is lower
bounded by $(\alpha / k)$ for some $\alpha > 0$, then there is an estimation algorithm that outputs
$\hat{\vec{w}}_1$, $\dots$, $\hat{\vec{w}}_k$ with
$\norm{\hat{\vec{w}}_i - \vec{w}_i^*} \le \varepsilon$, when
$n \ge \mathrm{poly}(d, k, 1/\alpha, 1/\varepsilon)$. Furthermore, the running time of the algorithm is
$\mathrm{poly}(d, k, 1/\alpha, 1/\varepsilon)$.
\end{inftheorem}
The case of unknown-index is significantly more challenging and as we already
mentioned even the with infinite number of samples it is unclear if we have
enough information to estimate $\vec{w}_j$'s. We define the setting formally in
Definition \ref{defn:index_unobserved} and informally below:
\begin{description}
\item[Unknown-index Setting:] We observe $n$ samples of the
form $(\vec{x}^{(i)}, y^{(i)})$, where $\vec{x}^{(i)} \sim \mathcal{N}(0,
\bm{I}_d)$, and
$y^{(i)} = \max_{j \in [k]} y_{j}^{(i)}$.
\end{description}
Our first result below shows this
identifiability and its formal version is Theorem \ref{thm:identifiability_no_index}.
\begin{inftheorem}[Unknown-index Identifiability] \label{infthm:unknownIndex:identifiability}
If we have infinitely many samples from the unknown-index
setting, then we can identify
all $k$ linear models $\vec{w}^*_1$, $\dots$, $\vec{w}^*_k$.
\end{inftheorem}
Next we continue with finite-sample and finite-time algorithms for the
unknown-index case. To achieve this problem we need some separability assumption
between the $\vec{w}_i^*$.
\begin{infassumption}[Separability Assumption -- See Assumption \ref{as:no_index}] \label{infasp:separability}
The separability assumption that we use is that the projection of any vector
$\vec{w}^*_i$ to the direction of any other vector $\vec{w}^*_j$ cannot be
larger that the norm of $\vec{w}_j^*$ (and in fact, must be at least
$\Delta$ smaller).
\end{infassumption}
Our first result for arbitrary $k$ guarantees the estimation of the parameter
vectors $\vec{w}^*_i$ within accuracy $1/\mathrm{poly}(k)$. For a formal statement of
the theorem below we refer to Theorem \ref{thm:no_index}.
\begin{inftheorem}[Unknown-index Estimation -- General $k$ Case] \label{infthm:unknownIndex:general}
Let $(\vec{x}^{(i)}, y^{(i)})_{i = 1}^n$ be $n$ observations from a
self-selection setting with $k$ linear models $\vec{w}^*_1$, $\dots$,
$\vec{w}^*_k$ as described in the unknown-index setting of Section
\ref{sec:intro:setup}. If we also assume Informal Assumption
\ref{infasp:separability} and that the probability of observing each model is
lower bounded by $(\alpha / k)$ for some $\alpha > 0$, then there exists an estimation algorithm that outputs
$\hat{\vec{w}}_1$, $\dots$, $\hat{\vec{w}}_k$ with
$\norm{\hat{\vec{w}}_i - \vec{w}_i^*} \le 1/\mathrm{poly}(k)$, assuming that
$n \ge \exp(\mathrm{poly}(k)) \cdot \mathrm{poly}(d, 1/\alpha)$. Furthermore, the running time of our
algorithm is also $\exp(\mathrm{poly}(k)) \cdot \mathrm{poly}(d, 1/\alpha)$.
\end{inftheorem}
Finally, we study the case where $k = 2$ and show how to estimate the
parameters to arbitrary accuracy. For this we only need a more direct
separability assumption on the parameter vectors:
\begin{infassumption}[$k = 2$ Separability Assumption -- See Assumption \ref{as:no_index:2}] \label{infasp:separability:2}
The two weight vectors are separated in Euclidean distance, i.e.,
$\norm{\vec{w}_1^* - \vec{w}_2^*}_2 \geq \Delta$.
\end{infassumption}
Under the above assumption we can recover the parameters efficiently: for the
formal version of the informal theorem below we refer to Theorem
\ref{thm:no_index:2}.
\begin{inftheorem}[Unknown-index Estimation -- $k = 2$ Case] \label{infthm:unknownIndex:2}
Let $\varepsilon > 0$ and $(\vec{x}^{(i)}, y^{(i)})_{i = 1}^n$ be $n$ observations from a
self-selection setting with $2$ linear models $\vec{w}^*_1$, $\vec{w}^*_2$
as described in the unknown-index setting of Section \ref{sec:intro:setup}. Under Informal Assumption \ref{infasp:separability:2}, there exists an estimation
algorithm that outputs $\hat{\vec{w}}_1$, $\hat{\vec{w}}_2$ with
$\norm{\hat{\vec{w}}_i - \vec{w}_i^*} \le \varepsilon$, assuming that
$n \ge \mathrm{poly}(d, 1/\alpha, 1/\varepsilon)$. Furthermore, the running time of the
algorithm is also $\mathrm{poly}(d, 1/\alpha, 1/\varepsilon)$.
\end{inftheorem}
\subsection{Our Contributions and Techniques} \label{sec:intro:contributions}
We initiate a line of work on attaining statistical and
computational efficiency guarantees in the face of structured self-selection
bias.
\paragraph{Known-index case.}
In the known-index case, we require $\mathrm{poly}(1/\varepsilon, k, d)$ sample and time
complexity to estimate all $k$ model parameters to accuracy $\varepsilon$ in $d$
dimensions, and can accommodate quite general selection criteria. To prove this
known-index result, we construct a log-likelihood-inspired objective function that has the true set of parameters as optimum. The key difficulty associated with this formulation is that unlike ``nice'' settings (for example, the data generating model belongs to an exponential family), the log-likelihood involves an integration over possible outputs of the \emph{unobserved} models. This scenario is reminiscent of latent-variable models where strong structural properties on the objective functions are uncommon. Nevertheless, we show that this objective function is \emph{strongly convex} where we crucially rely on the variance reduction properties of log-concave densities conditioned on convex sets. Our next goal is to run projected stochastic gradient descent (PSGD) on this objective function. Unfortunately, in contrast to standard settings in stochastic optimization, we do not have simple
access to unbiased stochastic estimates of the gradient due to the integrating out of the unobserved models in the objective function. Consequently, the gradient in this case involves sampling from the conditional distribution over outputs from the unobserved models given the observed sample at the candidate parameter set currently being considered. To sample from this conditional distribution, we show that a projected version of the Langevin Monte Carlo sampling algorithm due to Bubeck, Eldan and Lehec \citep{bubecksampling} mixes fast and produces an approximate stochastic gradient. Finally, we show that this
approximate stochastic gradient suffices for the PSGD algorithm to converge.
We provide the details of this algorithm and its analysis in Section
\ref{sec:known}.
\paragraph{Identification with unknown index.}
In the more challenging unknown-index case, it is not even clear whether the
parameters $\vec{w}_j$ are even \emph{identifiable} from the sample that we have. First, we show that this is in fact the case. Our proof uses an novel identification argument which we believe can be applied to other self-selection settings beyond the Max-selection criterion considered in this work. Formally speaking, we would like to exhibit the existence of a
mapping $f$ from $\Phi$ to the set of parameters $\vec{W}$ given access to the distribution function, $\Phi$, of the pairs $(\vec{x}, y)$ generated according to the self-selection model with unknown indices. Our construction of $f$ is based on a conditional moment calculation where we analyze the moments of $y$ conditioned on $\vec{x}$ lying in various one-dimensional subspaces. The main observation is that while closed form solutions are not known for the conditional moments, the higher order moments of $y$ still determine the length of the projection of the parameter vector with the largest projection along $x$. Concretely, we show that the higher-order moments of $y$ conditioned on $x$ being parallel to a unit vector $v$ are upper and lower bounded (up to constants) by the moments of normal distribution with $0$ and variance $\max_i (v^\top w_i)^2 + 1$. While a single direction does not uniquely determine any of the underlying parameter vectors, the direction maximizing this quantity over all one dimensional subspaces corresponds to the unit vector along the longest parameter vectors allowing recovery of \emph{one} of $k$ vectors. In the next step, we show that we may effectively ``peel off'' the single identified model from the distribution function, $\Phi$, reducing the problem of recovering the remaining parameter vectors to a self-selection problem with $k - 1$ parameter vectors. A recursive application of this argument allows identification of the remaining parameter vectors one by one.
\paragraph{Estimation with unknown index.}
We then move our attention to estimation with finite time and samples. We target
the common $\max$-selection criterion and, under some separability assumption
among the $\vec{w}_j$'s, we provide an algorithm with
$\mathrm{poly}(d) \cdot \exp(\mathrm{poly}(k))$ sample and time complexity to estimate the
regression parameters up to error $1/\mathrm{poly}(k)$.
Our technique to prove these finite-time and finite-sample results is to try to
develop a finite accuracy version of our identifiability argument. Unfortunately,
this requires an exponential (in the ambient dimension $d$) sized grid search over the unit sphere. The traditional approach to address such difficulties is to restrict our search to a
suitably chosen $k$ dimensional subspace, $U$, identifiable from the data and
crucially \emph{contains} the parameter vectors, $\vec{w}_i$. We identify this
subspace through the spectrum of a suitably chosen matrix but we encounter an additional
\emph{key} difficulty, compared to the previous applications of this method. While
our choice of matrix, $M \triangleq \mathbb{E} [y^2 \vec{x} \vec{x}^\top]$, is natural, $M$ does not decompose into a independently weighted sum of matrices each corresponding to a single parameter vector, in stark contrast to the scenario encountered in simpler problems such as mixtures of linear regressions. Hence, showing that the top singular singular subspace of $M$
contains the $\vec{w}_i$ is significantly more involved and requires novel
approximation ideas. Despite these difficulties, we derive a closed form expression for $M$ as a (positively) weighted sum of the identity matrix, $I$ and outer products $\vec{w}_i (\vec{w}_i)^\top$ where the coefficients for a single parameter vector \emph{depend} on the other vectors. To obtain this closed form characterization, we replace the max function in the self-selection criterion with a \textit{smooth maximization} function resulting
in a matrix $M'$ approximating $M$ and analyze its spectrum through several careful applications of Stein's Lemma facilitated by the differentiability of the smooth maximization function. Applying a limiting argument to $M'$, we obtain the closed form characterization of $M$ and consequently, show that the span of the parameter vectors $\vec{w}_i$ is contained in the top-$k$ singular subspace of $M$ and the singular values associated with these directions are bounded away from those for the orthogonal complement establishing a strict spectral gap. Having identified the
low-dimensional subspace containing the $w_i$, we may now restrict our search to
this subspace. We conclude our estimation argument with a careful finite-sample
adaptation of our identifiability argument highlighted above.
\paragraph{Efficient estimation for $k=2$.} For the specific yet well-studied case of $k = 2$, e.g. \cite{fair1972methods}, we develop a estimation algorithm based on the method of moments that achieves estimation error $\varepsilon$ with $\mathrm{poly}(d, 1/\varepsilon)$ time and sample complexity. It is an interesting open problem whether a similar procedure may be derived for the general case.
\medskip
\subsection{Related Work} \label{sec:intro:related}
\input{related}
\subsection{Objective Function for Linear Regression with Self-Selection Bias}
\label{sec:log_lik}
The objective function that we use is inspired by the log-likelihood function.
We show that our objective function is convex (even though linear regression with
self-selection bias does not belong to any exponential family). Suppose we have a
given parameter estimate for the $[\vec{w}_j^*]_{j=1}^k$ given by
$\bm{W} = [\vec{w}_j]_{j=1}^k$ then we define its objective value
$\overline{\ell}(\vec{W})$ as follows.
\begin{align}
\overline{\ell}(\vec{W})
&\triangleq
\frac{1}{n} \sum_{i = 1}^n
\mathbb{E}_{(y, j_*) \sim \mathcal{D}(\vec{x}^{(i)}; \vec{W}^*)}
\left[ \ell(\vec{W}; \vec{x}^{(i)}, y, j_*) \right] \notag \\
&\triangleq \frac{1}{n} \sum_{i = 1}^n \mathbb{E}_{(y, j_*)} \left[
\log\lr{f_{\sigma}(y - \bm{w}_{j_*}^\top \vec{x}^{(i)})} +
\log\lr{\int_{C_{j_*}(y)} \prod_{j \neq j_*}
f_{\sigma}(\bm{z}_j - \vec{w}_j^\top \vec{x}^{(i)})\, d\bm{z}_{-j_*} }
\right] \label{eq:objectiveDefinition},
\end{align}
where we recall that $f_\sigma$ is the density function of the standard normal
distribution.
The above expression is based on the population
likelihood under the current estimate $\vec{W}$ of the pair $(y, j_*)$
conditioned on the value of $\vec{x}$: see Appendix
\ref{app:computationsLogLikelihood} for the exact derivation.
The gradient of $\overline{\ell}$ can then be expressed in the following form:
\begin{align}
\label{eq:ll_grad_i:main}
\nabla_{\vec{w}_j} \overline{\ell}(\vec{W}) &=
\frac{1}{n\sigma^2} \sum_{i = 1}^n \Exp_{(y, j_*)} \left[
\bm{1}_{j = j_*}
\cdot y +
\bm{1}_{j \neq j_*} \cdot
\mathbb{E}_{\vec{z}_{-j_*} \sim \mathcal{N}((\vec{W}^\top x^{(i)})_{-j_*}, \sigma^2 \bm{I}_{k-1})}\left[z_j|\bm{z}_{-j_*} \in C_{j_*}(y)\right]
- \vec{w}_{j}^\top \vec{x}^{(i)}
\right] \vec{x}^{(i)}.
\end{align}
The first thing to verify is that the set of true parameters $\vec{W}^*$ are
a stationary point of the objective function that we proposed above. The
proof of the following lemma can be found in Appendix \ref{app:known}.
\begin{lemma} \label{lem:known:stationary}
It holds that $\nabla \overline{\ell}(\vec{W}^*) = 0$, where
$\bm{W}^* = [\bm{w}_j^*]_{j=1}^k$ is the set of true parameters
of the known-index self-selection model described in Definition
\ref{defn:index_observed}.
\end{lemma}
\begin{proof}
See Appendix \ref{app:known:stationary}.
\end{proof}
Our goal is to apply projected stochastic gradient descent (PSGD) on
$\overline{\ell}$. To this end, we need to prove that our objective function is
actually strongly concave and hence the optimum of $\overline{\ell}$ is unique and
equal to $\vec{W}^*$. We show this strong convexity in Section
\ref{sec:convexity}. Next, we need to show that we actually apply PSGD and hence
need to find a procedure to sample unbiased estimates of the gradient of
$\bar{\ell}$. Unfortunately the form of the objective function does not allow us
to find such an efficient procedure. For this reason we relax our requirement to
finding approximately unbiased estimates of $\nabla \bar{\ell}$. To achieve this
we use a projected version of Langevin dynamics as we show in Section
\ref{sec:Langevin}. Additionally, we need to show that the second moment of our
gradient estimates cannot be very large which we also show in Section
\ref{sec:Langevin}. Finally, we need to adapt the proof of convergence of PSGD
to show that the small bias that Langevin dynamics introduces can be controlled
in a way that does not severely affect the quality of the output estimation
which we show in Section \ref{sec:biased_sgd}. In Section
\ref{sec:known:main:proof} we combine everything together to prove our
estimation result.
\subsection{Strong Concavity} \label{sec:convexity}
The Hessian of $\overline{\ell}$ is difficult to analyze directly. We thus start
with the Hessian of the log-likelihood for a single sample $(\vec{x}^{(i)},
y^{(i)}, j_*^{(i)})$. In particular, in Appendix
\ref{app:computationsLogLikelihood} we derive the Hessian of this function
$\ell(\vec{W}; \vec{x}^{(i)}, y^{(i)}, j_*^{(i)})$,
which comprises blocks $\bm{H}_{jl}$ such that
\[ (\bm{H}_{j,l})_{ab} = \ddt{(\bm{w}_l)_a}{(\bm{w}_j)_b} \ell(\bm{W};
\vec{x}, y, j_*). \]
Following the computation in Appendix \ref{app:computationsLogLikelihood}, it
follows that for a single sample $\vec{x}, y, j_*$, the matrix block
$\bm{H}_{j,j_*} = 0$ for all $j \neq j_*$. Thus, it remains to consider only the
blocks $\bm{H}_{j_*,j_*}$ and $\bm{H}_{j,l}$ for $j,l \neq j_*$. In Appendix
\ref{app:computationsLogLikelihood} we show that
\begin{align*}
\bm{H}_{j_*,j_*} = -\frac{1}{\sigma^2}\vec{x}\vec{x}^\top
\end{align*}
and that
\begin{equation*}
\bm{H}_{\vec{W}_{-j_*}} = \frac{1}{\sigma^4} \lr{
\text{Cov}_{\bm{z}_{-j_*} \sim \mathcal{N}((\bm{W}^\top \vec{x})_{-j_*},\, \sigma^2 \bm{I}_{k-1})}
\left[z_j, z_l\,|\, \bm{z}_{-j_*} \in C_{j_*}(y)\right]
- \sigma^2 \bm{I}
} \otimes \vec{x}\vec{x}^\top,
\end{equation*}
where $\otimes$ represents the Kronecker product. Now, the key property of
convex-inducing selection functions in our proof is that, for Gaussian random
variables over $\mathbb{R}^{k-1}$, the variance is non-increasing when the
variable is restricted to a convex set.
\begin{lemma}[Corollary 2.1 of \citep{kanter1977reduction}]
Let $\vec{X} \in \mathbb{R}^n$ be a random vector with Gaussian density
$f_{\vec{X}}$. For a convex set $A \subseteq \mathbb{R}^n$ with positive mass
under the distribution of $\vec{X}$, define $\vec{X}_A$ to be $X$ restricted to
$A$, i.e., a random variable with density
$f_{\vec{X}_A}(\vec{x}) = f_{\vec{X}}(\vec{x}) \cdot (\int_{A} f_X(\vec{z})\, d\vec{z})^{-1}$.
Then, for all $\vec{v} \in \mathbb{R}^n$,
\[ \text{Var}[\vec{v}^\top \vec{X}_A] \leq \text{Var}[\vec{v}^\top \vec{X}]. \]
\end{lemma}
In particular, together with our thickness assumption and properties of the
Kronecker product, this implies that
$\bm{H}_{\vec{W}_{-j_*}} \preceq 0$. Thus, the complete Hessian of the
function $\ell$ can be expressed as a block matrix of the form:
\[ \bm{H}
= \left[\begin{matrix}
-\frac{1}{\sigma^2} \vec{x}\vec{x}^\top & \bm{0} \in \mathbb{R}^{d \times d(k-1)} \\
\bm{0} \in \mathbb{R}^{d(k-1) \times d} & \bm{H}_{\vec{W}_{-j_*}}
\end{matrix}\right]
\preceq \left[\begin{matrix}
-\frac{1}{\sigma^2} \vec{x}\vec{x}^\top & \bm{0} \\
\bm{0} & \bm{0}
\end{matrix}\right].
\]
We are now ready to upper bound the Hessian $\bm{H}_{pop}$ of our objective
function $\bar{\ell}$.
In particular,
at this point we can use our minimum-probability assumption
(Assumption \ref{asp:known:1}) and our thickness of covariates assumption
(Assumption \ref{asp:known:2}) from which we get that for the Hessian
$\bm{H}_{pop}$ it holds that
\[ \bm{H}_{pop} \preceq -\frac{\alpha}{\sigma^2 \cdot k} \bm{I}. \]
From the above we conclude that the following lemma
\begin{lemma} \label{lem:strongConcavity}
The objective function $\bar{\ell}$ is
$\left(\frac{\alpha}{\sigma^2 \cdot k}\right)$-strongly-concave.
\end{lemma}
\subsection{Approximate Stochastic Gradient Estimation} \label{sec:Langevin}
In this section we describe an algorithm for sampling approximate stochastic
estimates of our objective function $\bar{\ell}$. Our algorithm is based on
projected Langevin dynamics. We start with the expression of the gradient of
$\bar{\ell}$ based on \eqref{eq:ll_grad_i:main}.
\begin{align}
\label{eq:ll_grad_i:main:2}
\nabla_{\vec{w}_j} \overline{\ell}(\vec{W}) &=
\frac{1}{n\sigma^2} \sum_{i = 1}^n \Exp_{(y, j_*)} \left[
\bm{1}_{j = j_*}
\cdot y +
\bm{1}_{j \neq j_*} \cdot
\mathbb{E}_{\vec{z}_{-j_*} \sim \mathcal{N}((\vec{W}^\top x^{(i)})_{-j_*}, \sigma^2 \bm{I}_{k-1})}\left[z_j|\bm{z}_{-j_*} \in C_{j_*}(y)\right]
- \vec{w}_{j}^\top \vec{x}^{(i)}
\right] \vec{x}^{(i)}.
\end{align}
where we remind that $\mathbb{E}_{(y,j^*)}$ denotes the expectation of the pair
$(y^{(i)}, j_*^{(i)})$ conditioned on $\vec{x}^{(i)}$ and $\vec{W}^*$.
To obtain stochastic gradient estimates, we will replace these expectations
with their
corresponding observed values as we will see below. The more difficult step of
the gradient estimation process is
sampling the last term of \eqref{eq:ll_grad_i:main:2}, for which
it suffices to be able to sample the truncated normal distribution
\begin{align}
\mathcal{N}((\vec{W}^\top x^{(i)})_{-j_*}, \sigma^2 \bm{I}_{k-1};\ C_{j_*}(y))
\label{eq:difficultTermToSample}
\end{align}
given some set of parameter $\vec{W}$, a vector of covariates
$\vec{x}^{(i)}$ and a pair $(y, j_*)$ drawn from
$\mathcal{D}(\vec{x}^{(i)}; \vec{W}^*)$. The simplest way to get a sample from
\eqref{eq:difficultTermToSample} is to first sample from
$\mathcal{N}((\vec{W}^\top x^{(i)})_{-j_*}, \sigma^2 \bm{I}_{k-1}),$
and then
apply rejection sampling until we get a sample inside $C_{j_*}(y)$.
This is feasible
information-theoretically but it might require a lot of computational steps if
the survival probability of $C_{j_*}(y)$ is small. In particular, the rejection
sampling might require time that is exponential in the norm of the $\vec{W}_j$'s.
For this reason, if we require statistical efficiency we need to apply a more
elaborate technique. In particular, we use projected Langevin
dynamics. Let $K = S_{j_*}(y) \cap \mathcal{B}(R)$ for some sufficiently large
constant $R$ and let $\vec{\mu}_{-j_*} = (\vec{W}^\top \vec{x}^{(i)})_{-j_*}$
for the rest of this section. The iteration of projected Langevin algorithm for
sampling
is the following \cite{bubecksampling}:
\begin{align}
\vec{z}^{(t + 1)} = \Pi_K\left(\vec{z}^{(t)} - \frac{\gamma}{2 \cdot \sigma^2}
(\vec{z}^{(t)} - \vec{\mu}_{-j_*}) + \sqrt{\gamma} \cdot \vec{\xi}^{(t)} \right) \label{eq:projectedLangevin}
\end{align}
where $\vec{\xi}^{(1)}, \vec{\xi}^{(2)}, \dots$ are i.i.d. samples from the
standard normal distribution in $(k - 1)$-dimensions. The next lemma describes
the sampling guarantees of the Langevin iteration \eqref{eq:projectedLangevin}.
\begin{lemma} \label{lem:samp_trunc}
Let $L \subset \mathbb{R}^{(k - 1)}$ be a convex set with
$\mathbb{P}_{\mathcal{N}(\vec{0}, \sigma^2 \vec{I})}(L) \geq \alpha$ for $\alpha > 0$.
Then, for any $\vec{\mu}_{-j_*} \in \mathbb{R}^{k - 1}$ and $\epsilon \in (0, 1/2]$, the
projected Langevin sampling algorithm \eqref{eq:projectedLangevin} with
$K = L \cap \mathcal{B}(R)$ for some appropriate value $R$, and initialized
with $\bm{z}^{(0)} = \Pi_K(\bm{0}_{k-1})$, generates a random
variable $\hat{\vec{X}} = \vec{z}^{(m)}$ satisfying
\begin{equation*}
\mathrm{TV} \left(\widehat{\vec{X}}, \mathcal{N}(\vec{\mu}_{-j_*}, \sigma^2 \cdot \vec{I}_{k-1}, K)\right) \leq
\epsilon
\end{equation*}
assuming that the number of steps $m$ is larger that
$\mathrm{poly} (k, \norm{\vec{w}}, 1 / \epsilon, 1 / \alpha, \sigma^2, 1/\sigma^2)$.
\end{lemma}
\begin{proof}
The proof of this lemma can be found in Appendix \ref{app:lem:samp_trunc}.
\end{proof}
\smallskip
Now that we can sample from the distribution
\eqref{eq:difficultTermToSample}, we can move to approximately estimating a
stochastic gradient of $\overline{\ell}$. First, we sample uniformly $i \in [n]$a
uniformly at random, and we fix the corresponding $\vec{x}^{(i)}$. Then, we use
the $i$-th sample from the true model to substitute in the pair $(y, j_*)$.
Finally, we use the Langevin algorithm that we described above to sample
\eqref{eq:difficultTermToSample}. Before moving to bounding the bias of our
estimator there is one more thing that we need to take care of, and this is that
for every $\vec{x}^{(i)}$ we only have one sample of the pair $(y, j_*)$. Hence,
we need to make sure that during the execution of the algorithm while we pick
the indices $i \in [n]$ uniformly at random we will never pick the same
index $i$ twice. To ensure that we are going to require more samples than the
ones we need.
Let $n$ be the total number of samples that we have and $T$ be total number of
samples that we need for our PSGD algorithm. A straightforward birthday paradox
calculation
yields that the probability
of sampling the same $i$ twice is at most $2 T^2/n$. Thus,
if we pick $n \ge 2 T^2/\zeta$, then the collision probability
during the execution of the PSGD algorithm is at most $\zeta$.
\smallskip
We are now ready to put everything together in algorithm that
describes our combined estimation procedure. The following lemma whose proof be
found in Appendix \ref{app:lem:samplingGradient} describes the performance
guarantees of the above estimation algorithm.
\begin{algorithm}[t]
\caption{Approximate Stochastic Gradient Estimation Algorithm}\label{alg:langevin}
\begin{algorithmic}[1]
\Procedure{EstimateGradient}{$\vec{W}$}
\State sample $i$ uniformly from $[n]$
\State $K \gets C_{j_*^{(i)}}(y^{(i)}) \cap \mathcal{B}(R)$
\State $\vec{\mu} \gets (\vec{W}^\top \vec{x}^{(i)})_{-j_*^{(i)}}$
\State $\vec{z}^{(0)} \gets \Pi_K(\vec{0})$
\For{$t = 1,\ldots, m$}
\State sample $\vec{\xi}^{(t-1)}$ from $\mathcal{N}(\vec{0}, \vec{I})$
\State $\vec{z}^{(t)} \gets \Pi_{K}\left(\vec{z}^{(t-1)} -
\frac{\gamma}{2 \cdot \sigma^2} (\vec{z}^{(t-1)} - \vec{\mu}) +
\sqrt{\gamma} \cdot \vec{\xi}^{(t-1)} \right)$
\EndFor
\For{$j = 1,\ldots, k$}
\State $\vec{g}_j \gets \frac{1}{\sigma^2} \left(
\bm{1}_{j_*^{(i)} = j} \vec{y}^{(i)} + \bm{1}_{j_*^{(i)} \neq j} \cdot
z^{(m)}_j
- \vec{w}_j^\top \vec{x}^{(i)}
\right) \cdot \vec{x}^{(i)}$
\EndFor
\State \Return $\vec{g} = (\vec{g}_1, \dots, \vec{g}_k)$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{Projected Stochastic Gradient Descent}\label{alg:psgd}
\begin{algorithmic}[1]
\Procedure{PSGD}{}
\State $\vec{W}^{(0)} \gets 0$
\For{$t = 1, \ldots, T$}
\State $\eta_t \gets 1/\lambda \cdot t$
\State $\vec{g} \gets \textsc{EstimateGradient}(\vec{W}^{(t)})$
\State $\vec{w}^{(t)}_j \gets \Pi_{\mathcal{K}}\lr{\vec{w}^{(t)}_j
- \eta_t \cdot \vec{g}^{(t)}_j}$ for all $j \in [k]$
\EndFor
\State \Return $\bar{\vec{W}} \triangleq \frac{1}{T} \sum_{t = 1}^{T} \left(\vec{w}^{(t)}_j\right)_{j = 1}^k$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{lemma} \label{lem:samplingGradient}
Let $\vec{g}^{(1)}, \dots, \vec{g}^{(T)}$ be a sequence of outputs of
Algorithm \ref{alg:langevin} when used with input
$\vec{W}^{(1)}, \dots, \vec{W}^{(T)}$, where
$\norm{\vec{W}^{(p)}}_2 \le k \cdot B$ and $\vec{W}^{(p)}$ can
depend on $\vec{W}^{(p - 1)}, \vec{g}^{(p - 1)}$. If $n \ge 2 T^2/\zeta$,
for the hyperparameters $\eta$, $R$, it holds that
$R, 1/\eta \le \mathrm{poly} (k, B, 1 / \beta, 1 / \alpha, \sigma^2, 1/\sigma^2)$, and $m \ge \mathrm{poly} (k, B, 1 / \beta, 1 / \alpha, \sigma^2, 1/\sigma^2)$
then with probability at least $1 - \zeta$ it holds that for every $p \in [T]$
\begin{align}
\norm{\Exp \left[\vec{g}^{p} \mid \vec{W}^{(p - 1)}, \vec{g}^{(p - 1)} \right] - \nabla \bar{\ell}(\vec{W}^{(p)})}_2 \le \beta. \label{eq:boundOnBias}
\end{align}
\end{lemma}
\subsection{Stochastic Gradient Descent with Biased Gradients}
\label{sec:biased_sgd}
In the previous section we showed that we can compute approximate stochastic
gradients of the strongly-concave function $\bar{\ell}$. In this section we show
that this is enough to approximately optimize $\bar{\ell}$ using projected
gradient descent. We start with a description of the PSGD algorithm.
\begin{lemma}%
\label{lemma:sgd}
Let $f: \mathbb{R}^k \to \mathbb{R}$ be a convex function, $\mathcal{K} \subset
\mathbb{R}^k$ a convex set, and fix an initial estimate $w^{(0)} \in \mathcal{K}$.
Now, let $\vec{x}^{(1)}, \ldots, \vec{x}^{(T)}$ be the iterates generated by
running $T$
steps of projected SGD using gradient estimates
$\vec{g}^{(1)}, \ldots, \vec{g}^{(M)}$
satisfying $\mathbb{E}[\vec{g}^{(i)} | \vec{x}^{(i-1)}] =
\nabla f(\vec{x}^{(i-1)}) + \vec{b}^{(i)}
$. Let $\vec{x}_* = \arg\min_{\vec{x} \in
\mathcal{K}} f(w)$ be a minimizer of $f$. Then, if we assume:
\begin{enumerate}
\item[(i)] \textbf{Bounded step variance:}
$\mathbb{E}\left[\|\vec{g}^{(i)}\|_2^2 \right] \leq \rho^2$,
\item[(ii)] \textbf{Strong convexity:}
$f$ is $\lambda$-strongly convex, and
\item[(iii)] \textbf{Bounded gradient bias:}
$\|\vec{b}^{(i)}\| \leq
\frac{\rho^2}{2\cdot \lambda\cdot \mathrm{diam}(\mathcal{K}) \cdot i},$
\end{enumerate}
then the average iterate $\hat{\vec{x}} = \frac{1}{T} \sum_{t=1}^T \vec{x}^{(t)}$
satisfies
$\mathbb{E}[f(\hat{\vec{x}}) - f(\vec{x}_*)]
\leq \frac{\rho^2}{\lambda T}(1
+ \log(T))$.
\end{lemma}
\begin{proof}
See Appendix \ref{app:bias_sgd_proof}.
\end{proof}
\subsection{Proof of Theorem \ref{thm:knownIndex}}
\label{sec:known:main:proof}
We are now ready to combine the results of the previous sections into a recovery
guarantee for $\bm{W}^* = \{\vec{w}_i\}$. In particular, we will apply Lemma \ref{lemma:sgd}
to show that Algorithm \ref{alg:psgd} converges to an average iterate $\hat{\vec{W}}$
that is close to $\bm{W}^*$.
First, observe that the norm of the gradient estimates outputted by Lemma \ref{app:lem:samp_trunc}
are bounded in norm by
\begin{align*}
\mathbb{E}\left[\|v\|_2^2\right] \leq \mathbb{E}\left[
\sum_{j=1}^k \left\| \nabla_{\vec{w}_j} \ell(\bm{W}; (\vec{x}, i, y)) \right\|^2
\right] + \beta,
\end{align*}
where $\beta$ is as in Lemma \ref{lem:samp_trunc}.
Our bounds on the norm of the weights and covariates directly implies
\[
\mathbb{E}\left[\|v\|_2^2\right] \in O\lr{k\cdot \text{poly}(B, C)} + \beta.
\]
Next, Lemma \ref{lem:strongConcavity} guarantees that $f$ in Lemma \ref{lemma:sgd} is
strongly convex with $\lambda = \alpha/(\sigma k)$.
Finally, Lemma \ref{lem:samp_trunc} ensures access to gradients with appropriately
bounded bias (i.e., satisfying assumptions (i) and (iii) in Lemma \ref{lemma:sgd})
in $\text{poly}(k, B, C, T, 1/\alpha, \sigma^2, 1/\sigma^2)$-time.
We are thus free to apply Lemma \ref{lemma:sgd} to our
problem---after averaging $T$ steps of projected stochastic gradient
descent, we will find $\bm{W} = \{\vec{w}_j\}_{j=1}^k$ such that
\begin{align}
\label{eq:sgd_bound}
\mathbb{E}[\overline{\ell}(\bm{W})] - \overline{\ell}(\bm{W}^*)
\leq \frac{\sigma^2 \cdot k^2\cdot \mathrm{poly}(B, C)}{2\cdot \alpha\cdot T} (1 + \log(T)).
\end{align}
An application of Markov's inequality shows that, with probability at least $1 -
\delta$,
\[
\overline{\ell}(\bm{W}) - \overline{\ell}(\bm{W}^*)
\leq \frac{\sigma^2 \cdot k^2\cdot \mathrm{poly}(B, C, 1/\delta)}{2\cdot \alpha\cdot T} (1 + \log(T)).
\]
Thus, we can condition on the event in \eqref{eq:sgd_bound} while only losing a
factor of $1 - \delta$ in success probability. Finally, a parameter-space
recovery bound follows from another application of convexity:
\[
\left\|\bm{W} - \bm{W}^* \right\|_F
\leq \frac{\sigma^4 \cdot k^3\cdot \mathrm{poly}(B, C, 1/\delta)}{\alpha^2\cdot T} (1 + \log(T)).
\]
\section{Introduction}
\label{sec:intro}
\input{intro}
\section{Model and Main Results} \label{sec:model}
\input{setup}
\section{Parameter Estimation for the Known-Index Setting} \label{sec:known}
\input{known_index}
\section{Parameter Estimation for the Unknown-index Setting} \label{sec:unknown}
\input{unknown_index}
\section{Acknowledgements}
This work is supported by NSF Awards CCF-1901292, DMS-2022448 and
DMS2134108, a Simons Investigator Award, the Simons Collaboration on the Theory
of Algorithmic Fairness, a DSTA grant, the DOE PhILMs project
(DE-AC05-76RL01830), an Open Philanthropy AI Fellowship and a Microsoft
Research-BAIR Open Research Commons grant.
\printbibliography
\clearpage
\subsection{Known-Index Setting} \label{sec:model:knownIndex}
We start with the definition of a self-selection rule that is fundamental in
the modeling of the linear regression problem with self-selection bias in the
known-index setting.
\begin{definition}[Self-Selection Rule] \label{def:selfSelectionRule}
A \textit{self-selection rule} is a function
$S: \mathbb{R}^k \to [k]$. We assume throughout this work that we have query
access to $S$, i.e. for every $\bm{y} \in \mathbb{R}^k$ there is an oracle
that outputs $S(\bm{y})$.
We also define a \underline{slice} $S_j(a)$ of $S$ as follows:
\[
S_j: \mathbb{R} \to \mathbb{R}^{k} \qquad
S_j(a) \triangleq \{\bm{y} \in \mathbb{R}^k: S(\bm{y}) = j, \bm{y}_j = a\}.
\]
\end{definition}
The first setting that we consider is the \textit{known-index setting}, where
the observed data for each covariate include the response variable, as well as
the index of the corresponding regressor.
\begin{definition}[Self-Selection with Observed Index]
\label{defn:index_observed}
Self-selection with observed index is parameterized by an unknown set of
weight vectors $\vec{w}^*_1,\ldots \vec{w}_k^* \in \mathbb{R}^d$, a known
variance $\sigma > 0$,
and a self-selection rule
$S: \mathbb{R}^k \rightarrow [k]$. For $i \in [n]$, an observation
$(\vec{x}^{(i)}, y^{(i)}, j_*^{(i)})$ in this model is a triplet, comprising
a feature vector $\vec{x}^{(i)}$, and a pair $(y^{(i)},j_*^{(i)})$ sampled as
follows conditioning on $\vec{x}^{(i)}$:
\begin{enumerate}
\item[(1)] Sample the latent variables $y_{j}^{(i)} \sim \mathcal{N}(\vec{w}_j^{* \top} \vec{x}^{(i)},
\sigma^2)$ for each $j \in [k]$.
\item[(2)] Reveal the observation index $j_*^{(i)} =
S(y_{1}^{(i)},\ldots,y_{k}^{(i)})$
and the response variable $y^{(i)} = y_{j_*^{(i)}}^{(i)}$
\end{enumerate}
For a fixed $\vec{x}$ and $\vec{W}^*=(\vec{w}^*_1,\ldots \vec{w}_k^*)$, we use $\mathcal{D}(\vec{x}; \vec{W}^*)$ to denote the
probability distribution of the pair $(i, y)$ sampled according to Steps (1) and (2) above. For example, if
$S(y_{1}^{(i)},\ldots,y_{k}^{(i)}) = \arg\max_{j \in [k]} y_{j}^{(i)}$, then under
this model we observe only the largest $y_{j}^{(i)}$ and its index. Our goal is
to obtain an accurate estimate $\hat{\vec{w}}_j$ for each $\vec{w}^*_j$ given
only samples from the above model.
\end{definition}
\noindent In most of this paper we are mainly concerned with estimating
the weights $\vec{w_1^*}$, $\ldots$, $\vec{w_k^*}$ from observation of the
covariates and the {\em maximal} response variable
\[y = \max_{j \in [k]}\ \{\vec{w_j}^\top \vec{x} + \bm{\varepsilon}_j\}, \qquad \text{
where } \bm{\varepsilon} \sim \mathcal{N}(0, \sigma^2 \cdot \bm{I}_{k}). \]
In the observed-index setting, however, it turns out that our efficient estimation
can be applied to a much larger set of selection functions $S(\{y_j\})$, which we
call {\em convex-inducing}.
\begin{definition}[Convex-inducing Self-Selection Function]
\label{defn:convex-inducing}
We call a self-selection function $S: \mathbb{R}^k \to [k]$ {\em
convex-inducing} if, for each $j \in [k]$ and $a \in \mathbb{R}$, there
exists a convex set $C_j(a)$ such that
\[
S_{j}(a) = \bm{1}\{\bm{y}_{-j} \in C_j(a)\},
\]
where $S_j(a)$
are the slices defined in Definition \ref{def:selfSelectionRule}.
\end{definition}
Notably, setting $S(\bm{y}) = \arg\max_{j \in [k]} \bm{y}_j$ recovers the maximum-response
observation model, and this choice of $S(\cdot)$ also satisfies Definition
\ref{defn:convex-inducing} with $C_j(a) = (-\infty, a]^{k-1}$. The definition
also allows us to capture cases beyond the maximum; for example, convex-inducing
functions also include $S(\bm{y}) = \arg\max_{j \in [k]} f(\bm{y}_j)$ for any
monotonic function $f$.
In order for our estimation procedure to succeed we use the following assumptions
on the feature and parameter vectors which are classical even in many standard
linear regression instances.
\begin{assumption}[Feature and Parameter Vectors] \label{asp:known:1}
For the set of feature vectors $\{\vec{x}^{(i)}\}_{i = 1}^n$ that we have
observed we assume that:
\begin{align} \label{eq:asp:known}
\norm{\vec{x}^{(i)}}_2 \le C \quad \text{for all $i \in [n] \quad$ and } \quad \quad
\frac{1}{n} \sum_{i = 1}^n \vec{x}^{(i)} \vec{x}^{(i) \top} \succeq \mathbf{I}.
\end{align}
For the true parameter vectors $\vec{w}_1^*$, $\ldots$, $\vec{w}_k^*$ we
assume that $\norm{\vec{w}_j^*}_2 \le B$.
\end{assumption}
Although our theorem allows for a wide range of self-selection rules, we need some
additional assumptions that allow the recovery of every parameter vector
$\vec{w}_j$. To see why this is needed imagine the setting where the
self-selection rule $S$ is the $\argmax$ and for some $j$ the coordinates of
$\vec{w}_j$ are extremely small. In this case it is impossible to hope to estimate
$\vec{w}_j$ since there is a huge probability that we do not even observe one
sample of the form $y = \vec{x}^T \vec{w}_j + \varepsilon_j$. For this reason we need the
following assumption.
\begin{assumption}[Survival Probability -- Self-Selection Rule] \label{asp:known:2}
For every sample $(\vec{x}, y, j_*)$ that we have observed,
the following properties hold:
\begin{itemize}
\item[(i)] for every sample $\vec{x}^{(i)}$ and every $j \in [k]$, the
probability that $j_* = j$ is at least $\alpha / k$ for some constant $\alpha$;
\item[(ii)] the mass of the set $C_{j_*}(y)$ with respect to
$\mathcal{N}((\vec{W}^{* \top} \vec{x})_{-j_*}, \sigma^2 \cdot \vec{I}_{k-1})$
is at least $\alpha$.
\end{itemize}
\end{assumption}
Finally, we need to assume oracle access to a self-selection rule $S(\cdot)$
that is convex-inducing, in the same sense introduced in Definition
\ref{defn:convex-inducing}:
\begin{assumption}[Self-Selection Rule] \label{asp:known:3}
We assume that the self-selection rule $S: \mathbb{R}^k \to [k]$ is
convex-inducing, and that for every $j \in [k]$ and $y \in \mathbb{R}$ we have access
to the following:
\begin{itemize}
\item[(i)] a membership oracle for $S(y)$, i.e., for every
$\vec{y} \in \mathbb{R}^{k}$ and $j \in [k]$, we can answer whether $S(\vec{y}) = j$
or not, and
\item[(ii)] a projection oracle to the convex sets $C_j(y)$ that correspond
to slices $S_j(y)$ of the self-selection rule, i.e., for every
$\vec{y} \in \mathbb{R}^{k}$ such that $\vec{y} \not\in S_j(a)$, we can efficiently
compute the nearest vector $\vec{y}'$ such that $\vec{y}'_j$
$\argmin_{\vec{s} \in S_i(y)} \norm{\vec{z} - \vec{s}}$.
\end{itemize}
\end{assumption}
Using the above definitions and assumptions we are now ready to state our main
theorem for the known-index setting.
\begin{theorem}[Known-Index Estimation] \label{thm:knownIndex}
Let $(\vec{x}^{(i)}, y^{(i)}, i^{(i)})_{i = 1}^n$ be $n$ observations from
a self-selection setting with $k$ linear models $\vec{w}^*_1$, $\dots$,
$\vec{w}^*_k$ as described in Definition \ref{defn:index_observed}. Under Assumptions \ref{asp:known:1}, \ref{asp:known:2}(i), and \ref{asp:known:3}(i),
there exists an estimation algorithm that outputs $\hat{\vec{w}}_1$, $\dots$,
$\hat{\vec{w}}_k$ such that with probability at least $0.99$, for every
$j \in [k]$
\[ \norm{\hat{\vec{w}}_i - \vec{w}_i^*}_2^2 \le \mathrm{poly}(\sigma, k, 1/\alpha, B, C) \cdot \frac{\log(n)}{n}.\]
If we additionally assume the Assumptions \ref{asp:known:2}(ii), and
\ref{asp:known:3}(ii), then the running time of the algorithm is
$\mathrm{poly}(n, d, k, 1/\alpha, B, C, \sigma, 1/\sigma)$.
\end{theorem}
In Section \ref{sec:known}, we explain our algorithm for proving Theorem
\ref{thm:knownIndex} and we describe the main ideas and techniques for the proof.
To construct an efficient algorithm we use an interesting combination of
projected gradient descent and the Langevin algorithm, after we develop an
appropriate objective function.
\subsection{Unknown-Index Setting} \label{sec:model:unknownIndex}
We next consider the more challenging {\em unknown-index} setting, where we have
sample access to the response variable, but do not observe the index of the
corresponding weight vector. In this more challenging setting, we make a few
additional assumptions on the structure of the problem, namely that the
covariates $\vec{x}^{(i)}$ are drawn from an mean-zero identity-covariance
Gaussian distribution (rather than being arbitrary); that the noise terms
$\vec{\varepsilon}^{(i)}$ are also identity-covariance (rather than $\sigma\cdot
\bm{I}_k$ for $\sigma > 0$); and that the self-selection rule is the maximum
response rule $S(\bm{y}) = \arg\max_{j \in [k]} \bm{y}_j$ (rather than an arbitrary
convex-inducing rule).
\begin{definition}[Self-Selection with Unknown Index]
\label{defn:index_unobserved}
Just as for Definition \ref{defn:index_observed}, we have a set of weight
vectors $\vec{w}^*_1,\ldots \vec{w}^*_k \in \mathbb{R}^d$.
We assume that vectors $\vec{x}^{(i)}$ are drawn i.i.d. from the standard
multivariate normal distribution $\mathcal{N}(\vec{0}, \bm{I}_d)$. For
each sampled covariate $\vec{x}^{(i)}$, we:
\begin{enumerate}
\item[(1)] Sample $y_{j}^{(i)} \sim \mathcal{N}(\vec{w_j}^{* \top}
\vec{x}^{(i)}, 1)$ for each $j \in [k]$.
\item[(2)] Compute the maximum response $y^{(i)} = \max_{j \in [k]} y_{j}^{(i)}$.
\item[(3)] Observe only the covariate-response pair $(\vec{x}^{(i)},
y^{(i)})$: in particular, we do not observe the index of the maximum
response.
\end{enumerate}
Our goal is to obtain an accurate estimate $\hat{\vec{w}}_j$ for each
$\vec{w}^*_j$ given only samples from the above model.
\end{definition}
This model resembles mixtures of linear regressions, with the key
difference being that in the latter, the index of the weight vector used for
each covariate is sampled i.i.d. from a categorical distribution with fixed
mixture probability. In contrast, here the probability of observing a response
from a given weight vector depends on both the covariate and the sampled noise
$\eta_i$.
\smallskip
In this model it is not even clear that estimation of $\vec{w}^*_j$'s is
possible:
indeed, our first result (Theorem \ref{thm:identifiability_no_index}) is an
information-theoretic one that shows (infinite-sample) identifiability for
unknown-index model.
To show this information-theoretic result, we use a novel argument that might be
of independent interest---we believe that it can be used to show the
identifiability of other statistical problems with self-selection bias as well.
The proof of the theorem below is presented in Section
\ref{ssec:identifiability}.
\begin{restatable}{theorem}{unknownidentify} \label{thm:identifiability_no_index}
Let $\vec{W}^* = [\vec{w}^*_j]_{j = 1}^k \in \mathbb{R}^{d \times k}$ and
$\Phi_{\vec{W}^*}$ be the distribution function of the pairs
$(\vec{x}^{(i)}, y^{(i)})$ associated with the self-selection model with
unknown indices as described in Definition~\ref{defn:index_unobserved}.
Then, there exists a function $f$ satisfying:
\begin{equation*}
\forall~\vec{W}^* = [\vec{w}^*_j]_{j = 1}^k \subset \mathbb{R}^d ~ \text{it holds that} ~ f(\Phi_{\vec{W}^*}) = \vec{W}^*.
\end{equation*}
\end{restatable}
\noindent In order to transform the above information-theoretic result to a
finite-sample and finite-time bound in the unknown-index case, we need the
following separability assumption among the $\vec{w}^*_j$'s.
\begin{assumption}[Separability Assumption] %
\label{as:no_index}
For some known real values $B$, $\Delta$ it holds that:
\begin{equation*}
\forall i \neq j: \frac{\abs*{{\vec{w}_i^{* \top} \vec{w}^*_j}}}{\norm{\vec{w}^*_j}} + \Delta \leq \norm{\vec{w}^*_j} \text{ and } \max_{j \in [k]} \norm{\vec{w}^*_j} \leq B.
\tag{A} \label{eq:asump_no_index}
\end{equation*}
\end{assumption}
\begin{theorem}[Unknown-Index Estimation]
\label{thm:no_index}
Let $(\vec{x}^{(i)}, y^{(i)})_{i = 1}^n$ be $n$ observations from a
self-selection setting with $k$ linear models $\vec{w}^*_1$, $\dots$,
$\vec{w}^*_k$ as described in the unknown-index setting in Definition
\ref{defn:index_unobserved}. If we assume Assumption \ref{asp:known:2}(i) and
Assumption \ref{as:no_index}, then there exists an estimation algorithm that
outputs $\hat{\vec{w}}_1$, $\dots$, $\hat{\vec{w}}_k$ such that with
probability at least $0.99$, for every $j \in [k]$
\[ \norm{\hat{\vec{w}}_j - \vec{w}_j^*}_2 \le \varepsilon,\]
as long as
$n \geq \mathrm{poly} (d) \cdot \exp \blr{\mathrm{poly} (B / \varepsilon) \cdot \widetilde{O}(k)}$ and
$\varepsilon \le \Delta/16$.
Furthermore, the running time of the algorithm is at most
$n \cdot \exp \blr{\mathrm{poly} (B / \varepsilon) \cdot \widetilde{O} (k)}$.
\end{theorem}
Our estimation algorithm above only makes sense for $\varepsilon = 1/\mathrm{poly}(k)$ and
resembles the corresponding results of mixtures of linear regressions
\cite{li2018learning, chen2020learning}. What we are missing for this model is a
local analysis corresponding to \cite{kwon2020converges} that will enable us to
get error $\varepsilon$ with running time and number of samples that are polynomial in
$1/\varepsilon$. In the special, yet very relevant case of $k = 2$, we are able to
improve our aforementioned result and show that a moment-based algorithm
efficiently recovers $\vec{w}_1$ and $\vec{w}_2$ up to $\varepsilon$ with running time
and number of samples that are polynomial in $1/\varepsilon$, under the following
slightly different separability assumption.
\begin{assumption}[Separability Assumption -- $k = 2$] %
\label{as:no_index:2}
For some \emph{known} $B, \Delta > 0$ it holds that
\begin{equation*}
\norm{\vec{w}_1^* - \vec{w}_2^*}_2 \ge \Delta \quad \text{ and } \quad \max_{j \in [k]} \norm{\vec{w}_j^*} \leq B ~~~ \forall i \in [2]. \tag{B}
\label{eq:asump_no_index:2}
\end{equation*}
\end{assumption}
\begin{theorem}[Unknown-Index Estimation -- $k = 2$]
\label{thm:no_index:2}
Let $(\vec{x}^{(i)}, y^{(i)})_{i = 1}^n$ be $n$ observations from a
self-selection setting with $2$ linear models $\vec{w}^*_1$, $\vec{w}^*_2$ as
described in the unknown-index setting of Section \ref{sec:model}. If we assume
Assumption \ref{asp:known:2} and Assumption \ref{as:no_index:2}, then there
exists an estimation algorithm that outputs $\hat{\vec{w}}_1$,
$\hat{\vec{w}}_2$ such that with probability at least $0.99$, for every
$j \in [2]$
\[ \norm{\hat{\vec{w}}_j - \vec{w}_j^*}_2 \le \varepsilon,\]
as long as $n \geq \mathrm{poly} (d, 1/\varepsilon, \Delta, B, 1/\alpha)$ and
$\varepsilon \le \Delta/10$. Furthermore, the running time of the algorithm is at most
$n \cdot \mathrm{poly} (d, 1/\varepsilon, \Delta, B, 1/\alpha)$.
\end{theorem}
In Section \ref{sec:unknown} we describe the algorithms and proofs for Theorem
\ref{thm:no_index} and Theorem \ref{thm:no_index:2}.
\medskip
\noindent \textbf{Remark.} (High-Probability Results)
All of the above results are expressed in term of constant probability of
error. We can boost this probability to $\delta$ by paying an additional
$\log(1/\delta)$ factor in the sample and time complexities. This boosting can
be done because we are solving a parametric problem and it a folklore idea that
any probability of error less that $1/2$ can be boosted to $\delta$. Roughly
the way that this boosting works is that we run the algorithm independently
$\log(1/\delta)$ times and from the $\log(1/\delta)$ different estimates we
keep one that contains at least half of all the others within a ball of radius
$2 \varepsilon$.
\subsection{Identifiability with Unknown Indices}
\label{ssec:identifiability}
Here, we establish the information theoretic identifiability of the self-selection model with unknown indices. Recall, that we receive samples generated according to $y = \max_{j = 1}^k w_j^\top x + \eta_j$ where $x \thicksim \mathcal{N} (0, I)$ and $\eta_j \overset{iid}{\thicksim} \mathcal{N} (0, 1)$. We now establish the following theorem:
\unknownidentify*
\begin{proof}
Our proof will be based on an inductive argument on the number of components, $k$. We will use a peeling argument to reduce the parameter recovery problem with $k$ components to one with $k - 1$ components. The base case when $k = 1$, reduces to standard linear regression where, for example, $\mathbb{E} [xy] = w_1$ suffices. For the inductive argument, suppose $k > 1$ and consider the following function:
\begin{equation*}
\forall v \in \mathbb{R}^d, \norm{v} = 1: F(v) = \lim_{\text{Even } p \to \infty} \lim_{\gamma \to 0} \lr{\frac{\mathbb{E} \sqlr{y^p \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma}}{(p-1)!!}}^{1 / p}.
\end{equation*}
We will now show that the above function is well defined for all $\norm{v} = 1$. The conditional moments may be evaluated with access to the distribution function $\Phi_{\mc{W}}$. Defining $j^* = \argmax_{j \in [k]} \abs{v^\top w_j}$ and $\sigma_j = \abs{v^\top \vec{w}_{j}}$, we now lower bound the conditional moment:
\begin{align*}
\mathbb{E} [y^p \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma] &\geq \mathbb{E} [y^p \cdot \bm{1} \blr{\vec{w}_{j^*}^\top x + \eta_{j^*}\geq 0} \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma] \geq \frac{1}{2} \cdot \mathbb{E} [y_{j^*}^p\mid \norm{\mathcal{P}^\perp_v x} \leq \gamma] \\
&\geq \frac{1}{2} \cdot \mathbb{E} [(\vec{w}_{j^*}^\top \mathcal{P}_v (x) + \eta_{j^*} + \vec{w}_{j^*}^\top \mathcal{P}^\perp_v (x))^p\mid \norm{\mathcal{P}^\perp_v x} \leq \gamma] \\
&\geq \frac{1}{2} \cdot \mathbb{E} \sqlr{\sum_{l = 0}^{p / 2} \binom{p}{2l} (\vec{w}_{j^*}^\top \mathcal{P}_v (x) + \eta_{j^*})^{p - 2l} (\vec{w}_{j^*}^\top \mathcal{P}^\perp_v (x))^{2l} \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma} \\
&\geq \frac{1}{2} \cdot (p - 1)!! \cdot (\sigma_{j^*}^2 + 1)^{p / 2}.
\end{align*}
Through a similar computation, we obtain an upper bound on the conditional moment:
\begin{align*}
&\mathbb{E} [y^p \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma] \\
&\leq \sum_{j = 1}^k \mathbb{E} [(\vec{w}_j^\top x + \eta_j)^p \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma] \\
&= \sum_{j = 1}^k \mathbb{E} [(\vec{w}_j^\top \mathcal{P}_v x + \eta_j + \vec{w}_j^\top \mathcal{P}^\perp_v x)^p\mid \norm{\mathcal{P}^\perp_v x} \leq \gamma] \\
&= \sum_{j = 1}^k \mathbb{E} \sqlr{\sum_{l = 0}^{p / 2} \binom{p}{2l} (\vec{w}_{j}^\top \mathcal{P}_v (x) + \eta_{j})^{p - 2l}(\vec{w}_{j}^\top \mathcal{P}^\perp_v (x))^{2l} \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma} \\
&\leq \sum_{j = 1}^k (p - 1)!! \cdot (\sigma_j^2 + 1)^{p / 2} + \sum_{l = 1}^{p / 2} \mathbb{E} \binom{p}{2l} \sqlr{(\vec{w}_{j}^\top \mathcal{P}_v (x) + \eta_{j})^{p - 2l} (\vec{w}_{j}^\top \mathcal{P}^\perp_v (x))^{2l} \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma} \\
&\leq k \cdot (p - 1)!! \cdot (\sigma_{j^*}^2 + 1)^{p / 2} + \sum_{j = 1}^k \sum_{l = 1}^{p / 2} \binom{p}{2l} \mathbb{E} \sqlr{(\vec{w}_{j}^\top \mathcal{P}_v (x) + \eta_{j})^{p - 2l} (\gamma \cdot \norm{\vec{w}_j})^{2l} \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma}.
\end{align*}
From the previous two displays, we get by taking $p^{th}$ roots and taking the limit as $\gamma \to 0$:
\begin{equation*}
\lr{\frac{1}{2}}^{1 / p}\cdot \sqrt{(\sigma_{j^*}^2 + 1)} \leq \lim_{\gamma \to 0} \lr{\frac{\mathbb{E} [y^p \mid \norm{\mathcal{P}^\perp_v x} \leq \gamma]}{(p - 1)!!}}^{1 / p} \leq k^{1 / p}\cdot \sqrt{(\sigma_{j^*}^2 + 1)}.
\end{equation*}
Taking $p \to \infty$, we obtain:
\begin{equation}
\label{eq:f_def}
F(v) = \sqrt{\max_{j} \abs{w_j^\top v}^2 + 1}.
\end{equation}
Now, let $v^*$ be such that $v^* = \argmax_{\norm{v} = 1} F(v)$. From \eqref{eq:f_def}, we get that $v^* = \pm w_{j^*}$ for some $j^* \in [k]$ satisfying $j^* = \argmax_{j} \norm{w_j}$. Furthermore, $\sigma^* \coloneqq \norm{w_{j^*}} = \sqrt{F(v^*)^2 - 1}$. To identify the correct sign, consider the random variable $(x, y - x^\top (\sigma^* v^*))$. Note that this is a max-selection model with parameter set $\{w_j - \sigma^* v^*\}_{j \in [k]}$ with associated function $\wt{F}$ defined analogously to $F$. Now, we have the following two cases:
\begin{itemize}
\item[] \textbf{Case 1: } $\wt{F} (v^*) = \sqrt{4 \cdot (\sigma^*)^2 + 1}$. In this case, there exists $w \in \mc{W}$ with $w = - \sigma^* v^*$
\item[] \textbf{Case 2: } $\wt{F} (v^*) < \sqrt{4 \cdot (\sigma^*)^2 + 1}$. In this case, we must have $w_{j^*} = \sigma^* v^*$.
\end{itemize}
In either case, we identify a single $w \in \mc{W}$. To complete the reduction, note that $(x, y - w^\top x)$ is a self-selection model with parameter set $\mc{W}^\prime = \{w_j - w\}_{j \in [k]}$. Defining the $k - 1$ sized point set $\mc{W}^\dagger = \{w_j - w\}_{\substack{j \in [k], j \neq j^*}}$, we note the relationship between the distribution functions $\Phi_{\mc{W}^\prime}$ and $\Phi_{\mc{W}^\dagger}$ for all $S \subset \mathbb{R}^d$ with $\P \blr{x \in S} \neq 0$:
\begin{equation*}
\forall t \in \mathbb{R}: \Phi_{W^\dagger} (S \times (-\infty, t]) = \frac{\Phi_{W^\prime} (S \times (-\infty, t])}{\Phi (t)}.
\end{equation*}
Hence, the distribution function $\Phi_{\mc{W}^\dagger}$ is a function of the distribution function of $\Phi_{\mc{W}^\prime}$ which is in turn a function of the distribution function of $\Phi_{\mc{W}}$. From our induction hypothesis, we have that $\mc{W}^\dagger$ is identifiable from $\Phi_{\mc{W}^\dagger}$ and consequently, from $\Phi_{\mc{W}}$. The proof of the inductive step now follows from the observation that $\mc{W} = \{z + w\}_{z \in \mc{W}^\prime} \cup \{w\}$.
\end{proof}
\subsection{Finding an $O(k)$-subspace}
\label{ssec:identifying_subspace}
\newcommand{\bm{M}}{\bm{M}}
We now move towards a finite-sample estimation algorithm for the unknown-index
case. The first step in our approach is an algorithm for approximately
identifying a size-$k$ subspace that has high overlap with $\text{span}(w_1,
\ldots, w_k)$.
In order to estimate the subspace, we will consider the matrix
$\bm{M} = \mathbb{E}\lr{\max(0, y)^2\cdot \vec{x}\vec{x}^\top}$. The following Lemma
shows that the top $k$ eigenvectors of $\bm{M}$ capture the span of the weight
vectors $w_k$:
\begin{lemma}[Weighted covariance]
\label{lemma:svd_cov}
Consider the matrix $\bm{M} = \mathbb{E}\lr{\max(0, y)^2 \cdot
\vec{x}\vec{x}^\top}$, and let $$p_i = \mathbb{P}\lr{\left\{
i = \argmax_{j \in [k]} w_j^\top \vec{x} + \eta_j\right\}
\text{ and } \left\{ w_i^\top \vec{x} + \eta_i > 0\right\}}.$$ Then,
\[
\bm{M} = \mathbb{E}\lr{\max(0, y)^2} \cdot \bm{I} + 2\sum_{i=1}^{k} p_i \cdot w_iw_i^\top.
\]
\end{lemma}
\begin{proof}
\input{subspace_sq}
\end{proof}
We now use the spectral gap shown in the previous Lemma, in combination with a
matrix concentration argument, to argue that $k$-SVD on $\bm{M} -
\mathbb{E}[y^2]\cdot \bm{I}$ (where $\bm{M}$ is as defined in the previous Lemma)
will approximately recover the span of the $\{\vec{w_i}\}$.
We use as a primitive the following result about $k$-SVD:
\begin{fact}[\citep{rokhlin2010randomized}]
\label{fac:k_svd}
Let $M \in \mathbb{R}^{k \times k}$, and let $\sigma_1 \geq \sigma_2,
\geq \ldots \geq \sigma_d$ denote the non-zero singular values of M. For
any $j \in [k-1]$, define the spectral gap $g_j = \sigma_j /
\sigma_{j+1}$. Furthermore, suppose we have access to an oracle which
computes $Mv$ for any $v \in \mathbb{R}^k$ in time $R$. Then, for any
$\eta, \delta > 0$, there is an algorithm $\textsc{ApproxSVD}(M,
\eta, \delta)$ which runs in time $\tilde{O}(\frac{j \cdot R}{\min(1, g_j - 1)} \cdot log(k/(\eta \delta))$
and with probability at least $1 - \delta$ outputs $U \in R^{k\times j}$
with orthonormal columns so that $\|U - U_j\|_2 < \eta$, where $U_k$ is
the matrix whose columns are the top $j$ right singular vectors of M.
\end{fact}
The combination of Lemma \ref{lemma:svd_cov} and Fact \ref{fac:k_svd} imply that it
suffices to show concentration of the average of a sequence $M^{x_i,y_i}$. The
following Lemma helps us establish this concentration:
\begin{restatable}{lemma}{svdconc}
\label{lem:svd_conc}
Suppose we generate $n$ samples $(\vec{x}^{(l)}, y^{(l)}) \sim X, Y$ from the
self-selected linear regression model, i.e., $\vec{x}^{(l)} \sim
\mathcal{N}(0, \bm{I}_d)$, then $y^{(l)}_i = \bm{w}_i^\top \vec{x}^{(l)} +
\mathcal{N}(0, 1)$ for vectors $\vec{w_i} \in \mathbb{R}^d$ with $\|\vec{w_i}\|_2 \leq B$, and $y^{(l)} = \max_{i \in [k]} y^{(l)}_i$. Define the
empirical second-moment matrix
\begin{align*}
\widehat{\bm{M}} = \frac{1}{n}\sum_{l=1}^n \max(0, y^{(l)})^2 \cdot \vec{x}^{(l)} {\vec{x}^{(l)}}^\top.
\end{align*}
Fix any $\delta \in (0, 1)$. Then, if $n \geq \Omega(\max(1/\delta, d))$,
with probability at least $1 - \delta$,
\begin{align*}
\left\|\widehat{\bm{M}} - \mathbb{E}[\widehat{\bm{M}} ]\right\|_2 \in O\lr{
\mathrm{poly}(B)\cdot \frac{\log(kn)}{\sqrt{n}}
\max\left\{
\sqrt{\log(2/\delta)},\sqrt{d}
\right\}
}.
\end{align*}
\end{restatable}
\begin{proof}
See Appendix \ref{app:proof:svd_conc}.
\end{proof}
\noindent We can thus
approximately identify the
relevant subspace in polynomial time.
\subsection{Estimating Parameters using the Low-Dimensional Subspace}
\label{ssec:est_low_dim}
\input{ui_estimation_old}
\subsection{Estimation in the $k = 2$ case}
We will now demonstrate how, when $k = 2$, we can use a moment-based
algorithm to estimate $\{\vec{w_1}, \vec{w_2}\}$ in $\text{poly}(1/\epsilon$)
time. The algorithm will operate as follows:
\begin{enumerate}
\item Using the procedure outlined in Section \ref{ssec:identifying_subspace}, find an
approximation $U$ to the linear subspace $U^*$ containing
$\text{span}(\vec{w_1}, \vec{w_2})$.
\item Set up an $\epsilon$-covering over the span of $U$. Since we assume
$\|\vec{w_i}\| \leq 1$, the covering is of size $O(1/\epsilon^2)$.
\item For each element $\widehat{\vec{w}}$ of the
covering, collect samples $(\vec{x}, y - \vec{x}^\top \vec{v}_t)$ where $(x, y)$ are from the
no-index self selection model.
\item Using the moments of $y - \vec{x}^\top \widehat{\vec{w}}$ estimate
$\min_{i \in \{0, 1\}} \|\vec{w_i} - \vec{w}\|^2$. We will show that
$O(\frac{1}{\delta\varepsilon^2\text{poly}(B)})$ suffices to get an
$\epsilon$-close approximation to this quantity $1-\delta$. Setting
$\delta = O(\rho \epsilon^2)$ for some $\rho < 1$ ensures that we get
accurate estimates of this quantity for each element $\vec{w}$ of our covering.
\item As long as $\vec{w_1}$ and $\vec{w_2}$ are sufficiently separated,
we can estimate $\vec{w_1}$ to be the minimum of our estimate over the
$\epsilon$-covering, i.e., $\widehat{\vec{w_1}} = \arg\min_{w} \min_{i \in \{0, 1\}}
\|\vec{w_i} - \vec{w}\|$. We can then estimate $\vec{w_2}$ to be the
minimizer of the estimate over points that are $\Delta$-far from
$\widehat{\vec{w_1}}$.
\end{enumerate}
Turning this outline into an efficient algorithm entails tackling a few distinct
technical challenges. First, we will show how to estimate $\min_{i \in \{0, 1\}}
\|\vec{w_i} - \vec{v}_t\|$ using samples $(\vec{x}, y)$ from our data-generating
process. Then, we will show that our sequential approach to estimating
$\vec{w_1}$ and $\vec{w_2}$ indeed suffices to recover both with good enough
accuracy. Finally, we will show that the error incurred by the subspace-finding
step does not adversely affect our estimation of the weight vectors.
\paragraph{Distance to the nearest weight vector.}
When $k = 2$, direct integration allows us to compute the moment generating
function of $y = \max \{\vec{w_1}^\top \vec{x} + \eta_1, \vec{w_2}^\top
\vec{x} + \eta_2\}$ in terms of the covariance matrix between $y_1$ and
$y_2$. This in turn allows us to accurately estimate the lesser of
$\text{Var}[y_1]$ and $\text{Var}[y_2]$, as captured by the following Lemma:
\begin{lemma} \label{lem:minVariance}
Given a two-dimensional Gaussian random variable $z \sim \mathcal{N}(0,
\Sigma)$ with $0 \prec \Sigma \in \mathbb{R}^2$, there exists an
algorithm $\textsc{MinVariance}(\delta, \epsilon)$ which given
$O(\frac{1}{\delta\epsilon^2})$ samples, outputs an estimate
$\widehat{\sigma}$ of $\min_{0, 1} \Sigma_{ii}$ satisfying
\[
(1 - \beta)\lr{\min_{0, 1} \Sigma_{ii}} \leq \widehat{\sigma} \leq (1 + \beta) \lr{\min_{0, 1} \Sigma_{ii}}
\]
with probability $1 - \delta$.
\end{lemma}
\begin{proof}
\input{k_equals_two}
\end{proof}
As a corollary of this result, we can estimate $\min_{i \in \{0, 1\}}
\|\vec{w_i} - \vec{w}\|^2$ to $\epsilon$-precision with probability at least $1
- \delta$:
\begin{corollary}
Suppose we have samples $\{(x, y)\}$ generated from the self-selection model
with unobserved index. Suppose further that $\|\vec{w_i}\| \leq B$ for $i
\{0, 1\}$. Then, we can use the \textsc{MinVariance} algorithm of Lemma
\ref{lem:minVariance} to recover $\sigma$ such that with probability at least
$1 - \delta$,
\[
\left|\sigma - \min_{i \in \{0, 1\}} \|\vec{w_i} - \vec{w}\|^2 \right| \leq \epsilon.
\]
\end{corollary} \label{cor:minVariance}
\begin{proof}
Define the random variable $X = y - \vec{w}^\top x$. Then, $X$ is the
maximum of the two Gaussian random variables
$X_1 = (\vec{w_1} - \vec{w})^\top \vec{x} + \eta_i$ and
$X_2 = (\vec{w_2} - \vec{w})^\top \vec{x} + \eta_i$.
In particular, $\text{Var}[X_i] = \|\vec{w_i} - \vec{w}\|^2 + 1$. Thus, by
applying the \textsc{MinVariance} algorithm to $X$, we can recover the
desired quantity.
\end{proof}
\paragraph{Accuracy loss due to subspace recovery.}
From Section \ref{ssec:identifying_subspace}, we may assume the existence of a
$2$-dimensional subspace $\hat{U}$ satisfying the following:
\begin{equation*}
\forall i \in [2]: \frac{\norm{w_i - \mathcal{P}_{\hat{U}} w_i}}{\norm{w_i}} \leq \varepsilon/2
\end{equation*}
where to find $\hat{U}$ we need $\mathrm{poly}(d, 1/\varepsilon, B, 1/\alpha)$ sample and time
complexity. Now from Corollary \ref{cor:minVariance} we have that if we estimate
$\sigma_{\min}$ within accuracy $\varepsilon/10$ then we can identify the point on the
$2$-dimensional $\varepsilon/2$-cover that is at least $\varepsilon$ close to $\vec{w}_1$,
assuming also that $\vec{w}_2$ is at least $\Delta \ge \varepsilon$ far away from
$\vec{w}_1$. Once we have identified $\vec{w}_1$ we can substract
$\vec{w}_1^{\top} \vec{x}$ from all the next samples of the form $(\vec{x}, y)$
find the minimum variance among the grid points that are at least $\Delta - \varepsilon$
far away from the estimated $\vec{w}_1$ and this way we are guaranteed to recover
the remaining $\vec{w}_2$ as well. |
\section{Introduction}
In living cells, the shapes of biomembranes are changed dynamically depending on their functions~\cite{mcma05,suet14}.
For example, spherical vesicles are formed via membrane budding during endocytosis, exocytosis, and intracellular traffic.
Curvature-inducing proteins such as clathrin and coat protein complexes bend membranes into spherical buds~\cite{joha15,bran13,hurl10,mcma11,kaks18}.
In contrast, Bin/Amphiphysin/Rvs (BAR) superfamily proteins have a crescent binding domain to bend the membrane along the domain axis;
thus, their binding generates cylindrical membrane tubes rather than spherical buds~\cite{mcma05,suet14,joha15,itoh06,masu10,baum11,mim12a,fros08,adam15}.
Many BAR proteins also have intrinsically disordered domains. These domains are unfolded in physiological conditions. They ave been often removed for {\it in vitro} experiments to investigate
the interactions between the BAR domains and membranes.
However, recent experiments~\cite{stac12,busc15,zeno18,zeno19,snea19} have shown that such intrinsically disordered domains can significantly influence the membrane shapes.
They changed the length of disordered domains in BAR and other proteins using gene manipulation and revealed that the disordered domains induce curvature sensing
and also the long disordered domains induce the formation of small vesicles.
Thus, it is important to understand the role of disordered domains in membrane remodeling.
A disordered domain behaves like a linear polymer chain in a good solvent~\cite{hofm12},
i.e., its mean radius of gyration is scaled as $R_{\rm g} \sim {n_{\rm poly}}^{0.6}$,
where $n_{\rm poly}$ is the number of Kuhn segments~\cite{doi86,stro97}.
Interactions between anchored polymer chains and membrane have been intensively studied in theory~\cite{lipo95a,hier96,bick01,mars03,bick06},
simulations~\cite{auth03,auth05,wern10,wu13}, and experiments~\cite{tsaf01,tsaf03,akiy03,niko07,koth11}.
Polymer anchoring induces a positive spontaneous curvature of the membrane and increases the bending rigidity. These
relations are analytically derived using Green's function method for a low polymer density (mushroom region) \cite{lipo95a,hier96,bick01} and
scaling methods for a high density (brush region) \cite{lipo95a,hier96,mars03} and are confirmed by Monte Carlo (MC) simulations~\cite{auth03,auth05,wern10}.
In Ref.~\citenum{zeno18},
an MC method similar to that developed in Ref.~\citenum{auth03} was used to analyze experimental data.
Under phase-separation conditions, the polymer anchoring can induce the division of a large domain into smaller domains to gain a larger conformational entropy of chains~\cite{wu13}.
Formation of caveolae-like dimple-shaped membranes by the polymer anchoring was discussed in Ref.~\citenum{evan03a}.
In a poor solvent, the polymer anchoring can induce a negative spontaneous curvature~\cite{wern10,evan03a}.
Experimentally, the formation of spherical buds~\cite{tsaf01,tsaf03,niko07} and membrane tubes~\cite{tsaf03,akiy03,koth11} have been observed.
Thus, the effects of the polymer anchoring on membrane deformation are well understood.
However, to the best of our knowledge, the combination with the anisotropic bending of the BAR domains has not been studied yet in theory and simulations.
The aim of this study is to clarify the collective effects of the anisotropic curvature-inducing domains and anchored polymer chains.
Previously, we have simulated anisotropic curvature-inducing proteins using a meshless membrane method~\cite{nogu22a,nogu14,nogu15b,nogu16,nogu16a,nogu17,nogu17a,nogu19a,nogu22}.
The assembly of crescent rods induces tubulation from a flat membrane~\cite{nogu16,nogu17a}, and the protein chirality enhances this ability~\cite{nogu19a}.
Moreover, vesicles are deformed into disks and polyhedral shapes, since the edges are stabilized by protein assemblies~\cite{nogu14,nogu15b,nogu16}.
In this study, anchored chains were added to the chiral protein rod model~\cite{nogu19a} to investigated how these shape deformations are changed by polymer anchoring to proteins.
The simulation model and method are described in Sec.~\ref{sec:method}.
The results are presented and discussed in Sec.~\ref{sec:results}.
Deformations of the vesicles and flat membranes are described
in Sec.~\ref{sec:ves} and Sec.~\ref{sec:flat}, respectively.
Finally, a summary is presented in Sec.~\ref{sec:sum}.
\section{Simulation Model}\label{sec:method}
A fluid membrane is represented by a self-assembled single-layer sheet of $N$ particles.
The position and orientation vectors of the $i$-th particle are ${\bm{r}}_{i}$ and ${\bm{u}}_i$, respectively.
The membrane particles interact with each other via a potential $U=U_{\rm {rep}}+U_{\rm {att}}+U_{\rm {bend}}+U_{\rm {tilt}}$.
The potential $U_{\rm {rep}}$ is an excluded-volume interaction with a diameter $\sigma$ for all pairs of particles.
The effective attractive potential $U_{\rm {att}}$ implicitly represents the interaction involving the solvent.
The details of the meshless membrane model and protein rods are described
in Ref.~\citenum{shib11} and Refs.~\citenum{nogu14,nogu19a}, respectively.
We employ the parameter sets used in Ref.~\citenum{nogu19a}.
The bending and tilt potentials
are given by
$U_{\rm {bend}}/k_{\rm B}T=(k_{\rm {bend}}/2) \sum_{i<j} ({\bm{u}}_{i} - {\bm{u}}_{j} - C_{\rm {bd}} \hat{\bm{r}}_{i,j} )^2 w_{\rm {cv}}(r_{i,j})$
and $U_{\rm {tilt}}/k_{\rm B}T=(k_{\rm{tilt}}/2) \sum_{i<j} [ ( {\bm{u}}_{i}\cdot \hat{\bm{r}}_{i,j})^2 + ({\bm{u}}_{j}\cdot \hat{\bm{r}}_{i,j})^2 ] w_{\rm {cv}}(r_{i,j})$, respectively,
where ${\bm{r}}_{i,j}={\bm{r}}_{i}-{\bm{r}}_j$, $r_{i,j}=|{\bm{r}}_{i,j}|$,
$\hat{\bm{r}}_{i,j}={\bm{r}}_{i,j}/r_{i,j}$, $w_{\rm {cv}}(r_{i,j})$ is a weight function, and $k_{\rm B}T$ is the thermal energy.
The spontaneous curvature $C_0$ of the membrane is
given by $C_0\sigma= C_{\rm {bd}}/2$. \cite{shib11}
In this study, $C_0=0$ and $k_{\rm {bend}}=k_{\rm{tilt}}=10$ are used except for the membrane particles belonging to the protein rods.
A BAR domain and the membrane underneath it are modeled together as a rod,
that is a linear chain of $n_{\rm {sg}}$ membrane particles~\cite{nogu14} (see the left snapshot in Fig.~\ref{fig:snapves}(b)).
We use $n_{\rm {sg}}=10$, i.e., the rod length $r_{\rm rod}=10\sigma$.
The protein rods have spontaneous curvatures $C_{\rm {rod}}$ along the rod axis
and no spontaneous (side) curvatures perpendicular to the rod axis.
The protein-bound membrane is more rigid than the bare membrane as $k_{\rm {bend}}=k_{\rm{tilt}}=80$.
Since a BAR domain is dimmer and chiral, we model it as follows.
Two particles are connected to the end and next particles by a bond potential
and located laterally on opposite sides to represent the protein chirality~\cite{nogu19a}
(see cyan particles in Fig.~\ref{fig:snapves}(b)).
Two linear chains consisting of $n_{\rm poly}$ particles are anchored to the third rod particles from both ends
(see green particles in Fig.~\ref{fig:snapves}(b)).
These protein particles have excluded-volume interactions with diameter $\sigma$ as the membrane particles.
Neighboring segments in the chains are connected by a harmonic potential $U_{\rm bond}= (k_{\rm poly}/2)(r_{ij}- r_{\rm p0})^2$.
In this study, $k_{\rm poly}=10k_{\rm B}T$ and $r_{\rm p0}=1.2\sigma$ are used.
These excluded-volume chains are set on one side of the membrane (outside of a vesicle or upper side of a flat membrane).
The membrane has typical mechanical properties of lipid membranes:
the bare membrane has a bending rigidity $\kappa/k_{\rm B}T=15 \pm 1$,
area of the tensionless membrane per particle $a_0/\sigma^2=1.2778\pm 0.0002$,
area compression modulus $K_A\sigma^2/k_{\rm B}T=83.1 \pm 0.4$,
edge line tension $\Gamma\sigma/k_{\rm B}T= 5.73 \pm 0.04$~\cite{nogu14},
and the Gaussian modulus $\bar{\kappa}/\kappa=-0.9\pm 0.1$~\cite{nogu19}.
We primarily use a rod curvature of $C_{\rm rod}r_{\rm rod}=3$ that corresponds to that of BAR-PH~\cite{zhu07}.
The length of the excluded-volume chains are varied as $n_{\rm poly}=0$, $10$, $25$, $50$, $100$, and $200$:
the mean end-to-end distance of a free chain is $r_{\rm end} \simeq r_{\rm p0}{n_{\rm poly}}^{0.6}= 0$,
$4.8\sigma$, $8.3\sigma$, $13\sigma$, $19\sigma$, and $29\sigma$, respectively,
where the exponent $0.6$ is for the excluded-volume chains in a good solvent~\cite{doi86,stro97}.
The excluded-volume chains are attached to all protein rods unless otherwise specified.
For tubulation from a flat membrane, the case of the excluded-volume chains attached only to 10\% of protein rods is also examined.
The total number of membrane particles $N=25600$ is used for both vesicles and flat membranes.
The radius of the spherical vesicle is $50\sigma$.
The protein density varies from $\phi= n_{\rm {sg}}N_{\rm {rod}}/N=0.0375$ to $0.1$ for the vesicles
and $\phi=0.05$ and $0.2$ for flat membranes, respectively.
Zero surface tension is used for the flat membranes.
To generate an initial conformation,
straight chains are anchored first and then the chain conformation is relaxed with or without fixing the membrane position.
Molecular dynamics with a Langevin thermostat is employed~\cite{shib11,nogu11}.
For the time unit, $\tau= r_{\rm {rod}}^2/D$ is used,
where $D$ is the diffusion coefficient of the membrane particles in the tensionless membrane~\cite{nogu16a}.
Error bars are estimated from three and ten independent runs for vesicles and tubulations, respectively.
\begin{figure}[t]
\includegraphics[]{fig1.eps}
\caption{
Snapshots of vesicles at $C_{\rm rod}r_{\rm rod}=3$ and $\phi=0.1$.
A protein is displayed as a crescent rod (red and yellow)
connected with two particles (cyan) on the laterally opposite sides
and with two excluded-volume chains (transparent green) consisted of $n_{\rm poly}$ particles, as shown in the left panel in (b).
The orientation vector ${\bf u}_i$ lies along the direction from the yellow to red hemispheres.
The cylinder width and transparency of anchored chains are adjusted for each snapshot for clarity.
The membrane particles are displayed as transparent gray spheres.
Spherical vesicles at $C_{\rm rod}=0$ are used as initial states.
(a) Time evolution for the protein rods with no anchored chains ($n_{\rm poly}=0$).
From left to right, $t/\tau=0$, $30$, and $450$.
The membrane particles are not displayed in the right snapshot for clarity.
(b) A vesicle shape at $t/\tau=800$ for $n_{\rm poly}=50$.
(c) A vesicle shape at $t/\tau=800$ for $n_{\rm poly}=200$.
In the left snapshot,
the anchored chains are not displayed and the membrane particles are displayed as small spheres for clarity.
}
\label{fig:snapves}
\end{figure}
\begin{figure}[t]
\includegraphics[]{fig2.eps}
\caption{
Time development of (a) the asphericity $\alpha_{\rm sp}$, (b) aplanarity $\alpha_{\rm pl}$, and (c) mean cluster size $\langle N_{\rm cl} \rangle$ at $C_{\rm rod}r_{\rm rod}=3$ and $\phi=0.1$.
The blue, green, and red curves in (a),(b) correspond to those in Figs.~\ref{fig:snapves}(a), (b), and (c), respectively.
(c) The cluster size is averaged for three independent runs,
and the error bars are displayed at several data points.
From top to bottom, $n_{\rm poly}=0$, $10$, $25$, $50$, and $200$.
}
\label{fig:sphrg}
\end{figure}
A protein is considered to belong to a cluster
when the distance between the centers of mass of the rod region
and one of the rods in the cluster is less than $r_{\rm {rod}}/2$.
The mean size of the clusters is given by
$N_{\rm {cl}}= (\sum_{i_{\rm {cl}}=1}^{N_{\rm {rod}}} i_{\rm {cl}}^2 n^{{\rm {cl}}}_i)/N_{\rm {rod}}$,
where $n^{{\rm {cl}}}_i$ is the number of clusters of size $i_{\rm {cl}}$.
To quantify vesicle shapes, we calculate two shape parameters based on
the gyration tensor $a_{\alpha\beta}= (1/N)\sum_i (\alpha_{i}-\alpha_{\rm G})(\beta_{i}-\beta_{\rm G})$
of membrane particles including the bound region of the proteins (i.e., excluding anchored chains), where $\alpha,\beta\in \{x,y,z\}$:
\begin{eqnarray}
\alpha_{\rm {sp}} &=& \frac{({\lambda_1}-{\lambda_2})^2 +
({\lambda_2}-{\lambda_3})^2+({\lambda_3}-{\lambda_1})^2}{2(\lambda_1+\lambda_2+\lambda_3)^2}, \\
\alpha_{\rm {pl}} &=& \frac{9\lambda_1\lambda_2\lambda_3} {(\lambda_1+\lambda_2+\lambda_3)
(\lambda_1\lambda_2+\lambda_2\lambda_3+\lambda_3\lambda_1)},
\end{eqnarray}
where ${\lambda_1} \leq {\lambda_2} \leq {\lambda_3}$ are
three eigenvalues of the gyration tensor.
Spherical, disk, and rod shapes are well distinguished by $\alpha_{\rm {sp}}$ and $\alpha_{\rm {pl}}$.
The asphericity, $\alpha_{\rm {sp}}$, indicates the shape deformation from a spherical
shape to a thin-rod shape:~\cite{rudn86} $\alpha_{\rm {sp}}=0$, $0.25$, and $1$
for a sphere (${\lambda_1}={\lambda_2}={\lambda_3}$), thin disk (${\lambda_1}=0, {\lambda_2}={\lambda_3}$),
and thin rod (${\lambda_1}={\lambda_2}=0$), respectively.
A discocyte (red-blood-cell) shape takes $\alpha_{\rm {sp}}\simeq 0.2$.~\cite{nogu05}
The aplanarity, $\alpha_{\rm {pl}}$, indicates the deviation from a plane:~\cite{nogu06}
$\alpha_{\rm {pl}}=0$ and $1$ for a flat membrane (${\lambda_1}=0$) and sphere, respectively.
In addition to vesicles~\cite{nogu05,nogu15a,tame20} and polymer chains~\cite{nogu98},
these quantities were used to characterize the cup-shaped membrane closing into a vesicle~\cite{nogu19} and
the arrangements of pores of the vesicle~\cite{nogu16b}.
\begin{figure}[t]
\includegraphics[]{fig3a.eps}
\includegraphics[]{fig3b.eps}
\caption{
Transition between arc-shaped and circular assemblies of protein rods at $C_{\rm rod}r_{\rm rod}=3$.
(a),(b) Snapshots of (a) the arc-shaped and (b) circular assemblies for $n_{\rm poly}=25$ and $100$ at $\phi=0.05$, respectively.
The membrane particles are displayed as small transparent spheres for clarity.
(c) Phase diagram. The circular (arc-shaped) assembly exists in the region above the blue line (under the red line),
when the circular (arc-shaped) assembly is used as an initial conformation. These two states coexist in the region between two lines.
}
\label{fig:pd}
\end{figure}
\begin{figure}[t]
\includegraphics[]{fig4.eps}
\caption{
Sequential snapshots of tubulation from a flat membrane at $C_{\rm rod}r_{\rm rod}=3$, $\phi=0.2$, and $n_{\rm poly}=10$.
From top left to bottom right, $t/\tau=0$, $30$, $70$, $90$, $120$, and $465$.
}
\label{fig:snapn512n10}
\end{figure}
\section{Results and Discussion}\label{sec:results}
\subsection{Vesicle Deformation}\label{sec:ves}
First, we consider the deformation of vesicles by the proteins (Figs.~\ref{fig:snapves}--\ref{fig:pd}).
At $C_0=0$, the proteins are randomly disturbed on the vesicle either with or without anchored chains (see the left snapshot in Fig.~\ref{fig:snapves}(a)).
As the curvature is changed to $C_0r_{\rm rod}=3$ at $t=0$, the proteins self-assemble on the vesicle from this random state (Figs.~\ref{fig:snapves} and \ref{fig:sphrg}).
Before discussing the effects of the anchored chains, we describe vesicle deformation in their absence ($n_{\rm poly}=0$).
A parallel pair of proteins have an attraction in the direction perpendicular to the protein axis via membrane-mediated interactions~\cite{nogu17} so that the proteins form a side-by-side assembly along one direction.
At a sufficiently low protein density $\phi$, the protein assembly exhibits an arc shape, leading to a disk-shaped vesicle~\cite{nogu14} (see Figs.~\ref{fig:snapves}(a)--(c) and Movie S1). At a higher density, the arc is closed to a circular shape, and with increasing density, the assembly elongates to an elliptic shape~\cite{nogu15b,nogu16}. At an even higher density, polyhedral vesicles are formed~\cite{nogu15b}.
Anchored chains change the vesicle shape.
As anchored chains become longer at $\phi=0.1$, vesicles exhibit a twisted-arc shape at $n_{\rm poly}= 50$ (Fig.~\ref{fig:snapves}(b)) and multiple arc-shaped protrusions at $n_{\rm poly}\gtrsim 100$ (Fig.~\ref{fig:snapves}(c) and Movie S2).
Although this twisted-arc shape can be temporally formed by proteins without chains at a higher density~\cite{nogu16},
it quickly transforms into a disk shape. Conversely, the twisted shape is maintained for a long period (at least until $t/\tau=1000$) at $n_{\rm poly}= 50$ owing to the chain interactions.
The anchored chains generate repulsion between neighboring proteins as well as between protein assemblies.
Moreover, they induce an effective spontaneous curvature, such that the protein assembly is more curved along the arc.
These interactions suppress fusion of protein assemblies and stabilize the shape, as shown in Fig.~\ref{fig:snapves}(c).
To investigate the assembly dynamics in more details, the time development of the vesicle shapes and protein cluster sizes is plotted in Fig.~\ref{fig:sphrg}.
The disk-shaped vesicle with no anchored chains ($n_{\rm poly}=0$) has the asphericity $\alpha_{\rm sp} \simeq 0.2$ and aplanarity $\alpha_{\rm pl} \simeq 0.3$.
The chain anchoring makes vesicles more spherical shaped in the ellipsoidal approximation
(Figs.~\ref{fig:sphrg}(a) and (b)).
The initial stage of the protein assembly becomes slower as the chain length increases from $n_{\rm poly}=0$ to $50$,
and it is almost constant at $n_{\rm poly} \geq 50$ (see Fig.~\ref{fig:sphrg}(c)).
This threshold length $n_{\rm poly}=50$, the average distance ($\simeq 10\sigma$) between randomly distributed protein rods is close to the end-to-end distance of free chains.
Hence, the initial assembly is slowed by the repulsive interaction between the anchored chains,
but this effect is saturated at a chain length equal to the protein distance.
To clarify the interaction between neighboring protein rods in the assembly,
the closing transition of the arc-shaped assembly is investigated (Fig.~\ref{fig:pd}).
When long chains are anchored to an arc-shaped assembly,
the assembly transforms into a circular shape, as shown in Fig.~\ref{fig:pd}(b).
Conversely, when the chain length is reduced,
the circular assembly opens into an arc shape, as shown in Fig.~\ref{fig:pd}(a).
The thresholds of these changes are different, as shown in Fig.~\ref{fig:pd}(c),
and the two states coexist at intermediate chain lengths.
Therefore, this is a first-order transition.
The mean distance between neighboring protein rods increases with increasing chain length.
Even for low density $\phi\simeq 0.04$, where less than a half of the circumference is occupied by the protein rods,
a circular assembly is formed at long chains (Fig.~\ref{fig:pd}(c)).
\begin{figure}[t]
\includegraphics[]{fig5.eps}
\caption{
Snapshots of tubulation from a flat membrane at $C_{\rm rod}r_{\rm rod}=3$ and $\phi=0.2$
(a) $n_{\rm poly}=0$. (b) 10\% of the protein rods have anchored chains of $n_{\rm poly}=100$.
(c) $n_{\rm poly}=25$. (d) $n_{\rm poly}=100$.
The anchored chains are not displayed in the left snapshots in (c) and (d) for clarity.
The right snapshots in (c) and (d) are from the top view.
}
\label{fig:snapn512s}
\end{figure}
\begin{figure}[t]
\includegraphics[]{fig6.eps}
\caption{
Time development of (a) mean cluster size $\langle N_{\rm {cl}} \rangle$ and
(b) vertical membrane span $\langle {z_{\rm {mb}}}^2 \rangle^{1/2}$ at $C_{\rm rod}r_{\rm rod}=3$ and $\phi=0.2$.
The solid lines represent the data of the chains anchoring to all protein rods for $n_{\rm poly}=0$, $10$, $25$, $50$, and $100$.
The dashed lines represent the data of the chains anchoring to 10\% of protein rods for $n_{\rm poly}=100$.
}
\label{fig:flatn512}
\end{figure}
\begin{figure}[t]
\includegraphics[]{fig7a.eps}
\includegraphics[]{fig7b.eps}
\caption{
Tubulation from a flat membrane at $C_{\rm rod}r_{\rm rod}=3$ and low density $\phi=0.05$.
(a),(b) Snapshots for (a) $n_{\rm poly}=25$ and (b) $n_{\rm poly}=100$ at $t/\tau=600$ and $800$, respectively.
(c),(d) Time development of (c) mean cluster size $\langle N_{\rm {cl}} \rangle$ and
(d) vertical membrane span $\langle {z_{\rm {mb}}}^2 \rangle^{1/2}$.
The solid lines represent the data of the chains anchoring to all protein rods for $n_{\rm poly}=0$, $25$, $50$, and $100$.
The dashed lines represent the data of the chains anchoring to 10\% of protein rods for $n_{\rm poly}=100$.
}
\label{fig:flatn128}
\end{figure}
\subsection{Deformation of Flat Membrane}\label{sec:flat}
Next, we consider tubulation from a flat tensionless membrane (Figs.~\ref{fig:snapn512n10}--\ref{fig:flatn128}).
At the high density ($\phi=0.2$), tubulation occurs for all values of $n_{\rm poly}$ at $C_{\rm rod}r_{\rm rod}=3$.
For short chains ($n_{\rm poly} \leq 10$), tubules are formed via bending of the protein assembly and fuse into a larger tubule
(see Fig.~\ref{fig:snapn512n10} and Movie S3 for $n_{\rm poly} =10$ and Fig.~\ref{fig:snapn512s}(a) for $n_{\rm poly}=0$).
The protein chirality promotes tubulation via the helical assembly on tubules~\cite{nogu19a}.
With increasing chain length $n_{\rm poly}$, the mean cluster size $\langle N_{\rm {cl}} \rangle$ and membrane vertical span $\langle {z_{\rm {mb}}}^2 \rangle^{1/2}$ decrease until $n_{\rm poly}=25$ (Fig.~\ref{fig:flatn512}).
The vertical span of the membrane is calculated from
the membrane height variance as
${z_{\rm mb}}^2=\sum_{i}^{N} (z_i-z_{\rm G})^2/N$,
where $z_{\rm G}$ is the center of mass in the vertical ($z$) direction.
Interestingly, the vertical span increases from $n_{\rm poly}=25$ to $n_{\rm poly}=100$.
At $n_{\rm poly}=25$, the short tubules are separated by the repulsive interaction between anchored chains,
where the distance between tubules is greater than the end-to-end distance $r_{\rm end}$ (see the right snapshot in Fig.~\ref{fig:snapn512s}(c)).
A similar stabilization of small domains was previously obtained in a phase-separated membrane, in which polymer chains were anchored to one of the phases~\cite{wu13}.
At $n_{\rm poly}=100$, $r_{\rm end}$ is greater than the tubule distance,
so that the chains push the tubules in the upper direction and also lead to narrower shapes to gain more conformational entropy.
Hence, vertically elongated tubules are obtained as shown in Fig.~\ref{fig:snapn512s}(d) and Movie S4.
For the low density ($\phi=0.05$), tubulation is suppressed with increasing chain length from $n_{\rm poly}=0$ to $n_{\rm poly}=100$ (see Fig.~\ref{fig:flatn128}).
Since fewer protein assemblies are more separated,
the dependency for $n_{\rm poly}\leq 25$ at $\phi=0.2$ are seen even at $n_{\rm poly}=100$.
Until this point, the chains are anchored to all proteins.
When the chains of $n_{\rm poly}=100$ are anchored to 10\% of the proteins,
their effects are negligibly small.
For $\phi=0.2$, the difference from the unanchored proteins is not recognized (compare the magenta and black dashed lines in Fig.~\ref{fig:flatn512}, and also compare the snapshots in Figs.~\ref{fig:snapn512n10}(a) and (b)).
For $\phi=0.05$, there are small differences after tubules are formed (compare the magenta and black dashed lines in Fig.~\ref{fig:flatn128}).
Thus, the interactions between anchored chains are essential for the tubulation,
whereas the effects of isolated chains are marginal.
Next, we consider the formation of spherical buds.
At $C_{\rm rod}=0$, the flat membrane is stable on our simulation time scale ($\sim 500\tau$, see Fig.~\ref{fig:snapn512n10}(a)).
However, when a tubulated membrane is used as an initial state,
one to three spherical buds are formed at $n_{\rm poly}=50$ (see Fig.~\ref{fig:snapc0}(b) and Movie S5).
Note that the bud shapes are deviated form a perfect sphere.
In the bottom half of the buds (close to bud neck),
the membrane is elongated in the vertical direction, because the protein rods are orientated in the vertical direction.
When the chain length is reduced, one large bud remains at $n_{\rm poly}=25$, but all buds disappear at $n_{\rm poly}=10$.
Thus, the chain anchoring can induce buds, although a high free-energy barrier exists for their formation.
Finally, we consider the case in which protein rods and anchored chains bend the membrane in the opposite directions,
i.e., the chains are anchored to the concave side of the rods ($C_{\rm rod}<0$).
When these opposite bending effects are compatible,
proteins form branched assemblies on the membrane as shown in Fig.~\ref{fig:cn6}(a).
Thus, the membrane is rugged by these bumped assemblies.
This behavior agrees well with the experimental observation of vesicles induced by a chimeric protein of I-BAR and disordered domains~\cite{snea19}.
They showed that the disordered domain suppressed inward membrane invagination, and vesicles became a rugged shape.
For longer chains ($n_{\rm poly}=50$), the anchored chains have stronger bending effects;
proteins are distributed more randomly at the low density ($\phi=0.05$)
and form shish-kebab-shaped tubules, in which concave ring-shaped rod assemblies are formed periodically, at the high density ($\phi=0.2$), as shown in Figs.~\ref{fig:cn6}(b)--(d).
The membrane vertical span $z_{\rm mb}$ increases accordingly (Fig.~\ref{fig:cn6}(e)).
The formation of this tubule is very slow and it continues to grow at $t/\tau=500$ (see Figs.~\ref{fig:cn6}(c),(e) and Movie S6).
To examine longer tubular structures,
spherical buds are used as an initial state and the shish-kebab-shaped tubules are obtained (see Fig.~\ref{fig:cn6}(d)).
Small buds and tubules disappear or merge into larger tubules, as shown in Movie S7.
Although a single ring assembly was previously obtained at the neck of a spherical bud~\cite{nogu17a},
the periodic ring structure was found here for the first time.
Since I-BAR proteins do not contain long disordered domains,
this shish-kebab-shaped tubule is not likely to exist in living cells
but can be experimentally constructed by designing chimeric proteins of I-BAR and longer disordered domains.
\begin{figure}[t]
\includegraphics[]{fig8.eps}
\caption{
Snapshots of flat membranes at $C_{\rm rod}=0$, $\phi=0.2$, and $n_{\rm poly}=50$.
(a) Randomly distributed proteins in a flat membrane.
(b) Formation of spherical buds.
}
\label{fig:snapc0}
\end{figure}
\begin{figure}[t]
\includegraphics[]{fig9a.eps}
\includegraphics[]{fig9b.eps}
\caption{
Deformation of flat membranes at $C_{\rm rod}r_{\rm rod}=-3$.
Excluded-volume chains are anchored to the concave side of protein rods.
(a)--(d) Snapshots of membranes at (a) $(\phi, n_{\rm poly})=(0.05,25)$,
(b) $(\phi, n_{\rm poly})=(0.05,50)$, and (c),(d) $(\phi, n_{\rm poly})=(0.2,50)$.
Randomly mixed states at $C_{\rm rod}=0$ are used as initial states for (a)--(c),
while a tubulated state at $C_{\rm rod}r_{\rm rod}=3$ is used for (d).
(e) Time development of the vertical span of the membrane.
Top to bottom: $(\phi, n_{\rm poly})=(0.2,50)$, $(0.05,25)$, and $(0.05,50)$.
The snapshots in (a)--(c) correspond to the data at $t/\tau=500$.
}
\label{fig:cn6}
\end{figure}
\section{Summary}\label{sec:sum}
We have studied the membrane deformation induced by the proteins consisting of a chiral crescent domain and two anchored chains.
We have shown that these two types of domains can bend membranes both cooperatively and uncooperatively.
Anchored chains induce an effective spontaneous curvature of the membrane but also generate repulsion between protein rods as well as between protein assemblies.
The latter can weaken or suppress the membrane deformation by the protein rods, depending on the conditions.
For vesicle deformation, long chains induce multi-spindle vesicles and transition from arc-shaped protein assembly to circular assembly.
For tubulation from a flat membrane, small tubules are formed by repulsion between the crowded chains of tubules.
However, long vertical tubules are formed by very long chains that are longer than the mean distance between tubules.
Thus, the chains suppress the lateral assembly but promote vertical elongation.
Spherical buds can be formed by straight protein rods with sufficiently long anchored chains.
When the rod and chain domains bend membranes in opposite directions at similar amplitudes,
bumped network assemblies are formed on the membrane.
In contrast, shish-kebab-shaped tubules are formed when the long chains have stronger bending ability.
Our results suggest that the interactions between chains are more significant than the chain-membrane interactions.
The highly packed brush states of the chains induce large repulsion between protein assemblies and membrane bending.
In particular, brushed chains strongly push tubules away from the membrane, when the membrane is completely covered by the chain crowds.
Our findings provide new insights into the mechanism of membrane shape regulation by curvature-inducing proteins with disordered domains
and the design of curvature-inducing proteins.
\begin{acknowledgments}
This work was supported by JSPS KAKENHI Grant Number JP21K03481.
\end{acknowledgments}
|
\section{Introduction}
Different studies for the behavior of magnetic fields in stars and planets, and in particular the Earth, have supported the hypothesis that inside of these systems there is a flow of highly
conducting materials that produces magnetic induction \cite{Glatzmaiers1995,Braginsky1995,Roberts2000,Monchaux2009,Christensen2010}.
The usual framework for modelling this generation of magnetic fields from the kinetic energy of the flow
is that of magnetohydrodynamic (MHD) dynamos.
According to those studies, the rotatory movement of the planet can be one of the physical ingredients for
these chaotic, and often, turbulent flows \cite{stewartson1963,Busse1968,Noir2003,Tilgner2005}.
Suggested drivers for these flows include thermal and/or chemical reactions effects. A distinctive driver that was
also suggested is the precession effect in the otherwise constant
rotation of the planet\cite{stewartson1963,Busse1968,Noir2003,Tilgner2005,Malkus1968,Keke}, a mechanism that can be responsible for the injection of kinetic energy
in the system. In the dynamics of the fluid (or magnetofluid) precession acts
essentially as a forcing, through a term in the equations of motion (Navier-Stokes or MHD) that
it is proportional to the time derivative of the angular velocity.
For some dynamo systems that have a preferential direction, as is the case of rotating celestial bodies,
the magnetic fields arising from flow dynamics are known to undergo polarity reversals in their magnetic dipole
moment\cite{Babcock1961,Ponty2004,Olson2009,Amit2010,Monchaux2009}. We have previously studied these reversals in the context of
the magnetohydrodynamic equations with steady rotation \cite{Dmitruk2011,Dmitruk2014,Fontana2018}.
Here we consider the influence of a non-steady component in the angular velocity (i.e. of precession) on that phenomenon.
In this work we analyze the effect of precession on a conducting fluid inside a spherical cavity using 75 direct numerical simulations (DNS)
of the MHD equations. We first consider the purely hydrodynamic equations (i.e. no magnetic field) under rotation and precession at constant frequency,
and perform a parametric study to analyze the asymptotic behavior of the flow once a statistically stationary state is reached.
Afterwards we proceed to analyze the magnetohydrodynamic case, starting from an initial random magnetic field and observing its evolution.
We then characterize the conditions in which a turbulent regime is reached and the presence of dynamo action,
focusing on the sensitivity to the rotation and precession frequency values.
The organization of the paper is as follows. Section \ref{sec:level1} contains the model equations, including the initial conditions, and a description of the numerical method employed in the simulations. The results for these numerical simulations are found in section \ref{sec:Results}, distinguishing between the hydrodynamics (HD) and magnetohydrodynamics (MHD) cases. Finally, in the section \ref{conclusions} we summarize the main results of this study.
\section{\label{sec:level1}Simulation setup}
\subsection{\label{sec:level2}Model equations and numerical scheme}
In this work we study a rotating spherical domain $V$ filled with an incompressible conducting fluid. The boundary is considered to be rotating in a non-steady way (i.e. precessing), with an angular velocity $\boldsymbol{\Omega} = \Omega_0[\cos(\gamma t)\sin(\alpha)\Hat{\boldsymbol x}+\sin(\gamma t)\sin(\alpha)\Hat{\boldsymbol y}+\cos(\alpha)\Hat{\boldsymbol z}]$, where $\Omega_0$ is the amplitude of the angular velocity, $\gamma$ its precession frequency and $\alpha$ its precession angle. The density of the fluid is taken to be uniform and equal to unity. In a non-inertial reference frame that rotates fixed with the domain's boundary, the usual magnetohydrodynamic equations can be written as follows
\begin{align}
\label{NS}
&\frac{\partial\boldsymbol{v}}{\partial t}=\boldsymbol{v}\times\boldsymbol{\omega}-\boldsymbol{\nabla}\mathcal{P}+\boldsymbol{J}\times\boldsymbol{B}+\nu\nabla^2\boldsymbol{v}-2\boldsymbol{\Omega}\times\boldsymbol{v}+\boldsymbol{r}\times\dot{\boldsymbol{\Omega}},\\
\label{induction}
&\frac{\partial\boldsymbol{B}}{\partial t}=\boldsymbol{\nabla}\times(\boldsymbol{v}\times\boldsymbol{B})+\eta\nabla^2\boldsymbol{B},\\
\label{div_v}
&\boldsymbol{\nabla}\cdot\boldsymbol{v}=0,\\
\label{div_B}
&\boldsymbol{\nabla}\cdot\boldsymbol{B}=0,
\end{align}
where $\boldsymbol{v}$, $\boldsymbol{\omega}$ and $\boldsymbol{B}$ are the velocity, vorticity and magnetic fields respectively. $\boldsymbol{J}=\boldsymbol{\nabla}\times\boldsymbol{B}$ is the current density, the total pressure is $\mathcal{P}$, and $\nu$ and $\eta$ are the kinematic viscosity and the magnetic diffusivity. ${\boldsymbol r}$ is the position vector, whereas ${\dot{\boldsymbol \Omega}}$ is the time derivative of the angular velocity. Note that the last term in the RHS of Eq. (\ref{NS}) correspond to the precession term. All quantities are expressed in dimensionless Alfvénic units.
The boundary conditions that we consider are that the normal components of $\boldsymbol{v}$, $\boldsymbol{B}$, $\boldsymbol{\omega}$ and $\boldsymbol{J}$ must all vanish in the surface of the spherical domain, $r=R=1$ (using the ratio of the sphere $R$ to normalize lengths). The vanishing normal component of $\boldsymbol{v}$ represents the fact that there is no mass flux across the surface, whereas the vanishing normal components of $\boldsymbol{B}$ and $\boldsymbol{J}$ can be regarded as modelling a thin layer of dielectric material that is coated on the outside by a perfect conductor. Finally, the condition that results in a vorticity field tangential to the surface is implied by, but does not imply, no-slip boundary conditions.
For the boundary conditions under consideration, the total energy balance is determined as
\begin{equation}
\frac{dE}{dt}= \int_V\boldsymbol{v}\cdot(\boldsymbol{ r}\times\dot{\boldsymbol \Omega})dV-2\nu Z+\int_V\eta|\boldsymbol{J}|^2dV+\int_S \nu(\boldsymbol{v}\times\boldsymbol{\omega})\cdot \hat{\boldsymbol n}dS ,
\label{energy_balance}
\end{equation}
where $\hat{\boldsymbol n}$ is an outward-pointing unit vector normal to the sphere's surface. This equation can be obtained from the dynamical Eqs. (\ref{NS})-(\ref{div_B})
after multiplying by the velocity $\boldsymbol{v}$ and the magnetic field $\boldsymbol{B}$ respectively and integrating in volume space to obtain the total energy
\begin{equation}
E = \frac{1}{2} \int_V \boldsymbol{v}^2 + \boldsymbol{B}^2 dV
\end{equation}
The right hand side of equation \ref{energy_balance} contains the injection energy $\epsilon$ given, by the precession term (first term containing the time derivative of the angular velocity), and dissipative terms like the fluid volume dissipation that involves the enstrophy $Z=1/2 \int_V|\boldsymbol{\omega}|^2dV$,
the magnetic field dissipation involving the current density. The surface term corresponds to the injection or dissipation of energy through the boundary. This energy balance equation will be the base for the
scaling models presented in the forthcoming sections.
We numerically integrate Eqs. (\ref{NS})-(\ref{div_B}) for the magnetohydrodynamic case and with null $\boldsymbol{B}$ for the purely hydrodynamic case using the \texttt{SPHERE} code\cite{SPHERE}. In both scenarios we expand $\boldsymbol{v}$ and $\boldsymbol{B}$ in terms of Chandrasekhar-Kendall (CK) functions which constitute a spectral basis, as reported in \cite{Mininni2006,Mininni2007}. These functions are the eigenfunctions of the curl with linear eigenvalue, so they obey the following expression:
\begin{equation}
\boldsymbol{\nabla}\times\boldsymbol{K}_i = k_i \boldsymbol{K}_i.
\label{eigenf_curl}
\end{equation}
Equation (\ref{eigenf_curl}) can be converted into a vector Helmholtz equation, for which three indices \textit{q}, \textit{l} and \textit{m} are needed to express all the solutions. The vector fields $\boldsymbol{K}_i$ can be succinctly expressed as
\begin{equation}
\boldsymbol{K}_{qlm}=k_{ql}(\boldsymbol{\nabla}\times\psi_{qlm}\hat{\boldsymbol{r}})+\boldsymbol{\nabla}\times(\boldsymbol{\nabla}\times\psi_{qlm}\hat{\boldsymbol{r}}),
\end{equation}
with $\psi_{qlm}$ a solution to the scalar Helmholtz equation. Considering that our spherical domain contains the origin, the collection $\psi_{qlm}$ is given by
\begin{equation}
\psi_{qlm}=C_{ql}j_l(|k_{ql}|r)\mathrm{Y}_l^m(\theta,\varphi).
\end{equation}
where $j_l$ is the spherical Bessel function of order $l$ and $Y^m_l$ is the spherical harmonic of degree $l$ and order $m$. $C_{ql}$ is a normalization constant which we adjust for the base components to be orthonormal to each other with respect to the usual inner product. The indices are limited as follows: $l$ and $m$ must obey usual rules for spherical harmonic indexing, that is $l>1$ and $-l<m<l$, whereas the index $q$ can be any integer except for $q=0$. Considering this, if the fields are decomposed using the CK basis, their eigenvalues are $k_{ql}$ and satisfy $k_{-ql}=k_{ql}$ and hence CK functions with opposing values of the index $q$ correspond to fields with opposing helicity. Making an analogy with a Fourier decomposition, the eigenvalues $k_{ql}$ can be thought of as an analog of the wavenumber. The velocity and magnetic field can be therefore expressed as
\begin{align}
\label{u_CK}
\boldsymbol{v}(\boldsymbol{r},t)&=\sum_{\substack{q=-\infty\\q\ne0}}^\infty\sum_{l=1}^\infty\sum_{m=-l}^l\xi_{qlm}^v(t)\boldsymbol{K}_{qlm}(\boldsymbol{r}),\\
\label{B_CK}
\boldsymbol{B}(\boldsymbol{r},t)&=\sum_{\substack{q=-\infty\\q\ne0}}^\infty\sum_{l=1}^\infty\sum_{m=-l}^l\xi_{qlm}^B(t)\boldsymbol{K}_{qlm}(\boldsymbol{r}).
\end{align}
The coefficients $\xi^v_i$ and $\xi^B_i$ only depend on time and to obtain its evolution we use Eqs. (\ref{NS}) and (\ref{induction}) but with the fields expanded in the CK basis, obtaining the following set of ordinary differential equations
\begin{align}
\label{vcoeff_evolution}
\frac{d\xi_n^u}{dt}&=\sum_{ij}k_jI_{ij}^n\left(\xi_i^u\xi_j^u-\xi_i^B\xi_j^B \right)+2\sum_i\xi_i^u\boldsymbol{\Omega}\cdot\boldsymbol{\mathcal{O}}_i^n-\nu k_n^2\xi_n^u+\mathcal{B}_n \delta_{l,1},\\
\label{Bcoeff_evolution}
\frac{d\xi_n^B}{dt}&=\sum_{ij}k_nI_{ij}^n\xi_i^u\xi_j^B-\eta k_n^2\xi_n^B.
\end{align}
Here, for notation clarity, $n$, $i$ and $j$ each represent a $(q,l,m)$, and $I_{ij}^n$ and $\boldsymbol{\mathcal{O}}_i^n$ are coupling arrays which obey the following expressions
\begin{align}
\label{I_ij^n}
I_{ij}^n &= \int_V\boldsymbol{K}_n^*\cdot(\boldsymbol{K}_i\times\boldsymbol{K}_j)dV,\\
\label{O_i^n}
\boldsymbol{\mathcal{O}}_i^n&=\int_V\boldsymbol{K}_n^*\times\boldsymbol{K}_idV.
\end{align}
On the other hand, the term that comes from the forcing given by the precession is defined as
\begin{equation}
\mathcal{B}_n=4\sqrt{\frac{\pi}{3}}C_{q,l}\operatorname{sg}(k_{q,l})j_l'(|k_{q,l}|)\left(\dot{\Omega}_z\delta_{m,0}-\frac{1}{\sqrt{2}}(\dot{\Omega}_x-i\dot{\Omega}_y)\delta_{m,1}\right)
\label{Poincare_force_CK}
\end{equation}
with $\operatorname{sg}$ the sign function and $\delta$ the Kronecker delta.
To solve Eqs. (\ref{vcoeff_evolution}) and (\ref{Bcoeff_evolution}) in a computer, a numerical resolution $q_\text{max}$ and $l_\text{max}$ has to be chosen. Given a fixed resolution, all the normalization constant $C_{ql}$, the eigenvalues $k_i$ and the coupling arrays $I_{ij}^n$ and $\boldsymbol{\mathcal{O}}_i^n$ are computed. These tables can then be stored and used for the computation of the time-dependent coefficients. Due to the high precision used in the spatial discretization (i.e. the CK basis) Eqs. (\ref{vcoeff_evolution}) and (\ref{Bcoeff_evolution}) are integrated in time employing a fourth-order Runge-Kutta scheme.
\subsection{\label{subsec:Simulations}Simulations performed}
We carried out 75 direct numerical simulations, varying different parameters, as explained below. All of the simulations used $q_\text{max}=l_\text{max}=7$, which implies that 882 expansion coefficients are evolved in time. The temporal resolution was $\Delta t=10^{-3}$. Half of the simulations were for the hydrodynamic case and the other half for the magnetohydrodynamic case.
The initial conditions for the HD case consisted of exciting only the following helical modes:
\begin{equation}
\xi^v_{qlm}\Big|_{t=0} = 0.3 \ \ \text{with}\
\begin{cases}
q = \pm 1, \pm 2\\
l = 1,2 \\
m = 1,2
\end{cases}
\label{initial_v}
\end{equation}
Here, the $0.3$ value was chosen \textit{ad hoc} seeking to approximately normalize the initial energy.
The angular speed $\Omega_0$, the precession frequency $\gamma$ and the kinematic viscosity $\nu$ were varied in each simulation. Table \ref{table_HD} contains the simulations that we show in the figures in Section \ref{sec:Results} with their corresponding names (ID) and relevant adimensional parameters. The different $\nu$ values considered were the following ones: $\nu = 0.01$, $\nu = 0.03$, $\nu = 0.06$ and finally $\nu = 0.1$. It should be noted that not all the simulations are contained in the table, for better understanding only those shown in the figures are included. Other parameters that appear in the table are the initial Reynolds number ($Re_0$), the final Reynolds number ($Re_f$), the Ekman number ($Ek$), the final Rossby number ($Ro_f$) and the Poincaré number ($Po$). All of them were calculated considering the sphere radius as the length scale. Those characteristic numbers differ according to the moment of the run in which the typical velocity is computed. The final velocity was determined as the time average of the flow speed in the time window where the kinetic energy is steady. The parameters can be calculated as:
\begin{equation} \label{adim_param}
\begin{aligned}
Re_f &= \frac{v_f R}{\nu}, \\
Ek &= \frac{\nu}{R^2 \Omega_0}, \\
Ro_f &= \frac{v_f}{R \Omega_0}, \\
Po &= \frac{|\gamma|}{\Omega_0}.
\end{aligned}
\end{equation}
The table also includes $Re_0$, which is defined similarly to $Re_f$ but changing the final velocity $v_f$ for the initial velocity $v_0$. As indicated $Po$ is the ratio between the precession frequency and the angular velocity.
\begin{table}[H]
\begin{center}
\begin{tabular*}{.9\textwidth}{ c @{\extracolsep{\fill}} c c c c c c c }
\toprule
\toprule
ID & $\Omega_0$ & $\gamma$ & $Po$ & $Re_0$ & $Re_f$ & $Ek$ & $Ro_f$\\ \hline
\midrule
HD01 & $1$ & $0.1$ & $1.0\times10^{-1}$ & $1.0\times10^{2}$ & $1.2\times10^{0}$ & $1.0\times10^{-2}$ & $1.2\times10^{-2}$ \\
HD02 & $4$ & $0.1$ & $2.5\times10^{-2}$& $1.0\times10^{2}$ & $2.0\times10^{0}$ & $2.5\times10^{-3}$ & $5.0\times10^{-3}$ \\
HD03 & $10$ & $0.1$ & $1.0\times10^{-2}$ & $1.0\times10^{2}$ & $2.1\times10^{0}$ & $1.0\times10^{-3}$ & $2.1\times10^{-3}$\\
HD04 & $1$ & $1$ & $1.0\times10^{0}$ & $1.0\times10^{2}$ & $9.7\times10^{0}$ & $1.0\times10^{-2}$ & $9.7\times10^{-2}$\\
HD05 & $4$ & $1$ & $2.5\times10^{-1}$ & $1.0\times10^{2}$ & $1.7\times10^{1}$ & $2.5\times10^{-3}$ & $4.2\times10^{-2}$\\
HD06 & $10$ & $1$ & $1.0\times10^{-1}$ & $1.0\times10^{2}$ & $2.0\times10^{1}$ & $1.0\times10^{-3}$ & $2.0\times10^{-2}$\\
HD07 & $1$ & $10$ & $1.0\times10^{1}$ & $1.0\times10^{2}$ & $1.8\times10^{1}$ & $1.0\times10^{-2}$ & $1.8\times10^{-1}$\\
HD08 & $4$ & $10$ & $2.5\times10^{0}$ & $1.0\times10^{2}$ & $5.7\times10^{1}$ & $2.5\times10^{-3}$ & $1.4\times10^{-1}$\\
HD09 & $10$ & $10$ & $1.0\times10^{0}$ & $1.0\times10^{2}$ & $1.1\times10^{2}$ & $1.0\times10^{-3}$ & $1.1\times10^{-1}$\\
HD10 & $1$ & $0.1$ & $1.0\times10^{-1}$ & $3.5\times10^{1}$ & $4.1\times10^{-1}$ & $3.0\times10^{-2}$ & $1.2\times10^{-2}$ \\
HD11 & $4$ & $0.1$ & $2.5\times10^{-2}$ & $3.5\times10^{1}$ & $6.2\times10^{-1}$ & $7.5\times10^{-3}$ & $4.6\times10^{-3}$\\
HD12 & $10$ & $0.1$ & $1.0\times10^{-2}$ & $3.5\times10^{1}$ & $6.7\times10^{-1}$ & $3.0\times10^{-3}$ & $2.0\times10^{-3}$\\
HD13 & $1$ & $1$ & $1.0\times10^{0}$ & $3.5\times10^{1}$ & $2.8\times10^{0}$ & $3.0\times10^{-2}$ & $8.4\times10^{-2}$\\
HD14 & $4$ & $1$ & $2.5\times10^{-1}$ & $3.5\times10^{1}$ & $5.2\times10^{0}$ & $7.5\times10^{-3}$ & $3.9\times10^{-2}$\\
HD15 & $10$ & $1$ & $1.0\times10^{-1}$ & $3.5\times10^{1}$ & $6.2\times10^{0}$ & $3.0\times10^{-3}$ & $1.9\times10^{-2}$\\
HD16 & $1$ & $10$ & $1.0\times10^{1}$ & $3.5\times10^{1}$ & $5.8\times10^{0}$ & $3.0\times10^{-2}$ & $1.7\times10^{-1}$\\
HD17 & $4$ & $10$ & $2.5\times10^{0}$ & $3.5\times10^{1}$ & $1.9\times10^{1}$ & $7.5\times10^{-3}$ & $1.4\times10^{-1}$\\
HD18 & $10$ & $10$ & $1.0\times10^{0}$ & $3.5\times10^{1}$ & $3.5\times10^{1}$ & $3.0\times10^{-3}$ & $1.0\times10^{-1}$\\
\bottomrule
\bottomrule
\end{tabular*}
\caption{Runs corresponding to the HD case with their respective names (ID) and relevant adimensional parameters. The difference between the two Reynolds numbers is the time in which each one was calculated, the initial $Re_0$ and the final $Re_f$. The characteristic length considered to calculate all adimensional parameters was the sphere radius, and the initial velocity is of the order of unity.}
\label{table_HD}
\end{center}
\end{table}
On the other hand the MHD simulations were initially excited in the following modes for the velocity field
\begin{equation}
\xi^v_{qlm}\Big|_{t=0} = 0.5 \ \ \text{with $q=3$}\
\begin{cases}
\text{if}\ l=1,2 \Rightarrow m=0,1 \\
\text{if}\ l=3 \Rightarrow m\in [0,3]
\end{cases}
\label{initial_v_MHD}
\end{equation}
and for the magnetic field:
\begin{equation} \label{initial_B}
\begin{aligned}
\xi_{110}^B\Big|_{t=0}&=0.5, \\
\xi_{111}^B\Big|_{t=0}&=0.5 (1-i),
\end{aligned}
\end{equation}
in order to have fields with net initial helicity, a feature known to favor dynamo action. The corresponding Table \ref{table_MHD} shows the different runs for the MHD case. We take the magnetic diffusivity $\eta$ equal to the kinematic viscosity $\nu$ so that a unit magnetic Prandtl number $P_m$ is prescribed. In contrast to the HD case, only the values of $\Omega_0$ and $\gamma$ were varied. In this case we also considered negative values of the precession frequency, an important fact that we discuss in Section \ref{sec:Results}.
\begin{table}[H]
\begin{center}
\begin{tabular*}{.9\textwidth}{ c @{\extracolsep{\fill}} c c c c c c c c }
\toprule
\toprule
ID & $\Omega_0$ & $\gamma$ & $Po$ & $Re_0$ & $Re_f$ & $Ek$ & $Ro_f$\\ \hline
\midrule
MHD01 & 10 & 0.1 & $1.0\times10^{-2}$ & $1.3\times10^{3}$ & $2.9\times10^{2}$ & $1.0\times10^{-4}$ & $2.9\times10^{-2}$ \\
MHD02 & 10 & 1 & $1.0\times10^{-1}$& $1.3\times10^{3}$ & $3.7\times10^{2}$ & $1.0\times10^{-4}$ & $3.7\times10^{-2}$ \\
MHD03 & 10 & 10 & $1.0\times10^{0}$ & $1.3\times10^{3}$ & $1.24\times10^{3}$ & $1.0\times10^{-4}$ & $1.24\times10^{-1}$ \\
MHD04 & 1 & -1 & $1.0\times10^{0}$ & $1.3\times10^{3}$ & $6.3\times10^{2}$ & $1.0\times10^{-3}$ & $6.3\times10^{-1}$ \\
MHD05 & 1 & -3 & $3.0\times10^{0}$ & $1.3\times10^{3}$ & $4.2\times10^{2}$ & $1.0\times10^{-3}$ & $4.2\times10^{-1}$ \\
MHD06 & 1 & -5 & $5.0\times10^{0}$ & $1.3\times10^{3}$ & $3.85\times10^{2}$ & $1.0\times10^{-3}$ & $3.85\times10^{-1}$ \\
MHD07 & 8 & -1 & $1.25\times10^{-1}$ & $1.3\times10^{3}$ & $4.1\times10^{2}$ & $1.25\times10^{-4}$ & $5.1\times10^{-2}$\\
MHD08 & 8 & -3 & $3.75\times10^{-1}$ & $1.3\times10^{3}$ & $1.3\times10^{3}$ & $1.25\times10^{-4}$ & $1.6\times10^{-1}$ \\
MHD09 & 8 & -3.5 & $4.4\times10^{-1}$ & $1.3\times10^{3}$ & $1.6\times10^{3}$ & $1.25\times10^{-4}$ & $2.0\times10^{-1}$ \\
MHD10 & 8 & -4 & $5.0\times10^{-1}$ & $1.3\times10^{3}$ & $1.9\times10^{3}$ & $1.25\times10^{-4}$ & $2.4\times10^{-1}$\\
MHD11 & 8 & -4.5 & $5.6\times10^{-1}$ & $1.3\times10^{3}$ & $2.3\times10^{3}$ & $1.25\times10^{-4}$ & $2.9\times10^{-1}$\\
MHD12 & 8 & -5 & $6.25\times10^{-1}$ & $1.3\times10^{3}$ & $2.8\times10^{3}$ & $1.25\times10^{-4}$ & $3.5\times10^{-1}$\\
MHD13 & 16 & -1 & $6.2\times10^{-2}$ & $1.3\times10^{3}$ & $4.2\times10^{2}$ & $6.25\times10^{-5}$ & $2.6\times10^{-2}$\\
MHD14 & 16 & -3 & $1.9\times10^{-1}$ & $1.3\times10^{3}$ &$1.1\times10^{3}$ & $6.25\times10^{-5}$ & $6.85\times10^{-2}$\\
MHD15 & 16 & -3.5 & $2.2\times10^{-1}$ & $1.3\times10^{3}$ & $1.3\times10^{3}$ & $6.25\times10^{-5}$ & $8.35\times10^{-2}$\\
MHD16 & 16 & -4 & $2.5\times10^{-1}$ & $1.3\times10^{3}$ & $1.6\times10^{3}$ & $6.25\times10^{-5}$ & $1.0\times10^{-1}$\\
MHD17 & 16 & -4.5 & $2.8\times10^{-1}$ & $1.3\times10^{3}$ & $1.9\times10^{3}$ & $6.25\times10^{-5}$ & $1.2\times10^{-1}$\\
MHD18 & 16 & -5 & $3.1\times10^{-1}$ & $1..3\times10^{3}$ & $2.2\times10^{3}$ & $6.25\times10^{-5}$ & $1.4\times10^{-1}$\\
\bottomrule
\bottomrule
\end{tabular*}
\caption{Information of the MHD simulations with their name ID and adimensional parameters. The viscosity is the same for all runs and, in consequence, the initial Reynolds number too.}
\label{table_MHD}
\end{center}
\end{table}
\section{\label{sec:Results}Results}
\subsection{\label{subsec:Results_HD}HD Results}
We first report the results for the purely hydrodynamic case (i.e., no magnetic field).
We performed a parametric study changing three of the parameters of the system: the rotation rate $\Omega_0$, the precession frequency $\gamma$ and the kinematic viscous coefficient $\nu$.
Since the torque given by the precession is forcing the system, whereas viscosity is acting as a dissipation force, it is natural to expect that at certain time the energy and enstrophy have a statistically stabilization value. This is evidenced in Figure \ref{ene_enst_HD_prec} where for the two quantities a transitory regime followed by a stabilization is observed for our simulations. The transient initial decaying regime observed here is similar to previous results for the case without precession \cite{Mininni2007}. On the other hand, the presence of precession gives an asymptotic value for both the energy $\int_V \boldsymbol u^2 dV$ and enstrophy $\int_V \boldsymbol \omega^2 dV$. This asymptotic value is different for each simulation, increasing when either $\Omega_0$ or $\gamma$ are increased (i.e. more energy injected) and decreasing for larger values of $\nu$ (i.e. more energy dissipated).
In an attempt to model this asymptotic scaling we argue the following: looking at the energy equation \ref{energy_balance} without the current density that appears in the MHD case, we consider a scaling for the dissipation rate
$D\sim \nu U^2/l^2$ where we chose $U=\langle v \rangle$ as a characteristic velocity of the system, and with the Taylor scale as the characteristic length $l\sim\sqrt{\nu R/U}$. This last expression is the result of considering the sphere radius ($R=1$) as the injection length scale in the expression $l \sim Re^{-1/2} R$.
In the scaling for the dissipation rate we neglect any effect of the boundary term.
On the other hand we take a scaling for the energy injection rate as coming uniquely from the precession term,
$\epsilon \sim R U \gamma \Omega_0$.
Assuming a balance between the dissipation and the injection rate in the
statistically stationary state $D\sim\epsilon$
it follows then that the energy scales as $U^2\sim\gamma\Omega_0$.
The proposed scaling is evaluated in Figure \ref{HD_scaling} where we show the mean kinetic energy versus the corresponding $\gamma\Omega_0$ value for that run and compare with the model (represented by a linear behavior in this plot). It can be seen that the results obtained in most of the simulations are consistent with the scaling proposed.
As a final diagnosis for the pure hydrodynamic case in Figure \ref{spectra_HD} we show the kinetic energy spectra for different runs varying $\Omega_0$. It can be seen that the
spectra follow a Kolmogorov-like behavior ($\sim k^{-5/3}$) at the large scales, which is consistent with the development of stationary turbulence, although the range of scales is very limited.
\begin{figure}[H]
\centering
\includegraphics[scale=0.9]{fig1.pdf}
\caption{Energy $E_k$ (left) and averaged enstrophy $\langle \omega^2 \rangle$ (right) as a function of time for the runs HD04, HD05 and HD06.}
\label{ene_enst_HD_prec}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.9]{fig2.pdf}
\caption{Mean kinetic energy $\langle E_k \rangle$ of all the runs in Table \ref{table_HD} as a function of $\gamma \Omega_0$. The dashed line corresponds to the scaling of the proposed model.}
\label{HD_scaling}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.9]{fig3.pdf}
\caption{Kinetic power spectra $E$ as a function of the CK wavenumber $k$ for different runs wih varying $\Omega_0$.
The reference line corresponds to the Kolmogorov spectrum $k^{-5/3}$.}
\label{spectra_HD}
\end{figure}
\subsection{\label{subsec:Results_MHD}MHD Results}
We now consider adding a initial magnetic field to act as a seed, in order to study the feasibility of dynamo action, as well as the existence
of magnetic dipole reversals. The forcing is again given by the precession term and we vary the parameters, in this case $\Omega_0$ and
the precession frequency $\gamma$ in order to study the different behavior of the system for each case.
In Figure \ref{energies_MHD_prograde} we show the results for the kinetic and magnetic energy as a function of time for three different runs, labeled MHD01, MHD02 and MHD03 which differ in the
value of the precession frequency $\gamma=0.1, 1, 10$, respectively, and with the same value for the
rotating frequency $\Omega_0=10$. As it can be observed, the magnetic energy is not sustained, decaying to very low values, except for the high $\gamma$ case, where nevertheless a final
stationary value lower than the initial value is reached.
To explore a different regime we consider now the possibility of retrograde precession, that is, negative values of $\gamma$. The case of retrograde precession
is particularly interesting, as it has been
experimentally studied in several works related to the Earth's liquid core \cite{Vanyo1995,Pais2001,Noir2003,stewartson1963}.
The results for different runs with retrograde precession are shown in Fig. \ref{energy_retrograde_MHD}.
The leftmost panel, where the magnetic energy is not sustained, corresponds to the case with $\Omega_0=1$.
On the other hand, in the middle panel with $\Omega_0=8$
a dynamo action is observed, for the larger values (in absolute value) of $\gamma=-3, -5$.
A similar result is obtained for the case with $\Omega_0=16$ in the rightmost panel.
A case of clear magnetic field generation (i.e. dynamo action) is observed for the $\Omega_0 > 1$ cases and for the
largest values of (negative) $\gamma$. It seems that a critical value of $\gamma=\gamma_c=-3$ exists, beyond which the magnetic energy is mantained at a stable level, i.e, a self-sustaining dynamo is attained.
\begin{figure}[H]
\centering
\includegraphics[scale=0.8]{fig4.pdf}
\caption{Magnetic and kinetic energy as a function of time for MHD01 (left), MHD02 (middle) and MHD03 (right) which differ in the precession frequency, $\gamma=0.1, 1, 10$, respectively, and have the same value of $\Omega_0=10$. It can be observed that these runs do not generate dynamos, except for very high values of $\Omega_0$ and $\gamma$ as is the case for
the MHD03 run (rightmost panel).}
\label{energies_MHD_prograde}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.8]{fig5.pdf}
\caption{Comparison of the time evolution of the magnetic energy for different $\gamma$ values. Each panel is distinguished by their corresponding value of $\Omega_0$. On the left with $\Omega_0=1$ are the simulations MHD04, MHD05, MHD06, on the middle with $\Omega_0=8$ are MHD07, MHD08 and MHD12, and finaly on the right with $\Omega_0=16$ are the runs MHD13, MHD14 and MHD18.}
\label{energy_retrograde_MHD}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.8]{fig6.pdf}
\caption{Total energy as a function of the quantity $\epsilon/\sqrt{\gamma \Omega_0}$ for different runs (MHD08-MHD12 and MHD14-MHD18). The proposed scaling is indicated with the dashed line.}
\label{scaling_energy_MHD}
\end{figure}
In order to obtain a scaling for the total energy (kinetic plus magnetic) we proceed in
a similar way as was presented in the hydrodynamic results subsection.
We assume in this case the injection rate is the same as the energy transfer rate (in a statistically stationary state), so $\epsilon \sim (U^3+B^2U)/l \sim U^3/l$, taking
$U\sim B$ as the results of the runs seems to indicate. Here $l$ is a characteristic
length which can be linked with the injection rate and the rotation parameters from the scaling $\epsilon \sim l \Omega_0 \gamma$ so
we obtain $l \sim \epsilon / (\Omega_0 \gamma )$.
Replacing this scaling for $l$ in the expression
$\epsilon \sim U^3/l$
it follows then that the total energy scales as
$E = U^2+B^2 \sim 2 U^2 \sim \epsilon / \sqrt{\gamma \Omega_0}$.
This scaling is analyzed in Fig. \ref{scaling_energy_MHD}, where the final average energy $E$ is plotted against $\epsilon/ \sqrt{\gamma \Omega_0}$ for several runs. The straight line corresponding to the perfect scaling is shown as a reference and it is notable that for this case a turbulent scaling reproduces satisfactorily the behavior.
\begin{figure}[H]
\centering
\includegraphics[scale=0.8]{fig7.pdf}
\caption{Magnetic energy spatial spectra for runs with $\Omega_0=8$ (MHD08-MHD12) on the left and with $\Omega_0=16$ (MHD14-MHD18) on the right, all of them are for the time $t=150$. It can be observed a predominance in the high values of $k$ for all of the runs. This suggests that these are small scale dynamos.}
\label{MHD_spectra}
\end{figure}
Another type of diagnosis is presented in Fig. \ref{MHD_spectra}, which shows the
magnetic energy spectra for different runs, all with sustained magnetic energy (i.e. self-sustaining dynamos).
As can be seen the magnetic energy seems to be dominated by the larger values of $k$,
meaning the dynamos are in a small scale regime.
An interesting phenomenon are the reversals of the magnetic dipole moment component parallel to the mean rotation axis, $m_z$. We show the behavior of the normalized value $m_z/|m|$ vs time in Figure \ref{MHD_reversals} for a short time
window (to better appreciate the dynamics) for a set of runs with $\Omega_0=8$
and different values of $\gamma=-3, -3.5, -4, -4.5, -5$, from top to bottom.
The plots reveal the existence of reversals in all these dynamo runs and it can also be seen
that the dynamics seems to be faster for the larger values of $|\gamma|$. The right panel
of Fig. \ref{MHD_reversals} shows the normalized histograms corresponding to the distribution of times between reversals ({\it waiting times}). The histograms are consistent with
the fact that the reversals occurs faster when the value of the precession frequency $|\gamma|$ is greater, because there is a greater domain of long waiting times for cases with smaller values of $|\gamma|$.
\begin{figure}[H]
\centering
\includegraphics[scale=0.9]{fig8.pdf}
\caption{Time evolution of the normalized $z$ component of the magnetic dipole moment
$m_z/|m|$ for a time window between t=100 and t=120, for all the dynamo runs with $\Omega_0=8$ (left) and values of $\gamma=-3, -3.5, -4, -4.5, -5$, from top to bottom. Normalized histograms of the time $\tau$ between reversals (waiting times) are shown on the right. The simulations that are shown in this figure are MHD08-MHD12.}
\label{MHD_reversals}
\end{figure}
\section{Conclusions}\label{conclusions}
In the scenario of a rotating and precessive sphere filled with fluid, the incompressible HD and MHD equations were studied by direct numerical simulations. For both cases, a Galerkin spectral code was used. This method decomposes the fields in orthogonal Chandrasekhar-Kendall eigenfunctions. Being purely spectral, this numerical technique allowed us to integrate the system with high accuracy.
In a purely hidrodynamic setting, we found that the kinetic energy and enstrophy present a transitory regime that behaves like the non-precessional case, after which a steady state is reached. This regime was dominated by dissipation even though the spectrum present a power law of $k^{-5/3}$ which corresponds to the development of turbulence. We presented a model for the scaling of the kinetic energy with
the parameters of the system which showed very good agreement with the results from the simulations.
For the MHD case, we separated the study into in two scenarios, by taking into account the direction of precession rotation. For the case of prograde precession we could generate dynamos only with high values of $\Omega_0$ and $\gamma$. On the other hand, using retrograde precession if the $\Omega_0$ value is large ($\Omega_0 = 8$ or $\Omega_0 = 16$) we showed that there is a critical value of the precession frequency in order to attain a stabilization in the magnetic energy. This critical frequency is around $\gamma_c=-3$ for either of the two angular velocity amplitude $\Omega_0$ considered, and flows operating at $\gamma < \gamma_c$ showcased self-sustaining dynamos in both cases. For low rotation amplitude, $\Omega_0=1$, a monotonic decay of the magnetic energy was observed in the region of parameter space explored.
For the MHD cases which presented sustained dynamo action, a scaling relation for the total energy in the steady state was proposed using the parameters of the problem.
The proposed scaling showed a very good agreement with the results as well.
Finally we studied the behaviour of the magnetic field generated. We found that all dynamos present dominance of the small scales in the spectra of the magnetic energy, this suggest that the simulations operate in the regime of small scale dynamos. Another interesting feature of all dynamos with a preferential direction is the presence of multiple dipole moment reversals. Moreover, doing a statistical analysis we observed slower dynamics in the dipole moment when the precession is also slower, i.e., when the absolute value of the frequency precession ($|\gamma|$) is smaller.
The presented results constitutes a first numerical study of the influence of the precession in
the dynamics of HD and MHD in a rotating sphere filled with a fluid or magnetofluid. We believe that these results show interesting and novel features of rotating fluids with precession which deserve further study in the future.
\begin{acknowledgments}
The authors acknowledge support from CONICET and ANPCyT through PIP, Argentina Grant No. 11220150100324CO, and PICT, Argentina Grant No. 2018-4298.
\end{acknowledgments}
|
\section*{References}}
\begin{document}
\title{Tunable directional photon scattering from a pair of superconducting qubits}
\author{Elena S. Redchenko}
\email{<EMAIL>}
\affiliation{Institute of Science and Technology Austria, 3400 Klosterneuburg, Austria}
\author{Alexander V. Poshakinskiy}
\affiliation{Ioffe Institute, St.~Petersburg 194021, Russia}
\author{Riya Sett}
\affiliation{Institute of Science and Technology Austria, 3400 Klosterneuburg, Austria}
\author{Martin Zemlicka}
\affiliation{Institute of Science and Technology Austria, 3400 Klosterneuburg, Austria}
\author{Alexander N. Poddubny}
\affiliation{Ioffe Institute, St.~Petersburg 194021, Russia}
\author{Johannes M.~Fink}
\email{<EMAIL>}
\affiliation{Institute of Science and Technology Austria, 3400 Klosterneuburg, Austria}
\date{\today}
\begin{abstract}
The ability to control the direction of scattered light
is crucial to provide the flexibility and scalability for a wide range of on-chip applications, such as integrated photonics, quantum information processing and nonlinear optics.
In the optical and microwave frequency ranges tunable directionality can be achieved by applying external magnetic fields that modify optical selection rules~\cite{Sllner2015,Spitzer2018}, by using nonlinear effects~\cite{hamann2018nonreciprocity}, or interactions with vibrations~\cite{poulton2012design, kittlaus2017chip, safavi2019controlling}. However, these approaches are less suitable to control propagation of microwave photons inside integrated superconducting quantum devices~\cite{Reuer2022}, that is highly desirable.
Here, we demonstrate on-demand tunable directional scattering based on two periodically modulated transmon qubits coupled to a transmission line at a fixed distance \cite{astafiev2010resonance,vanLoo1494,Kannan2022} in close analogy to two oscillating mirrors \cite{poshakinskiy2019optomechanical,mirhosseini2019cavity}. By continuously changing the
relative phase
between the local modulation tones,
we realize unidirectional forward or backward photon scattering. Such an in-situ switchable mirror represents a versatile tool for intra- and inter-chip microwave photonic processors. In the future, a lattice of qubits can be used to realize topological circuits that exhibit stronger nonreciprocity or chirality \cite{Tomoki2019}.
\end{abstract}
\maketitle
One of the simplest ways to realize directional light scattering relies on the Kerker effect~\cite{kerker1983electromagnetic,Liu2018}. It is based on the interference between different multipole components of scattered light, for example electric and magnetic dipoles, and has been demonstrated for Si nanoparticles~\cite{Staude2013,fu2013directional,person2013demonstration}. However, the nanoparticle scattering pattern is fixed after fabrication and dictated by its shape. Tunable light routing is typically enabled by an external magnetic field that leads to the Zeeman splitting of optical transitions for clockwise- and counter-clockwise- propagating photons~\cite{Sllner2015} or a modification of optical selection rules~\cite{Spitzer2018}. In the optical domain, the routing can be reversed as well without changing the magnetic field by flipping the spin of the atom \cite{scheucher2016quantum}. Such structures are now actively studied in the domain of chiral quantum optics~\cite{lodahl2017chiral,Prasad2020}. Tunable directional scattering can also be achieved by using moving boundary conditions~\cite{poshakinskiy2019optomechanical,zhang2018thermal}.
For example, the trembling of a small particle with only an electric dipole resonance can induce a magnetic dipole resonance~\cite{poshakinskiy2019optomechanical}, which in turn provides directional scattering in analogy to the Kerker effect. Several compact nonmagnetic realizations of nonreciprocal devices using Raman and Brilliouin scattering \cite{muller2010measurement, poulton2012design, kittlaus2018non,safavi2019controlling} have been presented for optical frequencies.
Tunable directional interactions are also highly desired for superconducting quantum circuits in the microwave spectral range. For instance, isolators and circulators are commonly used for superconducting quantum computing to protect fragile qubits states. Cascaded photon processing in the chiral setup is also beneficial for the creation of complex entangled quantum states of qubits~\cite{Guimond2020}. However, it is difficult to directly extend existing approaches for visible light to microwave photons.
For example the classical Kerker approach is not applicable to a typical transmon qubit that behaves just as an electric dipole~\cite{astafiev2010resonance}, without magnetic dipole resonances.
Devices, based on an external magnetic field,~\cite{mahoney2017chip} are often bulky and always require additional shielding to protect superconducting qubits. While there exist demonstrations of directionality in parametrically driven, compact mechanical systems \cite{peterson2017demonstration,bernier2017,barzanjeh2017mechanical}, integration with superconducting circuitry is challenging due to limited bandwidth and tunability. Thus, there is a need for flexible to use on-chip microwave photon routers which do not require strong magnetic fields or moving mechanical parts.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figure1_v2.pdf}
\caption{\textbf{Experimental realization. a,} Schematic showing the scattering direction of the $\omega\pm\Omega$ component for in-phase (up-up) and out-of-phase (up-down) modulation of the qubits' transition frequencies $\omega$. \textbf{b,} Optical microscope image and simplified experimental setup. Two transmon qubits are capacitively coupled to a 50\,$\Omega$ transmission line, and each qubit has a local flux bias line connected to an arbitrary waveform generator channel (AWG), which is used to generate a sinusoidal wave with an amplitude $A_V$ that is filtered with a 80 MHz low pass filter and applied to ground via a 1\,k$\Omega$ resistor. We use an RF source, analog downconversion
and digitization (DIG) to back out the scattering parameters of the device cooled to 10\,mK.
\textbf{c,} Enlarged view of Qubit 2 and local flux bias line inductively coupled to the qubit SQUID. \textbf{d (e),} Individually measured and normalized transmission spectra $|t_0|^2$ of elastically scattered radiation from Qubit 1(2) with fit to theory (solid line).} \label{fig1}
\end{figure}
Here, our goal is to demonstrate an easy-to-fabricate circuit providing frequency and directionality tunable photon scattering with the minimum number of components required.
Our approach is based on the sinusoidal time-modulation of the qubit frequency \cite{chapman2017widely, rosenthal2017breaking, Shadrivov2018}, which is a standard technique in circuit and waveguide QED. The modulated qubit strongly coupled to a waveguide can be mapped onto the problem of light scattering from the trembling mirror \cite{jackel1992fluctuations,poshakinskiy2019optomechanical}.
By altering the relative phase $\alpha$ between the modulation tones of two qubits, we change the effective phase shift between the scattered sidebands resulting in different interference patterns for forward and backward scattering as schematically shown in Fig.~\ref{fig1}a.
Here we do not focus on the elastic scattering nonreciprocity~\cite{chapman2017widely, rosenthal2017breaking} or directional emission from the initial qubit state~\cite{oliver2020b,Gheeraert2020} but on the switching between forward and backward inelastic coherent scattering. Thus, although elastically (Rayleigh) scattered radiation remains almost unaffected, we gain the flexibility to choose the frequency of the scattered photons.
\vspace{0.25cm}
\noindent
\section*{Experimental implementation}
We fabricate the sample with two transmon qubits coupled to a 1D coplanar transmission line separated by $d=5$\,mm as shown in Fig.~\ref{fig1}b. The maximum frequency of the $|0\rangle \rightarrow |1\rangle$ transition is $9.129$ ($9.577$) GHz for Qubit 1(2). We tune both qubits to $\omega_0/(2\pi) = 6.129$ GHz corresponding to an effective distance of $d = \lambda/4$, with $ \lambda$ the wavelength of photons at $\omega_0$, using bias coils mounted on top of the sample box. Currents for the periodic frequency modulation are applied via on-chip bias lines inductively coupled to the SQUID loops as shown in Fig. \ref{fig1}c. Both ports of the transmission line are connected to separate microwave in- and output lines to measure reflection and transmission spectra simultaneously.
Firstly, we characterize the qubits individually at $\omega_0$ where $d=\lambda/4$ using a weak resonant probe tone and measuring the coherently and elastically scattered radiation, i.e. at the same frequency. We determine the normalized transmission spectrum of each qubit shown in Fig.~\ref{fig1}d(e) and find the radiative decay rates to be $\Gamma_{1} / (2\pi) \approx 4.4$ MHz
with the dephasing rates of $\Gamma_{2} / (2\pi) \approx 3.9$ $(4.3)$ MHz for Qubit 1 (2) which is dominated by flux noise due to the relatively high flux dispersion at this bias point.
An applied sinusoidal bias current makes the qubit resonance frequency tremble in time and the coherent
transmission amplitude
is then given by
\begin{eqnarray}\label{coh_oneQ}
&& t_0 = 1 + \sum_{n=-\infty}^{\infty}{\frac{i\Gamma_{1}/2}{\omega_0+n\Omega-\omega-i\Gamma_{2}}J_{n}^2\left( \frac{A_m}{\Omega}\right)},
\end{eqnarray}
where $J_{n}\left( \frac{A_m}{\Omega}\right)$ are Bessel functions of the first kind, $A_m$ is the modulation amplitude in frequency units, and $\Omega$ is the modulation frequency.
We measure the normalized transmission spectrum $|t_0|^2$ as a function of modulation frequency $\Omega$ as shown in Fig.~\ref{fig2}a. For the fixed signal amplitude at the AWG output $A_V = 50$\,mV$_\text{pp}$, the system undergoes a transition from the strong ($A_m/2>\Omega$) to the weak ($A_m/2<\Omega$) modulation regime. We fit similarly measured data to Eq.~\ref{coh_oneQ} for both qubits individually, as shown in Fig.~\ref{fig2}b for different applied $A_V$ and $\Omega/ (2\pi)=20$\,MHz and find that the fitted $A_m$ scales approximately linearly with $A_{V}$, which is shown in the Supplementary Information.
\begin{figure*}[t]
\centering
\includegraphics[width=1.5\columnwidth]{figure2_v2.pdf}
\caption{\textbf{Singe qubit properties. a}, Normalized transmission spectrum $|t_0|^2$ of qubit 1 measured as a function of the modulation frequency $\Omega$ and the probe frequency $\omega$ at the fixed modulation amplitude $A_V = 50$\,mV$_\text{pp}$. \textbf{b}, Measured $|t_0|^2$ of the modulated qubit with $\Omega/(2\pi)=20$\,MHz for different $A_m$ and fits to Eq.~\ref{coh_oneQ} (solid lines). \textbf{c}, Measured resonance fluorescence emission spectrum of qubit 1 as a function of the modulation frequency $\Omega$ and detuning of the detected inelastically scattered radiation from the drive applied at $\omega_0$ for a Rabi frequency $\Omega_\text{R}/(2\pi)=52$\,MHz and modulation amplitude $A_m=0.2\,\Omega_\text{R}$. \textbf{d}, Theoretically predicted Mollow spectrum in the presence of frequency modulation for the same parameters. \textbf{e}, Level splitting schematics of the dressed and modulated qubit, which explains the origin of the observed nested Mollow triplets at $\Omega=\Omega_R$.
} \label{fig2}
\end{figure*}
\vspace{0.25cm}
\noindent
\section*{Modulated Mollow resonance fluorescence}
One of the hallmark characteristics of quantum two-level systems is the observation of the incoherent resonance fluorescence spectrum taking the form of a Mollow triplet for an applied resonant drive of sufficient power $\Omega_{\rm R}> \Gamma_{1}$~\cite{astafiev2010resonance}. Here we observe this effect for a frequency modulated qubit with $A_m = 0.2\,\Omega_{\text R}$ and Rabi frequency $\Omega_{\text R}/(2\pi) = 52$ MHz. The measured power spectral density (PSD) as a function of the modulation frequency $\Omega$
is shown in Fig.~\ref{fig2}c and the corresponding theory in Fig.~\ref{fig2}d. Dressing with the drive leads to the well known emission spectrum with three maxima at $\omega_0$ and $\omega_0\pm \Omega_R$. However, the additional frequency modulation leads to
the formation of avoided crossings at $\Omega=\Omega_R$,
which can be qualitatively interpreted as a formation of nested Mollow triplets following the level scheme shown in Fig.~\ref{fig2}e. Specifically, each of the levels of the original Mollow triplet is split into two levels due to the modulation. Next, the photon transitions between the split levels lead to the formation of additional Mollow triplets. For example, the transition from the original triplet having the largest energy, and shown by the thick vertical blue arrow, is transformed by the modulation into three distinct transition energies shown by the thin blue lines. The observed splitting between the outermost transitions of the inner Mollow triplets for $\Omega=\Omega_R$ is equal to $\Delta\omega/(2\pi)\approx 20$\,MHz, in excellent agreement with the numerical calculation.
Similar formations of nested Mollow triplets in the electron spin-noise spectrum have been predicted for the conditions of electron paramagnetic resonance when the electron is subject to a the time-modulated magnetic field \cite{poshakinskiy2020spin}, but have so far not been observed directly to the best of our knowledge.
\vspace{0.25cm}
\noindent
\section*{Directional scattering}
Now we consider the system of two qubits both tuned to $\omega_0$ and located at a distance $\lambda/4$. For any odd multiple of $\lambda/4$ a single resonant microwave tone drives the two qubits with opposite phase, which leads to a coherent exchange interaction mediated by virtual photons \cite{vanLoo1494} forming a coupled two-qubit molecule \cite{Kannan2022}. In the absence of modulation the backscattering is suppressed by destructive interference~\cite{Guimond2020}, while the interference for forward scattering is constructive.
The addition of frequency modulation of both qubits with $\Omega/(2\pi)=20$\,MHz and $A_m/(2\pi) = 20$\,MHz results in nontrivial interference conditions for the Stokes and anti-Stokes side-bands, as shown in the insets of Fig.~\ref{fig3}.
The blue and green arrows correspond to the incident light (dashed) and the inelastically scattered light (solid) at $\pm 20$ MHz from the first and second qubit, respectively.
If the two modulation tones are in phase ($\alpha=0$), illustrated in the insets of Fig.~\ref{fig3}a and c by red arrows inside the qubits (up-up), the device continues to scatter light only in the forward direction since its symmetry is not modified by the modulation. Accordingly, we observe side-bands mostly scattered forward and almost fully suppressed in back scattering (dashed circles in panels a and c). However, if the modulation has a phase difference of $\alpha=\pi$, the situation is reversed. This is illustrated by the blue arrows inside the second qubit (up-down) in the insets of Fig.~\ref{fig3}b and d, corresponding to an additional phase factor of $-1$. While the inelastic backscattering is now highly likely as shown in Fig.~\ref{fig3}d, the side-bands scattered forward from the first qubit destructively interfere with the ones scattered from the second one due to the additional phase shift and thus preventing forward scattering as shown in Fig.~\ref{fig3}b.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figure3_v2.pdf}
\caption{\textbf{Resonance fluorescence spectra.} Power spectral density (PSD) measured in transmission (\textbf{a, b}) and reflection (\textbf{c, d}) at the digitizer for in-phase $\alpha = 0$ (\textbf{a, c}) and out-of-phase $\alpha = \pi$ (\textbf{b, d}) modulation. The Stokes components are highlighted with dashed circles. Scattering schematics are shown as insets where blue (green) arrows represent the light scattered from qubit 1 (2) at $\omega_0 \pm \Omega$ leading to constructive interference in \textbf{a} and \textbf{d} or destructive interference in \textbf{b} and \textbf{c}. Full Rayleigh peak heights are 1.9 and 1.6\,$V^2$ for the chosen settings in \textbf{a} and \textbf{b}.} \label{fig3}
\end{figure}
In order to better illustrate the phase and detuning dependence of the interference conditions
we extract the coherent scattering power of the Stokes component over the full range of $\alpha$ and for finite detuning from the qubit resonances at $\omega_0$. For this measurement the detection frequency is always detuned by the chosen modulation frequency $\Omega/(2\pi)=20$\,MHz from the probe tone at frequency $\omega$.
Here we detect both the transmitted and reflected scattered Stokes light with the two channels of the digitizer simultaneously for $A_m/(2\pi)=30$\,MHz. The obtained intensity in transmission and reflection is shown in Fig.~\ref{fig4}a and b.
We observe resonances at probe frequencies $\omega_0$, $\omega_0\pm \Omega$, and $\omega_0-2\Omega$ and their overall dependence on $\alpha$ is clearly pronounced and opposite in sign for forward and backward scattering.
These experimental results are in very good agreement with the theoretical model shown in Fig.~\ref{fig4}(b,d), see Methods for details.
The measured dependence of the scattering parameters on the phase difference $\alpha$ as well as the directivity $D = (P_{\rightarrow}-P_{\hookleftarrow})/(P_{\rightarrow}+P_{\hookleftarrow})$ is shown in Fig.~\ref{fig4}e for $(\omega-\omega_0)/(2\pi)=\Omega/(2\pi) = - 20$\,MHz (dashed white lines in panels a-d) together with theory. This shows phase selective control to enter the regimes when light is mostly scattered back ($D < 0$), forward ($D > 0$), or symmetrically in both directions ($D = 0$). The measured directivity demonstrates high diode efficiency that can be set continuously between 0.84 and -0.99.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figure4_v2.pdf}
\caption{\textbf{Coherent inelastic scattering spectrum of the Stokes component.}
\textbf{a(b)}, Measured and normalized Stokes power as a function of probe frequency detuning $\omega-\omega_0$ and relative phase between modulation tones $\alpha$ measured in transmission (reflection) at a fixed modulation amplitude $A_m/ (2\pi) = 30$\,MHz and modulation frequency $\Omega/ (2\pi) = 20$\,MHz. \textbf{c(d)}, Theoretically predicted transmission (reflection) spectrum on the same scale. \textbf{e}, Measured coherent inelastic scattering as a function of $\alpha$ at $(\omega - \omega_0)/ (2\pi) = \Omega/ (2\pi) = -20$ MHz (points) and theory (solid lines). Scattering directivity $D$ is shown in light green.} \label{fig4}
\end{figure}
\vspace{0.25cm}
\noindent
\section*{Discussion}
Besides the demonstrated high scattering directionality, at the relative phase $\alpha/\pi = \pm 1$, our system exhibits a gyrator-like behavior and transmits a signal unchanged in one direction whereas the reverse traveling signal experiences a $\pi$ phase-shift. For $\alpha/\pi = \pm 0.3$, our device demonstrates isolator properties, which might be enhanced by further tuning the device or by extending the principle to a larger number of qubits, details can be found in the Supplementary Discussion. This nonreciprocity relies on the qubits working in the linear regime, which would naturally limit the operation of the device to low powers $(\Omega_{R}/\Gamma_{1})^2 \ll 1$. In contrast, our calculations indicate that scattering remains directional up to intermediate drive powers
$(\Omega_{R}/\Gamma_{1})^2 \lesssim 9$ beyond which the inelastic scattering is fully suppressed, see Supplementary Figures.
We have demonstrated
an on-chip microwave photon router that can be switched \textit{in-situ} between scattering photons back, forward, or symmetrically in both directions.
While it is limited in dynamic range, it is fully compatible with modern superconducting quantum computing devices \cite{oliver2020b,Reuer2022} that operate in the single and few photon regime.
The suppression strength can be modified by the modulation amplitude and the signal frequency can be shifted and fine-tuned \textit{in-situ} by changing the modulation frequency. A larger range of frequency bands can be accessed by working at
odd multiples of the $\lambda/4$ boundary condition.
In the future, this device principle could be extended to multiple nodes, see Methods and Supplementary Information, in order to realize topologically protected states \cite{lu2016topological}, as a part of a hardware implementation of Gottesman-Kitaev-Preskill codes \cite{rymarz2021hardware}, or to route microwave radiation for the realization of chiral networks \cite{lodahl2017chiral}.
\vspace{0.25cm}
\noindent
\normalsize
\section*{Methods}
\footnotesize
\textbf{Calculation of scattering spectra.}
In this section we present the general approach to calculate photon scattering from an array of qubits with time-modulated resonance frequencies. Such a device is characterized by the following effective non-Hermitian Hamiltonian ~\cite{Caneva2015}:
\begin{multline}\label{eq:H}
H(t)=\sum\limits_{j}[\omega_0^{(j)}(t)-{\rm i} \Gamma_2^{(j)}]\sigma_j^\dag \sigma_j-\frac{{\rm i}\Gamma_1}2 \sum\limits_{j,k}{\rm e}^{{\rm i} \varphi|j-k|}\sigma_j^\dag \sigma_k
\\+\frac{\Omega_{\rm R}}{2}\sum\limits_j (\sigma_j^\dag{\rm e}^{{\rm i}\varphi j-{\rm i} \omega t}+{\rm H.c.})\:.
\end{multline}
Here $\sigma_j$ are the raising operators, $\Gamma_1$ is the (radiative) relaxation rate between the $|1\rangle$ and $|0\rangle$ qubit states, $\Gamma_2^{(j)}$ is the decay rate of the coherence between the $|1\rangle$ and $|0\rangle$ states, $\varphi=\omega_0d/c$ is the phase gained by light traveling between the qubits with propagation velocity $c$. The Rabi frequency $\Omega_R$ quantifies the incident wave amplitude and
\begin{equation}
\omega_0^{(j)}(t)=\omega_0+A_m\cos(\Omega t+\alpha_j)
\end{equation}
are the time-dependent qubit resonance frequencies.
The Hamiltonian Eq.~\eqref{eq:H} assumes the usual rotating wave and Markovian approximations.
Here, we are interested in the case of weak coherent driving
where the wavefunction can be approximately written as
\begin{equation}
\psi=|0\rangle+\sum\limits_j p_j \sigma_j^\dag |0\rangle\:.
\end{equation}
The amplitudes $p_j$ describe the coherence between the ground and excited states and can be found from the following Schr\"odinger equation:
\begin{multline}
{\rm i}\frac{{\rm d} }{{\rm d} t}p_j(t)=[\omega_0^{(j)}(t)-{\rm i}\Gamma_2^{(j)}]p_j\\-\frac{{\rm i}\Gamma_1}2 \sum\limits_{k}{\rm e}^{{\rm i} \varphi|j-k|}p_k+\frac{\Omega_{\rm R}}{2}\,{\rm e}^{{\rm i}\varphi j-{\rm i} \omega t} \:.
\end{multline}
It is convenient to seek the solution in the form
\begin{equation}
p_j(t)=\sum\limits_{n=-\infty}^\infty p_j^{(n)}{\rm e}^{-{\rm i} (\omega+n\Omega)t}\:,
\end{equation}
so that the amplitudes
$p_j^{(n)}$ are determined by the linear system \cite{Poddubny2021Ratchet}
\begin{multline}\label{eq:pjm}
(\omega+n\Omega)p_j^{(n)}=(\omega_0-{\rm i} \Gamma_2^{(j)})p_j^{(n)}
+\frac{A_m}{2}({\rm e}^{{\rm i} \alpha_j}p_j^{(n-1)}+{\rm e}^{{\rm i} \alpha_j}p_j^{(n+1)})
\\-\frac{{\rm i}\Gamma_1}2 \sum\limits_{k}{\rm e}^{{\rm i} \varphi|j-k|}p_k^{(n)}+\frac{\Omega_{\rm R}}{2}\, {\rm e}^{{\rm i}\varphi j}\delta_{m,0}\:.
\end{multline}
After the amplitudes $p_j^{(n)}$ have been found numerically, we calculate the coefficients $r^{(n)}$ and $t^{(n)}$
\begin{align}
r^{(n)}&=-\frac{{\rm i}\Gamma_1}{\Omega_{\rm R}}\sum\limits_j {\rm e}^{{\rm i}\varphi j}p_j^{(n)},\quad \\ t^{(n)}&=\delta_{n,0}-\frac{{\rm i}\Gamma_1}{\Omega_{\rm R}}\sum\limits_j {\rm e}^{-{\rm i}\varphi j}p_j^{(n)}\:,
\end{align}
that describe the backward (forward) scattering process with the frequency change $\omega \to \omega+n\Omega$. In the general case, the system of equations
\eqref{eq:pjm} is to be solved numerically. However, it is possible to obtain an analytical solution in the particular case of a single qubit~\cite{Marquardt2006}. In this case we find
\begin{equation}\label{eq:pnanswer}
p^{(n)}=\frac{\Omega_{\rm R}}{2}\,
\sum_{n'=-\infty}^\infty
\frac{ J_{n'-n}(A_m/\Omega)J_{n'}(A_m/\Omega)}{\omega+n'\Omega-\omega_0+{\rm i}\Gamma_2}\:.
\end{equation}
For elastic scattering ($n=0$) Eq.~\eqref{eq:pnanswer} leads to Eq.~\eqref{coh_oneQ} in the main text.
\noindent
\textbf{Resonance fluorescence of the time-modulated device.}
Here we describe the procedure to calculate the nested Mollow triplets shown in Fig.~\ref{fig2}.
The state of the qubit can be conveniently represented as vector $\bm S$ of the spin 1/2, where $|1\rangle$ and $|0\rangle$ states correspond to $S_z=1/2$ and $-1/2$, respectively. The dynamics $\bm S(t)$ is governed by the Bloch equation that reads
\begin{equation}\label{eq:dS}
\frac{{\rm d}\bm S}{{\rm d} t}=\bm S\times \widetilde{\bm \Omega}(t)-\bm\Gamma (\bm S - \bm S_0)
\end{equation}
where
\begin{align}
\widetilde{\bm \Omega}(t) = [\Omega_R \cos\omega t, \Omega_R \sin\omega t, \omega_0+\Delta\omega\cos(\Omega t+\alpha)]\,
\end{align}
is the time-dependent effective magnetic field,
$\bm S_0 = [0,0,-1/2]$ is the equilibrium spin, and the spin relaxation term reads $\bm\Gamma (\bm S - \bm S_0)\equiv [\Gamma_2 S_x,\Gamma_2 S_y,\Gamma_1 (S_z+1/2)] $.
The emission spectrum is determined by the spin correlation function
\begin{equation}\label{eq:Is}
I(\omega)\propto {\rm Re}\!\int_0^\infty {\rm d} t\:{\rm e}^{-{\rm i} \omega \tau}\langle\!\langle S_+(t+\tau)S_-(t)\rangle\!\rangle_t,
\end{equation}
where $S_\pm=S_x\pm {\rm i} S_y$ and the double angular brackets denote averaging over the absolute time $t$. Equation~\eqref{eq:Is} establishes the correspondence between the emission spectrum in the considered quantum optics problem and the electron spin noise spectrum in the conditions of electron paramagnetic resonance, when the electron is subject to two magnetic fields, a constant one and an oscillating one~\cite{poshakinskiy2020spin}.
In the theory of magnetic resonance, the standard trick to solve Eq.~\eqref{eq:dS} analytically is to switch to a reference frame rotating around the $z$ axis with the drive frequency $\omega$. There, the spin dynamics is governed by the same Eq.~\eqref{eq:dS} but $ \widetilde{\bm \Omega} (t)$ shall be replaced with
\begin{align}
\bm \Omega'(t) = [\Omega_R , 0, \omega_0-\omega+\Delta\omega\cos\Omega t]\,.
\end{align}
In the absence of modulation, $\Delta\omega = 0$, the effective magnetic field $ \bm \Omega'(t)$ would be constant and its amplitude
\begin{align}
\Omega_R' = \sqrt{\Omega_R^2 +(\omega_0-\omega)^2}
\end{align}
would determine the splitting in the conventional Mollow triplet.
The presence of modulation can be accounted for by repeating the trick and switching to yet another frame rotating with frequency $\Omega_R'$ with respect to the previous one. There, $\bm\Omega' (t)$ is replaced with
\begin{align}
\bm \Omega'' =\frac{\Omega_R \Delta\omega}{2{\Omega_R'}^2} [\omega-\omega_0 , 0, \Omega_R]+ \Big(1-\frac{\Omega}{\Omega_R'}\Big)[\Omega_R , 0, \omega_0-\omega]\,,
\end{align}
where we neglected all oscillating terms, since they average to zero. The amplitude of $\bm \Omega''$ determines the splitting of the nested Mollow triplet
\begin{align}
\Omega_R'' =\sqrt{\left(\frac{\Omega_R \Delta\omega}{2 \Omega_R'}\right)^2 + (\Omega_R' - \Omega)^2}\,.
\end{align}
Returning back to the initial reference frame, we obtain nine possible emission frequencies
\begin{align}
\omega_{p,q} = \omega + p \Omega_R' + q \Omega_R''\,,
\end{align}
where $p,q = 0,\pm 1$ enumerate the components of the two nested Mollow triplets. In the above analytical solution, we used twice the rotating wave approximation, which is valid provided $\Delta\omega \ll \Omega_R \ll \omega_0$.
\noindent
\textbf{Data normalization.}
We normalize the transmission spectra $|t_0|^2$ shown in Fig.~\ref{fig1}d, e and Fig.~\ref{fig2}a, b by dividing the background transmission coefficient $|t_0|^2 = |t|^2/|t_{bg}|^2$. Here, $|t_{bg}|^2$ is measured with both qubits tuned out of the frequency range of interest and $|t|^2$ is measured with the qubit tuned to the desired frequency. This method normalizes the gain in the system and compensates for the frequency-dependent transmission properties of the drive and detection lines.
The power spectral density of the measured resonance fluorescence spectrum $S(\omega)$ shown in Fig.~\ref{fig2}c, as well as the coherent inelastic scattering spectra shown in Fig.~\ref{fig4}a, b were scaled to the numerically predicted value. The latter relies on the qubit parameters extracted from the transmission measurements, the chosen modulation frequency and the independently calibrated modulation amplitude.
\vspace{0.25cm}
\noindent
\normalsize
\textbf{Data availability}\\
\footnotesize
All datasets and analysis files used in this study will be made available on the Zenodo repository before publication.
\vspace{0.25cm}
\noindent
\normalsize
\textbf{Acknowledgments}\\
\footnotesize
The authors thank L. Drmic and P. Zielinski for software development, and the MIBA workshop and the IST nanofabrication facility for technical support. This work was supported by the Austrian Science Fund (FWF) through BeyondC (F7105) and IST Austria. E.R. is the recipient of a DOC fellowship of the Austrian Academy of Sciences at IST Austria. J.M.F. and M.Z. acknowledge support from the European Research Council under grant agreement No 758053 (ERC StG QUNNECT) and a NOMIS foundation research grant. The work of A.N.P. and A.V.P. has been supported by the Russian Science Foundation under the grant No 20-12-00194.
\vspace{0.25cm}
\noindent
\normalsize
\textbf{Author contributions} \\
\footnotesize
E.S.R. designed and fabricated the samples, worked on the setup and performed the measurements. A.V.P. and A.N.P. developed the theory and together with E.S.R conducted the data analysis. M.Z. contributed to building the measurement setup and R.S. to the resonance fluorescence data acquisition code. E.S.R. and A.N.P. wrote the manuscript with contributions from all authors. J.M.F. supervised this work.
\vspace{0.25cm}
\noindent
\normalsize
\textbf{Competing interests} \\
\footnotesize
The authors declare no competing interests.
\renewcommand*{\bibfont}{\footnotesize}
|
\section{#1}\setcounter{equation}{0}}
\newtheorem{th}{Theorem}[section]
\newtheorem{cor}[th]{Corollary}
\newtheorem{lem}[th]{Lemma}
\newtheorem{prop}[th]{Proposition}
\newtheorem{example}[th]{Example}
\newtheorem{definition}[th]{Definition}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\def\sep{\;\vrule\;}
\def\proof#1. {\par
\ifdim\lastskip<15pt
\removelastskip\penalty-200
\vskip15pt plus3pt minus3pt
\fi
{\def\a{#1}
\ifx\a\empty
{\noindent\bf Proof.}
\else
{\noindent\bf Proof of #1.}
\fi}\enspace}
\def\restr#1{\,\vrule\,\lower1.75ex\hbox{$#1$}}
\def\endproof{\hfill\hspace{-6pt}\rule[-14pt]{6pt}{6pt}
\vskip22pt plus3pt minus 3pt}
\def\proofend{\endproof}
\def\be{\begin{equation}}
\def\ee{\end{equation}}
\def\bea{\begin{eqnarray}}
\def\eea{\end{eqnarray}}
\def\bean{\begin{eqnarray*}}
\def\eean{\end{eqnarray*}}
\def\SS{{$\cal \char'170$}}
\def\a{\alpha}
\def\b{\beta}
\def\d{\delta}
\def\D{\Delta}
\def\e{\varepsilon}
\def\f{\varphi}
\def\F{\Phi}
\def\g{\gamma}
\def\G{\Gamma}
\def\i{\infty}
\def\k{\kappa}
\def\l{\lambda}
\def\L{\Lambda}
\def\o{\omega}
\def\O{\Omega}
\def\r{\rho}
\def\s{\sigma}
\def\S{\Sigma}
\def\te{\theta}
\def\t{\theta}
\def\te{\theta}
\def\z{\zeta}
\def\c{{\rm cap}}
\def\li#1{\lim_{#1\to\i}}
\def\setm{\setminus}
\def\emp{\emptyset}
\def\empt{\emptyset}
\def\ov{\overline}
\def\c{{\rm cap}}
\def\li#1{\lim_{#1\to\i}}
\def\ss{{\cal S}}
\def\uw{U^{\mu_w}}
\def\inff{\inf_{P_n\in{\bf \Pi}_n}}
\font\tenopen = cmbx10
\font\sevenopen = cmbx7
\font\fiveopen = cmbx5
\newfam\openfam
\def\open{\fam\openfam\tenopen}
\textfont\openfam = \tenopen
\scriptfont\openfam = \sevenopen
\scriptscriptfont\openfam = \fiveopen
\def\R{{\open R}}
\def\N{{\open N}}
\def\C{{\open C}}
\def\u{\underline}
\def\Q{{\bf Q}}
\def\Z{\bf Z}
\def\rz{{\R/\Z}}
\title{Weyl's uniform distribution under periodic perturbation\footnote{AMS Classification: 11K06; Keywords: Weyl's theory, density (mod 1),
uniform distribution (mod 1),
polynomial sequences, vector sequences, periodic perturbations}}
\author{Vilmos Totik}
\date{}
\begin{document}
\maketitle
\begin{abstract} We examine the uniform distribution theory of H. Weyl when
there is a periodic perturbation present. As opposed to the classical
setting, in this case the conditions for (mod 1) density and (mod 1)
uniform distribution turn out to be different.
\end{abstract}
\sect{Introduction}
In connection with asymptotics of orthogonal polynomials Peter Yuditskii \cite{Yud} asked if
the numbers $n\a-\cos(n\pi \a+\phi_0)$, $n=1,2,\ldots$, where $\a$ is irrational
and $\phi_0$ is a given number, are dense modulo 1 (in short (mod 1)). Recall that
if $\a$ irrational, then the sequence $n\a$, $n=1,2,\ldots$, is dense in
$[0,1)$ (mod 1), so the problem asks if the same remains true if we perturb
this sequence in the fashion described.
This paper has emerged from this question, and will examine classical uniform distribution
theory when there is a periodic (or almost periodic) perturbation present.
We shall arrive at the most general results through simpler cases because those
simpler cases are interesting in themselves and because the proofs will be more
transparent by not repeating simpler arguments as we proceed to higher generality.
Let us begin with a simple result with an elementary proof that answers Yuditskii's question.
\begin{th}\label{th1} If $f$ is a continuous periodic function and $\a$ is irrational,
then the sequence $\{n\a+f(n)\}_{n=1}^\i$ is dense in $[0,1)$ {\rm (mod 1)}.
\end{th}
Thus, mod 1 density is always the case, but we shall see later that (mod 1) uniform
distribution may not be true. This is in contrast with the classical setting, where
the conditions for density and uniform distribution are mostly the same.
We shall have much more general density results, but for a start we give an elementary proof of
Theorem \ref{th1}.
\proof. Let $\e>0$ be small, and let $L$ be a large integer. Choose $\d<1/2$ such that $|f(u)-f(v)|<\e$ if
$|u-v|<\d$, and let $|f|\le M$, $M$ integer. If $\b$ is the/a period of $f$, then
choose a $k$ such that $k(\a/L)\in [-\e/L,\e/L]$ (mod 1)
and $k(1/\b)\in [-\d/\b,\d/\b]$ (mod 1) (possible
by Dirichlet's simultaneous approximation theorem, see \cite[Theorem 201]{Hardy}). Then $k\a\in[-\e,\e]$ (mod $L$)
and $k\in [-\d,\d]$ (mod $\b$).
Assume, for example, that
$k\a\in [0,\e]$ (mod $L$), say $k\a=k_1L+\a_1$, where $k_1$ is an integer and
$\a_1\in (0,\e]$ (since $\a$ is irrational,
$\a_1$ cannot be 0). Consider the numbers
\[x_m=mk\a,\qquad 1\le m\le N,\]
where $N=[L/\a_1]$. We have
$x_m=m(k\a-k_1L)=m\a_1$ (mod $L$),
and these last numbers $m\a_1$ are equidistant on the interval $[0,L]$ with
neighboring distance $\a_1\le \e$. At the same time
\[|f(k(m+1))-f(km))|=|f(km+(k\ {\rm mod}\ \b))-f(km)|
\le \e\]
because of the choice of $\d$ and because of $(k \ {\rm mod}\ \b)\in [-\d,\d]$,
where $(k \ {\rm mod}\ \b)$ denotes the number $k'$ of smallest absolute
value for which $k=k'$ (mod $\b$).
Let $M_1=[M/\a_1]+1$. Since
\be M\le M_1\a_1\le M+\a_1<M+1\label{m1}\ee
and
\be L-2-M<L-\a_1 -(M+\a_1)<(N-M_1)\a_1<L-M,\label{m2}\ee
it follows that if
$y_m=x_m+f(km)$ (mod $L$), $y_m\in [0,L)$, then the sequence
$\{y_m\}_{m=M_1}^{N-M_1}$ (which is the same as
$\{m\a_1+f(km)\}_{m=M_1}^{N-M_1}$ in view of (\ref{m1}), (\ref{m2}) and $|f|\le M$) starts from a positive number $<2M+1$ and ends with a number
$>L-2M-2$, and the distance in between consecutive terms is at most $2\e$.
Therefore, for every $y\in [2M+1,L-2M-2]$ there is a term $y_m$ in the
sequence with $|y-y_m|\le 2\e$, and for $L>4M+5$ (which guarantees that
$[2M+1,L-2M-2]$ contains two consecutive integers) this shows that
to any point $y'\in [0,1)$ there is a $y_m$ closer than $2\e$ (mod 1), and (mod 1) such a $y_m$ equals $km\a+f(km)$.\endproof
To illustrate that there is room for generalization, let us consider almost periodic functions.
Bohr's almost periodic (real) functions (sometimes called uniformly almost periodic
functions) coincide with the functions which can be uniformly approximated on $\R$ by generalized
(real) trigonometric polynomials
of the form $\sum c_je^{i\l_j x}$, see \cite[Section 84]{Bohr}\footnote{The original definition of almost periodicity is as follows:
a continuous real function $f$ is said to be almost periodic if for every $\e>0$ there is an
$L(\e)>0$ such for in every subinterval $I$ of $\R$ of length $\ge L(\e)$ there is a $t\in I$
for which $|f(x+t)-f(x)|\le \e$ for all $x\in \R$. That the space of almost periodic functions
coincides with the uniform closure of generalized trigonometric polynomials was one of Bohr's main theorems.}. Proposition \ref{th1} is the special case of
\begin{prop}\label{th2} If $f$ is almost periodic and $\a$ is irrational,
then the sequence $\{n\a+f(n)\}_{n=1}^\i$ is dense in $[0,1)$ {\rm (mod 1)}.
\end{prop}
\proof. First of all notice that the proof of Proposition \ref{th1} easily implies the following:
if $f_1,\ldots,f_\nu$ are continuous periodic functions and $\a$ is irrational,
then the sequence $\{n\a+f_1(n)+\cdots+f_\nu(n)\}_{n=1}^\i$ is dense in $[0,1]$ (mod 1).
Indeed, let $\b_1,\ldots,\b_\nu$ be the periods of $f_1,\ldots,f_\nu$, respectively.
As before, let $\e>0$ be given, and let $L$ be again a large integer. Choose $\d$ such that $|f_\tau(u)-f_\tau(v)|<\e/\nu$
for all $\tau$ if
$|u-v|<\d$, and let $\sum_\tau|f_\tau|\le M$, $1\le \tau\le \nu$, $M$ integer.
Dirichlet's simultaneous approximation theorem can again be applied and we can choose
a $k$ such that $k(\a/L)\in [-\e/L,\e/L]$ (mod 1)
and $k(1/\b_\tau)\in [-\d/\b_\tau,\d/\b_\tau]$ (mod 1) for all $\tau$.
The rest of the proof proceeds as before with minor modifications.
In particular, it follows that if $T=\sum_{\tau=1}^\nu (a_\tau\cos (\l_\tau x)+b_\tau
\sin (\l_\tau x))$ is a generalized trigonometric polynomial, then
the sequence $\{n\a+T(n)\}_{n=1}^\i$ is dense in $[0,1]$ (mod 1).
But every almost periodic function is the uniform
limit of such generalized trigonometric polynomials, and
the claim in the theorem follows: if $y\in (0,1)$ and $\e>0$ are given, then let
$\eta=\min\{\e,y,1-y\}$, choose
a $T$ with $|f-T|<\eta/2$ and then an $n$ with\footnote{$\{\cdot\}$ denotes fractional part.}
$|y-\{n\a+T(n)\}|<\eta/2$, from which we obtain
$|y-\{n\a+f(n)\}|<\eta<\e$.\endproof
\sect{Preliminaries, Weyl's theorems}\label{sectprel}
Let $\{\cdot\}$ denote fractional part. There are two slightly different ways
to talk about (mod 1) values of a sequence $x_n$, $n=1,2,\ldots$, namely one can look
at the sequence $\{x_n\}$, $n=1,2,\ldots$, of the fractional parts as elements of $[0,1)$
or as elements of the torus $\R/{\bf Z}$ (i.e. when we identify $0$ and $1$ in $[0,1]$).
From the point of view of (mod 1) denseness or (mod 1) uniform distribution it does
not matter which view we take. In our discussion it will be more convenient to work
on $[0,1)$ and not on the torus because the proofs are easier that way and because
the perturbing function $f$ may not be conveniently reduced to a function on the torus.
We shall briefly discuss in Section \ref{secttorus}
one question when the torus vs. $[0,1)$ makes a difference.
Let $\{x_n\}_{n=1}^\i$ be a sequence of real numbers. This sequence is said to be uniformly
distributed on $[0,1)$ (mod 1) if for all $0\le a<b<1$ we have
\[\lim_{n\to\i}\frac{1}{n}\#\bigl\{{1\le k\le n\sep \{x_k\}\in [a,b]}\bigr\}=b-a.\]
This is the same that for every continuous function $f$ on $[0,1]$ we have
\[\lim_{n\to\i}\frac{1}{n}\sum_{k=1}^nf(\{x_k\})=\int_0^1 f.\]
In a similar fashion, if $\{X_n\}_{n=1}^\i$ is a sequence of vectors from $\R^d$, then
its (mod 1) uniform
distribution (on $[0,1)^d$) means that for all $0\le a_i<b_i<1$ we have
\[\lim_{n\to\i}\frac{1}{n}\#\bigl\{{1\le k\le n\sep \{X_k\}\in }\prod_{i=1}^d[a_i,b_i]\bigr\}=\prod_{i=1}^d(b_i-a_i),\]
which is the same that for every continuous function $f$ on $[0,1]^d$ we have
\[\lim_{n\to\i}\frac{1}{n}\sum_{k=1}^nf(\{x_k\})=\int_{[0,1]^d} f.\]
Weyl's first theorem \cite[Satz 1]{Weyl} (\cite[Theorem 2.1]{K}) on uniform
distribution asserts that $\{x_n\}_{n=1}^\i$ is uniformly distributed (mod 1) on $[0,1)$ if and only if for all integers $a\not=0$
\[\lim_{n\to\i}\frac{1}{n}\sum_{k=1}^n e^{i2\pi a x_k}=0.\]
More generally, if
\be \{(x_{1,n},\ldots,x_{d,n})\}_{n=1}^\i\label{xnk}\ee
is a sequence from
$\R^d$, then
this sequence is uniformly distributed (mod 1) if and only if for all integer $d$-tuples
$(a_1,\ldots,a_d)$ with $\sum a_i^2\not=0$ we have
\[\lim_{n\to\i}\frac{1}{n}\sum_{k=1}^n e^{i2\pi (a_1x_{1,k}+\cdots+a_dx_{d,k})}=0,\]
see \cite[Satz 3]{Weyl} (\cite[Theorem 6.2]{K}).
In particular, (\ref{xnk}) is uniformly distributed on $[0,1)^d$ if an only if
every non-trivial integer combination sequence $y_n=a_1x_{1,n}+\cdots+a_dx_{d,n}$, $n=1,2,\ldots$,
is uniformly distributed on $[0,1)$.
This is not true if we talk about density ((\ref{xnk}) may not be dense
in $[0,1)^d$ even though every $\{y_n\}_{n=1}^\i$ is dense in $[0,1))$, but we shall see that for polynomial
sequences density and uniform distribution are the same.
Let $P_1,\ldots,P_d$ be real polynomials. It is clear that from the point of view
of the (mod 1) density or (mod 1) uniform distribution of the sequence
\[(P_1(n),\ldots,P_d(n)),\qquad n=1,2,\ldots,\]
the constant terms in the polynomials are irrelevant, therefore
in what follows we shall always assume that the polynomials
are without constant term.
For polynomial
sequences Weyl proved in \cite[Satz 9]{Weyl} (\cite[Theorem 3.2]{K}) that if $P$ (without constant term)
has at least one irrational coefficient, then $\{P(n)\}_{n=1}^\i$
is uniformly distributed on $[0,1)$ (mod 1). On the other hand, if all coefficients of $P$
are rational, then it is clear that $\{P(n)\}_{n=1}^\i$ has
only finitely many different terms (mod 1), so this sequence is
not dense in $[0,1)$ (mod 1).
To deal with the vector case we introduce the following definition.
\begin{definition} \label{def1} {\rm We call polynomials $P_1,\ldots,P_d$ (without constant term) $\Q$-independent, if
no non-trivial rational combination $r_1P_1+\cdots+r_d P_d$ is a polynomial
with rational coefficients. }
\end{definition}
Clearly, this is the same that no non-trivial
linear combination of $P_1,\ldots,P_d$ with integer coefficients
is a polynomial with integer coefficients.
Combining Weyl's theorems discussed so far we can see that if $P_1,\ldots,P_d$ are
polynomials without constant term, then
\be \{(P_1(n),\ldots,P_d(n))\}_{n=1}^\i\label{xnkd}\ee
is uniformly distributed in $[0,1)^d$ (mod 1) precisely when $P_1,\ldots,P_d$
are $\Q$-inde\-pen\-dent, i.e.
every non-trivial rational linear combination $\sum_j\a_j P_j$ has at least one
irrational coefficient. Furthermore, the condition is the same for the (mod 1) density of
the sequence (\ref{xnkd}) in $[0,1)^d$. Indeed, if there is a non-trivial linear combination
$r_1P_1+\cdots+r_dP_d$ with integer coefficients which has integer coefficients, then
$r_1\{P_1(n)\}+\cdots+r_d\{P_d(n)\}$ is an integer $r$ lying in the interval
$[-(|r_1|+\cdots+|r_d|),|r_1|+\cdots+|r_d|]$. Therefore, $(\{P_1(n)\},\ldots,\{P_d(n)\})$
lies in finitely many hyperplanes $r_1x_1+\cdots+r_dx_d=r$ of $\R^d$, hence this sequence
is not dense in $[0,1)^d$.
We can specialize these results to the case when $P_j(x)=\a_jx$ with some given $\a_j$ and obtain that
the sequence
\be \{(n\a_1,\ldots,n\a_d )\}_{n=1}^\i\label{xnkv}\ee
of vectors is uniformly distributed (mod 1) on $[0,1)^d$ if and only if
it is dense (mod 1) in $[0,1)^d$, and this happens precisely
if the numbers $1,\a_1,\ldots,\a_d$ are rationally independent,
i.e. no non-trivial rational linear combination of them is 0.
\sect{Density}
In this section we shall deal with the problem discussed in the introduction
in more generality. We shall see that periodic or almost periodic perturbations
never prevent denseness (mod 1). The story will be different when we
shall deal with uniform distribution in the next section.
The following proposition is equivalent to Proposition \ref{th2}, but
we present a formulation and a more streamlined proof for it that will be the
basis of the polynomial and vector cases to be followed.
\begin{prop} \label{th2*} If $\a$ is irrational and $f_1,\ldots,f_\nu$ are almost periodic functions,
then the sequence $\{n\a+\sum_jf_j(n)\}_{n=1}^\i$ is dense in $[0,1)$ {\rm (mod 1)}.
\end{prop}
\proof. As at the end of the proof of Proposition \ref{th2}, we may select each $f_j$
to be a trigonometric polynomial, and then, by considering individual terms
$\cos (\l x)$, $\sin (\l x)$ in the $f_j$'s and by increasing $\nu$
we may also assume that each $f_j$ is periodic
(say of the form $c\cos (\l x)$ or $c\sin (\l x)$ ). Let $\b_j$ be the period of
$f_j$, and set $\g_j=1/\b_j$. By replacing $f_j$ by $g_j(x)=f(\b_j x)$ it is sufficient
to show that if each $g_j$ is periodic with period 1,
then the sequence $\{n\a+\sum_jg_j(n\g_j)\}_{n=1}^\i$ is dense in $[0,1)$ (mod 1).
Let us choose a maximal subset $\{\g_\l\}_{\l\in \L}$ of $\{\g_1,\ldots,\g_\nu\}$
for which $\{1,\a\}\cup \{\g_\l\}_{\l\in \L}$ is a rationally independent set, i.e. 0 can be written
as a rational linear combination of these numbers only when all coefficients are 0. Thus, for every
$j\not\in \L$ there is a non-trivial dependence of the form
\[c_{j,0}\a+\sum_{\l\in \L} c_{j,\l}\g_\l+d_j\g_j=c_j,\]
where the $c$'s are integers and $d_j\not=0$ is an integer. With some large integer $L$
set $A=L\prod_{j\not \in \L} |d_j|$ (when $\L=\{1,\ldots,\nu\}$, then interpret the
product as 1), and consider numbers $n$ of the form
$n=Am$, $m=1,2,\ldots$. If $x\in (0,1)$ is given, then, by the rational independence of the set
$\{1,\a\}\cup \{\g_\l\}_{\l\in \L}$,
there is a subsequence ${\cal M}$ of the integers such that for $m\to\i$, $m\in {\cal M}$ we have
\[m\a\to x\ \ \mbox{(mod 1)}, \qquad m\g_\l \to \frac{1}{2}\ \ \mbox{(mod 1)}, \ \l\in \L.\]
Now
\[n\a+\sum_{j=1}^\nu g_j(n\g_j)=n\a+\sum_{j\in \L}g_j(n\g_j)+
\sum_{j\not \in \L}g_j\left( \frac{c_jn-c_{j,0}\a n-\sum_{\l\in \L} c_{j,\l}\g_\l n}{d_j}\right)\]
and this, by the 1 periodicity of the functions $g_j$, for $n=Am$ equals
\[A\{m\a\}+\sum_{j\in \L}g_j(A\{m\g_j\})+
\sum_{j\not \in \L}g_j\left(-c_{j,0}(A/d_j)\{m\a\}-\sum_{\l\in \L} c_{j,\l}(A/d_j)\{m\g_\l\}\right)\]
(mod 1), and this latter expression tends to
\be Ax+\sum_{j\in \L}g_j(A(1/2))+
\sum_{j\not \in \L}g_j\left(-c_{j,0}(A/d_j)x-\sum_{\l\in \L} c_{j,\l}(A/d_j)(1/2)\right)\label{89}\ee
as $m\to\i$, $m\in {\cal M}$.
From this we can conclude that the (mod 1) taken sequence
\[\{n\a+\sum_{j=1}^\nu g_j(n\g_j)\}_{n=1}^\i\]
is dense in the (mod 1) taken range of the continuous function
\[h(x)=Ax+\sum_{j\in \L}g_j(A(1/2))+
\sum_{j\not \in \L}g_j\left(-c_{j,0}(A/d_j)x-\sum_{\l\in \L} c_{j,\l}(A/d_j)(1/2)\right)\]
defined on $(0,1)$, and it is sufficient to show that this (mod 1) taken range is dense in $[0,1)$.
The function $h$ can also be considered on the interval $[0,1]$ where it is a continuous function, and,
if $|f_j|\le M$, $1\le j\le \nu$, then
$h(0)\le \nu M$, while $h(1)\ge A-\nu M$. Therefore, if $A>2\nu M+2$, then its range contains
an interval $[B,B+1]$ for some integer $B$, and hence if we take its range (mod 1), then we get the whole interval $[0,1)$.
This completes the proof.
\endproof
\subsection{The polynomial case}\label{pol}
As a generalization of what has been said before, we shall now examine the density of the
sequence
\be \Bigl\{P_0(n)+\sum_{j=1}^\nu f_j(P_j(n))\Bigr\}_{n=1}^\i,\label{00}\ee
(mod 1), where $P_j$, $1\le j\le \nu$, are
polynomials. As before, we may assume
that $P_j(0)=0$ for all $j$.
\begin{prop} \label{th2pol} Let $f_1,\ldots,f_\nu$ be almost periodic functions and
let $P_0,P_1,\ldots,P_\nu$ be polynomials without constant term. If $P_0$ has at least
one irrational coefficient, then the sequence {\rm (\ref{00})} is dense in $[0,1)$ {\rm (mod 1)}.
\end{prop}
\begin{cor} With the assumptions of the preceding proposition
\[\limsup_{n\to\i}\Bigl|\{P_0(n)\}-\bigl\{\sum_{j=1}^\nu f_j(P_j(n))\bigr\}\Bigr|\ge \frac{1}{2}.\]
\end{cor}
This corollary says that the fractional parts $\{P_0(n)\}$, $n=1,2,\ldots$, cannot be imitated by sums of the fractional
part of the values of almost periodic functions taken at polynomial places.
Note also that for $\nu=1$ and $f_1\equiv 1/2$ the equality is attained, so more than what is
claimed cannot be stated.
The corollary is immediate, for if the limsup in question was smaller than $1/2$, then
the fractional part of $P_0(n)-\sum_{j=1}^\nu f_j(P_j(n))$, $n=1,2,\ldots$,
would not accumulate at 1/2, so this sequence (mod 1) would not be dense in $[0,1)$
contradicting Proposition \ref{th2pol}.
\proof Proposition \ref{th2pol}. We follow the preceding proof, and exactly as there we may assume that each $f_j$ is a continuous
periodic function with period 1.
Let us choose a maximal subset $\{P_\l\}_{\l\in \L}$ of $\{P_1,\ldots,P_\nu\}$
for which $\{P_0\}\cup \{P_\l\}_{\l\in \L}$ is a $\Q$-independent set. Thus, for every
$j\not\in \L$ there is a non-trivial dependence of the form
\[c_{j,0}P_0+\sum_{\l\in \L} c_{j,\l}P_\l+d_jP_j=R_j,\]
where the $c$'s are integers, $d_j\not=0$ is an integer and $R_j$ is a polynomial
with integer coefficients. With some large integer $L$
set again $A=L\prod_{j\not \in \L} |d_j|$ and define
\[P_j^*(x)=\frac{1}{A}P_j(Ax), \qquad R_j^*(x)=\frac{1}{A}R_j(Ax).\]
The $R_j^*$ has integer coefficients, and we claim that
$\{P_0^*\}\cup \{P_\l^*\}_{\l\in \L}$ is again a $\Q$-independent set.
Indeed, if this was not the case, then we could find integers $a_0, a_\l$, not all zero,
such that
\[Q(x)=a_0P_0^*(x)+\sum_{\l\in \L}a_\l P_\l^*(x)=
\frac{a_0}{A}P_0(Ax)+\sum_{\l\in \L}\frac{a_\l}{A} P_\l(Ax)\]
has integer coefficients. But then for $D$ equal to the degree of $Q$,
the coefficients of
\[a_0A^{D-1}P_0(x)+\sum_{\l\in \L}a_\l A^{D-1} P_\l(x)=A^D Q(x/A)\]
would have integer coefficients, which is impossible by the choice of
$\L$.
Consider numbers $n$ of the form
$n=Am$, $m=1,2,\ldots$. If $x\in (0,1)$ is given, then, by the $\Q$-independence of the set
$\{P_0^*\}\cup \{P_\l^*\}_{\l\in \L}$
and by Weyl's theorem,
there is a subsequence ${\cal M}$ of the integers such that for $m\to\i$, $m\in {\cal M}$, we have
\[P_0^*(m)\to x\ \ \mbox{(mod 1)}, \qquad P_\l^*(m) \to \frac{1}{2}\ \ \mbox{(mod 1)}, \ \l\in \L.\]
Now
\bean P_0(n)&+&\sum_{j=1}^\nu f_j(P_j(n))=P_0(n)+\sum_{j\in \L}f_j(P_j(n))\\
&+&
\sum_{j\not \in \L}f_j\left( \frac{R_j(n)-c_{j,0}P_0(n)-\sum_{\l\in \L} c_{j,\l}P_\l(n)}{d_j}\right).\eean
Here for all $j$ and $n=Am$ we have
\[P_j(n)=AP_j^*(m), \qquad R_j(n)=AR_j^*(m),\]
and since $A$ is divisible by $d_j$, it follows that each $R_j(n)/d_j$ has integer coefficients. Thus,
by the 1-periodicity of the functions $f_j$, for $n=Am$ the sum
$ P_0(n)+\sum_{j=1}^\nu f_j(P_j(n))$ equals
\bean A\{P_0^*(m)\}&+&\sum_{j\in \L}f_j(A\{P_j^*(m)\})\\
&+&
\sum_{j\not \in \L}f_j\left(-c_{j,0}(A/d_j)\{P_0^*(m)\}-\sum_{\l\in \L} c_{j,\l}(A/d_j)\{P_\l^*(m)\}\right)\eean
(mod 1), and this latter expression tends to
\[Ax+\sum_{j\in \L}f_j(A(1/2))+
\sum_{j\not \in \L}f_j\left(-c_{j,0}(A/d_j)x-\sum_{\l\in \L} c_{j,\l}(A/d_j)(1/2)\right),\]
as $m\to\i$, $m\in {\cal M}$.
This is now the analogue of (\ref{89}), and from here the proof is completed exactly as
the preceding proof was completed from (\ref{89}).\endproof
Proposition \ref{th2pol} can be written in a seemingly more general, but actually equivalent form.
\begin{prop} \label{th2pol1} Let $F$ be an almost periodic function of $\nu$ variables\footnote{The set of these functions
$f: \R^\nu\to \R$
is the closure in the uniform norm of the algebra generated by the monomials $\cos \l x_j$, $\sin \l x_j$, $\l\in \R$, $1\le j\le \nu$.}
and
let $P_0,P_1,\ldots,P_\nu$ be polynomials without constant term. If $P_0$ has at least
one irrational coefficient, then the sequence
\be \Bigl\{P_0(n)+F(P_1(n),\ldots,P_\nu(n))\Bigr\}_{n=1}^\i,\label{001}\ee
is dense in $[0,1)$ {\rm (mod 1)}.
\end{prop}
\proof. One can replace $F$ by a generalized trigonometric polynomial of several variables, and then one
can follow the preceding proof without much change.
An alternative is to use that if $F$ is a generalized trigonometric polynomial of $\nu$
variables, then
$F(P_1(n),\ldots,P_\nu(n))$ is of the form
\[\sum_{j=1}^{\nu_1} f_j(R_j(n))\]
where each $f_j$ is a $c\cdot\cos \l x$ or a $c\cdot \sin \l x$,
and where the $R_j$'s are polynomials (just use trigonometric
identities to convert products to sums).
This reduces then the claim to Proposition \ref{th2pol}.\endproof
\subsection{The vector case}
Let $\a_1,\ldots,\a_d$ be real numbers. Recall Weyl's theorem from Section \ref{sectprel}
according to which
for the (mod 1) density in $[0,1)^d$ of the vectors
$(n\a_1,\ldots,n\a_d)$, $n=1,2,\ldots$, it is necessary and sufficient that the
numbers $1,\a_1,\ldots,\a_d$
are rationally independent.
\begin{prop}\label{th1v} If $f_1,\ldots,f_d$ are almost periodic functions and
the numbers $1,\a_1,\ldots,\a_d$ are rationally independent,
then the sequence
\be \bigl(n\a_1+f_1(n),\ldots,n\a_d+f_d(n)\bigr), \qquad n=1,2,\ldots,\label{seq}\ee
is dense in $[0,1)^d$ {\rm (mod 1)}.
\end{prop}
\proof. The argument is familiar by now, see e.g. the proof of Proposition \ref{th2*}.
By replacing each $f_j$ by a trigonometric polynomial close to it, it is enough to prove that if
$f_{i,j}$, $1\le i\le d$, $1\le j\le \nu$, are continuous periodic functions then for
any $\b_{i,j}$ the sequence
\be \Bigl(n\a_1+\sum_{j=1}^\nu f_{1,j}(n\b_{i,j}),\ldots,n\a_d+\sum_{j=1}^\nu f_{d,j}(n\b_{d,j})\Bigr), \qquad n=1,2,\ldots,\label{seq111}\ee
is dense in $[0,1)^d$ (mod 1). By applying dilation in $f_{i,j}$ and changing at the same
time the corresponding $\b_{i,j}$, we may assume that each $f_{i,j}$ is of period 1.
Select a maximal subset $\L\subset \{(i,j)\sep 1\le i\le d,\ 1\le j\le \nu\}$
such that
\be \{1,\a_1,\ldots,\a_d\}\cup \{\b_\l\sep \l\in \L\}\label{oi}\ee
is a rationally independent set. Then every $\b_{i,j}$ with $(i,j)\not\in \L$
can be expressed as a rational combination
of the numbers in (\ref{oi}), say (with $\a_0=1$)
\[\b_{i,j}=\sum_{\tau=0}^d c_\tau^{i,j}\a_\tau+\sum_{\l\in \L}d_\l^{i,j} \b_\l.\]
If $q$ is the least common denominator of the coefficients in these representations,
then for some large $L$ set $A=Lq$ and consider the numbers $n=Am$, $m=1,2,\ldots$, for which
\[\Bigl(n\a_1+\sum_{j=1}^\nu f_{1,j}(n\b_{i,j}),\ldots,n\a_d+\sum_{j=1}^\nu f_{d,j}(n\b_{d,j})\Bigr)\]
is of the form
\be \Bigl(Am\a_1+\sum_{j=1}^\nu f_{1,j}(m(\cdots))\ldots,Am\a_d+\sum_{j=1}^\nu f_{d,j}(m(\cdots))\Bigr)\label{kk}\ee
where $(\cdots)$ represent integral combinations of the numbers in (\ref{oi}).
By Weyl's theorem, if $(x_1,\ldots,x_d)\in (0,1)^d$, then there is a subsequence ${\cal M}$ of the natural numbers such
that, as $m\to\i$, $m\in {\cal M}$, we have
\[\{m\a_j\}\to x_j,\quad 1\le j\le d, \qquad m\b_\l\to \frac12, \quad \l\in \L,\]
and we obtain from (\ref{kk}) that the sequence (\ref{seq}) is dense (mod 1) in the (mod 1) taken range of the
function
\be H(x_1,\ldots,x_d)=\Bigl(Ax_1+\sum_{j=1}^\nu f_{1,j}(\cdots)\ldots,Ax_d+\sum_{j=1}^\nu f_{d,j}(\cdots)\Bigr)\label{kk1}\ee
for $(x_1,\ldots,x_d)\in [0,1]^d$, where now $\cdots$ stand for some integral linear combinations of $x_1,\ldots,x_d, 1/2$.
Let $|f_{i,j}|\le M$ and $A>2\nu M+2$. By the following lemma (apply it to
\[h(x_1,\ldots,x_d)=H(x_1/A,\ldots,x_d/A)\]
and to the cube $Q=[0,A]^d$) the range of $H$ over
$[0,1]^d$ contains a cube of side-length 1, hence the (mod 1) taken range of $H$ contains $[0,1)^d$,
from which the claim in the proposition follows.\endproof
In the next lemma we set $\underline x=(x_1,\ldots,x_d)$.
\begin{lem}\label{lem} If
\[h(\underline x)=(x_1+h_1(\underline x),\ldots,x_d+h_d(\underline x))\]
is a continuous mapping of a cube $Q\subset \R^d$ of side-length $A$ into $\R^d$, where
$|h_i|\le A/2-1$, then the range of $h$ contains a {\rm (}closed{\rm )} cube of side-length 1.
\end{lem}
\proof. This is an easy consequence of the Brouwer fixed point theorem. Indeed, we may assume
that $Q=[-A/2,A/2]^d$, and we show that then the range contains the cube $[-1/2,1/2]^d$.
Suppose to the contrary that this is not the case, and there is a point $\u y\in [-1/2,1/2]^d$
which is not in the range. For an $\u x\in Q$ connect
$h(\u x)$ with $\u y$, and let $\Phi(\u x)$ be the intersection of the extension
of this segment (beyond $\u y$) with the boundary $\partial Q$ of $Q$. This defines a continuous
mapping $\Phi: Q\to \partial Q$. Note that if $\u x$ belongs to the boundary of $Q$, say $x_1=-A/2$, then,
since $x_1+h_1(\u x)\le -1$, in forming $\Phi(\u x)$ we are connecting the point $h(\u x)$ lying in
the half-space $x_1\le -1$
with the point $\u y$ lying in the half-space $x_1\ge -1/2$. This implies (say by
the convexity of the half-space $x_1<-1/2$) that the first coordinate of $\Phi(\u x)$ is
$\ge -1/2$, and hence $\Phi(\u x)\not= \u x$. Thus, $\F$ is a continuous mapping
of $Q$ into itself without a fixed point, which is impossible by the Brouwer fixed point theorem.
This contradiction proves the lemma.\endproof
\subsection{The polynomial-vector case}
The most general theorem in this paper on density is the following (cf. also Theorem \ref{thmaintor} below).
\begin{th}\label{thmain} Let $F_1,\ldots,F_d$ be
almost periodic functions of $\nu$ variables, and let
$P_{i,j}$ and $P_1,\ldots,P_d$ be polynomials without constant term.
If $\{P_1,\ldots,P_d\}$ are $\Q$-independent, then the sequence of vectors
\be \left\{\Bigl(P_1(n)+F_1(P_{1,1}(n)), \ldots,P_{1,\nu}(n)),\ldots,
P_d(n)+F_d(P_{d,1}(n),\ldots,P_{d\nu}(n))\Bigr)\right\}_{n=1}^\i\label{00101}\ee
is dense in $[0,1)^d$ {\rm (mod 1)}.
\end{th}
Recall that the $\Q$-independence of $\{P_1,\ldots,P_d\}$ is necessary
for density when each $F_i$ is identically 0.
\proof. All components of the proof have already been demonstrated in our previous discussions.
First of all, we may replace each $F_i$ by a trigonometric polynomial of $\nu$ variables (see the end of
the proof of Proposition \ref{th2}), then
using trigonometric identities, each $F_i(P_{i,1}(n),\ldots,P_{i,\nu}(n))$ can be written
in the form $\sum_{j=1}^\mu\f_i(S_{i,j})$ where each $\f_i$ is a $c\cdot \cos\l t$ or a $c\cdot \sin\l t $
and $S_{i,j}$ are polynomials,
and finally such an expression is of the form $\tilde F_i(S_{i,1}(n),\ldots,S_{i,\mu}(n))$,
where $\tilde F_i$ is a continuous function of $\mu$ variables which is periodic in each variable.
Thus, we may assume from the start that $F_i$ is periodic in each of its variable, and by scaling
we may even assume that the periods are 1.
Select a maximal subset $\L\subset \{(i,j)\sep 1\le i\le d,\ 1\le j\le \nu\}$
such that
\be \{P_1,\ldots,P_d\}\cup \{P_\l\sep \l\in \L\}\label{oi1}\ee
is a $\Q$-independent set of polynomials. Then for every $P_{i,j}$ with $(i,j)\not\in \L$
there is a relation with the polynomials in (\ref{oi1}) of the form
\[\sum_{\tau=1}^d c_\tau^{i,j}P_\tau+\sum_{\l\in \L} c_\l^{i,j}P_\l+d_{i,j}P_{i,j}=R_{i,j},\]
where the $c$'s are integers, $d_{i,j}\not=0$ is an integer and $R_{i,j}$
is a polynomial with integer coefficients. We consider again
with some large integer $L$ the number
$A=L\prod_{(i,j)\not \in \L} |d_{i,j}|$ and define for a polynomial $P$ the polynomial
$P^*(x)=\frac{1}{A}P(Ax)$. The $*$-transform of the system (\ref{oi1}) is again
$\Q$-independent (see the proof of Proposition \ref{th2pol}), so, by Weyl's theorem,
if $(x_1,\ldots,x_d)\in (0,1)^d$ is given, then
there is a subsequence ${\cal M}$ of the integers such that for $m\to\i$, $m\in {\cal M}$ we have
\[P_j^*(m)\to x_j\ \mbox{(mod 1)}, \ j=1,\ldots,d,\qquad P_\l^*(m) \to
\frac{1}{2}\ \ \mbox{(mod 1)}, \ \l\in \L.\]
By considering numbers of the form $n=Am$, $m=1,2,\ldots$, and follow the steps in the proof of
Propositions \ref{th2pol} and \ref{th1v} we can conclude that the closure of the (mod 1)
taken sequence (\ref{00101}) includes the (mod 1) taken range of the function
\[ H(x_1,\ldots,x_d)=\Bigl(Ax_1+F_1(\cdots)\ldots,Ax_d+ F_d(\cdots)\Bigr)\]
for $(x_1,\ldots,x_d)\in [0,1]^d$, where, for each of the variables
of each $F_i$, the $\cdots$ stands for some integral
linear combinations of $x_1,\ldots,x_d, 1/2$. That this (mod 1) taken range is
$[0,1)^d$ follows from Lemma \ref{lem} precisely as in the proof of Proposition \ref{th1v}.\endproof
\sect{Uniform distribution}
In this section we shall examine
the uniform distribution of the sequences that have been considered before.
In the classical case Weyl's theorems implies that the conditions for
denseness and uniform distribution are the same. We shall see that
when a periodic perturbation is present the situation is
different: denseness is always true (as we have seen in the preceding
section), but that is not the case for uniform distribution.
\subsection{The scalar case} Let us start with the simplest possible situation.
\begin{prop} \label{th3} Let $\a$ be irrational and let $f$ be a $\b$-periodic continuous function. Then
the sequence $\{n\a+f(n)\}_{n=1}^\i$ is uniformly distributed {\rm (mod 1)} if $\b$ is rational
or $1,\a,1/\b$ are rationally independent.
\end{prop}
Let us mention that Proposition \ref{th3} is precise in the sense that if $\b$ is irrational and
$1,\a,1/\b$ are rationally dependent, then there is a $\b$-periodic continuous function $f$ such that
the sequence $\{n\a+f(n)\}_{n=1}^\i$ is not uniformly distributed (mod 1).
See Proposition \ref{th4}.
\proof. Set $\g=1/\b$ and $g(x)=f(\b x)$. Then $g$ is 1-periodic
and $n\a+f(n)=\{n\a\}+g(\{n\g\})$ (mod 1).
Consider first the case when $\g=1/\b$ is rational, say $\g=p/q$. Then for $n=mq+s$, $0\le s<q$, we have
\[n\a+g(n\g)=\{m(q\a)\}+(g(s\g)+s\a)\qquad \mbox{(mod 1)},\]
and since here the second term on the right is independent of
$m$ and $q\a$ is irrational, it follows from the uniform distribution of $\{m(q\a)\}_{m=1}^\i$
(mod 1) that
the sequence $\{(mq+s)\a+f((mq+s)\a)\}_{m=1}^\i$ is uniformly distributed (mod 1). Since this is true
for all $0\le s<q$, the (mod 1) uniform distribution of $\{n\a+f(n)\}_{n=1}^\i$ follows.
Let us now consider the case when $1,\a,1/\b$, i.e. $1,\a,\g$ are rationally independent. Then
the vector sequence $\{(n\a,n\g)\}_{n=1}^\i$ is uniformly distributed
in $[0,1)^2$ (mod 1).
Let $0<a<b<1$ be fixed, and for an $0<\e<\min\{a,1-b\}$ let $m$ be so large that the
oscillation\footnote{Here and in what follows, the oscillation of a function on a set
is the difference of its supremum and infimum on that set} of $g$ over any interval of length $<1/m$ is smaller
than $\e$. For $s=0,1,\ldots,m-1$ and for an $n$ consider the sets
\[K_{1,s}=\left\{1\le k\le n\sep k\g\in \left[\frac{s}{m},\frac{s+1}{m}\right],\
k\a\in \left[a-g\left(\frac{s}{m}\right)+\e,b-g\left(\frac{s}{m}\right)-\e\right]\right\},\]
and
\[K_{2,s}=\left\{1\le k\le n\sep k\g\in \left[\frac{s}{m},\frac{s+1}{m}\right],\
k\a\in \left[a-g\left(\frac{s}{m}\right)-\e,b-g\left(\frac{s}{m}\right)+\e\right]\right\},\]
where the relations are understood (mod 1). Since for
\[k\g\in \left[\frac{s}{m},\frac{s+1}{m}\right]\qquad \mbox{(mod 1)}\]
we have
\[g(k\g)\in \left[g\left(\frac{s}{m}\right)-\e,g\left(\frac{s}{m}\right)+\e\right]\]
by the choice of $m$, it follows that for $k\in K_{1,s}$ we have
$k\a+g(k\g)\in [a,b]$ (mod 1), and conversely, if
$k\g\in \left[\frac{s}{m},\frac{s+1}{m}\right]$ and
$k\a+g(k\g)\in [a,b]$ (mod 1), then
$k\in K_{2,s}$. By the (mod 1) uniform distribution of the sequence of vectors $\{(k\a,k\g)\}_{k=1}^\i$ we have
\[\#K_{1,s}=(1+o(1))n\frac{b-a-2\e}{m},\qquad \#K_{2,s}=(1+o(1))n\frac{b-a+2\e}{m},\]
hence it follows by summing these for all $s=0,1,\ldots,m-1$ that
\[\liminf_{n\to\i}\frac1{n}\#\left\{1\le k\le n\sep k\a+g(k\g)\in [a,b] \mbox{\ \ (mod 1)}\right\}\ge b-a-2\e\]
and
\[\limsup_{n\to\i}\frac1{n}\#\left\{1\le k\le n\sep k\a+g(k\g)\in [a,b] \mbox{\ \ (mod 1)}\right\}\le b-a+2\e.\]
Since $\e>0$ and $0<a<b<1$ are arbitrary, the uniform distribution of $\{n\a+f(n)\}_{n=1}^\i$ (mod 1) follows.\endproof
We shall need the extension of Proposition \ref{th3} to more than one term. To this end, note that
for an irrational $\a$ the condition
``$\b$ is rational or $1,\a,1/\b$ are rationally independent" can be phrased as
``$\a$ cannot be written as a rational linear combination of $1$ and $1/\b$". It turns out that this is the right
formulation in the general case.
\begin{prop} \label{th31} Let $\a$ be irrational and let $f_1,\ldots,f_\nu$ be continuous periodic functions with periods
$\b_1,\ldots,\b_\nu$. If $\a$ cannot be written as a rational linear combination of the numbers
$1,1/\b_1,\ldots,1/\b_\nu$, then
the sequence $\{n\a+f_1(n)+\cdots f_\nu(n)\}_{n=1}^\i$ is uniformly distributed {\rm (mod 1)}.
\end{prop}
This is best possible in the sense that if $\a$ and $\b_j$ are given and
$\a$ can be written as a
rational linear combination of the numbers
$1,1/\b_1,\ldots,1/\b_\nu$, then there are $\b_j$-periodic continuous functions $f_j$ for $j=1,\ldots,\nu$,
such that the sequence $n\a+f_1(n)+\cdots f_\nu(n)$, $n=1,2,\ldots$,
is not uniformly distributed (mod 1).
See Proposition \ref{th4}.
\proof. A large part of the proof has been already established in the preceding proof.
Set $\g_i=1/\b_i$ and $g_i(x)=f_i(\b_i x)$. Then $g_i$ are 1-periodic
and
\[n\a+f_1(n)+\cdots f_\nu(n)=\{n\a+g_1(n\g_1)+\cdots g_\nu(n\g_\nu)\}\qquad \mbox{(mod 1)}.\]
Select a maximal subset of $\g_1,\ldots,\g_\nu$ that forms with 1 a rationally independent
set. We may assume that $\{\g_1,\ldots,\g_\mu\}$ is this maximal subset. Then, by the assumption of
the theorem, $1,\a,\g_1,\ldots,\g_\mu$ are rationally independent, furthermore every $\g_j$ with $j>\mu$ is a rational
combination of the form
\be \g_j=d_j+\sum_{i=1}^\mu c_{j,i}\g_i.\label{tt00}\ee
Let $q$ be a common multiple of the denominators of these $d_j$. If $k$ is of the form
$k=ql+r$, $0\le r<q$, then
\[k\g_j=rd_j+\sum_{i=1}^\mu c_{j,i}k\g_i\quad \mbox{(mod 1)},\]
and it is sufficient to show the uniform distribution in question along
every sequence ${\cal N}_r=\{ql+r\}_{l=1}^\i$ where $0\le r<q$ is fixed.
For a large $m$ partition $[0,1]^\mu$ by the hyperplanes $x_i=t/m$, $t=1,\ldots,m-1$, $1\le i\le \mu$,
in $\R^\mu$ into $m^\mu$
little cubes $I_s$, $1\le s\le m^\mu$. The image of such an $I_s$ under the mapping
\be (x_1,\ldots,x_\mu)\to rd_j+\sum_{i=1}^\mu c_{j,i}x_i\label{tt100}\ee
(cf. (\ref{tt00})) is an interval $I_{j,s}$ of length $\le C/m$.
Now let $0<a<b<1$ be fixed, and for an $0<\e<\min\{a,1-b\}$ let $m$ be so large that the oscillation of
$g_1(x_1)+\cdots + g_\mu(x_\mu)$ over any $I_s$, as well as the oscillation of each
$g_j$, $j>\mu$, over any $I_{j,s}$ is smaller than $\e/2\nu$. Let, furthermore, $Q_s=(Q_{1.s},\ldots,Q_{\mu,s})$ be an
arbitrary point from $I_s$, let $\tilde Q_{j,s}$ be the image of $Q_s$ under the mapping (\ref{tt100}),
and set
\[T_s=\sum_{i=1}^\mu g_i(\g_i Q_{i,s})+\sum_{j=\nu+1}^d g_j(\g_j \tilde Q_{j,s}).\]
For $s=1,\ldots,m^\mu$ consider the sets
\[K_{1,s}=\left\{1\le k\le n,\ k\in {\cal N}_r\sep (k\g_1,\ldots,k\g_\mu)\in I_s,\
k\a\in [a-T_s+\e,b-T_s-\e]\right\},\]
and
\[K_{2,s}=\left\{1\le k\le n,\ k\in {\cal N}_r\sep (k\g_1,\ldots,k\g_\mu)\in I_s,\
k\a\in [a-T_s-\e,b-T_s+\e]\right\},\]
where the relations are understood (mod 1). Exactly as in the preceding proof,
by the choice of $m$, it follows that for $k\in K_{1,s}$ we have
\be k\a+g_1(\g_1 k)+\cdots g_\nu(\g_\nu k)\in [a,b] \quad {\rm (mod}\ 1),\label{fgh}\ee
and conversely, if
$(k\g_1,\ldots,k\g_\mu)\in I_s$ and (\ref{fgh}) is true, then
$k\in K_{2,s}$.
By the (mod 1) uniform distribution of the sequence of vectors
\[\{(k\a,k\g_1,\ldots,k\g_\mu)\}_{k=1,\; k\in{\cal N}_r}^\i\]
(which follows from the rational independence of $(1,\a,\g_1,\ldots,\g_\mu)$), we have
\[\#K_{1,s}=(1+o(1)n\frac{b-a-2\e}{q\cdot m^\mu},\qquad \#K_{2,s}=(1+o(1))n\frac{b-a+2\e}{q\cdot m^\mu}.\]
It follows by summing these for all $s=1,\ldots,m^\mu$ that
\bean \liminf_{n\to\i}\frac{q}{n}\#\left\{1\le k\le n,\ k\in {\cal N}_r\sep k\a+g_1(\g_1 k)+\cdots g_\nu(\g_\nu k)\in [a,b]
\mbox{\ \ (mod 1)}\right\}\\
\ge b-a-2\e\eean
and
\bean \limsup_{n\to\i}\frac{q}{n}\#\left\{1\le k\le n,\ k\in {\cal N}_r\sep k\a+g_1(\g_1 k)+\cdots g_\nu(\g_\nu k)\in [a,b]
\mbox{\ \ (mod 1)}\right\}\\
\le b-a+2\e.\eean
Since $\e>0$ and $0<a<b<1$ are arbitrary and since there are $n/q+O(1)$ numbers $k\in {\cal N}_r$ with
$1\le k\le n$, the uniform distribution of
\[\{k\a+g_1(\g_1 k)+\cdots g_\nu(\g_\nu k)\}_{k=1,\;k\in {\cal N}_r}^\i\]
(mod 1) follows.\endproof
Next we state a somewhat more general form (but actually, since the number $\nu$ of terms in Proposition
\ref{th31} is arbitrary, and since, by Weierstrass theorem, every periodic continuous function
of several variables can be uniformly approximated by trigonometric polynomials of the same variables,
one can show that this seemingly more general form is equivalent to
Proposition \ref{th31}).
\begin{prop} \label{th31000} Let $\a$ be irrational and let $F$ be a continuous
function of the variables $x_1,\ldots,x_\nu$ which is periodic in each variable with periods
$\b_1,\ldots,\b_\nu$, respectively. If $\a$ cannot be written as a rational linear combination of the numbers
$1,1/\b_1,\ldots,1/\b_\nu$, then
the sequence $\{n\a+F(n,\ldots,n)\}_{n=1}^\i$ is uniformly distributed {\rm (mod 1)}.
\end{prop}
In particular, if
\[F(x_1,\ldots,x_\nu)=f_1(x_1)+\cdots+f_\nu(x_\nu),\] where each $f_j$ is a continuous periodic function
of period $\b_j$, then we obtain Proposition \ref{th31}.
\proof. We follow the preceding proof.
Set $\g_i=1/\b_i$ and
\[G(x_1,\ldots,x_\nu)=F(\b_1 x_1,\ldots,\b_\nu x_\nu).\]
Then $G$ is 1-periodic in each variable
and
\[n\a+F(n,\ldots,n)=\{n\a+G(\{n\g_1\},\ldots,\{n\g_\nu\})\}\qquad \mbox{(mod 1)}.\]
Select again a maximal subset of $\g_1,\ldots,\g_\nu$ that forms with 1 a rationally independent
set. We may assume that $\{\g_1,\ldots,\g_\mu\}$ is this maximal subset. Then, by the assumption of
the theorem, $1,\a,\g_1,\ldots,\g_\mu$ are rationally independent, furthermore every $\g_j$ with $j>\mu$ is a rational
combination of the form (\ref{tt00}).
For a large $m$ partition $[0,1]^\mu$ by the hyperplanes $x_i=t/m$, $t=1,\ldots,m-1$ in $\R^\mu$ into $m^\mu$
little cubes $I_s$, $1\le s\le m^\mu$, as before. The image of such an $I_s$ under the mapping (\ref{tt100})
(cf. (\ref{tt00})) is again an interval $I_{j,s}$ of length $\le C/m$.
Now for fixed $0<a<b<1$ and for an $0<\e<\min\{a,1-b\}$ let $m$ be so large that the oscillation of
$G$ over any set of diameter $\le C\nu/m$ is smaller than $\e/2$. Let furthermore, $Q_s=(Q_{s,1},\ldots,Q_{s,\mu})$ be an
arbitrary point from $I_s$, let $\tilde Q_{j,s}$ be the image of $Q_s$ under the mapping (\ref{tt100}),
and set
\[T_s=G(\g_1 Q_{s,1},\ldots,\g_\mu Q_{s,\mu},\g_{\mu+1}\tilde Q_{\mu+1,s},\ldots, \g_{\nu}\tilde Q_{\nu,s}).\]
With these modifications for the sets
\[K_{1,s}=\left\{1\le k\le n,\ k\in{\cal N}_r\sep (k\g_1,\ldots,k\g_\mu)\in I_s,\
k\a\in [a-T_s+\e,b-T_s-\e]\right\},\]
and
\[K_{2,s}=\left\{1\le k\le n,\ k\in{\cal N}_r\sep (k\g_1,\ldots,k\g_\mu)\in I_s,\
k\a\in [a-T_s-\e,b-T_s+\e]\right\},\]
(where the relations are understood (mod 1)) we get again that for $k\in K_{1,s}$ we have
$k\a+G(\g_1 k,\cdots ,\g_\nu k)\in [a,b]$ (mod 1), and conversely, if
$(k\g_1,\ldots,k\g_\mu)\in I_s$ and
$k\a+G(\g_1 k,\cdots,\g_\nu k)\in [a,b]$ (mod 1), then
$k\in K_{2,s}$.
By the (mod 1) uniform distribution of the sequence of vectors $\{(k\a,k\g_1,\ldots,k\g_\mu)\}_{k=1,\;k\in {\cal N}_r}^\i$
(which is a consequence of the rational independence of $(1,\a,\g_1,\ldots,\g_\mu)$), we get again
\[\#K_{1,s}=(1+o(1)n\frac{b-a-2\e}{q\cdot m^\mu},\qquad \#K_{2,s}=(1+o(1))n\frac{b-a+2\e}{q\cdot m^\mu},\]
and from here the uniform distribution of
\[\{k\a+G(k\g_1,\cdots ,k\g_\nu)\}_{k=1,\;k\in{\cal N}_r}^\i\]
(mod 1) for each $0\le r<q$ follows as before, and that proves
Proposition \ref{th31000}.\endproof
\subsection{The polynomial case}
Now we extend Proposition \ref{th31000} to the polynomial case. As before, we may
assume without loss of generality that the polynomials we are dealing with are without constant term.
\begin{prop} \label{th312} Let $P_0,P_1,\ldots,P_d$ be polynomials without constant term and
let $F$ be a continuous
function of the variables $x_1,\ldots,x_\nu$ which is periodic in each variable with periods
$\b_1,\ldots,\b_\nu$, respectively. If $P_0$ cannot be written as a rational linear combination of the polynomials
$P_1/\b_1,\ldots,P_d/\b_d$ and of a polynomial with rational coefficients, then
the sequence
\be \{P_0(n)+F(P_1(n),\cdots,P_\nu(n))\}_{n=1}^\i\label{subs}\ee
is uniformly distributed {\rm (mod 1)}.
\end{prop}
Note that Proposition \ref{th31000} is the special case of this when $P_0(x)=\a x$ and
$P_j(x)=x$ for all other $j$.
In particular, if
\[F(x_1,\ldots,x_\nu)=f_1(x_1)+\cdots+f_\nu(x_\nu),\]
where each $f_j$ is a continuous periodic function
of period $\b_j$, then under the conditions of the theorem the (mod 1) density of the sequence
\[\{P_0(n)+f_1(P_1(n))+\cdots+f_\nu(P_\nu(n))\}_{n=1}^\i\]
follows.
In Proposition \ref{th41} we shall show that this is best possible in the sense that if $P_0,P_1,\ldots,P_\nu$
are given and $P_0$ can be written as a rational linear combination of the polynomials
$P_1/\b_1,\ldots,P_\nu/\b_\nu$ and of a polynomial with rational coefficients, then there are
continuous functions $f_j$ of period $\b_j$, $1\le j\le \nu$, such that
the sequence
\[P_0(n)+f_1(P_1(n))+\cdots f_d(P_\nu(n)), \qquad n=1,2,\ldots,\]
is not
uniformly distributed (mod 1).
\proof. The proof is almost identical to that of Proposition \ref{th31}. Indeed, as there
set $\g_i=1/\b_i$ and
\[G(x_1,\ldots,x_\nu)=F(\b_1 x_1,\ldots,\b_\nu x_\nu).\]
Then $F(P_1(n),\ldots,P_\nu(n))=G(\g_1 P_1(n),\ldots,\g_\nu P_\nu(n))$.
Select a maximal subset of $\g_1P_1,\ldots,\g_\nu P_\nu$ for which no non-trivial linear combination with integer coefficients produces
a polynomial with integer coefficients. We may assume that $\g_1P_1,\ldots,\g_\mu P_\mu$ is this maximal subset.
Then, by the assumption of
the theorem, $P_0,\g_1P_1,\ldots,\g_\mu P_\mu$ is $\Q$-independent (see Definition \ref{def1}), but every $\g_jP_j$ with $j>\nu$ is a
combination of the form
\be \g_jP_j(x)=R_j(x)+\sum_{i=1}^\mu c_{j,i}\g_iP_i(x),\label{tt11}\ee
where the $c_{j,i}$ are rationals and $R_j$ has rational coefficients.
Let $q$ be a common multiple of the denominators of all the coefficients
in all $R_j$, $\mu< j\le \nu$. It is sufficient to verify the uniform
distribution of each subsequence of (\ref{subs}) for which the indices $n$
belong to ${\cal N}_r=\{ql+r\}_{l=1}^\i$ with a fixed $0\le r<q$. If
$k=ql+r$ is such an index, then
\[ \g_jP_j(k)=R_j(r)+\sum_{i=1}^\mu c_{j,i}\g_iP_i(k)\qquad \mbox{(mod 1)},\]
and on the right $R_j(r)$ is independent of $l=1,2,\ldots$. This $R_j(r)$
plays the role of $rd_j$ in (\ref{tt100}).
Since
$P_0,\g_1P_1,\ldots,\g_\mu P_\mu$ are $\Q$-independent, it follows that the polynomials
$P_0(qx+r),\g_1P_1(qx+r)\ldots,\g_\mu P_\mu(qx+r)$ are also $Q$-independent (note that if a polynomial
$R$ has rational coefficients then so does $R((y-r)/q)$). Hence Weyl's theorem ensures that
the sequence
\[(P_0(k),\g_1P_1(k),\ldots,\g_\mu P_\mu(k)),\qquad k=ql+r,\ l=1,2,\ldots,\]
of vectors is uniformly distributed (mod 1) on $[0,1)^{\nu+1}$.
This is the analogue of the uniform distribution of the sequence
$\{(k\a,n\g_1,\ldots,k\g_\mu)\}_{k=1,\;k\in{\cal N}_r}^\i$ in the preceding proof
(while (\ref{tt11}) is the analogue of (\ref{tt00})), and by replacing the
latter by the previous one, that proof goes through without much change. \endproof
\subsection{The vector case}
Now we shall consider the vector case, namely the uniform distribution
of vector sequences of the form
\be \Bigl\{\bigl(n\a_1+f_1(n), \ldots,n\a_d+f_d(n)\bigr)\Bigr\}_{n=1}^\i\label{vvec}\ee
(mod 1), where the $f_i$ are continuous periodic functions. To this end we introduce the following
definition.
\begin{definition} {\rm We say that a system $(\a_1,\b_1),\ldots,(\a_d,\b_d)$ of number pairs is totally
$\Q$-independent, if for any non-empty subset $\L\subset \{1,\ldots,d\}$
it is true, that if $a_\l$, $\l\in \L$, are {\it nonzero} integers, then
the number $\sum_{\l\in \L} a_\l \a_\l$ cannot be written as a rational linear combination of the numbers
$1,\b_\l$, $\l\in \L$.}
\end{definition}
As an example consider the pairs
$(\sqrt 2,\sqrt 3)$, $(\sqrt 3,\pi)$, $(\pi^2,\sqrt 2)$. This is easily seen
to be totally $\Q$-independent, though $\a_1+\a_2=\sqrt 2+\sqrt 3=\b_1+\b_3$ (note that total $\Q$-independence requires
that $\a_1+\a_2=\sqrt 2+\sqrt 3$ should not be a rational
linear combination of $\b_1=\sqrt 2$ and $\b_2=\pi$ -- we cannot
use in this combination $\b_3$).
\begin{prop} \label{th311} Let $f_1,\ldots,f_d$ be continuous periodic functions with periods
$\b_1,\ldots,\b_d$. If $(\a_1,1/\b_1),\ldots,(\a_d,1/\b_d)$ are totally $\Q$-independent,
then
the sequence {\rm (\ref{vvec})} is uniformly distributed on $[0,1)^d$ {\rm (mod 1)}.
\end{prop}
In Proposition \ref{Th42} we shall show that one cannot do more than Proposition \ref{th311},
for uniform distribution is not true
(for some functions $f_j$ of period $\b_j$) if
$(\a_1,1/\b_1),\ldots,(\a_d,1/\b_d)$ are not totally $\Q$-independent.
\proof. By Weyl's criterion, a sequence
\be (v_{1,n},\ldots,v_{d,n}),\qquad n=1,2,\ldots,\label{df}\ee
of vectors is uniformly distributed on $[0,1)^d$ (mod 1) precisely if for all integers $a_1,\ldots,a_d$, $\sum a_j^2>0$,
the sequence of numbers
\be a_1v_{1,n}+\cdots+a_nv_{d,n},\qquad n=1,2,\ldots,\label{df1}\ee
is uniformly distributed on $[0,1)$ (mod 1). Apply this to
$v_{j,n}=\a_j n+f_j(n)$. If in the combination
the non-zero $a_j$ are $a_\l$, $\l\in \L$, then we need to prove the (mod 1) uniform
distribution on $[0,1)$ of the sequence
\[\left(\sum_{\l\in \L} a_\l \a_\l\right) n+\sum_{\l\in \L} a_\l f_\l(n),\qquad n=1,2,\ldots,\]
which follows from Proposition \ref{th31} because, by total $\Q$-independence,
$\sum_{\l\in \L} a_\l \a_\l$ cannot be written as a rational linear combination of the numbers
$1, 1/\b_\l$, $\l\in \L$.\endproof
We can state a more general form of the preceding proposition. To this end we introduce the following
definition.
\begin{definition} {\rm We say that a system
\[(\a_1,\b_{1,1},\ldots,\b_{1,\nu}),\ldots,(\a_d,\b_{d,1},\ldots,\b_{d,\nu})\]
of $(\nu+1)$-tuples of numbers is totally
$\Q$-independent, if for any non-empty subset $\L\subset \{1,\ldots,d\}$
it is true, that if $a_\l$, $\l\in \L$, are {\it nonzero} integers, then
the number $\sum_{\l\in \L} a_\l \a_\l$ cannot be written as a rational linear combination of the numbers
$1,\b_{\l,j}$, $\l\in \L,1\le j\le \nu$.}
\end{definition}
\begin{prop} \label{th311*} Let $F_1,\ldots,F_d$ be continuous periodic functions of $\nu$ variables
such that $F_i$ is periodic in the $j$-th variable with period
$\b_{i,j}$. If
\[(\a_1,1/\b_{1,1},\ldots,1/\b_{1,\nu}),\ldots,(\a_d,1/\b_{d,1},\ldots,1/\b_{d,\nu})\]
are totally $\Q$-independent, then
the sequence
\be \Bigl\{\bigl(n\a_1+F_1(n,\ldots,n), \ldots,n\a_d+F_d(n,\ldots,n)\bigr)\Bigr\}_{n=1}^\i\label{vvecv}\ee
of vectors is uniformly distributed on $[0,1)^d$ {\rm (mod 1)}.
\end{prop}
\proof. Just follow the proof of Proposition \ref{th311}. In this case we apply
Weyl's criterion to a linear combination $\sum_j a_j v_{j,n}$ with
$v_{j,n}=\a_j n+F_j(n,\ldots,n)$. If in the combination
the non-zero $a_j$ are $a_\l$, $\l\in \L$, then we need to prove the (mod 1) uniform
distribution on $[0,1)$ of the sequence
\[\left(\sum_{\l\in \L} a_\l \a_\l\right) n+\sum_{\l\in \L} a_\l F_\l(n,\ldots,n),\qquad n=1,2,\ldots,\]
which follows from Proposition \ref{th31000} because, by total $\Q$-independence,
$\sum_{\l\in \L} a_\l \a_\l$ cannot be written as a rational linear combination of the numbers
$1,\ 1/\b_{\l,j}, \l\in \L,\ 1\le j\le \nu$. Indeed, if, say, $\L=\{1,\ldots,\k\}$ (which
we may assume), then one should replace $\nu$ in Proposition \ref{th31000} by $\k\nu$ and
apply Proposition \ref{th31000} to
the function
\[F(x_1,\ldots,x_{k\nu})=F_1(x_1,\ldots,x_\nu)+F_2(x_{\nu+1},\ldots,x_{2\nu})+
\cdots+F_k(x_{(k-1)\nu+1},\ldots,x_{k\nu})\]
which is $\b_{i,j}$-periodic in the variable $x_{(i-1)\nu+j}$.
\endproof
The polynomial version of Proposition \ref{th311} is our next result.
\begin{definition} {\rm We say that a system
\[(P_1,P_{1,1},\ldots,P_{1,\nu}),\ldots,(P_d,P_{d,1},\ldots,P_{d,\nu})\]
of polynomial $(\nu+1)$-tuples is totally
$\Q$-independent, if for any non-empty subset $\L\subset \{1,\ldots,d\}$
it is true, that if $a_\l$, $\l\in \L$ are {\it nonzero} integers, then
the polynomial $\sum_{\l\in \L} a_\l P_\l$ cannot be written as
a linear combination with rational coefficients of the polynomials
$P_{\l,j}$, $\l\in \L,1\le j\le\nu$, and of a polynomial with integer coefficients.}
\end{definition}
\
The most general result on distribution in this paper is the following.
\begin{th} \label{th3110} Let $F_1,\ldots,F_d$ be continuous functions
of $\nu$ variables such that each $F_i$ is periodic in each of its variables with
periods
$\b_{i,1},\ldots,\b_{i,\nu}$, respectively. Let furthermore, $P_1,\ldots,P_d$ and $P_{i,j}$, $1\le i\le d$, $1\le j\le \nu$,
be polynomials without constant term.
If the system
\[(P_1,P_{1,1}/\b_{1,1},\ldots,P_{1,\nu}/\b_{1,\nu}),\ldots,(P_d,P_{d,1}/\b_{d,1},\ldots,P_{d,\nu}/\b_{d,\nu})\]
is totally $\Q$-independent, then
the sequence
\be \Bigl(P_1(n)+F_1\bigl(P_{1,1}(n)), \ldots,P_{1,\nu}(n)\bigr),\ldots,
P_d(n)+F_d\bigl(P_{d,1}(n),\ldots,P_{d\nu}(n)\bigr)\Bigr),\label{vvec1}\ee
$n=1,2,\ldots,$ of vectors is uniformly distributed on $[0,1)^d$ {\rm (mod 1)}.
\end{th}
See Proposition \ref{Th43} for the sharpness of this theorem.
\proof. The claim follows from Proposition \ref{th312} exactly as Proposition \ref{th311} resp., \ref{th311*} followed
from Proposition \ref{th31} resp., Proposition \ref{th311}, just use again that uniform distribution of a sequence (\ref{df}) is equivalent to
the uniform distribution of all the sequences (\ref{df1}), and apply Proposition \ref{th312}.\endproof
\sect{Exactness of the conditions for uniform distribution}
In this section we show that the conditions for uniform distribution set forth in the
preceding section are exact. Let us start by showing that the conditions in Proposition \ref{th3}
and Proposition \ref{th31} cannot be relaxed.
\begin{prop}\label{th4} If $\a$ and $\b_1,\ldots,\b_\nu$ are given and
$\a$ can be written as a
rational linear combination of the numbers
$1,1/\b_1,\ldots,1/\b_\nu$, then there are $\b_j$-periodic continuous functions $f_j$ for $j=1,\ldots,\nu$,
such that the sequence $n\a+f_1(n)+\cdots +f_\nu(n)$, $n=1,2,\ldots$,
is not uniformly distributed {\rm (mod 1)}.
\end{prop}
This proposition does not tell us if, in the case when uniform distribution does not happen,
the sequence has a distribution or not. We shall briefly discuss that question in the next section.
\proof Proposition \ref{th4}. First of all, we may assume that $\a$ is irrational, for
otherwise the claim is trivial by setting $f_j\equiv 0$.
Let $\g_j=1/\b_j$. By the assumption there are integers $q>0$
and $p_0,p_1\ldots,p_d$ such that
\be \a+\sum_{j=1}^\nu \frac{p_j}{q}\g_j=\frac{p_0}{q}.\label{uu}\ee
Multiply this equality by $n$ and write each number on the left as its integral and fractional part
to obtain
\[[n\a]+\{n\a\}+\sum_{j=1}^\nu \frac{p_j}{q}([n\g_j]+\{n\g_j\})=\frac{p_0}{q},\]
from which we can see that
\[\{n\a\}+\sum_{j=1}^\nu \frac{p_j}{q}\{n\g_j\}=\frac{s}{q},\]
where $s$ is an integer. Since on the left we have a linear combination of
numbers lying in $[0,1)$, actually
\[|s|<|q|+|p_1|+\cdots+|p_\nu|=:M.\]
Therefore,
for every $N$ there is an integer $t\in [-M,M]$ such that
for at least $N/(2M+1)$ of the numbers $1\le n\le N$ we have
\be \{n\a\}+\sum_{j=1}^\nu \frac{p_j}{q}\{n\g_j\}=\frac{t}{q},\label{hh}\ee
and we may assume that $t$ is the same for infinitely many $N$.
Let $\e>0$ be a small number. Since $\a$ is irrational, by the uniform distribution
of the sequence $\{n\a\}$, $n=1,2,\ldots$, for large $N$ at least $(1-2\e)N$ of the
numbers $\{n\a\}$, $n=1,2,\ldots,N$, belong to $[0,1-\e]$. In a similar vein, if for some
$j$ the number $\g_j$ is irrational, then for large $N$ at least $(1-2\e)N$ of the
numbers $\{n\g_j\}$, $n=1,2,\ldots,N$, belong to $[0,1-\e]$. On the other hand,
if $\g_j$ is rational, then for small $\e>0$ all the
numbers $\{n\g_j\}$, $n=1,2,\ldots,N$, belong to $[0,1-\e]$.
Thus, for at least $(1-2(\nu+1)\e)N$ of the numbers $1\le n\le N$ these relations
hold simultaneously. Choose $N$ sufficiently large so that all these are satisfied.
If, in addition, $2(\nu+1)\e<1/2(2M+1)$, then aut of these
$\ge (1-2(\nu+1)\e)N$ numbers $1\le n\le N$ at least $N/2(2M+1)$ also satisfy (\ref{hh})
(recall that (\ref{hh}) was true for at least $N/(2M+1)$ of the $1\le n\le N$).
Let now $g_j$, $j=1,\ldots,\nu$, be 1-periodic continuous functions such that
\[g_j(x)=\frac{p_j}{q}x\qquad\mbox{for $x \in [0,1-\e]$}.\]
In view of (\ref{hh}), for at least $N/2(2M+1)$ of the numbers $1\le n\le N$
we have
\[n\a+\sum_{i=1}^d g_i(n\g_i)=\frac{t}{q}\qquad \mbox{(mod 1)},\]
and since this is true for infinitely many $N$, it follows that the numbers
on the left for $n=1,2,\ldots$ are not uniformly distributed (mod 1)
(the uniform distribution does not have point masses).
We can complete the proof by setting $f_j(x)=g_j(\g_j x)=g_j(x/\b_j)$\endproof
Next we show that the condition given in Proposition \ref{th312} is sharp.
\begin{prop} \label{th41} If the numbers $\b_1,\ldots,\b_\nu$ and the polynomials $P_0,P_1,\ldots,P_\nu$
are given and $P_0$ can be written as a rational linear combination of the polynomials
$P_1/\b_1,\ldots,P_\nu/\b_\nu$ and of a polynomial with rational coefficients, then for $j=1,\ldots,\nu$
there are periodic continuous functions $f_j$ of period $\b_j$ such that
sequence
\[P_0(n)+f_1(P_1(n))+\cdots +f_\nu(P_\nu(n)), \quad n=1,2,\ldots,\]
is not
uniformly distributed {\rm (mod 1)}.
\end{prop}
\proof. The proof is almost identical to that of Proposition \ref{th4}.
We may assume that $P_0$ has an irrational coefficient (otherwise just set
$f_j\equiv 0$).
Now start with a relation
\be P_0(x)+\sum_{j=1}^d \frac{p_j}{q}P_j(x)\g_j=\frac{p_0}{q}R_j(x),\label{GG}\ee
where $\g_j=1/\b_j$ and $R_j$ is a polynomial with integer coefficients.
This is the analogue of (\ref{uu}), and from here proceed as we reasoned from (\ref{uu}).
We only need to
mention that since $P_0$ has at least one irrational coefficient, by Weyl's theorem,
for large $N$, at least $(1-2\e)N$ of the
numbers $\{P_0(n)\}$, $n=1,2,\ldots,N$, belong to $[0,1-\e]$, and similarly, if for some
$j$ the polynomial $P_j\g_j$ has an irrational coefficient,
then for large $N$ at least $(1-2\e)N$ of the
numbers $\{P_j(n)\g_j\}$, $n=1,2,\ldots,N$, belong to $[0,1-\e]$. On the other hand,
if all coefficients of $P_j\g_j$ are rational, then for sufficiently small $\e>0$ all the
numbers $\{P_j(n)\g_j\}$, $n=1,2,\ldots,N$, belong to $[0,1-\e]$. \endproof
The conditions set forth in the vector case in Proposition \ref{th311}
are also best possible:
\begin{prop} \label{Th42} Let $\a_1,\ldots,\a_d$ and
$\b_1,\ldots,\b_d$ be given numbers.
If the pairs $(\a_1,1/\b_1),\ldots,(\a_d,1/\b_d)$ are not totally $\Q$-independent,
then there are $\b_j$-periodic continuous functions $f_j$ such that
the sequence {\rm (\ref{vvec})} is not uniformly distributed {\rm (mod 1)} on $[0,1)^d$.
\end{prop}
\proof. We may assume that $1,\a_1,\ldots,\a_d$ are rationally independent (if this is not the
case, then for $f_j\equiv 0$ the sequence (\ref{vvec}) is not uniformly distributed
by Weyl's theorem).
By assumption there is a non-empty set $\L\subset \{1,\ldots,d\}$ and
non-zero integers $a_\l$, $\l\in \L$, such that
\be \sum_{\l\in \L}a_\l\a_\l+\sum_{\l\in \L}\frac{p_\l}{q} \g_\l=\frac{p_0}{q}\label{HH}\ee
where $\g_\l=1/\b_\l$ and $p_\l,q,p_0$ are integers. This is the analogue of (\ref{hh}), and
if we follow the reasoning after (\ref{hh}) (note that we can, since $\sum_\l a_\l \a_\l$ is irrational by the
rational independence of $1,\a_1,\ldots,\a_d$), then we obtain that there are
continuous functions $h_\l$ of period $\b_\l$ such that the sequence
\[\left(\sum_{\l\in \L}a_\l\a_\l\right)n+\sum_{\l\in \L}h_\l(n),\qquad n=1,2,\ldots,\]
is not (mod 1) uniformly distributed. If we set $f_\l =h_\l/a_\l$, then it follows that
the numerical sequence
\[\left(\sum_{\l\in \L}a_\l\a_\l\right)n+\sum_{\l\in \L}a_\l f_\l(n),\qquad n=1,2,\ldots,\]
is not (mod 1) uniformly distributed on $[0,1)$. But then, by Weyl's theorem, the vector
sequence
(\ref{vvec}) (where, say, we set $f_j\equiv 0$ if $j\not\in \L$) is not uniformly
distributed (mod 1) in $[0,1)^d$, either.\endproof
Finally, we show that Theorem \ref{th3110} is sharp. For simplicity we shall formulate
the sharpness only for the case when the functions $F_i$ are of a single variable.
\begin{th} \label{Th43}. Let $\b_1,\ldots,\b_d$ be given numbers
and let $Q_1,\ldots,Q_d$ and $P_1,\ldots,P_d$ be given polynomials without constant term.
If the system of polynomial pairs $(Q_1,P_1/\b_1),\ldots,(Q_d,P_d/\b_d)$ is not totally
$\Q$-independent, then there are continuous periodic functions $f_j$ of period $\b_j$, $1\le j\le d$,
such that the sequence
\be \Bigl\{\bigl(Q_1(n)+f_1(P_1(n)), \ldots,Q_d(n)+f_d(P_d(n))\bigr)\Bigr\}_{n=1}^\i\label{vvec1**}\ee
is not uniformly distributed {\rm (mod 1)} on $[0,1)^d$.
\end{th}
\proof. We may assume that $Q_1,\ldots,Q_d$ are $\Q$-independent (otherwise
set $f_j\equiv 0$ and apply Weyl's theorem).
By assumption there is a non-empty set $\L\subset \{1,\ldots,d\}$ and
non-zero integers $a_\l$, $\l\in \L$, such that with $\g_\l=1/\b_\l$ we have
\[\sum_{\l\in \L}a_\l Q_\l(x)+\sum_{\l\in \L}\frac{p_\l}{q} P_\l(x)\g_\l=\frac{p_0}{q}R(x),\]
where $p_0, p_\l,q$ are integers and the polynomial $R$ has integer coefficients.
This is now the analogue of (\ref{HH}) and (\ref{GG}), and from here the
reasoning is the same that was given after (\ref{HH}) (taking into account
the necessary modifications in the polynomial case as were given after (\ref{GG})
and taking also into account that $\sum a_\l Q_\a$ has
an irrational coefficient since $Q_1,\ldots,Q_d$ are $\Q$-independent):
by Proposition \ref{th41} there are
continuous functions $h_\l$ of period $\b_\l$ such that the sequence
\[\left(\sum_{\l\in \L}a_\l Q_\l(n)\right)+\sum_{\l\in \L}h_\l(P_\l(n)),\qquad n=1,2,\ldots,\]
is not (mod 1) uniformly distributed. Now set $f_\l =h_\l/a_\l$, and complete the proof
as in Proposition \ref{Th42}.\endproof
\sect{Density and distribution on the torus}\label{secttorus}
In this section we change somewhat our perspective. So far we considered modulo 1 values
as fractional parts on the interval $[0,1)$. But we can consider them also on
the torus $\R/\Z$, i.e., when we identify the points 0 and 1 in $[0,1]$.
All our results so far
on the (mod 1) density and uniform distribution hold also on the torus without any change. Until now we
were considering sequences (or vector sequences) of the form
\be P_0(0)+F(P_1(n),\ldots,P_\nu(n)),\quad n=1,2,\ldots,\label{tor1}\ee
where, in some cases, $F$ was allowed to be almost periodic. If we only restrict
our attention only to periodic $F$, then it is natural to work with mappings
$G:(\R/\Z)^\nu\to \R/\Z$ (or with
$G:(\R/\Z)^\nu\to (\R/\Z)^d$ in the vector case) and consider
sequences
\be G(P_1(n),\ldots,P_\nu(n)),\quad n=1,2,\ldots\label{tor2}\ee
on $\R/\Z$ (or on $(\R/\Z)^d$ in the vector case), which leads to a somewhat more
general setting.
Indeed, if the $F$ in (\ref{tor1}) is periodic in each variable with periods
$\b_1,\ldots,\b_\nu$, respectively, then by setting
\be G(x_0,x_1,\ldots,x_\nu)=x_0+F(\b_1x_1,\ldots,\b_\nu x_\nu),\quad (x_0,x_1,\ldots,x_\nu)\in (\R/\Z)^{\nu+1},
\label{tor3}\ee
and $P_j^*(x)=P_j(x)/\b_j$, we have
\[P_0(0)+F(P_1(n),\ldots,P_\nu(n))=G(P_0(n),P_1^*(n),\ldots,P_\nu^*(n))\qquad \mbox{(mod 1)},\]
so (\ref{tor2}) include, indeed, (mod 1) sequences of the form (\ref{tor1}).
Any function $G:(\R/\Z)^\nu\to \R/\Z$ can also be considered as a function $G:\R^\nu\to \R/\Z$ by stipulating
$G(x_1,\ldots,x_\nu):=G(\{x_1\},\ldots,\{x_\nu\})$, and this is indeed how we interpreted the values in (\ref{tor2}).
The torus $\R/\Z$ can also be identified with the unit circle under the mapping $x\to e^{2\pi i x}$.
This allows us to speak of the rotation (or winding) number $w$ of a continuous mapping $g:\R/\Z\to \R/\Z$ which
is defined
as the total change of the argument in $e^{2\pi i g(x)}$ as $x$ runs trough the interval $[0,1]$. Intuitively,
$w$ tells us how many times $e^{2\pi i g(x)}$ circles the origin as $e^{2\pi i x}$ makes one full circle
in the positive (counterclockwise) direction. For example, if $m$ is an integer, then the mapping
$x\to mx$ has rotation number $m$, and all $g:\R/\Z\to \R/\Z$ with rotation number $m$
is homotope with the mapping $x\to mx$.
If $G:=(\R/\Z)^\nu\to \R/\Z$ is continuous and we fix $x_2,\ldots,x_\nu$, then $g(x_1)=G(x_1,x_2,\ldots,x_\nu)$ is a mapping
from $\R/\Z$ into itself, and let $w_1$ be its rotation number. Since this is an integer which changes
continuously as we change $x_2,\ldots,x_\nu$ continuously, it follows that $w_1$ is independent of how $x_2,\ldots,x_\nu$ are
fixed, and we call $w_1$ the rotation number of $G$ with respect to $x_1$. The rotation number $w_j$ with respect
the variable $x_j$. $2\le j\le \nu$, is defined similarly. For example, the mapping $G$ in (\ref{tor3})
has rotation number $w_0=1$ with respect to the variable $x_0$ and (because of the periodicity of the function $F$)
rotation number 0 ($w_j=0$) with respect to all other variables $x_j$, $1\le j\le \nu$.
\subsection{Density on the torus}
With the just given definitions we are now ready to extend Proposition \ref{th2pol1} (at least for periodic functions).
\begin{prop} \label{th2pol1tor} Let $G:=(\R/\Z)^\nu\to \R/\Z$ be a continuous
function with rotation numbers $w_1,\ldots,w_\nu$, respectively, and
let $P_1,\ldots,P_\nu$ be arbitrary polynomials without constant terms. If $w_1P_1+\cdots+w_\nu P_\nu$ has at least
one irrational coefficient, then the sequence
\be \Bigl\{G(P_1(n),\ldots,P_\nu(n))\Bigr\}_{n=1}^\i,\label{001tor}\ee
is dense in $\rz$.
\end{prop}
Note that this is exact in the following sense: if all coefficients of
$w_1P_1+\cdots+w_\nu P_\nu$ are rational, then for
\[G(x_1,\ldots,x_\nu)=w_1x_1+\cdots+w_\nu x_\nu\]
(which clearly has rotation numbers $w_1,\ldots,w_\nu$), the sequence in (\ref{001tor})
is not dense in $\rz$ (for it has only finitely many different terms).
\proof Proposition \ref{th2pol1tor}. The claim easily follows from Proposition \ref{th2pol1}.
In fact, let
\[F(x_1,\ldots,x_\nu)=G(x_1,\ldots,x_\nu)-(w_1x_1+\cdots w_\nu x_\nu).\]
Since all rotation numbers of this $F$ are zero, $F$ (considered as a function from $\R^\nu$ to
$\R$) is a 1-periodic function in each of its variable. Furthermore,
if we set
\[P_0:=w_1P_1+\cdots+w_\nu P_\nu,\]
then for all $n$ we have
\[G(P_1(n),\ldots,P_\nu(n))=P_0(n)+F(P_1(n),\ldots,P_\nu(n)),\]
hence the claim in the proposition is a consequence of Proposition
\ref{th2pol1}.\endproof
In a similar vein can one obtain the vector case.
\begin{th}\label{thmaintor} Let $G_i(x_{i,1},\ldots x_{i,\nu})$, $1\le i\le d$, be
continuous functions of $\nu$ variables on $\rz$, and let
$P_{i,j}$, $1\le j\le \nu$, $1\le i\le d$, be polynomials without constant term.
If the rotation number of $G_i$ with respect to the variable
$x_{i,j}$ is $w_{i,j}$ and if the polynomials
\be P_i=w_{i,1}P_{i,1}+\cdots w_{i,\nu}P_{i,\nu},\qquad i=1,\ldots,d,\label{tor4}\ee
are ${\bf Q}$-independent, then the sequence of vectors
\be \Bigl(G_i\bigl(P_{i,1}(n), \ldots,P_{i,\nu}(n)\bigr)\Bigr)_{i=1}^d,\quad n=1,2,\ldots
\label{tor5}\ee
is dense in $(\rz)^d$.
\end{th}
This is exact again, for if the polynomials (\ref{tor4}) are not ${\bf Q}$-independent, i.e.
some linear combination of them with rational coefficients has rational coefficients,
then for
\[G_i(x_{i,1},\ldots,x_{i,\nu})=w_{i,1}x_{i,1}+\cdots w_{i,\nu} x_{i,\nu},\qquad i=1,\ldots,d,\]
the sequence (\ref{tor5}) is not dense in $(\rz)^d$ by Weyl's theorem.
\proof. If we set
\[F_i(x_{i,1},\ldots,x_{i,\nu})=G_i(x_{i,1},\ldots,x_{i,\nu})-(w_{i,1}x_{i,1}+\cdots w_{i,\nu} x_{i,\nu}),\]
then the claim follows from Theorem \ref{thmain} exactly as we deduced Proposition \ref{th2pol1tor}
from Proposition \ref{th2pol1}.\endproof
\subsection{Existence of distribution}
We have seen that when considering the Weyl theory with periodic perturbations, we do not always get
uniform distribution. This raises the question if in those cases the distribution of the sequences
in question exist at all. We shall see that the answer depends on if we work on the torus or on
$[0,1)$ with fractional parts.
Let $X_1,X_2,\ldots$ be a sequence on the torus $({\rz})^d$. We say that this sequence has
distribution $\s$ if for every continuous function $f:(\rz)^d\to \R$ we have
\be \lim_{n\to\i}\frac{1}{n}\sum_{k=1}^n f(X_k)=\int f\;d\s.\label{int1}\ee
An equivalent formulation is that for a dense set\footnote{In fact, for all $a_i,b_i$
for which the boundary of the box $\prod_{i=1}^d[a_i,b_i]$ has zero $\s$-measure.} of $0< a_i<b_i<1$, $1\le i\le d$,
we have
\be \lim_{n\to\i}\#\left\{1\le k\le n\sep X_k\in \prod_{i=1}^d[a_i,b_i]\right\}=\s\left(\prod_{i=1}[a_i,b_i]\right).\ee
Uniform distribution is when $\s$ is the Haar measure on $(\rz)^d$ (which is the $d$-fold product of
the normalized arc measure on the unit circle). In proving the existence of such a $\s$ it
is sufficient to show that the limit
\[L_f:=\lim_{n\to\i}\frac{1}{n}\sum_{k=1}^n f(X_k)\]
exists for all such $f$, for then $L_f$ is clearly a positive linear functional
and the Riesz representation theorem (cf. \cite[Theorem 2.14]{Rudin}) gives (\ref{int1}) with some measure $\s$.
Now on the torus the existence of the distribution for polynomial sequences holds in all situations.
\begin{th} \label{thtor2} Let $G_1,\ldots,G_d$ be continuous mappings of $(\rz)^\nu$ into $\rz$,
and let $P_{i,j}$, $1\le i\le d$, $1\le j\le \nu$,
be arbitrary polynomials. Then the sequence
\[\Bigl(G_1\bigl(P_{1,1}(n)), \ldots,P_{1,\nu}(n)\bigr),\ldots,
G_d\bigl(P_{d,1}(n),\ldots,P_{d,\nu}(n)\bigr)\Bigr)_{n=1}^\i\]
of vectors has a distribution.
\end{th}
\proof. By introducing $(d-1)\nu$ dummy variables in each $G_i$ (and hence
replacing $\nu$ by $\nu d$) and by listing
$P_{i,j}$ into a single sequence $P_j$ it is sufficient to consider the case when $P_{i,j}$
does not depend on $i$, i.e., it is sufficient to consider sequences
of the form
\[\Bigl(G_1\bigl(P_1(n)), \ldots,P_\nu(n)\bigr),\ldots,
G_d\bigl(P_1(n),\ldots,P_\nu(n)\bigr)\Bigr)_{n=1}^\i.\]
As before, we may assume that each $P_j$ is without a constant term.
Choose a maximal ${\bf Q}$-independent subset of $P_1,\ldots,P_\nu$. We may assume
that this is $P_1,\ldots,P_\mu$ for some $\mu$, and first consider the case when
$\mu\ge 1$. Every $P_j$, $\mu<j\le \nu$, can be expressed as
\[P_j=R_j+\sum_{s=1}^\mu c_{s,j}P_s,\]
where the $c_{s,j}$ and the coefficients of the polynomial $R_j$ are rational numbers.
Choose an integer $q>0$ such that every $qc_{s,j}$ is an integer, and all coefficients of
all $qR_j$
are integers. It is sufficient to show that the distribution of every subsequence
\[\Bigl(G_1\bigl(P_1(k)), \ldots,P_\nu(k)\bigr),\ldots,
G_d\bigl(P_1(k),\ldots,P_\nu(k)\bigr)\Bigr)_{k=1.\;k\in {\cal N}_r}^\i\]
exists, where ${\cal N}_r=\{k\sep k=ql+r\}_{l=1}^\i$ and $0\le r<q$ is a fixed integer.
If $k\in {\cal N}_r$, then
\[P_j(k)=R_j(r)+\sum_{s=1}^\mu c_{s,j}P_s(k)\qquad \mbox{(mod 1)},\]
so if
\[G_i^*(x_1,\ldots,x_\mu)=G_i\Bigl(x_1,\ldots,x_\mu, R_{\mu+1}(r)+\sum_{s=1}^\mu c_{s,\mu+1}x_s,\ldots,
R_{\nu}(r)+\sum_{s=1}^\mu c_{s,\nu}x_s\Bigr),\]
then
\[ G_i(P_1(k)), \ldots,P_\nu(k))=G_i^*(P_1(k)), \ldots,P_\mu(k))\]
for all $i$.
Now if $f$ is any continuous function on $(\rz)^d$ and we set
\[f^*(x_1,\ldots,x_\mu)=f(G_1^*(x_1,\ldots,x_\mu),\ldots,G_d^*(x_1,\ldots,x_\mu)),\]
then we have to show the existence of the limit
\[\lim_{n\to\i}\frac{q}{n}\sum_{k=1,\; k\in {\cal N}_r}^n f^*(P_1(k),\ldots,P_\mu(k)).\]
Since, by Weyl's theorem, $\{(P_1(k),\ldots,P_\mu(k))\}_{k=1,\; k\in {\cal N}_r}^\i$
is uniformly distributed in $(\rz)^\mu$ (a consequence of the ${\bf Q}$-independence of
the polynomials $P_1,\ldots,P_\mu$\footnote{Note that
any ${\bf Q}$-dependence of $P_j(qx +r)$, $1\le j\le \mu$, would automatically extend to a
${\bf Q}$-dependence of $P_1,\ldots,P_\mu$}), the preceding limit exists (and is actually
$\int f^*d\s$ where $\s$ is the Haar measure on $(\rz)^\mu$).
If $\mu=0$, i.e., when all $P_i$ have rational coefficients, then for a $q$ for which all $qP_j$ have integer
coefficients and for $k=ql+r$ we have $P_j(k)=P_j(r)$ (mod 1), i.e.
$(P_1(k),\ldots,P_\nu(k))$ is a constant vector in $(\rz)^\nu$ for $k\in {\cal N}_r$, from which the
claim in the theorem immediately follows.\endproof
Finally, we show that the situation is different if we consider distribution not on the torus $\rz$, but the distribution of the
fractional parts (that belong to $[0,1)$. The following proposition shows that in this case the distribution
may not exist in very simple situations.
\begin{prop}\label{prop} Let $\a$ be irrational. Then there is a $1/\a$-periodic continuous function $f$
such that the distribution of the fractional parts $\{n\a+f(n)\}$, $n=1,2,\ldots$, does not exist.
\end{prop}
The non-existence of the distribution in this case will mean that for all $b\in [7/8,15/16]$ the limit
\[\lim_{n\to\i}\frac{1}{n}\#\{{1\le k\le n\sep \{k\a+f(k)\}\in [0,b]}\}\]
does not exist.
Recall, however, that if $f$ is $\b$-periodic for some irrational $\b$ and $1,\a$ and $\b$ are rationally independent, then
this phenomenon cannot happen since then the fractional parts $\{n\a+f(n)\}$, $n=1,2,\ldots$, are uniformly distributed
(see Proposition \ref{th3}).
\proof Proposition \ref{prop}. It will be sufficient to construct a 1-periodic function $h$
for which the sequence
$\{n\a+h(n\a)\}$, $n=1,2,\ldots$, of fractional parts has no distribution on $[0,1)$,
for then $f(x):=h(\a x)$ is suitable in Proposition \ref{prop}.
Let $x_m=\{m\a\}$, and $S=\{x_m\}_{m=1}^\i$.
Let $h_0$ be the 1-periodic continuous function
for which
\[h_0(x)=\left\{
\begin{array}{lll}
-x & \mbox{$x\in [0,1/2]$}, \\
-1/2+2(x-1/2) & \mbox{$x\in (1/2,3/4]$},\\
0&\mbox{$x\in (3/4,1)$}.
\end{array}\right.\]
Then
\[x+h_0(x)=\left\{
\begin{array}{lll}
0 & \mbox{$x\in [0,1/2]$}, \\
3(x-1/2) & \mbox{$x\in (1/2,3/4]$},\\
x&\mbox{$x\in (3/4,1)$}.
\end{array}\right.\]
For $k=0,1,\ldots$ we shall define a 1-periodic continuous function $h_k$ and numbers $M_{k-1}<N_k<M_k$, where
$-1/16<h_{k+1}\le h_k$ will be true for all $k$, and $h_{k+1}$ will be obtained from $h_k$
by modifying $h_k$ on certain finitely many disjoint open intervals $(a_{k,s},b_{k,s})\subset(0,1/2)$, $s\in L_k$,
over which $h_k(x)=-x$. The modification will be done with the "triangle" function
\[\chi_{k,s}(x)=\left\{
\begin{array}{lll}
a_{k,s}-x & \mbox{$x\in (a_{k,s},c_{k,s}]$}, \\
x-b_{k,s} & \mbox{$x\in (c_{k,s},b_{k,s})$},\\
0&\mbox{elsewhere in $[0,1)$,}
\end{array}\right.\]
where $c_{k,s}$ is the midpoint of $(a_{k,s},b_{k,s})$, i.e. we set
\be h_{k+1}=h_k+\sum_{s\in L_k} \chi_{k,s}(x),\label{io}\ee
and of course continue this $h_{k+1}$ periodically with period 1.
We shall also require that
the endpoints $a_{k,s},b_{k,s}$ do not belong to the set $S$, and
$h_{k+1}(x)=h_0(x)$ outside a set consisting of finitely many closed intervals $\subset (0,1/2)$ of total length $<1/8$.
Note that with such a construction the functions $\{h_k\}_{k=1}^\i$
uniformly converge monotone
decreasingly to a continuous function $h$ such that
$x+h(x)=x$ on $[3/4,1)$, and
for $x\in [0,1)$ the fractional part $\{x+h(x)\}$ belongs to the interval
$(7/8,1)$ precisely if $x\in (7/8,1)$ or if $x+h(x)<0$, and in the
latter case the fractional part $\{x+h(x)\}$ actually belongs to $(15/16,1)$.
$h_0$ has already been given, and set $N_0=1$, $M_0=2$. Suppose $h_k,N_k,M_k$ have already
been defined with the property that $h_k(x)=h_0(x)$ outside a set of closed intervals $\subset(0,1/2)$
of total length $<1/8$, say of
total length $1/8-\d_k$.
The measure of the set
\be \{x\in [0,1)\sep \{x+h_k(x)\}\in (7/8,1)\}\label{set}\ee
is at most the measure of $(7/8,1)$ plus the measure of those $x\in [0,1/2]$ where $h_k(x)<h_0(x)$,
hence (\ref{set}) has measure $<1/4$.
Since $h_k$ is 1-periodic, we have
\[\{m\a+h_k(m\a)\}=\{x_m+h_k(x_m)\},\]
and we get from the uniform distribution of the sequence $\{x_m\}_{m=1}^\i$ in $[0,1)$
that there is an $N_{k+1}>M_k$ such that
\be \#\left\{1\le m\le N_{k+1}\sep x_m+h_k(x_m)\in (7/8,1)\ \ \mbox{(mod 1)}\right\}\le \frac{N_{k+1}}{4}.\label{kl}\ee
For the same reason there is an $M_{k+1}>N_{k+1}$ such that the set
\[L_k=\left\{N_{k+1}< m\le M_{k+1}\sep x_m\in (0,1/2), \ x_m+h_k(x_m)=0\right\}\]
has at least
\be \#L_k\ge \left(\frac12-\frac18\right)M_{k+1}\label{13}\ee
elements. For each such $x_m$ we have $h_k(x)=-x$ in a neighborhood of $x_m$ (this follows from
the fact that $a_{l,s},b_{l,s}\not\in S$ for $l<k$ by the induction hypothesis and by the fact
that $h_k(x)<-x$ only on the unions of the intervals $(a_{l,s},b_{l,s})$ with $l<k$), so we can choose
for each $s\in L_k$ an interval $(a_{k,s},b_{k,s})\subset(0,1/2)$ with center at $x_s$ such that
\begin{itemize}
\item $a_{k,s},b_{k,s}\not \in S$,
\item $(a_{k,s},b_{k,s})$ does not contain any of the points $x_m$, $1\le m\le N_{k+1}$,
\item the total length of these $(a_{k,s},b_{k,s})$, $s\in L_k$ is smaller than $\d_k/16(k+1)$,
\item $h_k(x)=-x$ on $(a_{k,s},b_{k,s})$ and
\item these intervals
are pairwise disjoint for different $s$.
\end{itemize}
Finally set $h_{k+1}$
as in (\ref{io}), and this finishes the induction.
The sequence $\{h_k\}_{k=1}^\i$ converges decreasingly and uniformly to a 1-periodic continuous function $h$, and by
the construction we have $h(x_m)=h_k(x_m)$ for all $1\le m\le N_{k+1}$. Hence, (\ref{kl}) yields
\be \liminf_{k\to\i}\frac{1}{N_{k+1}}\#\left\{1\le m\le N_{k+1}\sep x_m+h(x_m)\in (7/8,1)\ \ \mbox{(mod 1)}\right\}\le \frac{1}{4}.\label{kl1}\ee
On the other hand, for all $m\in L_k$ we have
\[\frac{1}{16}\le x_m+h(x_m)\le x_m+h_{k+1}(x_m)<0,\]
and hence
for all such $m$ the fractional part $\{x_m+h(x_m)\}$ belongs to $(15/16,1)$. But then
(\ref{13}) implies
\be \limsup_{k\to\i}\frac{1}{M_{k+1}}\#\left\{1\le m\le M_{k+1}\sep x_m+h(x_m)\in (15/16,1)\ \ \mbox{(mod 1)}\right\}\ge
\frac12-\frac18.\label{kl2}\ee
Since $(15/16,1)$ is a proper subinterval of $(7/8,1)$, the relations
(\ref{kl1}) and (\ref{kl2}) show that the sequence $\{x_m+h(x_m)\}_{m=1}^\i$,
and hence also the sequence $\Bigl\{\{m\a+h(m\a)\}\Bigr\}_{m=1}^\i$
(mod 1) has no distribution.\endproof
|
\section{Introduction}
In the recent decade, many high-energy astrophysical neutrinos (HANs) events in the TeV - PeV energy range have been observed by the IceCube observatory\cite{1,2,3,4,5}.
The energy spectrum, flavor ratio, and sky distribution of HANs are analysed\cite{6,7,8,9,10}. With help of the synergetic observation of electromagnetic wave messengers, the
promising sources of HANs are indicated\cite{11,12,13,14}. Although considerable progress has been made in neutrino astronomy,
there are several important problems to be solved: The uncertainty of flavor ratio measured at Earth is large\cite{15,16}; No source of HANs is identified above the $5\sigma$ discovery threshold; How to explain the multi-messenger emission from the blazar TXS0506+056 is challenging\cite{17,18,19}; The production mechanism of HANs is unclear. Among these problems, the first one would be overcomed
with the operations of the next-generation neutrino telescopes\cite{20,21,22,23,24} in the next two decades. Furthermore, because the energy spectrum and the flavor ratio of HANS at the source are determined by the production mechanism,
the progress of the first one would also provide clues for the last one.
As is known, the flavor ratio at the source is inferred from that at Earth and the flavor transition probability of the HANs. Under the assumption that
the leptonic mixing matrix at the source and that in the propagation path are the same as the one in vacuum,
after the decoherence of neutrinos, the so-called standard flavor transition matrix (FTM) is obtained, i.e.,
\begin{equation}
\label{eq:1}
\overline{P}_{\alpha\beta}^{s}=\sum_{i}|U_{\alpha i}|^{2}|U_{\beta i}|^{2},
\end{equation}
where $U$ is the leptonic mixing matrix in vacuum with $\alpha, \beta = e, \mu, \tau$, $i = 1, 2, 3$\cite{25}.
At present, the constraint on the FTM from the measured flavor ratio is loose. So this minimal theoretical scenario works well. However, precise measurements of the flavor ratio would be available in the near future.
Then a more realistic scenario which includes the matter effects or new physics effects should be considered. Various FTMs based on new physics have been proposed, see \cite{26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47} for example. These new physics effects could be tested by the IceCube observatory and the next-generation neutrino telescopes.
In this paper, we are concerned about impacts of the matter effects on the flavor transition of HANs. In general, the matter effects arise from the coherent forward elastic and incoherent scattering of HANs
with the medium particles. The former affects the flavor conversion through the neutrino mixing matrix in matter. The latter changes not only the flavor composition but also the energy spectrum of HANs.
So it might play an important role in the recombination process of HANs in the dense medium. We focus on the matter effects from the elastic scattering in this paper.
Particularly, we consider the case that the HANs propagate in gas which is a realistic environment around the possible sources such as blazars and tidal disruption events (TDEs)\cite{19}.
Considering the energy of HANs, the size of the medium, and the profile of the electron density, the FTMs are dependent on the decoherence schemes of neutrinos in the propagation.
Even for the same medium, different schemes might lead to different FTMs\cite{48,49}. In order to understand the matter effects on the flavor transition of HANs comprehensively, we consider five typical types of decoherence schemes in this paper. On the basis of the specific parameters of HANs and the medium, the predictions from these schemes are examined.
The paper is organised as follows. In Sec. \uppercase\expandafter{\romannumeral2}, the FTMs are given on the basis of the decoherence schemes in Refs.\cite{48,49,50,51} and the modified version.
In Sec. \uppercase\expandafter{\romannumeral3}, the predictions on the flavor ratio of HANs at Earth from special sources are given with these FTMs.
Identifications of these matter effects by the next generation of neutrino telescopes are discussed. Finally, a summary is given.
\section{FTMs based on different decoherence schemes.}
In this section, we introduce five FTMs on the basis of the decoherence schemes of HANs in the propagation. These schemes could be viewed as the boundary conditions
in the propagation problem of HANs.
First we consider the case that the electron density in the medium varies slowly and the flavor transition of HANs is
adiabatic. If the propagation distance in the medium is large enough, the oscillation effects can be neglected. Then we can obtain a simple FTM of the form\cite{51}
\begin{equation}
\label{eq:2}
\overline{P}^{\uppercase\expandafter{\romannumeral1}}_{\alpha\beta}=\sum_{i}|U_{\alpha i}^{m}|^{2}|U_{\beta i}|^{2},
\end{equation}
were $U^{m}$ is the leptonic mixing matrix in matter.
In this propagation scheme, there is a decoherence process that changes the flavor state of neutrinos to the eigen-state of the Hamiltonian in matter.
Furthermore, in the interface of matter and vacuum the eigen-state of the Hamiltonian in matter correspondingly changes into that in vacuum.
In the case that the oscillation effects should be considered, the adiabatic FTM could be modified as follows
\begin{equation}
\label{eq:3}
\overline{P}^{\uppercase\expandafter{\romannumeral2}}_{\alpha\beta}=\sum_{\gamma}P_{\alpha\gamma}^{m}\overline{P}^{\uppercase\expandafter{\romannumeral1}}_{\gamma\beta},
\end{equation}
where $P_{\alpha\gamma}^{m}$ is the neutrino oscillation probability in matter. This FTM shows the process that the flavor of HANs in matter first oscillates and then changes adiabaticly, see Fig.\ref{fig:1}.
\begin{figure}[t]
\centering
\includegraphics[width=.45
\textwidth]{fig1.pdf}
\caption{\label{fig:1} The schematic figure for the propagation Scheme \uppercase\expandafter{\romannumeral2} in matter. }
\end{figure}
Furthermore, the decoherence of neutrinos takes place two times in the propagation path. In detail, after the first decoherence, the superposition of the oscillation amplitude $A^{m}_{\alpha\gamma}$ is lost.
After the second, that of $U^{m}_{\alpha i}$ is lost. This decoherence scheme is the adiabatic version of that introduced
in Ref.\cite{48}, where the FTM is written as
\begin{equation}
\label{eq:4}
\overline{P}^{\uppercase\expandafter{\romannumeral3}}_{\alpha\beta}=\sum_{\gamma}P_{\alpha\gamma}^{m}\overline{P}^{S}_{\gamma\beta}.
\end{equation}
Note that $\overline{P}^{S}_{\gamma\beta}$ is the standard FTM shown in Eq.\ref{eq:1}.
In contrast, the flavor state in Scheme \uppercase\expandafter{\romannumeral3} changes into the mass eigen-state in vacuum after the second decoherence.
Another scheme including the oscillation effects can be obtained from the modification of the Scheme \uppercase\expandafter{\romannumeral3}. The corresponding FTM is expressed as \cite{49}
\begin{equation}
\label{eq:5}
\overline{P}^{\uppercase\expandafter{\romannumeral4}}_{\alpha\beta}=\sum_{i}|\sum_{\gamma}A^{m}_{\alpha\gamma}U_{\gamma i}|^{2}|U_{\beta i}|^{2},
\end{equation}
where $A^{m}_{\alpha\gamma}$ is the neutrino oscillation amplitude in matter. In comparison with $\overline{P}^{\uppercase\expandafter{\romannumeral3}}_{\alpha \beta}$, which can be rewritten
as
\begin{equation}
\label{eq:6}
\overline{P}^{\uppercase\expandafter{\romannumeral3}}_{\alpha\beta}=\sum_{i}\sum_{\gamma}|A^{m}_{\alpha\gamma}|^{2}|U_{\gamma i}|^{2}|U_{\beta i}|^{2},
\end{equation}
there is a superposition of the oscillation amplitude in $\overline{P}^{IV}_{\alpha \beta}$.
So in the propagation there is only one decoherence which changes the flavor state to the mass eigen-state.
Let us introduce the fifth decoherence scheme which is applied in the case that the electron density is constant.
The decoherence is obtained through the integral of the oscillation probability with the whole propagation path, i.e.\cite{50},
\begin{equation}
\label{eq:7}
\overline{P}^{\uppercase\expandafter{\romannumeral5}}_{\alpha\beta}=\int_{0}^{L_{vac}}\int_{0}^{L}P_{\alpha\beta}(l_{vac},l)dl_{vac}dl_{m}L_{vac}^{-1}L^{-1},
\end{equation}
with
\begin{equation}
\label{eq:8}
P_{\alpha\beta}(l_{vac},l)=|\langle\nu_{\beta}|Ue^{-iH_{vac}l_{vac}}U^{+}U^{m}e^{-iH_{m}l_{m}}U^{m +}|\nu_{\alpha} \rangle|^{2},
\end{equation}
where $L$ ($L_{vac}$) is the propagation distance in matter (vacuum), $H_{m}$ ($H_{vac}$) is the diagonal neutrino Hamiltonian matrix in matter (vacuum).
When the distance is much larger than the corresponding oscillation length, i.e., $L\gg L^{osc}$, $L_{vac}\gg L^{osc}_{vac}$, the averaged oscillation probability could
be written as follows
\begin{equation}
\label{eq:9}
\overline{P}^{\uppercase\expandafter{\romannumeral5}}_{\alpha\beta}=\sum_{i}\sum_{j}|U^{m}_{\alpha j}|^{2}P_{ji}|U_{\beta i}|^{2}
\end{equation}
with
\begin{equation}
\label{eq:10}
P_{ji}=|\sum_{\gamma }U^{*}_{\gamma i}U_{\gamma j}^{m}|^{2}.
\end{equation}
In contrast to Scheme \uppercase\expandafter{\romannumeral1}, besides the first decoherence that changes the flavor state to the eigen-state of the Hamiltonian in matter,
there is another decoherence that changes the eigen-state of the Hamiltonian in matter to that in vacuum, i.e.,
$|\nu_{j}\rangle^{m}\rightarrow|\nu_{i}\rangle$.
Now we have introduced five typical deocherence schemes for the propagation problem of HANs in matter.
For the sake of inter-comparison,
their characteristic is listed in Tab.\ref{tab:1}.
\begin{table}
\caption{Characteristic of the decoherence schemes}
\label{tab:1}
\begin{tabular}{c c c c c c }
\noalign{\smallskip}\hline
~~Decoherence Scheme~~&~~\uppercase\expandafter{\romannumeral1}~~ &~~\uppercase\expandafter{\romannumeral2}~~&
~~\uppercase\expandafter{\romannumeral3}~~&~~ \uppercase\expandafter{\romannumeral4}~~&~~~ \uppercase\expandafter{\romannumeral5}~~ \\ \hline
\noalign{\smallskip}
~~Including adiabatic evolution~~ &~yes~&~~yes~~& ~no~&~~no~~&~no~\\
~~Including oscillation effects~~&~~no~~&~~yes~~~&~yes~~& ~~yes~&~~no~~\\
~~Times of decoherence in the propagation~~&~~one~~&~~two~~&~~two~~& ~~one~~&~~two~~\\
\hline
\noalign{\smallskip}
\end{tabular}
\vspace*{0.6cm}
\end{table}
In the following section we would examine the predictions from these schemes with specific parameters of HANs and matter.
\section{Predictions from FTMs}
As we have pointed out in Sec.\uppercase\expandafter{\romannumeral1}, we consider the matter effects in the gas environment around the promising source.
In this section we first analyse the dependence of the FTMs on the parameters of neutrinos and matter and then examine the predictions from the given FTMs.
\subsection{Dependence of the FTMs on the parameters of neutrinos and matter}
The specific parameters of the HANs and matter for the corresponding propagation scheme are listed in Tab.\ref{tab:2}.
\begin{table}
\caption{Parameters of neutrinos and matter for the propagation schemes}
\label{tab:2}
\begin{tabular}{c c c c c }
\noalign{\smallskip}\hline
~Scheme~&~Distance in vacuum ($L_{vac}$)~&~Distance in matter($L$) ~ &~Neutrino energy ($E$) ~~&
~Electron density ($N_{e}$)~~\\ \hline
\noalign{\smallskip}
\uppercase\expandafter{\romannumeral2}, \uppercase\expandafter{\romannumeral3}, \uppercase\expandafter{\romannumeral4}~&~$\gg L_{vac}^{osc}$~ &~[$10^{5}$, $4\times10^{8}$]km~&~[$10^{2}$, $10^{3}$]TeV~& ~[$10^{21}$, $10^{22}$] cm$^{-3}$~\\
\uppercase\expandafter{\romannumeral1}, \uppercase\expandafter{\romannumeral5}~&~$\gg L_{vac}^{osc}$~&~$\gg L^{osc}$~&~[$10^{2}$, $10^{3}$]TeV~& ~[$10^{21}$, $10^{22}$] cm$^{-3}$~\\
\hline
\noalign{\smallskip}
\end{tabular}
\vspace*{0.6cm}
\end{table}
We give comments on the ranges of the parameters as follows: (a). The range of $L/E$ is $[0.1, 4000]$km/GeV. The oscillation effects become noticeable when $L/E$ if of order of $10^{3}$ km/GeV.
When $L/E\ll10^{3}$km/GeV, Scheme \uppercase\expandafter{\romannumeral2} would
be reduced to \uppercase\expandafter{\romannumeral1}, with \uppercase\expandafter{\romannumeral3}, \uppercase\expandafter{\romannumeral4} reduced to the standard scheme. (b). The range of the paramerter $A_{cc}\equiv2\sqrt{2}EG_{F}N_{e}$ is $[0.0254, 2.54]$eV$^{2}$. So there is no matter resonance for the mixing angle $\theta_{ij}$ with $ij=12, 13, 23$.
For the given range of the neutrino energy, the resonance takes place in dilute gas. For example, if $E$ takes 1PeV, the resonance density $N_{e}$ for $\theta_{12}$ is $2\times10^{17}$cm$^{-3}$.
We note that Ref.\cite{50} has discussed the resonance effect with Scheme \uppercase\expandafter{\romannumeral5}. So we don't consider the case of dilute gas here.
On the basis of the ranges of the parameters, the electron neutrinos are decoupled. The neutrino (antineutrinos) oscillation probabilities are of simple forms, i.e.\cite{52},
\begin{equation}
\label{eq:11}
\begin{array}{c}
P^{m}_{\nu_{e}\nu_{e}}\approx P^{m}_{\overline{\nu}_{e}\overline{\nu}_{e}}\approx1,\\
~~P^{m}_{\nu_{e}\nu_{\mu}}\approx P^{m}_{\overline{\nu}_{e}\overline{\nu}_{\mu}}\approx
P^{m}_{\nu_{e}\nu_{\tau}}\approx P^{m}_{\overline{\nu}_{e}\overline{\nu}_{\tau}}\approx P^{m}_{\nu_{\mu}\nu_{e}}\approx P^{m}_{\overline{\nu}_{\mu}\overline{\nu}_{e}}\approx
P^{m}_{\nu_{\tau}\nu_{e}}\approx P^{m}_{\overline{\nu}_{\tau}\overline{\nu}_{e}}\approx0,\\
P^{m}_{\nu_{\mu}\nu_{\mu}}\approx P^{m}_{\overline{\nu}_{\mu}\overline{\nu}_{\mu}}\approx P^{m}_{\nu_{\tau}\nu_{\tau}}\approx P^{m}_{\overline{\nu}_{\tau}\overline{\nu}_{\tau}}\approx1-\sin^{2}2\widetilde{\theta}\sin^{2}\frac{\Delta \widetilde{m}^{2}_{32}L}{4E},\\
P^{m}_{\nu_{\mu}\nu_{\tau}}\approx P^{m}_{\overline{\nu}_{\mu}\overline{\nu}_{\tau}}\approx P^{m}_{\nu_{\tau}\nu_{\mu}}\approx P^{m}_{\overline{\nu}_{\tau}\overline{\nu}_{\mu}}\approx
\sin^{2}2\widetilde{\theta}\sin^{2}\frac{\Delta \widetilde{m}^{2}_{32}L}{4E},
\end{array}
\end{equation}
where $\widetilde{\theta}$ and $\widetilde{m}^{2}_{32}$ are effective oscillation parameters in matter written as \cite{52}
\begin{equation}
\label{eq:12}
\tan2\widetilde{\theta}=\frac{2|\Delta m^{2}_{21}U_{\mu2}U^{*}_{\tau2}+\Delta m^{2}_{31}U_{\mu3}U^{*}_{\tau3}|}{\Delta m^{2}_{21}(|U_{\tau2}|^{2}-|U_{\mu2}|^{2})+\Delta m^{2}_{31}(|U_{\tau3}|^{2}-|U_{\mu3}|^{2})},
\end{equation}
\begin{equation}
\label{eq:13}
\Delta \widetilde{m}^{2}_{32}\approx\pm\sqrt{(\Delta m^{2}_{31}c^{2}_{13}-\Delta m^{2}_{21}(c^{2}_{12}-s^{2}_{12}s^{2}_{13}))^{2}+(\Delta m^{2}_{21}\sin2\theta_{12}s_{13})^{2}},
\end{equation}
with $s_{ij}=\sin\theta_{ij}$, $c_{ij}=\cos\theta_{ij}$.
From the above expressions, we find that the oscillation probabilities of neutrinos approximate those of antineutrinos. They are not dependent on the parameter $A_{cc}$.
Neither are the FTMs in the given ranges of the parameters, see Fig.\ref{fig:2} and Fig.\ref{fig:3}.
\begin{figure}[t]
\label{fig:2}
\centering
\includegraphics[width=.49\textwidth]{fig2a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig2b.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig2c.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig2d.pdf}
\caption{\label{fig:2} Relation between the magnitude of the element in FTMs and the parameter $A_{cc}$ based on the best fit data with the normal neutrino mass ordering\cite{53}. Scheme \uppercase\expandafter{\romannumeral1}: red,
Scheme \uppercase\expandafter{\romannumeral2}: green, Scheme \uppercase\expandafter{\romannumeral3}: magenta, Scheme \uppercase\expandafter{\romannumeral4}: yellow,
Scheme \uppercase\expandafter{\romannumeral5}: blue. Neutrinos: real lines, Antineutrinos: dashed lines.}
\end{figure}
\begin{figure}[t]
\label{fig:3}
\centering
\includegraphics[width=.49\textwidth]{fig3a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig3b.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig3c.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig3d.pdf}
\caption{\label{fig:3} Relation between the magnitude of the element in FTMs and the parameter $A_{cc}$ based on the best fit data with the inverted neutrino mass ordering\cite{53}.
The convention of color is the same as that in Fig.\ref{fig:2}.}
\end{figure}
However, we can see that except $\overline{P}_{\nu_{e}\nu_{e}}$ ($\overline{P}_{\overline{\nu}_{e}\overline{\nu}_{e}}$), the elements in the FTM are dependent on the mass ordering
of neutrinos. Correspondingly, the flavor ratio at Earth is affected by the mass ordering of neutrinos.
Now we examine the oscillations of the elements in FTMs with the parameter $L/E$ in the cases of Scheme \uppercase\expandafter{\romannumeral2}, \uppercase\expandafter{\romannumeral3}, \uppercase\expandafter{\romannumeral4}.
\begin{figure}[t]
\label{fig:4}
\centering
\includegraphics[width=.48\textwidth]{fig4a.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4b.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4c.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4d.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4e.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4f.pdf}
\caption{\label{fig:4} Oscillation of the element in FTMs based on the best fit data with the normal neutrino mass ordering\cite{53}.
Scheme \uppercase\expandafter{\romannumeral2}: green, Scheme \uppercase\expandafter{\romannumeral3}: magenta, Scheme \uppercase\expandafter{\romannumeral4}: yellow.}
\end{figure}
\begin{figure}[t]
\label{fig:5}
\centering
\includegraphics[width=.48\textwidth]{fig5a.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5b.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5c.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5d.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5e.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5f.pdf}
\caption{\label{fig:5} Oscillation of the element in FTMs based on the best fit data with the inverted neutrino mass ordering\cite{53}.
The convention of color is the same as that in Fig.\ref{fig:4}.}
\end{figure}
They are shown in Fig.\ref{fig:4} and Fig.\ref{fig:5}.
Some necessary comments are given as follows:
(a). The FTMs of neutrinos are nearly identical to those of antineutrinos. So we just show the oscillations of neutrinos. (b). The elements in the first row of the FTM, namely $\overline{P}_{\nu_{e}\nu_{e}}$, $\overline{P}_{\nu_{e}\nu_{\mu}}$, $\overline{P}_{\nu_{e}\nu_{\tau}}$, do not oscillation with the parameter $L/E$. (c). The oscillation amplitudes
are small for most elements in the FTMs except for $\overline{P}_{\nu_{\mu}\nu_{e}}$, $\overline{P}_{\nu_{\mu}\nu_{\mu}}$,
$\overline{P}_{\nu_{\tau}\nu_{e}}$ in Scheme \uppercase\expandafter{\romannumeral4} with the inverted neutrino mass ordering. Correspondingly, the most noticeable amplitude of the apparent T violation (ATV), namely
$\Delta _{e\mu}\equiv\overline{P}_{\nu_{e}\nu_{\mu}}-\overline{P}_{\nu_{\mu}\nu_{e}}$,
also appears in that case, see Fig.\ref{fig:6}.
\begin{figure}[t]
\label{fig:6}
\centering
\includegraphics[width=.48\textwidth]{fig6a.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig6b.pdf}
\caption{\label{fig:6} Oscillation of the the apparent T violation $\Delta _{e\mu}\equiv\overline{P}_{\nu_{e}\nu_{\mu}}-\overline{P}_{\nu_{\mu}\nu_{e}}$ based on the best fit data \cite{43}.
Left panel: Normal mass ordering, Right panel: Inverted mass ordering. The convention of color is the same as that in Fig.\ref{fig:4}.}
\end{figure}
\subsection{Matter effects on flavor ratio at Earth}
Now we examine the matter effects on the flavor ratio of HANs at Earth.
We consider three typical sources, i.e.,
\begin{equation}
\label{eq:14}
\begin{array}{c}
neutron-decay ~~source ~~with~~\Phi^{s}=(1, ~~0,~~ 0),\\
pion-decay ~~source~~with ~~~\Phi^{s}=(1/3, ~~2/3, ~~0),\\
muon-damping~~ source~~with ~~~ \Phi^{s}=(0, ~~1, ~~0),
\end{array}
\end{equation}
where $\Phi^{s}$ is the flavor ratio at the source. Using $\Phi^{s}$ and the aforementioned FTMs, the flavor ratio of HANs at Earth are obtained. On the basis of the given parameters in Tab.\ref{tab:2}
and the recent global fit data of neutrinos oscillations\cite{53}, the ternary plots for the flavor ratio are shown in Fig.\ref{fig:7}-Fig.\ref{fig:9}.
For the sake of comparison, the flavor ratio predicted by the standard scheme is also shown in these figures.
From the figures, we could obtain following observations.
\begin{figure}[t]
\label{fig:7}
\centering
\includegraphics[width=.6\textwidth]{fig7a.pdf}
\caption{\label{fig:7} Ternary plot of the flavor ratio at Earth with the neutron-decay source on the basis of the global fit data
at $3\sigma$ level\cite{53} with the normal mass ordering and the parameters in Tab.\ref{tab:2}. The 2015 $3\sigma$ constraint is taken from Ref.\cite{54}. The 2020 and 2040 $3\sigma$ credible regions
with the pion-decay source
based on IceCube and
IceCube-Gen2 are taken from Ref.\cite{21}.
The $3\sigma$ credible region with the pion-decay source based on all TeV-PeV neutrino telescopes available in 2040 is taken from Ref.\cite{15}.
}
\end{figure}
\begin{figure}[t]
\label{fig:8}
\centering
\includegraphics[width=.49\textwidth]{fig8a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig8b.pdf}
\caption{\label{fig:8} Ternary plot of the flavor ratio at Earth with the pion-decay source.
Left panel: Normal mass ordering, Right panel: Inverted mass ordering. The conventions of parameters are the same as those in Fig.\ref{fig:7}.}
\end{figure}
\begin{figure}[t]
\label{fig:9}
\centering
\includegraphics[width=.49\textwidth]{fig9a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig9b.pdf}
\caption{\label{fig:9} Ternary plot of the flavor ratio at Earth with the muon-damping source.
Left panel: Normal mass ordering, Right panel: Inverted mass ordering. The conventions of parameters are the same as those in Fig.\ref{fig:7}.}
\end{figure}
For the neutron-decay source (see Fig.\ref{fig:7}), because the mass ordering of neutrinos and the oscillation effects do not impact the flavor transition
of $\nu_{e}$ ($\overline{\nu}_{e}$) in the given ranges of the parameters, the predicted flavor ratio from the FTMs could be classified into three regions as follows.
(a). The red region from Scheme \uppercase\expandafter{\romannumeral1}: It is the same as the green from Scheme \uppercase\expandafter{\romannumeral2}.
(b). The magenta region from Scheme \uppercase\expandafter{\romannumeral3}: It is identical to the yellow from Scheme \uppercase\expandafter{\romannumeral3} or the cyan from the standard scheme.
(c). The blue region from Scheme \uppercase\expandafter{\romannumeral5}: It overlaps with that from Scheme \uppercase\expandafter{\romannumeral3}.
All the predicted regions are at the margin of the 2015 ICeCube $3\sigma$ constraint\cite{54}. Considering the 2040 expected sensitivity on the flavor ratio\cite{15,21},
these regions would be precluded by the observations of the neutrino telescopes in the near future.
For the pion-decay source (see Fig.\ref{fig:8}), the predicted flavor could be classified into four regions as follows.
(a). The region from Scheme \uppercase\expandafter{\romannumeral1}:
Since the oscillation effect is not notable for Scheme \uppercase\expandafter{\romannumeral2}, this region is nearly identical to that from Scheme \uppercase\expandafter{\romannumeral2}.
Nearly half of the area overlaps with the 2040 $3\sigma$ credible region \cite{15}. So Scheme \uppercase\expandafter{\romannumeral1} and \uppercase\expandafter{\romannumeral2}
might work in the era of precise measurement of the flavor ratio.
(b). The region from scheme \uppercase\expandafter{\romannumeral3}: Compared with the prediction from the standard scheme, this region shows that the oscillation effects are not negligible.
However, the main part of this region is outside the $3\sigma$ credible region \cite{15}. So Scheme \uppercase\expandafter{\romannumeral3} with this source
would be stringently constrained by the observations of the neutrino telescopes in the future.
(c). The region from scheme \uppercase\expandafter{\romannumeral4}: Although the oscillation amplitudes of the elements in the FTM are noticeable with
the inverted neutrino mass ordering, the variation of the parameters in Tab.\ref{tab:2} and the global fit data of neutrinos undermine the oscillation effects.
So the predicted region from Scheme \uppercase\expandafter{\romannumeral4} approximates that from the standard scheme.
Considering the 2040 $3\sigma$ credible region \cite{15}, Scheme \uppercase\expandafter{\romannumeral4} and the standard scheme are as promising as Scheme \uppercase\expandafter{\romannumeral1},
\uppercase\expandafter{\romannumeral2}.
(d). The region from Scheme \uppercase\expandafter{\romannumeral5}: The mass ordering of neutrinos impacts the prediction of Scheme \uppercase\expandafter{\romannumeral5} noticeably.
Given the $3\sigma$ credible region \cite{15}, the prediction with the normal ordering is more favored. In this case, the blue region overlaps with the yellow, while its area is smaller.
In general, for this source, the identification of the matter effects is difficult because the region from the standard scheme noticeably overlaps with the regions from the other schemes.
For the muon-damping source (see Fig.\ref{fig:9}),
the predicted flavor could also be classified into four regions as those illustrated in case of the pion-decay source.
Even though the condition of the matter resonance is not satisfied, the matter effects in this case are noticeable for the Scheme
\uppercase\expandafter{\romannumeral1}, \uppercase\expandafter{\romannumeral2}, \uppercase\expandafter{\romannumeral3} with both mass orderings,
and Scheme \uppercase\expandafter{\romannumeral5} with the inverted mass ordering. Furthermore, because the region from Scheme \uppercase\expandafter{\romannumeral1}
or \uppercase\expandafter{\romannumeral2} does not overlap with that from the standard scheme or Scheme \uppercase\expandafter{\romannumeral4},
the matter effects from the former might be identified by the neutrinos telescopes in the near future.
On the basis of the above observations on the flavor ratio of the HANs at Earth,
we can see that even in gas where the matter resonance is not available the matter effects could not be neglected in general.
Although the matter effects are dependent on the decoherence schemes and the production sources, the identification of them is possible
in the near future. Then the precise measurement of the flavor ratio would set constraints on the production environment and the propagation path.
\section{Conclusions}
A precise measurement of the flavor ratio of HANs is available in the near future.
To test the matter effects on the flavor transition of HANs by the next-generation neutrino telescopes deserves serious consideration.
The matter effects are dependent on the propagation scheme of the HANs.
We introduced five typical decoherence schemes for the HANs in the propagation.
On the basis of the propagation parameters of the HANs in the realistic environment around the promising sources,
the impacts of the matter effects on the FTM were analysed.
When the matter resonance condition is not satisfied, the oscillation effect is negligible for the schemes except for the special one (Scheme \uppercase\expandafter{\romannumeral3}).
Even so, the matter effects shown through the flavor ratio of the HANs at Earth are noticeable for the typical sources. Particularly, on the base of the schemes (\uppercase\expandafter{\romannumeral1},
\uppercase\expandafter{\romannumeral2}) the identification of the matter effects
is promising in the case that the HANs are from the muon-damping source.
Therefore, from the measured flavor ratio in the near future, the matter parameters and propagation path might be deduced, which could give us clues on the production mechanism
of HANs.
\vspace{0.08cm}
\acknowledgments
This work is supported by the National Natural Science Foundation of China under grant No. 12065007, the Guangxi Scientific Programm Foundation under grant No. Guike AD19110045, the Research Foundation of Guilin University of Technology under grant No. GUTQDJJ2018103.\\
\section{Introduction}
In the recent decade, many high-energy astrophysical neutrinos (HANs) events in the TeV - PeV energy range have been observed by the IceCube observatory\cite{1,2,3,4,5}.
The energy spectrum, flavor ratio, and sky distribution of HANs are analysed\cite{6,7,8,9,10}. With help of the synergetic observation of electromagnetic wave messengers, the
promising sources of HANs are indicated\cite{11,12,13,14}. Although considerable progress has been made in neutrino astronomy,
there are several important problems to be solved: The uncertainty of flavor ratio measured at Earth is large\cite{15,16}; No source of HANs is identified above the $5\sigma$ discovery threshold; How to explain the multi-messenger emission from the blazar TXS0506+056 is challenging\cite{17,18,19}; The production mechanism of HANs is unclear. Among these problems, the first one would be overcomed
with the operations of the next-generation neutrino telescopes\cite{20,21,22,23,24} in the next two decades. Furthermore, because the energy spectrum and the flavor ratio of HANS at the source are determined by the production mechanism,
the progress of the first one would also provide clues for the last one.
As is known, the flavor ratio at the source is inferred from that at Earth and the flavor transition probability of the HANs. Under the assumption that
the leptonic mixing matrix at the source and that in the propagation path are the same as the one in vacuum,
after the decoherence of neutrinos, the so-called standard flavor transition matrix (FTM) is obtained, i.e.,
\begin{equation}
\label{eq:1}
\overline{P}_{\alpha\beta}^{s}=\sum_{i}|U_{\alpha i}|^{2}|U_{\beta i}|^{2},
\end{equation}
where $U$ is the leptonic mixing matrix in vacuum with $\alpha, \beta = e, \mu, \tau$, $i = 1, 2, 3$\cite{25}.
At present, the constraint on the FTM from the measured flavor ratio is loose. So this minimal theoretical scenario works well. However, precise measurements of the flavor ratio would be available in the near future.
Then a more realistic scenario which includes the matter effects or new physics effects should be considered. Various FTMs based on new physics have been proposed, see \cite{26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47} for example. These new physics effects could be tested by the IceCube observatory and the next-generation neutrino telescopes.
In this paper, we are concerned about impacts of the matter effects on the flavor transition of HANs. In general, the matter effects arise from the coherent forward elastic and incoherent scattering of HANs
with the medium particles. The former affects the flavor conversion through the neutrino mixing matrix in matter. The latter changes not only the flavor composition but also the energy spectrum of HANs.
So it might play an important role in the recombination process of HANs in the dense medium. We focus on the matter effects from the elastic scattering in this paper.
Particularly, we consider the case that the HANs propagate in gas which is a realistic environment around the possible sources such as blazars and tidal disruption events (TDEs)\cite{19}.
Considering the energy of HANs, the size of the medium, and the profile of the electron density, the FTMs are dependent on the decoherence schemes of neutrinos in the propagation.
Even for the same medium, different schemes might lead to different FTMs\cite{48,49}. In order to understand the matter effects on the flavor transition of HANs comprehensively, we consider five typical types of decoherence schemes in this paper. On the basis of the specific parameters of HANs and the medium, the predictions from these schemes are examined.
The paper is organised as follows. In Sec. \uppercase\expandafter{\romannumeral2}, the FTMs are given on the basis of the decoherence schemes in Refs.\cite{48,49,50,51} and the modified version.
In Sec. \uppercase\expandafter{\romannumeral3}, the predictions on the flavor ratio of HANs at Earth from special sources are given with these FTMs.
Identifications of these matter effects by the next generation of neutrino telescopes are discussed. Finally, a summary is given.
\section{FTMs based on different decoherence schemes.}
In this section, we introduce five FTMs on the basis of the decoherence schemes of HANs in the propagation. These schemes could be viewed as the boundary conditions
in the propagation problem of HANs.
First we consider the case that the electron density in the medium varies slowly and the flavor transition of HANs is
adiabatic. If the propagation distance in the medium is large enough, the oscillation effects can be neglected. Then we can obtain a simple FTM of the form\cite{51}
\begin{equation}
\label{eq:2}
\overline{P}^{\uppercase\expandafter{\romannumeral1}}_{\alpha\beta}=\sum_{i}|U_{\alpha i}^{m}|^{2}|U_{\beta i}|^{2},
\end{equation}
were $U^{m}$ is the leptonic mixing matrix in matter.
In this propagation scheme, there is a decoherence process that changes the flavor state of neutrinos to the eigen-state of the Hamiltonian in matter.
Furthermore, in the interface of matter and vacuum the eigen-state of the Hamiltonian in matter correspondingly changes into that in vacuum.
In the case that the oscillation effects should be considered, the adiabatic FTM could be modified as follows
\begin{equation}
\label{eq:3}
\overline{P}^{\uppercase\expandafter{\romannumeral2}}_{\alpha\beta}=\sum_{\gamma}P_{\alpha\gamma}^{m}\overline{P}^{\uppercase\expandafter{\romannumeral1}}_{\gamma\beta},
\end{equation}
where $P_{\alpha\gamma}^{m}$ is the neutrino oscillation probability in matter. This FTM shows the process that the flavor of HANs in matter first oscillates and then changes adiabaticly, see Fig.\ref{fig:1}.
\begin{figure}[t]
\centering
\includegraphics[width=.45
\textwidth]{fig1.pdf}
\caption{\label{fig:1} The schematic figure for the propagation Scheme \uppercase\expandafter{\romannumeral2} in matter. }
\end{figure}
Furthermore, the decoherence of neutrinos takes place two times in the propagation path. In detail, after the first decoherence, the superposition of the oscillation amplitude $A^{m}_{\alpha\gamma}$ is lost.
After the second, that of $U^{m}_{\alpha i}$ is lost. This decoherence scheme is the adiabatic version of that introduced
in Ref.\cite{48}, where the FTM is written as
\begin{equation}
\label{eq:4}
\overline{P}^{\uppercase\expandafter{\romannumeral3}}_{\alpha\beta}=\sum_{\gamma}P_{\alpha\gamma}^{m}\overline{P}^{S}_{\gamma\beta}.
\end{equation}
Note that $\overline{P}^{S}_{\gamma\beta}$ is the standard FTM shown in Eq.\ref{eq:1}.
In contrast, the flavor state in Scheme \uppercase\expandafter{\romannumeral3} changes into the mass eigen-state in vacuum after the second decoherence.
Another scheme including the oscillation effects can be obtained from the modification of the Scheme \uppercase\expandafter{\romannumeral3}. The corresponding FTM is expressed as \cite{49}
\begin{equation}
\label{eq:5}
\overline{P}^{\uppercase\expandafter{\romannumeral4}}_{\alpha\beta}=\sum_{i}|\sum_{\gamma}A^{m}_{\alpha\gamma}U_{\gamma i}|^{2}|U_{\beta i}|^{2},
\end{equation}
where $A^{m}_{\alpha\gamma}$ is the neutrino oscillation amplitude in matter. In comparison with $\overline{P}^{\uppercase\expandafter{\romannumeral3}}_{\alpha \beta}$, which can be rewritten
as
\begin{equation}
\label{eq:6}
\overline{P}^{\uppercase\expandafter{\romannumeral3}}_{\alpha\beta}=\sum_{i}\sum_{\gamma}|A^{m}_{\alpha\gamma}|^{2}|U_{\gamma i}|^{2}|U_{\beta i}|^{2},
\end{equation}
there is a superposition of the oscillation amplitude in $\overline{P}^{IV}_{\alpha \beta}$.
So in the propagation there is only one decoherence which changes the flavor state to the mass eigen-state.
Let us introduce the fifth decoherence scheme which is applied in the case that the electron density is constant.
The decoherence is obtained through the integral of the oscillation probability with the whole propagation path, i.e.\cite{50},
\begin{equation}
\label{eq:7}
\overline{P}^{\uppercase\expandafter{\romannumeral5}}_{\alpha\beta}=\int_{0}^{L_{vac}}\int_{0}^{L}P_{\alpha\beta}(l_{vac},l)dl_{vac}dl_{m}L_{vac}^{-1}L^{-1},
\end{equation}
with
\begin{equation}
\label{eq:8}
P_{\alpha\beta}(l_{vac},l)=|\langle\nu_{\beta}|Ue^{-iH_{vac}l_{vac}}U^{+}U^{m}e^{-iH_{m}l_{m}}U^{m +}|\nu_{\alpha} \rangle|^{2},
\end{equation}
where $L$ ($L_{vac}$) is the propagation distance in matter (vacuum), $H_{m}$ ($H_{vac}$) is the diagonal neutrino Hamiltonian matrix in matter (vacuum).
When the distance is much larger than the corresponding oscillation length, i.e., $L\gg L^{osc}$, $L_{vac}\gg L^{osc}_{vac}$, the averaged oscillation probability could
be written as follows
\begin{equation}
\label{eq:9}
\overline{P}^{\uppercase\expandafter{\romannumeral5}}_{\alpha\beta}=\sum_{i}\sum_{j}|U^{m}_{\alpha j}|^{2}P_{ji}|U_{\beta i}|^{2}
\end{equation}
with
\begin{equation}
\label{eq:10}
P_{ji}=|\sum_{\gamma }U^{*}_{\gamma i}U_{\gamma j}^{m}|^{2}.
\end{equation}
In contrast to Scheme \uppercase\expandafter{\romannumeral1}, besides the first decoherence that changes the flavor state to the eigen-state of the Hamiltonian in matter,
there is another decoherence that changes the eigen-state of the Hamiltonian in matter to that in vacuum, i.e.,
$|\nu_{j}\rangle^{m}\rightarrow|\nu_{i}\rangle$.
Now we have introduced five typical deocherence schemes for the propagation problem of HANs in matter.
For the sake of inter-comparison,
their characteristic is listed in Tab.\ref{tab:1}.
\begin{table}
\caption{Characteristic of the decoherence schemes}
\label{tab:1}
\begin{tabular}{c c c c c c }
\noalign{\smallskip}\hline
~~Decoherence Scheme~~&~~\uppercase\expandafter{\romannumeral1}~~ &~~\uppercase\expandafter{\romannumeral2}~~&
~~\uppercase\expandafter{\romannumeral3}~~&~~ \uppercase\expandafter{\romannumeral4}~~&~~~ \uppercase\expandafter{\romannumeral5}~~ \\ \hline
\noalign{\smallskip}
~~Including adiabatic evolution~~ &~yes~&~~yes~~& ~no~&~~no~~&~no~\\
~~Including oscillation effects~~&~~no~~&~~yes~~~&~yes~~& ~~yes~&~~no~~\\
~~Times of decoherence in the propagation~~&~~one~~&~~two~~&~~two~~& ~~one~~&~~two~~\\
\hline
\noalign{\smallskip}
\end{tabular}
\vspace*{0.6cm}
\end{table}
In the following section we would examine the predictions from these schemes with specific parameters of HANs and matter.
\section{Predictions from FTMs}
As we have pointed out in Sec.\uppercase\expandafter{\romannumeral1}, we consider the matter effects in the gas environment around the promising source.
In this section we first analyse the dependence of the FTMs on the parameters of neutrinos and matter and then examine the predictions from the given FTMs.
\subsection{Dependence of the FTMs on the parameters of neutrinos and matter}
The specific parameters of the HANs and matter for the corresponding propagation scheme are listed in Tab.\ref{tab:2}.
\begin{table}
\caption{Parameters of neutrinos and matter for the propagation schemes}
\label{tab:2}
\begin{tabular}{c c c c c }
\noalign{\smallskip}\hline
~Scheme~&~Distance in vacuum ($L_{vac}$)~&~Distance in matter($L$) ~ &~Neutrino energy ($E$) ~~&
~Electron density ($N_{e}$)~~\\ \hline
\noalign{\smallskip}
\uppercase\expandafter{\romannumeral2}, \uppercase\expandafter{\romannumeral3}, \uppercase\expandafter{\romannumeral4}~&~$\gg L_{vac}^{osc}$~ &~[$10^{5}$, $4\times10^{8}$]km~&~[$10^{2}$, $10^{3}$]TeV~& ~[$10^{21}$, $10^{22}$] cm$^{-3}$~\\
\uppercase\expandafter{\romannumeral1}, \uppercase\expandafter{\romannumeral5}~&~$\gg L_{vac}^{osc}$~&~$\gg L^{osc}$~&~[$10^{2}$, $10^{3}$]TeV~& ~[$10^{21}$, $10^{22}$] cm$^{-3}$~\\
\hline
\noalign{\smallskip}
\end{tabular}
\vspace*{0.6cm}
\end{table}
We give comments on the ranges of the parameters as follows: (a). The range of $L/E$ is $[0.1, 4000]$km/GeV. The oscillation effects become noticeable when $L/E$ if of order of $10^{3}$ km/GeV.
When $L/E\ll10^{3}$km/GeV, Scheme \uppercase\expandafter{\romannumeral2} would
be reduced to \uppercase\expandafter{\romannumeral1}, with \uppercase\expandafter{\romannumeral3}, \uppercase\expandafter{\romannumeral4} reduced to the standard scheme. (b). The range of the paramerter $A_{cc}\equiv2\sqrt{2}EG_{F}N_{e}$ is $[0.0254, 2.54]$eV$^{2}$. So there is no matter resonance for the mixing angle $\theta_{ij}$ with $ij=12, 13, 23$.
For the given range of the neutrino energy, the resonance takes place in dilute gas. For example, if $E$ takes 1PeV, the resonance density $N_{e}$ for $\theta_{12}$ is $2\times10^{17}$cm$^{-3}$.
We note that Ref.\cite{50} has discussed the resonance effect with Scheme \uppercase\expandafter{\romannumeral5}. So we don't consider the case of dilute gas here.
On the basis of the ranges of the parameters, the electron neutrinos are decoupled. The neutrino (antineutrinos) oscillation probabilities are of simple forms, i.e.\cite{52},
\begin{equation}
\label{eq:11}
\begin{array}{c}
P^{m}_{\nu_{e}\nu_{e}}\approx P^{m}_{\overline{\nu}_{e}\overline{\nu}_{e}}\approx1,\\
~~P^{m}_{\nu_{e}\nu_{\mu}}\approx P^{m}_{\overline{\nu}_{e}\overline{\nu}_{\mu}}\approx
P^{m}_{\nu_{e}\nu_{\tau}}\approx P^{m}_{\overline{\nu}_{e}\overline{\nu}_{\tau}}\approx P^{m}_{\nu_{\mu}\nu_{e}}\approx P^{m}_{\overline{\nu}_{\mu}\overline{\nu}_{e}}\approx
P^{m}_{\nu_{\tau}\nu_{e}}\approx P^{m}_{\overline{\nu}_{\tau}\overline{\nu}_{e}}\approx0,\\
P^{m}_{\nu_{\mu}\nu_{\mu}}\approx P^{m}_{\overline{\nu}_{\mu}\overline{\nu}_{\mu}}\approx P^{m}_{\nu_{\tau}\nu_{\tau}}\approx P^{m}_{\overline{\nu}_{\tau}\overline{\nu}_{\tau}}\approx1-\sin^{2}2\widetilde{\theta}\sin^{2}\frac{\Delta \widetilde{m}^{2}_{32}L}{4E},\\
P^{m}_{\nu_{\mu}\nu_{\tau}}\approx P^{m}_{\overline{\nu}_{\mu}\overline{\nu}_{\tau}}\approx P^{m}_{\nu_{\tau}\nu_{\mu}}\approx P^{m}_{\overline{\nu}_{\tau}\overline{\nu}_{\mu}}\approx
\sin^{2}2\widetilde{\theta}\sin^{2}\frac{\Delta \widetilde{m}^{2}_{32}L}{4E},
\end{array}
\end{equation}
where $\widetilde{\theta}$ and $\widetilde{m}^{2}_{32}$ are effective oscillation parameters in matter written as \cite{52}
\begin{equation}
\label{eq:12}
\tan2\widetilde{\theta}=\frac{2|\Delta m^{2}_{21}U_{\mu2}U^{*}_{\tau2}+\Delta m^{2}_{31}U_{\mu3}U^{*}_{\tau3}|}{\Delta m^{2}_{21}(|U_{\tau2}|^{2}-|U_{\mu2}|^{2})+\Delta m^{2}_{31}(|U_{\tau3}|^{2}-|U_{\mu3}|^{2})},
\end{equation}
\begin{equation}
\label{eq:13}
\Delta \widetilde{m}^{2}_{32}\approx\pm\sqrt{(\Delta m^{2}_{31}c^{2}_{13}-\Delta m^{2}_{21}(c^{2}_{12}-s^{2}_{12}s^{2}_{13}))^{2}+(\Delta m^{2}_{21}\sin2\theta_{12}s_{13})^{2}},
\end{equation}
with $s_{ij}=\sin\theta_{ij}$, $c_{ij}=\cos\theta_{ij}$.
From the above expressions, we find that the oscillation probabilities of neutrinos approximate those of antineutrinos. They are not dependent on the parameter $A_{cc}$.
Neither are the FTMs in the given ranges of the parameters, see Fig.\ref{fig:2} and Fig.\ref{fig:3}.
\begin{figure}[t]
\label{fig:2}
\centering
\includegraphics[width=.49\textwidth]{fig2a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig2b.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig2c.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig2d.pdf}
\caption{\label{fig:2} Relation between the magnitude of the element in FTMs and the parameter $A_{cc}$ based on the best fit data with the normal neutrino mass ordering\cite{53}. Scheme \uppercase\expandafter{\romannumeral1}: red,
Scheme \uppercase\expandafter{\romannumeral2}: green, Scheme \uppercase\expandafter{\romannumeral3}: magenta, Scheme \uppercase\expandafter{\romannumeral4}: yellow,
Scheme \uppercase\expandafter{\romannumeral5}: blue. Neutrinos: real lines, Antineutrinos: dashed lines.}
\end{figure}
\begin{figure}[t]
\label{fig:3}
\centering
\includegraphics[width=.49\textwidth]{fig3a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig3b.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig3c.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig3d.pdf}
\caption{\label{fig:3} Relation between the magnitude of the element in FTMs and the parameter $A_{cc}$ based on the best fit data with the inverted neutrino mass ordering\cite{53}.
The convention of color is the same as that in Fig.\ref{fig:2}.}
\end{figure}
However, we can see that except $\overline{P}_{\nu_{e}\nu_{e}}$ ($\overline{P}_{\overline{\nu}_{e}\overline{\nu}_{e}}$), the elements in the FTM are dependent on the mass ordering
of neutrinos. Correspondingly, the flavor ratio at Earth is affected by the mass ordering of neutrinos.
Now we examine the oscillations of the elements in FTMs with the parameter $L/E$ in the cases of Scheme \uppercase\expandafter{\romannumeral2}, \uppercase\expandafter{\romannumeral3}, \uppercase\expandafter{\romannumeral4}.
\begin{figure}[t]
\label{fig:4}
\centering
\includegraphics[width=.48\textwidth]{fig4a.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4b.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4c.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4d.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4e.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig4f.pdf}
\caption{\label{fig:4} Oscillation of the element in FTMs based on the best fit data with the normal neutrino mass ordering\cite{53}.
Scheme \uppercase\expandafter{\romannumeral2}: green, Scheme \uppercase\expandafter{\romannumeral3}: magenta, Scheme \uppercase\expandafter{\romannumeral4}: yellow.}
\end{figure}
\begin{figure}[t]
\label{fig:5}
\centering
\includegraphics[width=.48\textwidth]{fig5a.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5b.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5c.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5d.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5e.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig5f.pdf}
\caption{\label{fig:5} Oscillation of the element in FTMs based on the best fit data with the inverted neutrino mass ordering\cite{53}.
The convention of color is the same as that in Fig.\ref{fig:4}.}
\end{figure}
They are shown in Fig.\ref{fig:4} and Fig.\ref{fig:5}.
Some necessary comments are given as follows:
(a). The FTMs of neutrinos are nearly identical to those of antineutrinos. So we just show the oscillations of neutrinos. (b). The elements in the first row of the FTM, namely $\overline{P}_{\nu_{e}\nu_{e}}$, $\overline{P}_{\nu_{e}\nu_{\mu}}$, $\overline{P}_{\nu_{e}\nu_{\tau}}$, do not oscillation with the parameter $L/E$. (c). The oscillation amplitudes
are small for most elements in the FTMs except for $\overline{P}_{\nu_{\mu}\nu_{e}}$, $\overline{P}_{\nu_{\mu}\nu_{\mu}}$,
$\overline{P}_{\nu_{\tau}\nu_{e}}$ in Scheme \uppercase\expandafter{\romannumeral4} with the inverted neutrino mass ordering. Correspondingly, the most noticeable amplitude of the apparent T violation (ATV), namely
$\Delta _{e\mu}\equiv\overline{P}_{\nu_{e}\nu_{\mu}}-\overline{P}_{\nu_{\mu}\nu_{e}}$,
also appears in that case, see Fig.\ref{fig:6}.
\begin{figure}[t]
\label{fig:6}
\centering
\includegraphics[width=.48\textwidth]{fig6a.pdf}
\hfill
\includegraphics[width=.48\textwidth]{fig6b.pdf}
\caption{\label{fig:6} Oscillation of the the apparent T violation $\Delta _{e\mu}\equiv\overline{P}_{\nu_{e}\nu_{\mu}}-\overline{P}_{\nu_{\mu}\nu_{e}}$ based on the best fit data \cite{43}.
Left panel: Normal mass ordering, Right panel: Inverted mass ordering. The convention of color is the same as that in Fig.\ref{fig:4}.}
\end{figure}
\subsection{Matter effects on flavor ratio at Earth}
Now we examine the matter effects on the flavor ratio of HANs at Earth.
We consider three typical sources, i.e.,
\begin{equation}
\label{eq:14}
\begin{array}{c}
neutron-decay ~~source ~~with~~\Phi^{s}=(1, ~~0,~~ 0),\\
pion-decay ~~source~~with ~~~\Phi^{s}=(1/3, ~~2/3, ~~0),\\
muon-damping~~ source~~with ~~~ \Phi^{s}=(0, ~~1, ~~0),
\end{array}
\end{equation}
where $\Phi^{s}$ is the flavor ratio at the source. Using $\Phi^{s}$ and the aforementioned FTMs, the flavor ratio of HANs at Earth are obtained. On the basis of the given parameters in Tab.\ref{tab:2}
and the recent global fit data of neutrinos oscillations\cite{53}, the ternary plots for the flavor ratio are shown in Fig.\ref{fig:7}-Fig.\ref{fig:9}.
For the sake of comparison, the flavor ratio predicted by the standard scheme is also shown in these figures.
From the figures, we could obtain following observations.
\begin{figure}[t]
\label{fig:7}
\centering
\includegraphics[width=.6\textwidth]{fig7a.pdf}
\caption{\label{fig:7} Ternary plot of the flavor ratio at Earth with the neutron-decay source on the basis of the global fit data
at $3\sigma$ level\cite{53} with the normal mass ordering and the parameters in Tab.\ref{tab:2}. The 2015 $3\sigma$ constraint is taken from Ref.\cite{54}. The 2020 and 2040 $3\sigma$ credible regions
with the pion-decay source
based on IceCube and
IceCube-Gen2 are taken from Ref.\cite{21}.
The $3\sigma$ credible region with the pion-decay source based on all TeV-PeV neutrino telescopes available in 2040 is taken from Ref.\cite{15}.
}
\end{figure}
\begin{figure}[t]
\label{fig:8}
\centering
\includegraphics[width=.49\textwidth]{fig8a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig8b.pdf}
\caption{\label{fig:8} Ternary plot of the flavor ratio at Earth with the pion-decay source.
Left panel: Normal mass ordering, Right panel: Inverted mass ordering. The conventions of parameters are the same as those in Fig.\ref{fig:7}.}
\end{figure}
\begin{figure}[t]
\label{fig:9}
\centering
\includegraphics[width=.49\textwidth]{fig9a.pdf}
\hfill
\includegraphics[width=.49\textwidth]{fig9b.pdf}
\caption{\label{fig:9} Ternary plot of the flavor ratio at Earth with the muon-damping source.
Left panel: Normal mass ordering, Right panel: Inverted mass ordering. The conventions of parameters are the same as those in Fig.\ref{fig:7}.}
\end{figure}
For the neutron-decay source (see Fig.\ref{fig:7}), because the mass ordering of neutrinos and the oscillation effects do not impact the flavor transition
of $\nu_{e}$ ($\overline{\nu}_{e}$) in the given ranges of the parameters, the predicted flavor ratio from the FTMs could be classified into three regions as follows.
(a). The red region from Scheme \uppercase\expandafter{\romannumeral1}: It is the same as the green from Scheme \uppercase\expandafter{\romannumeral2}.
(b). The magenta region from Scheme \uppercase\expandafter{\romannumeral3}: It is identical to the yellow from Scheme \uppercase\expandafter{\romannumeral3} or the cyan from the standard scheme.
(c). The blue region from Scheme \uppercase\expandafter{\romannumeral5}: It overlaps with that from Scheme \uppercase\expandafter{\romannumeral3}.
All the predicted regions are at the margin of the 2015 ICeCube $3\sigma$ constraint\cite{54}. Considering the 2040 expected sensitivity on the flavor ratio\cite{15,21},
these regions would be precluded by the observations of the neutrino telescopes in the near future.
For the pion-decay source (see Fig.\ref{fig:8}), the predicted flavor could be classified into four regions as follows.
(a). The region from Scheme \uppercase\expandafter{\romannumeral1}:
Since the oscillation effect is not notable for Scheme \uppercase\expandafter{\romannumeral2}, this region is nearly identical to that from Scheme \uppercase\expandafter{\romannumeral2}.
Nearly half of the area overlaps with the 2040 $3\sigma$ credible region \cite{15}. So Scheme \uppercase\expandafter{\romannumeral1} and \uppercase\expandafter{\romannumeral2}
might work in the era of precise measurement of the flavor ratio.
(b). The region from scheme \uppercase\expandafter{\romannumeral3}: Compared with the prediction from the standard scheme, this region shows that the oscillation effects are not negligible.
However, the main part of this region is outside the $3\sigma$ credible region \cite{15}. So Scheme \uppercase\expandafter{\romannumeral3} with this source
would be stringently constrained by the observations of the neutrino telescopes in the future.
(c). The region from scheme \uppercase\expandafter{\romannumeral4}: Although the oscillation amplitudes of the elements in the FTM are noticeable with
the inverted neutrino mass ordering, the variation of the parameters in Tab.\ref{tab:2} and the global fit data of neutrinos undermine the oscillation effects.
So the predicted region from Scheme \uppercase\expandafter{\romannumeral4} approximates that from the standard scheme.
Considering the 2040 $3\sigma$ credible region \cite{15}, Scheme \uppercase\expandafter{\romannumeral4} and the standard scheme are as promising as Scheme \uppercase\expandafter{\romannumeral1},
\uppercase\expandafter{\romannumeral2}.
(d). The region from Scheme \uppercase\expandafter{\romannumeral5}: The mass ordering of neutrinos impacts the prediction of Scheme \uppercase\expandafter{\romannumeral5} noticeably.
Given the $3\sigma$ credible region \cite{15}, the prediction with the normal ordering is more favored. In this case, the blue region overlaps with the yellow, while its area is smaller.
In general, for this source, the identification of the matter effects is difficult because the region from the standard scheme noticeably overlaps with the regions from the other schemes.
For the muon-damping source (see Fig.\ref{fig:9}),
the predicted flavor could also be classified into four regions as those illustrated in case of the pion-decay source.
Even though the condition of the matter resonance is not satisfied, the matter effects in this case are noticeable for the Scheme
\uppercase\expandafter{\romannumeral1}, \uppercase\expandafter{\romannumeral2}, \uppercase\expandafter{\romannumeral3} with both mass orderings,
and Scheme \uppercase\expandafter{\romannumeral5} with the inverted mass ordering. Furthermore, because the region from Scheme \uppercase\expandafter{\romannumeral1}
or \uppercase\expandafter{\romannumeral2} does not overlap with that from the standard scheme or Scheme \uppercase\expandafter{\romannumeral4},
the matter effects from the former might be identified by the neutrinos telescopes in the near future.
On the basis of the above observations on the flavor ratio of the HANs at Earth,
we can see that even in gas where the matter resonance is not available the matter effects could not be neglected in general.
Although the matter effects are dependent on the decoherence schemes and the production sources, the identification of them is possible
in the near future. Then the precise measurement of the flavor ratio would set constraints on the production environment and the propagation path.
\section{Conclusions}
A precise measurement of the flavor ratio of HANs is available in the near future.
To test the matter effects on the flavor transition of HANs by the next-generation neutrino telescopes deserves serious consideration.
The matter effects are dependent on the propagation scheme of the HANs.
We introduced five typical decoherence schemes for the HANs in the propagation.
On the basis of the propagation parameters of the HANs in the realistic environment around the promising sources,
the impacts of the matter effects on the FTM were analysed.
When the matter resonance condition is not satisfied, the oscillation effect is negligible for the schemes except for the special one (Scheme \uppercase\expandafter{\romannumeral3}).
Even so, the matter effects shown through the flavor ratio of the HANs at Earth are noticeable for the typical sources. Particularly, on the base of the schemes (\uppercase\expandafter{\romannumeral1},
\uppercase\expandafter{\romannumeral2}) the identification of the matter effects
is promising in the case that the HANs are from the muon-damping source.
Therefore, from the measured flavor ratio in the near future, the matter parameters and propagation path might be deduced, which could give us clues on the production mechanism
of HANs.
\vspace{0.08cm}
\acknowledgments
This work is supported by the National Natural Science Foundation of China under grant No. 12065007, the Guangxi Scientific Programm Foundation under grant No. Guike AD19110045, the Research Foundation of Guilin University of Technology under grant No. GUTQDJJ2018103.\\
|
\section{Introduction}
Minimizing power consumption is one of the most significant objectives for current digital circuit design. In many applications, such as image or signal processing, the users are also often willing to accept certain errors if they are accompanied by reduced power consumption or other improvements. This paradigm, called approximate computing~\cite{jieOrshanskyAXcomp}, systematically tries to find the best trade-offs between the quality of output and hardware parameters such as area, power consumption, or latency. One of the most widely used methods is replacing the standard arithmetic circuits with approximate implementations~\cite{jieOrshanskyAXcomp}. The circuit approximation techniques can be classified as manual (e.g.,~\cite{jiang}) and automated (e.g.,~\cite{hashemi2018blasys,ceska:iccad17}). From a broader perspective, the automated methods usually provide better trade-offs between the error and other circuit parameters. These methods typically start with an exact circuit and modify it using the following iterative scheme: (i) generate candidate approximate circuits, (ii) evaluate these circuits, and (iii) select the best candidate for the next iteration. As many candidate circuits are usually generated, it is essential to quickly analyze their error~\cite{ceska:iccad17}.
Let $f: \mathbb{B}^n \rightarrow \mathbb{B}^m$ be an $n$-input $m$-output Boolean function that describes the correct functionality (the accurate function) and $f': \mathbb{B}^n \rightarrow \mathbb{B}^m$ be an approximation of it with error $\varepsilon$, formally $\forall x \in \mathbb{B}^n: f(x) \approx f'(x); f'(x) = f(x) + \varepsilon(x)$. If there is no analytical model of the error function $\varepsilon$ (which can, in fact, be easily derived only for a few approximation techniques such as \textit{bit-truncation}), the error of $f'$ has to be determined for all possible input combinations, i.e. $|\mathbb{B}^n| = 2^n$. The evaluation time grows exponentially with $n$, but can effectively be reduced by parallel circuit simulation~\cite{hrbacek:gecco14} or formal verification techniques such as SAT, BDDs or algebraic solvers~\cite{ceska:iccad17,keszocze:access2022}.
The formal verification techniques achieve very good performance for the error evaluation~\cite{ceska:iccad17}. However, some of them do not scale for some types of circuits or error metrics. For example, SAT solving is applicable for determining the worst-case error but useless for the mean error. BDD-based methods are not applicable to structurally complex circuits such as multipliers.
In this paper, we deal with efficient error analysis methods using BDDs.
The calculation is based on the so-called characteristic function $|f - f'|$ represented as a Boolean equation~\cite{soeken:bdd}. Keszocze~\cite{keszocze:access2022} introduced a technique that does not need to compute the absolute value which represents the performance bottleneck. On the other hand, this algorithm assumes that $\forall x: f' (x) \le f(x)$ and can produce only such solutions that satisfy this assumption. However, this assumption is not realistic is general-purpose approximation engines (such as~\cite{ceska:iccad17}) in which arbitrary candidate approximations must be evaluated.
This paper aims to propose new algorithms for the exact error analysis that further simplify the characteristic function without postulating any specific constraint on the evaluated function $f'$.
We propose four error calculation algorithms (two for the worst-case error and two for the mean-absolute error). These algorithms omit the absolute value calculation while the average error evaluation time is reduced $3\times$ compared to the standard BDD-based approach. Moreover, they also reduce the memory complexity, and in some cases, the speedup is $10-30\times$.
The proposed error analysis algorithms were employed in a circuit approximation method based on CGP, which is inspired in~\cite{ceska:iccad17}.
In our experiments, the error evaluation was executed for billions of adders.
The resulting approximate adders exhibit better trade-offs than the approximate adders published in the EvoApprox library~\cite{mrazek:date16lib}. All approximate implementations are available for download (TBD). The C++ implementation of the error analysis algorithms is provided as an open-source software \textcolor{blue}{\url{https://github.com/ehw-fit/bdd-evaluation}}.
\section{Approximate computing}
Approximate computing aims to introduce an error into a software or hardware implementation. It can be done on various levels: algorithmic, functional, or physical. We focus on the functional approximation (i.e., modification of Boolean functions describing the logic behavior of digital circuits) of arithmetic circuits that may be used as basic building blocks in complex systems.
\subsection{Error metrics}
There are various application-independent error metrics used in the literature~\cite{jiang}. The \emph{worst-case arithmetic error}, sometimes denoted as \emph{error magnitude} or \emph{error significance}, is defined as
\begin{equation}
e_{wce}(f, f’) = \max_{\forall x \in \mathbb{B}^n} |int(f(x)) - int (f’(x))|,
\end{equation}
where $int(x)$ represents a function $int: \mathbb{B}^m \rightarrow \mathbb{Z}$ returning an integer value of the $m$-bit binary vector $x$. Typically, a natural unsigned binary representation is considered, i.e. $ int(x) = \sum_{i=1}^{m}{2^i \cdot x_i}$. The worst-case error represents the fundamental metric that is useful to guarantee that the approximate output differs from the correct output by at most the error bound $e$.
The \emph{average-case arithmetic error} (also known as \emph{mean absolute error}) is defined as the sum of absolute differences in magnitude between the original and approximate circuit, averaged over all inputs:
\begin{equation}
e_{mae}(f, f’) = 2^{-n} \sum_{\forall x \in \mathbb{B}^n} {| int(f(x)) - int(f’(x))| }.
\end{equation}
Note that the values produced by the ``absolute'' error metrics $e_{mae}$ and $e_{wce}$ can be very high. Hence, these values can be expressed relatively to the output range, using division by $2^{m} - 1$, i.e. the maximal output value.
In the metric error formulas, the enumeration of all possible input vectors is present. For $n$ greater than approx. 24, it is not feasible to enumerate $\mathbb{B}^n$ in a reasonable time even when a highly parallel circuit simulator is employed.
This issue can partly be eliminated by utilizing a formal verification approach for the error calculation. Formal techniques typically construct a virtual miter circuit to compute the characteristic function. Reduced Ordered Binary Decision Diagrams (BDD) or SAT solving-based methods were employed in the area of approximate circuits. ROBDDs are suitable for the formal error analysis of less complex circuits (such as adders) for which the BDD does not grow exponentially with $n$. SAT solving can effectively handle determining the worst-case error for up to 32-bit multipliers but does not provide suitable results for other error metrics. Details of BDDs are elaborated in Section~\ref{sec:bdds}.
\subsection{Automated design techniques}
The automated approximation algorithms iteratively modify a given accurate circuit to achieve the best trade-off between the error and other circuit parameters. The most recent algorithms use various approaches for circuit representation and error evaluation. Soeken et al. proposed an algorithm that directly modifies the BDD~\cite{soeken:bdd,keszocze:access2022}. This algorithm satisfies the assumption of the fast error evaluation, but its design space is limited. Another approach is to use some internal circuit representation. Hashemi et al. divided the original circuit into subcircuits and approximated each of them. Finally, the error was evaluated for many candidate solutions~\cite{hashemi2018blasys}. A specific internal circuit representation is used in \textit{Cartesian Genetic Programming} (CGP)~\cite{ceska:iccad17} which is a heuristic design-space exploration algorithm. Candidate circuits are represented as directed acyclic graphs in a fixed grid of nodes. The algorithm randomly changes the connections between nodes and the nodes' functions. It uses the following fitness function for a given error metric $e$ (e.g., $e_{mae}$ or $e_{wce}$) and the acceptable error threshold $\tau$
\begin{equation}\label{eq:fitness}
\mathbf{F}(f') =
\begin{cases}
size(f') & \text{if } e(|f - f'|) \leq \tau \\
\infty & \text{otherwise} \\
\end{cases}.
\end{equation}
This algorithm starts with an accurate solution and reduces its size while keeping the error constraint satisfied. The result of the search is the smallest circuit satisfying the constraint $\tau$.
\subsection{BDDs}
\label{sec:bdds}
This work exploits the Reduced Ordered Binary Decision Diagrams (ROBDDs, also referred to as BDDs). ROBDD is a canonical rooted acyclic graph-based representation of a Boolean function, where each \textbf{node} represents either an input variable or terminal node, and the edges represent the value assignment. There are two terminal nodes -- \textit{true} and \textit{false}. ROBDDs have been traditionally used to solve the equivalence checking problem due to their canonical property (i.e., the ROBDDs of two logic functions are isomorphic if the functions are functionally equivalent and the input variables have the same order). The tools, developed for constructing and manipulating ROBDDs, can exactly answer the following questions: (1) whether a variable (signal) in the circuit is satisfiable, i.e., whether an input vector exists which generates a true value of the signal; and (2) the probability that a selected variable is true. Although these questions are answered fast (by finding some paths in the ROBDD), the construction of the ROBDD is not trivial, and in some circuits such as multipliers, the ROBDDs grow exponentially with $n$.
Since there is no significant difference in the performance of major tools operating with ROBDDs~\cite{dijkBDD}, we selected BuDDy and CUDD tools, which also support dynamic variables reordering for evaluation. In these tools, the BDD node is represented by a structure. Starting with an input node, these tools allow adding new nodes by applying logical operations ($\neg, \wedge, \vee, \oplus, \dots$) to two existing nodes. Therefore, one can construct an arbitrary Boolean function.
\subsubsection{Conventions and notations}
Henceforth, we denote an $n$-input $m$-output Boolean function as $f$. Without loss of generality the functions can be represented by the BDDs $f=(f_n, f_{n-1}, \dots f_2, f_1)$ where each output bit $f_i$ is a BDD node. For integer functions, $\sign{f}$ denotes the sign bit $f_n$.
For one-output function (i.e., BDD node) the tools can provide the probability that a selected node $f_i$ is true
$$\satp{f_i} = 2^{-n} \cdot |\{x : f_i(x) = \textrm{True}\}| $$
and a Boolean value $\sat{f_i}$ answering whether the node $f_i$ becomes \textit{true} for any input.
We also implemented two arithmetic Boolean functions as BDDs. The Addition function $\add{a, b}$ creates a Ripple-carry adder with output nodes of functions $a$ and $b$ as inputs. Similarly, subtraction $\subtract{a,b}$ is constructed in the same way, but the $b$ outputs are inverted, and the carry input is set. Note that the input vectors must be (sign-) extended to handle both signed and unsigned variants of $a$ and $b$.
\subsubsection{Existing algorithms (\textbf{BASELINE})}
In \cite{soeken:bdd}, Soeken et al. proposed the error analysis based on calculating characteristic function $r = |f(x) - f'(x)|$. For the resulting vector $r$ they calculate the maximal (wce) and average (mae) values.
Their best-performing algorithm is based on a binary search (Alg. \ref{alg:wce_baseline}). Firstly, the values are subtracted (ln. 1), which is captured by a signed function $\varepsilon$, and the absolute value is obtained (unsigned $r$). The absolute value has to deal with the two's complement and increments the value if the result is negative (ln. 4). The binary search starts with the MSB, computes the worst-case error $wce$ and also the function $\mu: \mu(x) = 1$ if and only if $r(x) = wce$.
\begin{algorithm}[ht]
\caption{Baseline WCE calculation using two's complement}\label{alg:wce_baseline}
\Input{Functions $f$ and $f'$ as ROBDDs}
\Output{Worst-case error ($wce$)}
$\varepsilon \gets \subtract{f, f'}$ \tcp*{subtract results}
\For{$i \gets 1$ \KwTo $|\varepsilon| - 1$} {
$r_i \gets \varepsilon_i \oplus \sign{\varepsilon} $ \tcp*{absolute value}
}
$r \gets \add{r, \{\sign{\varepsilon}\}}$ \tcp*{two's complement}
$ \mu \gets $ True\;
$ wce \gets 0$\;
\For(\tcp*[h]{binary search}){$i \gets |r|$ \KwTo $1$ }{
\If{$\sat{\mu \wedge r_i}$} {
$wce \gets wce + 2^{i - 1}$\;
$\mu \gets \mu \wedge r_i$\;
}
}
\Return $wce$\;
\end{algorithm}
The average (mean absolute error) is calculated (Alg. \ref{alg:mae_baseline}) with the same characteristic function $r$. The weighted sum of positive-bit probabilities is computed. Finally, the sum is divided by $2^n$ while $n$ is the number of inputs of functions $f$ and $f'$.
\begin{algorithm}[ht]
\caption{Baseline MAE calculation using two's complement}\label{alg:mae_baseline}
\Input{Functions $f$ and $f'$ as ROBDDs}
\Output{Mean absolute error ($mae$)}
$\varepsilon \gets \subtract{f, f'}$ \tcp*{subtract results}
\For{$i \gets 1$ \KwTo $|\varepsilon| - 1$} {
$r_i \gets \varepsilon_i \oplus \sign{\varepsilon} $ \tcp*{absolute value}
}
$r \gets \add{r, \{\sign{\varepsilon}\}}$ \tcp*{two's complement}
$ mae \gets 0$\;
\For{$i \gets 1$ \KwTo $|r|$} {
$mae \gets mae + 2^{i-1} \cdot \satp{r_i}$ \;
}
\Return $mae$ \;
\end{algorithm}
Keszocze showed that the subtraction and absolute value calculation are the most time-consuming parts of the evaluation~\cite{keszocze:access2022},. He also proposed a fast algorithm for determining the MAE without the characteristic function. This algorithm needs to satisfy a condition that $\forall x: f'(x) \leq f(x)$. Automated iterative approximation algorithms typically cannot satisfy this condition, therefore, we are not able to employ this algorithm in a CGP-based approximation engine.
\section{Proposed MAE and WCE calculation}
We present improvements of the baseline error evaluation algorithms to omit the absolute value calculation. We propose two approaches --- one replaces the absolute value by approximation in ones' complement, the second one handles the positive and negative branches separately.
\subsection{Ones' complement \textbf{(ONES')}}
In the baseline Algorithms \ref{alg:wce_baseline},\ref{alg:mae_baseline} (ln. 4), the increment of $r$ vector by a sign bit is needless. In the proposed algorithm (Alg. \ref{alg:wce_ones}), we omitted this addition. We obtain the exact $wce$ for positive $r$; and ${wce}$ decreased by $1$ if the maximal error is in the negative $r$. Because this condition is expressed in the MSB node of $\varepsilon$, the error can be compensated (ln. 10 - 11).
\begin{algorithm}[ht]
\caption{Proposed WCE calculation using ones' complement}\label{alg:wce_ones}
\Input{Functions $f$ and $f'$ as ROBDDs}
\Output{Worst-case error (${wce}$)}
$\varepsilon \gets \subtract{f, f'}$ \tcp*{subtract results}
\For{$i \gets 1$ \KwTo $|\varepsilon| - 1$} {
$r_i \gets \varepsilon_i \oplus \sign{\varepsilon} $ \tcp*{absolute value}
}
$ \mu \gets $ True\;
$ wce \gets 0$\;
\For(\tcp*[h]{binary search}){$i \gets |r|$ \KwTo $1$ }{
\If{$\sat{\mu \wedge r_i}$} {
$wce \gets wce + 2^{i - 1}$\;
$ \mu \gets \mu \wedge r_i$\;
}
}
\If(\tcp*[h]{complement corr.}){$\sat{\mu \wedge \sign{\varepsilon}}$} {
$wce \gets wce + 1$ \;
}
\Return $wce$\;
\end{algorithm}
A similar approach is used in the average error calculation (Alg. \ref{alg:mae_ones}). The error by 1 can be corrected by adding the proportion of negative numbers $\varepsilon$, i.e. $\satp{\sign{\varepsilon}}$ (ln. 7).
\begin{algorithm}[ht]
\caption{Proposed MAE calculation using ones' complement}\label{alg:mae_ones}
\Input{Functions $f$ and $f'$ as ROBDDs}
\Output{Mean absolute error ($mae$)}
$\varepsilon \gets \subtract{f, f'}$ \tcp*{subtract results}
\For{$i \gets 1$ \KwTo $|\varepsilon| - 1$} {
$r_i \gets \varepsilon_i \oplus \sign{\varepsilon} $ \tcp*{absolute value}
}
$ mae \gets 0$\;
\For{$i \gets 1$ \KwTo $|r|$} {
$mae \gets mae + 2^{i-1} \cdot \satp{r_i}$ \;
}
$mae \gets mae + \satp{\sign{\varepsilon}} $ \tcp*{correction}
\Return $mae$ \;
\end{algorithm}
\subsection{Omitting the absolute values \textbf{(NOABS)}}
The second proposed approach exploits the fact that we can build two BDD trees for the resulting vector $\varepsilon$ --- one for the positive part $\neg \sign{\varepsilon} \wedge \varepsilon_i$ and one for the negative part $\sign{\varepsilon} \wedge \varepsilon_i$. The worst-case error calculation is shown in Alg. \ref{alg:wce_noabs}. Since the absolute error is calculated using XOR operations which make the resulting BDDs more complex, the WCE value is calculated directly from the signed result of subtraction $\varepsilon$. If the sign bit is negative, a binary search of the maximal value is performed as $wce_p$. On the other hand, the negative part is transformed to searching the minimal value $wce_n$. The subtraction result $\varepsilon$ is expressed as two's complement. Therefore, the negative $wce_n$ is incremented by 1, and the maximal value is returned (ln. 14).
\begin{algorithm}[ht]
\caption{Proposed WCE calculation without the absolute value (NOABS)}\label{alg:wce_noabs}
\Input{Functions $f$ and $f'$ as ROBDDs}
\Output{Worst-case error ($wce$)}
$\varepsilon \gets \subtract{f, f'}$ \tcp*{subtract results}
\tcc{positive error}
$ \mu_p \gets \neg \sign{\varepsilon}$\;
$ wce_p \gets 0$\;
\For(\tcp*[h]{binary search}){$i \gets |\varepsilon| - 1$ \KwTo $1$ }{
\If{$\sat{\mu_p \wedge \varepsilon_i}$} {
$wce_p \gets wce_p + 2^{i - 1}$\;
$\mu_p \gets \mu_p \wedge \varepsilon_i$\;
}
}
\tcc{negative error}
$ \mu_n \gets \sign{\varepsilon}$\;
$ wce_n \gets 0$\;
\For(\tcp*[h]{binary search}){$i \gets |\varepsilon| - 1$ \KwTo $1$ }{
\If{$\sat{\mu_n \wedge \neg \epsilon_i}$} {
$wce_n \gets wce_n + 2^{i - 1}$\;
$\mu_n \gets \mu_n \wedge \neg \varepsilon_i$\;
}
}
\Return $\max(wce_p, wce_n + 1)$\;
\end{algorithm}
The mean average error calculation (Alg. \ref{alg:mae_noabs}) employs two BDD trees --- for the positive part anded with the unset sign bit $\neg \sign{\varepsilon}$ and for the negative inverted anded with a set sign bit $\sign{\varepsilon}$. The negative part introduces an error by one, which can easily be compensated by adding the probability at line 6.
\begin{algorithm}[ht]
\caption{Proposed MAE calculation without the absolute value (NOABS)}\label{alg:mae_noabs}
\Input{Functions $f$ and $f'$ as ROBDDs}
\Output{Mean absolute error ($mae$)}
$\varepsilon \gets \subtract{f, f'}$ \tcp*{subtract results}
$ mae \gets 0$\;
\For{$i \gets 1$ \KwTo $|\varepsilon| - 1$} {
$mae \gets mae + 2^{i-1} \cdot \satp{\varepsilon_i \wedge \neg \sign{\varepsilon}}$ \;
$mae \gets mae + 2^{i-1} \cdot \satp{\neg \varepsilon_i \wedge \sign{\varepsilon}}$ \;
}
$mae \gets mae + \satp{\sign{\varepsilon}} $ \tcp*{complement corr.}
\Return $mae$ \;
\end{algorithm}
All the proposed algorithms follow the scheme of WCE and MAE calculation introduced by Soeken et al.~\cite{soeken:bdd} However, the absolute value calculation is partially (in ones') or totally (in noabs) transformed from the BDD function to mathematical corrections.
\section{Experiments}
\subsection{Experimental setup}
To evaluate the quality of the proposed error calculation algorithms, we employed them in the CGP-based approximation of adders (both signed and unsigned). The overall scheme of the approximation engine is given in Figure \ref{fig:experiments}. In the search loop, tens of thousands of candidate approximate adders were generated and their errors calculated. We collected all results and analyzed them statistically.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{fig/experiments.pdf}
\caption{Experimental setup of CGP-based search engine employing the proposed error calculation algorithms.}
\label{fig:experiments}
\end{figure}
The key component of the calculation is the BDD library, which provides an abstraction to BDD trees. It implements the functions deciding the satisfiability $\sat{d}$ and calculating the probability $\satp{d}$. We compared two state-of-the-art libraries --- BuDDy and CuDD. The cache memory of BuDDy was set to $10^7$ accordingly to the documentation. Similarly, default size of CuDD cache ($2^{18} = 262,144$) was used.
The maximal allowed MAE and WCE errors that are the key parameters of CGP exponentially vary from 2 to $0.2 \cdot 2^{n+1}$ (20\% of the output range). As a CGP seed, three adder types are considered: carry-look-ahead, carry-skip, and ripple-carry adders, unsigned and signed versions, for bit-widths 8 to 32 (of each operand). Since the CGP approximation is a stochastic approach, 5 independent runs are executed for each setup. Finally, the best adders are synthesized using Synopsys Design Compiler with 45~nm technology.
All experiments were executed on a computer cluster with nodes having 2 CPU Intel Xeon E5-2670 2,6 GHz, 16-core, 20 MB cache, and 64 GB RAM. Each run was a single thread (with 2 GB of memory), and 32 runs were executed on the node. In total, 49,716 independent runs were performed to evaluate the above-mentioned setups.
\subsection{CGP Approximation with BDD-based error evaluation}
Table~\ref{tab:speedlib} shows that BuDDy library leads to a faster error evaluations than CuDD, except the \textit{baseline} algorithm for 32-bit adders. The results are aggregated over all the error thresholds.
On average, CuDD is 2x slower than BuDDy. The addition of 8 inputs ($n$) leads to a 4-20x slow down for both libraries.
\begin{table}[ht]
\centering
\caption{The average evaluation time [s] for different BDD libraries}
\label{tab:speedlib}\vspace{-1em}
\resizebox{\columnwidth}{!}{\begin{tabular}{cc|rr|rr|rr}
\toprule
\textbf{Error} & \textbf{Bit-width} & \multicolumn{2}{c|}{\bf Baseline} & \multicolumn{2}{c|}{\bf Ones'} & \multicolumn{2}{c}{\bf Noabs}\\
& $n/2$ &\textit{BuDDy} & \textit{CuDD} & \textit{BuDDy} & \textit{CuDD} & \textit{BuDDy} & \textit{CuDD}\\
\midrule
\multirow{6}{*}{MAE} & 8 & \textbf{0.0001} & 0.0003 & \textbf{0.0001} & 0.0002 & \textbf{0.0001} & 0.0003\\
& 12 & \textbf{0.0009} & 0.0017 & \textbf{0.0003} & 0.0009 & \textbf{0.0004} & 0.0013\\
& 16 & \textbf{0.0044} & 0.0084 & \textbf{0.0013} & 0.0039 & \textbf{0.0017} & 0.0056\\
& 20 & \textbf{0.1818} & 0.2066 & \textbf{0.0063} & 0.0223 & \textbf{0.0113} & 0.1041\\
& 24 & \textbf{1.4713} & 1.8158 & \textbf{0.2438} & 0.6416 & \textbf{0.3537} & 0.8289\\
& 32 & 7.3208 & \textbf{6.1494} & \textbf{3.5416} & 4.7556 & \textbf{4.1580} & 4.7180\\\midrule
\multirow{6}{*}{WCE} & 8 & \textbf{0.0001} & 0.0002 & \textbf{0.0001} & 0.0001 & \textbf{0.0001} & 0.0001\\
& 12 & \textbf{0.0007} & 0.0010 & \textbf{0.0002} & 0.0005 & \textbf{0.0002} & 0.0005\\
& 16 & \textbf{0.0038} & 0.0051 & \textbf{0.0011} & 0.0023 & \textbf{0.0009} & 0.0018\\
& 20 & \textbf{0.0446} & 0.0910 & \textbf{0.0055} & 0.0114 & \textbf{0.0038} & 0.0072\\
& 24 & \textbf{0.8281} & 0.9924 & \textbf{0.1044} & 0.1896 & \textbf{0.0267} & 0.0657\\
& 32 & 5.4822 & \textbf{4.1673} & \textbf{2.2896} & 2.8677 & \textbf{2.0623} & 2.2698\\
\bottomrule
\end{tabular}}
\end{table}
Henceforth, only the BuDDy library is used in the following experiments. Table \ref{tab:speedup} shows the average evaluation time for different error evaluation algorithms. Both proposed implementations achieve significant speedup compared to the baseline implementation. On the average, the \textit{ones'} implementation performs better for MAE calculation, but WCE calculation is executed faster without the absolute value (\textit{noabs}). We can see an enormous speedup for 20-bit and 24-bit adders for MAE and WCE, respectively. This is caused by the fact that the proposed algorithms produce smaller BDD trees that can fit the cache. Increasing the cache size could move this peek for larger sizes, but we are also limited by CPU caches and RAM.
\begin{table}[ht]
\centering
\caption{An average evaluation time and speedup with BuDDy library}
\label{tab:speedup}\vspace{-1em}
\resizebox{\columnwidth}{!}{
\begin{tabular}{cc|rrr|rr}
\toprule
\textbf{Error} & \textbf{Bit-width} & \multicolumn{3}{c|}{\bf Evaluation time [s]} & \multicolumn{2}{c}{\bf Speedup}\\
& $n/2$ &\textit{baseline} & \textit{ones'} & \textit{noabs} & \textit{ones'} & \textit{noabs} \\
\midrule
\multirow{6}{*}{MAE} & 8 & 0.0001 & \textbf{0.0001} & 0.0001 & \textbf{2.58$\times$} & 1.91$\times$ \\
& 12 & 0.0009 & \textbf{0.0003} & 0.0004 & \textbf{3.11$\times$} & 2.21$\times$ \\
& 16 & 0.0044 & \textbf{0.0013} & 0.0017 & \textbf{3.47$\times$} & 2.55$\times$ \\
& 20 & 0.1818 & \textbf{0.0063} & 0.0113 & \textbf{28.76$\times$} & 16.06$\times$ \\
& 24 & 1.4713 & \textbf{0.2438} & 0.3537 & \textbf{6.03$\times$} & 4.16$\times$ \\
& 32 & 7.3208 & \textbf{3.5416} & 4.1580 & \textbf{2.07$\times$} & 1.76$\times$ \\\midrule
\multirow{6}{*}{WCE} & 8 & 0.0001 & 0.0001 & \textbf{0.0001} & 2.37$\times$ & \textbf{2.48$\times$} \\
& 12 & 0.0007 & 0.0002 & \textbf{0.0002} & 2.84$\times$ & \textbf{3.19$\times$} \\
& 16 & 0.0038 & 0.0011 & \textbf{0.0009} & 3.33$\times$ & \textbf{4.20$\times$} \\
& 20 & 0.0446 & 0.0055 & \textbf{0.0038} & 8.13$\times$ & \textbf{11.89$\times$} \\
& 24 & 0.8281 & 0.1044 & \textbf{0.0267} & 7.93$\times$ & \textbf{31.04$\times$} \\
& 32 & 5.4822 & 2.2896 & \textbf{2.0623} & 2.39$\times$ & \textbf{2.66$\times$} \\
\bottomrule
\end{tabular}}
\end{table}
We separately analyzed the impact of signed and unsigned versions of adders. The signed adders need the subtractor in the characteristic function to be extended by one bit. This extension slows down the error calculation by 3 - 17\% for the proposed algorithms.
\subsection{Error characterisation of 16-bit adders}
We analyzed the behavior of the proposed error calculation algorithms in the design of 16-bit approximate adders in greater detail.
The average evaluation time depends on the selected error analysis algorithm, error threshold, and the seed circuit. Figure \ref{fig:speederror} shows the average evaluation time for different error thresholds. The slowest evaluations are in the middle part of thresholds regardless of the evaluation algorithm. It is caused by the following: (i) For small thresholds, the BDDs representing $f'$s are not much different from the golden circuits $f$.(ii) For large thresholds, BDDs representing $f'$s are simplified (because of the approximation), and their sizes are reduced. We can see the proposed algorithms achieved significantly faster evaluations than the baseline.
\begin{figure}[ht]
\centering
\vspace*{-1em}
\includegraphics[width=\columnwidth]{fig/speed_error.pdf}
\vspace*{-2em}
\caption{Effect of the threshold value $\tau$ on the average evaluation time in CGP-based approximations.}
\label{fig:speederror}
\end{figure}
One of the inputs of the approximation algorithm is the seed circuit. This circuit acts as a parent while all candidate solutions are derived using a chain of mutations. We use this circuit as a golden circuit $r$ in the characteristic function $\varepsilon = f - f'$. Interestingly, the speed is not related to the size of the circuit (CLA:171, CSkA: 129, and RCA: 77 nodes, signed versions have two additional nodes). Although the RCA is the smallest one (and produces more approximated circuits), the evaluation speed is lower.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{fig/seeds.pdf}
\vspace*{-2em}
\caption{Effect of selected seed and golden model implementation for the average evaluation speed}
\vspace*{-1em}
\label{fig:seeds}
\end{figure}
We analyzed the error-metric evaluation for all resulting approximate implementations of 16-bit adders using considered BDD-based algorithms. Figure \ref{fig:parts} illustrates the execution time and the total number of created BDD nodes for three main parts of the error evaluation procedure. The part named \textit{loading} represents transforming the netlist of functions $f$ and $f'$ into BDDs. The second part, called \textit{subtracting}, consists of the ripple-carry subtractor creation. The last part, called \textit{calculating}, includes building BDDs that represent the calculation of the absolute value and the final evaluation. We can see that the first part is negligible in contrast to the subtracting and calculating parts. The \textit{baseline} implementation has an enormous proportion of the \textit{calculating} part, but the number of BDD nodes is relatively smaller. However, the overall evaluation time roughly correlates with the number of BDD nodes.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{fig/parts.pdf}
\vspace*{-2em}
\caption{Time spent and number of created nodes in three steps of the error calculation algorithms. Error probability $e_{ep}$~\cite{soeken:bdd} is included for complexity illustration.}
\vspace*{-1em}
\label{fig:parts}
\end{figure}
\subsection{Synthesis and comparison to SoA}
The resulting approximate adders are synthesized to obtain hardware parameters (we focus on energy). Both error metrics are evaluated. Figure \ref{fig:synthwce} shows trade-offs between $e_{wce}$ and the energy.
In addition to the area estimation in the fitness function (eq. \ref{eq:fitness}), we also used CGP with the area-delay product (ADP) in the fitness function. However, this setup did not improve the results in terms of energy.
The resulting approximate adders outperform relevant approximate adders published in the EvoApproxLib library~\cite{mrazek:date16lib}). Moreover, we designed larger approximate adders, including the signed and unsigned variants that are not included in the library. These circuits were introduced to the new version of EvoApproxLib.
\begin{figure}[ht]
\centering\vspace*{-1em}
\includegraphics[width=\columnwidth]{fig/synth_wce.pdf}
\vspace*{-2em}
\caption{Comparison of synthesized circuits designed with CGP with two fitness functions using proposed error calculations.}
\vspace*{-1em}
\label{fig:synthwce}
\end{figure}
\section{Conclusion}
We proposed new BDD-based algorithms for the fast error evaluation of approximate adders and subtractors. Compared with the baseline implementation, these algorithms reduced the evaluation time 3x times on average. Moreover, they do not need any assumption about $f'$. These algorithms enabled us to accelerate the approximate circuit design process. Our future work will be devoted to adapting the principle for different arithmetic error metrics and the application of the proposed circuits.
\vspace{1mm}
{\textit{Acknowledgement} This work has been supported the Czech Science Foundation grant (GJ20-02328Y).}
\bibliographystyle{./bibliography/IEEEtran}
|
\section{Introduction}
The macroscopic spin and charge state of solids can be manipulated by light pulses on few femtosecond \cite{Bigot2009, Siegrist2019} to picosecond time scales \cite{Kirilyuk2010,Carva2017, Kimel2020, Jeppson2021}. A wealth of fundamental new phenomena have been uncovered in this "femtoscape", including all-optical magnetization reversal \cite{Stanciu2007,Radu2011} and laser-driven spin transport \cite{Battiato2010,Eschenlohr2013,dewhurst2018}, effects that carry profound implications for technological application. Progress in the fields of femtomagnetism, ultrafast valleytronics\cite{Jin2018} and ultrafast spintronics, ultimately rely on the ability of experiments to decode the complex and non-equilibrium population of spin and charge on ultrafast time scales. Particularly important for building a microscopic understanding of such processes is the ability to obtain, from spectroscopy experiments, a state and element-selective view of the spin and charge dynamics, i.e., at what energies, and from which atoms, is spin optically excited to and from during the time evolving magnetic state.
Such charge and spin excitations are determined by the pump pulse parameters and the nature of the magnetic material, as encoded in the spin-resolved density of states (DOS). In a magnetic material possessing nearly full majority bands the dominant available optical excitations are from minority states below to minority states above the Fermi level ($E_F$), a process that leaves the macroscopic magnetic state unchanged. Concurrently occurring spin-flip transitions will, however, reoccupy these minority states leading to the observed ultrafast demagnetization\cite{krieger2015,ZH00}. In alloys and hetero-structures differences between the local DOS of distinct elemental species may themselves result in the laser induced transfer of spin from one atomic site to the other, adding a further layer of complexity to the spin dynamics\cite{dewhurst2018,Siegrist2019,hofherr2020, Willems2020a, Tengdin2020,Golias2021}. Assessing the role these basic principles of optical excitation and spin flips play in the complex landscape of ultrafast demagnetization is of crucial importance for understanding, and possibly controlling, the remarkable physics of ultrafast magnetization dynamics.
To unravel the complex pathways of non-equilibrium charge within a site-, energy- and spin dependent landscape during ultrafast demagnetization, requires techniques capable of resolving these degrees of freedom. Some of this information can be extracted from spin and time dependent angle-resolved photoemission spectroscopy \cite{Rhie2003,Gort2018, Eich2017a, Frietsch2020}, which essentially measures the transient electronic spectrum. However, such experiments are not in general element specific and, in addition, are strongest when probing bands and momentum directions on surface/interface and around the Fermi energy, not when providing a global picture (i.e. fully energy resolved spectrum of the entire bulk of the solid) of spin and charge excitation. Magnetic circular dichroism (MCD) experiments, on the other hand, provide an element specific tool that readily offers access to transient magnetization \cite{Stamm2007a, Boeglin2010} separated into the contribution from spin and orbital moments \cite{Thole1992,Carra1993,Stohr2006,Stamm2010, Bergeard2014, Hennecke2019a}.
Such experiments typically employ the XMCD contrast at fixed photon energy to probe magnetization dynamics, with recently a few experiments investigating spectral changes in absorption as a possible route to probe energy dependent magnetization dynamics \cite{Higley2019,Willems2020a,Hennes2020a,Rosner2020,LeGuyader2022}. While intriguing and as yet unexplained effects have been observed, a direct link to transient energy and spin-dependent electron occupations has not been established.
In this joint theory and experimental work, we demonstrate that transient helicity-dependent X-ray absorption spectroscopy (HXAS), the "raw data" of XMCD experiments that is typically discarded, encodes detailed information of the energy- spin- and species-resolved electronic transitions that occur both during the optical excitation as well as in the subsequent relaxation processes. In order to extract this information we provide a method based on well founded assumptions of the underlying optical transitions, requiring for its facilitation the {\it ab-initio} determination of 4 parameters per material system.
Employing this tool and with an example of Co, we unveil the energy and spin resolved dynamics of laser pumped demagnetization, revealing the pathways in energy-spin space in which the ultrafast loss of moment occurs. We find a richly structured energy resolved spin dynamics with, remarkably, the spin moment below the Fermi level first exhibiting an increase after the laser pulse (due to excitation of minority charge), before falling as spin-flip process in highly localized low energy states dominate. In contrast, the magnetic moment above Fermi energy shows a continuous increase in magnitude (but with the moment pointing in the opposite direction) from the outset of the laser pulse.
\section{HXAS, XMCD and going beyond}
\begin{figure}[ht]
\includegraphics[width=\columnwidth, clip]{Fig1.png}
\caption{
(a) Schematic illustration of helicity dependent X-ray absorption (HXAS) for the L$_{3}$ absorption resonance in Cobalt in which electrons are promoted from the $2p_{3/2}$ core level into the spin polarized $3d$ unoccupied valence levels. Light of positive helicity yields the yellow ($H^+$) XAS, negative helicity the blue ($X^-$) XAS. (b) Helicity dependent transient X-ray absorption. Laser induced changes to the spin dependent state occupation $\Delta$DOS, shown for majority ($\Delta u$) and minority ($\Delta d$) channels, are reflected in changes to the helicity dependent X-ray absorption $\Delta$XAS, with these two connected through the spin dependent excitation probabilities $A$, $B$, $A’$, and $B’$. Integration over the coloured energy regions yields the time, energy, and spin resolved changes in state occupation and XAS, shown on the right hand side. (c) Illustration of the spin-orbit spin flip process in ferromagnetic Cobalt after optical excitation. The stronger optical laser excitation of the minority electrons creates localized holes below $E_F$, enhancing subsequent spin-orbit coupling induced transitions, in turn leading to an increased spin flip rate and demagnetization at early times.
}\label{fig:scheme}
\end{figure}
When circular polarized X-ray or XUV photons are transmitted through a magnetic material the absorption strength of excitations from spin-orbit split core levels to the valence bands depends on the sign of the helicity, i.e. the material shows circular dichroism. A specific example of the absorption of circular polarized X-rays is the $L_{3}$ resonance (excitations from the $2p_{3/2}$ core level to the 3$d$ valence band) in ferromagnetic Cobalt, illustrated in Fig. \ref{fig:scheme}a. This response function represents a measure of magnetic moment of the material: light of positive helicity primarily couples to spin-up (majority spin) electrons (leading to a response function we call $H^+(E)$), while light of negative helicity couples predominantly to spin-down (minority spin) electrons (leading to a response function we call $H^-(E)$)\cite{Stohr2006}. The absorption probability for a particular spin channel depends additionally on the available unoccupied valence states of the respective spin. This allows for "counting" of available valence states of specific spin, which is a direct measure of the spin polarization in the case of isolated atoms. To obtain this spin polarization from $H^+(E)$ and $H^-(E)$ for a solid one uses the normalized difference between both response functions, which yields the X-ray circular dichroism (XMCD) spectrum, together with the XMCD sum-rules (i.e. integration of this spectrum over the $L_{3}$ and the $L_{2}$ resonances in case of $3d$ elements like Co)\cite{Thole1992, Carra1993,chen1995,Stohr2006}.
Element specificity of the XMCD response comes from the fact that the energies of these core levels strongly vary with the atomic species and can therefore be energetically separated.
In the transient case occupations of bands above and below the Fermi energy evolve as charge is excited by the laser pump pulse (compare Fig.~\ref{fig:scheme}b). Subsequent dynamical processes such as spin-orbit scattering, reduce the local spin moments and so further change these occupations. The XMCD spectra in combination with the sum rules\cite{chen1995} can then be employed to obtain the resulting dynamics of the magnetic moment. This procedure has been enormously successful, and has been used to map early time spin dynamics in a number of magnets including species resolved magnetic order in multi-component magnets\cite{Stamm2010,Radu2011,Bergeard2014, Hennecke2019a}. However, what is not readily accessible, yet is of great importance, is information concerning energy and time dependent changes of a specific spin channel (and not just the integrated difference as represented by the magnetic moment).
In obtaining magnetization dynamics from the XMCD spectra, the individual response functions to helical light, $H^+$ and $H^-$, are discarded. As we will now show, hidden within $H^+$ and $H^-$ lies detailed information on spin and energy resolved changes in the electronic structure of dynamical magnetic systems.
The helicity dependent absorption can be written as:
\begin{eqnarray}
H^{\pm}(E,t) &=& H_{\uparrow}^{\pm}(E,t) + H_{\downarrow}^{\pm}(E,t) \nonumber\\
&=&\Im \sum_{i,f} \Bigg[ p^{\pm}_{i,f,\uparrow}(t)\frac{(n_{i,\uparrow}-n_{f,\uparrow})(t)}{\Delta\epsilon_{if,\uparrow}(t)-E+i\eta} \nonumber \\
&+& p^{\pm}_{i,f,\downarrow}(t) \frac{(n_{i,\downarrow}-n_{f,\downarrow})(t)}{\Delta\epsilon_{if,\downarrow}(t)-E+i\eta}\Bigg]
\label{hpmc}
\end{eqnarray}
where in the first line we have decomposed the helicity dependent absorption into the spin projected responses $H^{\pm}_\sigma(E,t)$, and in the second line give the corresponding linear response formulae. In these latter equations $i$ and $f$ are joint indices of state and {\bf k}-point and represent the initial and the final state respectively, $\uparrow$ and $\downarrow$ refer to the majority and minority spins, and $p^{\pm}_{if}$ are the transition probabilities for transitions from state $i$ to $f$ with left/right circularly polarized light. $\epsilon_i$ is the energy and $n_i$ the occupation of the $i$'th state.
In the simplifying limit of a two state model, Eq.~\ref{hpmc} can be written as
\begin{eqnarray}
a(t)u(E,t)+b(t)d(E,t)&=H^+(E,t) \label{hxas0} \\
a'(t)u(E,t)+b'(t)d(E,t)&=H^-(E,t), \nonumber
\end{eqnarray}
i.e. in a way that separates the transition element information from that of state occupation and DOS (with $a(t) = p^+_{i f\uparrow}$, $b(t) = p^+_{i f\downarrow}$, $a'(t) = p^-_{i f\uparrow}$, $b'(t) = p^-_{i f\downarrow}$, and $u(E,t) = \Im (n_{i_\uparrow}-n_{f\uparrow})/(\Delta\epsilon_{if,\uparrow}-E+i\eta))$, and $d(E,t) = \Im (n_{i_\downarrow}-n_{f\downarrow})/(\Delta\epsilon_{if,\downarrow}-E+i\eta))$. The latter two quantities, $u(E,d)$ and $d(E,d)$, represent the DOS and occupation from which transient spin resolved charge can be extracted.
For a realistic solid such a decomposition cannot be rigorously made. However, as we now show \emph{ab-initio} theory allows us to obtain an approximate scheme providing a corresponding decomposition for solids. To this end we consider the spin projected helicity dependent XAS spectra ($H^\pm_\sigma(E,t)$), i.e. the XAS spectra with majority or minority spin transitions blocked, see Fig.~\ref{fig:ab}. We now make the assumption (to be numerically verified later) that an approximation to the average value over a band of the ratios of transition probabilities $p^\pm_{i f\sigma}/\left(\sum_\sigma p^\pm_{i f\sigma}\right)$ can be obtained by integration over the L3 edge the ratio $H^{\pm}_\sigma/H^\pm$. With these ratios in hand an approximation to Eq.~\ref{hxas0} can be written as
\begin{eqnarray}
\label{hxas}
A u(E,t)+B d(E,t)&=\frac{H^+(E,t)}{\int\!dE\, H^{+}(E,0)} \\
A' u(E,t)+B' d(E,t)&=\frac{H^-(E,t)}{\int\!dE\, H^{-}(E,0)}, \nonumber
\end{eqnarray}
where we have employed a second assumption in neglecting any time dependence of these ratios (again, to be validated later by numerical calculation). They thus may be obtained from the ground state as:
\begin{eqnarray}
A & = & \int\! dE\, H^+_\uparrow(E,0)/\int\! dE\, H^+(E,0), \\
B & = & \int\! dE\, H^+_\downarrow(E,0)/\int\! dE\, H^+(E,0), \nonumber \\
A' & = & \int\! dE\, H^-_\uparrow(E,0)/\int\! dE\, H^-(E,0), \nonumber \\
B' & = & \int\! dE\, H^+_\downarrow(E,0)/\int\! dE\, H^-(E,0) \nonumber.
\end{eqnarray}
where the energy integration is again over the L3 edge.
In this way we can identify the parts of the response function that come primarily from the bands and their occupation (the $u(E,t)$ and $d(E,t)$) separately from the parts that comes from the transition probabilities ($A$, $B$, $A'$ and $B'$).
From the transient-HXAS response functions of laser pumped systems Eq.~\ref{hxas} can then be solved transiently to obtain $u(E,t)$ and $d(E,t)$, a procedure schematically illustrated in Fig.\ref{fig:scheme}b. This will therefore provide us with the required information: changes in majority (u) and minority (d) charge as a function of energy and time.
In Fig.~\ref{fig:ab} we shown these response functions and their energy integration over the L3 edge for Cobalt. The calculated values of these parameters of Co found from these response functions are $A=0.846$, $B=0.154$, $A'=0.646$ and $B'=0.355$. For the corresponding values for Fe and Ni see \onlinecite{abc}. In the following sections we now describe the theoretical and experimental method used to determine the dynamics of magnetization from the HXAS spectra.
\begin{figure}[ht]
\includegraphics[width=\columnwidth, clip]{ab1.pdf}
\caption{Helicity dependent x-ray absorption spectra from right circularly polarized light for Cobalt for the L edge. The response function is calculated in three different ways: by blocking all majority 2$p$ states ($H^+_{\rm min}$), by blocking all minority 2$p$-states ($H^+_{\rm maj}$) and by including all transitions from core levels ($H^+_{\rm total}$, the full response), with the corresponding dashed lines the energy integration of these response functions. The parameters for extracting spin response functions from the response functions of circularly polarized light are given by the ratios of the integrated quantities $\int\! dE\, H^\pm_\sigma(E)/\int\! dE\, H^\pm(E)$ over the L3 edge.
}\label{fig:ab}
\end{figure}
\section{Theoretical methodology}
To calculate transient HXAS for laser pumped Co we have used the fully {\it ab-initio} state-of-the-art time dependent density functional theory (TDDFT)\cite{RG1984}, that rigorously maps the computationally intractable problem of interacting electrons to the Kohn-Sham (KS) system of non-interacting fermions in a fictitious potential, which can be solved by modern computing clusters. The time-dependent KS equation is:
\begin{align}
i &\frac{\partial \psi_{j{\bf k}}({\bf r},t)}{\partial t} =
\Bigg[
\frac{1}{2}\left(-i{\nabla} +\frac{1}{c}{\bf A}_{\rm ext}(t)\right)^2 +v_{s}({\bf r},t) \nonumber \\
& + \frac{1}{2c} {\sigma}\cdot{\bf B}_{s}({\bf r},t) + \frac{1}{4c^2} {\sigma}\cdot ({\nabla}v_{s}({\bf r},t) \times -i{\nabla})\Bigg]
\psi_{j{\bf k}}({\bf r},t)
\label{TDKS}
\end{align}
where ${\bf A}_{\rm ext}(t)$ is a vector potential representing the applied laser field. It is assumed that the wavelength of the applied laser is much greater than the size of a unit cell and the dipole approximation can be used, i.e. the spatial dependence of the vector potential is disregarded. The KS potential $v_{s}({\bf r},t) = v_{\rm ext}({\bf r},t)+v_{\rm H}({\bf r},t)+v_{\rm xc}({\bf r},t)$ is decomposed into the external potential $v_{\rm ext}$, the classical electrostatic Hartree potential $v_{\rm H}$ and the exchange-correlation (XC) potential $v_{\rm xc}$. Similarly, the KS magnetic field is written as ${\bf B}_{s}({\bf r},t)={\bf B}_{\rm ext}(t)+{\bf B}_{\rm xc}({\bf r},t)$ where ${\bf B}_{\rm ext}(t)$ is an external magnetic field and ${\bf B}_{\rm xc}({\bf r},t)$ is the exchange-correlation (XC) magnetic field. ${\sigma}$ are the Pauli matrices. The final term of Eq.~\eqref{TDKS} is the spin-orbit coupling term. In the fully non-collinear spin-dependent version of this theory\cite{krieger2015,dewhurst2016} the orbitals, $\psi$ are two component Pauli spinors and using these orbitals magnetisation density can be calculated as:
\begin{eqnarray}\label{st}
{\bf m}({\bf r},t)=\sum_j \psi^*_j({\bf r},t){\m \sigma}\psi_j({\bf r},t),
\end{eqnarray}
the integral of this vector field over space gives the magnetic moment as a function of time (M(t)).
In order to calculate the response function at various times, linear response formalism of the TD-DFT is the used\cite{RG1984,my-book,sharma2011}. This method of calculating the transient response function has been studied extensively\cite{clemens2020,dewhurst2020,sharma22}, and demonstrated to be in very good agreement with experiment. For the response function calculations a broadening of 0.9~eV was used. This choice of smearing is based on ground-state $GW$ calculations, being the average of the width of p$_{1/2}$ and p$_{3/2}$ states. Ground-state $GW$ calculations were also used to determine the position of the 2$p$-states and the KS states were then scissors corrected. These scissors corrected states are used for performing response function calculations. For performing these $GW$ calculations a finite temperature (a temperature of 500~K was used), all electron, spin-polarized $GW$ method\cite{gw} is used, where the spectral function on the real axis is constructed using a Pade approximation. Spin-orbit coupling was included in the $GW$ calculations and a Matsubara cutoff of 12~Ha was used.
In the present work we have used adiabatic local density approximation for XC potential and all calculations are performed using the highly accurate full potential linearized augmented-plane-wave method\cite{singh}, as implemented in the ELK\cite{elk,dewhurst2016} code. A face centred cubic unit cell with lattice parameter of $3.53\si{\angstrom}$ was used for Co. The Brillouin zone was sampled with a $20\times 20\times 20$ k-point mesh. For time propagation the algorithm detailed in Ref.~\onlinecite{dewhurst2016} was used with a time-step of $2.42$ atto-seconds. The final magnetization value converges with the above mentioned computational parameters to 1.61~$\mu_{\rm B}$. The theoretical pump pulse used in the present work has central frequency of 1.55~eV, duration of 24~fs and incident fluence of 6.7~mJ/cm$^2$. In comparison to the experimental pulse this is of shorter duration, a situation motivated by several factors: (a) non-inclusion of nuclear dynamics prevents us from describing physics at longer time scales and (b) while the method proposed above is state-of-the-art, there is a price to pay in computational burden that, with present computer power, precludes the calculation of longer pulses.
\section{Experimental method}
The ultrafast HXAS experiments were conducted at the BESSY~II Femtoslicing facility (beamline UE56/1-ZPM and DynaMaX end-station) at Helmholtz-Zentrum Berlin \cite{Holldack2014}. Femtosecond X-ray pulses of 100~fs temporal width and elliptical polarization are generated via femto-slicing from electron bunches stored in the BESSY~II electron storage ring by interaction with intense infrared laser pulses. Laser pulses from a second laser amplifier system (pump-laser) intrinsically synchronized with the fs-X-ray pulses are used as pump-pulses for sample excitation (wavelength 800~nm and pulse width 50~fs). The effective temporal resolution of this experimental setup is 120~fs. The experiment is operated at a repetition rate of 6~kHz, where only every second fs-X-ray pulse is synchronized to a pump-pulse (i.e. pump-laser running at 3~kHz). This allows for alternatingly probing the excited sample and the sample ground state on a shot-to-shot basis, providing excellent normalization opportunities. X-ray and pump-laser beams enclose an angle of about 1$^{\circ}$ when incident on the sample.
The sample was a Pt(3~nm)/Co(15~nm)/Al(300~nm) layer stack deposited on a 200~nm thin Si$_{3}$N$_{4}$ membrane by magnetron sputtering, which allowed measurement of XAS spectra across the Co L$_{3}$ resonance in transmission geometry. The X-ray photons hit the sample at 43$^{\circ}$ incidence angle to surface normal. The transmitted intensity is detected by an avalanche photo diode screened from the laser light by an Al membrane \cite{Holldack2014}. The relative orientation between X-ray helicity and magnetization was swapped by inversion of an external magnetic field (0.1~T field amplitude) pointing parallel to the X-ray beam. The X-ray absorption spectra were recorded quasi-simultaneously by field inversion for every single photon energy on a ten second basis. The energy resolution of the UE56/1-ZPM beamline achieved at the Co L$_{3}$ absorption edge is approximately 1.6~eV.
To determine time zero of the pump-probe experiment for the recorded transient X-ray spectra with highest possible precision, a MoSi multi-layer reference sample was used\cite{Schick2016,Jana2021}. To correct for slow drifts of time-zero we employed the following measurement procedure: Time zero was determined on the MoSi multi-layer by repeated delay scans with a net acquisition time of 1 hour. The data statistics of this measurement allows determination of time zero in an interval smaller than $\pm$15~fs. Subsequently, X-ray spectra for a set of delays in a $\pm$150~fs interval around nominal time zero from the Co sample were recorded during a 2.5~hours acquisition time. These two types of measurements were alternatingly repeated. The effective delay for any recorded X-ray spectrum was finally recalculated by linear interpolation between the two time-zero values from the measurements on the MoSi multi-layer before and after the XAS acquisitions. Such obtained distribution of effective delays and associated X-ray spectra were finally binned in effective delay windows of 50~fs width ranging from -150 to 200~fs delay, averaged, and evaluated.
\section{Results}
\begin{figure}[ht]
\includegraphics[width=0.9\columnwidth, clip]{Fig3.png}
\caption{Difference between the helicity dependent x-ray absorption spectra (HXAS) before pumping and after laser pumping showing (a) theoretical data and (d) experimental data. The theoretically calculated transient change in the (b) majority and (c) minority spins is obtained by using Eqs.~\ref{hxas} and integrating within the various energy windows, as indicated by different colours in panels shown in (a). For the experiment the integration is performed in the windows shown in (d), and the correspondingly obtained transient majority and minority spins shown in panels (e) and (f) respectively. The pump laser pulse of central wave length 800~nm is shown in grey for theory in panels (b) and (c), while the Gaussian envelope of the experimental pulse is shown in (e) and (f) in grey. Note that the theoretical calculations were performed using a pump pulse of 24~fs duration while in the experiment the laser pulse full width at half maximum is approximately 50~fs. Despite this one can see excellent qualitative agreement between theory and experiment.
In particular the turning point of the minority charge, that represents the point at which spin-flip induced transitions dominate over spin preserving optical transitions, is seen for both theory and experiment to occur close to half cycle of the pump pulse.
}\label{fig:hxas}
\end{figure}
Employing the method proposed above we now use Co as an example to examine the dynamics of spin and charge upon laser pumping. A pump laser pulse drives demagnetization of Co, a process reflected in the HXAS spectra. These changes are shown in Fig.~\ref{fig:hxas}(a) and \ref{fig:hxas}(d) where we show the difference between the unpumped and pumped HXAS spectra obtained by theoretical calculations (20~fs after laser pumping) and from experiment (50~fs after laser pumping)
respectively. Negative values indicate an increase and positive numbers a decrease in the HXAS signal after laser induced changes in the state occupation (note that a decrease in absorption is equivalent to an increase in state-occupation). As expected, both experiments and theory show a decrease in the state-occupation below and an increase in the state-occupation above Fermi energy (which is set to 0~eV in Fig.~\ref{fig:hxas}(a) and (d)), representing the excitation of charge from below to above the Fermi energy due to the laser pump. The amount of change varies with x-ray helicity, signifying spin-dependent pump effects.
In order to examine changes of majority and minority charge as a function of time and energy, we now solve Eqs.~\ref{hxas} to obtain $u$ and $d$ and integrate these changes of the spin dependent occupation in the energy windows indicated by the shaded rectangles, shown in panels (a) and (d) of Fig.~\ref{fig:hxas}.
These windows (width and position) are chosen to capture the dynamics of charge loss below the Fermi energy, the dynamics of charge gain above the Fermi energy, and the interesting region around the Fermi energy within which more complex behavior can take place.
Remarkably, we see that the integrated change in $u$ and $d$ (i.e state occupation) obtained from theory very closely follows changes in the corresponding experimental data(Fig.~\ref{fig:hxas} b, c, e, f), analyzed using exactly the same parameters $A$, $B$, $A'$, $B'$ as derived from and used in analyzing the theoretical data.
Below (green) the Fermi energy both majority and minority state occupations are lost and gained respectively, while for the window enclosing the Fermi energy (yellow) majority is lost and minority gained.
This indicates not just optical excitation of charge but also a large number of spin flips (also see Fig.~\ref{fig:dos} for further details) in this energy range, leading to an increase in the minority occupation and decrease in the majority spins. This leads to demagnetization of Co.
In the window above the Fermi energy (blue) a strong gain in minority charge carriers is seen both in theory and experiment. In our early time theoretical calculations a much smaller gain in majority is seen, reflecting the nearly full occupation of majority states in the high energy window above $E_F$, while for the longer times used in experiment almost no change in majority charge above $E_F$ is seen.
We note that both in experiment and theory we find charge excitations outside the energy window of $\pm$1.55~eV around the Fermi energy indicating a contribution from non-linear two-photon excitation processes. To prove this in experiment would entail a full probe of fluence and energy dependent spectra.
An interesting feature offering microscopic insight into the spin-flip process is the turning point in the minority charge of the low energy window (green). This turning point precisely demarks the time at which spin-flip processes (which will increase minority charge) begin to dominate over optical excitations (that reduce minority charge) in the low energy window. These results indicate that the spin-flip process play a very important role in the energy region of \emph{low-energy} states, as is indicated by the Hamiltonian which shows that the SOC term is strongest for such low energy localized states as the $\nabla v_s$ term is largest for these states\cite{krieger2015}.
Despite two different pump pulse lengths used in theory and experiment we see that the qualitative behavior of the spin dependent charge corresponds overall very well between theory and experiment in each energy window. This highlights the universality of the underlying physics of laser pumped demagnetizaton, as well as the robustness of the procedure used to extract spin- and energy- resolved excited charge proposed in the present work.
We note that the energy corresponding to the Fermi level in the experimental data in Fig.~\ref{fig:hxas} is not given in a natural way. We define zero energy in the latter case in analogy to theory at the zero crossing point of the transient $u$ spectrum (see Fig.~\ref{fig:hxas}). The energy resolution of this procedure may be inaccurate by about 0.2~eV; this, however, does not effect any of our conclusions given that the width of our energy window is much broader than this.
\begin{figure}[ht]
\includegraphics[width=\columnwidth, clip]{Co-dos.pdf}
\caption{(a) Change in transient occupied spin projected DOS. Results are shown at various times after laser pumping. Pump pulse induced (b) change in majority and minority charge integrated below and above the ground-state Fermi energy and (c) change in total spin magnetic moment and spin moment integrated below and above the Fermi energy. The pump pulse has central wave length of 800~nm, full width half maximum of 24~fs and incidence fluence of 6.7mJ/cm~$^2$.
}\label{fig:dos}
\end{figure}
The HXAS thus encodes a great deal of information of the underlying excitation processes in ultrafast spin dynamics, well beyond that which can be obtained directly from MCD spectra, which provides only the dynamics of the local moments. While attempts have been made in previous works to infer signatures of such excitation processes from MCD/MLD\cite{LeGuyader2022,clemens2020} spectra, it is only by systematically disentangling the spin degrees of freedom in HXAS, as we do here, that this information can be obtained. To do this the only extra ingredient required are the parameters $A$, $B$, $A'$ and $B'$ in Eqs.~\ref{hxas}.
As these can be obtained \emph{ab-initio} from the ground state, then this procedure can be preformed for all elemental and multi-component magnets of interest (values of these parameters for Fe, Co and Ni are given in this work).
Having experimental access to energy and spin resolved information via HXAS offers new opportunities to probe in-depth the physics of the excitation process, in particular energy and time resolved changes in magnetization. To explore this physics further in Fig.~\ref{fig:dos}(a) we show changes in the state-density for majority and minority electrons as a function of time. One observes a broad loss of majority states from below the Fermi energy, and a corresponding large gain over and energy range of $\sim1.5$~eV above the Fermi energy. Closer examination unveils a subtle interplay of spin orbit and optically induced transitions, in particular the large peak in the minority density of states difference observed 10~fs but not at 25~fs due to filling of these states by spin-orbit induced transitions from majority to minority. It is this transient peak that is responsible for the turning point in the integrated transient spin resolved DOS occupations seen in both theory and experiment above, and revealed also in the integrated minority charge below Fermi energy, see Fig.~\ref{fig:dos}b. These energy resolved spin channels below and above the Fermi energy also show clear evidence of the extent of spin flips as the increase in minority charge above the Fermi energy is significantly greater than the decrease in minority charge from below the Fermi energy.
In a similar way we can also energy resolve the dynamics of magnetic moments,
At early times minority electrons undergo optical excitation from below the Fermi energy, and the magnetic moment below the Fermi energy (orange curve in Fig.~\ref{fig:dos}c) shows a corresponding \emph{increase}. At $\sim20$~fs spin-flip processes begin to dominate, driving a reduction in the majority carriers and hence a reduction in the moment below the Fermi level. At the same time, this excitation of minority charge from below to above the Fermi level leads to a negative change in the moment to develop above the Fermi level (pink). However, the magnitude of the moment above the Fermi energy also starts to saturate as spin-flips dominate the physics of demagnetization. These results again point towards the physics seen in Fig.~\ref{fig:scheme}c and \ref{fig:hxas}; the spin flip process are not just around the Fermi level but are strongly energy dependent and show signatures well below the Fermi energy.
\begin{figure}[ht]
\includegraphics[width=\columnwidth, clip]{Fig5.png}
\caption{Internal theoretical consistency of HXAS and DOS derived spin, energy, and time resolved state occupation. (a) HXAS change 20~fs after laser pumping and (d) the difference in the occupied density of states 20~fs after pumping. Charge in (b) majority and (c) minority spins obtained by using Eqs.~\ref{hxas} and integration in various energy windows indicated by the different colours in panel (a). Change in (e) majority and (f) minority spins obtained from integrating the difference in DOS occupations in the corresponding energy windows shown in panel (d). Note that the energy windows for integration are exactly the same width and at the same position with respect to the Fermi energy in panels (a) and (d). As can be seen, both routes to deriving transient state occupations yield very similar results, validating the method proposed for disentangling spin, energy, and time resolved state occupation from the HXAS spectra. The {\bf A}-field of the pump pulse has an incidence fluence of 6.7mJ/cm~$^2$.
}\label{fig:th}
\end{figure}
\section{Theoretical validation of HXAS disentanglement}
One of the major approximations in solving Eqs.~\ref{hxas} was the assumption of very weak time dependence of the transition probabilities. This not only reduces the computational task, most importantly it allows calculation and tabulation of $A, B, A'$ and $B'$ from static ground-state calculations for the analysis of future experiments. While we have justified this assumption {\it post-hoc} by agreement between theory and experiment, within the theoretical framework itself this approximation can be justified. In order to do this in Fig.~\ref{fig:th} we present the changes in HXAS and DOS occupations 20~fs after laser pumping. The DOS is broadened by a Gaussian (FWHM $0.82$~eV) in order to have the same spectral width as the HXAS response function. Evidently, the two sets of data show the same trend in that occupation of both majority and minority spins decreases below and increases above the Fermi energy, while at the Fermi level the two curves cross zero. Closer inspection of the helicity dependent response function, Fig.~\ref{fig:th}(a), and the transient DOS occupation, Fig.~\ref{fig:th}(d), reveals slight differences arising from the fact that the response functions contain information both of changes in occupation as well as the respective transition probabilities.
Integrated changes in the occupation in three differently positioned energy windows are shown in panels \ref{fig:th}(b) and (e) for majority and (c) and (f) for minority spins. In this case, we see that the integrated charge from HXAS very closely follows changes in the corresponding state density. In both the positive (blue) and negative (green) energy windows both majority and minority states are lost and gained respectively, while for the window enclosing the Fermi energy (yellow) majority is lost and minority gained. This close agreement between the changes in spin projected occupations obtained from DOS and HXAS calculations provides a demonstration within the theoretical framework that the extraction from the HXAS of pure spin response functions using coefficients tabulated from the ground state is indeed a good approximation.
\section{Discussion}
Strong light-matter interaction in magnetic systems creates a highly non-equilibrium state of matter exhibiting remarkable phenomena such as all-optical switching and OISTR. Exploring this diversity of phenomena demands experimental probes able to precisely track the microscopic processes underlying ultrafast spin dynamics. In the present work we have proposed and verified a simple extension of the well established XMCD technique, allowing a transient spin- energy- and species-resolved probe of ultrafast demagnetization.
This extension is based upon disentangling the spin degrees of freedom that are, generally, woven in a complex way into the response of a magnetic system to circularly polarized light. We show that, by employing parameters obtained \emph{ab-initio}, the experimental HXAS spectra may be inverted to give spin specific response functions that, in turn, provide a window into the spin-energy landscape in which ultrafast demagnetization occurs.
Using the example of elemental cobalt we resolve the magnetization dynamics into energy regions that vividly illustrate the wide energy range on which the spin-orbit and optical transitions that drive demagnetization occur. Specific quantitative features, such as turning points in spin and energy resolved charge, further reveal critical points at which the balance between spin preserving optical transitions and spin-orbit induced spin-flips changes.
The \emph{ab-initio} procedure employed to disentangle the HXAS spectra is shown to be valid within the theoretically obtained HXAS spectra and state density, with the qualitative agreement between theory and experiment (which employ lasers pulses of significantly longer duration than those that can be treated theoretically) further bolstering the correctness of the spin projection of the HXAS spectra. As the parameters encoding spin specific transitions required to preform this transformation are, to an excellent approximation, time independent, they can be obtained from the ground state and so this procedure can, in principle, be applied to multi-component magnetic materials of almost any complexity. The extension of HXAS that we propose is therefore of great applicability.
Promising future experimental applications of this "extended HXAS" method include probing spin- and energy-resolved dynamics in all-optical spin switching magnets. Direct access to the transient spin- and energy resolved charge is expected to provide invaluable insight into still controversial features of the dynamics of spin switching, e.g. which moments rotate, at what times, and the in-plane versus out-of-plane components of magnetization. Understanding the still unclear microscopic world underpinning such well established phenomena may provide the basis for enriched exploration and control of the frontiers of ultrafast spin dynamics.
\section{Acknowledgements}
Sharma, CvKS, SE, NP and CSL would like to thank for funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project-ID 328545488 – TRR 227 (projects A04, A02 and A03).
Shallcross would like to thank DFG for funding through SPP 1840 QUTIF Grant No. SH 498/3-1. The authors acknowledge the North-German Supercomputing Alliance (HLRN) for providing HPC resources that have contributed to the research results reported in this paper.
|
\section{Introduction}
In 2013, the BESIII and Belle collaborations explored the process $e^+e^- \to \pi^+\pi^-J/\psi$, and observed a structure $Z_c^\pm(3900)$ in the $\pi^\pm J/\psi$ mass spectrum \cite{BES3900,Belle3900}.
Also in 2013, the BESIII collaboration investigated the process $e^+e^- \to \pi D \bar{D}^*$, and observed a structure $Z_c^\pm(3885)$ in the $(D \bar{D}^*)^{\pm}$
mass spectrum \cite{BES-3885}. The $Z_c(3900/3885)$ have the spin-parity $J^P=1^+$ \cite{BES-3885,JP-BES-Zc3900}.
In 2020, the BESIII collaboration observed a structure $Z_{cs}^-(3985)$ in the $K^{+}$ recoil-mass spectrum in the processes of the $e^+e^-\to K^+ (D_s^- D^{*0} + D^{*-}_s D^0)$ \cite{BES3985}. The Breit-Wigner mass and width are $3985.2^{+2.1}_{-2.0}\pm1.7\,\rm{MeV}$ and $13.8^{+8.1}_{-5.2}\pm4.9\,\rm{MeV}$, respectively \cite{BES3985}.
In 2021, the LHCb collaboration observed two new exotic states $Z_{cs}^+(4000)$ and $Z_{cs}^+(4220)$ in the $J/\psi K^+$ mass spectrum in the process $B^+ \to J/\psi \phi K^+$ \cite{LHCb-Zcs4000}. The most significant state, $Z_{cs}^+(4000)$, has the Breit-Wigner mass and width $4003 \pm 6 {}^{+4}_{-14}\,\rm{MeV}$ and $131 \pm 15 \pm 26\,\rm{MeV}$, respectively, and the spin-parity $J^P =1^+$ \cite{LHCb-Zcs4000}.
The $Z_c(3900/3885)$ and $Z_{cs}(3985/4000)$ have analogous decay modes,
\begin{eqnarray}
Z^{\pm}_c(3900) &\to & J/\psi \pi^\pm \, , \nonumber \\
Z^{+}_{cs}(4000) &\to & J/\psi K^+ \, ,
\end{eqnarray}
\begin{eqnarray}
Z_c^{\pm}(3885)&\to & (D\bar{D}^*)^\pm \, , \nonumber \\
Z_{cs}^{-}(3985) &\to & D_s^- D^{*0}\, , \, D^{*-}_s D^0 \, ,
\end{eqnarray}
which lead to the possible outcome that they should have analogous quark structures. Although the Particle Data Group takes it for granted that the $Z_c(3900)$ and $Z_c(3885)$ are the same particle according to the analogous masses and widths \cite{PDG}, however, it is difficult to explain the ratio
\cite{BES-3885},
\begin{eqnarray}
R_{exp}&=&\frac{\Gamma(Z_c(3885)\to D\bar{D}^*)}{\Gamma(Z_c(3900)\to J/\psi \pi)} =6.2 \pm 1.1 \pm 2.7 \, ,
\end{eqnarray}
in a satisfactory way. On the other hand, the widths of the $Z^-_{cs}(3985)$ and $Z_{cs}^+(4000)$
are inconsistent with each other. The $Z_c(3900/3885)$ and $Z_{cs}(3985/4000)$ may be four distinguished particles indeed.
The exotic states $Z_{cs}(3985/4000)$ lie above the $D_s^- D^{*0}$ and $D^{*-}_s D^0$ thresholds $3975.2\,\rm{MeV}$ and $3977.0\,\rm{MeV}$, respectively,
if they are molecular states indeed, we should introduce the coupled-channel effects to account for the mass gaps and decay widths \cite{Zcs-molecule-1,Zcs-molecule-2,Zcs-molecule-3,Zcs-molecule-4,Zcs-molecule-5}, or just re-scattering effects \cite{Chen:2013wca,Zcs-Rescatt-1}. It is more natural to reproduce their masses in the diquark-antidiquark type tetraquark pictures saving potential models are chosen \cite{Ebert:2008kb, Ferretti:2020ewe}.
In the QCD sum rules, we usually choose the color antitriplet-triplet type and singlet-singlet type four-quark currents to interpolate the hidden-charm exotic states, and can reproduce the masses of the $Z_{cs}(3985/4000)$ both in the pictures of the tetraquark states \cite{Zcs-CFQiao,Zcs-tetra-quark,WZG-Zcs3985-tetra,Zcs3985-Azizi} and molecular states \cite{Zcs-CFQiao,Lee:2008uy,Dias:2013qga,Zcs-molecule-6,WZG-Zcs3985-mole}. There maybe exist two $Z_{cs}$ states, an antitriplet-triplet type tetraquark state and singlet-singlet type molecular state, or only one $Z_{cs}$ state, which has both the antitriplet-triplet type and singlet-singlet type components. We have to explore the two-body strong decays to diagnose their nature.
In Ref.\cite{WZG-ZJX-Zc-Decay}, we assign the $Z^\pm_c(3900)$ to be the diquark-antidiquark type tetraquark state with the quantum numbers $J^{PC}=1^{+-}$, calculate the hadronic coupling constants $G_{ZJ/\psi\pi}$, $G_{Z\eta_c\rho}$, $G_{ZD \bar{D}^{*}}$ with the three-point QCD sum rules based on the rigorous duality for the first time, then obtain the partial decay widths to diagnose the nature of the $Z_c^\pm(3900)$. Thereafter, the rigorous duality has been successfully applied to study the strong decays of the exotic states $X(4140)$, $X(4274)$, $X(4660)$, $Z_c(4600)$ and $P_c(4312)$ \cite{WZG-Y4660-Decay,WZG-X4140-decay,WZG-X4274-decay,WZG-Z4600-decay,WZG-Pc4312-decay-mole,WZG-Pc4312-decay-tetra}.
In this work, we extend our previous work on the masses of the hidden-charm tetraquark states and molecular states with the strangeness \cite{WZG-Zcs3985-tetra,WZG-Zcs3985-mole}, to explore the two-body strong decays of the $Z_{cs}(3985/4000)$ with the possible assignments $J^{PC}=1^{+-}$, so as to diagnose their nature.
The article is arranged as follows: we acquire the QCD sum rules for the hadronic coupling constants $G_{ZJ/\psi K}$, $G_{Z\eta_cK^*}$, $G_{ZD^* \bar{D}_s}$ in section 2; in Sect.3, we present the numerical results and discussions; and Sect.4 is reserved for our
conclusion.
\section{The hadronic coupling constants of the $Z_{cs}(3985/4000)$ }
We investigate the two-body strong decays $Z_{cs}^-(3985/4000)\to J/\psi K^-$, $\eta_cK^{*-}$, $D^{*0} \bar{D}_s^{-}$, $D^0 \bar{D}_s^{*-}$ using the three-point correlation functions $\Pi_{\mu\nu}^{1}(p,q)$, $\Pi_{\mu\nu}^{2}(p,q)$ and $\Pi_{\mu\nu}^{3}(p,q)$, respectively,
\begin{eqnarray}
\Pi_{\mu\nu}^{1}(p,q)&=&i^2\int d^4xd^4y \, e^{ipx}e^{iqy}\, \langle 0|T\left\{J_\mu^{J/\psi}(x)J_5^{K}(y)J^\dagger_{\nu}(0)\right\}|0\rangle\, , \nonumber \\
\Pi_{\mu\nu}^{2}(p,q)&=&i^2\int d^4xd^4y\, e^{ipx}e^{iqy}\, \langle 0|T\left\{J_5^{\eta_c}(x)J_\mu^{K^*}(y)J^\dagger_{\nu}(0)\right\}|0\rangle \, , \nonumber \\
\Pi_{\mu\nu}^{3}(p,q)&=&i^2\int d^4xd^4y \, e^{ipx}e^{iqy}\, \langle 0|T\left\{J_\mu^{D^*}(x)J_5^{D_s}(y)J^\dagger_{\nu}(0)\right\}|0\rangle \, ,
\end{eqnarray}
where the currents
\begin{eqnarray}
J_\mu^{J/\psi}(x)&=&\bar{c}(x)\gamma_\mu c(x) \, ,\nonumber \\
J_5^{K}(y)&=&\bar{u}(y)i\gamma_5 s(y) \, , \nonumber \\
J_5^{\eta_c}(x)&=&\bar{c}(x)i\gamma_5 c(x) \, ,\nonumber \\
J_\mu^{K^*}(y)&=&\bar{u}(y)\gamma_\mu s(y) \, , \nonumber \\
J_\mu^{D^*}(x)&=&\bar{u}(x)\gamma_\mu c(x) \, ,\nonumber \\
J_5^{D_s}(y)&=&\bar{c}(y)i\gamma_5 s(y) \, , \nonumber \\
J^\dagger_\nu(0)&=&J_\nu^{T\dagger}(0)\, , \, J_\nu^{M\dagger}(0)\, ,
\end{eqnarray}
\begin{eqnarray}
J_\nu^{T\dagger}(0)&=&\frac{\varepsilon^{ijk}\varepsilon^{imn}}{\sqrt{2}}\Big\{c^{Tn}(0)C\gamma_\nu u^m(0)\, \bar{c}^k(0)\gamma_5 C \bar{s}^{Tj}(0)-c^{Tn}(0)C\gamma_5 u^m(0)\, \bar{c}^k(0)\gamma_\nu C \bar{s}^{Tj}(0) \Big\} \, , \nonumber\\
J_\nu^{M\dagger}(0)&=& \frac{1}{\sqrt{2}}\Big\{\bar{s}(0)\gamma_\nu c(0)\, \bar{c}(0)i\gamma_5 u(0)+\bar{s}(0)i\gamma_5 c(0)\, \bar{c}(0) \gamma_\nu u(0)\Big\}\, ,
\end{eqnarray}
interpolate the mesons $J/\psi$, $K$, $\eta_c$, $K^*$, $D^*$, $D_s$, $Z^T_{cs}$ and $Z^M_{cs}$, respectively, the superscripts $T$ and $M$ denote the tetraquark-type and molecule-type four-quark currents/tetraquarks, respectively \cite{WZG-Zcs3985-tetra,WZG-Zcs3985-mole}.
We insert a complete set of intermediate hadronic states with the same quantum numbers as the currents into the three-point correlation functions, and isolate the ground state contributions,
\begin{eqnarray}
\Pi_{\mu\nu}^{T,1}
(p,q)&=& \frac{f_{K}M_{K}^2f_{J/\psi}M_{J/\psi}\lambda_{Z}^T G^T_{ZJ/\psi K}}{m_u+m_s} \frac{i}{(M_{Z}^2-p^{\prime2})(M_{J/\psi}^2-p^2)(M_{K}^2-q^2)} g_{\mu\nu}+\cdots \nonumber\\
&=& \Pi_T^1(p^{\prime2},p^2,q^2)\,g_{\mu\nu}+\cdots\, ,
\end{eqnarray}
\begin{eqnarray}
\Pi_{\mu\nu}^{M,1}(p,q)&=& \frac{f_{K}M_{K}^2f_{J/\psi}M_{J/\psi}\lambda_{Z}^M G^M_{ZJ/\psi K}}{m_u+m_s} \frac{1}{(M_{Z}^2-p^{\prime2})(M_{J/\psi}^2-p^2)(M_{K}^2-q^2)} g_{\mu\nu}+\cdots \nonumber\\
&=& \Pi^1_M(p^{\prime2},p^2,q^2)\,g_{\mu\nu}+\cdots\, ,
\end{eqnarray}
\begin{eqnarray}
\Pi_{\mu\nu}^{T,2}(p,q)&=& \frac{f_{\eta_c}M_{\eta_c}^2f_{K^*}M_{K^*}\lambda_{Z}^TG_{Z\eta_c K^*}^T}{2m_c} \frac{i}{(M_{Z}^2-p^{\prime2})(M_{\eta_c}^2-p^2)(M_{K^*}^2-q^2)} g_{\mu\nu}+\cdots \nonumber\\
&=& \Pi_T^2(p^{\prime2},p^2,q^2)\,g_{\mu\nu}+\cdots\, ,
\end{eqnarray}
\begin{eqnarray}
\Pi_{\mu\nu}^{M,2}(p,q)&=& \frac{f_{\eta_c}M_{\eta_c}^2f_{K^*}M_{K^*}\lambda_{Z}^MG_{Z\eta_c K^*}^M}{2m_c} \frac{1}{(M_{Z}^2-p^{\prime2})(M_{\eta_c}^2-p^2)(M_{K^*}^2-q^2)} g_{\mu\nu}+\cdots \nonumber\\
&=& \Pi_M^2(p^{\prime2},p^2,q^2)\,g_{\mu\nu}+\cdots\, ,
\end{eqnarray}
\begin{eqnarray}
\Pi_{\mu\nu}^{T,3}(p,q)&=& \frac{f_{D_s}M_{D_s}^2f_{D^*}M_{D^*}\lambda_{Z}^TG^T_{Z D^*\bar{D}_s}}{m_c+m_s} \frac{i}{(M_{Z}^2-p^{\prime2})(M_{D^*}^2-p^2)(M_{D_s}^2-q^2)} g_{\mu\nu}+\cdots \nonumber\\
&=& \Pi_T^3(p^{\prime2},p^2,q^2)\,g_{\mu\nu}+\cdots \, ,
\end{eqnarray}
\begin{eqnarray}
\Pi_{\mu\nu}^{M,3}(p,q)&=& \frac{f_{D_s}M_{D_s}^2f_{D^*}M_{D^*}\lambda_{Z}^M G^M_{Z D^*\bar{D}_s}}{m_c+m_s} \frac{1}{(M_{Z}^2-p^{\prime2})(M_{D^*}^2-p^2)(M_{D_s}^2-q^2)} g_{\mu\nu}+\cdots \nonumber\\
&=& \Pi_M^3(p^{\prime2},p^2,q^2)\,g_{\mu\nu}+\cdots \, ,
\end{eqnarray}
where $p^\prime=p+q$, the $f_{J/\psi}$, $f_{K}$, $f_{\eta_c}$, $f_{K^*}$, $f_{D^*}$, $f_{D_s}$, $\lambda_Z^T$ and $\lambda_{Z}^M$ are the decay constants of the mesons $J/\psi$, $K$, $\eta_c$, $K^*$, $D^*$, $D_s$, $Z_{cs}^T$ and $Z_{cs}^M$, respectively,
\begin{eqnarray}
\langle0|J_{\mu}^{J/\psi}(0)|J/\psi(p)\rangle&=&f_{J/\psi}M_{J/\psi}\,\xi_\mu \,\, , \nonumber \\
\langle0|J_{\mu}^{K^*}(0)|K^*(q)\rangle&=&f_{K^*}M_{K^*}\,\varepsilon_\mu \,\, , \nonumber \\
\langle0|J_{\mu}^{D^*}(0)|D^*(p)\rangle&=&f_{D^*}M_{D^*}\,\varsigma_\mu \,\, , \nonumber \\
\langle0|J_{5}^{K}(0)|K(q)\rangle&=&\frac{f_{K}M_{K}^2}{m_u+m_s} \,\, , \nonumber \\
\langle0|J_{5}^{\eta_c}(0)|\eta_c(p)\rangle&=&\frac{f_{\eta_c}M_{\eta_c}^2}{2m_c} \,\, , \nonumber \\
\langle0|J_{5}^{D_s}(0)|D_s(q)\rangle&=&\frac{f_{D_s}M_{D_s}^2}{m_c+m_s} \,\, , \nonumber \\
\langle 0|J^{T/M}_\nu(0)|Z_{cs}^{T/M}(p^\prime)\rangle&=&\lambda_{Z}^{T/M}\,\zeta_\nu\,\, ,
\end{eqnarray}
and the $\xi$, $\varepsilon$, $\varsigma$ and $\zeta$ are polarization vectors of the $J/\psi$, $K^*$, $D^*$ and $Z_{cs}^{T/M}$, respectively, the hadronic coupling constants
$G^T_{ZJ/\psi K}$, $G^M_{ZJ/\psi K}$, $G^T_{Z\eta_cK^*}$, $G^M_{Z\eta_cK^*}$, $G_{Z D^*\bar{D}_s}^T$ and $G_{Z D^*\bar{D}_s}^M$ are defined by
\begin{eqnarray}
\langle J/\psi(p)K(q)|Z^T_{cs}(p^{\prime})\rangle&=&-\xi^*(p)\cdot\zeta(p^{\prime})\, G^T_{ZJ/\psi K} \, , \nonumber\\
\langle\eta_c(p)K^*(q)|Z_{cs}^T(p^{\prime})\rangle&=&-\varepsilon^*(q)\cdot\zeta(p^{\prime})\, G^T_{Z\eta_cK^*} \, , \nonumber\\
\langle D^*(p)D_s(q)|Z^T_{cs}(p^{\prime})\rangle&=&-\varsigma^*(p)\cdot\zeta(p^{\prime})\, G^T_{Z D^*\bar{D}_s} \, , \nonumber\\
\langle J/\psi(p)K(q)|Z^M_{cs}(p^{\prime})\rangle&=&i\xi^*(p)\cdot\zeta(p^{\prime})\, G^M_{ZJ/\psi K} \, , \nonumber\\
\langle\eta_c(p)K^*(q)|Z_{cs}^M(p^{\prime})\rangle&=&i\varepsilon^*(q)\cdot\zeta(p^{\prime})\, G^M_{Z\eta_cK^*} \, , \nonumber\\
\langle D^*(p)D_s(q)|Z^M_{cs}(p^{\prime})\rangle&=&i\varsigma^*(p)\cdot\zeta(p^{\prime})\, G^M_{Z D^*\bar{D}_s} \, .
\end{eqnarray}
In this work, we choose the tensor structure $g_{\mu\nu}$ to explore the hadronic coupling constants $G^{T/M}_{ZJ/\psi K}$, $G^{T/M}_{Z\eta_cK^*}$ and $G^{T/M}_{Z D^* \bar{D}_s}$.
We accomplish the operator product expansion up to the vacuum condensates of dimension 5 and neglect the tiny gluon condensate contributions, just like in our previous works \cite{WZG-ZJX-Zc-Decay,WZG-Y4660-Decay,WZG-X4140-decay,WZG-X4274-decay,WZG-Z4600-decay}, then obtain the QCD spectral densities through dispersion relation.
We match the hadron side with the QCD side of the correlation functions bellow the continuum thresholds according to the dispersion relation at the QCD side to warrant rigorous quark-hadron duality,
\begin{eqnarray}
\int_{\Delta_s^2}^{s_{0}}ds \int_{\Delta_u^2}^{u_0}du \frac{\rho_{QCD}(p^{\prime2},s,u)}{(s-p^2)(u-q^2)}&=& \int_{\Delta_s^2}^{s_0}ds \int_{\Delta_u^2}^{u_0}du \left[ \int_{\Delta_{s}^{\prime2}}^{\infty}ds^\prime \frac{\rho_H(s^\prime,s,u)}{(s^\prime-p^{\prime2})(s-p^2)(u-q^2)} \right]\, ,
\end{eqnarray}
where the $\Delta_s^2$, $\Delta_u^2$ and $\Delta_{s}^{\prime2}$ are the thresholds, and the $s_0$ and $u_0$ are the continuum thresholds, the $\rho_{QCD}(p^{\prime2},s,u)$ and $\rho_H(s^\prime,s,u)$ are the spectral densities at the QCD side and at the hadron side, respectively, we accomplish the integral over $ds^\prime$ firstly, and introduce some unknown parameters to parameterize the contributions involving the higher resonances and continuum states in the $s^\prime$ channel \cite{WZG-ZJX-Zc-Decay,WZG-Y4660-Decay}. We accomplish the integral over the variable $ds^\prime$ firstly, and do not need the continuum threshold parameters $s^{\prime Z}_0$, and the duality is rigorous.
We set $p^{\prime2}=p^2$ and $p^{\prime2}=4p^2$ in the correlation functions $\Pi^{1/2}_{T/M}(p^{\prime 2},p^2,q^2)$ and $\Pi^{3}_{T/M}(p^{\prime 2},p^2,q^2)$, respectively, and accomplish the double Borel transform in regard to the variables $P^2=-p^2$ and $Q^2=-q^2$ respectively, then set the Borel parameters $T_1^2=T_2^2=T^2$ to get the six QCD sum rules,
\begin{eqnarray} \label{JpsiK-Z-T}
&&\frac{f_{K}M_{K}^2f_{J/\psi}M_{J/\psi}\lambda_{Z}^TG^T_{ZJ/\psi K}}{m_u+m_s}\frac{1}{M_{Z}^2-M_{J/\psi}^2} \left[ \exp\left(-\frac{M_{J/\psi}^2}{T^2} \right)-\exp\left(-\frac{M_{Z}^2}{T^2} \right)\right]\exp\left(-\frac{M_{K}^2}{T^2} \right) \nonumber\\
&&+C^T_{J/\psi K} \exp\left(-\frac{M_{J/\psi}^2+M_{K}^2}{T^2} \right)=\frac{1}{64\sqrt{2}\pi^4}\int_{4m_c^2}^{s^0_{J/\psi}} ds \int_{0}^{u^0_{K}} du \sqrt{1-\frac{4m_c^2}{s}} u\left(2s+4m_c^2-3m_s m_c\right)\nonumber\\
&&\exp\left(-\frac{s+u}{T^2} \right)-\frac{m_s\left[2\langle\bar{q}q\rangle+\langle\bar{s}s\rangle\right]}{24\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{J/\psi}} ds \sqrt{1-\frac{4m_c^2}{s}}\left(s+2m_c^2\right)\exp\left(-\frac{s}{T^2}\right) \nonumber\\
&&-\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{288\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{J/\psi}} ds \frac{s+20m_c^2}{\sqrt{s\left(s-4m_c^2\right)}} \exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&-\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{24\sqrt{2}\pi^2T^2}\int_{4m_c^2}^{s^0_{J/\psi}} ds \sqrt{1-\frac{4m_c^2}{s}}\left(s+2m_c^2\right) \exp\left(-\frac{s}{T^2}\right) \, ,
\end{eqnarray}
\begin{eqnarray}
&&\frac{f_{K}M_{K}^2f_{J/\psi}M_{J/\psi}\lambda_{Z}^MG^M_{ZJ/\psi K}}{m_u+m_s}\frac{1}{M_{Z}^2-M_{J/\psi}^2} \left[ \exp\left(-\frac{M_{J/\psi}^2}{T^2} \right)-\exp\left(-\frac{M_{Z}^2}{T^2} \right)\right]\exp\left(-\frac{M_{K}^2}{T^2} \right) \nonumber\\
&&+C^M_{J/\psi K} \exp\left(-\frac{M_{J/\psi}^2+M_{K}^2}{T^2} \right)=\frac{1}{128\sqrt{2}\pi^4}\int_{4m_c^2}^{s^0_{J/\psi}} ds \int_{0}^{u^0_{K}} du \sqrt{1-\frac{4m_c^2}{s}} u\left(2s+4m_c^2-3m_s m_c\right)\nonumber\\
&&\exp\left(-\frac{s+u}{T^2} \right)-\frac{m_s\left[2\langle\bar{q}q\rangle+\langle\bar{s}s\rangle\right]}{48\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{J/\psi}} ds \sqrt{1-\frac{4m_c^2}{s}}\left(s+2m_c^2\right)\exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&-\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{288\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{J/\psi}} ds \frac{s+20m_c^2}{\sqrt{s\left(s-4m_c^2\right)}} \exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&-\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{48\sqrt{2}\pi^2T^2}\int_{4m_c^2}^{s^0_{J/\psi}} ds \sqrt{1-\frac{4m_c^2}{s}}\left(s+2m_c^2\right)\exp\left(-\frac{s}{T^2}\right) \, ,
\end{eqnarray}
\begin{eqnarray}
&&\frac{f_{\eta_c}M_{\eta_c}^2f_{K^*}M_{K^*}\lambda_{Z}^TG^T_{Z\eta_c K^*}}{2m_c } \frac{1}{M_{Z}^2-M_{\eta_c}^2}\left[ \exp\left(-\frac{M_{\eta_c}^2}{T^2} \right)-\exp\left(-\frac{M_{Z}^2}{T^2} \right)\right]\exp\left(-\frac{M_{K^*}^2}{T^2} \right) \nonumber\\
&&+C^T_{\eta_c K^*} \exp\left(-\frac{M_{\eta_c}^2+M_{K^*}^2}{T^2} \right)=-\frac{1}{64\sqrt{2}\pi^4}\int_{4m_c^2}^{s^0_{\eta_c}} ds \int_{0}^{u^0_{K^*}} du \sqrt{1-\frac{4m_c^2}{s}} u\left(2s-3m_sm_c \right)\nonumber\\
&&\exp\left(-\frac{s+u}{T^2} \right)+\frac{m_s\langle\bar{q}q\rangle}{8\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{\eta_c}} ds \sqrt{1-\frac{4m_c^2}{s}}\, s\, \exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&+\frac{m_s\langle\bar{s}g_s\sigma Gs\rangle}{96\sqrt{2}\pi^2T^2}\int_{4m_c^2}^{s^0_{\eta_c}} ds \sqrt{1-\frac{4m_c^2}{s}}\, s\, \exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&+\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{72\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{\eta_c}} ds \frac{s+2m_c^2}{\sqrt{s\left(s-4m_c^2\right)}}\exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&-\frac{m_c\left[\langle\bar{q}g_s\sigma Gq\rangle+\langle\bar{s}g_s\sigma Gs\rangle\right]}{24\sqrt{2}\pi^2} \int_{4m_c^2}^{s^0_{\eta_c}} ds \sqrt{1-\frac{4m_c^2}{s}}\exp\left(-\frac{s}{T^2}\right) \, ,
\end{eqnarray}
\begin{eqnarray}
&&\frac{f_{\eta_c}M_{\eta_c}^2f_{K^*}M_{K^*}\lambda_{Z}^MG^M_{Z\eta_c K^*}}{2m_c } \frac{1}{M_{Z}^2-M_{\eta_c}^2}\left[ \exp\left(-\frac{M_{\eta_c}^2}{T^2} \right)-\exp\left(-\frac{M_{Z}^2}{T^2} \right)\right]\exp\left(-\frac{M_{K^*}^2}{T^2} \right) \nonumber\\
&&+C^M_{\eta_c K^*} \exp\left(-\frac{M_{\eta_c}^2+M_{K^*}^2}{T^2} \right)=\frac{1}{128\sqrt{2}\pi^4}\int_{4m_c^2}^{s^0_{\eta_c}} ds \int_{0}^{u^0_{K^*}} du \sqrt{1-\frac{4m_c^2}{s}} u\left(2s-3m_sm_c \right)\nonumber\\
&&\exp\left(-\frac{s+u}{T^2} \right)-\frac{m_s\langle\bar{q}q\rangle}{16\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{\eta_c}} ds \sqrt{1-\frac{4m_c^2}{s}} s \exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&-\frac{m_s\langle\bar{s}g_s\sigma Gs\rangle}{192\sqrt{2}\pi^2T^2}\int_{4m_c^2}^{s^0_{\eta_c}} ds \sqrt{1-\frac{4m_c^2}{s}}\,s\,\exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&-\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{72\sqrt{2}\pi^2} \int_{4m_c^2}^{s^0_{\eta_c}} ds \frac{s+2m_c^2}{\sqrt{s\left(s-4m_c^2\right)}} \exp\left(-\frac{s}{T^2}\right)\nonumber\\
&&+\frac{m_c\left[\langle\bar{q}g_s\sigma Gq\rangle+\langle\bar{s}g_s\sigma Gs\rangle\right]}{48\sqrt{2}\pi^2}\int_{4m_c^2}^{s^0_{\eta_c}} ds \sqrt{1-\frac{4m_c^2}{s}}\exp\left(-\frac{s}{T^2}\right) \, ,
\end{eqnarray}
\begin{eqnarray}
&&\frac{f_{D_s}M_{D_s}^2f_{D^*}M_{D^*}\lambda_{Z}^TG^T_{Z D^*\bar{D}_s}}{4(m_c+m_s) } \frac{1}{\widetilde{M}_{Z}^2-M_{D^*}^2}\left[ \exp\left(-\frac{M_{D^*}^2}{T^2} \right)-\exp\left(-\frac{\widetilde{M}_{Z}^2}{T^2} \right)\right]\exp\left(-\frac{M_{D_s}^2}{T^2} \right) \nonumber\\
&&+ C^T_{ D^*\bar{D}_s} \exp\left(-\frac{M_{D^*}^2+M_{D_s}^2}{T^2} \right)=-\frac{m_c\langle\bar{s}g_s\sigma Gs\rangle}{96\sqrt{2}\pi^2}\int_{m_c^2}^{s^0_{D^*}} ds \left( \frac{9}{2}-\frac{10m_c^2}{s}+\frac{3m_c^4}{2s^2}\right)\exp\left(-\frac{s+m_c^2}{T^2}\right)\nonumber\\
&&+\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{96\sqrt{2}\pi^2}\int_{m_c^2}^{u^0_{D_s}} du \left(2+\frac{3m_c^2}{u}\right) \frac{m_c^2}{u}\exp\left(-\frac{u+m_c^2}{T^2}\right)\nonumber\\
&&+\frac{m_c\langle\bar{q}g_s\sigma Gq\rangle}{96\sqrt{2}\pi^2}\int_{m_c^2}^{u^0_{D_s}} du \left( \frac{3}{2}+\frac{4m_c^2}{u}-\frac{3m_c^4}{2u^2}\right)\exp\left(-\frac{u+m_c^2}{T^2}\right)\nonumber\\
&&-\frac{m_s\langle\bar{q}g_s\sigma Gq\rangle}{192\sqrt{2}\pi^2}\int_{m_c^2}^{u^0_{D_s}} du \frac{1}{u-m_c^2}\left(u-9m_c^2-\frac{5m_c^4}{u}-\frac{3m_c^6}{u^2}\right)\exp\left(-\frac{u+m_c^2}{T^2}\right)\, ,
\end{eqnarray}
\begin{eqnarray}
&&\frac{f_{D_s}M_{D_s}^2f_{D^*}M_{D^*}\lambda_{Z}^MG^M_{Z D^*\bar{D}_s}}{4(m_c+m_s) } \frac{1}{\widetilde{M}_{Z}^2-M_{D^*}^2}\left[ \exp\left(-\frac{M_{D^*}^2}{T^2} \right)-\exp\left(-\frac{\widetilde{M}_{Z}^2}{T^2} \right)\right]\exp\left(-\frac{M_{D_s}^2}{T^2} \right) \nonumber\\
&&+ C^M_{ D^*\bar{D}_s} \exp\left(-\frac{M_{D^*}^2+M_{D_s}^2}{T^2} \right)=-\frac{3}{64\sqrt{2}\pi^4}\int_{m_c^2}^{s^0_{D^*}} ds \int_{m_c^2}^{u^0_{D_s}} du \left(1-\frac{m_c^2}{s}\right)\left(1-\frac{m_c^2}{u}\right)\nonumber\\
&&\left(2s-m_c^2-\frac{m_c^4}{s}\right)\left(u-m_c^2+2m_s m_c\right)\exp\left(-\frac{s+u}{T^2}\right)\nonumber\\
&&+\frac{m_c\langle\bar{s}s\rangle}{8\sqrt{2}\pi^2}\int_{m_c^2}^{s^0_{D^*}} ds \left(1-\frac{m_c^2}{s}\right)\left(2s-m_c^2-\frac{m_c^4}{s}\right)\exp\left(-\frac{s+m_c^2}{T^2}\right)\nonumber\\
&&+\frac{3m_c\langle\bar{q}q\rangle}{8\sqrt{2}\pi^2}\int_{m_c^2}^{u^0_{D_s}} du \left(1-\frac{m_c^2}{u}\right)\left(u-m_c^2+2m_s m_c\right)
\exp\left(-\frac{u+m_c^2}{T^2}\right)\nonumber\\
&&+\frac{m_s\langle\bar{s}s\rangle}{16\sqrt{2}\pi^2}\left(1+\frac{m_c^2}{T^2}\right)\int_{m_c^2}^{s^0_{D^*}} ds \left(1-\frac{m_c^2}{s}\right)\left(2s-m_c^2-\frac{m_c^4}{s}\right)\exp\left(-\frac{s+m_c^2}{T^2}\right)\nonumber\\
&&-\frac{3m_c^3\langle\bar{q}g_s\sigma Gq\rangle}{32\sqrt{2}\pi^2T^4}\int_{m_c^2}^{u^0_{D_s}} du \left(1-\frac{m_c^2}{u}\right)\left(u-m_c^2+2m_s m_c\right) \exp\left(-\frac{u+m_c^2}{T^2}\right)\nonumber\\
&&-\frac{m_sm_c^4\langle\bar{s}g_s\sigma Gs\rangle}{96\sqrt{2}\pi^2T^6}\int_{m_c^2}^{s^0_{D^*}} ds \left(1-\frac{m_c^2}{s}\right) \left(2s-m_c^2-\frac{m_c^4}{s}\right)\exp\left(-\frac{s+m_c^2}{T^2}\right) \nonumber\\
&&+\frac{m_c\langle\bar{s}g_s\sigma Gs\rangle}{16\sqrt{2}\pi^2T^2}\left(1-\frac{m_c^2}{2T^2} \right)\int_{m_c^2}^{s^0_{D^*}} ds\left(1-\frac{m_c^2}{s}\right) \left(2s-m_c^2-\frac{m_c^4}{s}\right)\exp\left(-\frac{s+m_c^2}{T^2}\right) \, ,
\end{eqnarray}
where $\widetilde{M}_{Z}^2=\frac{M_{Z}^2}{4}$, the $s^0_{J/\psi}$, $u^0_{K}$, $s^0_{\eta_c}$, $u^0_{K^*}$, $s^0_{D^*}$ and $u^0_{D_s}$ are the continuum threshold parameters. There exist end-divergences at the endpoints $s=4m_c^2$ and $u=m_c^2$, we regularize the endpoint divergences with the simple replacements $\frac{1}{\sqrt{s-4m_c^2}} \to \frac{1}{\sqrt{s-4m_c^2+4m_s^2}}$ and $\frac{1}{u-m_c^2} \to \frac{1}{u-m_c^2+4m_s^2}$ by adding a small squared $s$-quark mass $4m_s^2$ \cite{WZG-ZJX-Zc-Decay,WZG-Y4660-Decay,WZG-X4140-decay,WZG-X4274-decay,WZG-Z4600-decay}.
Furthermore, we smear the dependencies of the parameters $C_{J/\psi K}^T$, $C_{J/\psi K}^M$, $C_{\eta_c K^*}^T$, $C_{\eta_c K^*}^M$, $C_{D^* \bar{D}_s}^T$, and $C_{D^* \bar{D}_s}^M$ on the Lorentz invariants $p^{\prime2}$, $p^2$, $q^2$, and take them as free parameters, and search for the best values to
delete the contaminations from the high resonances and continuum states to acquire stable QCD sum rules with variations of
the Borel parameters $T^2$. One can consult Refs.\cite{WZG-ZJX-Zc-Decay,WZG-Y4660-Decay} for the definitions of the unknown parameters $C_{J/\psi K}^T$, $C_{J/\psi K}^M$, $C_{\eta_c K^*}^T$, $C_{\eta_c K^*}^M$, $C_{D^* \bar{D}_s}^T$,
and $C_{D^* \bar{D}_s}^M$.
\section{Numerical results and discussions}
At the QCD side, we take the standard values
$\langle
\bar{q}q \rangle=-(0.24\pm 0.01\, \rm{GeV})^3$, $\langle\bar{s}s\rangle=(0.8\pm0.1)\langle\bar{q}q\rangle$,
$\langle\bar{q}g_s\sigma G q \rangle=m_0^2\langle \bar{q}q \rangle$,
$\langle\bar{s}g_s\sigma G s \rangle=m_0^2\langle \bar{s}s \rangle$,
$m_0^2=(0.8 \pm 0.1)\,\rm{GeV}^2$ at the energy scale $\mu=1\, \rm{GeV}$
\cite{SVZ79,Reinders85,Colangelo-Review}, and take the $\overline{MS}$ quark masses $m_{c}(m_c)=(1.275\pm0.025)\,\rm{GeV}$ and $m_s(\mu=2\,\rm{GeV})=(0.095\pm0.005)\,\rm{GeV}$ from the Particle Data Group \cite{PDG}. We set $m_u=m_d=0$ and take account of
the energy-scale dependence of the input parameters,
\begin{eqnarray}
\langle\bar{q}q \rangle(\mu)&=&\langle\bar{q}q \rangle({\rm 1GeV})\left[\frac{\alpha_{s}({\rm 1GeV})}{\alpha_{s}(\mu)}\right]^{\frac{12}{33-2n_f}}\, , \nonumber\\
\langle\bar{s}s \rangle(\mu)&=&\langle\bar{s}s \rangle({\rm 1GeV})\left[\frac{\alpha_{s}({\rm 1GeV})}{\alpha_{s}(\mu)}\right]^{\frac{12}{33-2n_f}}\, , \nonumber\\
\langle\bar{q}g_s \sigma Gq \rangle(\mu)&=&\langle\bar{q}g_s \sigma Gq \rangle({\rm 1GeV})\left[\frac{\alpha_{s}({\rm 1GeV})}{\alpha_{s}(\mu)}\right]^{\frac{2}{33-2n_f}}\, , \nonumber\\
\langle\bar{s}g_s \sigma Gs \rangle(\mu)&=&\langle\bar{s}g_s \sigma Gs \rangle({\rm 1GeV})\left[\frac{\alpha_{s}({\rm 1GeV})}{\alpha_{s}(\mu)}\right]^{\frac{2}{33-2n_f}}\, , \nonumber\\
m_c(\mu)&=&m_c(m_c)\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}(m_c)}\right]^{\frac{12}{33-2n_f}} \, ,\nonumber\\
m_s(\mu)&=&m_s({\rm 2GeV})\left[\frac{\alpha_{s}(\mu)}{\alpha_{s}({\rm 2GeV})}\right]^{\frac{12}{33-2n_f}} \, ,\nonumber\\
\alpha_s(\mu)&=&\frac{1}{b_0t}\left[1-\frac{b_1}{b_0^2}\frac{\log t}{t} +\frac{b_1^2(\log^2{t}-\log{t}-1)+b_0b_2}{b_0^4t^2}\right]\, ,
\end{eqnarray}
from the renormalization group equation, $t=\log \frac{\mu^2}{\Lambda_{QCD}^2}$, $b_0=\frac{33-2n_f}{12\pi}$, $b_1=\frac{153-19n_f}{24\pi^2}$, $b_2=\frac{2857-\frac{5033}{9}n_f+\frac{325}{27}n_f^2}{128\pi^3}$, $\Lambda_{QCD}=210\,\rm{MeV}$, $292\,\rm{MeV}$ and $332\,\rm{MeV}$ for the flavors $n_f=5$, $4$ and $3$, respectively \cite{PDG,Narison-mix}, and evolve all the input parameters to the typical energy scale $\mu=m_c(m_c)\approx 1.3\,\rm{GeV}$ to extract hadronic coupling constants \cite{WZG-Zcs3985-tetra,WZG-Zcs3985-mole,WZG-Hidden-charm-PRD}.
At the hadron side, we take the parameters as $M_{K}=0.4937\,\rm{GeV}$, $M_{K^*}=0.8917\,\rm{GeV}$,
$M_{J/\psi}=3.0969\,\rm{GeV}$, $M_{\eta_c}=2.9834\,\rm{GeV}$, $M_{D^*}=2.007\,\rm{GeV}$ and $D_{s}=1.969\,\rm{GeV}$ \cite{PDG}, $f_{K}=0.156\,\rm{GeV}$ \cite{PDG}, $f_{K^*}=0.220\,\rm{GeV}$, $\sqrt{u^0_{K}}=1.0\,\rm{GeV}$, $\sqrt{u^0_{K^*}}=1.3\,\rm{GeV}$ \cite{PBall-decay-Kv},
$f_{D_s}=240\,\rm{MeV}$, $\sqrt{u^0_{D_s}}=2.6\,\rm{GeV}$, $f_{D^*}=263\,\rm{MeV}$, $\sqrt{s^0_{D^*}}=2.5\,\rm{GeV}$ \cite{WZG-EPJC-decay},
$f_{J/\psi}=0.418 \,\rm{GeV}$, $f_{\eta_c}=0.387 \,\rm{GeV}$ \cite{Becirevic}, $\sqrt{s^0_{J/\psi}}=3.6\,\rm{GeV}$, $\sqrt{s^0_{\eta_c}}=3.5\,\rm{GeV}$ \cite{PDG}, $M_{Z}=3.99\,\rm{GeV}$, $\lambda_{Z}^T=2.85\times 10^{-2}\,\rm{GeV}^5$ \cite{WZG-Zcs3985-tetra}, $\lambda_{Z}^M=1.96\times 10^{-2}\,\rm{GeV}^5$ \cite{WZG-Zcs3985-mole}, $\frac{f_{K}M^2_{K}}{m_u+m_s}=-\frac{\langle \bar{q}q\rangle+\langle \bar{s}s\rangle}{f_{K}(1-\delta_K)}$ from the Gell-Mann-Oakes-Renner relation, $\delta_K=0.50$ \cite{GMOR-fK}.
In calculations, we fit the unknown parameters to be $C^T_{J/\psi K}=0.00045+0.00038\times T^2\,\rm{GeV}^8$,
$C_{\eta_c K^*}^T=-0.00184-0.00068\times T^2\,\rm{GeV}^8$, $C^T_{D^* \bar{D}_s}=0+0\times T^2\,\rm{GeV}^8$,
$C^M_{J/\psi K}=0.0004+0.00017\times T^2\,\rm{GeV}^8$,
$C^M_{\eta_c K^*}=0.00094+0.00033\times T^2\,\rm{GeV}^8$,
and $C^M_{D^* \bar{D}_s}=-0.023-0.00405\times T^2\,\rm{GeV}^8$ to acquire the flat Borel platforms $T^2_{max}-T^2_{min}=1\,\rm{GeV}^2$, where the max and min represent the maximum and minimum values, respectively, see Fig.\ref{ZcsPsiK-T-M} for example.
In the picture of tetraquark states, the Borel parameters are $T^2_{J/\psi K}=(2.7-3.7)\,\rm{GeV}^2$, $T^2_{\eta_c K^*}=(2.5-3.5)\,\rm{GeV}^2$ and $T^2_{D^*\bar{D}_s}=(3.1-4.1)\,\rm{GeV}^2$ for the hadronic coupling constants $G_{ZJ/\psi K}^T$,
$G^T_{Z\eta_cK^*}$ and $G^T_{ZD^*\bar{D}_s}$, respectively, while in the picture of molecular states, the Borel parameters $T^2_{J/\psi K}=(2.7-3.7)\,\rm{GeV}^2$, $T^2_{\eta_c K^*}=(2.3-3.3)\,\rm{GeV}^2$ and $T^2_{D^* \bar{D}_s}=(2.3-3.3)\,\rm{GeV}^2$ for the hadronic coupling constants $G_{ZJ/\psi K}^M$, $G^M_{Z\eta_cK^*}$ and $G^M_{ZD^*\bar{D}_s}$, respectively, where we add the subscripts $J/\psi K$, $\eta_c K^*$ and $D^*\bar{D}_s$ to denote the corresponding channels.
The uncertainties originate from the uncertainties of the input parameters at the QCD side, in general, can be absorbed into the decay constants and coupling constants together, for example, the $f_{J/\psi}$, $f_{K}$, $\lambda_Z^T$ and $G^T_{ZJ/\psi K}$ in the QCD sum rules in Eq.\eqref{JpsiK-Z-T}.
If we use the symbol $\xi$ to stand for the input parameters at the QCD side, the uncertainties $\xi \to \xi +\delta \xi$ lead to the uncertainties $f_{J/\psi}f_{K}\lambda^T_{Z}G^T_{ZJ/\psi K} \to f_{J/\psi}f_{K}\lambda^T_{Z}G^T_{ZJ/\psi K}+\delta\,f_{J/\psi}f_{K}\lambda^T_{Z}G^T_{ZJ/\psi K}$, $C^T_{J/\psi K} \to C^T_{J/\psi K}+\delta C^T_{J/\psi K}$,
where
\begin{eqnarray}\label{Uncertainty-4}
\delta\,f_{J/\psi}f_{K}\lambda^T_{Z}G^T_{ZJ/\psi K} &=&f_{J/\psi}f_{K}\lambda^T_{Z}G^T_{ZJ/\psi K}\left( \frac{\delta f_{J/\psi}}{f_{J/\psi}} +\frac{\delta f_{K}}{f_{K}}+\frac{\delta \lambda^T_{Z}}{\lambda^T_{Z}}+\frac{\delta G^T_{ZJ/\psi K}}{G^T_{ZJ/\psi K}}\right)\, ,
\end{eqnarray}
we can set $\frac{\delta f_{J/\psi}}{f_{J/\psi}} =\frac{\delta f_{K}}{f_{K}}=\frac{\delta \lambda^T_{Z}}{\lambda^T_{Z}}=\frac{\delta G^T_{ZJ/\psi K}}{G^T_{ZJ/\psi K}}
$ approximately, then
\begin{eqnarray}\label{Uncertainty}
\delta\,f_{J/\psi}f_{K}\lambda^T_{Z}G^T_{ZJ/\psi K} &=&f_{J/\psi}f_{K}\lambda^T_{Z}G^T_{ZJ/\psi K} \frac{4\delta G^T_{ZJ/\psi K}}{G^T_{ZJ/\psi K}} \, ,
\end{eqnarray}
to avoid overestimating the uncertainty of the hadronic coupling constant. In calculations, we estimate the uncertainties according to Eq.\eqref{Uncertainty} analogously and neglect the uncertainties of the unknown parameters $C^T_{J/\psi K}$, $C_{\eta_c K^*}^T$, $C^T_{D^* \bar{D}_s}$, $C^M_{J/\psi K}$, $C^M_{\eta_c K^*}$,
and $C^M_{D^* \bar{D}_s}$, except in the case of the $\delta m_c(m_c)$ for the $C^M_{D^* \bar{D}_s}$, where we have to take account of the uncertainty to acquire flat Borel platform.
Now let us obtain the values of the hadronic coupling constants routinely,
\begin{eqnarray}
G^T_{ZJ/\psi K} &=&1.79\pm0.10\,\rm{GeV}\, , \nonumber\\
|G^T_{Z\eta_cK^*}|&=&2.97\pm 0.22\,\rm{GeV}\, , \nonumber\\
|G^T_{ZD^*\bar{D}_s}|&=&0.71\pm0.04\,\rm{GeV}\, , \nonumber\\
G^M_{ZJ/\psi K} &=&1.23\pm0.07\,\rm{GeV}\, , \nonumber\\
G^M_{Z\eta_c K^*}&=&2.20\pm 0.16\,\rm{GeV}\, , \nonumber\\
|G^M_{ZD^*\bar{D}_s}|&=&9.60\pm1.27\,\rm{GeV}\, .
\end{eqnarray}
\begin{figure}
\centering
\includegraphics[totalheight=5cm,width=7cm]{ZcsPsiK-T.EPS}
\includegraphics[totalheight=5cm,width=7cm]{ZcsPsiK-M.EPS}
\caption{ The hadronic coupling constants $G^T_{ZJ/\psi K}$ (I) and $G^M_{ZJ/\psi K}$ (II) with variations of the Borel parameter $T^2$, where the regions between the two vertical lines are the Borel windows. }\label{ZcsPsiK-T-M}
\end{figure}
Then we choose the masses $M_{K^-}=0.4937\,\rm{GeV}$, $M_{K^{*+}}=0.8917\,\rm{GeV}$,
$M_{J/\psi}=3.0969\,\rm{GeV}$, $M_{\eta_c}=2.9834\,\rm{GeV}$,
$M_{D^{*0}}=2.0069 \,\rm{GeV}$, $M_{D^{-}_s}=1.9690\,\rm{GeV}$, $M_{D^0}=1.8648 \,\rm{GeV}$, $M_{D^{*-}_s}=2.1122\,\rm{GeV}$ \cite{PDG},
$M_{Z_c}=3.99\,\rm{GeV}$ \cite{WZG-Zcs3985-tetra,WZG-Zcs3985-mole}, and obtain the partial decay widths,
\begin{eqnarray}
\Gamma(Z_{cs}^T\to J/\psi K^-)&=&5.36\pm 0.60\,\rm{MeV} \, ,\nonumber\\
\Gamma(Z_{cs}^T\to\eta_c K^{*-})&=&9.54\pm 1.41\,\rm{MeV} \, , \nonumber\\
\Gamma(Z_{cs}^T\to D^{*0} \bar{D}^{-}_s)&=&0.21\pm 0.02\,\rm{MeV} \, ,\nonumber\\
\Gamma(Z_{cs}^T\to D^0 \bar{D}^{*-}_s)&=&0.20\pm 0.02\,\rm{MeV} \, ,
\end{eqnarray}
\begin{eqnarray}
\Gamma(Z_{cs}^M\to J/\psi K^-)&=&2.53\pm 0.29\,\rm{MeV} \, ,\nonumber\\
\Gamma(Z_{cs}^M\to\eta_c K^{*-})&=&5.23\pm 0.76 \,\rm{MeV} \, , \nonumber\\
\Gamma(Z_{cs}^M\to D^{*0} \bar{D}^{-}_s)&=&38.69\pm 10.24\,\rm{MeV} \, ,\nonumber\\
\Gamma(Z_{cs}^M\to D^0 \bar{D}^{*-}_s)&=&37.06\pm 9.81\,\rm{MeV} \, ,
\end{eqnarray}
and the total widths,
\begin{eqnarray}
\Gamma_{Z_{cs}}^T &=&15.31\pm 2.06\,\rm{MeV}\, ,\nonumber\\
\Gamma_{Z_{cs}}^M &=& 83.51\pm21.09\,\rm{MeV}\, ,
\end{eqnarray}
which are consistent with the experimental data $13.8^{+8.1}_{-5.2}\pm4.9\,\rm{MeV}$ from the BESIII collaboration \cite{BES3985} and
$131 \pm 15 \pm 26\,\rm{MeV}$ from the LHCb collaboration \cite{LHCb-Zcs4000}, respectively.
There are two $Z_{cs}$ states with analogous masses, just like in the case of the $Z_c^{\pm}$ states, there exist a tetraquark candidate $Z_c(3900)$ and a molecule candidate $Z_c(3885)$ \cite{WZG-ZJX-Zc-Decay}.
\section{Conclusion}
In this work, we explore the hadronic coupling constants $G_{ZJ/\psi K}$, $G_{Z\eta_c K^*}$, $G_{ZD^* \bar{D}_s}$ of the exotic states $Z_{cs}(3985/4000)$ both in the pictures of the tetraquark states and molecular states with the tentative assignments $J^{PC}=1^{+-}$ based on the rigorous quark-hadron duality. We write down the three-point correlation functions, and accomplish the operator product expansion up to the vacuum condensates of dimension-5, and neglect the tiny gluon condensate contributions, just like in our previous works, and obtain the spectral densities through dispersion relation, then we acquire the rigorous quark-hadron duality below the continuum thresholds $s_0$ and $u_0$ by accomplishing the integral over the variable $ds^\prime$ firstly, and obtain six QCD sum rules for the hadronic coupling constants. We investigate the two-body strong decays
of the axialvector tetraquark state and molecular state, respectively, and obtain the total widths
$\Gamma_{Z_{cs}}^T =15.31\pm 2.06\,\rm{MeV}$ and $\Gamma_{Z_{cs}}^M=83.51\pm21.09\,\rm{MeV}$,
which are consistent with the experimental data $13.8^{+8.1}_{-5.2}\pm4.9\,\rm{MeV}$ from the BESIII collaboration and
$131 \pm 15 \pm 26\,\rm{MeV}$ from the LHCb collaboration, respectively. The present calculations support assigning the $Z_{cs}(3985)$ and $Z_{cs}(4000)$ to be the hidden-charm tetraquark state and molecular state with the $J^{PC}=1^{+-}$, respectively.
\section*{Acknowledgements}
This work is supported by National Natural Science Foundation, Grant Number 12175068.
|
\section{Introduction}
The development and optimization of biopharmaceutical production processes with cell cultures is cost- and time-consuming, requiring substantial lab work. This necessitates thorough planning of experiments and processes, taking into account existing process knowledge.
The need of model-based decision support in biopharmaceutical manufacturing has been emphasized by the US Food and Drug Administration (FDA) \cite{Herwig.2015,U.S.DepartmentofHealthandHumanServicesFoodandDrugAdministration.}, including taking into account available prior know-how and experience within the decision process and uncertainties \cite{Sokolov.2020}.
Such methods are still not state of the art for cell culture processes during development or manufacturing \cite{Sokolov.2020, Xie.2019}, although first approaches have been proposed, for example for titer optimization of a mammalian cell culture process \cite{Liu.2017}.
This highlights a need for improved methods and tools for optimal experimental design, optimal and robust process design and process optimization for the purposes of monitoring and controlling during manufacturing.
Optimization for \emph{one} objective criterion (e.g.\ final titer) is relatively straight forward, i.e.\ building an objective function with a unique response variable and applying an appropriate optimization algorithm to maximize this function.
However, in industry, it is typically desired to optimize \emph{several} conflicting objectives at a time, leading to suitable trade-offs and compromises. For example when trying to maximize final titer via viable cell density while minimizing cultivation time.
Multi-objective optimization provides a decision-making tool for optimal decisions in the presence of trade-offs between two or more conflicting criteria.
However, multi-objective optimization is more challenging.
Its application is still not state of the art in the context of cell culture processes, probably due to a lack of related studies and instructions. Moreover, within the manufacturing life cycle of biopharmaceuticals, some phases are better investigated than others. Still very few investigations are reported concerning the cell expansion process (called seed train). It consists of several consecutive cultivation and passaging (transfer) steps, starting with a small amount of cell suspension because cells are frozen in small vials until they are used for a production process. The goal is to expand the number of viable cells in order to reach the required amount to inoculate (start) the production bioreactor (e.g.\ 10,000~L at industrial scale) while keeping them in a healthy and growing state. A high amount of operational requirements and constraints have to be fulfilled and, as reported in literature \cite{Le.2012,Bohl.2020}, the cell expansion process critically effects product quality and the amount of product at production scale. In \cite{Bohl.2020} for example, the passage duration as well as the initial viable cell density for each passage are reported as important parameters with high impact on process time and productivity at production scale. A careful and optimal planning of a seed train is therefore essential. However, this is not a trivial task due to the inherent variability concerning cell growth (cell growth differs from cell line to cell line and also from cultivation run to cultivation run) and uncertainty about the real state of the process due to considerable measurement uncertainties. This requires the design a reproducible process which is robust regarding viable cell density, meaning that despite (initial) variabilities concerning cell growth, low variations of viable cell density at the end of the seed train are obtained.
The goal of this paper is to close the gap between state of the art optimization techniques and modern techniques from machine learning to improve the biopharmaceutical production by allowing easy to use yet powerful multi-objective optimization.
In most multi-objective optimization problems no single best (unique optimal) solution exist, instead there is a set of optimal solutions (also called Pareto optimal solutions or non-dominated solutions), meaning for each solution that one criterion cannot be improved without degrading at least one of the other criteria. So, the decision maker has to choose from the set of non-dominated solutions according to the most preferred or important objective criterion.
A promising approach to optimize objective functions, which are expensive to evaluate, is Bayes optimization.
The methodology of Bayes optimization dates back to the work of Harold Kushner in 1964 \cite{Kushner.1964} and gained impact through the work of Jones et al.\ in 1998 \cite{Jones.1998}.
It is a probabilistic global optimization method for finding the maximum of expensive to evaluate objective functions or unknown (black-box) objective functions that are approximated using simulations \cite{Brochu.12.12.2010}.
In practice, the objective function could be the outcome of interest of a process, for example process productivity or control metrics to describe the quality of a product.
Input parameters can be process parameters needed to be optimized.
Bayesian optimization \cite{Shahriari.2016} creates a quick to evaluate model, the so-called \emph{surrogate model}, of the objective function. In order to reduce the objective function evaluations, the surrogate model is iteratively trained and updated on new data.
The positions of this new data are chosen by finding trade-off between exploration (improving the surrogate model) and exploitation (finding optimal points).
Typical surrogate models are Gaussian processes.
Gaussian processes (GP's) are popular machine learning models \cite{Rasmussen.2008}, because due to their Bayesian nature, they work well with few data points \cite{tulsyan2019industrial}.
Furthermore, they allow to include expert knowledge \cite{LH_AlgorithmicLinearlyConstrainedGaussianProcesses,LH_AlgorithmicLinearlyConstrainedGaussianProcessesBoundaryConditions} and can be used in dynamic systems \cite{Bradford.2020, Bradford.2018b}.
GP's are very flexible non-parametric models, hence they can approximate any function and do not assume a predefined set of modeling functions.
Bayes optimization is successfully applied in many fields of research and economics \cite{Petsagkourakis.04.06.2020}.
Also applications of Bayes optimization in the field of bioprocess engineering were published during the last decade \cite{Bradford.2018,Bradford.2018b,Clayton.2020,Narayanan.2022}.
Furthermore, this methodology was shown to be efficient in solving multi-objective optimization problems \cite{Yang.2019} and has also be applied for parameter estimation of kinetic parameters \cite{Manheim.2019}.
However, no applications are reported so far, applying model-based multi-objective Bayes optimization within biopharmaceutical process development.
This contribution aims to present the concept of a workflow which couples uncertainty-based upstream simulation and Bayes optimization using Gaussian processes and its application in form of a simulation case study to illustrate its applicability to a relevant industrial task in process development.\\
This simulation case study addresses the question if a reference seed train setup comprising 5 shake flask scales can be optimized through varying shake flask volumes and
how many shake flask scales, three, four or five, are recommendable in terms of two objective criteria, seed train duration and deviation rate.
Moreover it is investigated how the results change if cells grow with 5\% lower or 5\% higher maximum cell-specific growth rate. \\
Afterwards, two more objective criteria, titer (product concentration) and viability after 8 days in the production bioreactor, are added and seed train optimization is performed regarding four objective criteria simultaneously.\\
Furthermore, the suitability of the proposed method and the required number of iterations is evaluated with respect to the obtained information gain.
\section{Methods}
The main components of the applied methodology and the corresponding tools are described.
\subsection{Upstream simulation}\label{method: seed train simulation}
Upstream simulation comprises simulation of the cell expansion process (seed train) and simulation of the production scale.
The reference upstream process taken as an application example for the here presented simulation case study comprises five consecutive shake flask scales followed by three bioreactor scales and one production scale, similar to the upstream process investigated in \cite{HernandezRodriguez.2019}. Further specifications are listed in Table~\ref{tab: setup_bayes_opt_ex_2}. \\
A mathematical model is required, describing cell growth and interactions with the main limiting substrates and eventually inhibiting metabolites over time.
A cell growth model, a system of ordinary differential equations (ode), already adapted to an industrial cell culture upstream process using a CHO cell line \cite{HernandezRodriguez.2019} has been used, which describes the dynamic behavior of viable and total cell density, $X_{\rm v}$ and $X_{\rm t}$, concentrations of glucose $c_{\rm Glc}$, glutamine $c_{\rm Gln}$, lactate $c_{\rm Lac}$, ammonia $c_{\rm Amm}$ and product (volumetric titer) $c_{\rm titer}$ (see Table~\ref{tab: model_eq_novartis} in the supplementary material).
Moreover, such an upstream process includes several constraints, operation steps and process parameters (e.g.\ concerning passaging intervals, substrate/nutrient concentrations, initial viable cell densities and viable cell densities before transferring cells into the next cultivation vessel as well as the amount of cell suspension and fresh medium), which have to be considered in the simulation workflow.
A detailed description of the required components and calculation routines are described in \cite{HernandezRodriguez.2020b,HernandezRodriguez.2020}.
Besides these requirements, several passaging strategies can be applied, helping to decide at which point in time cells should be transferred from one cultivation vessel into the next larger one and how to perform these passaging steps (e.g.\ which amount of cell suspension should be mixed with how much fresh cell culture medium).
For the here presented simulation study, the passaging strategy for robust seed train design was chosen, where robustness refers to the reproducibility of the seed train regarding viable cell density, meaning that despite initial uncertainties and variabilities concerning cell growth, low variations of viable cell density at the end of the seed train are obtained. This strategy grounds on the objective of reaching the previously determined threshold of viable cell density and corresponding probability distributions of viable cell density at different points in time. These distributions are used in combination with a utility function following the mean-variance principle: The utility function $U(t)$ is defined as a function of viable cell density $X_{\rm v}$ including the expected value ${\rm E}(X_{\rm v})$ and the variance ${\rm Var}(X_{\rm v})$ of viable cell density as well as a risk aversion parameter $\alpha$ which controls the amount of risk (amount of uncertainty) the user is willing to bear. In the here presented example, the risk refers to the probability that viable cell density differs from the expected value (predicted mean). A risk aversion value of $\alpha\,=\,1$ would mean that the expected time profile minus one time the standard deviation of $X_{\rm v}$ is considered.
The utility function is defined through:
\begin{equation}\label{eq: utility}
U(t) = E(X_{\rm v}(t)) - \alpha \sqrt{{\rm Var}({X_{\rm v}(t)})}
\end{equation}
Based on the simulated time profiles of the current cultivation scale (by solving the corresponding ode system), Eq.~\ref{eq: utility} is used to calculate the utility function value $U(t)$ per hour and to check if this value reaches or exceeds the required transfer viable cell density $X_{\rm v,transfer}$ which is necessary to inoculate (start) the next cultivation scale fulfilling the required seeding (initial) viable cell density and the filling volume.
In a next step, it is evaluated if the calculated point in time lies within the range of practically feasible points in time for cell passaging, $T_{\rm p}$.
Thus, the objective is to find the minimum point in time out of the set of practically feasible points in time for passaging, $T_{\rm p}$, which fulfills
\begin{align}
&U(t) \ge X_{\rm v,transfer}, \\
&\mbox{subject to: \quad} t\in T_{\rm p}.
\end{align}
Based on the obtained point in time and the corresponding concentrations of viable cells, total cells, substrates and metabolites at this point in time, starting concentrations (=\,initial values of the system of ordinary differential equations) of the next cultivation scale are calculated based on the defined configurations and constraints (e.g.\ working volumes, acceptable range of seeding viable cell density and medium concentrations).
This calculation has to be performed for every cultivation scale and passaging step. For more details refer to \cite{HernandezRodriguez.2019, HernandezRodriguez.2020, HernandezRodriguez.2020b}.
{\small
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.0}
\caption[Specification of the exemplary seed train setup.]{Specification of the exemplary seed train setup providing information concerning cultivation vessels, required viable cell densities and the transfer of cells from one cultivation vessel into the next larger one, assumed in this work.}
\label{tab: setup_bayes_opt_ex_2}
\begin{tabular}{l|l}
\hline
Seed train setup & \\
\hline\\[-1.0em]
Flask scales: & 3, 4 or 5 flask scales between 0.014\,L and 8\,L filling volume \\[1.5mm]
Bioreactor scales: & 3 bioreactor scales, 38 L, 302 L and 2054 L filling volume \\[1.5mm]
Production bioreactor: & 9500 L filling volume \\[1.5mm]
Optimal range for & \\
viable seeding cell density: & $3\cdot\,10^{8}$ -- $3.5\cdot\,10^{8}$ cells L$^{-1}$ ($3\cdot\,10^{5}$ -- $3.5\cdot\,10^{5}$ cells mL$^{-1}$) \\ [1.5mm]
Optimal range for & \\
transfer viable cell density: & $0.1\cdot\,10^{10}$ -- $1\cdot\,10^{10}$ cells L$^{-1}$ ($0.1\cdot\,10^{7}$ -- $1\cdot\,10^{7}$ cells mL$^{-1}$ ) \\[1.5mm]
Target seeding (initial) & \\
viable cell density: & 3.15 $\cdot\,10^{8}$ cells L$^{-1}$ (3.15 $\cdot\,10^{5}$ cells~mL$^{-1}$) \\
& (= minimum viable seeding VCD + 5\%) \\
\hline \\[-1.0em]
Strategy concerning point & 'Xv transfer', i.e.\ passaging as soon as the calculated \\
in time for cell passaging: & required viable transfer cell density is reached \\[1.5mm]
Practically feasible & Passaging between 48 and 120 h possible \\
points in time for passaging: & (flexible ranges) \\[1.5mm]
Strategy concerning & Discard cell suspension during the passaging step, if\\
current and new volume: & required to start within an optimal seeding \\
& cell density range\\
\hline
\end{tabular}
\end{table}
}
\subsection{Bayes optimization}\label{method: Bayes_optimization}
A typical mathematical optimization problem is the following:
Given an objective function $f: \mathcal{X} \rightarrow \mathbb{R}$ over input space $\mathcal{X}\subseteq\mathbb{R}^d$, the aim is to find an argument $x^*\in\mathcal{X}$, which optimizes (minimizes or maximizes) $f$.
The idea behind Bayes Optimization consists in creating a simple, probabilistic and cheap to evaluate model, a so-called \emph{surrogate model} (substitute model), of the objective function $f$ \cite{Rasmussen.2008,Frazier.08.07.2018,Brochu.12.12.2010}.
Bayesian optimization reduces the number of evaluations of the objective $f$ via the following iterative approach:
Before sampling $f$ at another point, we take into account a trade-off between exploration (i.e.\ sampling of areas of high uncertainties) and exploitation (sampling from areas which are likely to move towards the optimum), which is encoded in a so-called acquisition function.
We can find such points quickly from evaluation of the surrogate model.
Within Bayes optimization the following steps are performed:
\begin{enumerate}
\item Generate a set of initial points and evaluate the objective function at these points.
\item Train the surrogate model based on all evaluated points.
\item Optimize the acquisition function, which determines the next candidate point $x_c$ to be evaluated.
\item Compute $f(x_c)$, the objective function $f$ at the candidate point $x_c$.
\item Repeat steps 2-4 for $N$ iterations
\end{enumerate}
The key of Bayesian optimization is not to rely on local approximations as many other optimization algorithms and instead to have a global viewpoint of also evaluating the function at unknown positions.
The acquisition function is used to propose the next candidate point to be evaluated based on specific criteria, for example the expected improvement of the optimization criteria, and on the reduction of predictive uncertainty.
As in the case of the kernels, there is also a wide variety of possible acquisition functions to choose from. In this study, the \emph{Expected Improvement} (EI) acquisition function is used \cite{Couckuyt.2014, AntonSekulic.2020}.
Gaussian processes (GPs) are well suited surrogate models when making few assumptions \cite{Brochu.12.12.2010}.
Just like a Gaussian distribution (a normal probability distribution) is fully described by its mean $m$ and variance $\sigma^2$, a GP is fully
described by a mean function $m(x)$ and a covariance function $k(x;x^{'})$ \cite{Rasmussen.2008}.
A GP is an extension of a multivariate Gaussian (or normal) distribution to distributions of functions in the sense that if a function $y$ follows a GP distribution, i.e.\ $y\sim \mathcal{GP}\left( m, k \right)$, then every evaluation of the function follows a Gaussian distribution $y(x) \sim \mathcal{N}\left( m(x), k(x,x) \right)$.
In particular, a GP returns mean and variance of the possible function values (instead of just returning a scalar), and hence also provides information about the uncertainty of a prediction.
Moreover, GPs can take into account uncertainty in form of noise, the class of Gaussian processes is closed under Bayesian updates, and such updates are computationally tractable \cite{Snoek.13.06.2012}.
The covariance function describes the assumed characteristics such as smoothness or periodicity of the objective function $f$ \cite{Shahriari.2016}.
They are so-called positive-definite functions, often also called kernels \cite{Rasmussen.2008, Duvenaud.2014}.
It specifies the relationship between two 'points' (vector of the input space) $x$ and $x'$ and the corresponding changes in $f$ at these points. A covariance function is described by a set of parameters, also called hyperparameters, describing a specific behavior. This is how prior information is embedded in the Bayes optimization procedure.
Also in this work, the most commonly used covariance function, the \emph{Squared Exponential} (SE) kernel (often also referred to as Gaussian kernel) is used. \cite{Frazier.08.07.2018}.
\subsection{Problem definition and computational procedure}
The goal of the presented application example is to propose a concept and a numerical procedure for optimal robust seed train design, where robustness refers to the reproducibility of the seed train regarding viable cell density, meaning that despite initial uncertainties and variabilities concerning cell growth, low variations of viable cell density at the end of the seed train are obtained.
First, seed train constraints are defined based on a chosen cell line and its characteristics concerning optimal cultivation conditions and based on the operative possibilities (e.g.\ feasible points in time for cell passaging).
Second, the optimizable input parameters and objective criteria (objective response variables) applied in this study are defined (as also illustrated in Figure~\ref{fig: motivation}), followed by the formulation of the mathematical optimization problem.
Thereafter, the optimization problem is solved using a workflow which connects seed train simulation and Bayes optimization. \\
\begin{figure}[h]
\centering
\includegraphics[width=0.9\linewidth]{figures/Fig_motivation.png}
\caption{Goal of the study is to propose a concept and a numerical framework for optimal robust seed train design (blue box in the middle), including optimizable inputs (first gray box) as well as objectives (objective criteria) used in this study (right gray box).}
\label{fig: motivation}
\end{figure}
The following objective criteria were chosen to represent an optimal seed train: I) a minimum duration ($d$) (=\,required cultivation time) of the seed train and II) a minimum deviation rate ($D$) regarding viable cell density, i.e.\ the probability that the seed train will run outside predefined ranges of viable cell density (for both, seeding viable cell density and transfer viable cell density)\footnote{This is important to consider because in case that specific constraints are not fulfilled, the performance of the cells could decrease. The growth rate could decrease and furthermore it has been observed that the violation of constraints could also cause less viability of the cells in the production phase \cite{Bohl.2020}.
} (compare to Figure~\ref{fig: motivation} right gray box). These two attributes shall enable an optimal start of the production scale. Note that, in addition to these criteria, the growth rate is another important parameter affecting an optimal start of the production scale and the growth rate should be high until the end of the seed train. However, in this first optimization study it is not set as optimization criterion because the here defined seed train setup (in terms of medium concentrations and possible cultivation volumes per scale) together with the aim to reduce cultivation time already supports a good growth during the entire cultivation. But for other seed train setups it might be advisable to include growth rate at the end of the seed train into the optimization problem.\\
After consideration of the two mentioned objective criteria, a third and fourth objective criterion, the product concentration (titer) and the viability at the end of the cultivation in the production bioreactor (in this simulation study: after 8 days in batch mode, i.e.\ without addition of nutrient feeds) are added to the optimization problem (see Figure~\ref{fig: motivation} right gray box (III)).
{\em Note: The authors are aware of the fact that cultivation in the production vessel itself, which is often performed in fed-batch mode, is also influenced by several process parameters having an impact on product quantity and quality. Moreover, data of further attributes would be necessary to describe product quality (e.g.\ of a recombinant therapeutic protein or antibody) but these are not provided and therewith not considered in this study}.\\
The input variables that can be varied to optimize the recently mentioned objective criteria, and thus the optimizable input variables, are the filling volumes in the first five shake flask scales, $V_1,...,V_5$ (compare to Figure~\ref{fig: motivation}, the part of the seed train between thawing cells from a small vial and inoculation of the first biorector). These target values are important inputs of the seed train simulation process because they are used to calculate points in time for cell passaging. Volumes in the finally proposed seed train protocol (output of the seed train simulation) may vary within allowed working volume ranges and these are also presented in this work.
\paragraph{Formulation of the mathematical optimization problem} \,
The optimizable variables and therewith inputs of the optimization problem are the filling volumes of the $n$ shake flask scales, $V_1,...,V_n$ which are included in the input vector
\begin{align}
x = (V_1,...,V_n)^T.
\end{align}
Outputs of the optimization problem are the defined objective criteria. These are seed train duration $d$ and deviation rate $D$ for the first optimization example. Thus, the unknown objective function (which should be minimized) can be written as follows:
\begin{align}
f(x) = (f_1(x),f_2(x))^T
\end{align}
with $f_1(x)\,\hat{=}\,d$ and $f_2(x)\,\hat{=}\,D$.
The second optimization example includes a third and fourth optimization criterion, product concentration and viability at the end of the production scale (here after 8 days in the production vessel). Thus $f(x)$ expands to
\begin{align}
f(x) = (f_1(x),f_2(x),f_3(x),f_4(x))^T
\end{align}
with $f_1(x)\,\hat{=}\,d$, $f_2(x)\,\hat{=}\,D$, $f_3(x)\,\hat{=}\,c_{\rm titer,end}$ and $f_4(x)\,\hat{=}\,\mbox{Viability}_{\rm end}$.
\subsection{Connecting seed train simulation and Bayes Optimization}
Uncertainty-based seed train simulation as described in Section~\ref{method: seed train simulation} was coupled with algorithms for Bayes optimization as described in Section ~\ref{method: Bayes_optimization}. The workflow integrating both components is illustrated in Figure~\ref{fig: scheme_bayes_opt}.
Input of the combined framework are the input variables:
Boundaries for the optimizable variables (here filling volumes) and objective criteria (here seed train duration, deviation rate and in the second example also product concentration at the end of production scale) given all required seed train configuration settings and constraints (e.g.\ initial concentrations, practically feasible points in time for cell passaging, acceptable ranges for viable cell density, ...).
First points (=\,combinations of optimizable variables) are determined using a Latin Hypercube design distributing these points within the design space (see Figure \ref{fig: scheme_bayes_opt}, Box A). Seed train simulations are performed at these points in order to obtain the corresponding objective criteria values. Input values together with output values form a data set.
An unknown model describing the relationship between inputs and outputs is approximated through a Gaussian process (GP) which has to be trained (see Figure \ref{fig: scheme_bayes_opt}, Box A) based on the given data set.
Therefore, the Gaussian process proposes a point that has to be evaluated next (see Figure \ref{fig: scheme_bayes_opt}, Box B).
A robust seed train is simulated, using a mechanistic process model, and the objective criteria are calculated. This output is then returned to the Bayes optimization (Box A) to update the GP.
Usually, experiments are performed to return the experimental output. The present approach instead exploits the advantages of the model-based upstream simulation in order to reduce the experimental effort to a minimum.
These steps are repeated various times, e.g.\ until a previously defined number of maximum iteration steps is reached. The latter depends on the resources (human and financial resources in case of laboratory experiments or computational resources in case of in-silico experiments).
In every iteration the Gaussian process chooses a new point aiming to move to the optimum and at the same time to reduce model uncertainty.
Results of this optimization framework are the set of Pareto optimal setups (also called Pareto front) and their corresponding response values.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{figures/Fig_Sim_and_Bayes_opt.jpg}
\caption{Scheme showing the applied computational workflow comprising {\bf A} a Bayes optimization algorithm which is coupled with {\bf B} a seed train simulation routine. Input and output values are shown in the blue boxes above and below.}
\label{fig: scheme_bayes_opt}
\end{figure}
\subsection{Numerical solvers and tools}\label{subsec: system}
The programming language and numeric computing environment MATLAB \cite{MATLAB.2019} was used for the seed train simulations. The code for the optimization workflow was written in Python \cite{vanRossum.2010} using the MATLAB Engine API for Python to call MATLAB as a computational engine from Python code. To perform Bayes optimization within this workflow, the library GPflow \cite{Knudde.10.11.2017} was used.
\section{Results \& Discussion}
\subsection{Optimization of cultivation vessels regarding number of shake flask scales and filling volumes for 5, 4 and 3 shake flask scales}
In this section, it is investigated which cultivation filling volumes should be used for the flask scales in order to obtain optimal results in terms of seed train duration and deviation rate, here defined as the probability that the seed train will run outside the predefined acceptable ranges for initial viable cell density (VCD) and transfer VCD (final VCD before transfer into the next cultivation vessel) per scale. The latter is a measure for the robustness of the seed train regarding viable cell density.\\
For assessment of the optimization results, a conventional reference seed train comprising 5 shake flask scales was simulated based on a non-optimized design. Therefore, a common passaging interval of 3 days per cultivation scale was fixed and filling volumes were determined following a conservative layout (i.e.\ choosing not too huge differences between one cultivation scale and the next to ensure that enough viable cells are generated even if they grow a little bit slower than expected). \\
In a first step, the optimal combination of filling volumes for 5 shake flask scales is investigated and the results are compared to the reference seed train.
Afterwards, it is investigated if a reduction of shake flask scales from 5 to 4 or 3 shake flask scales leads to similar or even better results in terms of seed train duration and deviation rate.
The number of bioreactor scales was kept fixed. Three bioreactors with filling volumes of 40 L, 320 L and 2100 L were used as pre-stages before inoculation of the production bioreactor with 9600 L.
The assumed seed train setup is given in Table \ref{tab: setup_bayes_opt_ex_2}.
\noindent To find the optimal solution, multi-objective Bayesian optimization coupled with uncertainty-based seed train simulation, as described in Section~\ref{method: seed train simulation}, was applied.
First, a Latin hypercube design for $n_{\rm lhs}$ design 'points' (combinations of filling volumes, here $n_{\rm lhs}=10$) was initiated and seed train simulation was applied to calculate the objective criteria values, here deviation rate $D$ and seed train duration $d$ (replacing the normally required experimental cultivation runs) at each point.
Within the Bayes optimization procedure, Gaussian processes were trained based on the simulation outcomes and an acquisition function was calculated in each iteration step in order to propose which point should be evaluated next.
The input space for shake flask filling volumes (here the optimizable variables) was defined as described in Table~\ref{tab: input_space}, assuming the possibility of using several shake flasks in parallel for one shake flask scale and also considering their working volumes ranges.
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.3}
\caption[Input space]
{Input space for the shake flask filling volumes, containing the possible filling volumes per scale, given for optimization runs with 5, 4 or 3 shake flask scales.}
\label{tab: input_space}
\begin{tabular}{l|lll}
\hline
& \multicolumn{3}{c}{Filling volumes} \\
& Range for & Range for & Range for \\
& 5 shake flask scale [L] & 4 shake flask scales [L] & 3 shake flask scales [L] \\
\hline
V1 & 0.014 -- 0.015 & 0.014 -- 0.015 & 0.014 -- 0.015 \\
V2 & 0.05 -- 0.15 & 0.1 -- 1 & 0.1 -- 2 \\
V3 & 0.15 -- 1.5 & 1.5 -- 4 & 4 -- 8 \\
V4 & 1.5 -- 4 & 4 -- 8 & - \\
V5 & 4 -- 8 & -- & - \\
\hline
\end{tabular}
\end{table}
\subsubsection{Optimization of 5 shake flask scales}\label{res: 5_sf} \, \\
The first optimization was performed for a seed train comprising 5 shake flask scales.
Figure~\ref{fig: pareto_front_sf5_mu1} shows the objective criteria values for each evaluated point, whereby the outcomes based on the initial Latin hypercube space are illustrated by blue dots and the outcomes for the proposed points based on the trained Gaussian processes are illustrated through yellow crosses. The optimal solutions are those near to the lower left corner aiming to minimize seed train duration and the deviation rate. The Pareto optimal solutions, also called non-dominated solutions are illustrated through green circles.
A solution (here seed train setup/combination of filling volumes) is called non-dominated if no solution exists leading to better (here lower) objective criteria values. As described previously, several Pareto optimal solutions can be obtained because when considering two or more objective criteria then for two different solutions one criterion might have better (here lower) value then the other solution for the same objective, while the other criterion has worse (here higher) values. The set of all Pareto optimal solutions is called Pareto-front.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\linewidth]{figures/Fig_pareto_front_5sf_mu100_2obj_lhs10_it20_appendix.png}
\caption[Pareto front showing the objective criterion seed train duration over the objective criterion deviation rate; optimizable variables are the filling volumes of the 5 shake flask scales (blue dots: based on the initial Latin hypercube design; yellow crosses: based on the proposed points; green circles: Pareto optimal solutions).]{Objective criterion {\em seed train duration} over objective criterion {\em deviation rate} obtained for evaluated points (optimizable variables, here combinations of 5 shake flask filling volumes); blue dots: based on the initial Latin hypercube design (LHC); yellow crosses: based on the points proposed by the algorithm; green circles: Pareto optimal solutions (=\,Pareto front)).}
\label{fig: pareto_front_sf5_mu1}
\end{figure}
\noindent For the investigated scenario (5 shake flask scales and the seed train configuration according to Table~\ref{tab: setup_bayes_opt_ex_2}) five Pareto optimal solutions were obtained (see green circles in Figure~\ref{fig: pareto_front_sf5_mu1}). It can be seen that comparing two of these solutions (green circles) each, one solution has a lower (here better) seed train duration value than the other solution and the opposite holds for the deviation rate.\\
The corresponding values for the optimizable variables, here shake flask filling volumes ($V_1$, $V_2$, $V_3$, $V_4$ and $V_5$), and the corresponding objective criteria values, here deviation rate $D$ and seed train duration $d$, are listed in Table~\ref{tab: bayes_opt_5sf_volumes}.
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.3}
\caption[Pareto optimal solutions concerning the choice of filling volumes in shake flask scales, for three scenarios for 5 flask scales. The following bioreactor filling volumes are 40\,L, 320\,L and 2210\,L. The averaged filling volumes in L and the resulting deviation rate ($D$) in \% and seed train duration $d$ in h are listed for each solution.]
{Pareto optimal solutions concerning the choice of filling volumes in shake flask scales, for three scenarios for 5 flask scales. The following bioreactor filling volumes are 40\,L, 320\,L and 2210\,L. The averaged filling volumes in L and the resulting deviation rate ($D$) in \% and seed train duration ($d$) in h are listed for each solution.}
\label{tab: bayes_opt_5sf_volumes}
\begin{tabular}{l|lllll|ll}
\hline
& \multicolumn{5}{c}{Filling volumes} & & \\
Solution & Vol. 1 & Vol. 2 & Vol. 3 & Vol. 4 & Vol. 5 & D & d \\
& [L] & [L] & [L] & [L] & [L] & [\%] & [h] \\
\hline
1 & 0.015 & 0.065 & 0.904 & 2.355 & 7.78 & 537 & 4.9 \\
2 & 0.015 & 0.115 & 0.451 & 1.672 & 7.89 & 521 & 6.1 \\
3 & 0.015 & 0.104 & 0.340 & 1.614 & 6.85 & 524 & 5.2 \\
4 & 0.014 & 0.103 & 0.369 & 1.582 & 7.87 & 523 & 5.3 \\
5 & 0.015 & 0.114 & 0.431 & 2.026 & 7.97 & 520 & 6.7 \\
\hline
& \multicolumn{5}{c}{Filling volumes} of reference seed train & & \\
Reference & 0.015 & 0.08 & 0.30 & 2 & 4 & 41.7 & 576 \\
\end{tabular}
\end{table}
\noindent The filling volume of the first scale was limited to a very narrow range (14--15 mL) (A higher variation after cell thawing was not expected).
Most obtained solutions start with the maximum value of this range (see Table~\ref{tab: bayes_opt_5sf_volumes}, first column).
The filling volume of flask scale 2 varies between 0.065 and 0.115 L, the filling volume of flask scale 3 between 0.340 and 0.904~L, the filling volume of flask scale 4 between 1.582 and 2.355 L and of flask scale 5 between 6.85 and 7.97~L. All five combinations lead to a deviation rate $D$ of less than 7\% and to a seed train duration between 520 to 537~h.
A more detailed illustration of the obtained results is presented in Figure~\ref{fig: contour_plots_D_5sf_mu1} and Figure~\ref{fig: contour_plots_duration_5sf_mu1}. For two optimizable variables and one objective criterion each (deviation rate in Figure~\ref{fig: contour_plots_D_5sf_mu1} and seed train duration in Figure~\ref{fig: contour_plots_duration_5sf_mu1}), a contour plot is shown which illustrates the objective value for each calculated point (combination of the two variables), using the trained Gaussian processes, through colored isolines. \\
For example, the diagram in the top left of Figure~\ref{fig: contour_plots_D_5sf_mu1} shows the deviation rate for each combination of $V_1$ (filling volume in flask scale~1) and $V_2$ (filling volume in flask scale~2) through colors representing the corresponding values in \%, as indicated in the color bar. The results obtained through seed train simulations are shown by dots. The red dots represent the non-dominated (optimal solutions), optimal with respect to the defined multi-objective optimization problem. The dark blue area indicates combinations of $V_1$ and $V_2$ leading to a lower deviation rate. It can be seen that values above 0.1 for $V_2$ combined with any value of $V_1$ (within the given range) lead to the lowest deviation rates (below 6.2\%, see dark blue area).
Also the optimal solutions (red dots) are mostly located in the area with higher filling volumes for shake flask 2, $V_2$, except one (red dot at $V_2\approx$\,0.065).
For some combinations, a closer delimitation is possible. For example, the middle diagram in the second row ($V_3$ over $V_2$) shows a limited region (dark blue area) and therewith a specific combination of $V_3$ and $V_2$ that leads to the lowest deviation rates (\textless~5.6\%)). These are around 0.3~L for $V_3$ and around 0.105~L for $V_2$.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{figures/deviation_rates.png}
\caption{Contour plots showing two optimizable variables on x and y-axis and one objective (here Deviation rate $D$ in \%), assigned to each combination of the two variables, through colored isolines. For example, the diagram top left shows the deviation rate for each combination of $V_1$ (filling volume in flask scale 1) and $V_2$ (filling volume in flask scale 2) through colors representing the corresponding values in \%, as indicated on the color bar. Moreover, the results obtained through seed train simulations are shown by dots. The red dots represent the non-dominated (optimal solutions).}
\label{fig: contour_plots_D_5sf_mu1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{figures/duration.png}
\caption{Contour plots showing two optimizable variables on x and y-axis and one objective (here seed train duration ($d$) in h), assigned to each combination of the two variables, through colored isolines. For example, the diagram top left shows the deviation rate for each combination of $V_1$ (filling volume in flask scale 1) and $V_2$ (filling volume in flask scale 2) through colors representing the corresponding values in \%, as indicated on the color bar. Moreover, the results obtained through seed train simulations are shown by dots. The red dots represent the non-dominated (optimal solutions)}
\label{fig: contour_plots_duration_5sf_mu1}
\end{figure}
Analogously, Figure~\ref{fig: contour_plots_duration_5sf_mu1} shows the contour plots for the second objective criterion, seed train duration. The dark blue areas show the combinations with the lowest seed train durations (approximately below 528\,h).
It can be seen in these diagrams that most red dots are located in the dark blue regions.
For some combinations the dark blue areas are wider, distributed over several possible values for one variable, e.g.\ the diagram in the top center, top left, center, and center right. \\
Other combinations show narrower regions with low seed train durations as can be seen in the diagram showing $V_4$ over $V_3$.
The lowest seed train duration is obtained for filling volumes between 1.5 and 2.5~L for shake flask 4 in combination with filling volumes between 0.2 and 0.8~L for shake flask 3. \\
Overall, these diagrams give an overview of the impact of two combined optimizable variables each on a specific objective criterion.
In addition to this information, simulated time profiles (predictive mean in green, 90\% prediction bands in blue) of viable and total cell density as well as concentrations of glucose, glutamine, lactate and ammonium (see Figure~\ref{fig: seed train sol1 5sf}) can be obtained for each solution as well as a seed train protocol containing information about the calculated passaging intervals, amount of medium, etc.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{figures/seed_with_ranges_sol_5.png}
\caption{Seed train showing viable cell density (VCD) and total cell density as well as substrate (glucose and glutamine) and metabolite (lactate and ammonium) concentrations over time and over the whole seed train (5 shake flask scales and three bioreactor scales), based on the shake flask filling volumes according to solution 1. The green lines represent the mean time course and the blue lines show the corresponding 90\%-prediction band (5\%- and 95\%-quantiles).
The plot (top left) also inlcudes the filling volumes and the acceptable ranges for seeding VCD and transfer VCD, illustrated through dashed lines.}
\label{fig: seed train sol1 5sf}
\end{figure}
It can be seen in Figure~\ref{fig: seed train sol1 5sf} top left that based on the given filling volumes in addition to the flexibility to choose individual points in time for cell passaging in each scale, it is possible to set the seeding viable cell density at the beginning of each cultivation scale on the desired value with low variability, allowing to stay within the corresponding acceptable ranges for seeding VCD (see yellow dashed lines). Also transfer VCDs lie within the corresponding acceptable range with high probability (see lower boundary, gray dashed line). Moreover, it can be seen that substrate concentrations are not depleted and according to \cite{HernandezRodriguez.2019}, values of 20 mmol/L lactate and 5 mmol/L ammonium are not yet inhibiting concentrations for this cell line.
For a better assessment, the obtained results are compared to the reference seed train which is also defined in this work for 5 shake flask scales and illustrated in Figure~\ref{fig: ref_seed train 5sf}. It grounds on a (non-optimized) configuration setup for 5 shake flask scales using fixed passaging intervals of 72\,h each (common practice) and filling volumes of 15\,mL (flask scale 1), 80\,mL (flask scale 2), 300\,mL (flask scale 3), 2,000\,mL (flask scale 4) and 4,000\,mL (flask scale 5). This choice grounds on a rather conservative approach aiming to avoid the risk of reaching too low transfer cell densities at the end of a cultivation scale but without inclusion of probabilistic simulations. \\
The proposed method instead includes risk calculations and a passaging strategy aiming to minimize this risk but at the same time identifying a seed train configuration which is optimal regarding further objectives like seed train duration in the present case.
A comparison of the seed train solutions obtained after optimization and the reference seed train shows that deviation rate is much lower after optimization (4.9--6.7\% instead of 41.7\%) and seed train duration could be reduced by 56 h from 576 h to 520 h.
Figure~\ref{fig: ref_seed train 5sf}, diagram top left shows where seeding or transfer viable cell density do not lie fully within the acceptable ranges (see red circles). This is different for the optimized solutions, e.g solution 5, as illustrated in Figure~\ref{fig: seed train sol1 5sf}, where seeding VCD lies within the acceptable range and also transfer VCD lies above the lower bound of the acceptable range for transfer VCD.
This significant reduction in time ($\approx$\,2 days per seed train) would contribute to a meaningful acceleration of the production process.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{figures/seed_with_ranges_reference.png}
\caption{Reference (non-optimized) seed train showing viable and total cell density as well as substrate (glucose and glutamine) and metabolite (lactate and ammonium) concentrations over time, based on a reference configuration setup for 5 shake flask scales using passaging intervals of 72\,h each. The green lines represent the mean time course and the blue lines show the corresponding 90\%-prediction band (5\%- and 95\%-quantiles).}
\label{fig: ref_seed train 5sf}
\end{figure}
\subsubsection{Optimization of 3 and 4 shake flask scales}\label{res: 3_4_sf} \, \\
In a next step, the number of shake flask scales was reduced from 5 to 4 and then to 3 shake flask scales and the same optimization procedure was applied. The aim was to investigate if less cultivation vessels would lead to comparable results and if so, which target and filling volumes should be chosen. This is of interest because less operations (like transferring cells from one scale into another one) signify less risk of failure and deviations.
Figure~\ref{fig: bayes_opt_3_4_sf} shows the obtained values for the objective criteria deviation rate and seed train duration for different combinations of filling volumes for 3 (left) and for 4 shake flask scales (right). Also here, the solutions based on the initial Latin hypercube design are shown by blue dots and Pareto optimal solutions are highlighted through green circles.
It can be seen that for both scenarios, combinations of filling volumes could be found leading to an overall seed train cultivation time between 519 and 530~h. However, the scenario of using 4 shake flask scales, leads to lower deviation rates ($D\,<\,10\%$) compared to the scenario of using 3 shake flask scales ($23\% < D < 26\%$).
\begin{figure}[h]
\begin{minipage}{0.5\textwidth}
\includegraphics[width=1\linewidth]{"figures/Fig_pareto_front_3sf_mu100_2obj_lhs10_it20_adapted5_scales.png"}
\end{minipage}
\begin{minipage}{0.5\textwidth}
\includegraphics[width=1\linewidth]{"figures/Fig_pareto_front_4sf_100_2obj_lhs10_it20_adapted5_scales.png"}
\end{minipage}
\caption[Pareto optimal solutions regarding seed train duration and deviation rate for 3 and 4 shake flask scales]{Objective criterion {\em seed train duration} over objective criterion {\em deviation rate} obtained for evaluated points (optimizable variables), here combinations of for 3 (left) and 4 shake flask scales (right); blue dots: based on the initial Latin hypercube design (LHC); yellow crosses: based on the points proposed by the algorithm; green circles: Pareto optimal solutions (=\,Pareto front).}
\label{fig: bayes_opt_3_4_sf}
\end{figure}
\noindent The corresponding filling volumes and the obtained filling volumes (based on the underlying passaging strategy) of the Pareto optimal solutions are listed in Table~\ref{tab: bayes_opt_3_4sf_volumes} together with the results for 5 shake flask scales from Table~\ref{tab: bayes_opt_5sf_volumes}. The results are sorted as discovered by the optimization algorithm. The obtained filling volumes for 4 shake flask scales are very similar, except for shake flask scale 4 ( $V_1$ = 15\,mL, $V_2$ = 158--200\,mL, $V_3$\,=\,1.51--1.60\,L and $V_4$\,=\,4.81--7.58\,L). Some of the obtained solutions would be seen or treated as equal in practice, because the differences are rather small. For example it would not be distinguished between 0.190 and 0.195~L. Probably 200~mL would be used instead.
However, the applied optimization algorithm works on a continuous input space and differentiates between the solutions listed in the Table~\ref{tab: bayes_opt_3_4sf_volumes}, even though the differences are very low.
The obtained optimal filling volumes for 3 shake flask scales also look similar, but with a bit more variation for shake flask 3 (4.45--5.58\,L).
Comparing the results for the three scenarios (3, 4 and 5 shake flask scales) endorse a decision against the 3 flask scales-scenario due to the higher deviation rates (>\,20\%), which stands for less process robustness. Between the other two scenarios (4 or 5 shake flask scales) only little differences with respect to deviation rates are observed for the determined optimal solutions (4.9\%--6.7\% for 5 shake flasks, 5.9\%--9.2\% for 4 shake flasks). Using 5 shake flask scales would lead to more or less similar
cultivations times (520--537~h) but one operational step more would be required.
This information together with the corresponding seed train protocol provides a solid basis to take a decision for one of the proposed optimal seed trains designs, taking into account seed train duration, robustness (expressed through deviation rates) and operational steps.
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.3}
\caption[Pareto optimal solutions concerning the choice of filling volumes in shake flask scales, for three scenarios for 5 flask scales. The following bioreactor filling volumes are 40\,L, 320\,L and 2210\,L. The averaged filling volumes in L and the resulting deviation rate ($D$) in \% and seed train duration $d$ in h are listed for each solution.]
{Pareto optimal solutions concerning the choice of filling volumes in shake flask scales, for 3, 4 and 5 shake flask scales. The following bioreactor filling volumes are 40\,L, 320\,L and 2210\,L. The averaged filling volumes in L, the resulting deviation rate ($D$) in \% and seed train duration $d$ in h are listed for each solution.}
\label{tab: bayes_opt_3_4sf_volumes}
\begin{tabular}{l|lllll|ll}
\hline
& \multicolumn{4}{c}{Filling volumes} & & &\\
Solution & Vol. 1 & Vol. 2 & Vol. 3 & Vol. 4 & Vol. 5 & D & d \\
& [L] & [L] & [L] & [L] & [L] & [\%] & [h] \\
\hline
& \multicolumn{3}{l}{5 flask scales} & & & &\\
1 & 0.015 & 0.065 & 0.904 & 2.355 & 7.78 & 4.9 & 537 \\
2 & 0.015 & 0.115 & 0.451 & 1.672 & 7.89 & 6.1 & 521 \\
3 & 0.015 & 0.104 & 0.340 & 1.614 & 6.85 & 5.2 & 524 \\
4 & 0.014 & 0.103 & 0.369 & 1.582 & 7.87 & 5.3 & 523 \\
5 & 0.0015 & 0.114 & 0.431 & 2.026 & 7.97 & 6.7 & 520 \\
\hline
& \multicolumn{3}{l}{4 flask scales} & & & &\\
1 & 0.015 & 0.195 & 1.60 & 7.58 & & 9.2 & 520 \\
2 & 0.015 & 0.190 & 1.51 & 5.39 & & 8.0 & 521 \\
3 & 0.015 & 0.169 & 1.52 & 6.33 & & 7.5 & 522 \\
4 & 0.015 & 0.158 & 1.59 & 4.81 & & 5.9 & 528 \\
\hline
& \multicolumn{3}{l}{3 flask scales} & & & &\\
1 & 0.015 & 0.733 & 4.45 & & & 23.0 & 522 \\
2 & 0.015 & 1.046 & 4.85 & & & 23.8 & 521 \\
3 & 0.015 & 1.103 & 5.26 & & & 24.8 & 520 \\
4 & 0.015 & 0.934 & 4.77 & & & 23.0 & 522 \\
5 & 0.015 & 1.110 & 4.65 & & & 22.2 & 523 \\
6 & 0.015 & 1.306 & 5.58 & & & 26.4 & 519 \\
\hline
\end{tabular}
\end{table}
\clearpage
\subsection{Application to further cell lines with potentially different growth rates}\label{res: variation_of_mu}
The optimization examples presented in the previous subsection were applied to a specific CHO cell line with growth characteristics described by a set of model parameters derived from an industrial cell culture process which was investigated in \cite{HernandezRodriguez.2019}.
If a different cell line or a clonal cell population with potentially differing growth behavior is used, then the optimization has to be performed for this specific cell line. In the following simulation study, a cell line having a 5\% lower and a cell line having a 5\% higher maximum cell-specific growth rate compared to the reference maximum growth rate ($\mu_{\rm max}$\,=\,0.028\,h$^{-1}$ for the first bioreactor scale and $\mu_{\rm max}$\,=\,0.029\,h$^{-1}$ for the remaining seed train scales) are assumed and the optimization is applied for both scenarios. \\
The results for the obtained/proposed filling volumes as well as the corresponding seed train duration and deviation rate are listed in Table~\ref{tab: bayes_opt_different_growth_rates}.
As expected, cells which grow faster (higher maximum growth rate $\mu_{\rm max}$) would require less time until reaching a specific target cell density. This can be seen in the right column of Table~\ref{tab: bayes_opt_different_growth_rates}. Using 5 flask scales, the optimal required seed train duration would lie between 494 and 503 h for a cell line with a 5\% higher growth rate compared to the reference cell line which would need 520--537\,h (see Table~\ref{tab: bayes_opt_5sf_volumes}). Correspondingly, cells with a 5\% lower growth rate would need more time (550--568 h). The same is observed when using 4 or 3 shake flasks. \\
With respect to the deviation rates which represent the robustness of the seed train design regarding variability of viable cells, it can be seen that low deviation rates of between 4.1\% and 11.6\% can be reached when using 5 or 4 flask scales, even if the maximum growth rate varies $\pm$ 5\%.
A critical limit was identified for the combination of using 3 shake flask scales for a slower growing cell line. The corresponding optimal solution shows a comparatively higher deviation rate (19.2--29.1\%) together with a high seed train duration (548--552\,h).
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.3}
\caption[]
{Pareto optimal solutions concerning the choice of filling volumes in shake flask scales, for 3, 4 and 5 shake flask scales, for two different scenarios. Scenario 1 assumes a 5\% lower and scenario 2 a 5\% higher cell-specific maximum growth rate compared to the reference maximum growth rate. The bioreactor filling volumes which follow after the shake flask scales are 40\,L, 320\,L and 2210\,L. The filling volumes in L, the resulting deviation rate ($D$) in \% and seed train duration ($d$) in h are listed for each solution (several Pareto optimal solutions can be obtained per setup).}
\small
\label{tab: bayes_opt_different_growth_rates}
\begin{tabular}{l|lllll|ll}
\hline
& \multicolumn{5}{c}{Filling volumes} & & \\
Solution & Vol. 1 & Vol. 2 & Vol. 3 & Vol. 4 & Vol. 5 & D & d \\
& [L] & [L] & [L] & [L] & [L] & [\%] & [h] \\
\hline
{\bf 5 flask scales} & \multicolumn{3}{l}{5 flask scales} & & & & \\
5\% lower growth rate & & & & & & & \\
1 & 0.015 & 0.083 & 0.45 & 2.06 & 6.67 & 7.3 & 550 \\
2 & 0.014 & 0.072 & 0.30 & 2.78 & 6.23 & 7.0 & 551 \\
3 & 0.014 & 0.105 & 0.45 & 2.53 & 6.77 & 6.1 & 553 \\
4 & 0.015 & 0.119 & 0.84 & 2.96 & 6.21 & 5.5 & 568 \\
5 & 0.014 & 0.122 & 0.56 & 2.64 & 6.62 & 5.8 & 559 \\
5\% higher growth rate & & & & & & & \\
6 & 0.015 & 0.08 & 0.354 & 1.52 & 7.24 & 6.2 & 495 \\
7 & 0.014 & 0.09 & 0.560 & 1.89 & 7.19 & 4.6 & 502 \\
8 & 0.014 & 0.14 & 0.545 & 1.89 & 7.16 & 4.1 & 503 \\
9 & 0.015 & 0.06 & 0.313 & 1.63 & 7.72 & 6.5 & 494 \\
10 & 0.014 & 0.09 & 0.312 & 1.82 & 7.08 & 5.3 & 496 \\
11 & 0.014 & 0.13 & 0.564 & 2.16 & 7.41 & 4.8 & 501 \\
\hline
{\bf 4 flask scales} & \multicolumn{3}{l}{4 flask scales} & & & & \\
5\% lower growth rate & & & & & & & \\
12 & 0.015 & 0.158 & 1.99 & 7.6 & & 8.8 & 551 \\
13 & 0.015 & 0.147 & 1.59 & 7.7 & & 7.7 & 552 \\
14 & 0.015 & 0.132 & 2.00 & 7.8 & & 7.6 & 553 \\
15 & 0.015 & 0.167 & 1.56 & 7.5 & & 9.8 & 550 \\
16 & 0.015 & 0.180 & 1.58 & 7.4 & & 11.6 & 548 \\
5\% higher growth rate & & & & & & & \\
17 & 0.015 & 0.199 & 1.59 & 5.2 & & 5.3 & 501 \\
18 & 0.015 & 0.246 & 1.53 & 7.9 & & 10.5 & 493 \\
19 & 0.015 & 0.215 & 1.53 & 7.2 & & 7.7 & 494 \\
20 & 0.015 & 0.193 & 1.56 & 6.8 & & 6.8 & 496 \\
21 & 0.015 & 0.191 & 1.60 & 5.6 & & 5.8 & 498 \\
22 & 0.015 & 0.210 & 1.56 & 7.5 & & 7.3 & 495 \\
23 & 0.014 & 0.183 & 1.71 & 6.1 & & 5.4 & 499 \\
\hline
{\bf 3 flask scales} & \multicolumn{3}{l}{3 flask scales} & & & & \\
5\% lower growth rate & & & & & & & \\
24 & 0.015 & 0.174 & 4.14 & & & 20.0 & 550 \\
25 & 0.015 & 1.011 & 4.03 & & & 25.7 & 549 \\
26 & 0.015 & 0.151 & 4.02 & & & 19.2 & 552 \\
27 & 0.015 & 0.929 & 4.34 & & & 29.1 & 548 \\
5\% higher growth rate & & & & & & & \\
28 & 0.015 & 0.235 & 4.277 & & & 11.13 & 496 \\
29 & 0.015 & 0.987 & 7.683 & & & 25.5 & 493 \\
30 & 0.015 & 0.267 & 4.589 & & & 14.9 & 495 \\
\hline
\end{tabular}
\end{table}
\subsection{Optimization regarding four objectives including product concentration}\label{res: 3_objectives}
To show the applicability of the proposed method to more than two objectives, a third and a fourth objective criterion, titer concentration and viability at the end of the production vessel (after 8 days) was added.
Whereas the first two objective criteria (seed train duration and deviation rate) are related to the seed train itself, the third and fourth criterion refer to the generated product in the production vessel and to the viability of the cells in the production vessel. Product concentration as well as product quality can be influenced by many factors (seeding cell density, substrate concentrations and nutrient feeds, metabolite production, temperature, pH, dissolved oxygen and carbon dioxide concentration, osmolality and more) and also by the amount and the state of the cells at the end of the seed train. Since no data describing product quality are available, product concentration and viability are considered in this study. A further simplification that was made is the assumption that the production vessel is performed in batch-mode (meaning without any addition of nutrient feeds or medium renewals). The reason for this simplification is to avoid confounding effects.
The authors are aware of the fact that many factors affect product concentration and product quality and when data of other critical process parameters or quality attributes are available, these could also be considered in the same manner. The main purpose of the present simulation example is to demonstrate how the proposed method can be applied to more than two objectives and how the corresponding results can be illustrated and interpreted.
To obtain a visual overview for multiple objective criteria in one figure, a so-called spider plot (or net plot) can be used, which is shown in Figure~\ref{fig: spider_4obj_5sf_20runs}.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{figures/Spider_plot_5sf_4obj_with_text.png}
\caption{Spider plot showing the objective criteria values (seed train duration, deviation rate, titer and viability after 8 days in the production vessel) for the Pareto optimal solutions for 5 shake flask scales}.
\label{fig: spider_4obj_5sf_20runs}
\end{figure}
The horizontal axis shows the values of the deviation rate (on the right) and of the viability (on the left). The vertical axis shows the values of the seed train duration (above) and of titer (below). The aim of the optimization was to minimize seed train duration and deviation rate and to maximize viability and titer.
Each color (hyperplane) represents one of the Pareto optimal seed train configurations (based on the optimal combinations of filling volumes in shake flask scales).
Since seed train duration and deviation rate should be minimal and titer and viability should be maximal, hyperplanes covering the lower left area would be desired. However, no such solution (hyperplane) was obtained. The reason is that the optimization problem contains conflicting objective criteria, meaning that an improvement of one criterion leads to a degradation of another criterion. The here presented solutions are all non-dominated (compare to the green circles in the figures for 2 objective criteria).
For all shown solutions, deviation rate is rather low (4.9--7.3\%), the seed train duration lies between 521 and 562 hours and a titer of approximately 430--433 mg/L (assuming here a cell-specific production rate of $q_{\rm titer, max}=3.9\cdot10^{-10}$ mg\,cell$^{-1}$\, h$^{-1}$, as reported in \cite{Hernandez.2021}) and a viability of 52--53\% is reached after 8 days in the production vessel (here via batch-mode).
Of course, the obtained values depend a lot on the real process conditions (production bioreactor probably performed in fed-batch model) and the model parameter values obtained after model validation.
However, the presented simulation example shall illustrate how the proposed approach can be applied for risk-based decision making under consideration of several criteria that should be optimal.
\subsection{Impact of performed iterations during Bayes optimization}\label{res: iterations}
For the example of 3 shake flask scales, (followed by 3 bioreactor scales) and optimizing filling volumes for all shake flask scales with respect to the two objective criteria: seed train duration and deviation rate, the number of performed iterations during the optimization procedure was varied. First, 10 initial points (combinations of filling volumes) distributed based on a Latin hypercube design were evaluated, followed by 10 Bayes iterations, which means that 10 times the algorithm updates the black box model (the Gaussian process), calculates the acquisition function and proposes the next point based on the outcome of this calculation.
Then, the optimization was performed again for the same seed train setup but using 20 and then 30 Bayes iterations. The obtained solutions are illustrated in Figure~\ref{fig: comparison_10_20_30_bayes_iterations}.
\begin{figure}[h]
\includegraphics[width=1\linewidth]{"figures/sf3_comparison_bayes_iterations_cut.png"}
\caption[]{Objective criterion {\em seed train duration} over objective criterion {\em deviation rate} obtained for evaluated points (optimizable variables, here combinations of 3 shake flask filling volumes) for 10 (top left), 20 (top right) and 30 (bottom left) Bayes iterations; blue dots: based on the initial Latin hypercube design (LHC); yellow crosses: based on the points proposed by the algorithm; green circles: Pareto optimal solutions (=\,Pareto front).}
\label{fig: comparison_10_20_30_bayes_iterations}
\end{figure}
Increasing the number of iterations from 10 to 20 helped to identify one solution that has not been discovered when running only 10 iterations. This can be seen when comparing the green circles in the diagram top left and the green circles in the diagram top right. The solution with $D\approx$\,22 and $d$\,=\,523 cannot been found in the diagram top left. \\
Increasing the number of iterations from 20 to 30 did not lead to an improved optimum as can be seen when comparing the green circles in Figure~\ref{fig: comparison_10_20_30_bayes_iterations} diagram top right and diagram bottom left. This underlines the efficiency of the Bayes optimization. In the present example only 10 initial points (distributed randomly according to a Latin hypercube design) and 20 Bayes optimization iteration steps were required to obtain the results which were confirmed when applying 30 iteration steps.
\section{Conclusion}
A concept has been developed to use process models in combination with algorithms for Bayes optimization using Gaussian processes to solve multi-objective optimization problems in the context of biopharmaceutical production processes.
To illustrate this approach, a relevant exemplary optimization problem was chosen and solved using the proposed method.
The goal was to find optimal combinations of filling volumes for the shake flask scales of a seed 577 train leading to a minimum deviation rate regarding viable cell densities and a minimum process 578 duration.
Compared to a non-optimized reference seed train, the optimized process showed much 579 lower deviation rates regarding viable cell densities (\textless~10\% instead of 41.7\%) using 5 or 4 shake flask 580 scales and seed train duration could be reduced by 56\,h from 576\,h to 520\,h.
Overall, it is shown that applying Bayes optimization to a multi-objective optimization function with several optimizable input variables and under a considerable amount of constraints, lead to revealing results with a low computational effort.
This approach provides the potential to be used in form of a decision tool, e.g.\ for the choice of an optimal and robust seed train design but also to further optimization tasks within process development.
It should be noted that Bayes optimization and the corresponding computational modules could also be applied, even if no mechanistic process model is available, following a slightly different workflow. Instead of performing model-based in-silico experiments (process simulations), real lab experiments would be performed and fed back to update the black box model (here the Gaussian process). This adaptive procedure (also called Bayesian experimental design or experimental design with Bayesian optimization \cite{Greenhill.2020}) or further related optimization methods might be promising tools to support experimental planning, process characterization, process transfer or optimization of cell culture processes but they still require further research and being embedded in software solutions easy to use for operators.
\clearpage
\vspace{6pt}
\section*{Acknowledgments}
{
The authors would like to express special thanks to Christoph Posch (Novartis Technical Research \& Development) for the fruitful scientific exchange regarding the here presented case study. Moreover, we acknowledge support for the Open Access fees by Ostwestfalen-Lippe University of Applied Sciences and Arts (TH OWL) in the funding program Open Access Publishing.}
\section*{Abbreviations}
{The following abbreviations are used in this manuscript:\\
\noindent
\begin{tabular}{@{}ll}
CHO & Chinese hamster ovary \\
EI & Expected improvement \\
FDA & Food and Drug Administration \\
GP & Gaussian process \\
LHS & Latin hypercube sampling \\
ode & Ordinary differential equations \\
SE & Squared exponential \\
VCD & Viable cell density \\
\end{tabular}}
\begin{table}
\label{tab:list_of_symbols}
\begin{tabular}{ll}
{\bf List of symbols} & \\
$\alpha$ & Risk aversion parameter (-) \\
$\mu$ & Cell-specific growth rate (h$^{-1}$) \\
$\mu_{\rm d}$ & Cell-specific death rate (h$^{-1}$ ) \\
$\mu_{\rm d,max}$ & Maximum cell-specific death rate (h$^{-1}$) \\
$\mu_{\rm d,min}$ & Minimum cell-specific death rate (h$^{-1}$) \\
$\mu_{\rm max}$ & Maximum cell-specific growth rate (h$^{-1}$)\\
$\mu_{\rm ref}$ & Reference maximum cell-specific growth rate (h$^{-1}$)\\
$\sigma^2$ & Variance \\
$c_{\rm Amm}$ ($c_{\rm Amm,0}$) & (Initial) ammonia concentration (mmol L$^{-1}$)\\
$c_{\rm Glc}$ ($c_{\rm Glc,0}$) & (Initial) glucose concentration (mmol L$^{-1}$)\\
$c_{\rm Gln}$ ($c_{\rm Gln,0}$) & (Initial) glutamine concentration (mmol L$^{-1}$) \\
$c_{\rm Lac}$ ($c_{\rm Lac,0}$) & (Initial) lactate concentration (mmol L$^{-1}$)\\
$c_{\rm titer}$ ($c_{\rm titer,0}$) & (Initial) volumetric titer (product concentration) (mg L$^{-1}$)\\
$d$ & Dimension of the input space, number of optimizable variables \\
& =~seed train duration (h)\\
$D$ & Data, Deviation rate \\
${\rm E}(\cdot)$ & Expectation value \\
$f$ & Objective function (-) \\
$f_i$ & Component $i$ of a multidimensional objective function (-) \\
$F_{\rm sample}$ & Change of volume due to sampling (L h$^{-1}$)\\
$i$ & Running index (-) \\
$k$ & Covariance function \\
$K_{ {\rm Amm}}$ & Correction factor for ammonia uptake (-)\\
$K_{{\rm Lys}}$ & Cell lysis constant (h$^{-1}$)\\
$K_{{\rm S,Glc}}$ & Monod kinetic constant for glucose (mmol L$^{-1}$) \\
$K_{{\rm S,Gln}}$ & Monod kinetic constant for glutamine (mmol L$^{-1}$) \\
$k_{{\rm Glc}}$ & Monod kinetic constant for glucose uptake (mmol L$^{-1}$)\\
$k_{{\rm Gln}}$ & Monod kinetic constant for glutamine uptake (mmol L$^{-1}$) \\
$m$ ($m(\cdot)$) & Mean (mean function) \\
$n$ & Number of shake flasks (-) \\
$N$ & Number of iterations (-) \\
$\mathcal{N}$ & Normal distribution (-) \\
$n_{\rm lhs}$ & Number of latin hypercube points (-) \\
$q_{Amm}$ ($q_{ {\rm Amm,uptake,max}}$) & (Maximum) cell-specific ammonia uptake rate (mmol cell$^{-1}$ h$^{-1}$) \\
$q_{ {\rm Glc}}$ ($q_{ {\rm Glc,max}}$) & (Maximum) cell-specific glucose uptake rate (mmol cell $^{-1}$ h$^{-1}$ )\\
$q_{ {\rm Gln}}$ ($q_{ {\rm Gln,max}}$) & (Maximum) cell-specific glutamine uptake rate (mmol cell $^{-1}$ h$^{-1}$ )\\
$q_{ {\rm Lac}}$ ($q_{ {\rm Lac,uptake,max}}$) & (Maximum) cell-specific lactate uptake rate (mmol cell$^{-1}$ h$^{-1}$ )\\
$q_{ {\rm titer}}$ ($q_{\rm titer,max}$) & (Maximum) cell-specific product production rate (mg cell$^{-1}$ h$^{-1}$)\\
$ \mathbb{R}$ & Set of real number \\
$t$ & Time (h) \\
$T_{\rm p}$ & Set of feasible points in time for passaging \\
$U(\cdot) $ & Utility function \\
$V$ & Volume (L) \\
$V_i$ & Volume in shake flask scale $i$ \\
${\rm Var}(\cdot)$ & Variance \\
$x_c$ & Candidate point \\
$x$, $x'$ & Multidimensional points (vectors) of the input space \\
$x*$ & Argument that maximizes $f(s)$ \\
$X_{\rm t}$ & Total cell density (cells L$^{-1}$) \\
$X_{\rm v}$ & Viable cell density (cells L$^{-1}$) \\
$X_{{\rm v},i}$ & Viable cell density at point in time with index $i$ (cells L$^{-1}$) \\
$\mathcal{X}$ & Input space \\
$y$ & Arbitrary function (-) \\
$Y$ & Arbitrary random variable (-) \\
$Y_{{\rm Amm/Gln}}$ & Kinetic production constant for ammonia (mmol mmol$^{-1}$) \\
$Y_{{\rm Lac/Glc}}$ & Kinetic production constant for lactate (mmol mmol$^{-1}$) \\
\end{tabular}
\end{table}
\clearpage
|
\section{Introduction} \label{sec:intro}
There is a growing interest in human tracking \cite{wu2020mmtrack}, \ac{har} \cite{singh2019radhar} and person identification \cite{zhao2019mid} using \ac{mmwave} devices, where the high carrier frequency and large available bandwidth allow for accurate Doppler estimation and precise localization and tracking. To fully exploit these properties, a large body of work has focused on dedicated \ac{mmwave} radars, that adopt specifically designed frequency modulated transmissions to extract the \ac{md} effect caused by human motion (a so-called \ac{md} signature) \cite{vandersmissen2018indoor, seifert2019toward}.
At the same time, given the increasing number of \ac{mmwave} network technologies such as $3$GPP $5$G-NR \cite{5G-NR_R15} and IEEE~802.11ad/ay \cite{802.11ad, 802.11ay}, \ac{isac} solutions are highly appealing. They effectively repurpose communication devices by endowing them with additional environment sensing capabilities, thus avoiding the cost of installing dedicated radar hardware.
This recent trend has led to the identification of sensing as a key feature of next generation $6$G mobile networks \cite{liu2021integrated} and the creation of the IEEE~802.11bf standardization group \cite{restuccia2021ieee}, aimed at enabling sensing features in \acp{wlan}. However, \ac{isac} system designs are still very limited, focusing on joint communication and sensing waveform design \cite{liu2018toward}, which would require significant modifications to existing communication protocols and a reduction in the achievable communication data rates. Other approaches \cite{liu2021integrated, li2021rethinking, pegoraro2021rapid, zhang2020mmeye} need to alternate communication and sensing phases according to a time-division scheme, where regularly spaced, radar-like transmissions are performed during dedicated sensing periods. This is needed to perceive the fine-grained \ac{md} effect of human motion, for which dense and regular sampling of the \ac{cir} is required, causing significant overhead and channel occupation.
In this paper, we address the problem of enabling \ac{isac} in realistic \ac{mmwave}
communication systems, by reusing existing communication traffic for sensing as much as possible and thus introducing only a minimal amount of additional overhead. To this end,
we propose SPARCS, the first \ac{mmwave} \ac{isac} system that reconstructs human \ac{md} signatures from \textit{irregular and sparse} \ac{cir} samples obtained from realistic traffic patterns.
The main insight of SPARCS is to leverage the intrinsic sparsity of the reflections in the \ac{mmwave} channel to pose the \ac{md} reconstruction as a sparse recovery problem. Indeed, \ac{mmwave} \ac{cir} estimation can naturally separate signal propagation paths with $<10$~cm resolution, leading to a sparse multi-path environment and consequently a sparse \ac{cir} in the Doppler domain.
This allows obtaining highly accurate \ac{md} signatures from only a small, randomly distributed fraction of the \ac{cir} samples that are currently needed by existing \ac{isac} methods.
To do so, SPARCS first performs \ac{cir} resampling to construct a regular grid of \ac{cir} samples with missing vales due to the irregularity of the sampling process in time. Next, a sparse reconstruction method is used to obtain the \ac{md} spectrum, decoupling different propagation paths to leverage their sparsity property.
Lastly, whenever communication traffic is absent or insufficient for the \ac{md} extraction, SPARCS supports a dynamic injection of very short \ac{cir} estimation fields into the (idle) channel. Given its sparse recovery capabilities, only a small number of additional \ac{cir} sensing units are needed to retrieve the \ac{md}, thus entailing a negligible overhead to the communication rate.
SPARCS is compatible with any \ac{mmwave} system that supports transmit beamforming for directional communication and
\ac{cir} estimation. This is the case, for example, for IEEE~802.11ay \acp{wlan} at $60$~GHz, which provide in-packet \ac{cir} estimation for beam tracking purposes, and for $3$GPP $5$G-NR, where base stations can send frequent downlink \ac{csirs} to estimate the channel using different \acp{bp}.
To evaluate SPARCS' performance, we implement it on a $60$~GHz IEEE~802.11ay \ac{sdr} experimentation platform. We then test it on sparse and irregular \ac{cir} samples derived from standard-compliant traces, both for synthetic traffic and traffic patterns obtained from datasets of operational real-world WiFi \acp{ap} deployments \cite{pdx-vwave-20090704}.
To assess the quality of the reconstructed \ac{md} signatures, we use them as input for a typical downstream task such as \ac{har}, which classifies human movement detected by the captured \ac{md} into different possible activities.
The main contributions of our work are summarized next.
\begin{enumerate}
\item We propose SPARCS, an \ac{isac} method for \ac{mmwave} systems that can reconstruct high-quality \ac{md} signatures of human movement from irregular and sparse \ac{cir} estimation samples. SPARCS reuses training fields appended to communication packets as sensing units, and injects additional sensing units if necessary, adapting to the underlying communication traffic and minimizing the sensing overhead.
\item We provide an original formulation of the \ac{md} extraction in communication systems as a sparse recovery problem, leveraging the intrinsic high distance resolution and sparsity properties of the \ac{mmwave} channel. As a side effect, this also improves the quality of the resulting spectrograms, making them more robust to noise and interference.
\item We design and validate an algorithm to perform the injection of additional sensing units when communication traffic is insufficient. The process is dynamic, requires no knowledge about future packet transmissions, and incurs minimal overall overhead.
\item We evaluate SPARCS by implementing it on an IEEE~802.11ay-compliant $60$~GHz \ac{sdr} platform and testing it on \ac{cir} measurements collected with realistic WiFi traffic patterns. For the common \ac{har} task, the \ac{md} signatures reconstructed by SPARCS achieve better F$1$ scores than existing methods, while reducing sensing overhead by a factor of $7$.
\end{enumerate}
The paper is organized as follows. In \secref{sec:prel} we discuss the necessary preliminaries for \ac{mmwave} human \ac{md} sensing using \ac{cir}.
SPARCS is introduced and explained in detail in \secref{sec:method}, describing the sparse recovery problem formulation and the involved processing steps.
In \secref{sec:implementation} we discuss the implementation of SPARCS on an \ac{sdr} platform, and \secref{sec:results} provides an evaluation of the system on real measurement traces. We summarize the related work in \secref{sec:rel} and give concluding remarks in \secref{sec:conclusion}.
\section{Primer on mmWave sensing} \label{sec:prel}
In this section we give a brief description of the \ac{cir} model for \ac{mmwave} communication systems that we use for sensin
. We then describe a baseline approach that allows tracking the movement of people in the environment and extract their \ac{md} signatures using \textit{regularly sampled} \ac{cir} information. This forms the basis of our SPARCS design, which
entirely \textit{eliminates} the requirement of fixed \ac{ifs} and enables ultra low-overhead \ac{isac}.
\subsection{Sensing in \ac{mmwave} systems}\label{sec:sensing-ay}
Capturing the movement features of humans in the environment requires an analysis of the reflections of the transmitted signal from their bodies, which is usually carried out applying signal processing techniques to the \ac{cir}. Due to the high path loss occurring at \ac{mmwave} frequencies, directional communication is employed by means of transmitter and receiver beamforming, typically using phased antenna arrays. The transmitter and the receiver use suitable \ac{bp} configurations of their antenna arrays to maximize the signal strength
\cite{giordani2018tutorial, 802.11ad, 802.11ay}.
To successfully sense with a \ac{mmwave} system, \textit{at least} one of the \acp{bp} has to illuminate the subjects of interest,
as only in this case the reflected signal carries detectable information about the movement signature. To this end, we consider a setup where an \ac{ap} transmits packets and is able to collect the reflections of \textit{its own} signal, after being reflected by objects (including humans). This reflection is collected by the receive array of the \ac{ap} itself using a quasi-omnidirectional \ac{bp}. This requires full-duplex capabilities, as is common in \ac{isac} scenarios \cite{kumari2017ieee}, which in the simplest form can be achieved with a \ac{mimo} system in a mixed configuration with one RF chain as transmitter and another as receiver. The \ac{cir} estimation fields used for sensing, which we denote by \textit{sensing units}, can either be piggybacked by appending them as a trailer to the \ac{phy} communication packets or transmitted independently (\textit{injected}).
\ac{mmwave} standards implement beam training mechanisms that help to establish a communication link by testing different \ac{bp} combinations and then selecting the best one. Such functionality is supported by all \ac{mmwave} standards. For example, $5$G-NR \cite{giordani2018tutorial}, use \ac{ssb} and \ac{csirs} for beam management, while \ac{wlan} systems adopting the IEEE~802.11ad/ay standards \cite{802.11ad, 802.11ay} use channel estimation and training fields (CEF and TRN, respectively) to obtain accurate \ac{cir} information. Our framework to extract sensing information from \ac{cir} measurements can be applied regardless of the specifics of the standards.
\subsection{\ac{mmwave} CIR model} \label{sec:cir-model}
Due to the large transmission bandwidth of \ac{mmwave} systems,
channel measurements
contain fine-grained information about the environment \cite{zhang2020mmeye, pegoraro2021rapid, liu2021integrated}. Depending on the communication system we consider, sensing could be performed using the $5$G-NR \ac{ofdm} \ac{csi}, which contains the channel gains for each \ac{ofdm} subcarrier, or the IEEE~802.11ad/ay \ac{sc} \ac{cir}. Both communication schemes are suitable for human sensing: \textit{(i)} in $5$G-NR, the base stations can send frequent downlink \ac{csirs} to estimate the channel using different \acp{bp}, while \textit{(ii)} in IEEE~802.11ay in-packet beam tracking is enabled, so that specific fields called training fields (TRN), each using a different \ac{bp}, can be appended to communication packets.
In the following, we focus on \ac{sc} \ac{cir}, and show how to extract the \ac{md} effect of human movement. However, previous works have demonstrated that similar processing can be performed with \ac{ofdm} \ac{csi} \cite{meneghello2021environment, liu2021integrated}, and SPARCS is general enough to be applied in both cases.
\rev{We consider a multipath propagation environment with a time-varying number of reflectors, $P(t)$. These cause physical signal propagation paths that can be separated in the \ac{cir} according to a finite \textit{ranging resolution}, i.e., the capability of the system to resolve the distance of the reflectors causing different signal paths. This is given by $\Delta d = c/2B$, where $B$ is the transmitted signal bandwidth and $c$ the speed of light. Thus, the \ac{cir} contains the complex channel gains for a discrete grid of possible signal paths (or \textit{distance bins}), with indices $\ell= 0,\dots, L-1$. Each path is associated with a specific distance from the \ac{ap}, according to the relation $d_{\ell}=c\tau_{\ell}/2$, with $\tau_{\ell}$ being the delay associated with path $\ell$.} Moreover, the \ac{cir} depends on the specific \ac{bp} used during the transmission, denoted by $b = 0, \dots, N_{\rm BP}-1$. For carrier frequency $f_o$, the \ac{cir} along $\ell$, using \ac{bp} $b$ at time $t$ is
\begin{equation}\label{eq:cir}
h_{\ell, b}(t) = \sum_{p=1}^{P_{\ell}(t)} a_{\ell, b}^p(t) \exp\left\{-j2\pi \frac{2f_o}{c} \left[d_{\ell} + \int_0^{t} v_{\ell}^p(x) dx\right]\right\}.
\end{equation}
In \eq{eq:cir}, $P_{\ell}(t)$ is the number of physical reflectors whose contributions overlap in the $\ell$-th \ac{cir} path, as their distances are within $d_{\ell} \pm \Delta d /2$, while $v_{\ell}^p$ is the radial velocity\footnote{By convention, $v_{\ell}^p$ has a positive sign when the reflector moves away from the \ac{ap}.} of reflector $p$. The quantity $a_{\ell, b}^p(t)$ is the complex gain due to the joint effect of the transmitter \ac{bp}, the object reflectivity and the signal attenuation.
\subsection{micro-Doppler extraction} \label{sec:md-extr}
The extraction of the \ac{md} spectrum from multiple, concurrently moving subjects requires tracking the position of each person in the physical space, in order to separate their individual contributions to the \ac{cir}. Then, a spectral analysis over different \ac{cir} samples yields the desired \ac{md} signature \cite{pegoraro2021rapid, korany2020multiple}.
\subsubsection{People tracking} \label{sec:track}
\rev{People tracking is performed by extracting measurements of each person's distance and angular position with respect to the \ac{ap} across time, as detailed, e.g., in \cite{pegoraro2021rapid}. This process consists of \textit{(i)} removing the background contribution to the \ac{cir} by subtracting the average \ac{cir} across a suitable time interval, \textit{(ii)} selecting the locally strongest reflection paths in the \ac{cir} (peaks) and obtaining the corresponding distance $d_{\ell}$, \textit{(iii)} computing the \ac{aoa}, $\theta$, of the reflection from the correlation between the different \acp{bp} gains and the strengths of the \acp{cir} across the whole angular \ac{fov} \cite{Lacruz_MOBISYS2020}. The approach in \textit{(iii)} requires the \ac{bp} shapes to be estimated in advance, and is based on the intuition that different \acp{bp} illuminate different possible reflectors in the environment, depending on their position. Therefore, we can expect to receive the reflected signal from a certain subject only when a \ac{bp} is pointing in his/her direction.}
The resulting set of distances and angles represent candidate positions of humans in the environment. A multi-target tracking method such as a \ac{jpdaf} \cite{shalom2009probabilistic} allows smoothing the trajectories of the subjects by rejecting noise and clutter.
The tracking phase provides an estimate of the position of each subject, at every time instant $t$, which we denote by $[\hat{d}(t), \hat{\theta}(t)]$
using symbol $\hat{}$ to differentiate between our estimates and the true position. Due to step $(ii)$ above, which selects the local peaks in the received power, typically only the reflections from the torso can be tracked when a person moves in the environment. In contrast, the head and the limbs cause much weaker reflections \cite{vandersmissen2018indoor}, whose contributions can only be detected in the \ac{md} spectrum.
Nevertheless, tracking the spatial location of the torso is crucial for the \ac{md} extraction as it allows separating the contributions of the multiple subjects in the environment.
Once $\hat{d}(t)$ and $\hat{\theta}(t)$ are determined, they are used to select the path $\ell^*$ and the \ac{bp} $b^*$ that correspond to the distance and angular position of the subject, respectively. Then, the \ac{cir} waveform that contains the \ac{md} effect of the person's movement is $h_{\ell^*, b^*}(t)$. \rev{Besides enabling the separation of multiple subjects, this operation makes mmWave sensing systems much more robust to changes in the environment than sub-$6$~GHz systems \cite{meng2020gait}. While the latter are heavily affected by second-order reflections on walls and objects, mmWave sensing mostly relies on the line-of-sight path between the person and the transmitter/receiver, with little contribution from the external environment.}
\subsubsection{micro-Doppler spectrum} \label{sec:md}
Human movement causes a small-scale Doppler effect on the reflected signal due to the different body parts, which possess different velocities and follow different trajectories \cite{chen2006micro}. This is referred to as \ac{md} effect and causes a measurable frequency modulation on the reflection. High frequency signals such as \ac{mmwave} communications are particularly affected by the \ac{md} modulation due to their small wavelengths. Various techniques from time-frequency analysis can be applied to analyze the \ac{md}, obtaining \textit{spectrograms} showing the time evolution of the signal energy contained in the different frequency bands of interest. The most popular and computationally efficient of such methods is the \ac{stft} of $h_{\ell, b}(t)$ \cite{boashash2015time}, which consists in applying a windowed \ac{ft} on partially overlapping portions of the \ac{cir}.
As shown in \cite{pegoraro2021rapid, meneghello2021environment, korany2020multiple}, \ac{stft} processing needs to be applied to a window of $W$ subsequent estimates of the \ac{cir} with a \textit{fixed} \ac{cir} sampling interval of $T_c$ seconds, provided that the time spanned by the window is short enough to consider the movement velocity of the reflectors \textit{constant} for its whole duration. Note that this operation allows detecting and separating the velocities of the $P_{\ell}(t)$ reflectors, whose contributions overlap in path $\ell$ when considering a single estimate of the \ac{cir}.
The choice of $T_c$ impacts the frequency resolution of the \ac{stft}, $\Delta f^{d} = 1/(WT_c)$, and its maximum measurable frequency, $f^d_{\rm max} = 1/(2T_c)$. Using the relationship between the Doppler frequency and the corresponding velocity, one can obtain the velocity resolution and the maximum observable velocity as $\Delta v = c/(2f_o W T_c)$ and $v_{\rm max} = c/(4f_o T_c)$. To fully capture the range of velocities of interest for human movement, the typical approach is to select $T_c$ such that $v_{\rm max}$ is sufficiently high that is covers the velocities that can occur in the human activities of interest, which may vary depending on the application \cite{vandersmissen2018indoor, pegoraro2021rapid, singh2019radhar}.
Previous work assumes that the constraint of a fixed $T_c$ is met, which does not hold in realistic communication scenarios, where packet transmissions are scheduled according to the needs of the communication protocols rather than sensing accuracy. Traffic patterns are typically bursty and irregular and thus cannot be used by existing methods for human sensing. Instead, dedicated time slots need to be reserved for the transmission of sensing units, which is incompatible with the random access CSMA/CA MAC commonly used in IEEE 802.11.
Conversely, SPARCS is the first approach that does not require any specific pattern in the transmission of the sensing units, enabling true \ac{isac} by exploiting communication packets for sensing whenever possible, and introducing minimal additional overhead when necessary.
\section{SPARCS methodology}
\label{sec:method}
\begin{figure}
\centering
\includegraphics[width=8.4cm]{FIGURES/workflow_sparcs_small.pdf}
\caption{Comparison between the traditional \ac{cir}-based human sensing and SPARCS.}
\label{fig:workflow}
\end{figure}
We now present the SPARCS algorithm to recover the \ac{md} spectrum from irregular and sparse \ac{cir} sampling patterns.
The processing steps of SPARCS compared to traditional \ac{cir}-based sensing methods are shown in \fig{fig:workflow}.
\textbf{$(1)$ \ac{cir} resampling:} after \ac{cir} estimation and people tracking, for which we adopt the standard \ac{jpdaf} technique \cite{shalom2009probabilistic}, we apply a resampling strategy to approximate the irregularly spaced \ac{cir} values with a regular sequence whose sampling interval is chosen according to the desired \ac{md} resolution (\secref{sec:cir-resample}). Due to irregularity of the original sampling process, the approximated regular sequence may contain missing values that need to be \textit{filled} in the subsequent processing steps.
\textbf{$(2)$ Sparse \ac{md} recovery:} we formulate the recovery of the \ac{md} spectrum from the incomplete \ac{cir} measurements as a sparse recovery problem. \rev{For this, we leverage two key aspects. On the one hand, the intrinsic sparsity of the \ac{mmwave} channel leads to a small number of signal reflections from the human body that carry information about different body parts. On the other hand, the high distance resolution of \ac{mmwave} systems makes the reflections from the different body parts separable. The combined effect of these two properties is that the resulting \ac{cir} is highly sparse in the Doppler frequency domain, as detailed in \secref{sec:sparse-rec}}
We then solve the sparse recovery problem using the \ac{iht} algorithm for each \ac{cir} path (\secref{sec:single-path}), and aggregate the results to obtain the final \ac{md} spectrum (\secref{sec:multipath-acc}).
\textbf{$(3)$ Sensing unit injection:} when communication traffic is absent or too scarce to obtain an accurate reconstruction, our system can inject short sensing units into the (idle) channel to overcome the problem, as described in \secref{sec:injection}. Thanks to the sparse reconstruction of point $(2)$, the amount of units that need to be injected is minimal and can be tuned to trade off between overhead and sensing accuracy.
\subsection{CIR resampling}
\label{sec:cir-resample}
Our system samples the \ac{cir} at time instants $t_i$, which coincide with the reception of the reflections from the $i$-th transmitted
packet.
To reconstruct the \ac{md} spectrum from \ac{cir} samples which are randomly distributed in the time domain, we first resample the \ac{cir} to obtain regularly spaced samples with a fixed granularity $T_c$, where possible.
To do so, we resort to the \textit{slotted resampling} technique, which allows approximating a sequence of randomly spaced samples into a regular grid with \textit{missing values} \cite{babu2010spectral}.
We consider $N_s$ consecutive samples obtained at the time instants $t_0, t_1, \dots, t_{N_s}$ and denote by $0, T_c, 2T_c, \dots, (K-1)T_c$ the regular grid with step size $T_c$.
Slotted resampling constructs a new \ac{cir} sample sequence $h_{\ell, b}(kT_c)$ where the \ac{cir} values are obtained from the original sequence $h_{\ell, b}(t_i)$ as follows.
Time \textit{bins} (or intervals) of length $T_c$ are centered on each time instant of the regular grid, i.e., bin $k$ is $\beta_k = [kT_c - T_c/2, kT_c + T_c/2)$, with center $kT_c$.
Then, the value of the \ac{cir} corresponding to the $k$-th grid value is either \textit{(i)} selected among the values of the original sequence whose sampling times fall inside bin $k$, taking the one whose sampling time is the closest to the bin center, or \textit{(ii)} considered as a missing value if no samples of the original sequence fall inside bin $k$.
Specifically,
\begin{equation}
h_{\ell, b}(kT_c) = \left\{
\begin{array}{cc}
0 &\mbox{ if } \{t_i | t_i \in \beta_k\} = \emptyset,\\
h_{\ell, b}(t_k) & \mbox{otherwise},
\end{array}\right.
\end{equation}
where the $0$ values represent missing samples and
\begin{equation}
t_k = \argmin_{\tau \in \{t_i | t_i \in \beta_k\}} |kT_c - \tau|.
\end{equation}
The resulting, regularly spaced sequence of \ac{cir} samples is used to reconstruct the \ac{md} spectrum of the subject. However, due to the missing samples which are set to $0$, a plain application of the \ac{stft} (as described in \secref{sec:md-extr}) would lead to a corrupted spectrum. In the next section we detail our solution to this problem, which is based on sparse recovery techniques.
\subsection{Sparse \ac{md} recovery problem formulation}\label{sec:sparse-rec}
Several methods exist to tackle the problem of computing the power spectrum of non-uniformly sampled signals \cite{babu2010spectral}. Our approach belongs to the category of sparsity-based approaches, in which the sparsity of the signal in the frequency domain is leveraged to drastically reduce the number of measurements needed for an accurate reconstruction of the spectrum.
We select windows of length $W$ samples (window size) every $\delta$ samples from the sequence $h_{\ell, b}(kT_c)$, choosing $\delta=W/2$. Due to the slotted resampling process, each window may contain missing samples. We denote by $\mathcal{U}_m$ the set of indices of the available
samples contained in the $m$-th window.
Then, we define vector $\mathbf{h}_{\ell, b}(m) \in \mathbb{C}^{|\mathcal{U}_m|}$, containing the available \ac{cir} samples in the $m$-th window, and vector $\tilde{\mathbf{h}}_{\ell, b}(m) \in \mathbb{C}^{W}$, representing the complete $m$-th \ac{cir} window, which is only partially known due to the missing samples.
We also denote by $\mathbf{F}_{\rm inv}$ the inverse Fourier matrix, whose element in position $(g, l)$ is given by $\left( \mathbf{F}_{\rm inv}\right)_{gl} = (1/\sqrt{W})\exp \left(j 2\pi gl/W\right), \,\, g, l = 0, \dots, W - 1$
while $\mathbf{U}_m=\left[ \mathbf{u}^T_{i}\right], \forall i \in \mathcal{U}_m$ is the matrix that selects the rows of $\mathbf{F}_{\rm inv}$ whose indices are in $\mathcal{U}_m$. $\mathbf{u}_{i}$ is the vector of all zeros but the $i$-th component, which equals $1$.
The following relation holds between the incomplete \ac{cir} window, $\mathbf{h}_{\ell, b}(m)$, and the \ac{ft} of the full \ac{cir} window, $\mathbf{H}_{\ell, b}(m) \in \mathbb{C}^{W}$, which we aim to recover in order to compute the \ac{md} spectrum,
\begin{equation}\label{eq:cs-model}
\mathbf{h}_{\ell, b}(m) = \mathbf{U}_m \tilde{\mathbf{h}}_{\ell, b}(m)= \mathbf{U}_m \mathbf{F}_{\rm inv}\mathbf{H}_{\ell, b}(m) = \mathbf{\Psi}_m \mathbf{H}_{\ell, b}(m),
\end{equation}
where in the last step we use matrix $\mathbf{\Psi}_m = \mathbf{U}_m \mathbf{F}_{\rm inv}$ as a shorthand notation.
Given \eq{eq:cs-model}, our aim is to recover $\mathbf{H}_{\ell, b}(m)$ from the incomplete measurement vector $\mathbf{h}_{\ell, b}(m)$, which is a typical sparse recovery or compressed sensing problem \cite{eldar2012compressed}. In this framework, it has been proven that recovering the \ac{ft} of the desired signal is possible if the latter is sparse in the frequency domain, i.e., the \ac{ft} only contains a low fraction of non-zero elements.
To verify that this sparsity assumption holds in our case, we rewrite \eq{eq:cir} after the resampling and windowing operations, so that the $i$-th sample of the complete $m$-th window is given by
\begin{equation}\label{eq:cir-discrete}
\left[\tilde{\mathbf{h}}_{\ell, b}(m)\right]_{i} = \sum_{p=1}^{P_{\ell}(m)}a_{\ell, b}^p(m) \exp\left\{-j4\pi \frac{f_o}{c} \left[d_{\ell}^p + (m\delta+i) T_c v_{\ell, m}^p \right]\right\},
\end{equation}
where $v_{\ell, m}^p$ is the radial velocity of the $p$-th reflector in path $\ell$ during window $m$, and $d_{\ell}^p$ its distance from the \ac{ap}.
Here, we use the assumption from \secref{sec:md-extr} that the velocity of each reflector can be considered constant during a window. In addition, we also consider that the reflective coefficients and the number of reflectors are constant. This is reasonable for the considered setup, where the reflectors are parts of the human body, which typically move slowly compared to the duration of a window $WT_c$ (see also \secref{sec:implementation}).
\begin{figure}
\centering
\includegraphics[width=8.5cm]{FIGURES/separation.pdf}
\caption{Visual representation of the \ac{md} spectrum computed using SPARCS on $2$ different \ac{cir} paths, one containing the reflection from a person's torso, the other capturing the \ac{md} signature of the leg. The total \ac{md} is obtained summing together these contributions.}
\label{fig:separation}
\end{figure}
From \eq{eq:cir-discrete}, one can see that as long as $P_{\ell}(m) \ll W$, the \ac{ft} of $\tilde{\mathbf{h}}_{\ell, b}(m)$ is indeed sparse, as it is composed of $P_{\ell}(m)$ spectral lines located at frequencies $2f_o v_{\ell, m}^p / c$.
Given the excellent distance resolution due to the high bandwidth of \ac{mmwave} systems and the \textit{intrinsic} sparsity and directionality of the channel, the different parts of the subject's body tend to contribute to the \ac{md} spectrum in different \ac{cir} paths as shown in \fig{fig:separation}. Therefore, $P_{\ell}(m)$ is generally close, if not equal, to $1$. Sometimes the number of reflectors in a single path can be larger than $1$, due to different body parts being closer than the distance resolution of the system, but this number is still much lower than $W$.
\rev{This even holds for multiple subjects. Assume that two subjects with labels $1$ and $2$ are present in the monitored physical space, and denote by $(\ell_1, b_1)$ and $(\ell_2, b_2)$ their \ac{cir} path-\ac{bp} pairs. According to \eq{eq:cir-discrete}, the sparsity assumption must hold for each pair \textit{independently}, and this is verified as long as the subjects occupy different spatial positions. Specifically, \textit{(i)} if $\ell_1 \neq \ell_2$ the \acp{cir} along \acp{bp} $b_1$ and $b_2$ are the combination of $P_{\ell_1}(m)\ll W$ and $P_{\ell_2}(m)\ll W$ complex exponentials each, and \textit{(ii)} if $\ell_1 = \ell_2$, but $b_1 \neq b_2$, the attenuation coefficient of $b_1$ will mostly remove the reflection from subject $2$ in $\tilde{\mathbf{h}}_{\ell_1, b_1}$ and vice versa, making the contributions from the subjects separable. The contributions from different subjects overlap only if they occupy the same \ac{cir} path \textit{and} share the same \ac{bp}, which is very unlikely to occur in real cases due to the high distance ($\sim 8$~cm) and angular (as low as $ 2^{\circ}$) resolutions of the \ac{mmwave} \ac{cir} \cite{pegoraro2021rapid}.
Therefore, the sparsity assumption in SPARCS still holds even if multiple subjects are present in the environment.}
Due to this, we can assume that $P_{\ell}(m) \ll W$ holds, and that sparse recovery techniques can be used to recover $\mathbf{H}_{\ell, b}(m)$, as detailed in the next section.
\subsection{Single-path sparse recovery}
\label{sec:single-path}
\begin{algorithm}[t!]
\footnotesize
\caption{Single path sparse recovery.}
\label{alg:sparse-recovery}
\begin{algorithmic}[1]
\REQUIRE $\mathbf{h}_{\ell, b^*}(m)$, $\eta, n_{\rm max}, \Omega, \xi$.
\ENSURE $\mathbf{H}_{\ell, b^*}(m)$.
\STATE Collect the set of available samples indices $\mathcal{U}_m$.
\STATE Build matrices $\mathbf{U}_m=\left[ \mathbf{u}^T_{i}\right], \forall i \in \mathcal{U}_m$ and $\mathbf{\mathcal{F}}_{\rm inv}$
\STATE Compute $\mathbf{\Psi}_m = \mathbf{U}_m \mathbf{\mathcal{F}}_{\rm inv}$.
\STATE Set $\hat{\mathbf{H}}^{(0)} = \mathbf{0}$, $ n = 0$, $\gamma^{(0)}$ to any value $> \xi$.
\WHILE{$ n < n_{\rm max}$ or $\gamma^{(n)} > \xi$}
\STATE $\hat{\mathbf{H}}^{(n+1)} \leftarrow $ \eq{eq:iht}
\STATE $\gamma^{(n + 1)} \leftarrow ||\hat{\mathbf{H}}^{(n+1)} - \hat{\mathbf{H}}^{(n)}||_2$
\STATE $n \leftarrow n+1$
\ENDWHILE
\RETURN $\hat{\mathbf{H}}^{(n)}$
\end{algorithmic}
\end{algorithm}
Given the model from \eq{eq:cs-model}, the reconstruction of the \ac{cir} \ac{ft} along each path can be posed as a sparse recovery problem. Specifically, we seek a vector $\mathbf{H}_{\ell, b}(m)$ which is a solution to \eq{eq:cs-model} while being as sparse as possible, coherent with the above discussion.
Considering the \ac{bp} $b^*$ pointing in the direction of the target, the desired \ac{ft} of $\tilde{\mathbf{h}}_{\ell, b^*}(m)$ is the solution of the optimization problem
\begin{equation}\label{eq:sp-opt}
\mathbf{H}_{\ell, b^*}(m)= \argmin_{\mathbf{H}} ||\mathbf{H}||_0 \quad \mbox{ s.t. } ||\mathbf{h}_{\ell, b^*}(m) - \mathbf{\Psi}_m\mathbf{H}||_2 \leq \varepsilon,
\end{equation}
where $||\cdot||_0$ denotes the $\ell_0$-norm of a vector, i.e., the number of its non-zero components. The constant $\varepsilon > 0$ can be estimated from the noise in the \ac{cir}, using a training dataset.
An approximate local solution to \eq{eq:sp-opt} can be found using fast greedy algorithms \cite{eldar2012compressed}. We adopt the \ac{iht}, which solves
\begin{equation}\label{eq:local-opt}
\mathbf{H}_{\ell, b^*}(m)= \argmin_{\mathbf{H}} ||\mathbf{h}_{\ell, b^*}(m) - \mathbf{\Psi}_m\mathbf{H}||_2^2 \quad \mbox{ s.t. } ||\mathbf{H}||_0 \leq \Omega,
\end{equation}
where $\Omega$ is a pre-defined sparsity level parameter.
The algorithm involves an iterative gradient descent step on the quadratic term in \eq{eq:local-opt}, followed by a thresholding operation:
\begin{equation}\label{eq:iht}
\hat{\mathbf{H}}^{(n+1)} \leftarrow \mathcal{T}_{\Omega}\left[ \hat{\mathbf{H}}^{(n)} + \eta \mathbf{\Psi}_m^T\left( \mathbf{h}_{\ell, b^*}(m) - \mathbf{\Psi}_m\hat{\mathbf{H}}^{(n)} \right)\right],
\end{equation}
where $n$ is the iteration index and $\mathcal{T}_{\Omega}$ is the hard-thresholding operator, which sets to $0$ all the components of the argument vector except the $\Omega$ largest ones in terms of the Euclidean norm. $\eta$ is a learning rate parameter which can be tuned to improve the convergence properties.
The iterative process is stopped whenever $||\hat{\mathbf{H}}^{(n+1)} - \hat{\mathbf{H}}^{(n)}||_2 < \xi$ or when a maximum number of iterations, $n_{\rm max}$, is reached.
In SPARCS, $\Omega$ is a key parameter, which is strictly related to the number of reflectors $P_{\ell}(m)$: as \ac{iht} reconstructs a vector which has at most $\Omega$ non-zero elements, $\Omega$ is an upper bound for $P_{\ell}(m)$, and it can be thought of as the maximum number of reflectors per path that we allow reconstructing.
$\Omega$ can be tuned in order to obtain better \ac{md} reconstruction (see \secref{sec:params-sens}).
The sparse recovery algorithm is summarized in \alg{alg:sparse-recovery}.
\rev{According to the compressive sensing theory \cite{foucart2013mathematical},
the reconstruction performance of \ac{iht} (and in general of any recovery algorithm) degrades as the number of available measurements, $|\mathcal{U}_m|$, decreases. Theoretical results show that the minimum number of measurements needed to reconstruct $\mathbf{H}_{\ell, b^*}(m)$ is $\mathcal{O}(\Omega \log(W/\Omega))$ \cite{foucart2013mathematical}, although the exact number has to be estimated empirically as it also depends on the level of noise present in the signal. In \secref{sec:results}, we show that SPARCS can achieve excellent \ac{md} reconstruction with as few as $W/8$ measurements per window, thanks to the high sparsity of the \ac{mmwave} \ac{cir}.}
\subsection{Multi-path aggregation}\label{sec:multipath-acc}
The moving body of a person causes several reflections that affect more than one \ac{cir} path, as discussed in \secref{sec:sparse-rec}. Using the procedure described in the previous sections, SPARCS is able to retrieve the contribution of each path $\mathbf{H}_{\ell, b^*}(m)$ to the \ac{md}. Since the different body parts contribute to the \ac{md} in different paths, to fully capture human movement we need to combine the information from the different paths.
Denote by $Q$ the number of distance bins we aggregate to obtain the \ac{md} spectrum. For convenience we assume $Q$ to be an odd integer, as this makes the following processing steps symmetric with respect to a central \ac{cir} path (corresponding to the torso), but the same steps can be applied for $Q$ being even. We aggregate the spectra obtained from the path caused by the torso, $\ell^*$, with the $\lfloor Q/2 \rfloor$ distance bins preceding $\ell^*$ and the $\lfloor Q/2 \rfloor$ subsequent distance bins, as they may contain the contributions of the other body parts.
The expression of the total \ac{md} spectrum is
\begin{equation}\label{eq:sparse-paths-sum}
\mathbf{D}(m) =\sum_{\ell=\ell^* - \lfloor Q/2 \rfloor }^{\ell^* + \lfloor Q/2 \rfloor} \left|\mathbf{H}_{\ell, b^*}(m)\right|^2,
\end{equation}
where the squared magnitude is applied element-wise. In addition, we apply normalization of the spectra in the range $[0, 1]$ by computing $\mathbf{D}(m) \leftarrow \frac{\mathbf{D}(m) - \min_i \mathrm{D}_i(m)} {\max_i \mathrm{D}_i(m) - \min_i \mathrm{D}_i(m)}$.
Note that \eq{eq:sparse-paths-sum} entails solving $Q$ optimization problems of the form in \eq{eq:local-opt}, however, the $Q$ problems can be parallelized as they are completely independent.
Decomposing the full \ac{md} spectrum reconstruction problem into $Q$ subproblems effectively allows applying sparse recovery techniques, which in turn leads to a significant reduction of the number of measurements that are needed.
The value of $Q$ is selected according to physical considerations and validated in practice, as described in \secref{sec:results}.
The \ac{md} vectors from \eq{eq:sparse-paths-sum} can be collected in sequences, one every $\delta$ slots, forming \ac{md} spectrograms of arbitrary length, depending on the specific application that is being performed, e.g., activity recognition, fall detection, gait segmentation, etc. In the following, we refer to the number of \ac{md} vectors considered in such spectrograms as $\Lambda$.
\subsection{Sensing unit injection}
\label{sec:injection}
SPARCS can exploit the sensing units in sparsely distributed communication packets to recover the \ac{md} spectrum of human movement. \rev{However, during communication between the \ac{ap} and one or more terminals it may happen that the \ac{ap} remains silent for longer than the duration of a processing window, $WT_c$, or that the received packets are fewer than the minimum number of measurements required for an accurate \ac{md} reconstruction. In these cases, the sparse recovery algorithm can not recover $\mathbf{H}_{\ell, b}(m)$ as the available sensing units are insufficient.}
To tackle this problem, we allow our system to \textit{inject} sensing units into the channel whenever the number of communication packets is not sufficient for \alg{alg:sparse-recovery} to work. Different from existing \ac{isac} frameworks,
our sparse recovery approach allows us to introduce a minimal amount of overhead, as the \ac{md} spectrum can be recovered from a number of \ac{cir} samples which is much lower than the full length of the window $W$. Note that for the injection of a sensing unit it is sufficient to transmit the necessary \ac{cir} estimation fields, without any preamble and header as used in conventional packets, since the unit is only received at the \ac{ap} itself and contains a known waveform.
\subsubsection{Basis of the injection algorithm} \label{sec:inj-setup}
In the following, we present the proposed injection procedure assuming that both communication packets and sensing units are transmitted at times that lie on a uniform grid with spacing $T_c$. This simplification is valid due to the fact that the slotted resampling process described in \secref{sec:cir-resample} is used.
Therefore, we can describe the injection process in terms of windows of size $W$, where each value in the window occupies a \textit{slot} which is a multiple of $T_c$. Due to slotted resampling, the slots can be empty if no packet was transmitted sufficiently close to it.
Our approach consists in setting a minimum number of sensing units \textit{per window}, termed $M_{s}$, that allows a sufficiently accurate reconstruction of the \ac{md} signatures. We then transmit additional units whenever the number of reflections of communication packets in the window is not sufficient to meet this minimum requirement.
The proposed method only requires the knowledge of whether a reflected communication packet is received in the \textit{current} slot, i.e., no information about the future traffic pattern is needed.
\subsubsection{Algorithm description} \label{sec:inj-algo}
\begin{algorithm}[t!]
\footnotesize
\caption{Injection of sensing units in window $m$.}
\label{alg:inj-algo}
\begin{algorithmic}[1]
\REQUIRE $M_s$.
\STATE \textcolor{gray}{\texttt{\# P1 - observation phase}}
\STATE $N_a(m) \leftarrow$ no. of sensing units received in the first half of the window (either from reflected communications packets or injected).
\STATE \textcolor{gray}{\texttt{\# P2 - scheduling phase}}
\STATE $N_w(m) \leftarrow \max(M_s - N_a(m), 0)$.
\STATE Schedule $\mathcal{S}_m=\left\{s_1, \dots, s_{N_w(m)}\right\}$.
\STATE \textcolor{gray}{\texttt{\# P3 - transmission phase}}
\FOR{$q = m W/2, \dots, (m+1)W/2 - 1$}
\IF{$q \in \mathcal{S}_m$}
\IF{no reflected comm. packet received}
\STATE Transmit the sensing unit.
\STATE $\mathcal{S}_m \leftarrow \mathcal{S}_m \setminus \{q\}$.
\ELSE
\STATE Use the sensing unit from the comm. packet
\STATE $\mathcal{S}_m \leftarrow \mathcal{S}_m \setminus \{q\}$.
\ENDIF
\ELSE
\IF{reflected comm. packet received}
\STATE Use the sensing unit from the comm. packet
\STATE $\mathcal{S}_m \leftarrow \mathcal{S}_m \setminus \{\min_{s \in \mathcal{S}_m} s\}$.
\ENDIF
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
The algorithm, summarized in \alg{alg:inj-algo}, operates in three phases, namely \textit{observation} (P$1$), \textit{scheduling} (P$2$) and \textit{transmission} (P$3$). Recall that the \ac{md} extraction described in \secref{sec:sparse-rec} follows a window-based approach, with subsequent windows overlapping by half of their length, as shown in \fig{fig:injection}.
Consider a time instant between the end of window $m-1$ and the start of window $m+1$. This coincides with the \textit{half} of window $m$, which is between slots $m W/2-1$ and $m W/2$.
In this time instant we can observe how many reflected communication packets were received in the first half of window $m$, which spans the indices from $(m-1)W/2$ to $m W/2-1$ (P$1$, line $2$ in \alg{alg:inj-algo}). We denote this number as $N_a(m)$.
The injection algorithm is executed on a half-window basis at the time when window $m-1$ has ended and window $m+1$ has not yet started, as this allows reasoning on the sole current window $m$.
Based on $N_a(m)$, we can compute how many sensing units we would need in the remaining half of window $m$ in order to meet the requirement of at least $M_{s}$ units, which we denote by $N_w(m) = \max(M_s - N_a(m), 0)$. However, the sensing process has no knowledge of when future communication packets will be received, so the best we can do is schedule the transmission of $N_w(m)$ sensing units in the next half-window. The slots in which these packets are scheduled can be selected according to a deterministic rule or a probability distribution. We call $\mathcal{S}_m = \left\{s_1, \dots, s_{N_w(m)}\right\}$ the set of indices of the slots in which we schedule the additional sensing units for the next half-window (P$2$, lines $4$-$5$ in \alg{alg:inj-algo}).
While P$1$ and P$2$ are performed in a single time slot, before the second half of window $m$ starts, P$3$ (lines $7$-$23$ of \alg{alg:inj-algo}) is a dynamic process that spans the whole second half of window $m$. The indices of the slots considered in this part of the algorithm are $q = mW/2, \dots, (m+1)W/2 - 1$. Note that some communication packets, of which we have no knowledge, may be received in this second half-window. The procedure iterates over the slots and in each of them checks if a sensing unit was scheduled for that slot, i.e., if $q \in \mathcal{S}_m$. There are four possible cases:
\begin{figure}
\centering
\includegraphics[width=7.5cm]{FIGURES/injection.pdf}
\caption{Example injection procedure with $M_s=8, W=16$. $4$ sensing units are scheduled after P$1$ and P$2$. Then, as three reflected communication packets are received, we reuse them and the first three scheduled sensing units are not transmitted. The fourth sensing unit is instead injected in the last slot.}
\label{fig:injection}
\end{figure}
\noindent $\mathbf{(1)}$ $q \in \mathcal{S}_m$ and no communication packet was received in this slot. In this case we transmit the sensing unit, then remove $q$ from $\mathcal{S}_m$.
\noindent $\mathbf{(2)}$ $q \in \mathcal{S}_m$ and a communication packet (or more) was received in this slot. In this case we reuse the sensing unit in the communication packet and remove $q$ from $\mathcal{S}_m$.
\noindent $\mathbf{(3)}$ $q \notin \mathcal{S}_m$ and no communication packet was received in this slot. In this case we just move to the next slot without taking action.
\noindent $\mathbf{(4)}$ $q \notin \mathcal{S}_m$ and a communication packet (or more) was received in this slot. In this case we reuse the sensing unit in the communication packet, then we remove the next sensing unit from the scheduled ones, i.e., we set $\mathcal{S}_m \leftarrow \mathcal{S}_m \setminus \{\min_{s \in \mathcal{S}_m} s\}$.
Note that, despite operating on a half-window basis, due to the overlap of adjacent windows, our algorithm only poses a constraint on the minimum number of packets sent per \textit{full window}. This means that half a window can be empty as long as enough sensing units are received in the other half.
\subsubsection{Scheduling the sensing units}\label{sec:scheduling}
While the scheduling of the sensing units in P$2$ can be done with any arbitrary policy that guarantees that exactly $N_w(m)$ packets are scheduled in the next half window, we want to maximize the number of sensing units that can be piggybacked on communication packets, rather than using a dedicated transmission.
From P$3$ in \alg{alg:inj-algo}, one can see that scheduling the sensing units towards the end of the half-window leaves more time for possible communication packets to become available and thus be reused instead of injecting a new sensing unit.
Consequently, in SPARCS we schedule the
sensing units for the second half of window $m$ as a burst of packets spaced by $T_c$, which occupy the last $N_w(m)$ slots of the window.
\section{Implementation}\label{sec:implementation}
In this section we describe the implementation of SPARCS on a \ac{mmwave} \ac{sdr} platform. We base our implementation on the IEEE~802.11ay WiFi protocol, as it operates in the unlicensed $60$~GHz band and supports \ac{cir} estimation for different \acp{bp}.
\noindent \textbf{Testbed.} We use the open-source mm-FLEX experimentation platform from \cite{Lacruz_MOBISYS2020} for a baseline design. It has an FPGA-based baseband processor which can generate, capture and process (custom or standard compliant) frames with up to $1.76$~GHz of bandwidth. The baseband processor is connected to \ac{mmwave} RF front-ends with phased antenna arrays and supports various front-ends to operate in different frequency bands, e.g., at $28$~GHz or $60$~GHz \cite{SIVERSIMA_BOTH}. In the remainder of this paper we use a $60$~GHz RF front-end which simplifies experimentation as this is an unlicensed band, but we note that simply by changing the RF front-end, SPARCS can operate in a different band, e.g., for 5G-NR compatibility.
To implement SPARCS, we augment the functionalities of the testbed to enable full-duplex operation, not only in the baseband processor but also in the RF front-end. A block diagram of the system is shown in \fig{fig:implementation}. In the baseband processor, the operation of the system is controlled by a state machine (SM) which triggers an AXI-DMA that reads the I/Q samples from the DDR-TX memory and feeds them to the DACs. The SM also triggers another AXI-DMA in the receiver datapath that saves the receives samples in the DDR-RX memory, i.e., \textit{both} datapaths are synchronized and no packet detection is required. To support the variable \ac{ifs} extracted from real (or artificially generated) traces, we include a block RAM memory (BRAM) in the FPGA logic that stores the \ac{ifs} that will be used in the experiments. The SM reads these values sequentially, introducing a delay in the system according to the value read from memory. The variable \ac{ifs} functionality can be disabled at runtime to configure a \textit{fixed} \ac{ifs}.
We remark that since we simultaneously use the up/down conversion stages from the \textit{same} \ac{mmwave} development kit,
the Tx/Rx sub-systems are fed by the same local oscillator and thus the \ac{cfo} is very low ($< 100$~Hz), which enables the extraction of the \ac{md} values required by SPARCS.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\linewidth]{FIGURES/implementation.pdf}
\caption{SPARCS implementation block diagram.}
\label{fig:implementation}
\end{figure}
\begin{table}[t!]
\footnotesize
\caption{Summary of the SPARCS implementation parameters. The suggested values based on experimental results are shown in bold. \label{tab:params}}
\begin{center}
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{{\bf System parameters }} \\
\midrule
Grid step & $T_c$ & $0.27$~ms\\
Window length & $W$ & $64$\\
Window shift & $\delta$ & $32$\\
Sparsity parameter & $\Omega$ & $\{1,2,\boldsymbol{3},4,5,6,7\}$\\
No. aggregated paths & $Q$ & $\{1,3,5,7,\boldsymbol{9},11,13,15\}$ \\
Min. no. measurements & $M_s$ & $\{4, \boldsymbol{8}, 16, 24, 32, 64\}$\\
\ac{iht} learning rate & $\eta$ & $1$\\
\ac{iht} convergence threshold & $\xi$ & $10^{-4}$\\
\ac{iht} maximum iteration number & $n_{\rm max}$ & $200$\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\begin{figure*}[t!]
\begin{center}
\centering
\subcaptionbox{Ground truth \ac{stft}.\label{fig:gt}}[2.8cm]{\includegraphics[width=2.7cm]{FIGURES/gt.pdf}}
\subcaptionbox{Full window (\ac{stft}).\label{fig:stft-64}}[2.8cm]{\includegraphics[width=2.7cm]{FIGURES/stft_64.pdf}}
\subcaptionbox{$1/4$ sparse (\ac{stft}).\label{fig:stft-sparse-64}}[2.8cm]{\includegraphics[width=2.7cm]{FIGURES/stft_16.pdf}}
\subcaptionbox{Full window (SPARCS).\label{fig:sparse-64}}[2.9cm]{\includegraphics[width=2.7cm]{FIGURES/sparcs_64.pdf}}
\subcaptionbox{$1/4$ sparse ( SPARCS). \label{fig:sparse-16}}[2.8cm]{\includegraphics[width=2.7cm]{FIGURES/sparcs_16.pdf}}
\subcaptionbox{$1/16$ sparse (SPARCS). \label{fig:sparse-4}}[3.2cm]{\includegraphics[width=3.198cm]{FIGURES/sparcs_4.pdf}}
\vspace{-0.8em}
\caption{\rev{Walking \ac{md} spectrograms and RMSE for different levels of sparsity, obtained by uniformly removing samples for each window.}}
\label{fig:sparse-test}
\end{center}
\end{figure*}
\noindent \textbf{IEEE~802.11ay \ac{cir} estimation details.}
\rev{In IEEE~802.11ay, \textit{in-packet} beam tracking \cite{Knightly_ICOMM2017} is introduced, where the \ac{cir} is estimated using different \acp{bp} \textit{within a single packet}. This is implemented by appending a given number of training (TRN) fields to the packet. A TRN field is composed of $6$ TRN \textit{units} formed by complementary Golay sequences of $128$~BPSK modulated samples, for a total of $768$~samples \cite{802.11ay}.
In our implementation, we use $n_{\rm TRN}$ TRN fields as the SPARCS sensing unit, where each TRN field employs a different \ac{bp}, and $n_{\rm TRN}$ is the number of subjects being tracked by the system, as a single TRN field per subject suffices. Considering the typical number of people that are to be simultaneously tracked in human sensing systems, reasonable $n_{\rm TRN}$ values range from $1$ to $10$. The \ac{cir} estimates obtained from the TRN fields are then used as the input to SPARCS sparse recovery algorithm.}
\noindent \textbf{System parameters.} In \tab{tab:params} we summarize the system parameters used in the implementation.
We set $T_c = 0.27$~ms and $W=64$, which lead to \textit{(i)} a velocity resolution of $\Delta v= c/(2f_o W T_c)\approx 0.14$~m/s and \textit{(ii)} aliasing-free velocity measurements up to $v_{\rm max}= \pm c/(4f_o T_c)\approx \pm 4.48$~m/s. These values are not critical to the functioning of our system, and can be modified according to specific implementation requirements. However, for reliable \ac{md} extraction without aliasing, it is advisable to adjust $T_c$ to a value that allows capturing the range of velocities typically covered by human movement, e.g., approximately $\pm 2-3$~m/s for a walking person, and up to $\pm 5$~m/s for running or other fast movements \cite{vandersmissen2018indoor}.
Note that suitable values of $T_c$ can also be obtained in $5$G-NR systems, where a base station can transmit downlink \ac{csirs} frames with a periodicity between $0.3125$~ms and $80$~ms. For a $5$G-NR carrier frequency of $28$~GHz, using $T_c=0.3125$~ms leads to $v_{\rm max}\approx \pm 8.57$~m/s, which is enough to capture fast human movement.
\rev{For people tracking, we
use periodically transmitted in-packet beam training frames with $12$ TRN units and antenna beams covering a \ac{fov} range from $-45^\circ$ to $45^\circ$. Then, we utilize the distance and \ac{aoa} estimation procedure described in \secref{sec:track}, as proposed in \cite{pegoraro2021rapid}, to which we refer for further details}.
We experimented with different values of $M_s, \Omega$ and $Q$, as reported in \tab{tab:params} and described in \secref{sec:har} and \secref{sec:params-sens}, while for the \ac{iht} algorithm we selected the parameters that led to the most accurate convergence results on our experiments, i.e., $\eta=1$, $\xi=10^{-4}$ and $n_{\max}=200$.
\section{Experimental results} \label{sec:results}
We now present the experimental results obtained with our SPARCS testbed implementation.
The experiments were performed in a laboratory of $6$ $\times$ $7$~meters with a complex multi-path environment due to additional reflections caused by furniture, computers, screens, and a wide whiteboard.
\subsection{Results on synthetic traces}
\label{sec:qualitative}
As a first qualitative result we show the \ac{md} spectrograms obtained by SPARCS on randomly sampled \acp{cir} of a walking subject (see \fig{fig:sparse-test}). For this, we use synthetic traces, generated by measuring the \ac{cir} using a uniform sampling interval equal to $T_c$, and then setting to $0$ a variable number of uniformly distributed values \textit{per window} to simulate missing samples. This is a simplified case, as \textit{(i)} the available (not removed) packets lie on a regular grid with spacing $T_c$, therefore no approximation error is introduced by slotted resampling, and \textit{(ii)} samples are removed on a per window basis, so a minimum number of packets in each window is guaranteed. Still, this evaluation is useful to highlight the impact of increasing the sparsity level of the measurements for SPARCS compared to standard \ac{stft} \cite{pegoraro2021multiperson}.
In the results presented in this section, no packet injection is performed, as we aim to assess the impact of the number of measurements per window on the reconstructed \ac{md}.
In \fig{fig:stft-64} we show the baseline walking spectrogram obtained using the standard \ac{stft} using the full window of $64$ samples, as done in \cite{pegoraro2021rapid}. The spectrogram shows a typical walking \ac{md} modulation, with the contribution of the static clutter (the strong component at $0$ velocity), of the torso (the strong oscillating component around $\pm 1.5$~m/s and the limbs (the faint contributions around the torso component). Moreover, a certain amount of noise and interference is present, as shown by the non-zero background level and the horizontal lines at around $\pm 2$~m/s and $\pm 3.7$~m/s. In \fig{fig:stft-sparse-64}, the same method is applied to windows with only 16 out of 64 the samples retained, while the rest is set to $0$. The impact is very strong as it completely corrupts the useful structure in the \ac{md} signature.
From \fig{fig:sparse-64} to \fig{fig:sparse-4} we show the results obtained by SPARCS, on the same sequence, with $64, 16$ and $4$ samples out of $64$, respectively.
\rev{At the top of each figure, we report the Root Mean-Squared Error (RMSE) of the \ac{md} with respect to a ground truth spectrogram, shown in \fig{fig:gt}. This ground truth was obtained from the \ac{stft} output with full measurement windows (\fig{fig:stft-64}), by manually isolating the useful \ac{md} spectrum containing the gait information and setting to $0$ any background noise and interference lines.}
We observe two interesting aspects. On the one hand, the SPARCS algorithm can successfully recover the \ac{md} spectrum even when a large fraction of the samples is missing, and the quality of the result decreases \textit{gracefully} with the sparsity of the available measurements. \rev{Unlike standard \ac{stft}, SPARCS almost completely eliminates the noise and interference in the estimated $\mu$D spectrogram. Such improvement is made possible by the sparsity constraint in \eq{eq:local-opt}, which allows for a lower RMSE than \ac{stft} operating on full measurement windows}. This is the main reason why SPARCS not only reduces the overhead needed for human sensing, but also improves its accuracy.
\subsection{Realistic traces: the \texttt{pdx/vwave} dataset}\label{sec:pdx-vwave}
\begin{figure}
\centering
\includegraphics[width=8cm]{FIGURES/pdx_examples.pdf}
\vspace{-0.5em}
\caption{Example traffic patterns from the \texttt{pdx/vwave} dataset.}
\label{fig:crawdad}
\end{figure}
\begin{table}[t!]
\footnotesize
\caption{Details of the $3$ sequences of the \texttt{pdx/vwave} dataset.} \label{tab:pdx-vwave}
\vspace{-0.25cm}
\begin{center}
\begin{tabular}{lccc}
\toprule
\textbf{Trace} & \textbf{Environment} & \textbf{No. frames} & \textbf{Duration}\\
\midrule
\texttt{psu cs} & University CS dept.&$260326$&$1:00$~h\\
\texttt{library} & Public library&$1300671$&$4:00$~h\\
\texttt{ug} &Coffee shop &$895721$&$2:34$~h\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
Next, we evaluate the performance of SPARCS on realistic WiFi \ac{ap} traces. This poses an experimental challenge, because commercial devices implementing the IEEE~802.11ay standard are not yet available, and no public datasets containing real traffic traces for the \ac{phy} layer of \ac{mmwave} WiFi (IEEE~802.11ay/ad) exist, to the best of our knowledge.
For this reason, we used the \texttt{pdx/vwave} dataset, containing real traffic traces captured in different real environments from WiFi \acp{ap} employing a legacy (sub-$6$~GHz) WiFi protocol \cite{pdx-vwave-20090704}.
Specifically, we use $3$, over $1$ hour long, traces from this dataset, called \texttt{psu cs}, \texttt{library} and \texttt{ug}, respectively. We select traces collected in different environments to represent different kinds of traffic patterns (see \tab{tab:pdx-vwave}).
The \texttt{pdx/vwave} dataset includes information about the transmission instants and packet sizes of all packets outgoing from the considered \ac{ap}. Exploiting this information, we perform our measurements transmitting packets according to these time patterns (see \secref{sec:har}), using the BRAM in the FPGA to store the desired transmission instants (see \secref{sec:implementation}).
On top of the existing \texttt{pdx/vwave} communication patterns we use the injection algorithm (\alg{alg:inj-algo}) to send additional sensing units when needed.
Even though the \texttt{pdx/vwave} dataset is based on a legacy sub-6 GHz WiFi protocol,
we argue that it is still reasonable to use it to obtain realistic packet transmission patterns.
While in the \texttt{pdx/vwave} dataset the maximum physical layer PDU size is $\mathrm{PPDU}_{\rm pdx} = 1.5$~kB (without packet aggregation), in IEEE~802.11ay three main transmission modes are defined, namely \ac{ht}, \ac{dmg} and \ac{vht}, with maximum physical layer PDU sizes, $\mathrm{PPDU}_{\rm ay}$, of $65$~kB, $262$~kB and
$4692$~kB, respectively \cite{802.11ad, 802.11ay}. With the increase in the packet sizes, the data rates of \ac{mmwave} systems have increased accordingly, and in IEEE~802.11ay they will range from $0.3$~Gbps to several Gbps. As a numerical example, the traffic patterns in \texttt{pdx/vwave} with a typical bitrate of $4$~Mbps would correspond to a bitrate of $0.7$~Gbps in \ac{dmg} IEEE~802.11ay when using an aggregated packet size of $262$~kB instead of $1.5$~kB. Note that traces with a larger number of packets and smaller PDU sizes (as will likely be the case in real deployments) will simply increase the sensing accuracy and further reduce the overhead.
\subsection{Human activity recognition results}\label{sec:har}
To evaluate the quality of the \ac{md} spectrograms extracted by SPARCS, we use them as the input to a \ac{har} method. Specifically, we follow a standard approach, training a deep neural network on a dataset of $\Lambda \times W$ dimensional \ac{md} spectrograms, with $\Lambda = 200$ (equivalent to $\approx 1.76$~s), in order to classify the movement performed by the person during that time. In order to provide a comparison with other IEEE~802.11ay \ac{har} methods based on regular \ac{cir} sampling, such as RAPID \cite{pegoraro2021rapid}, we consider the $4$ following activities: walking, running, sitting and waving hands.
For \ac{har}, we use a standard \ac{cnn} architecture, composed of $4$ \textit{inception modules} \cite{szegedy2015going} performing $1\times 1$, $3\times 3$ and $5\times 5$ convolutions. The number of filters used is $8, 16, 32$ and $64$ for the $4$ modules, respectively. The convolutional blocks are followed by a fully-connected layer with $64$ neurons, to which we apply Dropout \cite{srivastava2014dropout}, and a final Softmax layer with $4$ outputs \cite{goodfellow2016deep}. We use the \textit{exponential-linear unit} activation function after each layer \cite{clevert2015fast}.
\subsubsection{Training data.} We collected a training dataset involving $6$ different subjects performing the $4$ activities, for a total duration of about $12$ minutes each. This leads to over $400$ partially overlapping, $1.76$~s long, \ac{md} sequences per activity, which we then augmented as described shortly. Note that the training data only includes \textit{uniformly sampled} \ac{cir} traces with sampling period $T_c = 0.27$~ms.
The \ac{cnn} training is done for $80$ epochs, using a learning rate of $10^{-4}$, the Adam optimizer and the cross-entropy loss function \cite{goodfellow2016deep}.
In order to enhance the robustness of the \ac{cnn}, we apply an ad-hoc data augmentation strategy: we randomly remove some of the \ac{cir} samples in each window of the training dataset, and then apply SPARCS' \ac{iht} algorithm to reconstruct the spectrograms (see \secref{sec:qualitative}). We repeat the process using a sparsity level of $1/8, 1/4$ and $1/2$, enlarging the training dataset to $4$ times its original size, for a total of approximately $1600$ \ac{md} spectrograms per activity. \rev{A randomly selected subset of the training data (around $10$~\%) was used as a validation set to tune the \ac{cnn} hyperparameters.}
\subsubsection{Test data.} We test the \ac{cnn} on the \ac{md} spectrograms obtained from \ac{cir} samples collected using the \texttt{pdx/vwave} packet traces described in \secref{sec:pdx-vwave}. We collect four, randomly selected, $20$~s long traces (one per activity) for each of the $3$ sequences types (\texttt{psu cs}, \texttt{library} and \texttt{ug}). We repeat the experiments for different values of the minimum number of sensing units per window, $M_s = 4, 8, 16, 32, 64$, for a total of $60$ test sequences. The test data involves a single subject, which was not included in the training set.
\begin{figure}
\centering
\includegraphics[width=8.4cm]{FIGURES/f1scores.pdf}
\caption{Per-class F$1$ scores obtained by SPARCS (for different values of $M_s$) and RAPID on our test dataset \cite{pegoraro2021rapid}.}
\label{fig:f1scores}
\end{figure}
\subsubsection{\ac{har} F$1$ score.} We evaluate the performance of the \ac{cnn} with the \textit{per-class} F$1$ score metric \cite{chinchor1992muc}, which effectively summarizes the precision and recall and preserves the class-specific results.
\fig{fig:f1scores} shows the total average per-class F$1$ score over the $60$ sequences, for different values of the minimum number of sensing units per window, $M_s$. As a baseline for comparison, we also report the F$1$ score obtained by the RAPID algorithm from \cite{pegoraro2021rapid}, which extracts the \ac{md} signatures by regularly sampling the \ac{cir}. Our results show that SPARCS can reach over $0.9$ F$1$ scores on all activities with $M_s=8$ already, which corresponds to only $1/8$ of the full measurements window. Notably, with $M_s=4$, the low number of measurements per window affects significantly only the 'Sitting' and 'Waving hands' activities, which involve fine-grained movements and are therefore more difficult to classify.
Finally, we compare the results from SPARCS and RAPID \cite{pegoraro2021rapid}. For a fair comparison, we implemented RAPID's \ac{stft} to extract the \ac{md} and trained the \ac{cnn} on the resulting spectrograms without enlarging the dataset using different levels of sparsity described in the previous section. Instead, we directly use the training procedure of \cite{pegoraro2021rapid}, since we found that the sparsity-based data augmentation slightly reduced RAPID's performance. It can be seen that SPARCS' sparse recovery problem formulation (\secref{sec:sparse-rec}) and enforcing a sparsity constraint on the individual paths is beneficial to \ac{har} performance. The gap is particularly significant for 'Sitting' and 'Waving hands' as they involve lower energy traces in the spectrograms; these are more easily corrupted by noise and interference, that SPARCS is mostly able to reject (see, again, the comparison between \fig{fig:stft-64} and \fig{fig:sparse-64}).
\subsection{Overhead analysis}
\label{sec:overhead-res}
Increasing $M_s$ to improve the \ac{har} performance also increases the overhead of SPARCS.
A first general measure of this can be obtained comparing the maximum size of a PPDU in IEEE~802.11ay to the size of a sensing unit. Recalling the three different modes introduced in \secref{sec:pdx-vwave} and the size of an IEEE~802.11ay TRN field ($768$~bits), we obtain that a sensing unit, with $n_{\rm TRN} = 1$, is $0.1\%$, $0.03\%$ and $0.002\%$ of a PPDU in \ac{ht}, \ac{dmg} and \ac{vht}, respectively.
Moreover, the channel occupation time for a sensing unit with $n_{\rm TRN} = 1$ is $436$~ns \cite{802.11ay}, which is a negligible fraction ($0.16\%$) of a slot of duration $T_c$.
\begin{figure}[t!]
\begin{center}
\centering
\includegraphics[width=4.3cm]{FIGURES/overhead.pdf}
\includegraphics[width=4cm]{FIGURES/acc_vs_oh.pdf}
\caption{Overhead of SPARCS for different values of $M_s$ in the three traces of the \texttt{pdx/vwave} dataset (left). Overhead vs.\ average \ac{har} F$1$ score for different values of $M_s$ (right).
}
\label{fig:overhead}
\end{center}
\end{figure}
Next, to evaluate the overhead of SPARCS on a realistic communication scenario, we use the traces of the \texttt{pdx/vwave} dataset. In this way, we can also assess the impact of injecting sensing units, as they are not useful to the communication process.
Denote by $c_i$ the number of bits in the $i$-th communication packet transmitted in the trace.
As the number of bits transmitted in each trace refers to a legacy, lower bitrate, WiFi protocol, we rescaled the packet sizes according to the maximum PHY layer packet size in IEEE~802.11ay.
We rescaled the size of packet $i$ in each trace as $\tilde{c}_i = (\mathrm{PPDU}_{\rm ay} / \mathrm{PPDU}_{\rm pdx}) \times c_i$, with $\mathrm{PPDU}_{\rm ay} = 262$~kB.
We call $n_{c}$ the number of transmitted communication packets in a trace, by $\mathrm{TRN}_{\rm len}$ the length, in bits, of a piggybacked or injected TRN field, $n_{\rm inj}$ the number of injected sensing units and $n_{\rm TRN}$ the number of TRN fields used in every sensing operation (we consider it fixed, whereas in reality it is determined by the number of subjects in the environment). We define the overhead as a function of $n_{\rm inj}$ as
\begin{equation}
\mathrm{OH}(n_{\rm inj}) = \frac{n_{\rm TRN}\left(n_{C} + n_{\rm inj}\right)\mathrm{TRN}_{\rm len}}{\sum_{i=1}^{n_{c}}\tilde{c}_i}.
\end{equation}
In \fig{fig:overhead}, left, we show the overhead obtained on each of the three \texttt{pdx/vwave} traces, using $n_{\rm TRN} = 1$. The overhead for different values of $n_{\rm TRN}$ can be obtained by using it as a multiplicative factor on the values in \fig{fig:overhead}. We see that the overhead scales almost linearly as $M_s$ is increased from $4$ to $64$.
For values of $M_s < 32$, the entailed overhead is less than $4$\%, falling below $1$\% for $M_s=8$
As a reference, we report the overhead for $M_s=64$, which is the value obtained by injecting sensing units continuously into the channel, piggybacking them eventually on communication packets if possible. Note that existing approaches requiring uniform \ac{cir} sampling, like RAPID \cite{pegoraro2021rapid}, would require an even higher overhead, as not only do they need $64$ samples per window, but these samples have to be regularly spaced as no resampling procedure is carried out. This means they would have to take precedence over potential data packets so that they are sent exactly at the right sampling time.
From \fig{fig:overhead}, right, one can see that SPARCS can achieve an F$1$ score of over $0.9$ for every activity for a minimum of $M_s=8$ sensing units per window, resulting in a sensing overhead of less than $1$\%. With this configuration, SPARCS achieves a better F$1$ score than existing approaches, while reducing overhead by a factor of $7$ and being compatible with random access MAC protocols.
\subsection{Sensitivity to the choice of the parameters}
\label{sec:params-sens}
In \fig{fig:params-vary}, we show the effect of varying parameters $Q$, representing the number of paths aggregated around the person's position (see \eq{eq:sparse-paths-sum}), and $\Omega$, which is the maximum number of resolvable Doppler components, equal to the sparsity parameter in the \ac{iht} algorithm.
We computed the \ac{har} per-class F$1$ score using a random subset of the $60$ test sequences. The values adopted in our experiments are reported in \tab{tab:params}.
\subsubsection{Impact of changing $Q$} Our results show that SPARCS is robust to almost any value of $Q$ when considering walking and running, whereas sitting and waving hands are negatively affected by reducing $Q$ below $7$. This is due to the fact that while walking and running are, in most cases, distinguishable even from the sole contribution of the torso, this is not true for sitting and waving hands that require including the reflection paths coming from the limbs. Computational complexity considerations are also in order for high values of $Q$, as it leads to solving $Q$ times \eq{eq:local-opt} at each \ac{md} extraction process. As the problems are independent, they can be solved in parallel, and thus a reasonable approach is to tune $Q$ according to a trade-off between \ac{md} reconstruction accuracy and hardware resource availability for parallelization.
In the following, we use $Q=9$. Considering that we use $B=1.76$~GHz transmission bandwidth ($1$ IEEE~802.11ay channel), the range resolution of SPARCS is $c/2B = 8.5$~cm. This means that summing the contribution of $\lfloor Q/2\rfloor$ distance bins before and after the one corresponding to the torso, we include in the spectrum a region of $\pm 34$~cm around the person's position, which is a reasonable value considering typical body sizes and that the subjects are moving.
\subsubsection{Impact of changing $\Omega$} Fixing $Q=9$, in \fig{fig:params-vary} (right), we show that the best values for $\Omega$ are $2$ and $3$ for all the activities. This is because using $\Omega=1$ often leads to only reconstructing the $0$ Doppler component in the spectrogram, losing the information on the person's movement. On the other hand, choosing $\Omega$ too high makes the \ac{iht} reconstruction imprecise, as with a low number of measurements per window enforcing more sparsity is beneficial to restrict the number of possible solutions to \eq{eq:local-opt}.
\section{Related work}\label{sec:rel}
\begin{figure}
\centering
\includegraphics[width=4.6cm]{FIGURES/vary_Q.pdf}
\includegraphics[width=3.8cm]{FIGURES/vary_om.pdf}
\caption{Per-class F$1$ scores aggregating a different number of paths $Q$ (left) and changing $\Omega$, the \ac{iht} sparsity parameter (right).}
\label{fig:params-vary}
\end{figure}
\noindent\textbf{Dedicated \ac{mmwave} radars.}
The high sensitivity of \acp{mmwave} to micro-Doppler shifts, together with \ac{dl} methods for spectrogram analysis and classification, have been widely exploited to enable applications such as activity recognition \cite{singh2019radhar, lai2021radar}, person identification \cite{meng2020gait, pegoraro2021multiperson} and bio-mechanical gait analysis \cite{seifert2019toward}.
The typical approach in these works is to transmit sequences of large bandwidth signals (of $2$ to $4$~GHz), with a rate dictated by the desired sensing resolution \cite{patole2017automotive}. Thus, \ac{mmwave} radar sensors have two main drawbacks:
\noindent\textit{(i)} they are specifically tailored to sensing and cannot perform communication. Moreover, their deployment cost is relatively high as one single sensor can reliably cover a range of at most $8-10$~m due to the radial distortion and occlusion problems \cite{feng2020evaluating}. For this reason, ad-hoc radar sensor networks would need to be deployed in practical scenarios. Our method, in contrast, fully exploits existing \ac{mmwave} communication systems with no modifications to the \ac{cir} estimation process or packet structures.
\noindent\textit{(ii)} The fixed chirp transmission interval, which is related to the \ac{md} resolution, requires regular transmissions with continuous channel occupation.
Some works have explored the possibility of randomly subsampling the chirp transmission intervals using compressive sensing \cite{eldar2012compressed} to either save computational resources \cite{stankovic2014compressive} or reduce the effect of unwanted interference \cite{sejdic2018compressive}.
However, these works are based on a radar framework, where the transmission instants can be freely chosen and optimized. SPARCS, instead, reuses the given underlying communication traffic as much as possible and only injects small additional sensing units when necessary.
\noindent \textbf{$\mathbf{60}$~GHz WiFi sensing.} Research interest towards sensing with WiFi devices working in the \ac{mmwave} band has mostly focused on the $60$~GHz IEEE~802.11ad/ay standards \cite{wu2020mmtrack, zhang2020mmeye, pegoraro2021rapid}. These works target various applications, such as person tracking and gesture recognition, exploiting \ac{cir} estimation to detect humans in the environment. However, they require dedicated and regular sensing signal transmissions in order to function properly, entailing a significant overhead and channel utilization for sensing.
Our work significantly improves over the above mentioned studies in that it enables the reuse of randomly distributed communication packets via sparse recovery, whenever possible. This is of key importance to \emph{integrate} sensing capabilities in communication devices while maintaining low overhead and complexity.
\noindent \textbf{Integrated sensing and communication.}
A number of technical works address \ac{isac} systems in next generation $5$G/$6$G cellular networks \cite{liu2021integrated, li2021rethinking} and \acp{wlan} \cite{restuccia2021ieee, pegoraro2021rapid}. Many of those target the joint communication and sensing waveform design \cite{liu2018toward} and are mostly oriented to automotive applications to measure distance and velocity of nearby vehicles.
In contrast, few works focus on human sensing \cite{li2021rethinking}, which is the aim of the present work.
All of the above approaches alternate communication and sensing phases according to a time-division scheme, causing significant overhead and channel occupation. SPARCS instead, provides a full \ac{isac} scheme, as it passively exploits communication traffic while dynamically injecting sensing units to cover silent periods. As a result, our method significantly reduces sensing overhead while at the same time improving the sensing accuracy.
\section{Concluding remarks} \label{sec:conclusion}
In this paper, we have designed and implemented SPARCS, the first \ac{mmwave} \ac{isac} system that can sense human \ac{md} signatures from irregular and sparse \ac{cir} estimates. These are obtained in a standard compliant way by both reusing optional \ac{cir} estimation fields appended to communication packets and sporadically injecting sensing packets whenever communication traffic is absent.
Differently from the existing \ac{isac} methods, SPARCS is based on a sparse recovery approach to the \ac{md} reconstruction, which is theoretically grounded in the instrinsic sparse multi-path environment of the \ac{mmwave} channel. This enables an accurate \ac{md} extraction from a significantly lower number of randomly distributed \ac{cir} samples, thus drastically reducing the sensing overhead. After a \ac{cir} resampling step along the time domain, SPARCS performs an iterative sparse reconstruction in the frequency domain, decoupling different propagation paths at first, to leverage their sparsity property, and then combining them to obtain the final \ac{md} spectrum.
While SPARCS is compatible with different \ac{mmwave} systems (e.g., $3$GPP $5$G-NR, and IEEE $60$~GHz \acp{wlan}), for our implementation we used an IEEE~802.11ay \ac{sdr} platform working in the $60$~GHz band. We tested our system on a large set of standard-compliant \ac{cir} traces matching the traffic patterns of real WiFi access points, performing a typical downstream application such as \ac{har}. Our results show that SPARCS entails over $7$ times lower overhead compared to prior methods, while achieving better performance.
Future research directions include refining SPARCS by leveraging an exchange of information between the sensing and communication processes, e.g., utilizing sensing to optimize the beam selection, and exploiting communications for collaborative sensing.
\section*{Acknowledgments}
This research work was sponsored in part by the European Union’s Horizon 2020 research and innovation programme under grants No. 871249, on ``LOCalization and analytics on-demand embedded in the 5G ecosystem for Ubiquitous vertical applicationS'' (LOCUS) and No. 861222, on ``MIllimeter-wave NeTworking and Sensing for beyond 5G'' (MINTS), by the Spanish Ministry of Science and Innovation (MICIU) grant RTI2018-094313-B-I00 (PinPoint5G+), by the Region of Madrid through TAPIR-CM (S2018/TCS-4496) and by the Italian Ministry of Education, University and Research (MIUR) through the initiative “Departments of Excellence” (Law 232/2016).
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
The $n$- line splitting operation for graphs was introduced by Slater \cite{slt1} and it was extended to binary matroids by Azanchiler \cite{AH1}. Properties of splitting, element splitting, and $es$ -splitting operations on binary matroids are extensively discussed by several authors \cite{AH2,DMS, rsw,ma, shika,ba}.
Malavadkar et al. \cite{mjg, elt} defined splitting and element splitting operations for simple and coloopless $p$-matroids, i.e. for matroids representable over $GF(p).$
\begin{definition}\label{def0}
Let $M\cong M[A]$ be a $p$-matroid on ground set $E,$$\{a,b\} \subset E,$ and $\alpha\neq 0$ in $ GF(p)$. The matrix $A_{a,b}$ is constructed from $A$ by appending an extra row to $A$ which has coordinates equal to $\alpha$ in the columns corresponding to the elements $a,$$b,$ and zero elsewhere. Define the splitting matroid $M_{a,b}$ to be the vector matroid $M[A_{a,b}].$ The transformation of $M$ to $M_{a,b}$ is called splitting operation.
\end{definition}
\begin{definition}\label{def1}
Let $M\cong M[A]$ be a $p$-matroid on ground set $E,$$\{a,b\} \subset E,$ and $M_{a,b}$ be the corresponding splitting matroid. Let the matrix $A_{a,b}$ represents $M_{a,b}$ on $GF(p).$ Construct the matrix $A'_{a,b}$ from $A_{a,b}$ by adding an extra column to $A_{a,b},$labeled as $z,$ which has the last coordinate equal to $\alpha\neq 0$ and the rest are equal to zero. Define the element splitting matroid $M'_{a,b}$ to be the vector matroid $M[A'_{a,b}]$. The transformation of $M$ to $M'_{a,b}$ is called element splitting operation.
\end{definition}
\noindent In this paper, we define $es$-splitting operation on a simple and coloopless $p$-matroid, which is a single element extension of corresponding element splitting $p$-matroid, and study circuits, bases and connectivity of the resulting matroid.
\noindent A binary matroid is Eulerian if and only if it's dual matroid is bipartite. Equivalently, an Eulerian binary matroid can not have an odd cocircuit. The $es$-splitting matroid $M^e_{a,b}$of a bridge-less binary matroid $M$ always contains an odd cocircuit. Therefore $M^e_{a,b}$ can never be Eulerian. But when we consider matroids representable over $GF(p),(p>2),$ the $es$-splitting matroid of it contains an odd cocircuit and still can be Eulerian. In the present paper, we provide necessary and sufficient condition for an $es$-splitting $p$-matroid $M^e_{a,b}$ to be Eulerian. \\
\noindent A sufficient condition to obtain Hamiltonian $p$-matroid from Hamiltonian $p$-matroid under $es$-splitting operation is also provided. For undefined and standard terminologies, refer to \cite{ox}.
\section{$es$-splitting operation on $p$-matroids}
The present section generalizes the $es$-splitting operation to a $p$-matroid $M,$ and characterizes circuits and bases of the $es$-splitting matroid.
\begin{definition}\label{d3}
Let $M\cong M[A]$ be a $p$-matroid on ground set $E$ and $\{a,b\}\subset E$ with $e\in \{a,b\}.$ Let $A'_{a,b}$ represents the element splitting matroid $M'_{a,b}$ on $GF(p).$ Construct the matrix $A^e_{a,b}$ by adjoining an extra column $\gamma$ to the matrix $A'_{a,b}$ where $\gamma$ is subtraction$(e-z)$ of two column vectors corresponding to the elements $e$ and $z.$ Denote $M^e_{a,b}\cong M[A^e_{a,b}].$ The shift of $M$ to $M^e_{a,b}$ is called an $es$-splitting operation. We call the matroid $M^e_{a,b}$ as $es$-splitting $p$-matroid.
\end{definition}
\begin{remark}
For $p=2,$ the above definition coincides with the $es$-splitting operation on binary matroids \cite{AH1}.
\end{remark}
\begin{remark}
rank$(A)<$ rank$(A^e_{a,b})=$ rank$(A)+1.$ If the rank functions of $M$ and $M^e_{a,b}$ are denoted by $r$ and $r',$respectively, then $r(M) < r'(M^e_{a,b})= r(M) + 1.$
\end{remark}
\noindent Let $C_k=\{u_1, u_2,\ldots, u_l : u_i \in E, i=1,2,\dots,l \}$ be a circuit of $M$ for some positive integers $k,$ $l$ and $|C_k \cap \{a,b\}|=2.$ Suppose, without loss of generality, $u_1=a$,$u_2=b$. If there are non-zero constants $a_1^k, a_2^k, ..., a_l^k$ in $GF(p)$ such that $\sum_{i=1}^{i=l} a_i^k u_i \equiv 0 (mod~p)$ and $a_1^k + a_2^k \equiv 0 (mod~p),$ then we call $C_{k}$ a $p$-circuit of $M.$ However, if $|C_k \cap \{a,b\}|=1$ or $|C_k \cap \{a,b\}|=2$ and $a_1^k + a_2^k\not\equiv 0 (mod \: p),$ then we call $C_{k}$ an $np$-circuit of $M$.
\noindent We denote $\mathcal {C}_0=\{C\in \mathcal {C}(M): C$ is a $p$-circuit or $C\cap \{a,b\}=\phi\}$.
\begin{theorem}\label{circuits}
Let $M=(E,\mathcal{C})$ be a $p$-matroid, and $M'_{a,b}$ be its element splitting $p$-matroid. $e\in \{a,b\}$ and $z, \gamma \notin E$. Then $\mathcal{C}(M^e_{a,b})= \mathcal{C}(M'_{a,b}) \cup \mathcal{C}_4 \cup \mathcal{C}_5 \cup \mathcal{C}_6 \cup \mathcal{C}_7 \cup \mathcal{C}_8\cup \Delta,$ where
\begin{description}
\item $\mathcal {C}_4=\{C\cup \{e, \gamma\}: C$ is an $np$-circuit and $e\notin C\}$;
\item $\mathcal {C}_5=\{(C\setminus e)\cup \gamma: C$ is an $np$-circuit, $e\in C$ and $a\notin C\}$;
\item $\mathcal {C}_6=\{(C\setminus e)\cup \{z, \gamma \}: C$ is a $p$-circuit and $e\in C\}$;
\item $\mathcal {C}_7=\{C \cup \gamma : C$ is an $np$-circuit containing $e$ and $a$\};
\item $\mathcal {C}_8=\{(C\setminus e) \cup \{z, \gamma\} : C$ is an $np$-circuit containing $e$ and $a$\} and
\item $\Delta=\{e, z, \gamma\}.$
\end{description}
\end{theorem}
\begin{proof}
By the Definition \ref{d3} and $M'_{a,b}=M^e_{a,b}|_{E\cup z},$ the inclusion $ \mathcal{C}(M'_{a,b}) \cup \mathcal{C}_4 \cup \mathcal{C}_5 \cup \mathcal{C}_6 \cup \mathcal{C}_7 \cup \mathcal{C}_8\cup \Delta \subset \mathcal{C}(M^e_{a,b})$ follows.
\noindent To prove the other inclusion, assume $C \in \mathcal{C}(M^e_{a,b})$ and $\gamma \notin C.$ Then by Theorem 2.4 of \cite{elt}, $C \in \mathcal{C}(M'_{a,b}).$ Now assume that $\gamma \in C.$\\
$\mathbf{Case~1:}$ $z \in C.$ Note that if $e\in C,$ then $C=\Delta=\{e, z, \gamma\}$. Suppose $e \notin C.$ If $C_1= C\setminus\{z,\gamma\} $ is a dependent set of $M$ containing an $np$-circuit, say $C_2.$ Then $C_2 \cup z$ is a circuit of $M^e_{a,b}$ contained in $C,$ which is not possible. Therefore $C_1$ is an independent set of $M.$
Since the every coordinate of $e$ is equal to the corresponding coordinate of $\gamma,$ and all the coordinates of $z$ except the last one are zero, the set $C_1 \cup e $ is a dependent set of $M.$ Moreover $C_1 \cup e $ is a circuit of $M,$ otherwise $C_1 \cup e $ contains an $np$-circuit, say $C_3.$ And $C_3 \cup z $ is circuit of $M^e_{a,b}$ contained in $C$ which is not possible. If $C_1 \cup e $ is a $p$-circuit, then $C \in \mathcal {C}_6.$ And if $C_1 \cup e $ is an $np$-circuit, then $C \in \mathcal {C}_8.$
\noindent $\mathbf{Case~2:} $ $z\notin C$\\
\noindent$\mathbf{Subcase~(i):}$ $e \notin C\setminus \gamma.$ Then observe that $a$ is also not an element of $C.$ And by similar argument as in Case 1, $(C\setminus \gamma)\cup e$ is an $np$-circuit of $M.$ Therefore $C\in \mathcal{C}_5.$ \\
\noindent $\mathbf{Subcase~(ii):}$ $e \in C\setminus \gamma$. Note that $C\setminus\gamma$ is a dependent set of $M,$ containing $a$ and one of the following cases will occur.\\
$\mathbf{(a)}$ If $C\setminus \gamma$ is a circuit of $M,$ then it must be an $np$-circuit. Therefore $C\in \mathcal{C}_7.$\\
\noindent $\mathbf{(b)}$ If $C\setminus \gamma$ is not a circuit of $M,$ and $C\setminus \gamma=C_{np}\cup e,$ for some $np$-circuit $C_{np}$ excluding $e.$ Then $C_{np} \cup \{e, \gamma\}=C.$ In this case $C\in \mathcal{C}_4.$ \\
\noindent $\mathbf{(c)}$ If $C\setminus \gamma$ is not a circuit of $M,$ and $C\setminus \gamma=C_{np}\cup A,$ for some $np$-circuit $C_{np}$ excluding $e,$ and disjoint from $A \subset E(M),$ where $|A|\geq 2.$ Then $C_{np}\cup \{e, \gamma\}$ is the circuit, of type $\mathcal{C}_4,$ of $M^e_{a,b}$ contained in $C,$ a contradiction.\\
\noindent $\mathbf{(d)}$ If $C\setminus \gamma$ is not an $np$-circuit of $M,$ and $C\setminus \gamma=C_{np}\cup A,$ for some $np$-circuit $C_{np}$ containing $e,$ and $A\subset E(M).$ If $a \notin C_{np},$ then $(C_{np}\setminus e) \cup \gamma $ is a circuit of type $\mathcal{C}_5$ contained in $C$, a contradiction. And if $a \in C_{np},$ then $C_{np} \cup \gamma, $ a circuit of type $\mathcal{C}_7,$ is contained in $C,$ which is not possible.\\
Therefore we conclude that $C\in (\mathcal{C}(M'_{a,b}) \cup \mathcal{C}_4 \cup \mathcal{C}_5 \cup \mathcal{C}_6 \cup \mathcal{C}_7 \cup \mathcal{C}_8).$
\end{proof}
\begin{corollary}
The $es$-splitting matroid $M^e_{a,b}$ is not a bipartite matroid.
\end{corollary}
\begin{proof}
The proof follows from the fact, the $es$-splitting matroid $M^e_{a,b}$ contains the circuit $\Delta=\{e,z,\gamma\}$ of size three.
\end{proof}
\noindent The following example lists all the type of circuits of a $p$-matroid $M,$ and the corresponding $es$-splitting matroid $M^e_{a,b}.$
\begin{example}\label{ex1}
Consider the ternary sparse paving matroid $P_8$ which has the representation given by matrix $A.$
\begin{center}
$\mathbf{A} =
\begin{pNiceMatrix
[first-col,
first-row,
code-for-first-col = \color{blue},
code-for-first-row = \color{blue}]
& 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 \\
& 1 & 0 & 0 & 0 & 0 & 1 & 1 & 2 \\
& 0 & 1 & 0 & 0 & 1 & 0 & 1 & 1 \\
& 0 & 0 & 1 & 0 & 1 & 1 & 0 & 1 \\
& 0 & 0 & 0 & 1 & 2 & 1 & 1 & 0 \\
\end{pNiceMatrix} \qquad
\mathbf{A^e_{1,4}} =
\begin{pNiceMatrix
[first-col,
first-row,
code-for-first-col = \color{blue},
code-for-first-row = \color{blue}]
& 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\
& 1 & 0 & 0 & 0 & 0 & 1 & 1 & 2 & 0 & 0 \\
& 0 & 1 & 0 & 0 & 1 & 0 & 1 & 1 & 0 & 0 \\
& 0 & 0 & 1 & 0 & 1 & 1 & 0 & 1 & 0 & 0 \\
& 0 & 0 & 0 & 1 & 2 & 1 & 1 & 0 & 0 & 1 \\
& 1 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 0 \\
\end{pNiceMatrix}$
\end{center}
\noindent For $a=1$, $b=4,$ $e=4$ and $\alpha=1$ the representation of $es$-splitting matroid $M^e_{1,4}$ over $GF(3)$ is given by the matrix $A^e_{1,4}$.
The collection of circuits of $M,$ $M_{1,4}$ and $M^e_{1,4}$ is given in the following table.
\begin{longtable}{ | m{4.4cm}| m{4.5cm} | m{5.5cm} |}
\hline
\textbf{~~~~~~~~~Circuits of $M$} & \textbf{~~~~~~~~Circuits of $M_{1,4}$} & \textbf{~~~~~~~~Circuits of $M^e_{1,4}$}\\
\hline
$\{1, 2, 4, 5, 6\},$ $\{1, 2, 4, 6, 8\},$ $\{1, 3, 4, 5, 7\},$ $\{1, 3, 4, 7, 8\},$ $\{1, 4, 5, 8\},$ $\{2, 3, 6, 7\},$ $\{2, 5, 6, 8\},$~~~~~~~~~ $\{3, 5, 7, 8\}$& $\{1, 2, 4, 5, 6\},$ $\{1, 2, 4, 6, 8\},$ $\{1, 3, 4, 5, 7\},$ $\{1, 3, 4, 7, 8\},$ $\{1, 4, 5, 8\},$ $\{2, 3, 6, 7\},$ $\{2, 5, 6, 8\},$~~~~~~~~ $\{3, 5, 7, 8\}$ & $\{1, 2, 4, 5, 6\},$ $\{1, 2, 4, 6, 8\},$ $\{1, 3, 4, 5, 7\},$ $\{1, 3, 4, 7, 8\},$ $\{1, 4, 5, 8\},$ $\{2, 3, 6, 7\},$ $\{2, 5, 6, 8\},$~~~~~~~~~~~~~~ $\{3, 5, 7, 8\}$ \\
\hline
$\{1, 2, 3, 5, 6\},$ $\{1, 2, 3, 5, 7\},$ $\{1, 2, 3, 8\},$ $\{1, 2, 4, 7\},$ $\{1, 2, 5, 7, 8\},$ $\{1, 2, 6, 7, 8\},$ $\{1, 3, 4, 6\},$ $\{1, 3, 5, 6, 8\},$ $\{1, 3, 6, 7, 8\},$ $\{1, 5, 6, 7\},$ $\{2, 3, 4, 5\},$ $\{2, 3, 4, 6, 8\},$ $\{2, 3, 4, 7, 8\},$ $\{2, 4, 5, 6, 7\},$ $\{2, 4, 5, 7, 8\},$ $\{3, 4, 5, 6, 7\},$ $\{3, 4, 5, 6, 8\},$ ~~~~~~$\{4, 6, 7, 8\}$ & $-$ & $\{1, 2, 3, 5, 6, 9\},$ $\{1, 2, 3, 5, 7, 9\},$ $\{1, 2, 3, 8, 9\},$ $\{1, 2, 4, 7, 9\},$ $\{1, 2, 5, 7, 8, 9\},$ $\{1, 2, 6, 7, 8, 9\},$ $\{1, 3, 4, 6, 9\},$ $\{1, 3, 5, 6, 8, 9\},$ $\{1, 3, 6, 7, 8, 9\},$ $\{1, 5, 6, 7, 9\},$ $\{2, 3, 4, 5, 9\},$ $\{2, 3, 4, 6, 8, 9\},$ $\{2, 3, 4, 7, 8, 9\},$ $\{2, 4, 5, 6, 7, 9\},$ $\{2, 4, 5, 7, 8, 9\},$ $\{3, 4, 5, 6, 7, 9\},$ $\{3, 4, 5, 6, 8, 9\},$ ~~~~~~$\{4, 6, 7, 8, 9\}$ \\
\hline
$-$ & $-$ & $\{1, 2, 3, 4, 8, 10\},$~~~ $\{1, 4, 5, 6, 7, 10\}$\\
\hline
$-$ & $-$ & $\{2, 3, 5, 10\},$ $\{2, 3, 6, 8, 10\},$ $\{2, 3, 7, 8, 10\},$ $\{2, 5, 6, 7, 10\},$ $\{2, 5, 7, 8, 10\},$ $\{3, 5, 6, 7, 10\},$ $\{3, 5, 6, 8, 10\},$~~~~~~~~ $\{6, 7, 8, 10\}$\\
\hline
$-$ & $-$ & $\{1, 2, 5, 6, 9, 10\},$ $\{1, 2, 6, 8, 9, 10\},$ $\{1, 3, 5, 7, 9, 10\},$~~~~~$\{1, 3, 6, 9, 10\},$ $\{1, 3, 7, 8, 9, 10\},$~~~~~ $\{1, 5, 8, 9, 10\}$\\
\hline
$-$ & $-$ & $\{1, 2, 4, 7, 10\},$~~~~~~~ $\{1, 3, 4, 6, 10\}$\\
\hline
$-$ & $-$ & $\{1, 2, 7, 9, 10\}$\\
\hline
$-$ & $-$ & $\{4, 9, 10\}$\\
\hline
\end{longtable}
\end{example}
\noindent Consider subsets of $E$ of the type $C\cup I$ where $C=\{u_1,u_2,\ldots,u_l\}$ is an $np$-circuit of $M,$ which is disjoint from an independent set $I=\{v_1,v_2,\ldots,v_k\}$ and $\{a,b\}\subset (C\cup I)$. We say $C\cup I$ is $p$-dependent if it contains no member of $\mathcal {C}_0$ and there are non-zero constants $\alpha_1,\alpha _2,\ldots,\alpha_l$ and $\beta_1, \beta_2,\ldots,\beta_k$ such that $\sum_{i=0}^{l}\alpha_i u_i + \sum_{j=0}^{k}\beta_j v_j = 0 (mod \: p)$ and $coeff.(a)+coeff.(b) = 0(mod \: p).$\\
\noindent We use $cl$ and $cl'$ to denote closure operators on $M$ and $M^e_{a,b},$ respectively.
\begin{theorem}\label{basis}
Let $M$ be a $p$-matroid and $M^e_{a,b}$ be es-splitting matroid. $\mathcal{B}(M^e_{a,b})=\mathcal{B}(M'_{a,b})\cup \mathcal{B}_1\cup \mathcal{B}_2\cup \mathcal{B}_3,$ where
\begin{description}
\item $\mathcal{B}_1=\{B\cup \gamma:$ $B\in \mathcal{B}(M)$ and $e\in B$\};
\item $\mathcal{B}_2=\{B\cup \gamma:$ $B\in \mathcal{B}(M),$ $a\in B$ and $e\notin$ cl$(B\setminus a)$ \};
\item $\mathcal{B}_3=\{(C_{np}\cup I)\cup \gamma:$ $(C_{np}\cup I)$ is not a $p$-dependent set of $M,$ $rank(C_{np} \cup I)=rank(M),$ $e\notin cl(C_{np} \cup I)$\};
\item $\mathcal{B}_4=\{I\cup\{z,\gamma\}:$ $I\in \mathcal{I}(M),$ $e \notin cl(I),$$|I|=r(M)-1$ \}.
\end{description}
\end{theorem}
\begin{proof}
Let $B'\in \mathcal{B}(M^e_{a,b}).$ Then $|B'|=r(M)+1.$\\
$\mathbf{Case~1:}$ $z,\gamma \notin B'.$ Then $B' \in \mathcal{B}(M_{a,b}).$\\
$\mathbf{Case~2:}$ $z \in B',$$ \gamma \notin B'.$ Then $B' \in \mathcal{B}(M'_{a,b}).$\\
$\mathbf{Case~3:}$ $z \notin B',$$ \gamma \in B'.$ Then $B'\setminus\gamma$ is an independent set of $M'_{a,b}.$ Denote $B'\setminus\gamma=B.$ By Lemma 2.6 of \cite{elt}, following two subcases are possible:\\
$\mathbf{Subcase~(i):}$ $B \in \mathcal{I}(M).$ Since $|B|=r(M), B\in \mathcal{B}(M).$ Note that $a\in B$ or $e \in B.$ Otherwise $\gamma \in cl'(B),$ which is a contradiction to the fact that $B'$ is basis of $M^e_{a,b}.$\\
$\mathbf{(a)}$ $e \in B.$ Then $B'\in \mathcal{B}_1.$\\
$\mathbf{(b)}$ $e\notin B.$ If $e \notin cl(B\setminus a),$ then $B'\in \mathcal{B}_2.$ Suppose $e\in cl(B\setminus a),$ there are column vectors in $B\setminus a,$ say $v_1,v_2, \ldots, v_j,$ such that $\{e,v_1,v_2, \ldots, v_j\}$ forms a circuit in $M.$ Since the last coordinate of $v_1,v_2, \ldots, v_j,$ and $\gamma$ in $M^e_{a,b}$ are zero, and the coordinates of $e$ and the corresponding coordinates of $\gamma$ are identical implies $\{\gamma, v_1,v_2, \ldots, v_j\}$
is dependent in $M^e_{a,b}.$ Thus $B'$ contains a dependent set, which is a contradiction. \\
\noindent $\mathbf{Subcase~(ii):}$ $B = (C_{np}\cup I),$ where $(C_{np}\cup I)$ is not $p$ -dependent and it contains no union of two disjoint $np$-circuits.
Let $e\in(C_{np}\cup I).$ Then either $e\in C_{np}$ or $e\in I.$ If $e \in C_{np},$ then $(C_{np}\setminus e)\cup \gamma$ is a circuit, of type $ \mathcal{C}_7,$ contained in $B',$ a contradiction. Therefore $e \notin C_{np}.$ If $e \in I,$ then $C_{np} \cup \{e,\gamma\}$ is a circuit, of type $ \mathcal{C}_4,$ contained in $B',$ again a contradiction. Therefore $e \notin I.$ Furthermore $e \notin cl(C_{np} \cup I),$ otherwise $(C_{np} \cup I) \cup e$ contains a circuit, say $C,$ and $e \in C.$ Note that $C$ is an $np$-circuit. If $C$ contains $a,$ then $C \cup \gamma $ is a circuit of type $\mathcal{C}_7,$ contained in $B'$, a contradiction. If $a \notin C,$ then $(C\setminus e) \cup \gamma$ is a circuit, of type $\mathcal{C}_5$, contained in $B',$ a contradiction. Thus, $B' \in \mathcal{B}_3.$ \\
$\mathbf{Case~4:}$ $z \in B',$$ \gamma \in B'.$ As earlier denote $B'\setminus\gamma=B.$ Note that $e\notin B',$ otherwise $B'$ contains the circuit $\{e,z,\gamma\}.$ Observe that $B $ is independent in $M'_{a,b}.$ By Lemma 2.6 of \cite{elt}, following three are possible subcases:\\
$\mathbf{Subcase~(i):}$ $ B = I \cup z,I \in \mathcal{I}(M)$ and $|I|=r(M)-1.$ In this case, $B'\in \mathcal{B}_4.$ \\
$\mathbf{Subcase~(ii):}$ $B \in \mathcal{I}(M).$ Thus $B'= B \cup \gamma.$ In this case, $B'\in \mathcal{B}_1$ or $B'\in\mathcal{B}_2$\\
$\mathbf{Subcase~(iii):}$ $B = C_{np} \cup I,$ where $(C_{np}\cup I)$ is not $p$ -dependent and it contains no union of two disjoint $np$-circuits. Then $C_{np}\cup z$ is a circuit of $M^e_{a,b}$ contained in $B',$ a contradiction.
Therefore $B'\in (\mathcal{B}(M'_{a,b})\cup \mathcal{B}_1\cup \mathcal{B}_2\cup \mathcal{B}_4).$\\
\noindent Conversely, let $X\in (\mathcal{B}(M'_{a,b})\cup \mathcal{B}_1\cup \mathcal{B}_2\cup \mathcal{B}_3).$ We prove that $X$ is a basis of $M^e_{a,b}.$\\
If $X\in \mathcal{B}(M'_{a,b}),$ then $|X|=r(M)+1.$ Moreover $X$ is independent in $M^e_{a,b},$ therefore $X$ is basis of $M^e_{a,b}.$ If $X \in \mathcal{B}_1,$ then $X=B\cup \gamma,$ for some $B\in \mathcal{B}(M)$ containing $e.$ Let $B = \{u_1,u_2,\ldots,u_l\}.$ Note that $B$ is independent in $M^e_{a,b}.$ If $X=B\cup \gamma$ is dependent in $M^e_{a,b},$ then it contains a circuit, say $C,$ and $\gamma \in C.$ Following are the three possibilities regarding $C$:\\
$\mathbf{Case ~1}$ $C= B\cup \gamma = \{u_1,u_2,\ldots,u_l,\gamma\}.$ Since $e\in B,$ then $a$ must be in $B.$ Assume $u_1=a, u_2=e,$ WLOG. There exists non-zero scalars $\alpha_1,\alpha_2,\ldots,\alpha_l,\alpha_{\gamma},$ such that $\alpha_1 u_1+ \alpha_2 u_2+ \ldots+\alpha_l u_l+ \alpha_{\gamma} u_{\gamma} \equiv 0(mod\:p).$ Since every coordinate of $e$ is equal to the corresponding coordinate of $\gamma,$ restricting the column vectors of $B$ to the matrix $A,$ we get $\alpha_1 u_1+ (\alpha_2 + \alpha_{\gamma}) u_2+ \ldots+\alpha_l u_l \equiv 0(mod\:p).$ Therefore $B= \{u_1,u_2,\ldots,u_l\}$ is a dependent set of $M,$ a contradiction.\\
$\mathbf{Case ~2}$ $C \subset B \cup \gamma,$ and $a,e \in C \setminus \gamma.$ Using the arguments as in $\mathbf{(a)},$ we get a contradiction. \\
$\mathbf{Case ~3}$ $C \subset B \cup \gamma,$ and $a,e \notin C .$ Using the arguments as in $\mathbf{(a)},$ we get a contradiction.\\
Next, assume $X \in \mathcal{B}_2.$ Then $X=B\cup \gamma,$ where $B\in \mathcal{B}(M),$ $a\in B$ and $e\notin$ cl$(B\setminus a).$ Note that $B$ is independent in $M^e_{a,b}.$ If $X=B\cup \gamma$ is dependent in $M^e_{a,b},$ then it contains a circuit, say $C,$ and $\gamma \in C.$ Then $C \in \mathcal{C}_5$ or $C \in \mathcal{C}_7.$ Since $e\notin cl(B\setminus a),$ and hence $ e\notin C.$ Therefore $C$ cannot be a member of $\mathcal{C}_5$ and $\mathcal{C}_7.$ \\
Now if $X \in \mathcal{B}_3,$ then $X = (C_{np}\cup I)\cup \gamma $ for some $np$-circuit $C_{np},$ independent set $I$ of $M.$ Since $rank(C_{np} \cup I)=rank(M),$ and $(C_{np}\cup I)$ is not a $p$-dependent set of $M,$ implies $|(C_{np}\cup I)|=rank(M),$ and it is independent in $M^e_{a,b}.$ Therefore $|X|=rank(M)+1.$ Here it is enough to prove that $X=(C_{np}\cup I)\cup \gamma$ is independent in $M^e_{a,b}.$ On the contrary assume that $X$ is not independent in $M^e_{a,b}.$ Then $X$ contains a circuit, say $C,$ passing through $\gamma.$ Note that $e,z \notin C.$ Therefore $C$ is a circuit of type $\mathcal{C}_5.$ This implies $e \in cl(C) \subset cl(C_{np}\cup I),$ a contradiction again. \\
Let $X \in \mathcal{B}_4.$ Then $X=I \cup \{z,\gamma\},$ where $I \in \mathcal{I}(M).$ By Lemma 2.6 of \cite{elt} $I'=I \cup z$ is an independent set of $M'_{a,b},$ and hence of $M^e_{a,b}.$ Therefore it is enough to show that $I' \cup \gamma $ is an independent set of $M^e_{a,b}.$ If not, then $I' \cup \gamma$ contains a circuit $C,$ such that $\gamma \in C.$ Since $e \notin X,$ $C$ cannot be a circuit of the type $\mathcal{C}_4,\mathcal{C}_7$ and $\Delta=\{e,z, \gamma\}.$ Assume $C \in \mathcal{C}_5.$ Then $C=(C'\setminus e)\cup \gamma,$ where $C'$ is an $np$-circuit containing $e,$ and $C'\subset I.$ Thus $e \in cl(I),$ a contradiction. We get the same contradiction, if $C \in \mathcal{C}_6, \mathcal{C}_8.$ Therefore $I'\cup \gamma $ is an independent set of size $rank(M)+1.$
\end{proof}
\noindent The next theorem gives the rank function of a $es$-splitting matroid $M^e_{a,b}$ in terms of the rank function of corresponding $p$-matroid $M.$
\begin{theorem}
Let $r$ and $r'$ denote the rank functions of the $p$-matroids $M$ and $M^e_{a,b}$, respectively. Let $X \subseteq E(M).$ Then
\begin{equation}\label{r1}
r'(X)= \begin{cases}
r(X)+1, \text{if}~ X ~ \text{contains an}~ np-\text{circuit;}~~~ \\
r(X), \text{otherwise .}
\end{cases}
\end{equation}
\begin{equation}\label{r2}
r'(X\cup z) = r(X)+1~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\end{equation}
\begin{equation}\label{r3}
~~~~~~~~~~~~~~~~~~~~~~~~r'(X~\cup ~\gamma) = \begin{cases}
r(X), \text{not }X ~\text{but}~ X\cup e ~\text{contains an np-circuit;}\\
r(X)+2, \text{if}~ e\notin cl(X)\text{and}~ X \text{ contains an np circuit;}\\
r(X)+1, \text{otherwise.}
\end{cases}
\end{equation}
\begin{equation}\label{r4}
r'(X~\cup ~\{z,\gamma\}) = \begin{cases}
r(X)+1, \text{if}~ e\in cl(X);\\
r(X)+2, \text{if}~ e\notin cl(X);~~~~~~~~~~
\end{cases}
\end{equation}
\end{theorem}
\begin{proof}
Equation (\ref{r1}) and (\ref{r2}) are proved in \cite{mjg} and \cite{elt}.\\
\noindent To prove (\ref{r3}), consider not $X,$ but $X\cup e $ contains an $np$-circuit of $M$, say $C_{np}.$ Then $e \in C_{np}.$ Now $C_{np}\setminus e \subset X.$ In $X\cup \gamma,$$ C_{np}\setminus e$ forms a circuit, of type $\mathcal{C}_5,$ with $\gamma.$ Therefore $\gamma \in cl'(X).$ Thus $r'(X\cup \gamma)=r(X).$ \\
\noindent Next, assume that $e\notin cl(X)$ and $X$ contains an $np$-circuit of $M$, say $C_{np}.$ Now by \ref{r1}, $r'(X)=r(X)+1.$ We claim that $\gamma \notin cl'(X).$ On the contrary assume that $\gamma \in cl'(X).$ Then there exists a circuit $ C \subset X \cup \gamma$ in $M^e_{a,b},$ containing $\gamma.$ Such a circuit must be of the type $\mathcal{C}_4,\mathcal{C}_5$ or $\mathcal{C}_7.$ Therefore $e \in cl(C\setminus\gamma),$ that implies $e \in cl(X),$ a contradiction. Therefore $r'(X\cup \gamma) =(r(X)+1)+1= r(X)+2.$\\
\noindent Suppose $X$ contains an $np$-circuit $C_{np},$ and $e \in cl(X).$ Then there exists a circuit,say $C,$ containing $e.$ Such a circuit could be $p$ or $np$ circuit. In case when $C$ is an $np$-circuit, containing $e$ then we get a circuit of type $\mathcal{C}_5$ or $\mathcal{C}_7$ of $M^e_{a,b}$. Therefore $r'(X\cup \gamma)=r(X)+1.$ If $C$ is a $p$ circuit, then $\gamma$ cannot form a circuit with $C,$ since for such a circuit to form we need $z\in X\cup \gamma.$ Thus $\gamma \notin cl'(X)$ in $M^e_{a,b}.$ Therefore $r'(X\cup \gamma)=r(X)+1.$ \\
\noindent Next, assume that $X$ contains no $np$-circuit in $M,$ and $e\in cl(X).$ If $X\cup e$ contains an $np$-circuit, then as discussed earlier $r'(X\cup \gamma)=r(X).$ If $X\cup e$ contains a $p$-circuit, then $\gamma \notin cl'(X),$ as $z \notin X\cup \gamma.$ Therefore $r'(X\cup \gamma)=r(X)+1.$ Further if $e \notin cl(X),$ then $\gamma \notin cl'(X).$ Therefore $r'(X\cup \gamma)=r(X)+1.$ \\
\noindent To prove (\ref{r4}), consider $e \in cl(X)$ in $M.$ Note that $r'(X \cup z)=r(X)+1.$ Since $e \in cl(X), X \cup\{z,\gamma\}$ contains a circuits of the type $\Delta, \mathcal{C}_6,$ or $\mathcal{C}_8.$ Thus $ r'(X \cup\{z,\gamma\})=r'( X \cup z)=r(X)+1.$ If $e \notin cl(X),$ then $ \gamma \notin cl'(X\cup z);$ otherwise there exist a circuit $C\subset X\cup z,$ containing $\gamma,$ for such a circuit to exists $e$ must be in $cl(C\setminus\{z,\gamma\})$ in $M.$ Therefore $e\in X,$ a contradiction. Thus $r'(X\cup \{z,\gamma\})=r(X)+2.$\\
\end{proof}
\noindent Equation (\ref{r3}) and equation (\ref{r4}) can be verified with Ex.\ref{ex1} as follows:
\begin{enumerate}
\item $r'(\{2,3,5\}\cup 10)=3=r(\{2,3,5\}).$
\item $r'(\{1,2,3,8\}\cup 10)=5=r(\{1,2,3,8\})+2.$
\item $r'(\{1,2,3,5,6\}\cup 10)=5=r(\{1,2,3,5,6\})+1.$
\item $r'(\{1,2,4,7\}\cup \{9,10\})=4=r(\{1,2,4,7\})+1.$
\item $r'(\{1,2,3,8\}\cup \{9,10\})=5=r(\{1,2,3,8\})+2.$
\end{enumerate}
\begin{corollary}
In the $es$-splitting matroid $M^e_{a,b},$ $\{a,e,z\}$ forms a cocircuit.
\end{corollary}
\begin{proof}
Let $X = E(M^e_{a,b})\setminus \{a,e,z\}.$ We will show that $X$ is a hyperplane of $M^e_{a,b}.$ First note that the last coordinate of every vector in $X$ is zero and that of $a,e,$ and $z$ is $1.$ Therefore, no linear combination of vectors in $X$ yields $a,e,$ or $z.$ Thus, $a,e,z \notin cl'(X).$ Equivalently $cl'(X)= X.$ Next, we show that $r'(X)=r(M).$ As noted earlier that every vector in $X$ has last coordinate equal to zero, and every coordinate of $\gamma$ is equal to the corresponding coordinate of $e,$ therefore $X= E(M)\setminus a,$ and $r'(X)=r(X).$ Since $M$ is a coloopless, there exist a basis $B$ of $M$ such that $a \notin B.$ Then $r'(X)=r(X)=|B|=r(M).$ Therefore $X$ is a hyperplane of $M^e_{a,b}.$\
\end{proof}
\section{Connectivity of $es$-splitting $p$-matroids}
In this section, we show that the $es$-splitting operation on $p$-matroids preserves connectivity and 3-connectedness.
Let $M$ be a matroid having ground set $E,$ and $k$ be a positive integer. The $k$-separation of matroid $M$ is a partition $\{S, T\}$ of $E$ such that $|S|, |T|\geq k$ and $r(S) + r(T)- r(M) <k.$ For an integer $n \geq 2,$ we say $M$ is $n$-connected if $M$ has no $k$- separation, where $1 \leq k \leq n-1.$
\begin{lemma}
If $M$ is a connected $p$-matroid, then the corresponding $es$-splitting matroid $M^e_{a,b}$ is also connected.
\end{lemma}
\begin{proof}
As $M$ is a connected $p$-matroid, by Theorem 3.1 of \cite{elt} the element splitting matroid $M'_{a,b}$ is also connected. So it is enough to prove that for every element $x\in E\cup z,$ there exists a circuit passing through $x$ and $\gamma.$\\
$\mathbf{Case~1:}$ Suppose $x=z.$ Then $\Delta=\{e, z, \gamma\}$ is the required circuit.\\
$\mathbf{Case~2:}$ Suppose $x\in E.$ If $x=e,$ then the circuit $\Delta=\{e, z, \gamma\}$ contains $x$ and $\gamma.$ Now assume $x \notin e,$ since $M$ is a connected matroid, there exist a circuit, say $C,$ passing through $x$ and $e.$ Such a circuit $C$ is a $p$-circuit or $np$-circuit. If $C$ is a $p$-circuit, then $(C\setminus e) \cup \{z,\gamma\}$ is the desired circuit passing through $x$ and $\gamma.$ Next assume $C$ is an $np$-circuit. Then the desired circuit passing through $x$ and $\gamma$ will be a member of one of the collection $\mathcal{C}_5,$ $\mathcal{C}_7$ or $\mathcal{C}_8.$
\end{proof}
\noindent We use the following proposition to prove the next theorem:
\begin{proposition}\cite{ox}
If a matroid $M$ is $n$-connected matroid and $|E(M)|>2(n-1)$, then all circuits and all cocircuits of $M$ have at least $n$ elements.
\end{proposition}
\begin{theorem}
Let $M$ be a 3-connected $p$-matroid and $|E(M)|>4$. Then $M^e_{a,b}$ is also a 3-connected matroid.
\end{theorem}
\begin{proof}
We will show that $M^e_{a,b}$ cannot have a 1 and 2-separation. As $M^e_{a,b}$ is connected, it cannot have a 1-separation. Next assume that $M^e_{a,b}$ has a 2-separation $X,Y$ where $X,Y\subset E\cup\{z,\gamma\},$ $|X|,|Y|\geq 2,$ and
\begin{equation}\label{eq1}
r'(X)+r'(Y)-r'(M^e_{a,b})<2
\end{equation}
$\mathbf{Case~1:}$ $ z \in X$ and $\gamma \in Y$ \\
Let $X'=X\setminus z$ and $Y'=Y\setminus \gamma$\\
\noindent$\mathbf{Subcase~(i):}$ $|X|=2$ and $X=\{x,z\},$ where $x\in E(M).$ Then by equation (\ref{r2}), $r'(X)=r(X')+1.$ As $M$ is $3$-connected, no two element of $M$ are in series, which guarantees that there is a circuit containing $a$ or $e$ but not both. Such a circuit is an $np$-circuit in $Y$ and hence $r'(Y)=r(Y')+1,$ by equation (\ref{r1}) Therefore equation (\ref{eq1}) becomes $r(X')+r(Y')-r(M)<1,$ a contradiction to the fact that $M$ is $3$-connected.\\
$\mathbf{Subcase~(ii):}$ $|Y|=2$ and $Y=\{x,\gamma\},$ where $x\in E(M).$ Observe that $r'(X)= r(X')+1$ as $z\in X$, and by equation (\ref{r3}) $r'(Y)= r(Y')+1.$ In this case equation (\ref{eq1}) becomes $r(X')+r(Y')-r(M)<1,$ which implies $(X',Y')$ forms a $1$-separation of $M.$ A contradiction again.\\
$\mathbf{Subcase~(iii):}$ $|X|,|Y|>2$\\
Then $|X'|, |Y'|\geq 2$ and by equation (\ref{r2}) $r'(X) = r(X')+1,$ and by equation (\ref{r3}) $r'(Y)\geq r(Y').$ Therefore equation (\ref{eq1}) yields, $r(X')+r(Y')-r(M) < 2.$ That is $(X',Y')$ forms a $2$-separation of $M;$ a contradiction.\\
$\mathbf{Case~2:}$ $\{z,\gamma\}\subseteq X$\\
Let $X'=X\setminus\{z,\gamma\}$\\
$\mathbf{Subcase~(i):}$ $X=\{z,\gamma\}.$ Then $Y = E(M).$ Note that $r'(X)=2,$ and $r(Y)=r(M).$ Now equation (\ref{eq1}) gives, $2+ (r(Y)+1)-(r(M)+1)<2.$ Therefore, $r(Y)<r(M),$ which is not possible. \\
$\mathbf{Subcase~(ii):}$ $X=\{z,\gamma, x\}$ where $x\in E(M)$ \\
Here $X'=\{x\}$ we have\\
\begin{center}
$r'(X) = \left\{ \begin{array}{rcl}
r(X')+1, & \mbox{~~if~$e \in cl(X')$}
\\ r(X')+2, & \mbox{~~if~$e \notin cl(X')$}
\end{array}\right. $
\end{center}
If $x=e,$ then $r'(X)=r(X')+1.$ Again using the fact that there is an $np$-circuit excluding $e$ contained in $Y,$ we have $r'(Y)=r(Y)+1.$ Thus equation (\ref{eq1}) yields $r(X')+r(Y)-r(M)<1$ which means $(X', Y)$ forms a $1$-separation of $M,$ a contradiction.\\
If $x\neq e,$ then $r'(X)=r(X')+2,$ and $r'(Y)=r(Y')+1.$ Again equation (\ref{eq1}) yields $r(X')+r(Y)-r(M)<0,$ which is a contradiction.\\
$\mathbf{Subcase~(iii):}$ $|X|>3$\\
Here $|X'|,|Y|\geq 2.$ If $Y$ contains an $np$-circuit of $M,$ then, by equation (\ref{r1}) $r'(Y)=r(Y)+1,$ and by equation (\ref{r4}) $r'(X)\geq r(X')+1.$ Therefore equation (\ref{eq1}) gives $r(X')+r(Y)-r(M)<1,$ which means $(X',Y)$ forms a $1$-separation of $M;$ a contradiction.\\
If $Y$ contains no $np$-circuit of $M,$ then $r'(Y)=r(Y)$ and $r'(X)\geq r(X')+1.$ Again equation (\ref{eq1}) yields $r(X')+r(Y)-r(M)<2,$ implying $(X', Y)$ is a $2$-separation of $M,$ a contradiction.\\
So we conclude that $M^e_{a,b}$ cannot have such $2$-separation. Hence $M^e_{a,b}$ is $3$-connected.
\end{proof}
\section{Applications}
The present section characterizes Eulerian $es$-splitting $p$-matroids $M^e_{a,b},(p>2).$ Also we provide a sufficient condition to yield Hamiltonian $p$-matriods from Hamiltonian $p$-matriods under $es$-splitting operation.
\begin{theorem}\label{EU}
Let $M$ be a $p$-matroid $(p>2),$ and $a,b \in E(M).$ Then the $es$-splitting matroid $M^e_{a,b}$ is Eulerian if and only if there exists a collection of circuits $\{ C_{np},C_1,\ldots,C_k \},$ where $C_{np}$ is an $np$-circuit containing $a,b,$ $E(M)= C_{np} \cup C_1\cup C_2\cup \ldots \cup C_k,$ and all the circuits in the collection are pairwise disjoint, except $C_{np} \cap C_1 = \{b\}.$
\end{theorem}
\begin{proof}
First assume that $M^e_{a,b}$ is Eulerian $p$-matroid, and $E(M^e_{a,b})=C'_1 \cup C'_2\cup C'_3\ldots \cup C'_k,$ where $C'_1, C'_2, \ldots,C'_k$ are disjoint circuits of $M^e_{a,b}.$ Set $e=b$. Without loss of generality, we assume $\gamma \in C'_1.$\\
$\mathbf{Case~(i):}$ $C'_1=\Delta=\{e, z, \gamma\}$ and $a\in C'_2.$\\
Since $\gamma, \notin C'_2,$ $C'_2$ is not a circuit of type $ \mathcal{C}_i; i=4,5,6,7,8. $ Also $C'_2\notin \mathcal{C}_z,$ as $z \notin C'_2.$ No circuits of $M_{a,b}$ contains only $a,$ therefore $C'_2\notin \mathcal{C}(M_{a,b})$.Therefore $C'_2 $ is not a circuit of $M^e_{a,b},$ a contradiction.\\
\noindent $\mathbf{Case~(ii):}$ $C'_1\in \mathcal{C}_4,$ \\
Then $e, a, \gamma \in C'_1.$ Assume $z\in C_2'.$ Since $\gamma \notin C'_2,$ $C'_2$ is not a circuit of type $ \mathcal{C}_i; i=4,5,6,7,8$ and $C'_2\neq \Delta.$ Also $C_2'\notin \mathcal{C}_z,$ as both $a,e \in C'_1.$ No circuit of $M_{a,b}$ contains $z,$ therefore $C_2'\notin \mathcal{C}(M_{a,b}).$ Therefore $C'_2 $ is not a circuit of $M^e_{a,b},$ a contradiction.\\
\noindent By similar arguments as in $\mathbf{Case~(i)},$$\mathbf{Case~(ii)},$ we get a contradiction if $C'_1\in \mathcal{C}_6,$ or $ C'_1\in\mathcal{C}_7,$ or $C'_1\in \mathcal{C}_8.$\\
\noindent $\mathbf{Case~(iii):}$ $C'_1\in \mathcal{C}_5,$\\
$C'_1=(C\setminus e)\cup \gamma,$ where $C$ is an $np$-circuit containing $e$ but not $a.$ Thus $z, e, a\notin C'_1.$ Since $\gamma \notin C'_2,$ $C'_2$ is not a circuit of type $ \mathcal{C}_i; i=4,5,6,7,8,$ and $C'_2\neq \Delta.$ Assume,WLOG, $a \in C'_2.$ Then following are the possible subcases:\\
\noindent $\mathbf{Subcase~(i):}$ $e,z \in C'_2.$\\
Observe that $a,e,z \in C'_2.$ As $z \in C_2'$ but $\gamma \notin C_2',$ $C'_2$ is a circuit of type $\mathcal{C}_z.$ Therefore $C'_2\setminus z,$ is an $np$-circuit. Set $C_{np}= C'_2\setminus z.$ Note that $C$ and $C_{np}$ are $np$-circuits of $M$ and $C \cap C_{np}=\{e\}.$ Thus the collection $\{ C, C_{np}, C'_3,\ldots,C'_k \}$ is the desired collection of the circuits of $M,$ where all the circuits are pairwise disjoint, except $C \cap C_{np}=\{e\}.$ \\
\noindent $\mathbf{Subcase~(ii):}$ $e\in C'_2, $ but $z \notin C'_2.$\\
Let $z \in C'_i$ for some $i \in \{ 3,4,\dots,k\}.$ Since $a,e \notin C_i,$ $C_i$ cannot be a circuit of $M^e_{a,b},$ a contradiction.\\
\noindent $\mathbf{Subcase~(iii):}$ $e,z\notin C'_2.$ \\
Let $e \in C'_i$ and $z \in C'_j$ for some $i\neq j \in \{ 3,4,\dots,k\}.$ Since $a,e \notin C_j,$ $C_j$ cannot be a circuit of $M^e_{a,b},$ a contradiction.
\noindent Conversely, suppose there exists a collection of circuits $\{ C_{np},C_1,\ldots,C_k \},$ where $C_{np}$ is an $np$-circuit containing $a,b,$ $E(M)= C_{np} \cup C_1\cup C_2\cup \ldots \cup C_k,$ and all the circuits in the collection are pairwise disjoint, except $C_{np} \cap C_1 = \{b\}.$ Note that the circuits $C_2,C_3, \ldots,C_k $ are $p$-circuits, and $C_{np}$ and $C_1$ are $np$-circuits. Therefore $\{C_2,\ldots,C_k \} \subset \mathcal{C}(M^e_{a,b}).$ Observe that $e \in C_1,$ but $a \notin C_1.$ Thus $C'_1=(C_1 \setminus e) \cup \gamma $ forms a circuit of type $\mathcal{C}_5$ in $M^e_{a,b},$ and $C_{np} \cup z$ forms a circuit of type $\mathcal{C}_z.$ The collection $\{ C_{np} \cup z,C'_1,C_2, C_3,\ldots,C_k \}$ gives a circuit decomposition of $E(M^e_{a,b}).$ Therefore, $M^e_{a,b}$ is Eulerian.
\end{proof}
\noindent Theorem \ref{EU} can be verified with Eulerian $p$- matroid discussed in Example \ref{ex1}. Here $a=1,b=e=4.$ Take $C_{np}=\{1,2,4,7\}$ and $C_1=\{3,4,5,6,8\}.$ Note that,$1,4\in C_{np},$$ C_{np}\cap C_1=\{4\},$ and $E(M)=C_{np}\cup C_1.$ The $es$-splitting matroid $M^e_{1,4}$ is Eulerian with circuit decomposition $\{\{1,2,4,7,9\},\{3,5,6,8,10\}\}=\{C_{np}\cup z, (C_1\setminus e ) \cup \gamma \}.$\\
\noindent In the next example we verify Theorem \ref{EU} for a non-Eulerian $p$-matroid.
\begin{example} \label{counter}
Consider the vector matroid $M \cong M[A]$ represented by the matrix $A$ over field $GF(5)$.
\begin{center}
$\mathbf{A} =
\begin{pNiceMatrix
[first-col,
first-row,
code-for-first-col = \color{black},
code-for-first-row = \color{black}]
& 1 & 2 & 3 & 4 & 5 \\
& 1 & 0 & 0 & 1 & 1\\
& 0 & 1 & 0 & 1 & 1 \\
& 0 & 0 & 1 & 1 & 0 \\
\end{pNiceMatrix} \qquad
\mathbf{A^e_{3,5}} =
\begin{pNiceMatrix
[first-col,
first-row,
code-for-first-col = \color{black},
code-for-first-row = \color{black}]
& 1 & 2 & 3 & 4 & 5 & 6 & 7\\
& 1 & 0 & 0 & 1 & 1 & 0 & 1\\
& 0 & 1 & 0 & 1 & 1 & 0 & 1\\
& 0 & 0 & 1 & 1 & 0 & 0 & 0\\
& 0 & 0 & 1 & 0 & 1 & 1 & 0\\
\end{pNiceMatrix}$
\end{center}
\noindent The circuits of $M$ are $\{1,2,3,4\},\{1,2,5\},\{3,4,5\}.$ Therefore $M$ is a non-Eulerian matroid. Consider $es$-splitting matroid $M^e_{3,5} \cong M[A^e_{3,5}]$ corresponding to $a=3$, $b=e=5.$ Take $C_{np}= \{3,4,5\}, C_1= \{1,2,5\}.$ Note that $C_{np}\cap C_1= \{5\},$ and $C_{np}\cup C_1=E(M).$ The resulting $es$-splitting matroid $M^e_{3,5}$ is Eulerian with circuit decomposition $\{\{3,4,5,6\}, \{1,2,7\}\}= \{C_{np}\cup z, (C_1\setminus e) \cup \gamma\}.$
\end{example}
\begin{proposition}
Let $M$ be a $p$-matroid containing an $np$-circuit $C_{np}$ of size $r(M)+1.$ Then $M^e_{a,b}$ is Hamiltonian. In this case Hamiltonian circuits in $M^e_{a,b}$ are:
\begin{enumerate}
\item $C_{np}\cup z,$
\item $C_{np}\cup \gamma,$ if $a,e \in C_{np}.$
\end{enumerate}
\end{proposition}
|
\section{Introduction}
\label{sec:intro}
\subsection{Motivation}
\label{sec:motivation}
Convex functions are analytically and algorithmically very well-behaved. For example, every local minimizer of a convex function is also a global minimizer, and there are many efficient algorithms that can be used to find minima of convex functions under mild assumptions~\citep{nesterov_book,nemirovsky_yudin}. However, there are important examples in statistics, economics, and machine learning where standard assumptions might not hold, specifically, where the underlying objective function is convex, continuous, and bounded below, but nevertheless fails to attain its minimum at any finite point in ${\R^n}$.
When this happens, to approach its infimum, the function's argument must somehow be taken ``to infinity.''
This can happen even in the simplest of cases:
\begin{example}[Exponential]
\label{ex:exp}
Consider minimization of $f(x)=e^x$ over $x\in{\mathbb{R}}$. This function can only be ``minimized'' in a limit as $x\to-\infty$.
\end{example}
This example is by no means unusual. In statistics, for instance, many common parametric distributions, such as a multivariate Gaussian or a multinomial, can be expressed as elements of a suitable ``exponential family'' and their parameters can be fitted by maximizing log likelihood~\citep[Equation 3.38]{WainwrightJo08}. This modeling approach is both analytically and algorithmically appealing, because the log-likelihood function for exponential families is concave. However, there are cases when the log likelihood is maximized only as the parameter goes to infinity. Since the parameter is, in general, a multi-dimensional vector, there are a variety of ways how it can go to infinity. The analysis of these cases relies on machinery of convex duality~\citep[Section 3.6]{WainwrightJo08}. At the crux of the analysis, optimality conditions are developed for sequences of parameters diverging to infinity. The analysis is highly tailored to exponential families.
Another example, drawn from machine learning, is the problem of binary classification. Here, a learning algorithm is provided with a set of training examples (such as images of photographs, represented as vectors of pixels); some of the examples are labeled as positive and some as negative (indicating, for example, if the photograph is or is not of a person's face). The goal then is to find a rule for predicting if a new instance should be labeled as positive or negative. Many standard algorithms, including variants of boosting and logistic regression, can be recast as minimization algorithms for specific convex objective functions~\citep[Chapter 7]{schapire_freund_book_final}. Again, there are cases, when the minimum is only attained as the underlying parameter goes to infinity. The analysis of these cases is highly specialized to the classification setting, and, similar to the analysis of exponential families, relies on the machinery of convex duality
\citep{zhang_yu_boosting,mjt_margins,nati_logistic,riskparam,GLSS18}.
These highly tailored approaches suggest that there is a structure among the divergent sequences that could be perhaps described by a unified theory. Motivated by that observation, we ask in this manuscript: \emph{How can we extend ${\R^n}$ to include ``points at infinity'' that would lead to a more complete theory of convex functions?} Ideally, with such an extension, there would be no need to analyze various types of divergent sequences; instead, we could work directly in the extended space, which would contain ``infinite points'' to which these sequences converge.
\bigskip
The task of adding ``points at infinity'' is deceivingly simple in one dimension, where we only need to add $+\infty$ and $-\infty$. The resulting extension is typically referred to as the \emph{(affinely) extended real line} and denoted $\eR=[-\infty,+\infty]$~\citep[Chapter~4]{bourbaki_general_topology_part1}.
But it is far from obvious how to generalize this concept to multiple dimensions. To develop some intuition about what is required of such an extension, we next look at several examples of convex functions and their minimizing sequences.
Before proceeding, some brief notational comments:
Vectors ${\bf x}$ in ${\R^n}$ are usually written in bold, with components
$x_i$ given in italics so that
${\bf x}=\trans{[x_1,\ldots,x_n]}$.
We use $t=1,2,\ldots$ as the
index of sequences, unless noted otherwise,
and write $\seq{{\bf x}_t}$ to mean
the sequence ${\bf x}_1,{\bf x}_2,\ldots$. Limits and convergence are taken as $t\rightarrow+\infty$,
unless stated otherwise. For example, $\lim {\bf x}_t$ means
$\lim_{t\to\infty}{\bf x}_t$, and ${\bf x}_t\rightarrow{\bf x}$ means ${\bf x}_t$ converges to~${\bf x}$ as $t\rightarrow+\infty$.
\begin{example}[Log-sum-exp]
\label{ex:log-sum-exp}
We first consider the log-sum-exp function, which comes up, for example, when fitting a multinomial model by maximum likelihood. Specifically,
let
$
\Delta_n = \braces{ {\bf x}\in [0,1]^n : \sum_{i=1}^n x_i = 1 }
$
be the probability simplex in ${\R^n}$,
and
let $\boldsymbol{\lsec}$ be any point in $\Delta_n$.
Consider minimization of
\[
\textstyle
f({\bf x})=\ln\Parens{\sum_{i=1}^n e^{x_i}}-\boldsymbol{\lsec}\cdot{\bf x}.
\]
As is standard, $f$ can be rewritten as
\[
f({\bf x})=\sum_{i=1}^n -\alpha_i\ln\Parens{\frac{e^{x_i}}{\sum_{j=1}^n e^{x_j}}}
=\sum_{i=1}^n -\alpha_i\ln p_i,
\]
where $p_i=e^{x_i}/\sum_{j=1}^n e^{x_j}$, so ${\bf p}$ is a probability vector in $\Delta_n$. In this form, it can be argued that
\[
f({\bf x})\ge\sum_{i=1}^n -\alpha_i\ln\alpha_i,
\]
and the minimum is attained if and only if ${\bf p}=\boldsymbol{\lsec}$. So, if $\alpha_i>0$ for all $i$, the minimum is attained by the vector ${\bf x}$ with entries $x_i=\ln\alpha_i$.
However, if $\alpha_i=0$ for some $i$, then there is no ${\bf x}\in{\R^n}$ that attains the minimum. The infimum $\sum_{i=1}^n -\alpha_i\ln\alpha_i$ is in this case reached by any sequence $\seq{{\bf x}_t}$ whose corresponding sequence $\seq{{\bf p}_t}$ converges to $\boldsymbol{\lsec}$. For example, if $n=3$ and $\boldsymbol{\lsec}=[0,\frac13,\frac23]^\top$, then the sequence ${\bf x}_t=[-t,\ln\frac13,\ln\frac23]^\top$ converges to the infimum as $t\to+\infty$.
\end{example}
This example suggests that it perhaps would suffice to allow individual coordinates to take on values $\pm\infty$. In other words, we could consider a Cartesian product of extended reals, $(\eR)^n$. The minimizer of \Ex{log-sum-exp} would then be $[-\infty,\ln\frac13,\ln\frac23]^\top$. But the next example shows that this
is not enough, because minimization may take us in a direction that is not aligned with coordinate axes.
\begin{figure}[t]
\begin{minipage}[t]{0.46\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/astral_contours__1c.pdf}
\end{minipage}\hfill
\begin{minipage}[t]{0.46\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/astral_contours__0c.pdf}
\end{minipage}
\caption{Contours and minimizing paths of functions in Examples~\ref{ex:diagonal-valley}
and~\ref{ex:two-speed-exp}.
\emph{Left.} Diagonal valley: $e^{-x_1} + (x_2-x_1)^2$.
\emph{Right.} Two-speed exponential: ${e^{-x_1} + e^{-x_2 + x_1^2/2}}$.\looseness=-1
}\label{fig:intro:examples}
\end{figure}
\begin{example}[Diagonal valley]
\label{ex:diagonal-valley}
Consider $f:{\mathbb{R}}^2\to{\mathbb{R}}$, shown in Figure~\ref{fig:intro:examples}~(left), defined as
\begin{equation}
\label{eqn:eg-diag-val}
f({\bf x})=e^{-x_1}+(x_2-x_1)^2.
\end{equation}
The infimum is obtained in the limit of any sequence $\seq{{\bf x}_t}$ that satisfies $x_{t,1}\to+\infty$ while also $x_{t,2}-x_{t,1}\to 0$. One such sequence is ${\bf x}_t=[t,t]^\top$, which goes to infinity ``in the direction'' of the vector ${\bf v}=[1,1]^\top$.
If we were to just work in $(\overline{\R})^2$, we would find that ${\bf x}_t\to[+\infty,+\infty]^\top$. However, to minimize $f$, the direction in which $\seq{{\bf x}_t}$ goes to infinity is critical, and that direction is not represented by limit points in $(\overline{\R})^2$.
For instance,
the sequence ${\bf x}'_t=[2t,t]^\top$, which goes to infinity in the direction $[2,1]^\top$, also converges to $[+\infty,+\infty]^\top$ in $(\overline{\R})^2$, but it fails to minimize $f$ since $f({\bf x}'_t)\rightarrow+\infty$.
\end{example}
So, perhaps, we should extend ${\mathbb{R}}^n$ with ``limit points''
of sequences going to infinity in various ``directions'' in ${\mathbb{R}}^n$.
Such limit points are considered, for example,
by \citet[Section 3]{rock_wets}, under the name ``direction points'', and the topology and properties of the resulting ``cosmic space'' are developed. A related concept is that of ``ideal points'' in the real projective plane, where a separate ideal point is introduced for each class of parallel lines~\citep[see][Chapter 8]{cox_little_oshea__ideals_varieties_algorithms}.
However, these abstractions do not capture the minimizing sequences in the diagonal valley example. For example, in the cosmic space formalism, the sequence ${\bf x}''_t=[t+1,t]^\top$ converges to the same limit point as ${\bf x}_t=[t,t]^\top$, since they both go to infinity in the same direction, but the sequence $\seq{{\bf x}''_t}$ fails to minimize $f$, because $f({\bf x}''_t)\rightarrow 1$.
Maybe we just need to consider an ``offset'' in addition to a ``direction'' of a sequence, and consider limit points corresponding to following a ray with a specified starting point and direction. All the minimizing sequences in the preceding examples can be written as ${\bf x}_t=t{\bf v}+{\bf q}$ for a suitable choice of ${\bf v}$ and ${\bf q}$, so each indeed proceeds along a ray with a starting point ${\bf q}$ and a direction ${\bf v}$. It turns out that this is still not enough, as seen in the next example:
\begin{example}[Two-speed exponential]
\label{ex:two-speed-exp}
Consider $f:{\mathbb{R}}^2\to{\mathbb{R}}$, shown in Figure~\ref{fig:intro:examples}~(right), defined as
\[
f({\bf x})=e^{-x_1} + e^{-x_2+x_1^2/2}.
\]
The infimum is obtained in the limit of any sequence $\seq{{\bf x}_t}$ that satisfies $x_{t,1}\to+\infty$ while also $-x_{t,2}+x_{t,1}^2/2\to-\infty$. This means $f$ cannot be minimized along any ray; rather, $x_{t,2}$ must go to $+\infty$ at least quadratically faster
than $x_{t,1}$.
One such sequence is $\smash{{\bf x}_t=[t,t^2]^\top}$.
\end{example}
The above examples show that the task of adding ``infinite points'' is subtle already in ${\mathbb{R}}^2$. So perhaps we should just stick to sequences to maintain the broadest flexibility.
The downside is that sequences seem only indirect proxies for the underlying ``infinite points,'' so working with sequences makes it harder to discover their structural properties. Moreover, we hope that by adding the extra ``infinite points,'' the theory of convex functions will become more ``regular,'' for example, by ensuring that all lower-bounded and continuous convex functions (like all those from our examples) have minima. Our hope is to extend ${\mathbb{R}}^n$ analogously to how the set of rational numbers is extended to obtain reals, or the set of reals is extended to obtain complex numbers. When moving from rational numbers to reals to complex numbers,
basic arithmetic operations naturally extend to the enlarged space, and the enlarged space has a more regular structure;
for example, bounded sets of reals have real-valued infima (which is not true for rational numbers), and all polynomials with complex coefficients have complex roots (whereas polynomials with real coefficients might have no real roots).
Here, we seek an extension of ${\mathbb{R}}^n$ which would lend itself to natural extensions of the key concepts in convex analysis, like convex sets and functions, conjugacy, subdifferentials, and optimality conditions, but which would also exhibit more regularity, for example, when it comes to the existence of minimizers.
\bigskip
So, what might a ``more complete'' theory of convex functions look like, and what is there to gain from such a theory? To start, there might be some algorithmic benefits. For example, if the extension of ${\mathbb{R}}^n$ were a \emph{sequentially compact} topological space, then any sequence would have a convergent subsequence, making it much easier to establish convergence of optimization algorithms.
But such a theory might exhibit additional regularities. To get an initial glimpse of what these might be, we study a continuous extension of the exponential function to $\eR$.
\begin{example}[Exponential, continued]
\label{ex:exp:cont}
Let $f(x)=e^x$ over $x\in{\mathbb{R}}$. A natural, continuous, extension of $f$ to $\eR$ is the function
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu:\eR\to\eR$ defined as
\begin{align*}
&{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(x)=e^x\text{ when $x\in{\mathbb{R}}$},
\\
&\textstyle
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(-\infty)=\lim_{x\to-\infty}e^x=0,
\\
&\textstyle
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(+\infty)=\lim_{x\to+\infty}e^x=+\infty.
\end{align*}
Unlike $f$, which has no minimum,
the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is more ``regular'' because it does attain its minimum (at $-\infty$).
Now, let us attempt to generalize the notion of conjugacy to the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$. Recall that each convex function $f:{\R^n}\to\overline{\R}$ has an associated conjugate
$f^*:{\R^n}\to\overline{\R}$ defined as
\[
f^*({\bf u})=\sup_{{\bf x}\in{\R^n}}\bigBracks{{\bf x}\cdot{\bf u}-f({\bf x})},
\]
which is itself a convex function. The conjugate is closely related to the problem of minimization of $f$: if the conjugate is finite (or equal to $-\infty$) at the origin,
it means that $f$ is bounded below; and if the conjugate is differentiable at the origin,
then $\nabla f^*({\bf 0})$ must be the minimizer of $f$.
More generally, any subgradient of $f^*$ at the origin
is a minimizer of $f$. Since the subgradients at the origin
are exactly the slopes of nonvertical tangents of the (epi)graph of $f^*$ at the origin, the analysis of minimizers of $f$ can be conducted by analyzing the tangents of the epigraph of $f^*$ at the origin.
Is it possible to lift this style of reasoning---a conjugate, subgradients at the origin, tangents of epigraph at the origin---to the function ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ we defined above? Let's try!
First, the conjugate of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ could be perhaps defined as ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}:{\mathbb{R}}\to\overline{\R}$ such that
\[
{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}(u)
=\sup_{\e{x}\in\eR:\:{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\e{x})\in{\mathbb{R}}}
[\e{x}\cdot u - {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\e{x})].
\]
We restricted the supremum to the set where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\e{x})$ is finite to assure that the expression $\e{x}\cdot u - {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\e{x})$ is defined even when $\e{x}\in\set{-\infty,+\infty}$. This definition does not completely match the definition we provide in Section~\ref{sec:conjugacy}, but for the presently studied function (and more generally, whenever $f>-\infty$) it gives rise to the same ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}$. This resulting ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}$ coincides with the standard conjugate $f^*$, that is,
\[
{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}(u)=
f^*(u)=
\begin{cases}
u\ln u-u
&\text{if $u\ge0$}
\\
+\infty
&\text{if $u<0$,}
\end{cases}
\]
with the convention $0\ln 0=0$. The function ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}$ is differentiable at all $u>0$, with the derivative $\bigParens{{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}}{\strut}'(u)=\ln u$. The function is finite at $0$, but not differentiable, because its epigraph has no nonvertical tangents at $0$.
However, its epigraph has a \emph{vertical} tangent at $0$. We could represent this vertical tangent by positing the ``derivative'' equal to $-\infty$, corresponding to the observation that ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}$ is ``decreasing infinitely fast at $0$'' (faster than any finite slope). If we could now apply a theorem that says that subgradients of ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}$ at $0$ are minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$, we would obtain that $-\infty$ is the minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$, as is actually the case.
\end{example}
The above example suggests one area, besides the existence of minimizers, in which the theory of extended functions might give rise to a ``more complete'' convex analysis. By allowing subgradients to take on infinite values, extended subdifferentials would represent not only nonvertical, but also vertical tangents of the (epi)graphs of convex functions. As a result, convex functions with closed epigraphs would be subdifferentiable everywhere where they are finite, which is not the case in standard convex analysis.
The extended real line, $\overline{\R}$, appears suitable for developing such a more complete theory of convex functions in one dimension. But what might such an extension look like in multiple dimensions?
Our goal is not to merely extend the subgradients to include vertical tangents; that goal is already achieved by other frameworks, for example, by Rockafellar's ``horizon subgradients''~(\citealp{rockafellar1985extensions}; \citealp[Chapter 8]{rock_wets}).
Horizon subgradients have a well-developed variational theory, but beyond two dimensions, the approach misses some structures that our theory captures, and these structures are required to analyze Examples~\ref{ex:log-sum-exp}, \ref{ex:diagonal-valley}, and~\ref{ex:two-speed-exp}. We seek to capture these structures and develop a subdifferential that would fit like a puzzle piece into a more complete theory of convex analysis, including extended conjugates, optimality conditions, and other fundamental concepts.
In this manuscript, we propose such a theory. We extend ${\mathbb{R}}^n$ into a topological space called \emph{astral space}.
We study its properties, and develop a theory of convex functions on this space. Although astral space includes all of ${\mathbb{R}}^n$, it is not a vector space, nor even a metric space. However, it is sufficiently well-structured to allow useful and meaningful extensions of concepts of convexity, conjugacy, and subdifferentials. We develop these concepts and analyze various properties of convex functions on astral space, including the structure of their minimizers, characterization of continuity, and convergence of descent algorithms.
Although the conceptual underpinnings of astral space are simple, the full formal development is somewhat involved, because key topological properties need to be carefully established.
As a teaser of what the manuscript is about, we next present a condensed development of astral space. This will allow us to revisit our earlier multi-dimensional examples
and situate them within the theory of functions on astral space. After that, we will finish this introductory chapter with a high-level overview of the manuscript.
\subsection{A quick introduction to astral space}
\label{sec:intro:astral}
In constructing astral space, our aim is to derive
a topological extension of ${\mathbb{R}}^n$ in which various points ``at infinity'' have been added, corresponding to limits of sequences such as those from Examples~\ref{ex:exp}, \ref{ex:log-sum-exp}, \ref{ex:diagonal-valley}, and~\ref{ex:two-speed-exp}. In fact, we seek a \emph{compactification} of ${\mathbb{R}}^n$, in which every sequence has a convergent subsequence. There are many possible compactifications, which differ in how many new points they add and how finely they differentiate among different kinds of convergence to infinity.
We would like to add as few points as possible, but the fewer points we add, the fewer functions will be continuous in the new space, because there will be more sequences converging to each new point.
In this work,
we choose a tradeoff in which we add as few points as possible while still ensuring that all \emph{linear} functions, which are the bedrock of convex analysis, remain continuous.\looseness=-1
\paragraph{Convergence in all directions.}
To formalize this idea, we note that linear functions on ${\mathbb{R}}^n$ all take the form ${\bf x}\mapsto{\bf x}\cdot{\bf u}$ for some ${\bf u}\in{\mathbb{R}}^n$, and so each can be viewed as realizing a ``projection along a vector ${\bf u}$'' for some ${\bf u}\in{\mathbb{R}}^n$. We define a notion of a well-behaved sequence with respect to linear maps by saying that a sequence $\seq{{\bf x}_t}$ in ${\mathbb{R}}^n$ \emph{converges in all directions} if its projection $\seq{{\bf x}_t\cdot{\bf u}}$ along any vector ${\bf u}\in{\mathbb{R}}^n$ converges in $\eR$, meaning that $\lim({\bf x}_t\cdot{\bf u})$ exists in $\eR$ for all ${\bf u}\in{\R^n}$.
For example, every sequence that converges in ${\mathbb{R}}^n$ also converges in all directions, because ${\bf x}_t\to{\bf x}$ implies that ${\bf x}_t\cdot{\bf u}\to{\bf x}\cdot{\bf u}$. There are additional sequences that converge in all directions like all those appearing in Examples~\ref{ex:exp}, \ref{ex:log-sum-exp}, \ref{ex:diagonal-valley}, and~\ref{ex:two-speed-exp}.\looseness=-1
If two sequences $\seq{{\bf x}_t}$ and $\seq{{\bf y}_t}$ both converge in all directions
and also ${\lim({\bf x}_t\cdot{\bf u})}={\lim({\bf y}_t\cdot{\bf u})}$ for all ${\bf u}\in{\mathbb{R}}^n$, we say that
$\seq{{\bf x}_t}$ and $\seq{{\bf y}_t}$ are \emph{all-directions equivalent}.
\paragraph{Astral space.}
All-directions equivalence creates a partition into equivalence classes
of the set of sequences that converge in all directions.
To construct \emph{astral space}, we associate an \emph{astral point} with each such equivalence class; this point will then be exactly the common limit of every sequence in the associated class.
We write $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ for $n$-dimensional astral space, and use bar or overline notation to denote its
elements, such as $\overline{\xx}$ or $\overline{\yy}$.\looseness=-1
This definition allows us to naturally extend the
inner product. Specifically, for all $\overline{\xx}\in\extspac{n}$ and all ${\bf u}\in{\mathbb{R}}^n$, we define
\[
\overline{\xx}\cdot{\bf u}=\lim({\bf x}_t\cdot{\bf u})\in\eR,
\]
where $\seq{{\bf x}_t}$ is any sequence in the equivalence class associated with $\overline{\xx}$. Note that the value of $\overline{\xx}\cdot{\bf u}$ does not depend on the choice of $\seq{{\bf x}_t}$ because all the sequences
associated with $\overline{\xx}$ are all-directions equivalent and therefore have identical limits $\lim({\bf x}_t\cdot{\bf u})$. In fact, the values of these limits
uniquely identify each astral point, so every astral point $\overline{\xx}$ is uniquely identified by the values of $\overline{\xx}\cdot{\bf u}$ over all ${\bf u}\in{\mathbb{R}}^n$.
The space ${\mathbb{R}}^n$ is naturally included in $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$, coinciding with equivalence classes of sequences that converge to points in ${\mathbb{R}}^n$, as seen above.
But there are additional elements, which are said to be \emph{infinite} since they must satisfy
$\overline{\xx}\cdot{\bf u}\in\set{-\infty,+\infty}$ for at least one vector ${\bf u}$.
The simplest astral points (other than those in ${\R^n}$) are called \emph{astrons}, each obtained as the limit of a sequence of the form $\seq{t{\bf v}}$ for some ${\bf v}\in{\mathbb{R}}^n$. The resulting astron is denoted $\omega{\bf v}$, where $\omega$ is mnemonic for infinity. By reasoning about the limit of $\seq{t{\bf v}\cdot{\bf u}}$, it can be checked that
\[
(\limray{{\bf v}})\cdot{\bf u}
=
\begin{cases}
+\infty & \text{if ${\bf v}\cdot{\bf u}>0$} \\
0 & \text{if ${\bf v}\cdot{\bf u}=0$} \\
-\infty & \text{if ${\bf v}\cdot{\bf u}<0$.}
\end{cases}
\]
If ${\bf v}\neq{\bf 0}$, then the associated astron $\limray{{\bf v}}$ is infinite, and is exactly the limit when a ray is followed to infinity in the direction ${\bf v}$, beginning at the origin.
In one-dimensional astral space $\overline{{\mathbb{R}}^1}$, there are two distinct astrons (other than $0$) corresponding to $+\infty$ and $-\infty$. In multiple dimensions, there is a distinct astron $\omega{\bf v}$ associated with every unit vector ${\bf v}$.
The astron construction can be generalized in a way that turns out to yield all additional astral points, including the limits in Examples~\ref{ex:log-sum-exp}, \ref{ex:diagonal-valley}, \ref{ex:two-speed-exp}.
Let ${\bf q}\in{\R^n}$, let
${\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$, for some $k\geq 0$, and
consider the sequence
\begin{equation} \label{eqn:xt-poly-seq}
{\bf x}_t = t^k {\bf v}_1 + t^{k-1} {\bf v}_2 + \cdots + t {\bf v}_k + {\bf q}
= \sum_{i=1}^k t^{k-i+1} {\bf v}_i + {\bf q}.
\end{equation}
To verify that this sequence converges in all directions, and therefore
corresponds to an astral point, we can informally compute the limit of $\seq{{\bf x}_t\cdot{\bf u}}$
along any direction ${\bf u}\in{\R^n}$.
The evolution of the sequence $({\bf x}_t)$ is dominated
by its overwhelmingly rapid growth in the direction of ${\bf v}_1$.
As a result,
if ${\bf v}_1\cdot{\bf u}>0$ then ${\bf x}_t\cdot{\bf u}\rightarrow +\infty$,
and
if ${{\bf v}_1\cdot{\bf u}<0}$ then ${\bf x}_t\cdot{\bf u}\rightarrow -\infty$.
However, if ${\bf v}_1\cdot{\bf u}=0$,
then the term involving ${\bf v}_1$ vanishes when considering
${\bf x}_t\cdot{\bf u}$.
So, when projecting the sequence along vectors ${\bf u}$ perpendicular to
${\bf v}_1$, the direction ${\bf v}_2$ becomes dominant.
Considering these vectors ${\bf u}$, we find once again that ${\bf x}_t\cdot{\bf u}$ converges
to $+\infty$ or $-\infty$ if ${\bf v}_2\cdot{\bf u}>0$ or ${\bf v}_2\cdot{\bf u}<0$,
respectively.
This analysis can be continued, so that we next consider vectors
${\bf u}$ in the subspace orthogonal to both ${\bf v}_1$ and ${\bf v}_2$, where
${\bf v}_3$ is dominant.
And so on.
Eventually, for vectors ${\bf u}$ that are orthogonal to all
${\bf v}_1,\ldots,{\bf v}_k$, we find that
${\bf x}_t\cdot{\bf u}$ converges to the finite value ${\bf q}\cdot{\bf u}$.\looseness=-1
In summary, this argument shows that the sequence $\seq{{\bf x}_t}$ converges in all directions, and its corresponding astral point $\overline{\xx}$ is described by
\begin{equation}
\label{eq:intro-astral-point}
\newcommand{\mkern-.8mu\inprod\mkern-.8mu}{\mkern-.8mu\cdot\mkern-.8mu}
\!\!\!
\overline{\xx}\mkern-.8mu\inprod\mkern-.8mu{\bf u}=
\begin{cases}
+\infty
&\text{if ${\bf v}_i\mkern-.8mu\inprod\mkern-.8mu{\bf u}>0$ for some $i$, and ${\bf v}_j\mkern-.8mu\inprod\mkern-.8mu{\bf u}=0$ for $j=1,\dotsc,i-1$}
\\
-\infty
&\text{if ${\bf v}_i\mkern-.8mu\inprod\mkern-.8mu{\bf u}<0$ for some $i$, and ${\bf v}_j\mkern-.8mu\inprod\mkern-.8mu{\bf u}=0$ for $j=1,\dotsc,i-1$}
\\
{\bf q}\mkern-.8mu\inprod\mkern-.8mu{\bf u}
&\text{if ${\bf v}_i\mkern-.8mu\inprod\mkern-.8mu{\bf u}=0$ for $i=1,\dotsc,k$.}
\end{cases}
\end{equation}
Of course, this only applies to the particular sequence given in \eqref{eqn:xt-poly-seq}.
Nevertheless, we show (by a different style of reasoning) that every sequence that converges in all directions must have a similar structure as the sequence from \eqref{eqn:xt-poly-seq}. Namely, every
such sequence must have a most dominant direction ${\bf v}_1$ in which it is tending to infinity most rapidly, followed by a second most dominant direction ${\bf v}_2$, and so on, with residual convergence to some finite point ${\bf q}\in{\R^n}$.
As a result, as we prove
in \Cref{sec:astral:space:summary},
every astral point is characterized by \eqref{eq:intro-astral-point} for some choice of ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}$.
\paragraph{Astral topology.}
We endow astral space with an \emph{astral topology} (formally defined in Section~\ref{sec:astral-as-fcns}), which is the ``coarsest'' topology under which all linear maps ${\bf x}\mapsto{\bf x}\cdot{\bf u}$ defined over ${\mathbb{R}}^n$ can be extended cotinuously over $\extspac{n}$, specifically, by the map $\overline{\xx}\mapsto\overline{\xx}\cdot{\bf u}$ (for all ${\bf u}\in{\R^n}$).
In this topology, astral space has the following key topological properties (see Theorems~\ref{thm:i:1} and~\ref{thm:1count}):
\begin{enumerate}[noitemsep,label={(\alph*)}]
\item \label{intro:prop:compact}
$\extspac{n}$ is compact.\item \label{intro:prop:first}
$\extspac{n}$ is first-countable (meaning every point in $\extspac{n}$ has a countable basis).
\item \label{intro:prop:iad:1}
$\overline{\xx}_t\rightarrow\overline{\xx}$ if and only if
$\overline{\xx}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$ for all ${\bf u}\in{\R^n}$.
\item \label{intro:prop:iad:2}
If $\seq{\overline{\xx}_t\cdot{\bf u}}$
converges in $\overline{\R}$ for all ${\bf u}\in{\R^n}$, then
$\overline{\xx}_t\rightarrow\overline{\xx}$ for some $\overline{\xx}\in\extspac{n}$.
\item \label{intro:prop:dense}
${\R^n}$ is dense in $\extspac{n}$.
\item \label{intro:prop:R1}
$\extspac{1}$ is homeomorphic with $\eR$.
\end{enumerate}
The most important property is compactness, a key aspect in how astral space is more ``regular'' than ${\mathbb{R}}^n$. Because of compactness, every continuous function on astral space attains its minimum. Also, every closed subset is compact.
First-countability is another crucial structural property of astral space, allowing us to work with limits of sequences similar to ${\mathbb{R}}^n$, despite astral space not being metrizable for $n\ge 2$ (see Section~\ref{sec:not-second}).
First-countability implies, for example, that the closure of a set $S\subseteq\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ coincides with the set of limits of all sequences with elements in~$S$. Also, a map $f:\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}\to Y$ is continuous if and only if $\overline{\xx}_t\to\overline{\xx}$ implies $f(\overline{\xx}_t)\to f(\overline{\xx})$.
Since astral space is both compact and first-countable, it is also
sequentially compact, meaning every sequence in $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ has a convergent subsequence, greatly simplifying the analysis of optimization algorithms.
A great many of our proofs depend critically on compactness and first-countability.
Properties~\ref{intro:prop:iad:1} and~\ref{intro:prop:iad:2}
establish that convergence in astral space is the same as convergence in all directions. And, indeed, convergence in all directions is how we analyze convergence in astral space.
Finally, that ${\mathbb{R}}^n$ is dense in $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ means that $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ is the closure of ${\mathbb{R}}^n$, part of what it means for $\extspac{n}$ to be a compactification. As a result, every astral point $\overline{\xx}\in\extspac{n}$ is arbitrarily ``near'' to ${\R^n}$ in the topological sense that every neighborhood of $\overline{\xx}$ must include a point in ${\R^n}$.
In one dimension, $\extspac{1}$ is homeomorphic with $\eR$, so we can write $\eR$ instead of $\extspac{1}$ and work with the standard topology on $\eR$. In fact, we later define $\extspac{1}$ to be \emph{equal} to $\overline{\R}$.
\paragraph{Representing astral points.}
Although astral space is a topological extension of the vector space ${\mathbb{R}}^n$,
it is not a vector space itself, and astral points cannot be
added.
The root problem is that the sum of $-\infty$ and $+\infty$ is undefined, making it impossible, for example, to establish the identity $(\overline{\xx} + \overline{{\bf y}})\cdot{\bf u}=\overline{\xx}\cdot{\bf u}+\overline{{\bf y}}\cdot{\bf u}$, which is meaningless when $-\infty$ and $+\infty$ appear on the right-hand side.
While standard addition does not generalize to astral space, a non-commuta\-tive version does. For $\bar{x},\bar{y}\in\overline{\R}$, we write this operation, called \emph{leftward addition},
as $\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y}$.
It is the same as
ordinary addition except that,
when adding $-\infty$ and $+\infty$, the argument on the \emph{left} dominates.
Thus,
\begin{eqnarray*}
(+\infty) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (-\infty) &=& +\infty \\
(-\infty) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (+\infty) &=& -\infty \\
\bar{x} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y} &=& \bar{x} + \bar{y} \;\;\mbox{in all other cases.}
\end{eqnarray*}
This operation can then be extended from $\overline{\R}$ to $\extspac{n}$:
For $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
the leftward sum, written $\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$,
is defined to be that
unique point in $\extspac{n}$
(which must exist, by Proposition~\ref{pr:i:6})
for which
\[ (\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}})\cdot{\bf u} = \overline{\xx}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}}\cdot{\bf u} \]
for all ${\bf u}\in{\R^n}$.
While leftward addition is not commutative, it is associative.
Scalar multiplication of a vector and standard matrix-vector multiplication also extend to astral space and are distributive with leftward addition.
All astral points can be decomposed as the leftward sum of astrons
and a finite part, corresponding exactly to the
sequence of dominant directions as we saw in the example that led to
\eqref{eq:intro-astral-point}.
In that example,
\begin{equation}
\label{eq:intro-sum-astrons}
\overline{\xx} = \limray{{\bf v}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\dotsb\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}.
\end{equation}
This can be seen by comparing the definitions and observations above with \eqref{eq:intro-astral-point}.
Such a representation of an astral point is not unique. Nevertheless, every astral point does have a
unique \emph{canonical representation} in which the vectors ${\bf v}_1,\dotsc,
{\bf v}_k$ are orthonormal, and ${\bf q}$ is orthogonal to all of them.
Every astral point has an intrinsic astral rank, which is the
smallest number of astrons needed to represent it, and which is also equal to the number
of astrons appearing in its canonical representation.
Points in ${\R^n}$ have astral rank~$0$. Astral points of the form $\omega{\bf v}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ (including nonzero astrons) have astral rank $1$. They can be obtained as limits of sequences that go to infinity along a ray with the starting point ${\bf q}$ and the direction ${\bf v}$, such as ${\bf x}_t= t{\bf v}+{\bf q}$.
\subsection{Minimization revisited}
\label{sec:intro-min-revisit}
Building on the definition of astral space, we return to minimization problems considered in \Cref{sec:motivation}, using them to highlight a few general results about astral space.
Each of the functions appearing in the minimization problems of \Cref{sec:motivation} is minimized only via a sequence of points going to infinity. We would like to be able to take a limit of this sequence in astral space and evaluate the function at such an infinite limit point.
To achieve this goal, we focus especially on a natural extension of a convex function $f:{\mathbb{R}}^n\to{\mathbb{R}}$ to astral space, which we call the \emph{lower semicontinuous extension} (or simply,
\emph{extension}) of $f$ to $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$.
This function, written ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}:\extspac{n}\rightarrow\overline{\R}$, is defined at any $\overline{\xx}\in\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ as
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\overline{\xx}) = \adjustlimits\inf_{({\bf x}_t):\:{\bf x}_t\rightarrow \overline{\xx}\;}
\liminf_{t\rightarrow\infty} f({\bf x}_t),
\]
where the infimum is taken over all sequences $\seq{{\bf x}_t}$ in ${\R^n}$ that converge to $\overline{\xx}$.
In words, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\overline{\xx})$ is the infimum across all possible limit values achievable by any sequence $\bigParens{f({\bf x}_t)}$ for which ${\bf x}_t\to\overline{\xx}$. The extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is a lower semicontinuous function on a compact space, and therefore always has a minimizer~$\overline{\xx}$. Moreover, by the construction of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$, there must also exist a sequence in ${\R^n}$ converging to $\overline{\xx}$ that minimizes $f$.\looseness=-1
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is not just lower-semicontinuous, but actually continuous at $\overline{\xx}$, then \emph{every} sequence converging to $\overline{\xx}$ must minimize $f$. This makes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s continuity properties algorithmically appealing; understanding when and where the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous is therefore an integral part of this manuscript.
We now return to the
minimization problems from \Cref{sec:motivation},
illustrating, along the way,
some of the general results that appear later in the manuscript
(though somewhat specialized for simpler presentation).
In what follows, ${\bf e}_1,\ldots,{\bf e}_n$ denote the standard basis vectors in ${\R^n}$, so ${\bf e}_i$ is all $0$'s except for a $1$ in the $i$-th component.
\begin{example}[Diagonal valley, continued]
\label{ex:diagonal-valley:cont}
We derive the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of the diagonal valley function
from Example~\ref{ex:diagonal-valley}, which turns out to be continuous everywhere. First, we rewrite the function using inner products, since these can be continuously extended to astral space:
\[
f({\bf x})=e^{-x_1}+(x_2-x_1)^2=e^{{\bf x}\cdot(-{\bf e}_1)}+\bigBracks{{\bf x}\cdot({\bf e}_2-{\bf e}_1)}^2.
\]
Now to obtain the continuous extension, we can just rely on the continuity of extended inner products to obtain
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\overline{\xx})=e^{\overline{\xx}\cdot(-{\bf e}_1)}+\bigBracks{\overline{\xx}\cdot({\bf e}_2-{\bf e}_1)}^2.
\]
Here, implicitly, the functions $e^x$ and $x^2$ have been extended in the natural way to $\overline{\R}$ according to their limits as $x\rightarrow\pm\infty$.
Note that both terms of the summation are nonnegative (though possibly $+\infty$), so the sum is always defined.
The minimizing sequence ${\bf x}_t=[t,t]^\top$ from Example~\ref{ex:diagonal-valley} follows a ray, and converges in astral space to the point $\overline{\xx}=\limray{{\bf v}}$ where ${\bf v}=[1,1]^\top$. This point is an astron and its astral rank is one.
That ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is both continuous everwhere and has a rank-one minimizer is not a coincidence:
In Section~\ref{subsec:rank-one-minimizers},
we prove that if the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere, then it must have a minimizer of astral rank at most one.
\looseness=-1
\end{example}
\begin{example}[Two-speed exponential, continued]
\label{ex:two-speed-exp:cont}
Recall the two-speed exponential function from
Example~\ref{ex:two-speed-exp}:
\[
f({\bf x})=e^{-x_1} + e^{-x_2+x_1^2/2}.
\]
Unlike the previous example, this function's extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous everywhere.
Earlier, we argued that the sequence ${\bf x}_t=[t,t^2]^\top$ minimizes $f$, satisfying $f({\bf x}_t)\to 0$. The sequence $({\bf x}_t)$ converges to the astral point $\overline{\xx}=\omega{\bf e}_2\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\omega{\bf e}_1$ with astral rank~2. On the other hand, the sequence ${\bf x}'_t=[2t,t^2]^\top$
also converges to $\omega{\bf e}_2\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\omega{\bf e}_1$, but $f({\bf x}'_t)\to +\infty$.
This shows that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at $\overline{\xx}$, and means, more specifically, that the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ satisfies ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\overline{\xx})=0$, but not all sequences converging to $\overline{\xx}$ minimize $f$.\looseness=-1
It turns out that $\overline{\xx}$ is the only minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$, so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is also an example of a function that does not have a rank-one minimizer, and so cannot be minimized by following a ray to infinity.
As discussed in the previous example, the fact that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ has no rank-one minimizer implies generally that it cannot be continuous everywhere.
Of course, for minimizing a function, we mainly care about continuity at its minimizers; the function $f$ is discontinuous exactly at its only minimizer.
In general, we might hope that functions that are not continuous everywhere are at least continuous at their minimizers.
For functions $f$ that are finite everywhere, such as this one, however, it turns out that that is impossible.
In Theorem~\ref{thm:cont-conds-finiteev},
we prove in this case that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere if and only if it is continuous at all its minimizers.
So if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is discontinuous anywhere, then it must be discontinuous at least at one of its minimizers, as in this example.
In Section~\ref{subsec:charact:cont}, we determine the exact set of astral points where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is both continuous and not $+\infty$.
For a function $f$ that is finite everywhere, this set consists of all points of the form given in \eqref{eq:intro-sum-astrons} where ${\bf q}\in{\R^n}$, and ${\bf v}_1,\ldots,{\bf v}_k$ are in $f$'s \emph{recession cone} from standard convex analysis, which consists of all vectors or ``directions'' ${\bf z}\in{\R^n}$ in which $f$ never increases
(so that $f({\bf x}+{\bf z})\leq f({\bf x})$ for all ${\bf x}\in{\R^n}$).
In this case, the recession cone is the cone generated by the single vector ${\bf e}_2$,
that is, $\{ \lambda {\bf e}_2 : \lambda\geq 0\}$.
Therefore, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is both finite and continuous exactly at points in ${\mathbb{R}}^2$ and all points of the form $\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$, for some ${\bf q}\in{\mathbb{R}}^2$.
In particular, this means it is not continuous at its minimizer, $\overline{\xx}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$, as we determined earlier.
\end{example}
\begin{example}[Log-sum-exp, continued]
\label{ex:log-sum-exp:cont}
The log-sum-exp function from
Example~\ref{ex:log-sum-exp}
has a continuous extension similar to the diagonal valley function, but a direct construction is slightly more involved. Instead, we take a different route that showcases an exact dual characterization of continuity in terms of a geometric property related to $f$'s conjugate, ${f^*}$.
In Section~\ref{subsec:charact:cont:duality}, we show that when $f$ is finite everywhere, its extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is continuous everywhere if and only if
the effective domain of ${f^*}$
has a conic hull that is \emph{polyhedral}, meaning that it is equal to the intersection of finitely many halfspaces in ${\R^n}$.
(The \emph{effective domain} of a function with range $\overline{\R}$ is the set of points where the function is not $+\infty$.
The \emph{conic hull} of a set $S\subseteq{\R^n}$ is obtained by taking all the nonnegative combinations of points in $S$.)
For the log-sum-exp function, we have
\[
{f^*}({\bf u})=\begin{cases}
\sum_{i=1}^n (u_i+\alpha_i)\ln(u_i+\alpha_i)
&\text{if ${\bf u}+\boldsymbol{\lsec}\in\Delta_n$}
\\
+\infty
&\text{otherwise,}
\end{cases}
\]
where, as before, $\Delta_n$ is the probability simplex in ${\R^n}$.
Hence, the effective domain of ${f^*}$ is $\Delta_n-\boldsymbol{\lsec}$, which, being a translation of a simplex, is polyhedral. Its conic hull is therefore also polyhedral, so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is continuous everywhere (by Corollary~\ref{cor:thm:cont-conds-finiteev}).
Earlier, we considered the particular case that $n=3$ and
$\boldsymbol{\lsec}=[0,\frac13,\frac23]^\top$.
We saw, in this case, that $f$ is minimized by the sequence
${\bf x}_t=t{\bf v}+{\bf q}$, where
${\bf v}=[-1,0,0]^\top$ and
${\bf q}=[0,\ln\frac13,\ln\frac23]^\top$.
In astral terms, the sequence $\seq{{\bf x}_t}$ converges to
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$, which does indeed minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
(as does every sequence converging to $\overline{\xx}$,
since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere).
These facts can be related to differential properties of $f^*$.
In Example~\ref{ex:exp:cont},
we discussed that any subgradient of ${f^*}$ at ${\bf 0}$
minimizes $f$ (with mild conditions on $f$).
By its standard definition, ${\bf x}\in{\R^n}$ is a
\emph{subgradient} of ${f^*}$ at ${\bf u}_0\in{\R^n}$ if
\begin{equation} \label{eqn:intro-stand-subgrad}
{f^*}({\bf u}) \geq {f^*}({\bf u}_0) + {\bf x}\cdot ({\bf u}-{\bf u}_0)
\end{equation}
for all ${\bf u}\in{\R^n}$, so that ${f^*}({\bf u})$ is supported at ${\bf u}_0$ by the
affine function (in ${\bf u}$) on the right-hand side of the inequality.
In this case,
as in Example~\ref{ex:exp:cont},
${f^*}$ has \emph{no} subgradients at ${\bf 0}$, even though
${\bf 0}$ is in the effective domain of ${f^*}$, corresponding to $f$
having no finite minimizers.
In Section~\ref{sec:gradients},
we will see how subgradients extend to astral space.
Although our formulation there is more general, when
${f^*}({\bf u}_0)\in{\mathbb{R}}$, we simply need to replace the inner product in
\eqref{eqn:intro-stand-subgrad} by its extension; thus, a point
$\overline{\xx}\in\extspac{n}$ is an \emph{astral dual subgradient} of ${f^*}$
at ${\bf u}_0$ if
\[
{f^*}({\bf u}) \geq {f^*}({\bf u}_0) + \overline{\xx}\cdot ({\bf u}-{\bf u}_0)
\]
for all ${\bf u}\in{\R^n}$.
Whereas, as just seen, it is possible for a function to have no
standard subgradients at a particular point, in astral space, \emph{every}
point must have an astral dual subgradient (even points not in the function's
effective domain).
In this case, from the preceding development, it can be checked that
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$, as defined above, does indeed satisfy this
condition at ${\bf u}_0={\bf 0}$, and so is an astral dual subgradient.
By general results given in Section~\ref{sec:gradients}, this implies
that ${\bf 0}$ is a (primal) astral subgradient of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at $\overline{\xx}$,
which in turn implies that $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\end{example}
These examples illustrate some of the key questions that we study in
astral space, including continuity, conjugacy, the structure of
minimizers, and differential theory.
\subsection{Overview of the manuscript}
We next overview some of the highlights and main topics of
this manuscript,
as a kind of condensed roadmap.
Depending on the reader's interests and preferences, this overview can be read sequentially as a more detailed summary and sampling of much of what this manuscript is about;
alternatively, it can be skipped or skimmed on first reading, and set aside for later reference.
At the highest level, the manuscript defines astral space, studies its properties, then develops the theory of convex functions on astral space.
We follow the conceptual framework developed for convex analysis on ${\mathbb{R}}^n$ by \citet{ROC} (in particular, Parts I, II, III, and V), which in turn grew out of Fenchel's lecture notes~\citeyearpar{fenchel1953convex}.
Our construction of astral space can be viewed as a special case of Loeb's $\mathcal{Q}$-compactification~\citeyearpar{q_compactification}. Our development of convexity could be classified as an ``inner approach'' in the terminology of \citet{singer_book}, or an ``interval space'' approach in the terminology of \citet{vandeVel}. Our development of conjugacy builds on the abstract framework of~\citet{singer_book}.
\subsubsection*{Astral space
(\Cref{sec:astral-space-intro,,sec:astral-as-fcns,,sec:representing,,sec:countable})}
The manuscript begins with a full development of the concepts and results briefly introduced in \Cref{sec:intro:astral}.
We construct astral space
in Section~\ref{sec:astral-space-intro}
as a compactification of ${\mathbb{R}}^n$, in which limits of various sequences have been added, including those from Examples~\ref{ex:exp}, \ref{ex:log-sum-exp}, \ref{ex:diagonal-valley}, and~\ref{ex:two-speed-exp}.
A key property of this compactification is that it allows the continuous
extension of the inner product; that is, all linear maps ${\bf x}\mapsto{\bf x}\cdot{\bf u}$ defined over ${\mathbb{R}}^n$ can be continuously extended to the maps $\overline{\xx}\mapsto\overline{\xx}\cdot{\bf u}$ defined over astral space $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$. Furthermore, astral points $\overline{\xx}$ are entirely characterized by the values of the inner product $\overline{\xx}\cdot{\bf u}$ across all ${\bf u}\in{\mathbb{R}}^n$.
We establish the bedrock topological properties of astral space,
including compactness, in Section~\ref{sec:astral-as-fcns},
many of which are summarized in Theorem~\ref{thm:i:1}.
We show in Section~\ref{sec:countable}
that while astral space is not metrizable for $n\ge 2$, it is first-countable, and so we can rely on sequential definitions of continuity, closure, and compactness, as in ${\mathbb{R}}^n$.
As seen in \Cref{sec:intro:astral},
\emph{astrons} are
fundamental building blocks of astral space, obtained as limits of sequences of the form $(t{\bf v})$ for ${\bf v}\in{\mathbb{R}}^n$, and denoted $\omega{\bf v}$.
Astral space is not a vector space, and astral points cannot be
added.
However, as seen previously,
it is possible to combine astral points using a non-commutative, associative operation, called \emph{leftward addition}, and denoted $\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$.
We also have seen how
every astral point can be decomposed as the leftward sum of astrons
and a finite part, corresponding to the
sequence of dominant directions, as we saw in \eqref{eq:intro-astral-point}. Thus, every astral point $\overline{\xx}\in\extspac{n}$ has the form:
\begin{equation}
\label{eq:intro-sum-astrons:again}
\overline{\xx} = \limray{{\bf v}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\dotsb\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}
\end{equation}
for some ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in{\R^n}$
(Corollary~\ref{cor:h:1}).
Every astral point has an intrinsic \emph{astral rank}, which is the
smallest number of astrons needed to represent it.
Points in ${\R^n}$ have astral rank~$0$. Points of the form $\omega{\bf v}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ have astral rank $1$ and correspond to limits of sequences that go to infinity along a ray.
Many of our proofs are by induction on astral rank.
Since every astral point $\overline{\xx}\in\extspac{n}$ has
the form given in \eqref{eq:intro-sum-astrons:again},
we can more simply write
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$, where
${\bf q}\in{\R^n}$
and
$\overline{\ee}=\limray{{\bf v}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\cdots\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf v}_k}$.
An astral point, such as $\overline{\ee}$, that is the leftward sum of
(finitely many) astrons is called an \emph{icon}.
Equivalently,
$\overline{\ee}$ is an icon if and only if it is
an \emph{idempotent} (with respect to leftward addition),
meaning $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}=\overline{\ee}$.
Thus, every astral point $\overline{\xx}$ can be decomposed,
as was done here, into an
\emph{iconic part} $\overline{\ee}$,
and a \emph{finite part} ${\bf q}$.
Icons play a central role in our development.
Every point's iconic part is uniquely determined.
As a result, astral space can be partitioned into disjoint sets
called \emph{galaxies}, each
associated with an icon $\overline{\ee}$ and consisting exactly of those points
with $\overline{\ee}$ as their iconic part, that is,
$\{ \overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q} : {\bf q}\in{\R^n} \}$.
Within this galaxy,
leftward addition behaves like ordinary vector
addition with $\overline{\ee}$ as its zero.
Every such galaxy is a topological copy of $(n-k)$-dimensional Euclidean space, ${\mathbb{R}}^{n-k}$, where $k$ is the astral rank of the associated icon, $\overline{\ee}$.
(See Section~\ref{sec:core:zero}.)
\bigskip
Throughout our analysis,
and especially in Section~\ref{sec:astral-as-fcns},
we crucially rely on the dual
view of astral space, in which every astral point $\overline{\xx}$ is associated with a function
$\ph{\overline{\xx}}:{\R^n}\rightarrow\overline{\R}$ given by $\ph{\overline{\xx}}({\bf u}) = \overline{\xx}\cdot{\bf u}$,
for ${\bf u}\in{\R^n}$.
These turn out to be exactly those functions that are convex,
concave and that vanish at the origin.
Such functions are closely related to hemispaces in ${\mathbb{R}}^{n+1}$ (i.e.,
convex sets whose complements are also convex), whose structure was characterized by \citet{hemispaces}.
Astral space is
homeomorphic to this space of functions
$\set{\ph{\overline{\xx}} : \overline{\xx}\in\extspac{n}}$ under the standard
product topology.
Thus, compactness can be proved using
Tychonoff's theorem. This construction is a special case of Loeb's $\mathcal{Q}$-compactification~\citeyearpar{q_compactification}, similar
to Stone-\v{C}ech compactification.
\subsubsection*{Lower-semicontinuous extension and conjugacy
(\Cref{sec:functions,,sec:conjugacy})}
Lower-semicontinuous functions play a prominent role in standard convex analysis, because they are exactly the functions whose epigraphs are closed. Since astral space is compact, lower-semicontinuous functions on astral space have an additional appealing property: they always attain a minimum.
Our study of functions on astral space is motivated by the problem
of minimizing a convex function $f$ on ${\R^n}$, and in particular the fact that even when $f$ is bounded below, it might be minimized only via a sequence of points going to infinity.
To address this issue,
in Section~\ref{sec:lsc:ext},
we construct a \emph{lower-semicontinuous extension} ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ of $f$ to $\extspac{n}$,
as was seen in Section~\ref{sec:intro-min-revisit},
a main focus of our study.
The extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ always has a minimizer $\overline{\xx}$ such that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\overline{\xx})=\inf f$. Furthermore, if $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$, then there exists a sequence in ${\R^n}$ converging to $\overline{\xx}$ that minimizes $f$. If, in addition, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is continuous at $\overline{\xx}$, then \emph{every} sequence converging to $\overline{\xx}$ minimizes $f$.
For the remainder of this overview, we assume $f:{\R^n}\rightarrow\overline{\R}$ is convex and not identically $+\infty$.
Some of the mentioned results also require $f$ itself to be lower-semicontinuous; details are provided later in the manuscript.
In analyzing ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, it is often helpful to focus on its
behavior restricted to individual galaxies by considering the function
${\bf x}\mapsto{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})$, for ${\bf x}\in{\R^n}$, where
$\overline{\ee}$ is some icon.
Such a function, introduced in Section~\ref{sec:shadow},
is called a \emph{reduction} of $f$.
Although defined over all of ${\R^n}$, reductions are effectively functions of only a linear subspace whose dimension depends on the astral rank of the associated icon $\overline{\ee}$ (corresponding to galaxies being copies of lower-dimensional Euclidean space).
Reductions are used heavily in many of our proofs.
\bigskip
Conjugacy of a convex function in standard convex analysis expresses the fact that a ``closed'' convex function is a pointwise supremum of a set of affine functions that it majorizes.
Taking the conjugate of the conjugate (that is, the double conjugate) of a closed convex function yields that same function. Following \citet{moreau__convexity} and \citet{singer_book}, we translate the concepts of conjugacy and double conjugacy to astral space.
The resulting (astral) double conjugate of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is denoted $f^{*\dualstar}$.
We give a condition that exactly characterizes when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ coincides with its double conjugate, that is, when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
Specifically, this condition requires \emph{all} of $f$'s reductions to be closed, so that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is a closed function when restricted to any single galaxy (rather than just on ${\mathbb{R}}^n$ as would be required for a standard closed function). A function with this property is said to have \emph{all reductions closed}
(Section~\ref{sec:ent-closed-fcn}).
Functions with all reductions closed play a role similar to closed functions on~${\R^n}$. While the condition is more restrictive than just closedness, it is satisfied by many functions of interest in optimization. For example, all convex functions that are lower-bounded or finite everywhere have all reductions closed, including functions from Examples~\ref{ex:exp}, \ref{ex:log-sum-exp}, \ref{ex:diagonal-valley}, and~\ref{ex:two-speed-exp}.
Moreover, results that might be easier to prove initially for functions with this property can usually be generalized to all convex functions using a simple trick of composing the function with an exponential
(Section~\ref{sec:exp-comp}).
We also give in Section~\ref{sec:dual-char-ent-clos}
an exact dual condition for when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$,
stated in terms of a set called the \emph{barrier cone} of $f$
(defined in Section~\ref{sec:slopes}),
which arises repeatedly in our development.
Thus, the equality of these two astral functions can be
characterized entirely in terms of sets and notions from standard convex analysis.
\subsubsection*{Convex sets
(\Cref{sec:convexity,,sec:convex-set-ops,,sec:mono-passages})}
Even though astral space is not a vector space, convexity can be
naturally extended to it, as we briefly sketch. A \emph{closed halfspace} consists of all points $\overline{\xx}$ for which $\overline{\xx}\cdot{\bf u}\leq b$ for some
${\bf u}\in{\R^n}$ and $b\in{\mathbb{R}}$.
A \emph{segment} joining two endpoints is then the intersection of all
closed halfspaces that include the two points. A subset of astral space is defined to be
\emph{convex} if it contains the segment between every pair of its elements
(see Section~\ref{sec:def-convexity}).
The segment between two astral points $\overline{\xx}$ and $\overline{{\bf y}}$ can be formulated equivalently as the limit of sequences of ordinary segments in ${\mathbb{R}}^n$ whose endpoints converge to $\overline{\xx}$ and~$\overline{{\bf y}}$
(Corollary~\ref{cor:e:1}).
A different characterization of segments and convexity in terms of path-like, linearly ordered sets, called \emph{monotone passages}, is given in Section~\ref{sec:mono-passages}.
A construction of convex sets that begins with ``segments'' is referred to as an ``inner approach'' by \citet{singer_book} and as an ``interval space'' approach by \citet{vandeVel}.
When restricted to ${\R^n}$, astral convexity is the same as standard convexity. Convex hulls in astral space are defined like standard convex hulls and have many similar properties.
Many, but not all, properties of convex sets in ${\mathbb{R}}^n$ also hold in $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$.
For example,
in ${\R^n}$, both the interior and the closure of a convex set are convex. In $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$, the interior of a convex set is convex
(Theorem~\ref{thm:interior-convex}).
Also, if $S$ is a convex subset of ${\R^n}$, then its astral closure is a convex subset of $\extspac{n}$
(Theorem~\ref{thm:e:6}).
However, if $S$ is an arbitrary convex subset of $\extspac{n}$ (rather than ${\R^n}$),
then its closure is not necessarily convex
(Theorem~\ref{thm:closure-not-always-convex}).
\subsubsection*{Convex cones and functions
(\Cref{sec:cones,,sec:conv:fct})}
An important class of convex sets in astral space are convex cones.
A \emph{cone}, as in standard convex analysis, is simply a set in $\extspac{n}$
that is closed under multiplication by any positive scalar.
Convex cones and their topological properties play a key role in the study of minimization and continuity of functions on astral space.
Any set of icons is a cone, regardless of the cardinality of the set, providing an example of an astral cone of a rather different nature than standard cones.
Consequently, the convex hull of any set of icons is a convex cone.
Furthermore, every closed convex cone in astral space is equal
to the convex hull of its icons, so closed convex cones are characterized
by their icons.
(See Section~\ref{sec:cones-basic}.)
We study fundamental operations for constructing convex cones in astral space.
First, in Section~\ref{sec:cones:astrons},
we define the \emph{representational closure} of a set $K\subseteq{\R^n}$ to be the set of all astral points $\overline{\xx}$ that can be represented, as in \eqref{eq:intro-sum-astrons:again}, using only vectors
${\bf v}_1,\ldots,{\bf v}_k,{\bf q}$ that are all in $K$.
When $K$ is itself a (standard) convex cone,
its representational closure can be equivalently obtained as the convex hull of a particular set of astrons derived from $K$, which, as just mentioned, must be an astral convex cone.
This operation turns out to be central in determining the exact set of points where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous.
We also show,
in Section~\ref{sec:ast-pol-cones},
how the standard notion of a polar cone can be straightforwardly extended to astral space, yielding the \emph{astral polar cone}. Similar to a standard polar cone, the astral polar cone of a set in ${\R^n}$ is always a closed convex cone.
Astral polar cones arise in the study of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s minimizers, as will be discussed in a moment.
We prove key properties about both these operations.
We determine the exact condition under which the representational closure of a convex cone $K\subseteq{\R^n}$ is closed
(Theorem~\ref{thm:repcl-polyhedral-cone}).
We also determine the exact condition for when the astral polar cone of a set is equal to the closure of its standard polar cone
(Theorem~\ref{thm:k-closed-equals-closed-apol}).
Later, we combine and apply both these properties:
although more general,
for a function $f$ that is finite everywhere,
we use them together to give a precise dual characterization for when the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere
(Section~\ref{subsec:charact:cont:duality}).
Building on the definition of convex sets in astral space,
in Section~\ref{sec:conv:fct},
we define a function
on astral space to be convex if its epigraph is
convex. For a convex function $f$, the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ and its astral double conjugate $f^{*\dualstar}$ are both always convex.
Also, the effective domain and sublevel sets of a convex function on astral space
are convex.
\subsubsection*{Minimizers
(\Cref{sec:minimizers,,sec:univ-red-and-min,,sec:minimizers-examples})}
With the theory of astral convex cones as foundation, we next study in detail the structure of minimizers of an extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$.
Since all astral points can be decomposed into iconic and finite parts, we can likewise decompose minimization into the separate problems of finding a minimizer's iconic part and finding its finite part.
We begin with the former.
As mentioned in Example~\ref{ex:two-speed-exp:cont},
the (standard) \emph{recession cone} of a function $f$ is the set
of vectors ${\bf z}\in{\R^n}$ which, when added to the function's argument,
never cause it to increase, so that $f({\bf z}+{\bf x})\leq f({\bf x})$
for all ${\bf x}\in{\R^n}$.
Analogously, we define, in Section~\ref{sec:arescone},
the \emph{astral recession cone} of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
to consist of all astral points
$\overline{{\bf z}}\in\extspac{n}$ which, when leftwardly added to
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s argument, never cause an increase,
so that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}+\overline{\xx})\leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$
for all $\overline{\xx}\in\extspac{n}$.
It turns out that
the iconic part of every minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ must belong to
its astral recession cone.
In fact, the same is true for every point in ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s
effective domain
(Corollary~\ref{cor:a:4}).
Thus, for finding the iconic parts of points that minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
the astral recession cone plays a central role.
In standard convex analysis, the recession cone of a function $f$ is
exactly the polar cone of the function's barrier cone,
mentioned above, so that these cones are, in this sense, duals
of one another
(Corollary~\ref{cor:rescpol-is-slopes}).
Correspondingly, the \emph{astral} recession cone of the {extension} ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
turns out to be the \emph{astral} polar cone of this same barrier cone
(Theorem~\ref{th:ares-is-apolslopes}).
These two facts together create an important link between the
recession cone of $f$ and the astral recession cone of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
which we later exploit.
In Section~\ref{sec:find-all-min},
we give a complete accounting of all minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
To do so, we characterize the
astral recession cone in a way that yields a technique for constructing
an icon $\overline{\ee}$ in the set, one astron at a time.
When this technique is applied,
the reduction (restriction to a galaxy)
associated with $\overline{\ee}$ must eventually have a finite
minimum ${\bf q}\in{\R^n}$, at which point their combination,
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$, must be a minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
In a certain nondeterministic sense, the resulting procedure effectively enumerates \emph{all} minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Delving deeper into the structure of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s minimizers,
we focus next on their finite parts.
In Section~\ref{sec:fullshad},
we construct a particular convex function with the property that a finite point in ${\R^n}$ minimizes this function if and only if that point is the finite part of some minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
In this sense, this one function,
called the \emph{universal reduction} of $f$,
captures \emph{all} finite parts of \emph{all} of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s minimizers.
Furthermore,
the universal reduction has compact sublevel sets (ignoring the subspace along which it is constant), so all its minima are at finite points.
In this way, finding the finite parts of (possibly infinite) minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ can be reduced to minimizing an ordinary convex function whose minimizers are all finite and in a compact region.
The universal reduction of $f$ is defined, for any ${\bf x}\in{\R^n}$,
to be the minimum possible value of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ when ${\bf x}$ is combined with any icon.
In other words, we fix the finite part ${\bf x}$, and consider minimizing the function ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ over possible iconic parts.
In this way, all of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s behavior at infinity is effectively ``washed out.''
For different ${\bf x}\in{\R^n}$, we might expect that different icons would be required to attain this minimum (or even that the minimum is sometimes not attained).
Remarkably, it turns out that there exist icons that \emph{simultaneously} attain this minimum for \emph{all} ${\bf x}\in{\R^n}$
(Section~\ref{subsec:univ-min}).
Such icons are called \emph{universal reducers}.
This means the universal reduction is itself a reduction, specifically, the restriction of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ to the galaxy corresponding to any single universal reducer.
Universal reducers have the ``most minimizing'' effect of any astral points, and generate the entire astral recession cone, which is exactly its conic hull
(Theorem~\ref{thm:res-convhull-unimin}).
Any icon that is a universal reducer can be combined with any finite minimizer of the universal reduction to obtain a kind of canonical minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ called a \emph{universal minimizer}.
We describe procedures for constructing the universal reduction, and for finding all universal reducers and all universal minimizers
(Section~\ref{sec:construct-univ-reduct}).
\bigskip
As we saw in Example~\ref{ex:two-speed-exp:cont} (the two-speed exponential), there exist convex functions whose only minimizers in astral space are of rank~2, so it is not always possible to simplify
the process of minimization to sequences of the form ${\bf x}_t=t{\bf v}+{\bf q}$. More generally,
for any $n$, there exist convex functions $f:{\mathbb{R}}^n\to{\mathbb{R}}$ whose unique minimizer is of astral rank $n$, so any minimizing sequence of such a function ``diverges in all directions,'' meaning that it will wander arbitrarily far from
any proper linear subspace of ${\mathbb{R}}^n$
(Section~\ref{sec:max-rank-minimizers}).
On the other hand,
in Section~\ref{subsec:rank-one-minimizers},
we show that
there are important special cases when it suffices to consider minimizers of astral rank at most one.
In particular, all functions $f$ with a continuous extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ can be minimized by sequences along a ray, that is, of the form ${\bf x}_t=t{\bf v}+{\bf q}$.
This includes a broad class of
``empirical risk'' functions commonly encountered in machine learning and statistics, whose astral properties we
study in detail in Section~\ref{sec:emp-loss-min}.
\subsubsection*{Continuity
(\Cref{sec:continuity})}
Continuity of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ plays a key role in the minimization of $f$.
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is continuous at a minimizer $\overline{\xx}$, then
any sequence $\seq{{\bf x}_t}$ converging to $\overline{\xx}$ attains values
arbitrarily close to the minimum,
plausibly making such functions
more amenable to algorithmic minimization.
Perhaps surprisingly,
continuity of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is not assured by the continuity of $f$.
Nevertheless, regardless of whether $f$ itself is continuous,
in Section~\ref{subsec:charact:cont},
we give a precise description of exactly the points in its effective domain where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous.
(Note that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is continuous at all points outside its effective domain, that is, where it is $+\infty$.)
First, we show that this set coincides with the interior of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s effective domain,
analogous to a standard convex function being finite and continuous
exactly on the interior of its effective domain.
We then provide a second characterization
showing that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous and not $+\infty$ exactly at those astral points of a
particular form, namely, whose iconic part is
in the representational closure of the standard recession cone of $f$
(and therefore a leftward sum of astrons, each associated with a
vector in the standard recession cone of $f$), and whose
finite part is in the interior of the effective domain of $f$.
In this way, continuity of the astral function ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
can be characterized entirely in terms of the
original function $f$ and its behavior on ${\R^n}$.
We discuss applications of this result in Section~\ref{subsec:cond-for-cont}.
For example, we prove that if the function $f$ is
finite everywhere, then its extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is continuous everywhere if
and only if it is continuous at all its minimizers---a rather
surprising property.
Thus, if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is discontinuous anywhere, it must be discontinuous at one or more of its minimizers.
We also give a dual characterization regarding continuity
in Section~\ref{subsec:charact:cont:duality}.
Although more general, when $f$ is finite everywhere,
this characterization implies that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is continuous everywhere
if and only if the conic hull of
the effective domain of ${f^*}$
is polyhedral.
This means that continuity of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ can be characterized entirely by a geometric property regarding the original function $f$'s conjugate.
The proof uses general properties of astral convex cones,
applied to the standard recession cone and barrier cone,
which, as we have seen, are linked both to minimization and continuity
of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\subsubsection*{Subdifferentials and descent methods
(\Cref{sec:gradients})}
Subgradients and the subdifferential mapping are fundamental in describing
properties of convex functions. Standard subgradients correspond to
supporting nonvertical hyperplanes of epigraphs. One of their key properties is that they characterize minimizers: ${\bf x}$~is a minimizer of $f$ if and only if ${\bf 0}$ is a subgradient of $f$ at ${\bf x}$, that is, if and only if ${\bf 0}$ is in $\partial f({\bf x})$, the subdifferential of $f$ at ${\bf x}$.
We generalize subgradients in two different ways.
First, in Section~\ref{sec:gradients-def},
we extend the standard definition to
include subgradients ``at infinity.'' This, for example, allows us to state generally (Proposition~\ref{pr:asub-zero-is-min})
that $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ if and only if ${\bf 0}\in\partial{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\overline{\xx})$, where $\partial{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ is the \emph{astral subdifferential}, mapping $\overline{\xx}$ to the set of associated \emph{astral subgradients}.
Second, in Section~\ref{sec:astral-dual-subgrad},
we show how it is possible, at finite points,
to include ``infinite'' subgradients,
corresponding to vertical supporting hyperplanes. These
are called \emph{astral dual subgradients}, and the associated \emph{astral dual subdifferential} mapping is denoted $\e{\partial}$, to distinguish it from the (primal) astral subdifferential.
Thus, the primal subdifferential $\partial{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ maps points in $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ to subsets of ${\R^n}$ (the subgradients), whereas the dual subdifferential $\e{\partial} f$ maps points in ${\R^n}$ to subsets of $\overline{\R^n}}\newcommand{\eRd}{\overline{\R^d}}\newcommand{\eR}{\overline{\R}$ (the dual subgradients).
Since there exist either vertical or nonvertical supporting hyperplanes at every point of the graph of a function, a fundamental property of the astral dual subdifferential,
proved in Section~\ref{sec:dual-subdiff-not-empty},
is that it is nonempty at \emph{every} point of the function's effective domain, which is not true for the standard subdifferential.
In fact, our definition yields a nonempty dual subdifferential even outside the function's effective domain, where its values correspond to vertical ``barrier hyperplanes,'' that is, vertical hyperplanes that do not intersect the function's epigraph.
In standard convex analysis \citep[Corollary~23.5.1]{ROC}, subdifferentials of $f$
and $f^*$ act as inverses of one another in
the sense that if $f$ is a closed proper convex function, then
\[
{\bf u}\in\partial f({\bf x})
\quad
\Longleftrightarrow
\quad
{\bf x}\in\partial {f^*}({\bf u}).
\]
In Section~\ref{sec:subgrad-inv}, we prove analogous results for functions on astral space.
In particular, regarding the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, we show
(Theorem~\ref{thm:adif-fext-inverses})
that
\[
{\bf u}\in\partial{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu(\overline{\xx})
\quad
\Longleftrightarrow
\quad
\overline{\xx}\in\e{\partial}{f^*}({\bf u}),
\]
provided $\overline{\xx}$ is in the closure (in $\extspac{n}$) of $f$'s effective domain (as will hold always if $f$ is finite everywhere).
These results can be applied to obtain a dual characterization of minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$.
For example, if $f$ is finite everywhere, then the minimizers of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu$ are exactly the slopes of the hyperplanes that support the epigraph of $f^*$ at the origin. This is exactly what we saw in Example~\ref{ex:exp:cont}.
Our notions of subdifferential differ
from the ``horizon subgradients''
of \citet{rockafellar1985extensions}
and \citet[Chapter 8]{rock_wets},
and also differ
from those introduced by \citet[Chapter 10]{singer_book},
even as our development of conjugacy builds on his abstract framework.
\bigskip
The differential theory and the theory of continuity on astral space are brought together
in the analysis of minimization algorithms.
Many such methods construct a
sequence $\seq{{\bf x}_t}$ in ${\R^n}$ whose gradients $\nabla f({\bf x}_t)$ converge to~${\bf 0}$.
Such a sequence is guaranteed to converge to a minimizer whenever it converges to a point in~${\R^n}$~\citep[Theorem~24.4]{ROC}, and more generally, whenever the sequence $\seq{{\bf x}_t}$ remains within a bounded region of ${\R^n}$.
This is not surprising since any finite point where $f$'s gradient is zero is guaranteed to be a global minimum, as a result of $f$ being convex.
We might, therefore, expect the same to hold when $\seq{{\bf x}_t}$ is not bounded.
In other words, even in the absence of additional assumptions
(but still assuming $f$ is convex),
if the gradients $\nabla f({\bf x}_t)$ are converging to ${\bf 0}$,
we might reasonably expect the function values $f({\bf x}_t)$ to be correspondingly converging to $f$'s infimum. However,
this turns out to be false, in general, as shown in \Cref{sec:subdif:cont}
with an example of a well-behaved, convex function $f$
with $\inf f = 0$,
and a sequence $\seq{{\bf x}_t}$
on which $\nabla f({\bf x}_t)\rightarrow{\bf 0}$ but
$f({\bf x}_t)\rightarrow+\infty$.
Not coincidentally, the sequence used in that example converges to a point where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is discontinuous.
Indeed, when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ can be assumed to be continuous (and not $+\infty$) at points of convergence, we prove that driving the gradients to zero suffices for minimizing the function; that is, if
$\nabla f({\bf x}_t)\to{\bf 0}$ then $f({\bf x}_t)\rightarrow\inf f$
(Theorem~\ref{thm:subdiff-min-cont}).
Apparently, continuity of the astral function ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is linked to the effectiveness of general gradient-based methods for minimizing a convex function.
Moving from sequences to algorithms,
Theorem~\ref{thm:fact:gd}
broadly proves the convergence of iterative-descent methods (such as gradient descent and certain types of steepest descent and coordinate descent) whose iterates
$\seq{{\bf x}_t}$ and some associated subgradients ${\bf u}_t\in\partial f({\bf x}_t)$ satisfy a generic progress condition,
provided ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is suitably continuous at points of convergence
(otherwise, such a result cannot hold in general, as just discussed).
This generic result is then applied to address our original motivation, the analysis of minimization methods
on standard objective functions.
Having come full circle, we prove, rather easily at this point,
the convergence
of several standard algorithms on objectives that are of practical importance,
but which do not always have finite minimizers,
such as likelihood functions and empirical-risk functions.
\section{Preliminaries}
\label{sec:more-prelim}
This short section introduces notational conventions, standard definitions, and some basic results used in the manuscript.
Throughout this manuscript, we usually use $t=1,2,\ldots$ as the
index of the sequences and write $\seq{{\bf x}_t}$ to mean
the sequence ${\bf x}_1,{\bf x}_2,\ldots$. Limits and convergence are taken as $t\rightarrow+\infty$,
unless stated otherwise. For example, $\lim({\bf x}_t\cdot{\bf u})$ refers to
$\lim_{t\to\infty}{\bf x}_t\cdot{\bf u}$.
We generally denote scalars, like $x$, in italics;
vectors in ${\R^n}$, like ${\bf x}$, in bold;
and points in astral space, like $\overline{\xx}$, in bold
with a bar or overline.
Vectors ${\bf x}\in{\R^n}$ are usually understood to have components
$x_i$ so that
${\bf x}=\trans{[x_1,\ldots,x_n]}$.
The usual Euclidean norm of a vector ${\bf x}\in{\R^n}$
is denoted $\norm{{\bf x}}$.
We write ${\bf e}_1,\ldots,{\bf e}_n\in{\R^n}$ for the
standard basis vectors
(so the $i$-th component of ${\bf e}_i$ is equal to $1$, and all others
are $0$).
We write $\eR$ for the set of extended reals $[-\infty,+\infty]$.
The sum of $-\infty$ and $+\infty$ is undefined, whereas
$0\cdot(\pm\infty)$ is defined to be $0$.
All the other sums and products involving $\pm\infty$ are
defined in the standard way.
For any
$f:{\cal X}\rightarrow \overline{\R}$,
we define the \emph{epigraph} of $f$ as the set of pairs $\rpair{x}{y}$, with $x\in{\cal X}$ and $y\in{\mathbb{R}}$,
for which $f(x)\leq y$:
\begin{equation}
\label{eqn:epi-def}
\epi{f} = \{ \rpair{x}{y} \in {\cal X}\times{\mathbb{R}}:\: f(x) \leq y \}.
\end{equation}
The \emph{effective domain} of $f$
(or simply its \emph{domain}, for short)
is the set of points where $f(x)<+\infty$:
\[
\dom{f} = \{ x \in {\cal X}:\: f(x) < +\infty \}.
\]
We write $\inf f$ for the function's infimum,
$\inf_{x\in {\cal X}} f(x)$, and similarly define
$\sup f$, as well as $\min f$ and $\max f$,
when these are attained.
For any $a\in\overline{\R}$, we write
$f\equiv a$
to mean $f(x)=a$ for all $x\in{\cal X}$.
Likewise, $f>a$ means
$f(x)>a$ for all $x\in{\cal X}$, with
$f\geq a$, $f<a$, $f\leq a$ defined similarly.
For any
$f:{\cal X}\rightarrow \overline{\R}$
and
$g:{\cal X}\rightarrow \overline{\R}$,
we write $f=g$ to mean $f(x)=g(x)$ for all $x\in{\cal X}$, and similarly
define $f<g$, $f\leq g$, etc.
We say that $f$ \emph{majorizes} $g$ if $f\geq g$.
A set $S\subseteq{\R^n}$ is \emph{convex} if
the line segment joining any two points in $S$ is also
entirely included in $S$.
A function $f:{\R^n}\rightarrow\overline{\R}$ is \emph{convex} if its epigraph is
convex.
The function is \emph{concave} if $-f$ is convex.
The function is \emph{lower semicontinuous} if
$\liminf f({\bf x}_t)\geq f({\bf x})$ for every sequence $\seq{{\bf x}_t}$
in ${\R^n}$ converging to ${\bf x}$.
The function is \emph{finite everywhere} if $f>-\infty$ and $f<+\infty$.
If convex, the function
is said to be \emph{proper} if $f > -\infty$ and $f\not\equiv+\infty$,
and it is \emph{closed} if it is lower
semicontinuous and either $f > -\infty$ or $f\equiv-\infty$.
(This definition of a closed, convex function follows
\citet[Section~7]{ROC}; note importantly, however, that there is not full
agreement in the literature on this terminology.)
Here is a standard characterization of convexity, applicable to both
proper and improper functions, that is taken
from \citet[Theorem~4.2]{ROC}:
\begin{proposition} \label{pr:rock:4.2}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then $f$ is convex if and only if it has the property that
for all ${\bf x},{\bf y}\in{\R^n}$ and all $a,b\in{\mathbb{R}}$ and all $\lambda\in [0,1]$,
if $f({\bf x})<a$ and $f({\bf y})<b$ then
\[
f(\lambda {\bf x} + (1-\lambda) {\bf y})
< \lambda a + (1-\lambda) b.
\]
\end{proposition}
We write $\zerov{n}$ for the all-zeros vector in ${\R^n}$, or simply
${\bf 0}$ when the dimension $n$ is clear from context.
We also write $\zeromat{m}{n}$ for the $m\times n$ matrix with entries
that are all equal to zero.
We will find it mathematically convenient to allow vectors
to be zero-dimensional, and matrices to have zero rows or columns,
while still being of particular dimensions that define when they can
be combined with other matrices in the usual way.
As such, we define $0$-dimensional Euclidean space, ${\mathbb{R}}^0$, to consist
of a single point, the ``empty'' vector, denoted $\zerov{0}$.
This point behaves like a vector with all components equal to zero;
thus, $\zerov{0}+\zerov{0}=\zerov{0}$,
$\zerov{0}\cdot\zerov{0}=0$ (in ${\mathbb{R}}$),
and
$\lambda\zerov{0}=\zerov{0}$ for all $\lambda\in{\mathbb{R}}$.
The topology on ${\mathbb{R}}^0$ is the only one possible for a single-point
space.
Likewise, for $n\geq 0$, we define
$\zeromat{n}{0}$ and $\zeromat{0}{n}$
to be the unique $n\times 0$ and $0\times n$
matrices, respectively; thus,
${\mathbb{R}}^{n\times 0}=\{\zeromat{n}{0}\}$
and
${\mathbb{R}}^{0\times n}=\{\zeromat{0}{n}\}$.
As before, these behave as if all entries were equal to zero, so that
$\zeromat{n}{0}+\zeromat{n}{0}=\zeromat{n}{0}$,
and
$\lambda\zeromat{n}{0}=\zeromat{n}{0}$
for $\lambda\in{\mathbb{R}}$;
similarly for $\zeromat{0}{n}$.
For multiplying such matrices, for all $n\geq 0$, $m\geq 0$, $k\geq 0$,
and for all ${\bf A}\in {\mathbb{R}}^{n\times k}$ and ${\bf B}\in{\mathbb{R}}^{k\times m}$,
we define
\[ {\bf A} \zeromat{k}{m} = \zeromat{n}{k} {\bf B} = \zeromat{n}{m}. \]
In particular, this implies
$\zeromat{n}{0} \, \zeromat{0}{m} = \zeromat{n}{m}$,
and that
$\zeromat{0}{0}$ is the
$0\times 0$ identity matrix.
Also,
just as vectors in ${\R^n}$ are treated as $n\times 1$ matrices,
we identify $\zerov{0}$ with $\zeromat{0}{1}$ so that
$\zeromat{n}{0} \zerov{0} = \zerov{n}$.
It can be checked that these definitions preserve all of the standard
properties of matrix operations, such as associativity and
distributivity.
\section{Astral space} \label{sec:astral-space-intro}
\subsection{Motivation}
\label{subsec:astral-intro-motiv}
We begin our development with the formal construction
of {astral space},
a compact enlargement of Euclidean space ${\R^n}$
in which various points ``at infinity'' have been added,
as briefly introduced in Section~\ref{sec:intro:astral}.
Astral points represent the destinations of infinitely
long trajectories through ${\R^n}$.
Therefore,
to construct the space, we need to decide which points to add, and
which sequences should be regarded as having the same or different
limits in the new space.
To motivate our approach, we consider a few simple examples,
focusing now on sequences rather than minimization of functions,
as in Section~\ref{sec:intro}.
In the simplest case of $n=1$ dimensions, the only reasonable
destinations for an unbounded sequence $\seq{x_t}$ of points in ${\mathbb{R}}$
are $+\infty$ and $-\infty$,
leading to the standard compactification of ${\mathbb{R}}$ in which these two
points have been added,
namely,
$\overline{\R} = {\mathbb{R}} \cup \{-\infty,+\infty\} = [-\infty,+\infty]$.
When we move to $n=2$ dimensions, there are many more possibilities
and issues to consider.
Consider first the sequence ${\bf x}_t=\trans{[2t,t]}=t{\bf v}$, where
${\bf v}=\trans{[2,1]}$, for $t=1,2,\ldots$.
This sequence follows along a ray in the direction of ${\bf v}$.
In ${\mathbb{R}}^2$, this sequence does not converge.
But in our compactification, sequences of this form along rays from
the origin do converge, with a particular limit that we will later
denote $\limray{{\bf v}}$.
Intuitively, ``similar'' sequences that ``eventually follow''
this same ray should have the same limit; for instance, we might expect
sequences like
$\sqrt{t}{\bf v}$ or $\trans{[2t,t+1/t]}$ to have the same limit (as
will be the case in astral space).
Consider next the sequence $\seq{{\bf y}_t}$ with elements
${\bf y}_t=\trans{[2t-1,t+1]}=t{\bf v}+{\bf w}$, where
${\bf w}=\trans{[-1,1]}$, which is certainly similar to the $\seq{{\bf x}_t}$
sequence with both following parallel, unbounded trajectories along rays
in the direction of ${\bf v}$, a tendency that seems to be overwhelming as
$t$ gets large.
The sequences differ only in their starting points,
with $\seq{{\bf x}_t}$ beginning its trajectory at the origin,
but $\seq{{\bf y}_t}$ beginning at ${\bf w}$.
Does this seemingly small difference matter as $t$ gets large?
In other words, do we want to regard these two sequences as having the
same or different limits?
To see the issues involved,
let us consider how the trajectories progress in various directions.
In the direction of ${\bf v}$, the dominant direction of convergence, both
sequences are clearly heading to $+\infty$; that is, ${\bf x}_t\cdot{\bf v}$,
the projection of ${\bf x}_t$ in the direction of ${\bf v}$, is converging to
$+\infty$, and likewise for ${\bf y}_t$.
So in this direction, the sequences' behaviors are the same.
We can similarly consider projections of the sequences in other
directions ${\bf u}$.
Indeed, the direction ${\bf v}$ is so dominant that
if ${\bf v}\cdot{\bf u}>0$, then both ${\bf x}_t\cdot{\bf u}$ and ${\bf y}_t\cdot{\bf u}$ will
converge to $+\infty$ so that, in all these directions, the sequences
again appear asymptotically the same.
But if we project onto a direction ${\bf u}$ that is perpendicular to
${\bf v}$, say, ${\bf u}=\trans{[1,-2]}$, the situation changes, with
${\bf x}_t\cdot{\bf u}\rightarrow 0$ and
${\bf y}_t\cdot{\bf u}\rightarrow {\bf w}\cdot{\bf u} = -3$.
So viewed in this direction, the sequences appear rather
different in terms of their limits.
As a consequence, if $\seq{{\bf x}_t}$ and $\seq{{\bf y}_t}$
are regarded as having the same limit in our enlarged space, then the
simple function
$f(x_1,x_2)= x_1 - 2 x_2$
(that is, $f({\bf x})={\bf x}\cdot{\bf u}$, with ${\bf u}$ as above)
cannot be extended continuously to this space since
$\lim f({\bf x}_t) \neq \lim f({\bf y}_t)$.
Therefore, and more generally, if we want linear functions to be
continuous in the enlarged space we are constructing, then we need to
treat sequences as having distinct limits if they differ in their
limit in any direction.
In this example, this means that $\seq{{\bf y}_t}$ should have a different
limit from $\seq{{\bf x}_t}$, as will be the case in astral space.
As a last example, consider the sequence
${\bf z}_t= t {\bf v} + \sqrt{t} {\bf w} =
\trans{[2 t - \sqrt{t}, t + \sqrt{t}]}$.
Like the last two sequences,
this one is moving to infinity most rapidly in the direction of
${\bf v}$.
But this sequence is also growing to infinity, though at a slower
rate, in the direction of ${\bf w}$.
As before, we can examine how the sequence evolves in various directions.
As was the case earlier, since ${\bf v}$ is so dominant,
if ${\bf v}\cdot{\bf u}>0$, then ${\bf z}_t\cdot{\bf u}\rightarrow+\infty$
(and similarly,
if ${\bf v}\cdot{\bf u}<0$, then ${\bf z}_t\cdot{\bf u}\rightarrow -\infty$).
In these cases, the secondary direction ${\bf w}$ is
irrelevant (asymptotically).
But if we look in a direction ${\bf u}$ that is perpendicular to ${\bf v}$,
then ${\bf v}$'s effect on the sequence vanishes,
and ${\bf w}$'s secondary effect becomes apparent;
in particular,
${\bf z}_t\cdot{\bf u}=\sqrt{t} {\bf w}\cdot{\bf u}$ so the limit of
${\bf z}_t\cdot{\bf u}$ is determined by the sign of ${\bf w}\cdot{\bf u}$.
For example, if ${\bf u}=\trans{[1,-2]}$, as before, then
${\bf z}_t\cdot{\bf u}\rightarrow-\infty$ (since ${\bf w}\cdot{\bf u}=-3<0$),
which differs from the limits of either ${\bf x}_t\cdot{\bf u}$
or ${\bf y}_t\cdot{\bf u}$.
Accordingly, the limit of $\seq{{\bf z}_t}$ should be considered distinct
from that of either $\seq{{\bf x}_t}$ or $\seq{{\bf y}_t}$.
As suggested by this example, astral space will include not only the limits
of sequences along rays, but also sequences which are
growing in multiple directions at varying rates.
\subsection{The construction}
\label{sec:astral:construction}
As discussed in Section~\ref{sec:intro:astral},
our construction of astral space is based centrally on
the asymptotic behavior of sequences when projected in various
directions.
The space will consist of ${\R^n}$ together with newly constructed points
at infinity representing the limits of sequences.
Moreover, our construction will ensure that two convergent
sequences will have different limits in astral space
if and only if their projections
differ asymptotically in one or more directions.
Formally, let $\seq{{\bf x}_t}$ be a sequence in ${\R^n}$.
We say that $\seq{{\bf x}_t}$ \emph{converges in all directions} if
for all ${\bf u}\in{\R^n}$, the sequence ${\bf x}_t\cdot{\bf u}$ converges to a limit
in $\overline{\R}$.
All of the example sequences in
Section~\ref{subsec:astral-intro-motiv} converge in all directions.
This is also true for
any sequence $\seq{{\bf x}_t}$ which has a limit ${\bf x}\in{\R^n}$
(since then ${\bf x}_t\cdot{\bf u}\rightarrow{\bf x}\cdot{\bf u}$, for all ${\bf u}$).
Let ${\cal S}$ denote the set of all sequences that converge in
all directions.
In accord with our approach,
we say that two sequences $\seq{{\bf x}_t}$ and $\seq{{\bf y}_t}$, both in
${\cal S}$, are \emph{all-directions equivalent}
(or just \emph{equivalent}, when the context is clear)
if they have the same
limits in every direction, that is, if for all ${\bf u}\in{\R^n}$,
$\lim ({\bf x}_t\cdot{\bf u}) = \lim ({\bf y}_t\cdot{\bf u})$.
This is easily checked to be an equivalence relation.
As discussed above, we want two sequences to have the
same limit in astral space if and only if they are equivalent in
exactly this sense.
As such, we can use all-directions equivalence to partition ${\cal S}$
into equivalence classes, and use those classes to define the astral
space.
Let $\Pi$ be the resulting collection of equivalence classes;
that is, the sets in $\Pi$ are all nonempty, their union is all of
${\cal S}$, and two sequences are in the same set in $\Pi$ if and only if
they are all-directions equivalent.
Each equivalence class corresponds to a common destination of the
sequences included in that class.
As such, we will define astral space so that every point is
effectively identified with one of the equivalence classes of $\Pi$.
Note also that for every point ${\bf x}\in{\R^n}$, there exists one
equivalence class in $\Pi$ consisting exactly of all sequences that
converge to ${\bf x}$; naturally, we will want to identify ${\bf x}$ with this
class so that ${\R^n}$ is included in the space.
Formally, the \emph{astral space} is a set denoted $\extspac{n}$ such
that the following hold:
\begin{itemize}
\item
There exists a bijection $\pi:\extspac{n}\rightarrow\Pi$ identifying
each element of $\extspac{n}$ with an equivalence class in $\Pi$.
\item
${\R^n}\subseteq\extspac{n}$.
\item
For all ${\bf x}\in{\R^n}$, $\pi({\bf x})$ is the equivalence class consisting
of all sequences that converge to ${\bf x}$, establishing the natural
correspondence discussed above.
\end{itemize}
In the special case that $n=1$, we choose $\extspac{1}=\overline{\R}$, with
$\pi$ defined in the most natural way.
This is possible because, for every $\bar{x}\in\overline{\R}$ (including
$\pm\infty$), there is one equivalence class consisting of all sequences
$\seq{x_t}$ in ${\mathbb{R}}$ that converge to $\bar{x}$;
naturally, we define $\pi(x)$ to be equal to this class.
Furthermore, these are the only equivalence classes in $\Pi$.
When $n=0$, it follows from definitions that
$\extspac{0}={\mathbb{R}}^0=\{\zerov{0}\}$ since the only possible sequence
has every element equal to ${\mathbb{R}}^0$'s only point, $\zerov{0}$.
We will later define a natural topology for $\extspac{n}$.
For every point $\overline{\xx}\in\extspac{n}$, we will later see that, in this
topology, every sequence $\seq{{\bf x}_t}$ in the associated equivalence
class $\pi(\overline{\xx})$ converges to
$\overline{\xx}$, so that the astral point $\overline{\xx}$ truly can be understood as
the destination of sequences, as previously discussed.
We will also see that $\extspac{n}$ is compact, and thus a
compactification of ${\R^n}$.
Let $\overline{\xx}\in\extspac{n}$, and let ${\bf u}\in{\R^n}$.
By construction,
all of the sequences $\seq{{\bf x}_t}$ in $\pi(\overline{\xx})$ have the same
limit when projected in the direction
of ${\bf u}$, that is, $\lim ({\bf x}_t\cdot{\bf u})$ is the same for all
sequences in $\pi(\overline{\xx})$.
We use the notation $\overline{\xx}\cdot{\bf u}$ to denote this common limit.
That is, we define
\[ \overline{\xx}\cdot{\bf u} = {\bf u}\cdot\overline{\xx} = \lim ({\bf x}_t\cdot{\bf u}), \]
where $\seq{{\bf x}_t}$ is any sequence in $\pi(\overline{\xx})$ (noting that the
same value will result, regardless of which one is selected).
Note that if $\overline{\xx}={\bf x}$ where ${\bf x}\in{\R^n}$, then
we will have $\overline{\xx}\cdot{\bf u}={\bf x}\cdot{\bf u}$
since
${\bf x}_t\rightarrow{\bf x}$ for every $\seq{{\bf x}_t}$ in $\pi({\bf x})$;
in other words, this notation is compatible with the usual definition
of ${\bf x}\cdot{\bf u}$ as the inner product between ${\bf x}$ and ${\bf u}$.
Note also that $\overline{\xx}\cdot{\bf 0}=0$ for all $\overline{\xx}\in\extspac{n}$.
The operation $\overline{\xx}\cdot{\bf u}$ is critically central to our
development;
indeed, we will see that all properties of astral points
$\overline{\xx}\in\extspac{n}$ can be expressed in terms of the values of
$\overline{\xx}\cdot{\bf u}$ over all ${\bf u}\in{\R^n}$.
As a start,
these values uniquely determine $\overline{\xx}$'s identity:
\begin{proposition} \label{pr:i:4}
Let $\overline{\xx}$ and $\overline{\xx}'$ be in $\extspac{n}$.
Then $\overline{\xx}=\overline{\xx}'$ if and only if
for all ${\bf u}\in{\R^n}$,
$\overline{\xx}\cdot{\bf u}=\overline{\xx}'\cdot{\bf u}$.
\end{proposition}
\proof
Suppose $\overline{\xx}\cdot{\bf u}=\overline{\xx}'\cdot{\bf u}$ for all ${\bf u}\in{\R^n}$.
If $\seq{{\bf x}_t}$ is a sequence in $\pi(\overline{\xx})$, then
${\bf x}_t\cdot{\bf u}\rightarrow \overline{\xx}\cdot{\bf u}=\overline{\xx}'\cdot{\bf u}$,
for all ${\bf u}\in{\R^n}$, implying $\seq{{\bf x}_t}$ is also in $\pi(\overline{\xx}')$.
Being equivalence classes, since $\pi(\overline{\xx})$ and $\pi(\overline{\xx}')$ have a
nonempty intersection, they must actually be equal,
which implies $\overline{\xx}=\overline{\xx}'$ since $\pi$ is a bijection.
\qed
Despite the suggestiveness of the notation,
$\overline{\xx}\cdot{\bf u}$ is not actually an inner product.
But it has some properties that make it sometimes similar to one, as
seen in the next two propositions.
The first shows that it is partially distributive, except when
adding $-\infty$ with $+\infty$ would be involved.
As used in this proposition,
we say that two numbers $\alpha$ and $\beta$ in $\overline{\R}$ are
\emph{summable}
if $\{\alpha,\beta\}\neq\{-\infty,+\infty\}$,
that is, if their sum, $\alpha+\beta$, is defined.
(Equivalently, $\alpha$ and $\beta$ are summable
if either $\min\{\alpha,\beta\}>-\infty$
or $\max\{\alpha,\beta\}<+\infty$.)
\begin{proposition} \label{pr:i:1}
Let $\overline{\xx}\in\extspac{n}$ and ${\bf u},{\bf v}\in{\R^n}$.
Suppose
$\overline{\xx}\cdot{\bf u}$ and $\overline{\xx}\cdot{\bf v}$
are summable.
Then
\[ \overline{\xx}\cdot({\bf u}+{\bf v}) = \overline{\xx}\cdot{\bf u} + \overline{\xx}\cdot{\bf v}. \]
\end{proposition}
\proof
Let ${\bf x}_t\in{\R^n}$ be any sequence in $\pi(\overline{\xx})$.
Then
\begin{eqnarray*}
\overline{\xx}\cdot{\bf u} + \overline{\xx}\cdot{\bf v}
&=&
\lim {\bf x}_t\cdot{\bf u} + \lim{\bf x}_t\cdot{\bf v}
\\
&=&
\lim ({\bf x}_t\cdot{\bf u}+{\bf x}_t\cdot{\bf v})
\\
&=&
\lim {\bf x}_t\cdot({\bf u}+{\bf v})
\\
&=&
\overline{\xx}\cdot({\bf u}+{\bf v}).
\end{eqnarray*}
The first and last equalities are because $\seq{{\bf x}_t}$ is in
$\pi(\overline{\xx})$, and
the second equality is by continuity.
\qed
For any point $\overline{\xx}\in\extspac{n}$ and scalar $\lambda\in{\mathbb{R}}$, we define the
scalar product $\lambda\overline{\xx}$ to be the unique point in $\extspac{n}$ for which
$(\lambda\overline{\xx})\cdot{\bf u}=\lambda(\overline{\xx}\cdot{\bf u})$ for all ${\bf u}\in{\R^n}$.
The next proposition proves that such a point exists.
Note that when $\overline{\xx}={\bf x}$ is in ${\R^n}$,
$\lambda\overline{\xx}$ is necessarily
equal to the usual product $\lambda{\bf x}$ of scalar $\lambda$ with vector ${\bf x}$.
For the case $\lambda=0$, this proposition (combined with
Proposition~\ref{pr:i:4})
implies $0\overline{\xx}={\bf 0}$ for all $\overline{\xx}\in\extspac{n}$.
(Keep in mind that $0\cdot(\pm\infty)=0$.)
\begin{proposition} \label{pr:i:2}
Let $\overline{\xx}\in\extspac{n}$ and let $\lambda\in{\mathbb{R}}$.
Then there exists a unique point in $\extspac{n}$, henceforth denoted
$\lambda\overline{\xx}$, for which
\[ (\lambda\overline{\xx})\cdot{\bf u} = \lambda (\overline{\xx}\cdot{\bf u}) = \overline{\xx}\cdot (\lambda{\bf u}) \]
for all ${\bf u}\in{\R^n}$.
\end{proposition}
\proof
Let $\seq{{\bf x}_t}$ in ${\R^n}$ be any sequence in $\pi(\overline{\xx})$.
Then
\begin{eqnarray}
\overline{\xx}\cdot (\lambda{\bf u})
&=&
\lim {\bf x}_t\cdot (\lambda{\bf u})
\nonumber \\
&=&
\lim (\lambda{\bf x}_t) \cdot {\bf u}
\nonumber \\
&=&
\lim \lambda ({\bf x}_t\cdot{\bf u})
\nonumber \\
&=&
\lambda \lim {\bf x}_t\cdot{\bf u}
\label{eq:i:2a} \\
&=&
\lambda (\overline{\xx}\cdot{\bf u}),
\nonumber
\end{eqnarray}
where \eqref{eq:i:2a} is by continuity when multiplying a point in
$\overline{\R}$ by a constant $\lambda$.
Thus, the sequence $\seq{\lambda {\bf x}_t}$ is in ${\cal S}$ since,
for every ${\bf u}\in{\R^n}$,
$(\lambda {\bf x}_t) \cdot {\bf u}$ has a limit, namely,
$\overline{\xx}\cdot (\lambda{\bf u})=\lambda(\overline{\xx}\cdot{\bf u})$.
Thus, this sequence is in the equivalence class $\pi(\overline{{\bf y}})$ for some
$\overline{{\bf y}}\in\extspac{n}$ (which must be unique by Proposition~\ref{pr:i:4}).
Defining $\lambda\overline{\xx}$ to be $\overline{{\bf y}}$ proves the result.
\qed
Points in ${\R^n}$ are characterized by being finite in every direction
${\bf u}$:
\begin{proposition} \label{pr:i:3}
Let $\overline{\xx}\in\extspac{n}$.
Then $\overline{\xx}\in{\R^n}$ if and only if
for all ${\bf u}\in{\R^n}$,
$\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$.
\end{proposition}
\proof
Suppose $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$ for all ${\bf u}\in{\R^n}$.
Let ${\bf e}_1,\ldots,{\bf e}_n\in{\R^n}$ be standard basis vectors.
Let $x_i=\overline{\xx}\cdot{\bf e}_i$, and
${\bf x}=\trans{[x_1,\ldots,x_n]}$.
For any ${\bf u}\in{\R^n}$, we can write
${\bf u}=\sum_{i=1}^n u_i {\bf e}_i$.
Then by repeated application of Propositions~\ref{pr:i:1}
and~\ref{pr:i:2},
\[
\overline{\xx}\cdot{\bf u} = \sum_{i=1}^n u_i (\overline{\xx}\cdot{\bf e}_i)
= \sum_{i=1}^n u_i x_i = {\bf x}\cdot{\bf u}.
\]
Since this holds for all ${\bf u}\in{\R^n}$, $\overline{\xx}$ must be equal to ${\bf x}$,
by Proposition~\ref{pr:i:4},
and therefore is in ${\R^n}$.
The reverse implication is immediate.
\qed
It is crucial to distinguish astral space, $\extspac{n}$, from
$(\overline{\R})^n$, the $n$-fold Cartesian product of $\overline{\R}$ with itself.
Points in either space can be regarded as the limits of possibly
unbounded sequences in ${\R^n}$.
But astral points embody far more information.
To see this, suppose some sequence $\seq{{\bf x}_t}$ in ${\R^n}$
converges to a point ${\hat{\bf x}}$ in $(\overline{\R})^n$.
Then ${\hat{\bf x}}$, by its form, encodes the limit of
$({\bf x}_t\cdot{\bf e}_i)$, for each standard basis vector ${\bf e}_i$,
since this limit is exactly the $i$-th component of ${\hat{\bf x}}$.
In comparison, if instead the sequence $\seq{{\bf x}_t}$ converges to an
astral point $\overline{\xx}$ in $\extspac{n}$, then $\overline{\xx}$ encodes the limit
of $({\bf x}_t\cdot{\bf u})$ for \emph{all} directions ${\bf u}\in{\R^n}$, not just
along the coordinate axes
(as will be seen later in Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
For instance, if $n=2$ and ${\hat{\bf x}}=\trans{[+\infty,+\infty]}$
in $(\overline{\R})^2$,
then ${\bf x}_t\cdot{\bf e}_i\rightarrow+\infty$, for $i=1,2$.
From this information,
if ${\bf u}={\bf e}_1-{\bf e}_2$,
for example,
there is no way to deduce the limit of
$\seq{{\bf x}_t\cdot{\bf u}}$, or even if the limit exists.
On the other hand, this limit, in this or any other direction ${\bf u}\in{\mathbb{R}}^2$,
would be readily available from $\overline{\xx}\in\extspac{2}$.
Thus, as limits of sequences,
an astral point $\overline{\xx}\in\extspac{n}$
retains all of the information embodied
by a point ${\hat{\bf x}}$ in $(\overline{\R})^n$, and
usually far more.
This results in astral space having a remarkably rich, powerful structure,
as will be seen in the course of our development.
\subsection{Astral points}
\label{subsec:astral-pt-form}
What is the nature of points comprising astral space?
The space includes all of ${\R^n}$, of course.
All of the other ``new'' points are limits of unbounded sequences
$\seq{{\bf x}_t}$ in ${\R^n}$
(that is, for which $\norm{{\bf x}_t}\rightarrow+\infty$).
This follows from Proposition~\ref{pr:i:3} which shows that any
point $\overline{\xx}$ in $\extspac{n}\setminus{\R^n}$ must be infinite in some
direction ${\bf u}$, and therefore ${\bf x}_t\cdot{\bf u}$ converges to
$\pm\infty$ for any sequence $\seq{{\bf x}_t}$ in $\pi(\overline{\xx})$.
Thus, points in ${\R^n}$ are said to be \emph{finite}, and all other
points are \emph{infinite}.
We will see that all points in $\extspac{n}$ have a
specific structure, as suggested by the examples in
Section~\ref{subsec:astral-intro-motiv}.
Every astral point $\overline{\xx}$, outside those in ${\R^n}$, is the limit of
sequences which have a particular dominant direction ${\bf v}$ in which the
sequence is growing to infinity most rapidly.
In addition, these sequences may be growing to infinity in other
directions that are secondary, tertiary, etc.
These sequences may also have a finite part in the sense of converging
to a finite value in some directions.
Importantly, the details of this structure are entirely determined by
the point $\overline{\xx}$ itself so that every sequence that converges to
$\overline{\xx}$ will have the same dominant direction, same finite part, etc.
To be more concrete while also generalizing the examples in
Section~\ref{subsec:astral-intro-motiv}, let ${\bf q}\in{\R^n}$, and let
${\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$, for some $k\geq 0$.
Consider the sequence
\begin{equation} \label{eq:h:6}
{\bf x}_t = t^k {\bf v}_1 + t^{k-1} {\bf v}_2 + \cdots + t {\bf v}_k + {\bf q}
= \sum_{i=1}^k t^{k-i+1} {\bf v}_i + {\bf q}.
\end{equation}
As usual, we focus on the limit of ${\bf x}_t\cdot{\bf u}$ in various
directions ${\bf u}\in{\R^n}$.
The evolution of this particular sequence is clearly dominated
by its overwhelmingly rapid growth in the direction of ${\bf v}_1$.
As a result, similar to what was seen in our earlier examples,
if ${\bf v}_1\cdot{\bf u}>0$ then ${\bf x}_t\cdot{\bf u}\rightarrow +\infty$,
and
if ${\bf v}_1\cdot{\bf u}<0$ then ${\bf x}_t\cdot{\bf u}\rightarrow -\infty$.
If ${\bf v}_1\cdot{\bf u}=0$,
then the term involving ${\bf v}_1$ vanishes when considering
${\bf x}_t\cdot{\bf u}$.
So when projecting the sequence in directions perpendicular to
${\bf v}_1$, ${\bf v}_2$ becomes dominant.
In these directions, we find once again that ${\bf x}_t\cdot{\bf u}$ converges
to $+\infty$ or $-\infty$ if ${\bf v}_2\cdot{\bf u}>0$ or ${\bf v}_2\cdot{\bf u}<0$,
respectively.
This analysis can be continued, so that we next consider vectors
${\bf u}$ in the subspace orthogonal to both ${\bf v}_1$ and ${\bf v}_2$ where
${\bf v}_3$ is dominant.
And so on.
Eventually, for vectors ${\bf u}$ that are orthogonal to all
${\bf v}_1,\ldots,{\bf v}_k$, we find that
${\bf x}_t\cdot{\bf u}$ converges to the finite value ${\bf q}\cdot{\bf u}$.
Summarizing, this argument shows that the sequence ${\bf x}_t$ converges
to $\overline{\xx}\in\extspac{n}$ (or equivalently, $\seq{{\bf x}_t}$ is in $\overline{\xx}$'s
equivalence class, $\pi(\overline{\xx})$)
where, for ${\bf u}\in{\R^n}$,
\begin{equation} \label{eqn:pf:2}
\overline{\xx}\cdot{\bf u} = \left\{ \begin{array}{cl}
+\infty & \mbox{if ${\bf v}_i\cdot{\bf u}>0$ and
${\bf v}_{1}\cdot{\bf u}=\cdots={\bf v}_{i-1}\cdot{\bf u}=0$
for some $i\in\{1,\ldots,k\}$} \\
-\infty & \mbox{if ${\bf v}_i\cdot{\bf u}<0$ and
${\bf v}_{1}\cdot{\bf u}=\cdots={\bf v}_{i-1}\cdot{\bf u}=0$
for some $i\in\{1,\ldots,k\}$} \\
{\bf q}\cdot{\bf u} & \mbox{if
${\bf v}_1\cdot{\bf u}=\cdots={\bf v}_k\cdot{\bf u}=0$.}
\end{array}
\right.
\end{equation}
This example and argument show that we can construct points $\overline{\xx}$ in
$\extspac{n}$ that have the form given in \eqref{eqn:pf:2},
for any ${\bf q},{\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$.
In fact, we will later prove the converse, that every point
$\overline{\xx}\in\extspac{n}$ must have exactly this form for some
${\bf q},{\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$.
This specific structure is fundamental to understanding $\extspac{n}$,
and will be central to the foundation on which all of the results
appearing later in the manuscript are based.
\section{Astral topology}
\label{sec:astral-as-fcns}
Astral space was constructed to consist of points representing
destinations of sequences.
Here, we present an alternative view of the space in which
astral points are instead regarded as functions,
a perspective that will allow us
to prove several of the central properties listed earlier, such as
compactness of astral space as well as the structural form of all
astral points.
\subsection{Astral points as functions}
\label{subsec:astral-pts-as-fcns}
Every point $\overline{\xx}\in\extspac{n}$ can be naturally regarded as
defining a function $\ph{\overline{\xx}}:{\R^n}\rightarrow\overline{\R}$ given by
\begin{equation} \label{eq:ph-xbar-defn}
\ph{\overline{\xx}}({\bf u}) = \overline{\xx}\cdot{\bf u}
\end{equation}
for ${\bf u}\in{\R^n}$.
Every such function is included in
\[ {\cal F}=\overline{\R}^{{\R^n}}, \]
the space of all
functions mapping ${\R^n}$ to $\overline{\R}$.
Since $\overline{\R}$ is compact,
${\cal F}$ is also compact,
by Tychonoff's theorem~\citep[Theorem~37.3]{munkres},
since it is
a product of compact spaces, where we assume a standard product
topology on ${\cal F}$.
(By invoking Tychonoff's theorem, we are hereby assuming the axiom of
choice throughout this work.)
Let $\Phi\subseteq{\cal F}$ be the space of all such functions
$\ph{\overline{\xx}}$, that is,
\[ \Phi = \{ \ph{\overline{\xx}} : \overline{\xx}\in\extspac{n} \}. \]
Clearly, there is a natural correspondence between $\extspac{n}$ and
$\Phi$ that is given formally by the map
${\varphi}:\extspac{n}\rightarrow\Phi$
where ${\varphi}(\overline{\xx})=\ph{\overline{\xx}}$
for $\overline{\xx}\in\extspac{n}$.
This map is a bijection, by Proposition~\ref{pr:i:4}.
Furthermore,
we will soon define a topology on $\extspac{n}$ under which ${\varphi}$ is
also a homeomorphism so that topological properties proved for $\Phi$
will apply to $\extspac{n}$ as well.
In particular, we will prove below that $\Phi$ is closed, and
therefore compact (since ${\cal F}$ is compact), which will show that
$\extspac{n}$ is compact too.
We will also prove below that functions in $\Phi$ have specific
convexity properties and a particular functional form.
When translated back to $\extspac{n}$, this will prove the structural
form for all points in $\extspac{n}$ discussed
in Section~\ref{subsec:astral-pt-form}.
As already mentioned, we assume a standard product
topology on ${\cal F}$ with basis elements of the form
\begin{equation} \label{eq:h:3}
\{\psi\in{\cal F} : s_i \psi({\bf u}_i) < b_i \mbox{ for all }
i=1,\ldots,k \},
\end{equation}
for some finite $k$, and some $s_i\in\{-1,+1\}$, $b_i\in{\mathbb{R}}$, ${\bf u}_i\in{\R^n}$,
for $i=1,\ldots,k$.
(Here, we are using a standard definition of product topology, while
also noting that $\overline{\R}$ has basis elements $(a,b)$, $(a,+\infty]$,
$[-\infty,b)$, for $a,b\in{\mathbb{R}}$, all of which can be written as an
intersection of one or two open half-lines.)
In this topology, as is generally true for the product topology,
if $\seq{\psi_t}$ is any sequence in ${\cal F}$,
and $\psi\in{\cal F}$, then
$\psi_t\rightarrow\psi$ if and only if
$\psi_t({\bf u})\rightarrow\psi({\bf u})$ for all ${\bf u}\in{\R^n}$.
The topology on the subspace $\Phi$ has basis elements
$B\cap\Phi$, where $B$ has the form given in \eqref{eq:h:3}.
We already noted that ${\varphi}$ defines a bijection between
$\extspac{n}$ and $\Phi$.
To ensure that ${\varphi}$ defines a homeomorphism as well,
we can simply define a topology on $\extspac{n}$,
called the \emph{astral topology},
in which a set
$U\subseteq\extspac{n}$ is open if and only if its image ${\varphi}(U)$ is
open in $\Phi$.
From \eqref{eq:h:3}, it is apparent that this is the topology defined
by standard basis elements of the form
\begin{equation} \label{eq:h:3a}
\{\overline{\xx}\in\extspac{n} : s_i (\overline{\xx}\cdot{\bf u}_i) < b_i \mbox{ for all }
i=1,\ldots,k \},
\end{equation}
for some finite $k$, and some $s_i\in\{-1,+1\}$, $b_i\in{\mathbb{R}}$, ${\bf u}_i\in{\R^n}$,
for $i=1,\ldots,k$.
(Actually, by Proposition~\ref{pr:i:2},
$s_i(\overline{\xx}\cdot{\bf u}_i)=\overline{\xx}\cdot(s_i{\bf u}_i)$, so, when convenient,
we can take $s_i=+1$ for all $i$.)
Equivalently, this is the topology defined
by standard \emph{subbasis} elements of the form
\begin{equation} \label{eq:h:3a:subbasis}
\{\overline{\xx}\in\extspac{n} : s (\overline{\xx}\cdot{\bf u}) < b \},
\end{equation}
for $s\in\{-1,+1\}$, $b\in{\mathbb{R}}$ and ${\bf u}\in{\R^n}$.
Henceforth, throughout this manuscript, we assume this topology
for $\extspac{n}$.
As we show next,
if restricted to ${\R^n}$, this topology is the same as the
standard topology for Euclidean space, with basis elements given by
open balls,
\begin{equation} \label{eqn:open-ball-defn}
{\cal B}({\bf x},\epsilon) = \{ {\bf z}\in{\R^n} : \norm{{\bf z}-{\bf x}} < \epsilon \}
\end{equation}
for ${\bf x}\in{\R^n}$ and $\epsilon>0$.
For the purposes of this proposition, we say that a set $U$ is open in
$\extspac{n}$ if it is open in the astral topology just described, and open in
${\R^n}$ if it is open in the standard Euclidean topology.
This proposition also shows that ${\R^n}$, with its usual topology, is a
topological subspace of $\extspac{n}$.
(The subspace topology of $\extspac{n}$ on ${\R^n}$ consists of all sets
$U\cap{\R^n}$ where $U\subseteq\extspac{n}$ is open in $\extspac{n}$.)
\begin{proposition} \label{pr:open-sets-equiv}
Let $U\subseteq{\R^n}$.
Then $U$ is open in $\extspac{n}$ if and only if
$U$ is open in ${\R^n}$.
Therefore, the Euclidean topology on ${\R^n}$ is the same as the
subspace topology of $\extspac{n}$ on ${\R^n}$.
\end{proposition}
\proof
Suppose first that $U$ is open in $\extspac{n}$.
Let ${\bf y}\in U$.
Then there exists a basis element $B$ as in \eqref{eq:h:3a} such that
${\bf y}\in B \subseteq U$.
By Proposition~\ref{pr:i:2}, we can assume without loss of generality
that, for $i=1,\ldots,k$, $s_i=+1$ and $\norm{{\bf u}_i}=1$.
(If ${\bf u}_i={\bf 0}$, then the constraint $\overline{\xx}\cdot{\bf u}_i<b_i$ can be
discarded; otherwise, we can divide both sides of the inequality by
$\norm{{\bf u}_i}$.)
Let $\epsilon$ be such that
\[ 0 < \epsilon < \min \braces{b_i - {\bf y}\cdot{\bf u}_i : i=1,\ldots,k}, \]
which must exist since ${\bf y}\in B$.
If ${\bf x}\in {\cal B}({\bf y},\epsilon)$, then
for $i=1,\ldots,k$,
\[
{\bf x}\cdot{\bf u}_i - {\bf y}\cdot{\bf u}_i
\leq
\norm{{\bf x}-{\bf y}} \norm{{\bf u}_i}
< \epsilon.
\]
Thus, by our choice of $\epsilon$,
${\bf x}\cdot{\bf u}_i < {\bf y}\cdot{\bf u}_i + \epsilon < b_i$,
for $i=1,\ldots,k$,
so ${\bf x}\in B$.
This means
${\bf y}\in{\cal B}({\bf y},\epsilon)\subseteq B\subseteq U$,
and so $U$ is open in ${\R^n}$.
Conversely, suppose now that $U$ is open in ${\R^n}$.
Let ${\bf y}\in U$.
Then there exists $\epsilon>0$ such that
${\cal B}({\bf y},\epsilon) \subseteq U$.
Let $B$ be the following basis element (in $\extspac{n}$):
\[
B = \braces{\overline{\xx}\in\extspac{n} :
|\overline{\xx}\cdot{\bf e}_i - {\bf y}\cdot{\bf e}_i| < \delta
\mbox{ for } i=1,\ldots,n \}
}
\]
where ${\bf e}_1,\ldots,{\bf e}_n$ are standard basis vectors,
and $\delta = \epsilon/\sqrt{n}$.
We claim that $B\subseteq{\R^n}$.
For if $\overline{\xx}\in B$, then we must have $\overline{\xx}\cdot{\bf e}_i\in{\mathbb{R}}$, for
$i=1,\ldots,n$.
Further, Propositions~\ref{pr:i:1} and~\ref{pr:i:2} show that the set
$\{{\bf u}\in{\R^n} : \overline{\xx}\cdot{\bf u}\in{\mathbb{R}}\}$ is a linear
subspace.
Since this set includes the $n$ standard basis vectors, it must be all of
${\R^n}$, which implies that $\overline{\xx}\in{\R^n}$ by Proposition~\ref{pr:i:3}.
So let ${\bf x}\in B\subseteq{\R^n}$.
Then $|({\bf x}-{\bf y})\cdot{\bf e}_i|<\delta$, for $i=1,\ldots,n$, implying
$\norm{{\bf x}-{\bf y}}<\delta\sqrt{n} = \epsilon$.
Thus, ${\bf y}\in B\subseteq{\cal B}({\bf y},\epsilon)\subseteq U$,
proving $U$ is open in $\extspac{n}$.
For the final claim,
if $U\subseteq{\R^n}$ is open in ${\R^n}$, then it is open in $\extspac{n}$
(as just proved),
so that $U=U\cap{\R^n}$ is open in the subspace topology.
Conversely, if $U$ is open in the subspace topology,
then $U=V\cap{\R^n}$ for some $V$ that is open in $\extspac{n}$,
implying, since ${\R^n}$ is open (in ${\R^n}$, and so also in $\extspac{n}$),
that $U$ is open in $\extspac{n}$, and so in ${\R^n}$ as well.
\qed
\subsection{Characterizing functional form}
\label{sec:char:func}
Let $\phfcn(\Rn)$ be the image of ${\R^n}$ under ${\varphi}$, and let $\overline{\phimg}$
be its closure in ${\cal F}$, meaning that
a function $\psi\in{\cal F}$ is in $\overline{\phimg}$ if and only
if every neighborhood $U$ of $\psi$ intersects $\phfcn(\Rn)$.
We will be especially interested in $\overline{\phimg}$ and properties of the
functions comprising it.
Eventually, we will prove that $\overline{\phimg}=\Phi$, which immediately
implies that $\Phi$ is closed and
therefore compact (so that $\extspac{n}$ is as well).
Proving inclusion in one direction is straightforward:
\begin{proposition} \label{pr:h:7}
For any $\overline{\xx}\in\extspac{n}$, $\ph{\overline{\xx}}\in\overline{\phimg}$.
Therefore, $\Phi\subseteq\overline{\phimg}$.
\end{proposition}
\proof
Let $\overline{\xx}\in\extspac{n}$.
By construction of $\extspac{n}$, there exists a sequence $\seq{{\bf x}_t}$
in ${\R^n}$
such that, for all ${\bf u}\in{\R^n}$,
${\bf x}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$, that is,
$\ph{{\bf x}_t}({\bf u})\rightarrow\ph{\overline{\xx}}({\bf u})$.
This pointwise convergence implies that
$\ph{{\bf x}_t}\rightarrow\ph{\overline{\xx}}$.
Since $\ph{{\bf x}_t}\in\phfcn(\Rn)$, for all $t$, this shows that
$\ph{\overline{\xx}}\in\overline{\phimg}$.
\qed
The next proposition gives some basic but useful properties of
functions in $\overline{\phimg}$:
\begin{proposition} \label{pr:h:2}
Let $\psi\in\overline{\phimg}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:h:2:a}
$\psi({\bf 0})=0$.
\item \label{pr:h:2:b}
$-\psi\in\overline{\phimg}$.
\end{enumerate}
\end{proposition}
\proof
Part~(\ref{pr:h:2:a}):
Suppose to the contrary that $\psi({\bf 0})\neq 0$.
The set
\[
\{\xi\in{\cal F} : \xi({\bf 0}) < 0\}
\cup
\{\xi\in{\cal F} : \xi({\bf 0}) > 0\}
\]
is open, being a union of two basis elements.
Since it is a neighborhood of $\psi$, it must include a point
$\ph{\xx}\in\phfcn(\Rn)$, for some ${\bf x}\in{\R^n}$.
But this means $\ph{\xx}({\bf 0})={\bf x}\cdot{\bf 0}\neq 0$, which is obviously
impossible.
Part~(\ref{pr:h:2:b}):
Let $U$ be any neighborhood of $-\psi$.
Then there exists a basis element $B$ with $-\psi\in B\subseteq U$.
Suppose $B$ is as given in \eqref{eq:h:3}, and let $B'$ be a basis
element that is the same as $B$, but with each $s_i$ replaced by
$-s_i$, that is,
\[
B' =
\{\xi\in{\cal F} : -s_i \xi({\bf u}_i) < b_i \mbox{ for all }
i=1,\ldots,k \}.
\]
Then $\psi\in B'$ (since $-\psi\in B$),
implying, since $\psi\in\overline{\phimg}$, that there exists a
point $\ph{\xx}\in B'\cap\phfcn(\Rn)$, and therefore that
$\ph{-{\bf x}}=-\ph{\xx}\in B\cap\phfcn(\Rn) \subseteq U\cap\phfcn(\Rn)$.
Thus, $-\psi\in\overline{\phimg}$.
\qed
All functions in $\overline{\phimg}$ are both convex and concave, as we show
next (although they are not generally proper functions).
In fact, we will soon see that such properties exactly characterize
the functions in
$\overline{\phimg}$, and therefore in $\Phi$ as well.
\begin{theorem} \label{thm:h:3}
Let $\psi:{\R^n}\rightarrow\overline{\R}$ be a function in $\overline{\phimg}$.
Then $\psi$ is both convex and concave.
\end{theorem}
\proof
To show $\psi$ is convex, we prove that it satisfies the condition in
Proposition~\ref{pr:rock:4.2}.
Suppose that $\psi({\bf u})<a$ and $\psi({\bf v})<b$ for some
${\bf u},{\bf v}\in{\R^n}$ and $a,b\in{\mathbb{R}}$.
Let $\lambda\in[0,1]$.
We aim to show that
\[
\psi(\lambda {\bf u} + (1-\lambda) {\bf v})
< \lambda a + (1-\lambda) b.
\]
This is trivial when $\lambda\in\{0,1\}$, so assume henceforth
that $\lambda\in(0,1)$.
Let ${\bf w}=\lambda{\bf u} + (1-\lambda){\bf v}$,
let $b'\in{\mathbb{R}}$ be such that $\psi({\bf v})<b'<b$,
and let $c=\lambda a + (1-\lambda) b'$.
Suppose, by way of contradiction, that $\psi({\bf w})>c$.
Let
\[
U = \{\xi\in{\cal F} : \xi({\bf w})>c, \xi({\bf u})<a, \xi({\bf v})<b'\}.
\]
Then $U$ is a neighborhood of $\psi$.
Therefore, there exists a point $\ph{\xx}\in U\cap\phfcn(\Rn)$, which means, by
definition of $\ph{\xx}$, that
\[
c < {\bf x}\cdot{\bf w}
= \lambda {\bf x}\cdot{\bf u} + (1-\lambda) {\bf x}\cdot{\bf v}
< \lambda a + (1-\lambda) b' = c,
\]
clearly a contradiction.
Thus,
\[ \psi({\bf w})\leq c < \lambda a + (1-\lambda) b, \]
so $\psi$ is convex.
Since this holds for all of $\overline{\phimg}$,
by Proposition~\ref{pr:h:2}, this also means that $-\psi$ is convex,
and thus $\psi$ is concave.
\qed
Before proceeding, we introduce a useful operator.
As noted earlier, adding $-\infty$ and $+\infty$ is problematic
and generally undefined.
To handle this difficulty, we define a form of addition
which specifies explicitly which summand should be
treated as dominant.
This is meant only to resolve the sum of $-\infty$ and $+\infty$;
addition is defined in the usual commutative way in all other
situations.
Formally, for $\bar{x},\,\bar{y}\in\overline{\R}$, we write this non-commutative sum
as $\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y}$.
This operation, called \emph{leftward addition}, is the same as
ordinary addition except that,
when adding $-\infty$ and $+\infty$, the argument on the left is dominant.
Thus,
\begin{eqnarray*}
(+\infty) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (-\infty) &=& +\infty \\
(-\infty) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (+\infty) &=& -\infty \\
\bar{x} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y} &=& \bar{x} + \bar{y} \;\;\mbox{in all other cases.}
\end{eqnarray*}
Equivalently,
\begin{equation} \label{eq:h:5}
\bar{x} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y} = \left\{ \begin{array}{cl}
\bar{x} & \mbox{if $\bar{x}\in\{-\infty,+\infty\}$}\\
\bar{x} + \bar{y} & \mbox{otherwise.}
\end{array}\right.
\end{equation}
Leftward addition is assigned the same precedence as ordinary
addition.
It is associative and
distributive,
and, when one of the arguments is in ${\mathbb{R}}$,
it is also commutative (and equivalent to ordinary addition).
These and other basic properties are summarized in the next
proposition, whose proof is straightforward.
\begin{proposition} \label{pr:i:5}
For all $\bar{x},\bar{y},\bar{z},\bar{x}',\bar{y}'\in\overline{\R}$,
the following hold:
\begin{enumerate}
\item \label{pr:i:5a}
$(\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{z}=\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\bar{y}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{z})$.
\item \label{pr:i:5b}
$\lambda(\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y})=\lambda\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda\bar{y}$,
for $\lambda\in{\mathbb{R}}$.
\item \label{pr:i:5bb}
$\alpha\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \beta\bar{x}=\alpha\bar{x}+\beta\bar{x}
=(\alpha+\beta)\bar{x}$, for $\alpha,\,\beta\in{\mathbb{R}}$
with $\alpha\beta\geq 0$.
\item \label{pr:i:5c}
$\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax y = \bar{x} + y = y + \bar{x} = y\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{x}$, for $y\in{\mathbb{R}}$.
\item \label{pr:i:5ord}
If $\bar{x}\leq \bar{x}'$ and $\bar{y}\leq \bar{y}'$ then $\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y} \leq \bar{x}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \bar{y}'$.
\end{enumerate}
\end{proposition}
For $x\in{\mathbb{R}}$, we also define the notation
\[
\omsf{x} =
\left\{ \begin{array}{cl}
+\infty & \mbox{if $x>0$}\\
0 & \mbox{if $x=0$}\\
-\infty & \mbox{if $x<0$.}
\end{array}\right.
\]
This notation is meant to be suggestive of multiplication by
a scalar $\omega$ that is equal to $+\infty$.
Later, we will extend this notation in detail along the lines
suggested in Section~\ref{sec:astral-space-intro}.
We turn next to a central theorem showing that any function
$\psi:{\R^n}\rightarrow\overline{\R}$ that is convex, concave
and that vanishes at ${\bf 0}$ must have a particular structural form.
By Theorem~\ref{thm:h:3}, this will be the case for every function in
$\overline{\phimg}$, and so also every function in $\Phi$ by Proposition~\ref{pr:h:7}.
Furthermore, as detailed below, this will be the key in showing
that all astral points must have a correspondingly specific form, as
outlined in Section~\ref{sec:astral-space-intro}.
Indeed, we will see below that if $\overline{\xx}\in\extspac{n}$ corresponds to
$\psi$ (so that $\psi=\ph{\overline{\xx}}$ and therefore
$\psi({\bf u})=\overline{\xx}\cdot{\bf u}$ for all ${\bf u}\in{\R^n}$), then
Theorem~\ref{thm:h:4} shows that $\overline{\xx}$ has exactly the form given in
\eqref{eqn:pf:2}, a form that has merely been re-stated using the
notation above (and which we will later simplify further).
\begin{theorem} \label{thm:h:4}
Let $\psi:{\R^n}\rightarrow\overline{\R}$ be convex and
concave with $\psi({\bf 0})=0$.
Then there exists ${\bf q},{\bf v}_1\ldots,{\bf v}_k\in{\R^n}$, for some $k\geq 0$,
such that for all ${\bf u}\in{\R^n}$,
\begin{equation} \label{thm:h:4:eq1}
\psi({\bf u}) = \omsf{({\bf v}_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}_k\cdot{\bf u})} + {\bf q}\cdot{\bf u}.
\end{equation}
\end{theorem}
\proof
We prove the following by induction on $d=0,\ldots,n$:
For every linear subspace $L$ of ${\R^n}$ of dimension $d$,
there exists a function $\xi\in{\cal F}$ of the form
given on the right-hand side of \eqref{thm:h:4:eq1}
such that $\psi({\bf u})=\xi({\bf u})$ for all ${\bf u}\in L$.
The theorem will then follow by letting $L={\R^n}$, which implies $\psi=\xi$,
and shows that $\psi$ has the stated form.
In the base case that $d=0$, we must have $L=\{{\bf 0}\}$ and so
we can simply define
$\xi$ by choosing $k=0$ and ${\bf q}={\bf 0}$ so that $\xi({\bf 0})=0=\psi({\bf 0})$.
For the inductive step, let $d>0$, and assume the main claim holds for
all subspaces of dimension $d-1$.
Let $L\subseteq{\R^n}$ be any linear subspace of dimension $d$.
Let us define the sets
\begin{eqnarray*}
{E_{+}} &=& \{\rpair{{\bf u}}{z} \in L\times{\mathbb{R}} : z\geq \psi({\bf u})\}
\\
{E_{-}} &=& \{\rpair{{\bf u}}{z} \in L\times{\mathbb{R}} : z\leq \psi({\bf u})\}.
\end{eqnarray*}
The main idea of the proof is to separate these sets with a
hyperplane, which, together with our inductive hypothesis, will allow
us to derive a representation of $\psi$.
The sets ${E_{+}}$ and ${E_{-}}$ are nonempty (since they both contain
$\rpair{{\bf 0}}{0}$), and also are both convex subsets of ${\R^{n+1}}$.
Indeed,
the set ${E_{+}}$ is simply the epigraph of $\psi$,
intersected with $L\times{\mathbb{R}}$.
Therefore, ${E_{+}}$ is convex since $\psi$ is convex.
Likewise, ${E_{-}}$ can be obtained from the epigraph of $-\psi$ by
linearly mapping each point $\rpair{{\bf u}}{z}$ to $\rpair{{\bf u}}{-z}$, and then
intersecting with $L\times{\mathbb{R}}$.
So ${E_{-}}$ is also convex since $\psi$ is concave.
(Alternatively, the convexity of ${E_{+}}$ and ${E_{-}}$ can be proved
directly using Proposition~\ref{pr:rock:4.2}.)
We next claim that $(\relint{{E_{+}}})\cap(\relint{{E_{-}}})=\emptyset$.
For if $\rpair{{\bf u}}{z}\in\relint{{E_{+}}}\subseteq {E_{+}}$, then $\rpair{{\bf u}}{z+1}$ must also
be in ${E_{+}}$, implying, since $\rpair{{\bf u}}{z}\in\relint{{E_{+}}}$, that some point
$\rpair{{\bf u}}{z-\delta}$ must be in ${E_{+}}$ as well, for some $\delta>0$
\cite[Theorem~6.4]{ROC},
meaning $z-\delta\geq \psi({\bf u})$.
In other words, if $\rpair{{\bf u}}{z}\in\relint{{E_{+}}}$ then $z>\psi({\bf u})$.
Similarly, if $\rpair{{\bf u}}{z}\in\relint{{E_{-}}}$ then $z<\psi({\bf u})$.
Thus, no point can be both in $\relint{{E_{+}}}$ and in $\relint{{E_{-}}}$.
So ${E_{+}}$ and ${E_{-}}$ are nonempty convex subsets of
${\R^{n+1}}$ with disjoint relative interiors.
Therefore, there exists a hyperplane that properly separates them
\cite[Theorem~11.2]{ROC},
meaning that ${E_{+}}$ is included in the closed halfspace ${H_{+}}$ on one side of
the hyperplane, ${E_{-}}$ is included in the opposite closed halfspace ${H_{-}}$, and
furthermore, they are not both entirely included in the separating
hyperplane itself.
Thus, there exists ${\bf v}\in{\R^n}$, $b,c\in{\mathbb{R}}$ such that
\begin{eqnarray*}
{E_{+}}\! &\subseteq& \!{H_{+}} = \{ \rpair{{\bf u}}{z}\in {\R^{n+1}} : c \geq {\bf v}\cdot{\bf u} + bz \} \\
{E_{-}}\! &\subseteq& \!{H_{-}} = \{ \rpair{{\bf u}}{z}\in {\R^{n+1}} : c \leq {\bf v}\cdot{\bf u} + bz \}.
\end{eqnarray*}
In fact, we must have $c=0$ since
$\rpair{{\bf 0}}{0}\in{E_{+}}\cap{E_{-}}\subseteq {H_{+}}\cap{H_{-}}$.
The remainder of the proof considers separately the cases that $b=0$
or $b\neq 0$.
\setcounter{claimp}{0}
\medskip
\begin{claimp}
Suppose $b\neq 0$, and let ${\bf q}=-{\bf v}/b$.
Then
$\psi({\bf u})={\bf q}\cdot{\bf u}$ for all ${\bf u}\in L$
(implying the main claim is proved by defining
$\xi({\bf u})={\bf q}\cdot{\bf u}$ for ${\bf u}\in{\R^n}$).
\end{claimp}
\proof
Let ${\bf u}\in L$.
We claim first that $\psi({\bf u})<+\infty$.
Otherwise, if $\psi({\bf u})= +\infty$, then for all $z\in{\mathbb{R}}$,
$\rpair{{\bf u}}{z}\in{E_{-}}\subseteq {H_{-}}$.
That is, $0\leq {\bf v}\cdot{\bf u}+bz$ must hold for all $z\in{\mathbb{R}}$, which
is impossible since $b\neq 0$.
Similarly, $\psi({\bf u})>-\infty$.
Thus, $\psi({\bf u})\in{\mathbb{R}}$, so
\[\rpair{{\bf u}}{\psi({\bf u})}\in{E_{+}}\cap{E_{-}}\subseteq {H_{+}}\cap{H_{-}},\]
implying $0={\bf v}\cdot{\bf u}+b \psi({\bf u})$.
That is, $\psi({\bf u})=-{\bf v}\cdot{\bf u}/b = {\bf q}\cdot{\bf u}$, proving
the claim.
\qed
It remains to prove the main claim when $b=0$, which we assume
to be the case henceforth.
\medskip
\begin{claimp} \label{thm:h:4:c3}
Let ${\bf u}\in L$.
If ${\bf v}\cdot{\bf u}>0$, then $\psi({\bf u})=+\infty$.
And if ${\bf v}\cdot{\bf u}<0$, then $\psi({\bf u})=-\infty$.
\end{claimp}
\proof
We prove the contrapositive of the first statement.
Suppose $\psi({\bf u})<+\infty$.
Then $\psi({\bf u})<z$ for some $z\in{\mathbb{R}}$,
so that $\rpair{{\bf u}}{z}\in{E_{+}}\subseteq{H_{+}}$.
That is, ${\bf v}\cdot{\bf u}\leq 0$.
Likewise, if $\psi({\bf u})>-\infty$ then ${\bf v}\cdot{\bf u}\geq 0$.
\qed
The separating hyperplane is
\[ {H_{+}}\cap{H_{-}}=\{\rpair{{\bf u}}{z}\in {\R^{n+1}} : {\bf v}\cdot{\bf u}=0\} = M\times{\mathbb{R}} \]
where
\[ M=\{{\bf u}\in {\R^n} : {\bf v}\cdot{\bf u}=0\}. \]
Because ${E_{+}}$ and ${E_{-}}$ are \emph{properly} separated, they are not
both entirely contained in this hyperplane;
that is,
\[ L\times{\mathbb{R}} = {E_{+}}\cup{E_{-}} \not\subseteq M\times{\mathbb{R}}, \]
so
$L\not\subseteq M$.
We next apply our inductive hypothesis to the linear subspace
$ L' = L\cap M $.
Since $L\not\subseteq M$,
this space has dimension $d-1$.
Thus, there exists a function $\xi'$ such that, for ${\bf u}\in{\R^n}$,
\[ \xi'({\bf u}) =
\omsf{({\bf v}_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}_k\cdot{\bf u})} + {\bf q}\cdot{\bf u}
\]
for some ${\bf q},{\bf v}_1\ldots,{\bf v}_k\in{\R^n}$ and some $k\geq 0$.
Furthermore $\psi({\bf u})=\xi'({\bf u})$ for ${\bf u}\in L'$.
Let us define $\xi:{\R^n}\rightarrow\overline{\R}$ by
\begin{eqnarray*}
\xi({\bf u}) &=& \omsf{({\bf v}\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \xi'({\bf u})
\\
&=& \omsf{({\bf v}\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}_k\cdot{\bf u})} + {\bf q}\cdot{\bf u}
\end{eqnarray*}
for ${\bf u}\in{\R^n}$.
Then for ${\bf u}\in L$,
\[
\xi({\bf u}) = \left\{ \begin{array}{cl}
+\infty & \mbox{if ${\bf v}\cdot{\bf u}>0$}\\
\xi'({\bf u}) & \mbox{if ${\bf v}\cdot{\bf u}=0$}\\
-\infty & \mbox{if ${\bf v}\cdot{\bf u}<0$.}
\end{array}
\right.
\]
Thus, if ${\bf v}\cdot{\bf u}\neq 0$, then $\xi({\bf u})=\psi({\bf u})$ by
Claim~\ref{thm:h:4:c3}.
And if ${\bf v}\cdot{\bf u}=0$, then ${\bf u}\in L'$ so
$\xi({\bf u})=\xi'({\bf u})=\psi({\bf u})$ by
inductive hypothesis.
\qed
Pulling these results together, we can now conclude that
the sets $\Phi$ and $\overline{\phimg}$ are actually identical, each
consisting exactly of all of the functions on ${\R^n}$ that are convex,
concave and that vanish at the origin, and furthermore, that these are
exactly the functions of the form given in Theorem~\ref{thm:h:4}.
\begin{theorem} \label{thm:h:5}
Let $\psi\in{\cal F}$, that is, $\psi:{\R^n}\rightarrow\overline{\R}$.
Then the following are equivalent:
\begin{enumerate}
\item \label{thm:h:5a0}
$\psi\in\Phi$.
(That is, for some $\overline{\xx}\in\extspac{n}$,
$\psi({\bf u})=\ph{\overline{\xx}}({\bf u})=\overline{\xx}\cdot{\bf u}$ for all ${\bf u}\in{\R^n}$.)
\item \label{thm:h:5a}
$\psi\in\overline{\phimg}$.
\item \label{thm:h:5b}
$\psi$ is convex, concave and $\psi({\bf 0})=0$.
\item \label{thm:h:5c}
There exists ${\bf q},{\bf v}_1\ldots,{\bf v}_k\in{\R^n}$, for some $k\geq 0$,
such that for all ${\bf u}\in{\R^n}$,
\[
\psi({\bf u}) = \omsf{({\bf v}_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}_k\cdot{\bf u})} + {\bf q}\cdot{\bf u}.
\]
\end{enumerate}
\end{theorem}
\proof
Proposition~\ref{pr:h:7} proves that (\ref{thm:h:5a0}) implies
(\ref{thm:h:5a}).
That (\ref{thm:h:5a}) implies (\ref{thm:h:5b}) follows immediately
from Proposition~\ref{pr:h:2} and Theorem~\ref{thm:h:3}.
Theorem~\ref{thm:h:4} proves that (\ref{thm:h:5b}) implies
(\ref{thm:h:5c}).
To show that (\ref{thm:h:5c}) implies (\ref{thm:h:5a0}), let
$\psi$ have the form given in part~(\ref{thm:h:5c}), and let
$\seq{{\bf x}_t}$ be the sequence given in \eqref{eq:h:6}.
Then for all ${\bf u}\in{\R^n}$,
${\bf x}_t\cdot{\bf u}\rightarrow \psi({\bf u})$, as
can be seen by checking the various cases in \eqref{eqn:pf:2}
(and as argued informally in deriving that equation).
Since ${\bf x}_t\cdot{\bf u}$ converges in $\overline{\R}$, for all ${\bf u}\in{\R^n}$,
by construction of astral space, there exists some
$\overline{\xx}\in\extspac{n}$ that includes the sequence $\seq{{\bf x}_t}$ in its
equivalence class, $\pi(\overline{\xx})$, so that
\[
\ph{\overline{\xx}}({\bf u}) = \overline{\xx}\cdot{\bf u} = \lim {\bf x}_t\cdot{\bf u} = \psi({\bf u}).
\]
That is, $\ph{\overline{\xx}}=\psi$, proving $\psi\in\Phi$.
\qed
\subsection{Astral space}
\label{sec:astral:space:summary}
As already discussed, we can translate these results back to
$\extspac{n}$ which is homeomorphic to $\Phi$ via ${\varphi}$.
We summarize these topological implications in the following theorem.
Most importantly, we have shown that $\extspac{n}$ is a compactification
of ${\R^n}$:
\begin{theorem} \label{thm:i:1}
~
\begin{enumerate}
\item \label{thm:i:1a}
$\extspac{n}$ is compact and Hausdorff.
\item \label{thm:i:1b}
${\R^n}$ is dense in $\extspac{n}$.
\item \label{thm:i:1c}
Let $\seq{\overline{\xx}_t}$ be a sequence in $\extspac{n}$ and let
$\overline{\xx}\in\extspac{n}$.
Then $\overline{\xx}_t\rightarrow\overline{\xx}$ if and only if
for all ${\bf u}\in{\R^n}$,
$\overline{\xx}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$.
\item \label{thm:i:1d}
Let $\overline{\xx}\in\extspac{n}$.
Then there exists a sequence $\seq{{\bf x}_t}$ in ${\R^n}$ with
${\bf x}_t\rightarrow\overline{\xx}$.
\item \label{thm:i:1e}
Let $\seq{\overline{\xx}_t}$ be a sequence in $\extspac{n}$ and
suppose that for all ${\bf u}\in{\R^n}$, the sequence $\overline{\xx}_t\cdot{\bf u}$
converges in $\overline{\R}$.
Then there exists a point $\overline{\xx}\in\extspac{n}$ with
$\overline{\xx}_t\rightarrow\overline{\xx}$.
\end{enumerate}
\end{theorem}
\proof
~
Part~(\ref{thm:i:1a}):
By Theorem~\ref{thm:h:5}, $\Phi$ equals $\overline{\phimg}$ and is therefore compact,
being a closed subset of the compact space ${\cal F}$.
Since $\overline{\R}$ is Hausdorff, the product space ${\cal F}$ and its
subspace $\Phi$ are also Hausdorff.
Therefore, $\extspac{n}$, which is homeomorphic with $\Phi$, is
compact and Hausdorff as well.
Part~(\ref{thm:i:1b}):
That $\Phi=\overline{\phimg}$ means exactly that $\phfcn(\Rn)$ is dense in
$\Phi={\varphi}(\extspac{n})$.
So, by our homeomorphism, ${\R^n}$ is dense in $\extspac{n}$.
Part~(\ref{thm:i:1c}):
By our homeomorphism,
${\overline{\xx}_t}\rightarrow{\overline{\xx}}$ if and only if
$\ph{\overline{\xx}_t}\rightarrow\ph{\overline{\xx}}$,
which,
because we are using product topology in ${\cal F}$,
in turn holds if and only if
$\overline{\xx}_t\cdot{\bf u}=\ph{\overline{\xx}_t}({\bf u})\rightarrow\ph{\overline{\xx}}({\bf u})=\overline{\xx}\cdot{\bf u}$
for all ${\bf u}\in{\R^n}$.
Part~(\ref{thm:i:1d}):
Let $\seq{{\bf x}_t}$ in ${\R^n}$ be any sequence in $\pi(\overline{\xx})$.
Then ${\bf x}_t\rightarrow\overline{\xx}$ by part~(\ref{thm:i:1c}).
Part~(\ref{thm:i:1e}):
Let $\psi:{\R^n}\rightarrow\overline{\R}$ be defined by
$\psi({\bf u})=\lim \overline{\xx}_t\cdot{\bf u}$
for ${\bf u}\in{\R^n}$.
Then $\ph{\overline{\xx}_t}({\bf u})\rightarrow\psi({\bf u})$ for all ${\bf u}\in{\R^n}$,
so $\ph{\overline{\xx}_t}\rightarrow\psi$.
Since $\ph{\overline{\xx}_t}\in\Phi$, for all $t$, and since $\Phi$ is closed,
$\psi$ must also be in $\Phi$.
Thus, $\psi=\ph{\overline{\xx}}$ for some $\overline{\xx}\in\extspac{n}$, so
$\ph{\overline{\xx}_t}\rightarrow\ph{\overline{\xx}}$, implying,
by our homeomorphism, that
${\overline{\xx}_t}\rightarrow{\overline{\xx}}$.
\qed
Just as addition of scalars can be extended to vectors, we can extend
leftward addition, an operation already defined for scalars in
$\overline{\R}$, to all points in $\extspac{n}$.
For $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$, we define $\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$ to be that
unique point in $\extspac{n}$ for which
\[ (\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}})\cdot{\bf u} = \overline{\xx}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}}\cdot{\bf u} \]
for all ${\bf u}\in{\R^n}$.
The next proposition shows that such a point must exist.
This operation will turn out to be very natural for describing and
working with points in $\extspac{n}$, as we will see shortly.
It is much like taking the vector sum of two points in ${\R^n}$, but
assigns a kind of dominance to the summand on the left.
\begin{proposition} \label{pr:i:6}
Let $\overline{\xx}$ and $\overline{{\bf y}}$ be in $\extspac{n}$.
Then there exists a unique point in $\extspac{n}$, henceforth denoted
$\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$, for which
\[ (\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}})\cdot{\bf u} = \overline{\xx}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}}\cdot{\bf u} \]
for all ${\bf u}\in{\R^n}$.
\end{proposition}
\proof
By Theorem~\ref{thm:h:5}, since $\ph{\overline{\xx}}$ and $\ph{\overline{{\bf y}}}$ are both
in $\Phi$,
there exists ${\bf q},{\bf v}_1\ldots,{\bf v}_k\in{\R^n}$, for some $k\geq 0$,
and ${\bf q}',{\bf v}'_1\ldots,{\bf v}'_{k'}\in{\R^n}$, for some $k'\geq 0$,
such that for all ${\bf u}\in{\R^n}$,
\begin{eqnarray}
\ph{\overline{\xx}}({\bf u}) &=& \omsf{({\bf v}_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}_k\cdot{\bf u})} + {\bf q}\cdot{\bf u}
\nonumber
\\
\ph{\overline{{\bf y}}}({\bf u}) &=& \omsf{({\bf v}'_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}'_{k'}{\bf u}\cdot)} + {\bf q}'\cdot{\bf u}.
\label{eq:i:3}
\end{eqnarray}
Let $\psi:{\R^n}\rightarrow\overline{\R}$ be defined by
\begin{eqnarray*}
\psi({\bf u})
&=&
\overline{\xx}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}}\cdot{\bf u}
\\
&=&
\ph{\overline{\xx}}({\bf u})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \ph{\overline{{\bf y}}}({\bf u})
\\
&=&
\omsf{({\bf v}_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \omsf{({\bf v}_k\cdot{\bf u})}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}'_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \omsf{({\bf v}'_{k'}{\bf u}\cdot)}
+ ({\bf q}+{\bf q}')\cdot{\bf u},
\end{eqnarray*}
for ${\bf u}\in{\R^n}$,
where the last equality uses \eqref{eq:i:3} and
Proposition~\ref{pr:i:5}.
Thus, because it has the form given in Theorem~\ref{thm:h:5},
$\psi$ is in $\Phi$ and therefore is
equal to $\ph{\overline{{\bf z}}}$ for some $\overline{{\bf z}}\in\extspac{n}$.
Setting $\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$ equal to $\overline{{\bf z}}$ proves the proposition.
(Uniqueness follows from Proposition~\ref{pr:i:4}.)
\qed
Just as was the case for scalars, leftward addition of astral
points is associative, distributive and partially commutative
(specifically, if either of the summands is in ${\R^n}$).
It is also the same as vector addition when both summands are in
${\R^n}$.
We summarize these in the next proposition.
\begin{proposition} \label{pr:i:7}
For all $\overline{\xx},\overline{{\bf y}},\overline{{\bf z}}\in\extspac{n}$,
and for all ${\bf x},{\bf y}\in{\R^n}$,
the following hold:
\begin{enumerate}
\item \label{pr:i:7a}
$(\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}=\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}})$.
\item \label{pr:i:7b}
$\lambda(\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}})=\lambda\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda\overline{{\bf y}}$,
for $\lambda\in{\mathbb{R}}$.
\item \label{pr:i:7c}
$\alpha\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \beta\overline{\xx}=(\alpha+\beta)\overline{\xx}$,
for $\alpha,\beta\in{\mathbb{R}}$ with $\alpha\beta\geq 0$.
\item \label{pr:i:7d}
$\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf y} = {\bf y}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}$.
In particular,
$\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf 0} = {\bf 0}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx} = \overline{\xx}$.
\item \label{pr:i:7e}
${\bf x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf y} = {\bf x} + {\bf y}$.
\end{enumerate}
\end{proposition}
\proof
Part~(\ref{pr:i:7a}):
By Propositions~\ref{pr:i:6}
and~\ref{pr:i:5}(\ref{pr:i:5a}),
for ${\bf u}\in{\R^n}$,
\begin{eqnarray*}
((\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}})\cdot{\bf u}
&=&
(\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}})\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot{\bf u}
\\
&=&
(\overline{\xx}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}}\cdot{\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot{\bf u}
\\
&=&
\overline{\xx}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\overline{{\bf y}}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot{\bf u})
\\
&=&
\overline{\xx}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}) \cdot{\bf u}
\\
&=&
(\overline{\xx} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}})) \cdot{\bf u}.
\end{eqnarray*}
The claim now follows by Proposition~\ref{pr:i:4}.
The proofs of the other parts are similar.
\qed
The leftward sum of two sets $X$ and $Y$ in
$\extspac{n}$ is defined in the same natural way as for ordinary addition:
\[
X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax Y = \{ \overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}} : \overline{\xx}\in X, \overline{{\bf y}}\in Y \}.
\]
Similarly, we define
$X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}} = X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \{\overline{{\bf y}}\}$
and
$\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax X = \{\overline{{\bf y}}\} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax X$
for $X\subseteq\extspac{n}$ and $\overline{{\bf y}}\in\extspac{n}$.
Analogously, if $\Lambda\subseteq{\mathbb{R}}$ and $X\subseteq\extspac{n}$,
then we define the product
\[
\Lambda X = \{ \lambda \overline{\xx} : \lambda\in \Lambda, \overline{\xx}\in X \},
\]
with $\lambda X = \{\lambda\} X$ and
$\Lambda \overline{\xx} = \Lambda \{\overline{\xx}\}$
for $\lambda\in{\mathbb{R}}$ and $\overline{\xx}\in\extspac{n}$.
In Section~\ref{subsec:astral-intro-motiv}, we encountered
an important example of an astral point not in ${\R^n}$:
For ${\bf v}\in{\R^n}$, the point $\limray{{\bf v}}$,
called an \emph{astron}, is that point in $\extspac{n}$
that is the limit of the sequence $\seq{t{\bf v}}$, that is,
by following a ray to infinity (unless ${\bf v}={\bf 0}$),
beginning at the origin, and heading in the direction of ${\bf v}$.
As before, the notation $\limray{{\bf v}}$ is meant to be suggestive of
multiplying ${\bf v}$ by an infinitely large scalar, $\omega$.
In a moment,
we will see that astrons are the fundamental building blocks out of
which every point in $\extspac{n}$ can be constructed.
The next simple proposition gives their general form:
\begin{proposition} \label{pr:astrons-exist}
Let ${\bf v}\in{\R^n}$.
Then the sequence $\seq{t{\bf v}}$ has a limit in $\extspac{n}$,
henceforth denoted $\limray{{\bf v}}$.
Furthermore, for all ${\bf u}\in{\R^n}$,
\[
(\limray{{\bf v}})\cdot{\bf u}
= \omsf{({\bf v}\cdot{\bf u})}
= \left\{ \begin{array}{cl}
+\infty & \mbox{if ${\bf v}\cdot{\bf u}>0$} \\
0 & \mbox{if ${\bf v}\cdot{\bf u}=0$} \\
-\infty & \mbox{if ${\bf v}\cdot{\bf u}<0$.}
\end{array} \right.
\]
\end{proposition}
\proof
For ${\bf u}\in{\R^n}$, $(t{\bf v})\cdot{\bf u}=t({\bf v}\cdot{\bf u})\rightarrow\omsf{({\bf v}\cdot{\bf u})}$.
Since this limit exists for each ${\bf u}\in{\R^n}$, the sequence
$\seq{t{\bf v}}$ must have a limit, $\limray{{\bf v}}$, in $\extspac{n}$ by
Theorem~\ref{thm:i:1}(\ref{thm:i:1e}).
This also proves $(\limray{{\bf v}})\cdot{\bf u}$ has the form given
in the proposition.
\qed
Using astrons and leftward addition,
we can now succinctly state and prove the form of
every point in $\extspac{n}$ using Theorem~\ref{thm:h:5}, which is the
same form discussed in Section~\ref{subsec:astral-pt-form}.
\begin{corollary} \label{cor:h:1}
Astral space $\extspac{n}$ consists exactly of all points
of the form
\[ \limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}_k} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q} \]
for ${\bf q},{\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$, $k\geq 0$.
\end{corollary}
In other words, every point in $\extspac{n}$ has this form, and every
point that has this form is in $\extspac{n}$.
\proof
Every expression of the form given in the corollary is in $\extspac{n}$
since astral space includes all astrons and
since the space is closed under leftward addition
(by Propositions~\ref{pr:astrons-exist} and~\ref{pr:i:6}).
To show every point has this form,
suppose $\overline{\xx}\in\extspac{n}$.
Then $\ph{\overline{\xx}}\in\Phi$, so by Theorem~\ref{thm:h:5},
for some ${\bf q},{\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$,
\begin{eqnarray*}
\overline{\xx}\cdot{\bf u}
&=&
\ph{\overline{\xx}}({\bf u})
\\
&=&
\omsf{({\bf v}_1\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\omsf{({\bf v}_k\cdot{\bf u})} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}\cdot{\bf u}
\\
&=&
(\limray{{\bf v}_1})\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
(\limray{{\bf v}_k})\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}\cdot{\bf u}
\\
&=&
\paren{\limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}_k} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}}\cdot{\bf u}
\end{eqnarray*}
for ${\bf u}\in{\R^n}$.
This means, by Proposition~\ref{pr:i:4}, that
\[
\overline{\xx} = {\limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}_k} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}}.
\]
\qed
Theorem~\ref{thm:h:5} and Corollary~\ref{cor:h:1}
tell us much about the nature of
elements of astral space, which can be viewed
from two main perspectives, as we have now seen in detail.
First, as discussed in Section~\ref{sec:astral-space-intro},
we can think of astral points $\overline{\xx}\in\extspac{n}$
as the possible destinations of trajectories through ${\R^n}$.
Such a trajectory may have a finite destination ${\bf q}$ in ${\R^n}$.
But more interestingly, Corollary~\ref{cor:h:1} expresses exactly
all of the ways in which a convergent trajectory can ``go to
infinity'':
Every such trajectory has a primary dominant direction
${\bf v}_1$, and may also have a secondary direction ${\bf v}_2$, a tertiary
direction ${\bf v}_3$, etc.
Also, at convergence, every such trajectory has a finite part
${\bf q}\in{\R^n}$ as well.
This is exactly the form discussed in
Section~\ref{subsec:astral-pt-form};
the sequence constructed in \eqref{eq:h:6} is an example of a
trajectory with just these properties.
Alternatively, as we saw in Section~\ref{sec:astral-as-fcns},
astral points have a kind of dual form as functions mapping ${\R^n}$ to
$\overline{\R}$, and indeed their topology and other properties are largely
derived from this correspondence.
Astral space, we proved, is a homeomorphic copy of the function space
$\Phi=\overline{\phimg}$, a space that we also showed consists exactly of those
functions on ${\R^n}$ that are convex, concave and that vanish at the
origin.
We also proved that these functions have a very particular functional
form.
In the foregoing,
we followed an approach in which astral space was constructed
based on the first perspective.
But it would have been possible to instead construct this same space
based on the second perspective.
Such an approach would begin
with $\phfcn(\Rn)$, which is exactly the space of all
linear functions, and which is
equivalent topologically and as a vector space to ${\R^n}$.
The next step would be to form the closure of this space, $\overline{\phimg}$, which
is compact as a closed subset of the compact space ${\cal F}$.
The actual astral space $\extspac{n}$ could then be constructed as a
homeomorphic copy of $\overline{\phimg}$.
This is very much like the approach used, for instance,
in constructing the Stone-\v{C}ech compactification~\citep[Section 38]{munkres}, and
shows that astral space is a special case of the
${\cal Q}$-compactifications studied by~\citet{q_compactification}, in which we set
${\cal Q}$ to be the space of all linear functions.
\subsection{Closed sets and closures}
We saw in Proposition~\ref{pr:open-sets-equiv} that a set in ${\R^n}$ is
open in the topology we have defined for $\extspac{n}$ if and only if it
is open in the standard Euclidean topology.
For this reason, we can simply refer to a set in ${\R^n}$ as being open
or not, without specifying a topology.
The same is \emph{not} true for closed sets:
A set $U\subseteq{\R^n}$ can be closed with respect to
the standard Euclidean topology, but not closed in
the topology we defined on $\extspac{n}$.
We use the phrase \emph{closed in ${\R^n}$} or
\emph{closed in $\extspac{n}$} to make clear which topology is being
referred to, although
usually the topology will be clear from context, especially
since this ambiguity only arises when considering sets that are
necessarily in ${\R^n}$.
For instance, if ${\bf v}\neq{\bf 0}$, then the ray
$\{ \lambda {\bf v} : \lambda\geq 0 \}$
is closed in ${\R^n}$, but it is not closed in $\extspac{n}$ since it does
not include $\limray{{\bf v}}$, the limit of the sequence $\seq{t{\bf v}}$.
Indeed, ${\R^n}$ itself is closed in ${\R^n}$ but not in $\extspac{n}$.
In the same way, the closure of a set in ${\R^n}$ can in general be
different if the closure is with respect to the topology on ${\R^n}$ or
$\extspac{n}$.
We use the notation $\clgen{{\mathrm r}}$ or $\clgen{{\mathrm a}}$ for closure in ${\R^n}$ or closure
in $\extspac{n}$, respectively.
Furthermore, we usually use the alternative notation $\overline{U}$ for $\clgen{{\mathrm a}} U$, and the simplified notation $\cl U$ for $\clgen{{\mathrm r}} U$.
Here are some useful facts about closures and closed sets, all
standard results specialized to our current setting:
\begin{proposition} \label{pr:closed-set-facts}
\mbox{}
\begin{enumerate}
\item \label{pr:closed-set-facts:a}
Let $S \subseteq {\R^n}$.
Then $\cl S = \overline{S} \cap {\R^n}$.
\item \label{pr:closed-set-facts:aa}
Let $S \subseteq {\R^n}$.
Then $\overline{\cl S} = \overline{S}$.
\item \label{pr:closed-set-facts:b}
Let $U \subseteq \extspac{n}$ be open.
Then $\overline{(U \cap {\R^n})} = \overline{U}$.
\item \label{pr:closed-set-facts:c}
Let $\overline{\xx}\in\extspac{n}$ and $S\subseteq\extspac{n}$ where
$S$ is closed and $\overline{\xx}\not\in S$.
Then there exist disjoint open sets $U$ and $V$ in $\extspac{n}$
such that $\overline{\xx}\in U$ and $S \subseteq V$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:closed-set-facts:a}):
This is a special case of
\citet[Theorem 17.4]{munkres}.
Part~(\ref{pr:closed-set-facts:aa}):
On the one hand, $\overline{S}\subseteq \overline{\cl S}$
since $S\subseteq \cl S$.
On the other hand,
$\cl S = \overline{S}\cap{\R^n}\subseteq \overline{S}$
by part~(\ref{pr:closed-set-facts:a}).
Since $\overline{S}$ is closed, this implies
$\overline{\cl S} \subseteq \overline{S}$.
Part~(\ref{pr:closed-set-facts:b}):
First, $\overline{(U \cap {\R^n})} \subseteq \overline{U}$
since $U \cap {\R^n} \subseteq U$.
To prove the reverse inclusion, suppose $\overline{\xx}\in\overline{U}$.
Let $V$ be any neighborhood of $\overline{\xx}$.
Then there exists a point $\overline{{\bf y}}$ in the open set $U\cap V$,
which in turn, since
${\R^n}$ is dense in $\extspac{n}$, implies that there is a point
${\bf z}\in U\cap V\cap {\R^n}$.
Therefore, $\overline{\xx}\in \overline{(U \cap {\R^n})}$
since $U\cap{\R^n}$ intersects every neighborhood $V$ of $\overline{\xx}$.
Part~(\ref{pr:closed-set-facts:c}):
A topological space with exactly this property is said to be
\emph{regular}.
Since $\extspac{n}$ is compact and Haussdorf, it is also regular
\citep[Theorem 32.3]{munkres}.
\qed
\section{Representation and structure}
\label{sec:representing}
\subsection{Linear and affine maps}
\label{sec:linear-maps}
It will be very useful to extend linear maps to astral space.
The next proposition shows how to do this, effectively defining
the natural, continuous extension of
ordinary matrix multiplication on points in ${\R^n}$ to astral space,
$\extspac{n}$.
\begin{proposition} \label{pr:mat-mult-def}
Let ${\bf A}\in{\R^{m\times n}}$, and let $\overline{\xx}\in\extspac{n}$.
Then there exists a unique point in $\extspac{m}$,
henceforth denoted ${\bf A}\overline{\xx}$, with
the property that
if $\seq{{\bf x}_t}$ is any sequence in ${\R^n}$ with
${\bf x}_t\rightarrow\overline{\xx}$, then ${\bf A}{\bf x}_t\rightarrow {\bf A}\overline{\xx}$.
Furthermore,
\begin{equation} \label{eq:i:4}
({\bf A}\overline{\xx})\cdot{\bf u} = \overline{\xx}\cdot(\trans{{\bf A}} {\bf u})
\end{equation}
for all ${\bf u}\in{\R^m}$.
\end{proposition}
\proof
Let $\seq{{\bf x}_t}$ be any sequence in ${\R^n}$ that converges to $\overline{\xx}$
(which must exist by Theorem~\ref{thm:i:1}(\ref{thm:i:1d})).
Then for all ${\bf u}\in{\R^n}$,
\begin{equation} \label{eq:i:5}
({\bf A}{\bf x}_t)\cdot{\bf u}
= \trans{({\bf A}{\bf x}_t)}{\bf u}
= \trans{{\bf x}}_t \trans{{\bf A}} {\bf u}
= {\bf x}_t\cdot(\trans{{\bf A}} {\bf u})
\rightarrow \overline{\xx}\cdot (\trans{{\bf A}} {\bf u}),
\end{equation}
with the convergence following from
Theorem~\ref{thm:i:1}(\ref{thm:i:1c}).
Thus, this limit exists for all ${\bf u}\in{\R^n}$, and therefore,
by Theorem~\ref{thm:i:1}(\ref{thm:i:1e}), the sequence
$\seq{{\bf A}{\bf x}_t}$ has a limit in $\extspac{m}$, which we denote
${\bf A}\overline{\xx}$.
Applying
Theorem~\ref{thm:i:1}(\ref{thm:i:1c})
again yields
$({\bf A}{\bf x}_t)\cdot{\bf u}\rightarrow({\bf A}\overline{\xx})\cdot{\bf u}$
so that
\eqref{eq:i:4} now follows from \eqref{eq:i:5}.
This same argument also shows that this is true for every sequence
$\seq{{\bf x}_t}$ in ${\R^n}$ that converges to $\overline{\xx}$.
(Uniqueness follows from Proposition~\ref{pr:i:4}.)
\qed
Note that if $\overline{\xx}={\bf x}$, for some ${\bf x}\in{\R^n}$,
then ${\bf A}\overline{\xx}$, as here defined, is the same as the standard matrix
product ${\bf A}{\bf x}$ (as can be seen, for instance, by considering the
sequence ${\bf x}_t={\bf x}$, for all $t$, which trivially converges to
$\overline{\xx}$).
Also, if ${\bf u}\in{\R^n}$ and $\overline{\xx}\in\extspac{n}$,
then applying Proposition~\ref{pr:mat-mult-def}
to the matrix ${\bf A}=\trans{{\bf u}}$ shows that
$\trans{{\bf u}}\overline{\xx}=\overline{\xx}\cdot{\bf u}$, since if $\seq{{\bf x}_t}$ converges to
$\overline{\xx}$ then
\begin{equation} \label{eq:f:7}
\trans{{\bf u}}\overline{\xx}
= \lim \trans{{\bf u}} {\bf x}_t
= \lim {\bf x}_t\cdot{\bf u}
= \overline{\xx}\cdot{\bf u}
\end{equation}
(using $\extspac{1}=\overline{\R}$).
Here are some other basic properties:
\begin{proposition} \label{pr:h:4}
Let ${\bf A}\in{\R^{m\times n}}$, and let $\overline{\xx}\in\extspac{n}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:h:4c}
${\bf A}(\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}) = {\bf A}\overline{\xx} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf A}\overline{{\bf y}}$ for $\overline{{\bf y}}\in\extspac{n}$.
\item \label{pr:h:4d}
${\bf B}({\bf A}\overline{\xx}) = ({\bf B}{\bf A})\overline{\xx}$ for ${\bf B}\in{\mathbb{R}}^{\ell\times m}$.
\item \label{pr:h:4g}
$\lambda({\bf A}\overline{\xx}) = (\lambda{\bf A})\overline{\xx} = {\bf A} (\lambda\overline{\xx})$ for $\lambda\in{\mathbb{R}}$.
\item \label{pr:h:4e}
${\bf A} (\limray{{\bf v}}) = \limray{({\bf A}{\bf v})}$ for ${\bf v}\in{\R^n}$.
\end{enumerate}
\end{proposition}
\proof
Part~(\ref{pr:h:4c}):
By Propositions~\ref{pr:mat-mult-def} and~\ref{pr:i:6},
for all ${\bf u}\in{\R^n}$,
\begin{eqnarray*}
({\bf A}(\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}))\cdot{\bf u}
&=& (\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}})\cdot(\trans{{\bf A}}{\bf u})
\\
&=& \overline{\xx}\cdot(\trans{{\bf A}}{\bf u})
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\overline{{\bf y}}\cdot(\trans{{\bf A}}{\bf u})
\\
&=& ({\bf A}\overline{\xx})\cdot{\bf u}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
({\bf A}\overline{{\bf y}})\cdot{\bf u}
\\
&=& ({\bf A}\overline{\xx} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf A}\overline{{\bf y}})\cdot{\bf u}.
\end{eqnarray*}
The claim now follows by Proposition~\ref{pr:i:4}.
The proofs of parts~(\ref{pr:h:4d}),~(\ref{pr:h:4g})
and~(\ref{pr:h:4e}) are similar.
\qed
When multiplying a vector ${\bf x}\in{\R^n}$ by a matrix ${\bf A}\in{\R^{m\times n}}$,
the result ${\bf A}{\bf x}$ is a vector in ${\R^m}$ whose $j$-th component is
computed by multiplying ${\bf x}$ by the $j$-th row of ${\bf A}$.
In the same way, for an astral point $\overline{\xx}\in\extspac{n}$,
we might naively suppose that ${\bf A}\overline{\xx}$ in $\extspac{m}$
can be similarly viewed as the result of multiplying $\overline{\xx}$ by the
individual rows of ${\bf A}$.
However, this is not the case.
Indeed, matrix multiplication of astral points is far more holistic,
retaining and conveying much more information in effecting the linear
map $\overline{\xx}\mapsto{\bf A}\overline{\xx}$.
For instance, with $m=n=2$, suppose
$\overline{{\bf z}}={\bf A}\overline{\xx}$ where
\[
{\bf A} = \left[ \begin{array}{rr}
1 & 0 \\
-1 & 1 \\
\end{array}
\right],
\]
and $\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \beta {\bf e}_2$,
for some $\beta\in{\mathbb{R}}$
(with ${\bf e}_1,{\bf e}_2$ the standard basis vectors).
Let $\trans{{\bf a}}_1=[1, 0]$ and $\trans{{\bf a}}_2=[-1,1]$
denote the rows of ${\bf A}$.
Multiplying these rows separately by $\overline{\xx}$,
it can be checked that
$\trans{{\bf a}}_1\overline{\xx}=\overline{\xx}\cdot{\bf a}_1=+\infty$
and
$\trans{{\bf a}}_2\overline{\xx}=\overline{\xx}\cdot{\bf a}_2=-\infty$
(by \eqref{eq:f:7} and using
Propositions~\ref{pr:i:6}
and~\ref{pr:astrons-exist}).
From this,
it might appear that information about $\beta$ has been
entirely erased by the process of multiplying by ${\bf A}$.
But this is incorrect.
Indeed, the value of $\beta$ can be readily extracted from $\overline{{\bf z}}$
since
\[
\overline{{\bf z}}\cdot({\bf e}_1+{\bf e}_2)
=
({\bf A}\overline{\xx})\cdot({\bf e}_1+{\bf e}_2)
=
\overline{\xx}\cdot(\trans{{\bf A}}({\bf e}_1+{\bf e}_2))
=
\overline{\xx}\cdot{\bf e}_2
=
\beta,
\]
with the second equality following from
Proposition~\ref{pr:mat-mult-def}, and the third from a simple matrix
calculation.
Alternatively, since this matrix ${\bf A}$ is invertable,
$\overline{\xx}={\bf A}^{-1} \overline{{\bf z}}$, by
Proposition~\ref{pr:h:4}(\ref{pr:h:4d}),
which means $\overline{\xx}$ (including the value of $\beta$)
can be entirely recovered from $\overline{{\bf z}}$, though clearly not from
the row products,
$\trans{{\bf a}}_1\overline{\xx}$
and
$\trans{{\bf a}}_2\overline{\xx}$.
We next prove the continuity of linear maps, and more generally,
of \emph{affine maps}, functions that map each
point $\overline{\xx}$ in $\extspac{n}$ to $\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf A}\overline{\xx}$ in
$\extspac{m}$, for some matrix ${\bf A}\in{\mathbb{R}}^{m\times n}$ and point
$\overline{\bb}\in\extspac{m}$.
\begin{proposition} \label{pr:aff-cont}
Let ${\bf A}\in{\R^{m\times n}}$ and $\overline{\bb}\in\extspac{m}$,
and let $F:\extspac{n}\rightarrow\extspac{m}$
be defined by
$F(\overline{\xx})=\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{\xx}$.
Then $F$ is continuous.
Consequently,
if $\seq{\overline{\xx}_t}$ is a sequence in $\extspac{n}$ that converges to
$\overline{\xx}\in\extspac{n}$,
then $\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{\xx}_t\rightarrow\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{\xx}$.
\end{proposition}
\proof
To prove continuity, we show that the pre-image under $F$ of every
open set is also open;
indeed, it suffices to do so for every standard subbasis element
\citep[Section~18]{munkres}.
As in \eqref{eq:h:3a:subbasis},
let $V\subseteq\extspac{m}$ be the standard subbasis element
\[
V = \braces{\overline{{\bf y}}\in\extspac{m} : \overline{{\bf y}}\cdot{\bf u} < \beta}
\]
for some ${\bf u}\in{\R^m}$ and $\beta\in{\mathbb{R}}$.
We aim to show that $F^{-1}(V)$ is open.
We have
\begin{eqnarray}
F^{-1}(V)
&=&
\braces{\overline{\xx}\in\extspac{n} : F(\overline{\xx})\in V}
\nonumber
\\
&=&
\braces{\overline{\xx}\in\extspac{n} : (\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{\xx})\cdot{\bf u} < \beta}
\nonumber
\\
&=&
\braces{\overline{\xx}\in\extspac{n} :
\overline{\bb}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax({\bf A}\overline{\xx})\cdot{\bf u} < \beta}.
\label{eqn:pr:aff-cont:1}
\end{eqnarray}
Thus, if $\overline{\bb}\cdot{\bf u}=-\infty$, then the inequality in
\eqref{eqn:pr:aff-cont:1} always holds, so
$F^{-1}(V)=\extspac{n}$, which is open.
Likewise, if $\overline{\bb}\cdot{\bf u}=+\infty$, then that inequality never
holds, so $F^{-1}(V)=\emptyset$, which is also open.
And if $\overline{\bb}\cdot{\bf u}\in{\mathbb{R}}$, then that inequality holds if and only
if
$({\bf A}\overline{\xx})\cdot{\bf u} < \beta - \overline{\bb}\cdot{\bf u}$, so that, in this
case,
\[
F^{-1}(V)
=
\braces{\overline{\xx}\in\extspac{n} :
\overline{\xx}\cdot(\trans{{\bf A}}{\bf u})
< \beta - \overline{\bb}\cdot{\bf u}}
\]
by Proposition~\ref{pr:mat-mult-def}.
This set is open since it is a standard (sub)basis element.
Thus, in all cases,
$F^{-1}(V)$ is open,
proving that $F$ is continuous.
Since $F$ is continuous,
if $\seq{\overline{\xx}_t}$ is a sequence in $\extspac{n}$ that converges to
$\overline{\xx}\in\extspac{n}$,
then $F(\overline{\xx}_t)\rightarrow F(\overline{\xx})$.
\qed
Note that an affine map formed by adding a point $\overline{\bb}$ on the \emph{right},
that is, a map of the form $\overline{\xx}\mapsto{\bf A}\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\bb}$, need not
be continuous.
For example, with $n=m=1$, the function
$F(\bar{x})=\bar{x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(-\infty)$, where $\bar{x}\in\overline{\R}$, has this form.
On the sequence $x_t=t$, whose limit is $+\infty$, we see that
$F(x_t)=-\infty$ for all $t$, but $F(+\infty)=+\infty$, demonstrating
that this function is not continuous.
\subsection{Astral points in matrix form}
\label{sec:matrix:form}
We have seen already, for instance in Corollary~\ref{cor:h:1}, that
leftward sums of the form
\begin{equation} \label{eqn:pf:4}
\limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}_k}
\end{equation}
emerge naturally in studying astral space.
Using the matrix-product definition in Proposition~\ref{pr:mat-mult-def},
we will see next how such sums can be written more simply using
matrices, which will make for expressions that are more
compact, and easier to analyze and manipulate using tools from linear
algebra.
Specifically, we will see how to rewrite \eqref{eqn:pf:4} in the
compact form ${\bf V}{\bar{\bm{\omega}}}$ where
${\bf V}=[{\bf v}_1,\ldots,{\bf v}_k]$ denotes the matrix in ${\mathbb{R}}^{n\times k}$
whose $i$-th column is ${\bf v}_i$.
To see how to do this,
note first that
the astral point appearing in \eqref{eqn:pf:4} is the limit of any
sequence whose elements have the form
\[
{\bf V}{\bf b}_t = b_{t,1} {\bf v}_1 + \cdots + b_{t,k} {\bf v}_k,
\]
where
${\bf b}_t = \trans{[b_{t,1},\ldots,b_{t,k}]}$,
and, for $i=1,\ldots,k$, the sequence $\seq{b_{t,i}}$ grows
unboundedly to $+\infty$ with each sequence $\seq{b_{t,i}}$
growing faster than $\seq{b_{t,i+1}}$, so that
$b_{t,i}/b_{t,i+1}\rightarrow+\infty$
(see Theorem~\ref{thm:seq-rep}).
Thus, to express \eqref{eqn:pf:4} as ${\bf V}{\bar{\bm{\omega}}}$, the limit of a
sequence $\seq{{\bf V}{\bf b}_t}$ as above,
we informally want to think of ${\bar{\bm{\omega}}}$ as being like a column vector,
all of whose elements are infinite, but in which each element is
``more infinite'' than the next.
To do this formally, we define $\ommsub{k}$ to be that point in $\extspac{k}$
defined by
\begin{equation} \label{eq:i:7}
\ommsub{k} = \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf e}_k},
\end{equation}
where ${\bf e}_i\in{\mathbb{R}}^k$ is the $i$-th standard basis vector.
Then by Proposition~\ref{pr:h:4},
\begin{eqnarray}
{\bf V} \ommsub{k}
&=&
{\bf V} \paren{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf e}_k} }
\nonumber
\\
&=&
\limray{({\bf V}{\bf e}_1)} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{({\bf V}{\bf e}_k)}
\nonumber
\\
&=&
\limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}_k}.
\label{eqn:mat-omm-defn}
\end{eqnarray}
In other words, ${\bf V} \ommsub{k}$, with $\ommsub{k}$ defined as in
\eqref{eq:i:7}, is exactly equal to \eqref{eqn:pf:4}, now stated in
matrix form.
When clear from context,
we usually omit $\ommsub{k}$'s subscript and write simply ${\bar{\bm{\omega}}}$.
When $k=0$, we define
$\ommsub{0}=\zerov{0}$, the only element of
$\extspac{0}={\mathbb{R}}^0$.
In this case, ${\bf V}=\zeromat{n}{0}$, the only matrix
in ${\mathbb{R}}^{n\times 0}$.
Thus, ${\bf V}\ommsub{0}=\zeromat{n}{0} \zerov{0} = \zerov{n}$,
or more simply,
${\bf V}{\bar{\bm{\omega}}}={\bf 0}$.
(See Section~\ref{sec:more-prelim}.)
In matrix notation, Corollary~\ref{cor:h:1} states that astral space
consists exactly of all points of the form
${\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$ for some matrix ${\bf V}$ and vector ${\bf q}$.
That is,
\begin{equation} \label{eq:i:6}
\extspac{n} = \{ {\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q} : {\bf V}\in{\mathbb{R}}^{n\times k},
{\bf q}\in{\R^n}, k\geq 0 \}.
\end{equation}
Thus, every point
$\overline{\xx}\in\extspac{n}$ can be written in this form
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$.
We refer to ${\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ (or more formally, the pair
$({\bf V},{\bf q})$) as a
\emph{matrix representation}, or
\emph{matrix form}, of the point $\overline{\xx}$.
In matrix form, matrix product is simple to compute:
If ${\bf A}\in{\R^{m\times n}}$ and $\overline{\xx}\in\extspac{n}$, then we can write $\overline{\xx}$ in
the form
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ as above,
and can then immediately compute the corresponding form of
${\bf A}\overline{\xx}$ using Proposition~\ref{pr:h:4} as
\[ {\bf A}\overline{\xx} = ({\bf A}{\bf V}){\bar{\bm{\omega}}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf A}{\bf q}). \]
Here is another simple observation, which implies that if
$\overline{\xx}$ has the form above, then $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$ if and only if
${\bf u}$ is orthogonal to all of the columns of ${\bf V}$:
\begin{proposition} \label{pr:vtransu-zero}
Let $\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$
where ${\bf V}\in{\mathbb{R}}^{n\times k}$, for some $k\geq 0$, and ${\bf q}\in{\R^n}$.
For all ${\bf u}\in{\R^n}$,
if $\trans{{\bf V}} {\bf u} = {\bf 0}$ then
$\overline{\xx}\cdot{\bf u}={\bf q}\cdot{\bf u}\in{\mathbb{R}}$;
otherwise,
$\overline{\xx}\cdot{\bf u}={\bf V}{\bar{\bm{\omega}}}\cdot{\bf u} \in \{-\infty,+\infty\}$.
\end{proposition}
\proof
Suppose ${\bf v}_1,\ldots,{\bf v}_k$ are the columns of ${\bf V}$.
Then $\overline{\xx}\cdot{\bf u}$ is exactly as given in \eqref{eqn:pf:2}, and
${\bf V}{\bar{\bm{\omega}}}\cdot{\bf u}$ can be computed using this same equation with ${\bf q}$
replaced by ${\bf 0}$.
From these, the proposition now follows straightforwardly.
\qed
In general, an astral point's matrix representation need not be
uniquely determined by the point itself.
Nevertheless, we will see next how
standard notions from linear algebra
can be used to understand when two
representations are equivalent, and how every point can be
expressed in a unique and canonical form.
Recall that
a square matrix ${\bf R}\in{\mathbb{R}}^{k\times k}$ with entries $[r_{ij}]$
is \emph{upper triangular} if $r_{ij} = 0$ for all $i>j$.
We say ${\bf R}$ is \emph{positive upper triangular} if it is upper
triangular and if all of its diagonal entries are strictly positive.
Also, we say that a matrix ${\bf V}\in{\mathbb{R}}^{n\times k}$ is
\emph{column-orthogonal} if its columns are orthonormal (that is,
if the columns all have unit length and are orthogonal to one
another).
Matrices of
these types will play a significant role in manipulating and
comparing how astral points are represented.
Note that, vacuously,
$\zeromat{0}{0}$ is positive upper triangular,
and $\zeromat{n}{0}$ is column-orthogonal.
Here are some simple, well-known facts about positive upper triangular
matrices:
\begin{proposition} \label{thm:put-prop}
~
\begin{enumerate}
\item \label{thm:put-prop:a}
The product of two positive upper triangular matrices is also
positive upper triangular.
\item \label{thm:put-prop:b}
Every positive upper triangular matrix ${\bf R}\in{\mathbb{R}}^{k\times k}$ is
invertible, and its inverse, ${\RR^{-1}}$, is also positive upper
triangular.
\item \label{thm:put-prop:c}
Let ${\bf V}\in {\mathbb{R}}^{n\times k}$ have full column rank.
Then there exists a column-orthogonal matrix ${\bf W}\in{\mathbb{R}}^{n\times k}$
and a positive upper triangular matrix
${\bf R}\in{\mathbb{R}}^{k\times k}$ such that ${\bf V}={\bf W}{\bf R}$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{thm:put-prop:a}) is straightforward to check.
Part~(\ref{thm:put-prop:b}):
Let ${\bf R}$ have rows ${\bf r}_1,\ldots,{\bf r}_k$ and entries $r_{ij}$.
By induction on $j$, each standard basis vector
${\bf e}_j$ can be written as
${\bf e}_j=\sum_{i=1}^j s_{ij} {\bf r}_i$
for some $s_{1j},\ldots,s_{jj}\in{\mathbb{R}}$ with
$s_{jj}>0$.
This is because
${\bf r}_j=\sum_{i=1}^j r_{ij} {\bf e}_i$
so
\[
{\bf e}_j = \frac{1}{r_{jj}}
\brackets{ {\bf r}_j - \sum_{i=1}^{j-1} r_{ij}{\bf e}_i },
\]
which implies, by inductive hypothesis, that ${\bf e}_j$ can be written as
above as a linear combination of ${\bf r}_1,\ldots,{\bf r}_j$
with $s_{jj}=1/r_{jj}>0$.
Setting $s_{ij}=0$ for $i>j$ and letting ${\bf S}$
denote the $k\times k$ matrix with entries $s_{ij}$,
it follows that ${\bf S}$ is positive upper
triangular, and that ${\bf R}{\bf S}={\bf I}$, that is, ${\bf S}={\RR^{-1}}$.
Part~(\ref{thm:put-prop:c}) follows using a standard QR factorization.
In more detail, let ${\bf V}'\in{\mathbb{R}}^{n\times n}$ be a matrix whose first
$k$ columns are the same as ${\bf V}$, and whose remaining $n-k$ columns
are chosen so that all $n$ columns are linearly independent of one
another, that is, so that the entire matrix ${\bf V}'$ will have full rank
$n$.
Using Gram-Schmidt, the columns of ${\bf V}'$ can be transformed into an
orthonormal basis for ${\R^n}$.
Formally, this is equivalent to a QR factorization of ${\bf V}'$ which
states that ${\bf V}'={\bf W}'{\bf R}'$ for some column-orthogonal matrix
${\bf W}'\in{\mathbb{R}}^{n\times n}$ and some positive
upper triangular matrix ${\bf R}'\in{\mathbb{R}}^{n\times n}$.
(See \citet[Theorem 2.6.1]{horn_johnson}.)
The result now follows by
letting ${\bf W}$ be the first $k$ columns of ${\bf W}'$, and letting ${\bf R}$ be
the $k\times k$ submatrix of ${\bf R}'$ consisting of the first $k$ rows
and columns.
\qed
Returning to astral space,
we show that the astral point ${\bar{\bm{\omega}}}=\ommsub{k}$ defined in
\eqref{eq:i:7} is unchanged when multiplied by a positive upper
triangular matrix ${\bf R}$.
This will immediately imply that an astral point with representation
${\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
can be expressed alternatively with ${\bf V}$ replaced by ${\bf V}{\bf R}$.
\begin{lemma} \label{lem:i:1}
Let $k\geq 0$ and let ${\bf R}\in{\mathbb{R}}^{k\times k}$ be a positive upper
triangular matrix.
Then ${\bf R}{\bar{\bm{\omega}}}={\bar{\bm{\omega}}}$.
\end{lemma}
\proof
Let ${\bf r}_1,\ldots,{\bf r}_k$ be the columns of ${\bf R}$, and let $[r_{ij}]$
denote ${\bf R}$'s individual entries.
Then
\begin{eqnarray*}
{\bar{\bm{\omega}}} &=& \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf e}_k} \\
{\bf R}{\bar{\bm{\omega}}} &=& \limray{{\bf r}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \cdots \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf r}_k}
\end{eqnarray*}
by Proposition~\ref{pr:h:4}(\ref{pr:h:4c},\ref{pr:h:4e}).
To prove the lemma, we show that $({\bf R}{\bar{\bm{\omega}}})\cdot{\bf u} = {\bar{\bm{\omega}}}\cdot{\bf u}$
for all ${\bf u}\in{\mathbb{R}}^{k}$.
Let ${\bf u}\in{\mathbb{R}}^{k}$.
If ${\bf u}={\bf 0}$, this clearly holds since $\overline{\xx}\cdot{\bf 0}=0$ for all
$\overline{\xx}\in\extspac{k}$,
so we assume henceforth that ${\bf u}\neq{\bf 0}$.
Let us write ${\bf u}=\trans{[u_1,\ldots,u_k]}$, and let $\ell$ be the
smallest index for which $u_{\ell}\neq 0$ (so $u_j=0$ for $j<\ell$).
If $j<\ell$ then $u_j={\bf e}_j\cdot{\bf u}=0$, and also
${\bf r}_j\cdot{\bf u}=\sum_{i=1}^k r_{ij} \, u_i = 0$
since $u_i=0$ for $i\leq j$,
and $r_{ij}=0$ for $i>j$
(since ${\bf R}$ is upper triangular).
Thus,
$ \limray{{\bf e}_{j}}\cdot{\bf u} = \limray{{\bf r}_{j}}\cdot{\bf u} = 0$.
By this same reasoning,
\[
{\bf r}_{\ell}\cdot{\bf u}
=
\sum_{i=1}^k r_{i \ell} \, u_i
=
r_{\ell \ell} \, u_{\ell}.
\]
Since $r_{\ell\ell}>0$,
this means that ${\bf r}_{\ell}\cdot{\bf u}$ must have the same sign as
$u_{\ell}={\bf e}_{\ell}\cdot{\bf u}$, and both are nonzero.
Therefore,
$ \limray{{\bf e}_{\ell}}\cdot{\bf u} = \limray{{\bf r}_{\ell}}\cdot{\bf u} $,
and both must be either $+\infty$ or $-\infty$.
Combining these facts yields
\[
({\bf R}{\bar{\bm{\omega}}})\cdot{\bf u}
=
\limray{{\bf r}_{\ell}}\cdot{\bf u}
=
\limray{{\bf e}_{\ell}}\cdot{\bf u}
=
{\bar{\bm{\omega}}}\cdot{\bf u}.
\]
Since this holds for all ${\bf u}\in{\mathbb{R}}^k$, the lemma now follows by
Proposition~\ref{pr:i:4}.
\qed
From this lemma, as noted above, we can immediately give conditions
under which two representations specify the same astral point.
Later, in Theorem~\ref{thm:i:2}, we prove a converse.
In what follows, $\spn S$ denotes the span of any set $S\subseteq{\R^n}$
(where span of the empty set is defined to be $\{{\bf 0}\}$).
Also, recall that the column space of a matrix ${\bf V}$
is the span of its columns.
\begin{proposition} \label{pr:g:1}
Let
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$
and
$\overline{\xx}'={\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}'$
be two points in $\extspac{n}$, for some ${\bf V},{\bf V}'\in{\mathbb{R}}^{n\times k}$,
$k\geq 0$, and ${\bf q},{\bf q}'\in{\R^n}$.
Suppose ${\bf V}'={\bf V}{\bf R}$, for some positive upper triangular matrix
${\bf R}\in{\mathbb{R}}^{k\times k}$, and that ${\bf q}'-{\bf q}$ is in the column space of ${\bf V}$.
Then $\overline{\xx}=\overline{\xx}'$.
\end{proposition}
\proof
First, by Lemma~\ref{lem:i:1} and
Proposition~\ref{pr:h:4}(\ref{pr:h:4d}),
\[
{\bf V}'{\bar{\bm{\omega}}} = ({\bf V}{\bf R}){\bar{\bm{\omega}}} = {\bf V} ({\bf R}{\bar{\bm{\omega}}}) = {\bf V} {\bar{\bm{\omega}}}.
\]
(If $k=0$, then both sides are equal to ${\bf 0}$.)
Thus, $\overline{\xx}'={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}'$.
To prove the proposition, we show that $\overline{\xx}\cdot{\bf u}=\overline{\xx}'\cdot{\bf u}$
for all ${\bf u}\in{\R^n}$,
which will imply $\overline{\xx}=\overline{\xx}'$ by Proposition~\ref{pr:i:4}.
Let ${\bf u}\in{\R^n}$.
If $\trans{{\bf V}}{\bf u}\neq{{\bf 0}}$, then
$\overline{\xx}\cdot{\bf u}={\bf V}{\bar{\bm{\omega}}}\cdot{\bf u}=\overline{\xx}'\cdot{\bf u}$
by Proposition~\ref{pr:vtransu-zero}.
Otherwise,
if $\trans{{\bf V}}{\bf u} = {\bf 0}$, then
that same proposition implies that,
$\overline{\xx}\cdot{\bf u} = {\bf q}\cdot{\bf u}$
and
$\overline{\xx}'\cdot{\bf u} = {\bf q}'\cdot{\bf u}$.
By assumption, ${\bf q}'-{\bf q} = {\bf V}{\bf b}$ for some ${\bf b}\in{\mathbb{R}}^k$.
Thus,
\[
{\bf q}'\cdot{\bf u} - {\bf q}\cdot{\bf u}
= ({\bf q}' - {\bf q})\cdot{\bf u}
= \trans{{\bf u}}({\bf q}' - {\bf q})
= \trans{{\bf u}} {\bf V} {\bf b}
= \trans{{\bf 0}} {\bf b} = 0.
\]
Therefore,
$
\overline{\xx}\cdot{\bf u}
= {\bf q}\cdot{\bf u}
= {\bf q}'\cdot{\bf u}
= \overline{\xx}'\cdot{\bf u}
$.
\qed
The next proposition, which directly follows from
Proposition~\ref{pr:g:1}, shows that any point ${\bf V}{\bar{\bm{\omega}}}$ can be
rewritten using a linearly independent subset of the columns of ${\bf V}$.
\begin{proposition} \label{pr:g:2}
Let
$\overline{\xx}= [{\bf v}_1,\ldots,{\bf v}_k]{\bar{\bm{\omega}}}$
with $k>0$.
Then there exists a subset of indices
$ 1\leq i_1 < i_2 < \cdots < i_{\ell} \leq k $
such that
$\overline{\xx}=[{\bf v}_{i_1},\ldots,{\bf v}_{i_{\ell}}]{\bar{\bm{\omega}}}$
and
${\bf v}_{i_1},\ldots,{\bf v}_{i_{\ell}}$
are linearly independent.
\end{proposition}
\proof
Suppose ${\bf v}_i\in\spnfin{{\bf v}_1,\ldots,{\bf v}_{i-1}}$ so that
${\bf v}_i = a_1 {\bf v}_1 + \cdots + a_{i-1} {\bf v}_{i-1}$
for some $a_1,\ldots,a_{i-1}\in{\mathbb{R}}$.
Let ${\bf R}\in{\mathbb{R}}^{k\times k}$ be a positive upper triangular matrix that is
the same as the identity matrix, except the $i$-th column is equal
to $\trans{[-a_1, \ldots, -a_{i-1}, 1, 0,\ldots,0]}$.
Then Proposition~\ref{pr:g:1}, applied to
$[{\bf v}_1,\ldots,{\bf v}_k]$ and ${\bf R}$,
implies that
$\overline{\xx}$ is equal to the same
representation with ${\bf v}_i$ replaced by ${\bf 0}$, or equivalently, with ${\bf v}_i$
removed entirely.
By repeatedly removing vectors in this way, eventually we must end up
with a linearly independent subset, as claimed in the proposition.
\qed
For any point $\overline{\xx}\in\extspac{n}$,
we define the \emph{astral rank} of $\overline{\xx}$ to be
the smallest number of columns $k$ for which there exists a matrix
${\bf V}\in{\mathbb{R}}^{n\times k}$
and ${\bf q}\in{\R^n}$ with $\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$.
Informally, astral rank captures the number, or really the
dimensionality, of the directions in which a trajectory with limit
$\overline{\xx}$ is going to infinity.
The set of points with astral rank $0$ consists exactly of ${\R^n}$.
Points with astral rank $1$ are those that are reached by a trajectory
that follows a ray to infinity (that is, of the form
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$, for some ${\bf v},{\bf q}\in{\R^n}$).
Points with higher astral ranks are reached by following one dominant
direction to infinity, along with a secondary direction, possibly a
tertiary direction, and so on, as discussed in
Section~\ref{subsec:astral-pt-form}.
\begin{proposition} \label{pr:h:3}
Let $\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some
${\bf V}\in{\mathbb{R}}^{n\times k}$ and ${\bf q}\in{\R^n}$.
Then the astral rank of $\overline{\xx}$ is equal to the
matrix rank of ${\bf V}$.
\end{proposition}
\proof
Let $r$ be the matrix rank of ${\bf V}$.
By Proposition~\ref{pr:g:2}, the matrix ${\bf V}$ can be reduced down to a
linearly independent subset of columns, of which there cannot be more
than $r$.
Thus, the astral rank of $\overline{\xx}$ is at most $r$.
Let
$\overline{\xx}'={\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$,
where ${\bf q}'\in{\R^n}$, ${\bf V}'\in{\mathbb{R}}^{n\times r'}$, for some $r'<r$.
To prove the astral rank of $\overline{\xx}$ cannot be strictly smaller than $r$,
we will show that $\overline{\xx}$ cannot be equal to any such $\overline{\xx}'$.
Let $L$ be the linear subspace orthogonal to the columns of ${\bf V}$:
\begin{equation} \label{eqn:subspace-orth-cols-v}
L = \{{\bf u}\in{\R^n} : \trans{{\bf V}} {\bf u} = {\bf 0} \}.
\end{equation}
Then, by Proposition~\ref{pr:vtransu-zero},
$\overline{\xx}\cdot{\bf u}$ is in ${\mathbb{R}}$ if and only if ${\bf u}\in L$.
Let $L'$ be defined correspondingly for ${\bf V}'$, which similarly captures
exactly those ${\bf u}$ for which $\overline{\xx}'\cdot{\bf u}\in{\mathbb{R}}$.
The dimension of $L$ is
$n-r$, while the dimension of $L'$ is at least $n-r'>n-r$.
Thus, $L\neq L'$, implying $\overline{\xx}\cdot{\bf u}\neq\overline{\xx}'\cdot{\bf u}$ for some
${\bf u}\in{\R^n}$, so $\overline{\xx}\neq\overline{\xx}'$.
Therefore, the astral rank of $\overline{\xx}$ is exactly $r$.
\qed
\subsection{Canonical representation}
\label{sec:canonical:rep}
Although the same astral point may have multiple representations, every point
has a single representation that is unique and perhaps the most
natural.
We say that a representation ${\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ of an astral point
$\overline{\xx}$ is \emph{canonical} if ${\bf V}$ is column-orthogonal
and if ${\bf q}$ is orthogonal to all of ${\bf V}$'s columns (so that
$\trans{{\bf V}}{\bf q}={\bf 0}$).
By Proposition~\ref{pr:h:3}, the number of columns comprising ${\bf V}$
will necessarily be equal to $\overline{\xx}$'s astral rank.
This representation exists and is unique for every point in
$\extspac{n}$.
\begin{theorem} \label{pr:uniq-canon-rep}
Every point in $\extspac{n}$ has a unique canonical representation.
\end{theorem}
Before proving this theorem, we give a lemma that will be one of the
main steps in proving uniqueness.
We state the lemma in a form that is a bit more general than needed
at this point since it will get used again shortly.
\begin{lemma} \label{lem:uniq-ortho-prefix}
Let ${\bf V}$ and ${\bf V}'$ be column-orthogonal matrices in ${\mathbb{R}}^{n\times k}$,
and let $\overline{{\bf z}},\,\overline{{\bf z}}'\in\extspac{n}$.
If ${\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}={\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'$
then ${\bf V}={\bf V}'$.
\end{lemma}
\proof
Let
$\overline{\xx} = {\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$
and
$\overline{\xx}' = {\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'$.
Also, let us write out the orthonormal columns of the matrices as
${\bf V}=[{\bf v}_1,\ldots,{\bf v}_k]$
and
${\bf V}'=[{\bf v}'_1,\ldots,{\bf v}'_k]$.
To prove the lemma, we assume
${\bf V}\neq{\bf V}'$, and show this implies $\overline{\xx}\neq\overline{\xx}'$.
Let $j$ be the smallest index for which ${\bf v}_j\neq{\bf v}'_j$
so that ${\bf v}_i={\bf v}'_i$ for $i<j$.
Then
${\bf v}_j\cdot({\bf v}_j-{\bf v}'_j)=1-{\bf v}_j\cdot{\bf v}'_j>0$
since $\norm{{\bf v}_j}=\norm{{\bf v}'_j}=1$ and ${\bf v}_j\neq{\bf v}'_j$,
implying
$\limray{{\bf v}_j}\cdot({\bf v}_j-{\bf v}'_j)=+\infty$
(by Proposition~\ref{pr:astrons-exist}).
Also, for $i<j$,
${\bf v}_j$ and ${\bf v}'_j$ are both orthogonal to ${\bf v}_i={\bf v}'_i$,
so
${{\bf v}_i}\cdot({\bf v}_j-{\bf v}'_j)=0$,
implying
$\limray{{\bf v}_i}\cdot({\bf v}_j-{\bf v}'_j)=0$.
Together, these facts imply
$
\overline{\xx}\cdot({\bf v}_j-{\bf v}'_j)
= +\infty
$
(using $\overline{\xx}$'s form).
By a similar argument,
${\bf v}'_j\cdot({\bf v}_j-{\bf v}'_j) = {\bf v}'_j\cdot{\bf v}_j - 1 < 0$,
so
$ \overline{\xx}'\cdot({\bf v}_j-{\bf v}'_j)=-\infty $.
Thus,
$ \overline{\xx}\cdot({\bf v}_j-{\bf v}'_j) \neq \overline{\xx}'\cdot({\bf v}_j-{\bf v}'_j)$
so $\overline{\xx}\neq\overline{\xx}'$.
\qed
\proof[Proof of Theorem~\ref{pr:uniq-canon-rep}]
~
\underline{Existence}:
Let $\overline{\xx}\in\extspac{n}$.
Then by Corollary~\ref{cor:h:1}, there exists some
representation of $\overline{\xx}$ of the form
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$,
for some ${\bf V}\in{\mathbb{R}}^{n\times k}$ and ${\bf q}\in{\R^n}$, $k\geq 0$.
By Proposition~\ref{pr:g:2}, we can assume without loss of generality
that ${\bf V}$ has full column rank.
By Proposition~\ref{thm:put-prop}(\ref{thm:put-prop:c}), there exists
a column-orthogonal matrix $\tilde{\VV}\in{\mathbb{R}}^{n\times k}$ and a positive upper triangular
matrix ${\bf R}\in{\mathbb{R}}^{k\times k}$ with ${\bf V}=\tilde{\VV}{\bf R}$.
Also, let $\tilde{\qq}={\bf q}-\tilde{\VV}\trans{\tilde{\VV}}{\bf q}$, which is the projection of
${\bf q}$ onto the space perpendicular to the columns of $\tilde{\VV}$.
This implies $\trans{\tilde{\VV}}\tilde{\qq}={\bf 0}$
(since $\trans{\tilde{\VV}}\tilde{\VV}$ is the $k\times k$ identity matrix),
and that ${\bf q}-\tilde{\qq}=\tilde{\VV}(\trans{\tilde{\VV}}{\bf q})$ is in the
column space of $\tilde{\VV}$.
Therefore,
by Proposition~\ref{pr:g:1}, $\overline{\xx}=\tilde{\VV}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\tilde{\qq}$, a
representation that is in canonical form.
\underline{Uniqueness}:
Consider canonical representations of two points,
$\overline{\xx} = {\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
and
$\overline{\xx}' = {\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$,
where
${\bf V}\in{\mathbb{R}}^{n\times k}$,
${\bf V}'\in{\mathbb{R}}^{n\times k'}$,
and ${\bf q},{\bf q}'\in{\R^n}$.
To prove uniqueness, we assume henceforth that
$\overline{\xx}=\overline{\xx}'$, and show that this implies
that these representations must be identical.
First, since $\overline{\xx}=\overline{\xx}'$, these points must have the same astral rank.
Therefore, $k'=k$, by Proposition~\ref{pr:h:3}.
As a result, we can apply Lemma~\ref{lem:uniq-ortho-prefix}
(since ${\bf V}$ and ${\bf V}'$ are column-orthogonal, and
with $\overline{{\bf z}}={\bf q}$ and $\overline{{\bf z}}'={\bf q}'$),
implying ${\bf V}={\bf V}'$.
Finally, because ${\bf q}$ and ${\bf q}'$ are both orthogonal to all the columns
of ${\bf V}={\bf V}'$, Propostion~\ref{pr:vtransu-zero} yields
\begin{eqnarray*}
\overline{\xx}\cdot({\bf q}-{\bf q}') &=& {\bf q} \cdot({\bf q} - {\bf q}') \\
\overline{\xx}'\cdot({\bf q}-{\bf q}') &=& {\bf q}' \cdot({\bf q} - {\bf q}').
\end{eqnarray*}
Since $\overline{\xx}=\overline{\xx}'$, these quantities must be equal.
Taking their difference yields
$\norm{{\bf q}-{\bf q}'}^2=0$, and therefore ${\bf q}={\bf q}'$.
Thus, the representations are identical, as claimed,
completing the proof.
\qed
With these tools and concepts, we can now prove a converse to
Proposition~\ref{pr:g:1}, thus providing a complete characterization of
when two representations actually refer to the same astral point.
\begin{theorem} \label{thm:i:2}
Let
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$
and
$\overline{\xx}'={\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}'$
be two points in $\extspac{n}$, for some ${\bf V},{\bf V}'\in{\mathbb{R}}^{n\times k}$,
$k\geq 0$, and ${\bf q},{\bf q}'\in{\R^n}$.
Assume ${\bf V}$ and ${\bf V}'$ both have full column rank.
Then $\overline{\xx}=\overline{\xx}'$ if and only if
${\bf V}'={\bf V}{\bf R}$ for some positive upper triangular matrix
${\bf R}\in{\mathbb{R}}^{k\times k}$, and also ${\bf q}'-{\bf q}$ is in the column space of ${\bf V}$.
\end{theorem}
\proof
The fact that the given conditions imply $\overline{\xx}=\overline{\xx}'$ is proved in
Proposition~\ref{pr:g:1}.
To prove the converse, assume $\overline{\xx}=\overline{\xx}'$.
As in the proof of Theorem~\ref{pr:uniq-canon-rep},
there exists a canonical representation $\tilde{\VV}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\tilde{\qq}$ for
$\overline{\xx}$ where ${\bf V}=\tilde{\VV}{\bf R}$ for some positive upper triangular
matrix ${\bf R}\in{\mathbb{R}}^{k\times k}$
(by Proposition~\ref{thm:put-prop}(\ref{thm:put-prop:c})),
and where $\tilde{\qq}={\bf q}-\tilde{\VV}{\bf b}$ with ${\bf b}=\trans{\tilde{\VV}}{\bf q}$.
A corresponding canonical representation $\tilde{\VV}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\tilde{\qq}'$
exists for $\overline{\xx}'$ (with ${\bf V}'=\tilde{\VV}'{\bf R}'$, ${\bf R}'\in{\mathbb{R}}^{k\times k}$
positive upper triangular, $\tilde{\qq}'={\bf q}'-\tilde{\VV}'{\bf b}'$,
${\bf b}'=\trans{\mbox{$\tilde{\VV}'$}}{\bf q}'$).
Since $\overline{\xx}=\overline{\xx}'$, by
Theorem~\ref{pr:uniq-canon-rep}, these two canonical
representations must actually be the same, that is, $\tilde{\VV}=\tilde{\VV}'$
and $\tilde{\qq}=\tilde{\qq}'$.
Thus, by algebra,
$ {\bf V}'={\bf V} ({\RR^{-1}} {\bf R}') $
and
$ {\bf q}' - {\bf q} = {\bf V} {\RR^{-1}} ({\bf b}' - {\bf b}) $.
Proposition~\ref{thm:put-prop} implies both that ${\RR^{-1}}$ exists and that
${\RR^{-1}}{\bf R}'$ is positive upper triangular, completing the theorem.
\qed
\subsection{Representation and sequences}
\label{sec:rep-seq}
In addition to representing and characterizing astral points
themselves, we can also precisely describe the sequences that converge
to a particular point with a particular representation.
Suppose the columns of some matrix ${\bf V}\in{\mathbb{R}}^{n\times k}$ are linearly
independent.
Let $\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some ${\bf q}\in{\R^n}$;
without loss of generality, by Theorem~\ref{thm:i:2},
we can assume $\trans{{\bf V}}{\bf q}={\bf 0}$.
(This is because if $\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$ for some ${\bf q}'\in{\R^n}$, then we can
write ${\bf q}'={\bf V}{\bf c}+{\bf q}$ for some ${\bf c}\in{\R^k}$ and ${\bf q}\in{\R^n}$ with
$\trans{{\bf V}}{\bf q}={\bf 0}$, implying
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ since ${\bf q}'-{\bf q}$ is in the column space of
${\bf V}$.)
Let $\seq{{\bf x}_t}$ be any sequence in ${\R^n}$.
Then by linear algebra, each point ${\bf x}_t$ can be uniquely
represented in the form ${\bf x}_t={\bf V} {\bf b}_t + {\bf q}_t$ for some ${\bf b}_t\in{\R^k}$ and
some point ${\bf q}_t\in{\R^n}$ that is orthogonal to the columns of ${\bf V}$, so
that $\trans{{\bf V}}{\bf q}_t = {\bf 0}$.
The next theorem provides necessary and sufficient conditions for
when the sequence $\seq{{\bf x}_t}$ will converge to $\overline{\xx}$,
in terms of this representation.
(As usual, $b_{t,i}$ denotes the $i$-th component of ${\bf b}_t$,
and convergence means in the limit as $t\rightarrow+\infty$.)
\begin{theorem} \label{thm:seq-rep}
Let ${\bf V}\in{\mathbb{R}}^{n\times k}$ be a matrix with $k\geq 0$ columns that are
linearly independent.
Let $\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where ${\bf q}\in{\R^n}$ and
$\trans{{\bf V}}{\bf q}={\bf 0}$.
Let $\seq{{\bf x}_t}$ be a sequence in ${\R^n}$ with
${\bf x}_t={\bf V} {\bf b}_t + {\bf q}_t$ for some ${\bf b}_t\in{\R^k}$ and ${\bf q}_t\in{\R^n}$
with $\trans{{\bf V}}{\bf q}_t = {\bf 0}$, for all $t$.
Then ${\bf x}_t\rightarrow\overline{\xx}$ if and only if all of the following
hold:
\begin{enumerate}
\item \label{thm:seq-rep:a}
$b_{t,i}\rightarrow+\infty$, for $i=1,\ldots,k$.
\item \label{thm:seq-rep:b}
$b_{t,i+1}/b_{t,i}\rightarrow 0$, for $i=1,\ldots,k-1$.
\item \label{thm:seq-rep:c}
${\bf q}_t\rightarrow{\bf q}$.
\end{enumerate}
\end{theorem}
\proof
Let ${\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$ be the columns of ${\bf V}$.
Suppose first that
(\ref{thm:seq-rep:a}),
(\ref{thm:seq-rep:b}) and
(\ref{thm:seq-rep:c})
hold.
We aim to show that ${\bf x}_t\rightarrow\overline{\xx}$.
By Theorem~\ref{thm:i:1}(\ref{thm:i:1c}),
it suffices to show ${\bf x}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$ for all
${\bf u}\in{\R^n}$.
Note first that
(\ref{thm:seq-rep:a}) implies, for each $i=1,\ldots,k$, that
$b_{t,i}>0$ for all but finitely many values of $t$.
By discarding from the sequence any element ${\bf x}_t$ for which
any component of ${\bf b}_t$ is nonpositive, we can therefore assume in
what follows that $b_{t,i}>0$ for all $t$ and all
$i\in\{1,\ldots,k\}$.
Let ${\bf u}\in{\R^n}$.
Suppose first that $\trans{{\bf V}}{\bf u}={\bf 0}$.
Then
\[
{\bf x}_t\cdot{\bf u}
=
({\bf V} {\bf b}_t)\cdot{\bf u} + {\bf q}_t\cdot{\bf u}
=
{\bf q}_t\cdot{\bf u}
\rightarrow
{\bf q}\cdot{\bf u}
=
\overline{\xx}\cdot{\bf u}.
\]
The second equality is because
$({\bf V} {\bf b}_t)\cdot{\bf u} = \trans{{\bf u}}{\bf V}{\bf b}_t = {\bf 0}$.
The last equality is by Proposition~\ref{pr:vtransu-zero}.
Convergence is from (\ref{thm:seq-rep:c}).
Otherwise, in the alternative case, $\trans{{\bf V}}{\bf u}\neq{\bf 0}$.
Let $j\in\{1,\ldots,k\}$ be the first index for which
${\bf v}_j\cdot{\bf u}\neq 0$, so that
${\bf v}_i\cdot{\bf u}=0$ for $i=1,\ldots,j-1$.
Assume further that ${\bf v}_j\cdot{\bf u}>0$; the remaining case that
${\bf v}_j\cdot{\bf u}<0$ can be proved symmetrically
(or by applying the same argument to $-{\bf u}$).
Since $\overline{\xx}\cdot{\bf u}$ must have the form given in \eqref{eqn:pf:2},
these facts immediately imply $\overline{\xx}\cdot{\bf u}=+\infty$.
Also,
\begin{eqnarray}
{\bf x}_t\cdot{\bf u}
&=&
\sum_{i=1}^k b_{t,i} {\bf v}_i \cdot {\bf u} + {\bf q}_t\cdot{\bf u}
\nonumber
\\
&=&
b_{t,j} \brackets{ {\bf v}_j\cdot{\bf u}
+
\sum_{i=j+1}^k \frac{b_{t,i}}{b_{t,j}} {\bf v}_i\cdot{\bf u}
}
+ {{\bf q}_t\cdot{\bf u}}.
\label{eqn:thm:seq-rep:1}
\end{eqnarray}
For each $i=j+1,\ldots,k$,
\[
\frac{b_{t,i}}{b_{t,j}}
=
\prod_{\ell=j}^{i-1} {\frac{b_{t,\ell+1}}{b_{t,\ell}}}
\rightarrow 0
\]
from (\ref{thm:seq-rep:b}).
Therefore, the bracketed expression in
\eqref{eqn:thm:seq-rep:1} is converging to ${\bf v}_j\cdot{\bf u}>0$, and so
is at least ${\bf v}_j\cdot{\bf u} / 2 > 0$ for $t$ sufficiently large.
Since $b_{t,j}\rightarrow+\infty$, by (\ref{thm:seq-rep:a}),
it follows that the term on the left of
\eqref{eqn:thm:seq-rep:1} is converging to $+\infty$, while
${\bf q}_t\cdot{\bf u}$ is converging to ${\bf q}\cdot{\bf u}\in{\mathbb{R}}$
from (\ref{thm:seq-rep:c}).
Therefore,
${\bf x}_t\cdot{\bf u}\rightarrow+\infty=\overline{\xx}\cdot{\bf u}$.
Thus, in all cases, ${\bf x}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$ for all
${\bf u}\in{\R^n}$, so ${\bf x}_t\rightarrow\overline{\xx}$.
For the converse, assume for the rest of the proof that
${\bf x}_t\rightarrow\overline{\xx}$.
We prove the three parts of the theorem, beginning with
(\ref{thm:seq-rep:c}).
\underline{(\ref{thm:seq-rep:c})}:
By Proposition~\ref{thm:put-prop}(\ref{thm:put-prop:c}),
we can write ${\bf V}={\bf W}{\bf R}$ for some column-orthogonal matrix ${\bf W}\in{\mathbb{R}}^{n\times k}$
and some positive upper triangular matrix
${\bf R}\in{\mathbb{R}}^{k\times k}$.
Let ${\bf P}={\bf I} - {\bf W}\trans{{\bf W}}$
(where ${\bf I}$ is the $n\times n$ identity matrix),
which projects any vector ${\bf x}\in{\R^n}$ onto
the linear subspace that is orthogonal to the columns of ${\bf W}$.
Note that
${\bf W}\trans{{\bf W}}{\bf V} = {\bf W}\trans{{\bf W}}{\bf W}{\bf R} = {\bf W}{\bf R} = {\bf V}$,
since ${\bf W}$ is column-orthogonal.
Thus, ${\bf P}{\bf V}=\zeromat{n}{k}$, the $n\times k$ all-zeros matrix.
Also, if $\trans{{\bf V}}{\bf u}={\bf 0}$, then
$\trans{{\bf W}}{\bf u}=\trans{({\bf V}{\RR^{-1}})}{\bf u}=\trans{({\RR^{-1}})}\trans{{\bf V}}{\bf u}={\bf 0}$
so ${\bf P}{\bf u}={\bf u}-{\bf W}\trans{{\bf W}}{\bf u}={\bf u}$.
Combining, these show that
${\bf P}{\bf x}_t={\bf P}{\bf V}{\bf b}_t+{\bf P}{\bf q}_t={\bf q}_t$
and
${\bf P}\overline{\xx}={\bf P}{\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf P}{\bf q}={\bf q}$.
Therefore,
since ${\bf x}_t\rightarrow\overline{\xx}$,
${\bf q}_t={\bf P}{\bf x}_t\rightarrow{\bf P}\overline{\xx}={\bf q}$,
by Proposition~\ref{pr:aff-cont}, proving
(\ref{thm:seq-rep:c}).
Before continuing, we observe the following:
Since the columns of ${\bf V}$ are linearly
independent, the matrix ${\bf V}$ has full rank $k$, which means that its
transpose $\trans{{\bf V}}$ does as well.
Therefore, the column space of $\trans{{\bf V}}$ has dimension $k$, which
thus must be all of ${\R^k}$.
As a consequence, for all ${\bf y}\in{\R^k}$, there exists ${\bf u}\in{\R^n}$ such
that $\trans{{\bf V}}{\bf u} = {\bf y}$.
\underline{(\ref{thm:seq-rep:a})}:
Let $i\in\{1,\ldots,k\}$, and let ${\bf y}\in{\R^k}$ be such that
$y_i=1$ and $y_j=0$ for $j\neq i$.
By the preceding observation, there exists ${\bf u}\in{\R^n}$ such that
$\trans{{\bf V}}{\bf u} = {\bf y}$,
which means ${\bf v}_j\cdot{\bf u} = y_j$ for $j=1,\ldots,k$.
These choices imply $\overline{\xx}\cdot{\bf u}=+\infty$
(since $\overline{\xx}\cdot{\bf u}$ has the form given in \eqref{eqn:pf:2}).
Expanding ${\bf x}_t$'s representation, for each $t$, yields
\begin{eqnarray}
{\bf x}_t\cdot{\bf u}
&=&
\paren{\sum_{j=1}^k b_{t,j} {\bf v}_j + {\bf q}_t}\cdot{\bf u}
\nonumber
\\
&=&
\sum_{j=1}^k b_{t,j} {\bf v}_j\cdot{\bf u} + {\bf q}_t\cdot{\bf u}
\nonumber
\\
&=&
\sum_{j=1}^k b_{t,j} y_j + {\bf q}_t\cdot{\bf u}
\nonumber
\\
&=&
b_{t,i} + {\bf q}_t\cdot{\bf u}.
\label{eqn:thm:seq-rep:2}
\end{eqnarray}
Since ${\bf q}_t\rightarrow{\bf q}$ (as shown above),
${\bf q}_t\cdot{\bf u}\rightarrow{\bf q}\cdot{\bf u}\in{\mathbb{R}}$.
By assumption, ${\bf x}_t\rightarrow\overline{\xx}$, so
${\bf x}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}=+\infty$.
Therefore,
\[
b_{t,i} = {\bf x}_t\cdot{\bf u} - {\bf q}_t\cdot{\bf u} \rightarrow +\infty,
\]
proving (\ref{thm:seq-rep:a}).
\underline{(\ref{thm:seq-rep:b})}:
Having proved (\ref{thm:seq-rep:a}), we can infer that $b_{t,i}>0$
for $i=1,\ldots,k$, for all but finitely many values of $t$;
by discarding these, we can assume that $b_{t,i}>0$ for
all $i$ and for all $t$.
Suppose, by way of contradiction, that (\ref{thm:seq-rep:b}) is false.
Then for some $i\in\{1,\ldots,k-1\}$, and some $\epsilon>0$,
$b_{t,i+1}/b_{t,i} \geq \epsilon$ for infinitely many values of $t$.
By discarding all other sequence elements, we can assume that this holds
for all $t$.
Similar to the proof of (\ref{thm:seq-rep:a}),
let ${\bf y}\in{\R^k}$ be such that $y_i = \epsilon$, $y_{i+1}=-2$ and
$y_j=0$ for $j\not\in\{i,i+1\}$.
By the observation above, there exists ${\bf u}\in{\R^n}$ such that
$\trans{{\bf V}}{\bf u}={\bf y}$.
As before, ${\bf v}_j\cdot{\bf u}=y_j$ for $j=1,\ldots,k$.
This implies $\overline{\xx}\cdot{\bf u}=+\infty$.
Similar to \eqref{eqn:thm:seq-rep:2}, we can compute
\begin{eqnarray*}
{\bf x}_t\cdot{\bf u}
&=&
\sum_{j=1}^k b_{t,j} y_j + {\bf q}_t\cdot{\bf u}
\\
&=&
\epsilon b_{t,i} - 2 b_{t,i+1} + {\bf q}_t\cdot{\bf u}
\\
&\leq&
\epsilon b_{t,i} - 2 \epsilon b_{t,i} + {\bf q}_t\cdot{\bf u}
\\
&=&
-\epsilon b_{t,i} + {\bf q}_t\cdot{\bf u}
\rightarrow
-\infty.
\end{eqnarray*}
The inequality follows from our assumption that
$b_{t,i+1} \geq \epsilon b_{t,i}$
for all $t$.
The convergence is
because $b_{t,i}\rightarrow+\infty$ from (\ref{thm:seq-rep:a}),
and because ${\bf q}_t\cdot{\bf u}\rightarrow{\bf q}\cdot{\bf u}\in{\mathbb{R}}$ from (\ref{thm:seq-rep:c}).
Thus, ${\bf x}_t\cdot{\bf u}\rightarrow-\infty$ but
$\overline{\xx}\cdot{\bf u}=+\infty$, contradicting that ${\bf x}_t\rightarrow\overline{\xx}$.
\qed
\subsection{Icons and galaxies}
\label{sec:core:zero}
In algebraic terms, we have seen that astral space $\extspac{n}$ is
closed under leftward addition (Proposition~\ref{pr:i:6}), and that
this operation is associative
(Proposition~\ref{pr:i:7}(\ref{pr:i:7a})).
This shows that astral space is a semigroup under this operation, and
furthermore is a monoid since ${\bf 0}$ is an identity element.
As such, we will especially be interested in the idempotents of this
semigroup, that is, the points $\overline{\ee}\in\extspac{n}$ for which
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}=\overline{\ee}$.
An astral point with this property is called an \emph{icon},
or is said to be \emph{iconic}, terms
derived as a contraction of ``idempotent'' and ``cone'' to refer both
to this property, and the cone-like properties of such points
that will be discussed later (see Section~\ref{sec:cones}
and Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:c})).
We write ${{\cal E}_n}$ for the set of all icons:
\[
{{\cal E}_n} = \braces{ \overline{\ee}\in\extspac{n} :
\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee} = \overline{\ee} }.
\]
The next proposition gives three equivalent characterizations for this
property.
In particular, this proposition implies that all astrons are iconic.
\begin{proposition} \label{pr:icon-equiv}
Let $\overline{\ee}\in\extspac{n}$.
Then the following are equivalent:
\begin{enumerate}
\item \label{pr:icon-equiv:a}
$\overline{\ee}$ is an icon; that is, $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}=\overline{\ee}$.
\item \label{pr:icon-equiv:b}
For all ${\bf u}\in{\R^n}$, $\overline{\ee}\cdot{\bf u}\in\{-\infty,0,+\infty\}$.
\item \label{pr:icon-equiv:c}
$\overline{\ee}={\bf V}{\bar{\bm{\omega}}}$ for some
${\bf V}\in{\mathbb{R}}^{n\times k}$, $k\geq 0$.
\end{enumerate}
Furthermore, the same equivalence holds if the matrix in
part~(\ref{pr:icon-equiv:c}) is required to be column-orthogonal.
\end{proposition}
\proof
~
\underline{(\ref{pr:icon-equiv:a}) $\Rightarrow$ (\ref{pr:icon-equiv:b})}:
Suppose $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}=\overline{\ee}$.
Then for all ${\bf u}\in{\R^n}$,
by Proposition~\ref{pr:i:6},
$\overline{\ee}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}\cdot{\bf u}=\overline{\ee}\cdot{\bf u}$, which is impossible
if $\overline{\ee}\cdot{\bf u}$ is a positive or negative real number, that is,
unless $\overline{\ee}\cdot{\bf u}$ is in $\{-\infty,0,+\infty\}$.
\underline{(\ref{pr:icon-equiv:b}) $\Rightarrow$ (\ref{pr:icon-equiv:c})}:
Suppose $\overline{\ee}$ satisfies (\ref{pr:icon-equiv:b}).
Let
$\overline{\ee}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
be its canonical representation
(which exists by Theorem~\ref{pr:uniq-canon-rep}).
Then because ${\bf q}$ is orthogonal to all the columns of ${\bf V}$,
by Proposition~\ref{pr:vtransu-zero},
$\overline{\ee}\cdot{\bf q}={\bf q}\cdot{\bf q}=\norm{{\bf q}}^2$.
Since $\overline{\ee}\cdot{\bf q}\in\{-\infty,0,+\infty\}$, this quantity must be
$0$, implying ${\bf q}={\bf 0}$.
Note that this matrix ${\bf V}$ is column-orthogonal, thereby
satisfying the additional
requirement stated at the end of the proposition.
\underline{(\ref{pr:icon-equiv:c}) $\Rightarrow$ (\ref{pr:icon-equiv:a})}:
Suppose $\overline{\ee}={\bf V}{\bar{\bm{\omega}}}$ for some ${\bf V}\in{\mathbb{R}}^{n\times k}$, $k\geq 0$.
Then
\[
{\bf V}{\bar{\bm{\omega}}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf V}{\bar{\bm{\omega}}}
=
2 {\bf V}{\bar{\bm{\omega}}}
=
{\bf V} (2{\bf I}) {\bar{\bm{\omega}}}
=
{\bf V}{\bar{\bm{\omega}}}
\]
where ${\bf I}$ is the $k\times k$ identity matrix.
The first equality is by
Proposition~\ref{pr:i:5}(\ref{pr:i:5bb}).
The second uses
Proposition~\ref{pr:h:4}(\ref{pr:h:4d},\ref{pr:h:4g}).
And the third is by Proposition~\ref{pr:g:1} since $2{\bf I}$ is
positive upper triangular.
Therefore, $\overline{\ee}$ is an icon.
\qed
We saw in Corollary~\ref{cor:h:1} and as re-stated in \eqref{eq:i:6}
that every point $\overline{\xx}\in\extspac{n}$ can be represented as
${\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$.
As a result of Proposition~\ref{pr:icon-equiv}, ${\bf V}{\bar{\bm{\omega}}}$ is an icon,
which means that $\overline{\xx}$ can be written $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for
some icon $\overline{\ee}\in{{\cal E}_n}$ and some finite vector ${\bf q}\in{\R^n}$.
In other words, $\overline{\xx}$ has been effectively decomposed into an
\emph{iconic part}, $\overline{\ee}$, and a \emph{finite part}, ${\bf q}$.
This decomposition will be used very heavily in later sections.
Note that $\overline{\ee}$, the iconic part, is uniquely determined by $\overline{\xx}$,
but the finite part is not, in general.
The next proposition summarizes this and other properties:
\begin{proposition} \label{pr:i:8}
~
\begin{enumerate}
\item \label{pr:i:8b}
The only icon in ${\R^n}$ is ${\bf 0}$; that is, ${{\cal E}_n}\cap{\R^n}=\{{\bf 0}\}$.
\item \label{pr:i:8c}
Let $\overline{\xx}\in\extspac{n}$.
Then
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$
(so that $\extspac{n} = {{\cal E}_n}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\R^n}$).
Furthermore, $\overline{\ee}$ is uniquely determined by $\overline{\xx}$; that is,
if it also holds that $\overline{\xx}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$ for some
$\overline{\ee}'\in{{\cal E}_n}$ and ${\bf q}'\in{\R^n}$, then $\overline{\ee}=\overline{\ee}'$.
\item \label{pr:i:8d}
If $\overline{\ee}\in{{\cal E}_n}$, then $\lambda\overline{\ee}=\overline{\ee}$ for all $\lambda>0$.
\item \label{pr:i:8-leftsum}
The set of icons is closed under leftward addition; that is,
if $\overline{\ee}$ and $\overline{\dd}$ are in ${{\cal E}_n}$, then
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd}$ is also in ${{\cal E}_n}$.
\item \label{pr:i:8e}
The set of all icons, ${{\cal E}_n}$, is topologically closed
in $\extspac{n}$
(and therefore compact).
\end{enumerate}
\end{proposition}
\proof ~
Part~(\ref{pr:i:8b}):
Using Proposition~\ref{pr:i:3},
this is immediate from the characterization of icons given in
Proposition~\ref{pr:icon-equiv}(\ref{pr:icon-equiv:b}).
Part~(\ref{pr:i:8c}):
As noted already, that $\overline{\xx}$ can be written in the stated form follows
directly from
Corollary~\ref{cor:h:1} and \eqref{eq:i:6},
combined with the characterization of icons given in
Proposition~\ref{pr:icon-equiv}(\ref{pr:icon-equiv:c}).
To show uniqueness of $\overline{\ee}$, suppose
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$ for some
$\overline{\ee},\overline{\ee}'\in{{\cal E}_n}$ and ${\bf q},{\bf q}'\in{\R^n}$.
Then for all ${\bf u}\in{\R^n}$,
$ \overline{\ee}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}\cdot{\bf u} = \overline{\ee}'\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}'\cdot{\bf u}$.
Since ${\bf q}\cdot{\bf u}$ and ${\bf q}'\cdot{\bf u}$ are both in ${\mathbb{R}}$,
it follows that $\overline{\ee}\cdot{\bf u}=+\infty$ if and only if
$\overline{\ee}'\cdot{\bf u}=+\infty$,
and likewise,
$\overline{\ee}\cdot{\bf u}=-\infty$ if and only if
$\overline{\ee}'\cdot{\bf u}=-\infty$.
Since,
by Proposition~\ref{pr:icon-equiv}(\ref{pr:icon-equiv:b}),
$\overline{\ee}\cdot{\bf u}$ and $\overline{\ee}'\cdot{\bf u}$ are both in
$\{-\infty,0,+\infty\}$,
this implies
$\overline{\ee}\cdot{\bf u}=\overline{\ee}'\cdot{\bf u}$.
Therefore, $\overline{\ee}=\overline{\ee}'$,
since this holds for all ${\bf u}\in{\R^n}$
(by Proposition~\ref{pr:i:4}).
Part~(\ref{pr:i:8d}):
Let $\overline{\ee}\in{{\cal E}_n}$ and $\lambda>0$.
Then for all ${\bf u}\in{\R^n}$,
$(\lambda\overline{\ee})\cdot{\bf u} = \lambda(\overline{\ee}\cdot{\bf u}) = \overline{\ee}\cdot{\bf u}$
by Proposition~\ref{pr:i:2}
and since
$\overline{\ee}\cdot{\bf u}\in\{-\infty,0,+\infty\}$.
Therefore, $\lambda\overline{\ee}=\overline{\ee}$
(by Proposition~\ref{pr:i:4}).
Part~(\ref{pr:i:8-leftsum}):
Let $\overline{\ee}$ and $\overline{\dd}$ be in ${{\cal E}_n}$,
and let ${\bf u}\in{\R^n}$.
Then $\overline{\ee}\cdot{\bf u}$ and $\overline{\dd}\cdot{\bf u}$ are both
in $\{-\infty,0,+\infty\}$, by
Proposition~\ref{pr:icon-equiv}.
Combining with
Proposition~\ref{pr:i:6},
this implies that
$(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd})\cdot{\bf u}=\overline{\ee}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd}\cdot{\bf u}$
is in $\{-\infty,0,+\infty\}$ as well.
Therefore, $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd}\in{{\cal E}_n}$, again by
Proposition~\ref{pr:icon-equiv}.
Part~(\ref{pr:i:8e}):
Suppose $\overline{\xx}\in\extspac{n}$ is not in ${{\cal E}_n}$.
Then by Proposition~\ref{pr:icon-equiv}, $\overline{\xx}\cdot{\bf u}\not\in\{-\infty,0,+\infty\}$
for some ${\bf u}\in{\R^n}$, that is,
$\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}\setminus\{0\}$.
Let $\epsilon=|\overline{\xx}\cdot{\bf u}|/2>0$, and let
\[ U = \{ \overline{{\bf z}}\in\extspac{n} : |\overline{{\bf z}}\cdot{\bf u} - \overline{\xx}\cdot{\bf u}| < \epsilon \}. \]
Then $U$ is open, being a standard basis element, and $\overline{\xx}\in U$.
On the other hand, if
$\overline{{\bf z}}\in U$ then it is impossible, because of the constraint defining
$U$, that $\overline{{\bf z}}\cdot{\bf u}\in\{-\infty,0,+\infty\}$.
Thus, $U$ is a neighborhood of $\overline{\xx}$ that is disjoint from
${{\cal E}_n}$, which means $\overline{\xx}$ is not in ${\overline{\cal E}_n}$,
the closure of ${{\cal E}_n}$.
This proves ${\overline{\cal E}_n}\subseteq{{\cal E}_n}$, so ${{\cal E}_n}$ is closed.
\qed
Proposition~\ref{pr:i:8}(\ref{pr:i:8c})
shows that
every astral point has a uniquely determined iconic part.
As a result, astral space itself can be naturally partitioned into
disjoint sets called \emph{galaxies}
consisting of all points with a common iconic part, that
is, into sets
\[
\galax{\overline{\ee}}
= \braces{ \overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q} : {\bf q}\in{\R^n} }
= \overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\R^n}
\]
for each $\overline{\ee}\in{{\cal E}_n}$.
By Proposition~\ref{pr:i:8}(\ref{pr:i:8c}),
every astral point belongs to exactly one galaxy.
Every galaxy $\galax{\overline{\ee}}$ has its icon $\overline{\ee}$ at its center,
with ${\R^n}$ being exactly the galaxy $\galax{{\bf 0}}$ with icon ${\bf 0}$.
In algebraic terms, each galaxy $\galax{\overline{\ee}}$ is a
commutative subgroup under leftward addition, which acts like
vector addition on elements of the same galaxy; this is because
if $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q},{\bf q}'\in{\R^n}$ then
$(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}') = \overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax({\bf q}+{\bf q}')$
using Proposition~\ref{pr:i:7}.
We briefly explore some of the other properties of galaxies, their
closures and relationships to one another, and how these relate to the
overall topological structure of astral space.
As we show next,
the closure of a galaxy $\galax{\ebar}$ includes exactly those
astral points that can be written in the form $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$, for
some $\overline{{\bf z}}\in\extspac{n}$.
(For readability, we write the closure of galaxy
$\galax{\ebar}$ as $\galcl{\ebar}$, rather than
$\overline{\galax{\ebar}}$.)
\begin{proposition} \label{pr:galaxy-closure}
Let $\overline{\ee}\in{{\cal E}_n}$ be an icon.
Then the closure of galaxy $\galax{\ebar}$ is
\[
\galcl{\ebar}
= \braces{ \overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}} : \overline{{\bf z}}\in\extspac{n} }
= \overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \extspac{n}.
\]
\end{proposition}
\proof
Let $\overline{\xx}\in\extspac{n}$.
We prove $\overline{\xx}\in\galcl{\ebar}$
if and only if $\overline{\xx}\in\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\extspac{n}$.
Suppose $\overline{\xx}\in\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\extspac{n}$, meaning
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ for some $\overline{{\bf z}}\in\extspac{n}$.
Then by
Theorem~\ref{thm:i:1}(\ref{thm:i:1d}),
there exists a sequence $\seq{{\bf z}_t}$ in ${\R^n}$ that converges to
$\overline{{\bf z}}$.
By Proposition~\ref{pr:aff-cont}
(applied to the map $\overline{{\bf y}}\mapsto\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$),
it follows that
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}_t\rightarrow\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}=\overline{\xx}$.
Since each $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}_t$ is in $\galax{\ebar}$, this proves that
$\overline{\xx}\in\galcl{\ebar}$.
Conversely,
suppose now that $\overline{\xx}\not\in\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\extspac{n}$.
Then, in particular, $\overline{\xx}\neq\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$.
Therefore, by Proposition~\ref{pr:i:4},
there exists ${\bf u}\in{\R^n}$ such that
\begin{equation} \label{eqn:pr:galaxy-closure:1}
\overline{\xx}\cdot{\bf u}
\neq
(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})\cdot{\bf u}
=
\overline{\ee}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}\cdot{\bf u}.
\end{equation}
Since $\overline{\ee}$ is an icon, $\overline{\ee}\cdot{\bf u}$ must be in
$\{-\infty,0,+\infty\}$
(by Proposition~\ref{pr:icon-equiv}).
Thus, \eqref{eqn:pr:galaxy-closure:1} can only hold if
$\overline{\ee}\cdot{\bf u}\in\{-\infty,+\infty\}$
and
$\overline{\ee}\cdot{\bf u}\neq\overline{\xx}\cdot{\bf u}$.
Without loss of generality, we assume $\overline{\ee}\cdot{\bf u}=+\infty$,
since otherwise we can replace ${\bf u}$ with $-{\bf u}$.
Then $\overline{\xx}\cdot{\bf u}\neq+\infty$.
Let $\beta\in{\mathbb{R}}$ be such that $\overline{\xx}\cdot{\bf u}<\beta$,
and let $U=\{\overline{{\bf z}}\in\extspac{n} : \overline{{\bf z}}\cdot{\bf u} < \beta\}$.
Then $U$ is open, being a standard basis element,
and includes $\overline{\xx}$.
On the other hand, $U$ is disjoint from $\galax{\ebar}$ since if
$\overline{{\bf z}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y}$ is any point in $\galax{\ebar}$, for some ${\bf y}\in{\R^n}$,
then $\overline{{\bf z}}\cdot{\bf u}=\overline{\ee}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y}\cdot{\bf u}=+\infty$.
This proves that $\overline{\xx}\not\in\galcl{\ebar}$.
\qed
For any icons $\overline{\ee},\,\overline{\ee}'\in{{\cal E}_n}$,
the galaxies
$\galax{\ebar}$ and $\galax{\ebar'}$ are disjoint, unless $\overline{\ee}=\overline{\ee}'$,
as noted already.
Nevertheless, the \emph{closures} of these galaxies might not be disjoint.
Indeed,
the next proposition shows that
if $\galcl{\ebar}$ and $\galcl{\ebar'}$ intersect at even a single point, then one
must fully contain the other.
Furthermore, $\galcl{\ebar}$ is entirely included in $\galcl{\ebar'}$
if and only if $\overline{\ee}'$ is a \emph{prefix} of $\overline{\ee}$, meaning
$\overline{\ee}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd}$ for some icon $\overline{\dd}\in{{\cal E}_n}$.
This can be re-stated equivalently
in terms of canonical representations
since
if $\overline{\ee}={\bf V}{\bar{\bm{\omega}}}$ and $\overline{\ee}'={\bf V}'{\bar{\bm{\omega}}}$ where ${\bf V}\in{\mathbb{R}}^{n\times k}$
and ${\bf V}'\in{\mathbb{R}}^{n\times k'}$ are
column-orthogonal, then
it can be shown (using Lemma~\ref{lem:uniq-ortho-prefix})
that $\overline{\ee}'$ is a prefix of $\overline{\ee}$ if and only if
$k'\leq k$ and
the first $k'$ columns of ${\bf V}$ are identical to ${\bf V}'$
(so that the columns of ${\bf V}'$ are truly a prefix of the columns of
${\bf V}$).
\begin{proposition} \label{pr:galaxy-closure-inclusion}
Let $\overline{\ee},\,\overline{\ee}'\in{{\cal E}_n}$ be icons.
Then the following hold:
\begin{enumerate}
\item \label{pr:galaxy-closure-inclusion:a}
$\galcl{\ebar}\subseteq\galcl{\ebar'}$
if and only if
$\overline{\ee}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd}$ for some $\overline{\dd}\in{{\cal E}_n}$.
\item \label{pr:galaxy-closure-inclusion:b}
If $\galcl{\ebar}\cap\galcl{\ebar'}\neq\emptyset$
then
either
$\galcl{\ebar}\subseteq\galcl{\ebar'}$
or
$\galcl{\ebar'}\subseteq\galcl{\ebar}$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:galaxy-closure-inclusion:a}):
Suppose first that
$\overline{\ee}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd}$ for some $\overline{\dd}\in{{\cal E}_n}$.
Let $\overline{\xx}\in\galcl{\ebar}$.
Then,
by Proposition~\ref{pr:galaxy-closure},
there exists $\overline{{\bf z}}\in\extspac{n}$ for which
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}})$.
Thus,
$\overline{\xx}\in\galcl{\ebar'}$, again
by Proposition~\ref{pr:galaxy-closure}.
Conversely, suppose now that
$\galcl{\ebar}\subseteq\galcl{\ebar'}$.
Then, in particular, $\overline{\ee}\in\galcl{\ebar'}$, so
$\overline{\ee}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ for some $\overline{{\bf z}}\in\extspac{n}$,
by Proposition~\ref{pr:galaxy-closure}.
By Proposition~\ref{pr:i:8}(\ref{pr:i:8c}),
$\overline{{\bf z}}=\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some $\overline{\dd}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$,
so
$\overline{\ee}=(\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$.
Since the iconic part of this point is unique, it follows that
$\overline{\ee}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\dd}$
(by Proposition~\ref{pr:i:8}(\ref{pr:i:8c},\ref{pr:i:8-leftsum})).
Part~(\ref{pr:galaxy-closure-inclusion:b}):
Suppose there exists a point $\overline{\xx}$ in $\galcl{\ebar}\cap\galcl{\ebar'}$.
Then by Proposition~\ref{pr:galaxy-closure},
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'$,
for some $\overline{{\bf z}},\,\overline{{\bf z}}'\in\extspac{n}$.
By Proposition~\ref{pr:icon-equiv},
$\overline{\ee}={\bf V}{\bar{\bm{\omega}}}$ and $\overline{\ee}'={\bf V}'{\bar{\bm{\omega}}}$
for some column-orthogonal matrices
${\bf V}\in{\mathbb{R}}^{n\times k}$ and
${\bf V}'\in{\mathbb{R}}^{n\times k'}$.
Without loss of generality, we assume $k'\leq k$.
We can write out the columns of ${\bf V}$ as
${\bf V}=[{\bf v}_1,\ldots,{\bf v}_k]$,
for some ${\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$.
Also, let
${\bf V}_0=[{\bf v}_1,\ldots,{\bf v}_{k'}]$
be a matrix consisting of the the initial $k'$ columns of ${\bf V}$,
and let
${\bf V}_1=[{\bf v}_{k'+1},\ldots,{\bf v}_k]$
be a matrix consisting of the remaining columns.
Then ${\bf V}{\bar{\bm{\omega}}} = {\bf V}_0{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf V}_1{\bar{\bm{\omega}}}$,
using \eqref{eqn:mat-omm-defn}.
Thus, making substitutions,
\[
\overline{\xx}
=
{\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'
=
{\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}
=
{\bf V}_0{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf V}_1{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}.
\]
Since ${\bf V}'$ and ${\bf V}_0$ are each column-orthogonal matrices with $k'$
columns,
Lemma~\ref{lem:uniq-ortho-prefix}
implies they are equal.
Therefore,
\[
\overline{\ee}
=
{\bf V}{\bar{\bm{\omega}}}
=
{\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf V}_1{\bar{\bm{\omega}}}
=
\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf V}_1{\bar{\bm{\omega}}},
\]
so
$\galcl{\ebar}\subseteq\galcl{\ebar'}$
by part~(\ref{pr:galaxy-closure-inclusion:a}).
\qed
As a result, we can arrange the galaxies in a directed rooted tree
capturing these inclusion relationships.
The vertices of the tree consist exactly of all icons in ${{\cal E}_n}$
(corresponding to galaxies),
with ${\bf 0}$ as the root.
For all icons $\overline{\ee}\in{{\cal E}_n}$ and for all ${\bf v}\in{\R^n}$,
an edge is directed from $\overline{\ee}$ to
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf v}}$ (which is also an icon,
by Proposition~\ref{pr:i:8}(\ref{pr:i:8-leftsum})), unless this would
result in a self-loop, that is, unless
$\overline{\ee}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf v}}$.
That the resulting structure is a tree rather than a directed acyclic
graph can be proved using Lemma~\ref{lem:uniq-ortho-prefix}.
Equivalently, the tree can be formulated in terms of canonical
representations, so that, using
Proposition~\ref{pr:icon-equiv},
for column-orthogonal matrices ${\bf V}\in{\mathbb{R}}^{n\times k}$
and ${\bf V}'\in{\mathbb{R}}^{n\times (k+1)}$,
an edge is directed from ${\bf V}{\bar{\bm{\omega}}}$ to ${\bf V}'{\bar{\bm{\omega}}}$ if and only if
the first $k$ columns of ${\bf V}'$ are identical to ${\bf V}$.
Thus, the children of ${\bf 0}$, the root, are exactly the
astrons.
Their children are the icons of astral rank~2, and so on.
In general,
the depth (distance from the root) of every icon $\overline{\ee}$ is exactly
equal to its astral rank, so
the height of the tree is $n$.
Further, the tree captures all galactic inclusions since
a path exists from $\overline{\ee}$ to $\overline{\ee}'$ if and only if $\overline{\ee}$
is a prefix of $\overline{\ee}'$, that is,
if and only if
$\galcl{\ebar}\supseteq\galcl{\ebar'}$
(by
Proposition~\ref{pr:galaxy-closure-inclusion}(\ref{pr:galaxy-closure-inclusion:a})).
Finally, we consider the topology of galaxies and their closures.
The next theorem shows that the galaxy $\galax{\ebar}$ of an
icon $\overline{\ee}$ of astral rank $k$
is homeomorphic to $(n-k)$-dimensional Euclidean space,
${\mathbb{R}}^{n-k}$, and
furthermore, the closure of that galaxy is
homeomorphic to
$(n-k)$-dimensional astral space,
$\extspac{n-k}$.
In other words, all galaxies and their closures
are topological copies of
lower-dimensional Euclidean spaces and astral spaces
(respectively).
\begin{theorem} \label{thm:galaxies-homeo}
Let $\overline{\ee}\in{{\cal E}_n}$ be an icon of astral rank $k$.
Then the following hold:
\begin{enumerate}
\item \label{thm:galaxies-homeo:a}
$\galcl{\ebar}$ is homeomorphic to $\extspac{n-k}$.
\item \label{thm:galaxies-homeo:b}
$\galax{\ebar}$ is homeomorphic to ${\mathbb{R}}^{n-k}$.
\end{enumerate}
\end{theorem}
\proof
~
Part~(\ref{thm:galaxies-homeo:a}):
By Propositions~\ref{pr:h:3} and~\ref{pr:icon-equiv},
$\overline{\ee}={\bf V}{\bar{\bm{\omega}}}$, for
some column-orthogonal matrix ${\bf V}\in{\mathbb{R}}^{n\times k}$.
Let ${\bf w}_1,\ldots,{\bf w}_{n-k}\in{\R^n}$ be an orthonormal basis for the
$(n-k)$-dimensional
linear subspace $L$ orthogonal to the columns of ${\bf V}$, as given in
\eqref{eqn:subspace-orth-cols-v},
and let ${\bf W}=[{\bf w}_1,\ldots,{\bf w}_{n-k}]$.
Note that
\begin{equation} \label{eqn:thm:galaxies-homeo:3}
\trans{{\bf W}}{\bf W}={\bf I}
\end{equation}
since ${\bf W}$ is column-orthogonal
(where ${\bf I}$ is the $(n-k)\times(n-k)$ identity matrix).
Also,
\begin{equation} \label{eqn:thm:galaxies-homeo:4}
\trans{{\bf W}}\overline{\ee}=\trans{{\bf W}}{\bf V}{\bar{\bm{\omega}}}={\bf 0}
\end{equation}
since each ${\bf w}_j\in L$, so
$\trans{{\bf W}}{\bf V}=\zeromat{(n-k)}{k}$.
We explicitly construct a homeomorphism
$\gamma: \galcl{\ebar}\rightarrow\extspac{n-k}$
by the rule
\begin{equation} \label{eqn:thm:galaxies-homeo:hogal-rule}
\gamma(\overline{\xx}) = \trans{{\bf W}} \overline{\xx}
\end{equation}
for $\overline{\xx}\in\galcl{\ebar}$.
This function is a linear map, and therefore
continuous, by
Proposition~\ref{pr:aff-cont}.
Next, we define the function
$\gamma^{-1}: \extspac{n-k}\rightarrow\galcl{\ebar}$
by the rule
\begin{equation} \label{eqn:thm:galaxies-homeo:hogalinv-rule}
\gamma^{-1}(\overline{{\bf y}})=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf W}\overline{{\bf y}}
\end{equation}
for $\overline{{\bf y}}\in\extspac{n-k}$.
That $\gamma^{-1}$ maps into $\galcl{\ebar}$ follows from
Proposition~\ref{pr:galaxy-closure}.
Furthermore, $\gamma^{-1}$ is continuous since it is an affine map,
again by
Proposition~\ref{pr:aff-cont}.
For the remainder of the proof,
we show that $\gamma$ and $\gamma^{-1}$ are functional inverses of each other,
thereby proving they are continuous bijections that therefore define a
homeomorphism.
Let $\overline{{\bf y}}\in\extspac{n-k}$.
Then
\[
\gamma(\gamma^{-1}(\overline{{\bf y}}))
=
\trans{{\bf W}}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf W}\overline{{\bf y}})
=
\trans{{\bf W}}\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \trans{{\bf W}}{\bf W}\overline{{\bf y}}
=
\overline{{\bf y}}
\]
with the last equality following from
\eqref{eqn:thm:galaxies-homeo:3}
and
\eqref{eqn:thm:galaxies-homeo:4}.
Thus, $\gamma^{-1}$ is a right inverse of $\gamma$.
It remains to show it is also a left inverse.
Let $\overline{\xx}\in\galcl{\ebar}$,
implying $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ for some $\overline{{\bf z}}\in\extspac{n}$
by Proposition~\ref{pr:galaxy-closure}.
Let
$\overline{\xx}' = \gamma^{-1}(\gamma(\overline{\xx}))$.
We aim to show $\overline{\xx}'=\overline{\xx}$.
Plugging in definitions yields
\begin{equation} \label{eqn:thm:galaxies-homeo:1}
\overline{\xx}'
=
\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf W}\trans{{\bf W}} \overline{\xx}
=
\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf W}\trans{{\bf W}} \overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf W}\trans{{\bf W}} \overline{{\bf z}}
=
\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf W}\trans{{\bf W}} \overline{{\bf z}},
\end{equation}
with the last equality from
\eqref{eqn:thm:galaxies-homeo:4}.
Let ${\bf u}\in{\R^n}$.
If $\overline{\ee}\cdot{\bf u}\in\{-\infty,+\infty\}$,
then
$\overline{\xx}'\cdot{\bf u}=\overline{\ee}\cdot{\bf u}=\overline{\xx}\cdot{\bf u}$.
Otherwise, $\overline{\ee}\cdot{\bf u}=0$ (by Proposition~\ref{pr:icon-equiv},
since $\overline{\ee}\in{{\cal E}_n}$), so
$\trans{{\bf V}} {\bf u} = {\bf 0}$, by Proposition~\ref{pr:vtransu-zero}.
That is, ${\bf u}\in L$.
Since the columns of ${\bf W}$ are a basis for $L$, there exists
${\bf b}\in{\mathbb{R}}^{n-k}$ such that
${\bf u}={\bf W} {\bf b}$.
This implies
\begin{equation} \label{eqn:thm:galaxies-homeo:2}
{\bf W}\trans{{\bf W}} {\bf u}
=
{\bf W}\trans{{\bf W}} {\bf W} {\bf b}
=
{\bf W} {\bf b}
=
{\bf u},
\end{equation}
using
\eqref{eqn:thm:galaxies-homeo:3}
for the second equality.
Therefore,
\begin{eqnarray*}
\overline{\xx}'\cdot{\bf u}
&=&
\overline{\ee}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf W}\trans{{\bf W}} \overline{{\bf z}})\cdot{\bf u}
\\
&=&
\overline{\ee}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot({\bf W}\trans{{\bf W}} {\bf u})
\\
&=&
\overline{\ee}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot{\bf u}
\\
&=&
\overline{\xx}\cdot{\bf u}.
\end{eqnarray*}
The first equality is by
\eqref{eqn:thm:galaxies-homeo:1}.
The second is by
Proposition~\ref{pr:mat-mult-def}
(and since ${\bf W}\trans{{\bf W}}$ is symmetric).
The third is by
\eqref{eqn:thm:galaxies-homeo:2}.
Thus, in all cases, we have argued that
$\overline{\xx}'\cdot{\bf u}=\overline{\xx}\cdot{\bf u}$, for all ${\bf u}\in{\R^n}$.
Therefore, $\overline{\xx}'=\overline{\xx}$
(by Proposition~\ref{pr:i:4}),
proving that
$\gamma^{-1}$ is also a left inverse of $\gamma$, and completing the
proof.
Part~(\ref{thm:galaxies-homeo:b}):
We re-define
$\gamma$ and $\gamma^{-1}$ according to the same rules given in
\eqref{eqn:thm:galaxies-homeo:hogal-rule}
and
\eqref{eqn:thm:galaxies-homeo:hogalinv-rule}, but now with restricted
domain and range so that
$\gamma: \galax{\ebar}\rightarrow{\mathbb{R}}^{n-k}$
and
$\gamma^{-1}: {\mathbb{R}}^{n-k}\rightarrow\galax{\ebar}$.
If ${\bf y}\in{\mathbb{R}}^{n-k}$, then clearly
$\gamma^{-1}({\bf y})$ is in $\galax{\ebar}$, from
\eqref{eqn:thm:galaxies-homeo:hogalinv-rule}.
And
if $\overline{\xx}\in\galax{\ebar}$ then $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}$ for some
${\bf z}\in{\R^n}$, so
$\gamma(\overline{\xx})=\trans{{\bf W}}\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\trans{{\bf W}}{\bf z}=\trans{{\bf W}}{\bf z}$,
by
\eqref{eqn:thm:galaxies-homeo:4};
thus,
$\gamma(\overline{\xx})$ is indeed in ${\mathbb{R}}^{n-k}$.
Even with restricted domain and range,
the arguments used in
part~(\ref{thm:galaxies-homeo:a}) can be applied, proving that
$\gamma$ and $\gamma^{-1}$ are continuous and functional inverses of one
another;
therefore, $\gamma$ is a homeomorphism.
\qed
\subsection{Dominant directions}
\label{sec:dom:dir}
We next explore another way of decomposing astral points.
We have seen that every infinite point $\overline{\xx}\in\extspac{n}$ can be
regarded as the
destination of a trajectory heading to infinity in directions
ordered by dominance.
By considering sequences converging to $\overline{\xx}$, we can analyze this
structure, first identifying the most dominant direction, then
projecting to an effectively lower dimensional space and repeating to
find the next dominant direction, and so on, in this way, peeling the
point apart, one direction at a time.
As a first step in such an analysis, we focus on the strongest of these directions.
For any infinite point $\overline{\xx}\in\extspac{n}$,
we say that a direction ${\bf v}\in{\R^n}$, where $\norm{{\bf v}}=1$, is a
\emph{dominant direction} of $\overline{\xx}$ if
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$
for some $\overline{{\bf z}}\in\extspac{n}$.
As we show next,
if $\seq{{\bf x}_t}$ is any sequence in ${\R^n}$ converging to $\overline{\xx}$, then
the directions in which the vectors ${\bf x}_t$ point must converge to
$\overline{\xx}$'s dominant direction.
Moreover, every infinite astral point has exactly one dominant
direction.
\begin{theorem} \label{thm:dom-dir}
Let $\overline{\xx}\in\extspac{n}\setminus{\R^n}$, and let ${\bf v}\in{\R^n}$
with $\norm{{\bf v}}=1$.
Also,
let $\seq{{\bf x}_t}$ and $\seq{{\bf d}_t}$ be
sequences in ${\R^n}$ such that ${\bf x}_t\rightarrow\overline{\xx}$,
and ${\bf d}_t={\bf x}_t / \norm{{\bf x}_t}$
whenever ${\bf x}_t\neq {\bf 0}$
(or equivalently,
${\bf x}_t = {\bf d}_t \norm{{\bf x}_t}$ for all $t$).
Then the following are equivalent:
\begin{enumerate}
\item \label{thm:dom-dir:a}
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ for some $\overline{{\bf z}}\in\extspac{n}$.
That is, ${\bf v}$ is a dominant direction of $\overline{\xx}$.
\item \label{thm:dom-dir:b}
For all ${\bf u}\in{\R^n}$,
if ${\bf v}\cdot{\bf u}>0$ then $\overline{\xx}\cdot{\bf u}=+\infty$,
and
if ${\bf v}\cdot{\bf u}<0$ then $\overline{\xx}\cdot{\bf u}=-\infty$.
\item \label{thm:dom-dir:c}
${\bf d}_t \rightarrow {\bf v}$.
\end{enumerate}
Furthermore, every point $\overline{\xx}\in\extspac{n}\setminus{\R^n}$ has a
unique dominant direction.
\end{theorem}
\proof
~
\underline{Existence}:
Being in $\extspac{n}$, $\overline{\xx}$ must have the form given in
Corollary~\ref{cor:h:1}.
In the notation of that corollary, $k\geq 1$ (since
$\overline{\xx}\not\in{\R^n}$),
and
${\bf v}_1$ must be a dominant direction (assuming, without loss of
generality, that $\norm{{\bf v}_1}=1$).
\underline{Uniqueness}:
Suppose both ${\bf v}$ and ${\bf v}'$ are dominant directions of
$\overline{\xx}$.
Then $\norm{{\bf v}}=\norm{{\bf v}'}=1$, and
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}=\limray{{\bf v}'}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'$,
for some $\overline{{\bf z}},\,\overline{{\bf z}}'\in\extspac{n}$.
That is,
${\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}={\bf V}'{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'$,
where ${\bf V}=[{\bf v}]$ and ${\bf V}'=[{\bf v}']$.
Since these matrices are each column-orthogonal with a single column,
we can apply Lemma~\ref{lem:uniq-ortho-prefix},
yielding ${\bf V}={\bf V}'$.
Thus, ${\bf v}={\bf v}'$.
\underline{(\ref{thm:dom-dir:b}) $\Rightarrow$ (\ref{thm:dom-dir:a})}:
Suppose (\ref{thm:dom-dir:b}) holds.
Then
$\overline{\xx}\cdot{\bf u}=\limray{{\bf v}}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u}$
for all ${\bf u}\in{\R^n}$.
This is because if ${\bf v}\cdot{\bf u}= 0$, then
$\limray{{\bf v}}\cdot{\bf u}=0$.
And if ${\bf v}\cdot{\bf u}\neq 0$, then
$\overline{\xx}\cdot{\bf u}=\limray{{\bf v}}\cdot{\bf u}$, by assumption.
Therefore,
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$,
by Proposition~\ref{pr:i:4},
proving (\ref{thm:dom-dir:a}).
Before continuing, we note that because $\overline{\xx}\not\in{\R^n}$, we must
have $\overline{\xx}\cdot{\bf u}\in\{-\infty,+\infty\}$ for some ${\bf u}\in{\R^n}$
(by Proposition~\ref{pr:i:3}), implying
$\norm{{\bf x}_t}\rightarrow+\infty$
(since ${\bf x}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$,
by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
\underline{(\ref{thm:dom-dir:c}) $\Rightarrow$ (\ref{thm:dom-dir:b})}.
Suppose (\ref{thm:dom-dir:c}) holds.
Let ${\bf u}\in{\R^n}$, and suppose ${\bf v}\cdot{\bf u}>0$.
Then
${\bf x}_t\cdot{\bf u} = ({\bf d}_t\cdot{\bf u}) \norm{{\bf x}_t} \rightarrow +\infty$
since ${\bf d}_t\cdot{\bf u}\rightarrow {\bf v}\cdot{\bf u} >0$
and $\norm{{\bf x}_t}\rightarrow+\infty$.
Therefore, $\overline{\xx}\cdot{\bf u}=+\infty$,
by Theorem~\ref{thm:i:1}(\ref{thm:i:1c}),
since ${\bf x}_t\rightarrow\overline{\xx}$.
By a symmetric argument, if ${\bf v}\cdot{\bf u}<0$ then
$\overline{\xx}\cdot{\bf u}=-\infty$.
\underline{(\ref{thm:dom-dir:a}) $\Rightarrow$ (\ref{thm:dom-dir:c})}.
Suppose to the contrary that
(\ref{thm:dom-dir:a}) holds, but (\ref{thm:dom-dir:c}) does not.
Since $\norm{{\bf x}_t}\rightarrow +\infty$, we can have ${\bf x}_t={\bf 0}$ for
at most finitely many values of $t$.
Also, because we assume ${\bf d}_t\not\rightarrow{\bf v}$,
there must exist a neighborhood $U\subseteq{\R^n}$
of ${\bf v}$ that excludes infinitely
many of the ${\bf d}_t$'s.
By discarding all of the sequence elements indexed by any $t$
for which either
${\bf x}_t={\bf 0}$ or ${\bf d}_t\in U$, we
can assume henceforth that ${\bf x}_t\neq{\bf 0}$
and ${\bf d}_t\not\in U$ for all $t$.
As a result, the ${\bf d}_t$'s are all on the unit sphere in ${\R^n}$, which is
compact, implying that their sequence must have a convergent
subsequence.
By again discarding all of the other sequence elements,
we can assume this subsequence is the entire sequence.
What remains of the corresponding sequence of ${\bf x}_t$'s still
converges to $\overline{\xx}$.
Thus, the remaining ${\bf d}_t$'s converge, and so, as just argued
(in proving
(\ref{thm:dom-dir:c})
$\Rightarrow$
(\ref{thm:dom-dir:b})
$\Rightarrow$
(\ref{thm:dom-dir:a})),
their limit is $\overline{\xx}$'s unique dominant direction, which is ${\bf v}$ by
assumption.
But this is a contradiction since all the ${\bf d}_t$'s in the sequence
lie outside $U$, a neighborhood of ${\bf v}$.
\qed
\subsection{Projections}
\label{subsec:projections}
As we have just seen, every infinite point $\overline{\xx}\in\extspac{n}$ has a
dominant direction ${\bf v}$ which fully determines $\overline{\xx}\cdot{\bf u}$
in every direction ${\bf u}$ that is not orthogonal to ${\bf v}$.
What about in the directions that \emph{are} orthogonal to ${\bf v}$, where
${\bf v}\cdot{\bf u}=0$?
To answer this, consider a sequence $\seq{{\bf x}_t}$ in ${\R^n}$ that
converges to $\overline{\xx}$.
Every element ${\bf x}_t$ can be expressed as
${\bf x}_t=a_t {\bf v} + {\xx_t^\bot}$
where $a_t\in{\mathbb{R}}$ and ${\xx_t^\bot}\in{\R^n}$ is the projection of ${\bf x}_t$
onto the space orthogonal to ${\bf v}$.
In fact, we will see in a moment that the sequence $\seq{{\xx_t^\bot}}$
has a limit $\xbar^{\bot}\in\extspac{n}$, which can be viewed intuitively
as a projection of $\overline{\xx}$ itself.
Since ${\bf v}$ is $\overline{\xx}$'s dominant direction, ${\bf x}_t\cdot{\bf u}$ converges
to $\pm\infty$ for all ${\bf u}\in{\R^n}$ with ${\bf v}\cdot{\bf u}\neq 0$.
But if ${\bf v}\cdot{\bf u}=0$, then ${\bf x}_t\cdot{\bf u}={\xx_t^\bot}\cdot{\bf u}$ so that
in directions perpendicular to ${\bf v}$, the sequences ${\bf x}_t$ and
${\xx_t^\bot}$ are identical, that is,
$\overline{\xx}\cdot{\bf u}=\xbar^{\bot}\cdot{\bf u}$.
Thus, analogous to our decomposition of the sequence
elements ${\bf x}_t$, $\overline{\xx}$ has been effectively decomposed and can
now be written as
$\overline{\xx} = \limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \xbar^{\bot}$,
that is, in terms of its dominant direction ${\bf v}$ and its projection
onto the space orthogonal to ${\bf v}$.
We will use such decompositions extensively later in the manuscript.
Projection is a linear operation which means we can use the results
from Section~\ref{sec:linear-maps} to extend it to astral space.
In more detail, as above, we consider projection onto the space
orthogonal to ${\bf v}\in{\R^n}$, and
when the vector ${\bf v}$ is clear from context, we usually write ${\xx^\bot}\in{\R^n}$
for the projection of a point ${\bf x}\in{\R^n}$ onto that space.
If $\norm{{\bf v}}=1$, ${\xx^\bot}$ can be computed as
\[ {\xx^\bot}={\bf x}-({\bf v}\cdot{\bf x}){\bf v}, \]
ensuring ${\xx^\bot}\cdot{\bf v}=0$.
So ${\xx^\bot}={\bf P}{\bf x}$ where ${\bf P}={\bf I} - {\bf v}\trans{{\bf v}}$,
where ${\bf I}$ is the $n\times n$ identity matrix.
This means the projection operation, being a linear map, can be
extended directly to $\extspac{n}$ simply by defining
$\xbar^{\bot}$,
\emph{the projection of $\overline{\xx}$ perpendicular to ${\bf v}$} to be
$\xbar^{\bot} = {\bf P}\overline{\xx}$.
More generally,
if ${\bf v}\in{\R^n}$ but not necessarily a unit vector, projection is
defined in the same way after normalizing ${\bf v}$, assuming
${\bf v}\neq{\bf 0}$.
If ${\bf v}={\bf 0}$, we define $\xbar^{\bot}=\overline{\xx}$.
Thus, in general, $\xbar^{\bot} = {\bf P}\overline{\xx}$ where
\begin{equation} \label{eqn:gen-proj-mat}
{\bf P}= \left\{ \begin{array}{cl}
{\bf I} & \mbox{if ${\bf v}={\bf 0}$} \\
{\bf I} - \frac{{\bf v}\trans{{\bf v}}}{\norm{{\bf v}}^2}
& \mbox{otherwise.}
\end{array} \right.
\end{equation}
From Propositions~\ref{pr:mat-mult-def} and~\ref{pr:h:4}, this
operation inherits several useful properties:
\begin{proposition} \label{pr:h:5}
Let ${\bf v}\in{\R^n}$.
Let $\overline{\xx}\in\extspac{n}$, and let $\xbar^{\bot}$ denote its projection
onto the linear space orthogonal to ${\bf v}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:h:5a}
$\xbar^{\bot}\cdot{\bf u}=\overline{\xx}\cdot{\uu^\bot}$ for all ${\bf u}\in{\R^n}$.
\item \label{pr:h:5b}
For any sequence $\seq{{\bf x}_t}$ in ${\R^n}$, if ${\bf x}_t\rightarrow\overline{\xx}$ then
${\xx_t^\bot}\rightarrow \xbar^{\bot}$.
\item \label{pr:h:5c}
$(\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}})^\bot = \xbar^{\bot} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \ybar^{\bot}$ for $\overline{{\bf y}}\in\extspac{n}$.
\item \label{pr:h:5e}
$(\limray{{\bf w}})^\bot = \limray{({\bf w}^\bot)}$ for ${\bf w}\in{\R^n}$.
\item \label{pr:h:5f}
$\paren{[{\bf v}_1,\ldots,{\bf v}_k]{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}}^\bot =
[\vperp{1},\ldots,\vperp{k}]{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\qq}^{\bot}$
for ${\bf q},{\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$.
\item \label{pr:h:5g}
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx} = \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}$.
\end{enumerate}
\end{proposition}
\proof
Let ${\bf P}$ be as defined in \eqref{eqn:gen-proj-mat}.
Part~(\ref{pr:h:5a}):
For ${\bf u}\in{\R^n}$,
\[
\xbar^{\bot}\cdot{\bf u}
= ({\bf P}\overline{\xx})\cdot{\bf u}
= \overline{\xx}\cdot(\trans{{\bf P}}{\bf u})
= \overline{\xx}\cdot({{\bf P}}{\bf u}) = \overline{\xx}\cdot{\uu^\bot}
\]
by Proposition~\ref{pr:mat-mult-def} and since ${\bf P}$ is symmetric.
Parts~(\ref{pr:h:5b}),~(\ref{pr:h:5c}) and~(\ref{pr:h:5e})
are similar.
Part~(\ref{pr:h:5f}) follows from~(\ref{pr:h:5c}) and~(\ref{pr:h:5e}).
Part~(\ref{pr:h:5g}):
By Proposition~\ref{pr:i:4}, it suffices to show that
\begin{equation} \label{eq:j:3}
(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})\cdot{\bf u}
=(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot})\cdot{\bf u}
\end{equation}
for all ${\bf u}\in{\R^n}$.
If ${\bf v}\cdot{\bf u}\neq 0$, then both sides of \eqref{eq:j:3} are infinite
and equal to the common value $\limray{{\bf v}}\cdot{\bf u}$.
And if ${\bf v}\cdot{\bf u}=0$ then ${\uu^\bot}={\bf u}$, so
\[
(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})\cdot{\bf u}
=
\overline{\xx}\cdot{\bf u}=\overline{\xx}\cdot{\uu^\bot}=\xbar^{\bot}\cdot{\bf u}
=(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot})\cdot{\bf u}
\]
by part~(\ref{pr:h:5a}).
Thus, \eqref{eq:j:3} holds in this case as well.
\qed
As already noted,
these simple facts provide a decomposition of any infinite point
$\overline{\xx}\in\extspac{n}$ into its dominant direction and its projection
perpendicular to that direction.
This is especially useful as a proof technique since many properties
can be proved by induction on astral rank, which is reduced by
orthogonal projection:
\begin{proposition} \label{pr:h:6}
Let $\overline{\xx}$ be an infinite point in $\extspac{n}$.
Let ${\bf v}$ be its dominant direction, and
let $\xbar^{\bot}$ be $\overline{\xx}$'s projection
perpendicular to ${\bf v}$.
Then $\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}$.
Further, suppose $\overline{\xx}$'s canonical form is
$\overline{\xx}=[{\bf v}_1,\ldots,{\bf v}_k]{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$,
so that its astral rank is $k$.
Then $\xbar^{\bot}$'s canonical form is
$\xbar^{\bot}=[{\bf v}_2,\ldots,{\bf v}_k]{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$,
implying its astral rank is $k-1$.
\end{proposition}
\proof
Since ${\bf v}$ is a dominant direction for $\overline{\xx}$, and by
Proposition~\ref{pr:h:5}(\ref{pr:h:5g}),
\[
\overline{\xx} = \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx} = \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}.
\]
The canonical form of $\xbar^{\bot}$ follows from
Proposition~\ref{pr:h:5}(\ref{pr:h:5f}), and since, in canonical
form, $\overline{\xx}$'s dominant direction is ${\bf v}={\bf v}_1$, which must be
orthogonal to ${\bf q}$ and ${\bf v}_2,\ldots,{\bf v}_{k}$.
\qed
\section{Countable bases}
\label{sec:countable}
\subsection{First Countability}
We show next that the space $\extspac{n}$ is \emph{first countable}, that is,
that it satisfies the first axiom of countability, meaning that
every point $\overline{\xx}\in\extspac{n}$ has a \emph{countable basis},
a countable collection of sets, $B_1,B_2,\ldots$, each a
neighborhood of $\overline{\xx}$, such that every
neighborhood $U$ of $\overline{\xx}$ includes one among the countable
collection, so that $\overline{\xx}\in B_t\subseteq U$ for some $B_t$.
This is a fundamental property of the space, and an extremely useful
one since it will allow us to use sequences to study the space.
In particular, first countability implies that the closure of any set
$S\subseteq\extspac{n}$ consists exactly of those points that are the
limit of sequences of points in $S$.
Moreover,
the fact that $\extspac{n}$ is both compact and first
countable implies that $\extspac{n}$ is \emph{sequentially compact},
meaning that every sequence in $\extspac{n}$ must have a
convergent subsequence.
Throughout this section, we let $\overline{\xx}\in\extspac{n}$ have canonical
form
\begin{equation} \label{eq:h:7}
\overline{\xx}=[{\bf v}_1,\ldots,{\bf v}_k]{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q},
\end{equation}
implying the ${\bf v}_i$'s are orthonormal.
Also, let
\[ L = \{{\bf u}\in{\R^n} : {\bf v}_1\cdot{\bf u} = \cdots = {\bf v}_k\cdot{\bf u} = 0 \} \]
be the linear space orthogonal to the ${\bf v}_i$'s (which must include ${\bf q}$), and let
${\bf w}_1,\ldots,{\bf w}_{n-k}\in{\R^n}$ be any orthonormal basis for $L$.
In a moment, we will explicitly construct a sequence of neighborhoods
$B_1,B_2,\ldots$ of $\overline{\xx}$.
To prove that these form a countable basis, we will first show,
in the next theorem,
that
if we choose any ${\bf x}_t\in{\R^n}\cap B_t$, for $t=1,2,\ldots$, then the
resulting sequence $\seq{{\bf x}_t}$ must converge to $\overline{\xx}$.
We will then argue that this property implies that the collection of
neighborhoods is in fact a countable basis.
Thus, we need to choose the $B_t$'s in a way that effectively forces
any such sequence $\seq{{\bf x}_t}$ to converge to $\overline{\xx}$.
To do this, we want to ensure that the sequence is growing to $+\infty$
in all of the directions ${\bf v}_1,\ldots,{\bf v}_k$, but is growing faster
in the ${\bf v}_1$ direction than ${\bf v}_2$, and faster in the ${\bf v}_2$
direction than ${\bf v}_3$, etc.
Also, in the directions orthogonal to ${\bf v}_1,\ldots,{\bf v}_k$, the
sequence needs to converge to the point ${\bf q}\in{\R^n}$.
In other words, we want to ensure that the three conditions of
Theorem~\ref{thm:seq-rep}
are all satisfied.
Based on these intuitions, for $t=1,2,\ldots$, we define the set
$B_t$ to consist of all those points $\overline{{\bf z}}\in\extspac{n}$ satisfying
all of the following conditions:
\begin{itemize}
\item \label{it:count-base-cond:a}
$\overline{{\bf z}}\cdot{\bf v}_i > t \;$ for $i=1,\ldots,k$.
\item \label{it:count-base-cond:b}
$\overline{{\bf z}}\cdot({\bf v}_i - t\,{\bf v}_{i+1}) > 0 \;$ for $i=1,\ldots,k-1$.
\item \label{it:count-base-cond:c}
$|\overline{{\bf z}}\cdot{\bf w}_j - {\bf q}\cdot{\bf w}_j| < 1/t \;$ for $j=1,\ldots,n-k$.
\end{itemize}
Each $B_t$ is open since it has the form of a standard basis
element, as in \eqref{eq:h:3a}.
Also, $\overline{\xx}$ is in each $B_t$ since, by its canonical form (and using
\eqref{eqn:pf:2}), we have:
\begin{itemize}
\item
$\overline{\xx}\cdot{\bf v}_i=+\infty$ for $i=1,\ldots,k$
(since
${\bf v}_j\cdot{\bf v}_i=0$ for $j<i$ and ${\bf v}_i\cdot{\bf v}_i=1$);
\item
$\overline{\xx}\cdot({\bf v}_i - t\,{\bf v}_{i+1}) = +\infty$, for $i=1,\ldots,k-1$
(since ${\bf v}_j\cdot({\bf v}_i-t\,{\bf v}_{i+1})=0$ for $j<i$
and ${\bf v}_i\cdot({\bf v}_i-t\,{\bf v}_{i+1})=1$);
\item
$\overline{\xx}\cdot{\bf w}_j = {\bf q}\cdot{\bf w}_j$, for $j=1,\ldots,n-k$
(by Proposition~\ref{pr:vtransu-zero}).
\end{itemize}
\begin{theorem} \label{thm:h:7}
Let $\overline{\xx}\in\extspac{n}$ have the canonical form in \eqref{eq:h:7},
and let $L$, the ${\bf w}_j$'s, and the collection of sets $B_t$ be as
defined above.
For each $t=1,2,\ldots$, let ${\bf x}_t$ be any point in ${\R^n}\cap B_t$.
Then ${\bf x}_t\rightarrow\overline{\xx}$.
\end{theorem}
\proof
We prove that the sequence $\seq{{\bf x}_t}$ satisfies the
conditions given in Theorem~\ref{thm:seq-rep} for a
sequence to converge to an astral point.
By linear algebra, we can write each ${\bf x}_t$ uniquely as
\[
{\bf x}_t = \sum_{i=1}^k b_{t,i} {\bf v}_i + {\bf q}_t,
\]
where each $b_{t,i}={\bf x}_t\cdot{\bf v}_i$ (since the ${\bf v}_i$'s are
orthonormal), and for some ${\bf q}_t\in L$.
Since ${\bf x}_t\in B_t$, we have
$b_{t,i}={\bf x}_t\cdot{\bf v}_i > t$,
for $i=1,\ldots,k$,
so $b_{t,i}\rightarrow+\infty$.
Therefore,
condition~(\ref{thm:seq-rep:a}) of
Theorem~\ref{thm:seq-rep} is satisfied.
That ${\bf x}_t\in B_t$ also implies, for $i=1,\ldots,k-1$, that
${\bf x}_t\cdot({\bf v}_i - t\,{\bf v}_{i+1}) > 0$,
so that
\[
b_{t,i}={\bf x}_t\cdot{\bf v}_i > t\,({\bf x}_t\cdot{\bf v}_{i+1}) = t\,b_{t,i+1}.
\]
As a result,
$0<b_{t,i+1}/b_{t,i} < 1/t$, and thus,
$b_{t,i+1}/b_{t,i}\rightarrow 0$.
Therefore,
condition~(\ref{thm:seq-rep:b}) of
Theorem~\ref{thm:seq-rep} is satisfied as well.
Finally, for $j=1,\ldots,n-k$, each ${\bf w}_j\in L$ so
${\bf x}_t\cdot{\bf w}_j={\bf q}_t\cdot{\bf w}_j$.
Since ${\bf x}_t\in B_t$, this implies
\[
| {\bf q}_t\cdot{\bf w}_j - {\bf q}\cdot{\bf w}_j |
=
| {\bf x}_t\cdot{\bf w}_j - {\bf q}\cdot{\bf w}_j |
<
\frac{1}{t},
\]
so ${\bf q}_t\cdot{\bf w}_j \rightarrow {\bf q}\cdot{\bf w}_j$.
Since the ${\bf w}_j$'s form an orthonormal basis for $L$,
and since ${\bf q}_t$ and ${\bf q}$
are both in $L$, it follows that ${\bf q}_t\rightarrow{\bf q}$;
indeed,
\[
{\bf q}_t
=
\sum_{j=1}^{n-k} ({\bf q}_t\cdot{\bf w}_j) {\bf w}_j
\rightarrow
\sum_{j=1}^{n-k} ({\bf q}\cdot{\bf w}_j) {\bf w}_j
=
{\bf q}.
\]
Thus, condition~(\ref{thm:seq-rep:c}) of
Theorem~\ref{thm:seq-rep} is also satisfied.
Having satisfied Theorem~\ref{thm:seq-rep}'s
three conditions,
we conclude that ${\bf x}_t\rightarrow\overline{\xx}$.
\qed
To finish the argument of first countability, we show that any
collection of neighborhoods with the property proved in
Theorem~\ref{thm:h:7} must in fact be a countable basis.
\begin{theorem}
Let $\overline{\xx}\in\extspac{n}$ and let $B_1,B_2,\ldots$ be a countable
collection of neighborhoods of $\overline{\xx}$.
Suppose that for every sequence ${\bf x}_t$, if ${\bf x}_t\in{\R^n}\cap B_t$,
for $t=1,2,\ldots$, then ${\bf x}_t\rightarrow\overline{\xx}$.
Then the $B_t$'s form a countable basis for $\overline{\xx}$.
\end{theorem}
\proof
Suppose, contrary to the theorem's conclusion, that there exists a
neighborhood $N$ of $\overline{\xx}$ that does not contain any $B_t$.
In other words, for each $t$, we can find a point
$\overline{\xx}_t\in B_t\cap {N^c}$
where ${N^c}=\extspac{n} \setminus N$ is the complement of $N$.
Since ${N^c}$ is closed (being the complement of an open set),
and since $\overline{\xx}\not\in{N^c}$, there must
exist disjoint open sets $U$ and $V$ such that ${N^c}\subseteq U$ and
$\overline{\xx}\in V$
(by Proposition~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:c})).
Thus, each $\overline{\xx}_t\in B_t\cap{N^c} \subseteq B_t\cap U$.
In other words, $B_t\cap U$ is a neighborhood of $\overline{\xx}_t$, and
therefore, there exists a point
${\bf x}_t\in B_t\cap U\cap{\R^n}$
(since ${\R^n}$ is dense in $\extspac{n}$).
By assumption, the resulting sequence $\seq{{\bf x}_t}$ converges to $\overline{\xx}$.
Since $V$ is a neighborhood of $\overline{\xx}$, this means that all but
finitely many of the points ${\bf x}_t$ are in $V$.
But this is a contradiction since every ${\bf x}_t\in U$, and $U$ and $V$
are disjoint.
\qed
First countability follows immediately:
\begin{theorem} \label{thm:1count}
Let $\overline{\xx}\in\extspac{n}$, and let $B_1,B_2,\ldots$ be the countable
collection of neighborhoods of $\overline{\xx}$ constructed above and considered in
Theorem~\ref{thm:h:7}.
Then this collection is a countable basis for $\overline{\xx}$.
Therefore, $\extspac{n}$ is first countable and also
sequentially compact.
\end{theorem}
As corollary, every astral point
$\overline{\xx}$ has a countable basis that is \emph{nested}, meaning
$B_1\supseteq B_2\supseteq \cdots$.
Furthermore, any sequence whose points are selected respectively from
each set in a nested countable basis must converge to $\overline{\xx}$.
\begin{corollary} \label{cor:nest-count-basis}
Let $\overline{\xx}\in\extspac{n}$.
Then there exists a nested countable basis $B_1,B_2,\ldots$
for $\overline{\xx}$.
Furthermore, if $\seq{\overline{\xx}_t}$ is a sequence in $\extspac{n}$ with
$\overline{\xx}_t\in B_t$ for all $t$, then
$\overline{\xx}_t\rightarrow\overline{\xx}$.
\end{corollary}
\proof
By Theorem~\ref{thm:1count}, there exists a countable basis
$B'_1,B'_2,\ldots$ for $\overline{\xx}$.
For each $t$,
let $B_t=B'_1\cap\cdots\cap B'_t$.
Each of these sets is a neighborhood of $\overline{\xx}$.
And for any neighborhood $U$ of $\overline{\xx}$, we must have,
for some $t$,
$B'_t\subseteq U$, which implies $B_t\subseteq U$.
Thus, $B_1,B_2,\ldots$ is a nested countable basis.
Suppose $\seq{\overline{\xx}_t}$ is a sequence in $\extspac{n}$ with
$\overline{\xx}_t\in B_t$ for all $t$.
Let $U$ be a neighborhood of $\overline{\xx}$.
Since $B_1,B_2,\ldots$ is a countable basis,
there must exist $t_0$ with
$B_{t_0}\subseteq U$, implying
$\overline{\xx}_t\in B_t\subseteq B_{t_0}\subseteq U$
for all $t\geq t_0$.
Thus, every neighborhood $U$ of $\overline{\xx}$ includes all but finitely many
sequence elements $\overline{\xx}_t$.
Therefore, $\overline{\xx}_t\rightarrow\overline{\xx}$.
\qed
\subsection{Not second countable and not metrizable}
\label{sec:not-second}
Although $\extspac{n}$ is first countable, we show next that it does not
satisfy the more stringent second axiom of countability, a fact that
implies it also is not metrizable.
Second countability would require that there exist a
\emph{countable basis} for the entire space, that is, a countable
collection of open sets such that for every neighborhood $U$ of every
point $\overline{\xx}$ in the space, there exists a set $B$ in the collection with
$\overline{\xx}\in B\subseteq U$.
We show that $\extspac{n}$ does not satisfy this condition for
$n\geq 2$.
To prove this,
for each ${\bf v}\in{\R^n}$, we will define an open set ${U_{\vv}}$ with
the property that ${U_{\vv}}$ includes the astron $\limray{{\bf v}}$, but does
not include any other astron (nor even any other icon).
This will imply that a countable basis must include a different set
for every astron, which is impossible
since the set of all astrons is uncountable when $n\geq 2$.
These sets will be used again later in the manuscript, and we therefore
prove properties for these sets that are a bit stronger than is
necessary for our current purposes.
In particular, we prove that all of the points in ${U_{\vv}}$ must either be
in ${\R^n}$ or have the form $\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some
${\bf q}\in{\R^n}$; in other words, ${U_{\vv}}$ is entirely included in the union
of the galaxies $\galax{{\bf 0}}={\R^n}$ and $\galax{\limray{{\bf v}}}$.
This allows us to exclude the possibility of ${U_{\vv}}$ including any
astrons or icons other than $\limray{{\bf v}}$.
\begin{theorem} \label{thm:formerly-lem:h:1}
For every ${\bf v}\in{\R^n}$,
there exists an open set ${U_{\vv}}\subseteq\extspac{n}$ for which the
following hold:
\begin{enumerate}
\item \label{thm:formerly-lem:h:1:a}
The only icon in ${U_{\vv}}$ is $\limray{{\bf v}}$.
That is, ${U_{\vv}}\cap{{\cal E}_n}=\{\limray{{\bf v}}\}$.
\item \label{thm:formerly-lem:h:1:b}
If $\overline{\xx}\in{U_{\vv}}$, then either $\overline{\xx}\in{\R^n}$ or
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some ${\bf q}\in{\R^n}$.
That is, ${U_{\vv}}\subseteq {\R^n} \cup [\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\R^n}]$.
\item \label{thm:formerly-lem:h:1:c}
If ${\bf w}\in{\R^n}$ and $\norm{{\bf v}}=\norm{{\bf w}}=1$, then
$\limray{{\bf w}}\in{U_{\vv}}$ if and only if ${\bf w}={\bf v}$.
\end{enumerate}
\end{theorem}
\proof
Let ${\bf v}\in{\R^n}$.
If ${\bf v}={\bf 0}$, we can simply choose ${U_{\zero}}={\R^n}$, which
satisfies
part~(\ref{thm:formerly-lem:h:1:a}) by
Proposition~\ref{pr:i:8}(\ref{pr:i:8b}),
part~(\ref{thm:formerly-lem:h:1:b}) trivially,
and
part~(\ref{thm:formerly-lem:h:1:c}) vacuously.
So we assume henceforth that ${\bf v}\neq{\bf 0}$.
Also, it suffices to consider only the case that ${\bf v}$ is a unit
vector since in all other cases, when $\norm{{\bf v}}\neq 1$,
we can simply choose ${U_{\vv}}$ to be the same as the corresponding set
for a normalized version of ${\bf v}$; that is, we can choose
${U_{\vv}}={U_{\vv/\norm{\vv}}}$.
Therefore, we assume henceforth that
$\norm{{\bf v}}=1$.
Let ${\bf u}_1,\ldots,{\bf u}_{n-1}$ be
any orthonormal basis for the linear space perpendicular to ${\bf v}$;
thus, ${\bf v},{\bf u}_1,\ldots,{\bf u}_{n-1}$ form an orthonormal basis for all
of ${\R^n}$.
We define ${U_{\vv}}$ to be the set of all $\overline{\xx}\in\extspac{n}$ which satisfy
all of the following:
\begin{itemize}
\item
$\overline{\xx}\cdot{\bf v}>0$.
\item
$|\overline{\xx}\cdot{\bf u}_j| < 1$ for all $j=1,\ldots,n-1$.
\end{itemize}
Then ${U_{\vv}}$ is open, since it has the form of a standard basis element
as in \eqref{eq:h:3a}.
With this set-up, we can now prove the different parts of the theorem
(though not in the order stated).
Part~(\ref{thm:formerly-lem:h:1:b}):
Suppose
$\overline{\xx}\in{U_{\vv}}$, and let
$\overline{\xx}=\limrays{{\bf w}_1,\ldots,{\bf w}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
be $\overline{\xx}$'s canonical representation.
Suppose ${\bf w}_i\cdot{\bf u}_j\neq 0$
for some $i\in\{1,\ldots,k\}$ and some
$j\in\{1,\ldots,n-1\}$.
Then $\overline{\xx}\cdot{\bf u}_j\in\{-\infty,+\infty\}$ by
Proposition~\ref{pr:vtransu-zero}.
But this is impossible since $\overline{\xx}\in{U_{\vv}}$
so $|\overline{\xx}\cdot{\bf u}_j|<1$.
Therefore, for $i=1,\ldots,k$, ${\bf w}_i\cdot{\bf u}_j=0$ for
$j=1,\ldots,n-1$.
Since ${\bf w}_i$ is a unit vector, and
since ${\bf v},{\bf u}_1,\ldots,{\bf u}_{n-1}$ form an orthonormal basis for
${\R^n}$, the only possibility
is that each ${\bf w}_i$ is either equal to ${\bf v}$ or $-{\bf v}$.
Furthermore,
since the ${\bf w}_i$'s are orthogonal to one another, this further
implies that $k\leq 1$.
If $k=0$, then $\overline{\xx}={\bf q}\in{\R^n}$.
Otherwise, if $k=1$, then ${\bf w}_1$ is either ${\bf v}$ or $-{\bf v}$,
as just argued.
But if ${\bf w}_1=-{\bf v}$, then
$\overline{\xx}\cdot{\bf v}=\limray{{\bf w}_1}\cdot{\bf v}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\cdot{\bf v}=-\infty$,
contradicting that
$\overline{\xx}\cdot{\bf v}>0$ since $\overline{\xx}\in{U_{\vv}}$.
Thus, if $k=1$ then
$\overline{\xx}=\limray{{\bf w}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$.
We conclude that $\overline{\xx}$ is either in ${\R^n}$ or
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\R^n}$, as claimed.
Part~(\ref{thm:formerly-lem:h:1:a}).
From the preceding definitions,
$\limray{{\bf v}}\cdot{\bf v}=+\infty$ and
$\limray{{\bf v}}\cdot{\bf u}_j=0$ for $j=1,\ldots,n-1$;
therefore, $\limray{{\bf v}}$ is in ${U_{\vv}}$.
To show no other icons are in ${U_{\vv}}$,
suppose $\overline{\ee}\in{U_{\vv}}\cap{{\cal E}_n}$.
Then $\overline{\ee}$ must have the
form stated in part~(\ref{thm:formerly-lem:h:1:b}).
Since it is in ${U_{\vv}}$, $\overline{\ee}\cdot{\bf v}>0$, implying
$\overline{\ee}\neq{\bf 0}$, and
further implying $\overline{\ee}\not\in{\R^n}$
(by Proposition~\ref{pr:i:8}(\ref{pr:i:8b})).
So we must have
$\overline{\ee}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some ${\bf q}\in{\R^n}$.
Since every point's iconic part is unique
(Proposition~\ref{pr:i:8}(\ref{pr:i:8c})), this means the iconic parts on the left and right must be equal; that is,
$\overline{\ee}=\limray{{\bf v}}$.
Part~(\ref{thm:formerly-lem:h:1:c}):
Let ${\bf w}\in{\R^n}$ and assume $\norm{{\bf v}}=\norm{{\bf w}}=1$.
Then by part~(\ref{thm:formerly-lem:h:1:a}),
$\limray{{\bf w}}\in{U_{\vv}}$ if and only if $\limray{{\bf w}}=\limray{{\bf v}}$.
Also, if $\limray{{\bf w}}=\limray{{\bf v}}$ then ${\bf w}$ and ${\bf v}$ are dominant
directions for the same point, and so must be equal
(by Theorem~\ref{thm:dom-dir}).
Therefore, $\limray{{\bf w}}=\limray{{\bf v}}$ if and only if ${\bf w}={\bf v}$.
\qed
\begin{theorem} \label{thm:h:8}
For $n\geq 2$, astral space $\extspac{n}$ is not second countable.
\end{theorem}
\proof
Let $S=\{ {\bf v}\in{\R^n} : \norm{{\bf v}}=1 \}$ be the unit sphere in ${\R^n}$,
and for each ${\bf v}\in S$,
let ${U_{\vv}}$ be as in Theorem~\ref{thm:formerly-lem:h:1}.
Suppose, contrary to the theorem, that there exists a countable basis
$B_1,B_2,\ldots$ for $\extspac{n}$.
Then for each ${\bf v}\in S$, ${U_{\vv}}$ is a neighborhood of $\limray{{\bf v}}$,
so there exists an index $i({\bf v})\in{\mathbb{N}}$ such that
$\limray{{\bf v}}\in B_{i({\bf v})} \subseteq {U_{\vv}}$.
The resulting function $i: S\rightarrow{\mathbb{N}}$ is injective since if
$i({\bf v})=i({\bf w})$ for some ${\bf v},{\bf w}\in S$, then
\[ \limray{{\bf w}}\in B_{i({\bf w})} = B_{i({\bf v})} \subseteq {U_{\vv}}, \]
which implies ${\bf v}={\bf w}$ by
Theorem~\ref{thm:formerly-lem:h:1}(\ref{thm:formerly-lem:h:1:c}).
But such an injective function is impossible since $S$ is uncountable.
\qed
A space is \emph{metrizable} if there exists a metric that induces its
topology.
As a corollary to Theorem~\ref{thm:h:8},
$\extspac{n}$ also cannot be metrizable.
\begin{corollary}
For $n\geq 2$, astral space $\extspac{n}$ is not metrizable.
\end{corollary}
\proof
Since ${\R^n}$ is dense in $\extspac{n}$, ${\rats^n}$ is as well (since if $U$ is
a neighborhood of $\overline{\xx}\in\extspac{n}$, then there exists a point
in the open set $U\cap{\R^n}$, implying, in turn, that
$(U\cap{\R^n})\cap{\rats^n}=U\cap{\rats^n}\neq\emptyset$).
Thus, $\extspac{n}$ is \emph{separable}, meaning that it includes a
countable dense subset.
Since it is known~\citep[Lemma 3.4]{hitchhiker_guide_analysis}
that every separable metrizable space has a
countable basis, it follows from Theorem~\ref{thm:h:8}
that $\extspac{n}$ is not metrizable.
\qed
The sets ${U_{\vv}}$ constructed in proving
Theorem~\ref{thm:formerly-lem:h:1} highlight some important
topological properties of astrons.
In ${\R^n}$, any neighborhood of a point ${\bf x}\in{\R^n}$, no matter how tiny
and constrained, will include other points in ${\R^n}$, which means,
topologically, that no point can be isolated from all others.
Intuitively, one might expect astrons to behave in a similar way, that
any neighborhood of an astron $\limray{{\bf v}}$ will include other
``nearby'' astrons $\limray{{\bf w}}$, likely including those for which
${\bf w}$ is exceedingly close to ${\bf v}$ (in ${\R^n}$).
Theorem~\ref{thm:formerly-lem:h:1} shows that this is not true
for any astron.
Rather, every astron $\limray{{\bf v}}$ has a neighborhood that excludes
\emph{all} other astrons $\limray{{\bf w}}$, no matter how tiny the
distance between ${\bf v}$ and ${\bf w}$.
So although it might be tempting to picture astrons as points
``in the sky'', a seemingly continuous set in direct
correspondence with points on the unit sphere,
in fact, astrons are quite topologically discrete and isolated from
one another.
Also,
we saw in Proposition~\ref{pr:i:8}(\ref{pr:i:8e})
that ${{\cal E}_n}$, the set of all icons, is closed, which means that a
convergent sequence of icons will converge to an icon.
One might expect something similar for astrons.
Indeed, if $\seq{\limray{{\bf v}_t}}$ is a sequence of astrons, and if
${\bf v}_t\rightarrow{\bf w}$, for some ${\bf w}\in{\R^n}$,
then we might naturally expect
$\limray{{\bf v}_t}\rightarrow\limray{{\bf w}}$.
But this is not true in general.
For instance, in ${\mathbb{R}}^2$, suppose that
${\bf v}_t = {\bf e}_1 + (1/t) {\bf e}_2$
where ${\bf e}_1$ and ${\bf e}_2$ are the standard basis vectors.
Then ${\bf v}_t\rightarrow{\bf e}_1$, but
$\limray{{\bf v}_t}\not\rightarrow\limray{{\bf e}_1}$.
Rather,
$\limray{{\bf v}_t}\rightarrow\overline{{\bf z}}$
where $\overline{{\bf z}}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_2}$.
To see this, let ${\bf u}\in{\R^n}$.
If ${\bf e}_1\cdot{\bf u}>0$ then
${\bf v}_t\cdot{\bf u}\rightarrow{\bf e}_1\cdot{\bf u}$, so for all
sufficiently large $t$, ${\bf v}_t\cdot{\bf u}>0$,
implying $\limray{{\bf v}_t}\cdot{\bf u}=+\infty=\overline{{\bf z}}\cdot{\bf u}$.
The case ${\bf e}_1\cdot{\bf u}<0$ is symmetric.
And if ${\bf e}_1\cdot{\bf u}=0$, then
${\bf v}_t\cdot{\bf u}=(1/t){\bf e}_2\cdot{\bf u}$
so, in this case,
$\limray{{\bf v}_t}\cdot{\bf u}=\limray{{\bf e}_2}\cdot{\bf u}=\overline{{\bf z}}\cdot{\bf u}$ for all $t$.
In every case $\limray{{\bf v}_t}\cdot{\bf u}\rightarrow\overline{{\bf z}}\cdot{\bf u}$
so $\limray{{\bf v}_t}\rightarrow\overline{{\bf z}}$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
More generally, a sequence of icons $\seq{\overline{\ee}_t}$ converges
to some astron $\limray{{\bf v}}$ if and only if
$\overline{\ee}_t=\limray{{\bf v}}$ for all but finitely many values of $t$.
This is because, by Theorem~\ref{thm:formerly-lem:h:1}, there exists a
neighborhood ${U_{\vv}}$ of $\limray{{\bf v}}$ that excludes all other
icons.
Since the sequence converges to $\limray{{\bf v}}$,
for all sufficiently large $t$, every sequence element
$\overline{\ee}_t$ must be in ${U_{\vv}}$, and therefore must be equal to
$\limray{{\bf v}}$.
\section{Extending functions to astral space}
\label{sec:functions}
We are now ready to begin the study of functions
that have been extended to astral space.
We are especially motivated by the fundamental problem of minimizing a
convex function $f$ on ${\R^n}$.
In general, such a function might not be minimized at any
finite point in its domain so that its minimizers can only be attained
``at infinity'' by following the trajectory of a sequence.
To study this situation within our framework,
we focus particularly on an extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of $f$ to $\extspac{n}$
that is constructed in a way ensuring that $f$'s minimum over sequences in
${\R^n}$ coincides with ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s minimum, which is attained always
at some astral point in $\extspac{n}$.
Much of the rest of the manuscript studies ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s properties,
for example, where it is continuous and the nature of its minimizers.
\subsection{Lower semicontinuous extension}
\label{sec:lsc:ext}
We start by defining the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ and proving some basic,
useful properties.
Let $f:{\R^n}\rightarrow\overline{\R}$.
Later, we will usually require $f$ to be convex, but for now, we
allow it to be any function on ${\R^n}$.
As is standard,
the \emph{lower semicontinuous hull} of $f$, denoted $\lsc f$, is the
function on ${\R^n}$ defined by
\[
(\lsc f)({\bf x}) = \inf_{\seq{{\bf x}_t}:{\bf x}_t\rightarrow {\bf x}}
\;\liminf_{t\rightarrow\infty} f({\bf x}_t)
\]
for ${\bf x}\in{\R^n}$,
where the infimum is over all sequences $\seq{{\bf x}_t}$ in ${\R^n}$ converging to ${\bf x}$.
It is the greatest lower semicontinuous function that is majorized by
$f$.
In the same fashion,
we can extend any function $f$ to astral space $\extspac{n}$ by
defining its \emph{lower semicontinuous extension} (or simply,
\emph{extension}), denoted ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, which is defined by
\begin{equation} \label{eq:e:7}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) = \inf_{\seq{{\bf x}_t}:{\bf x}_t\rightarrow \overline{\xx}}
\;\liminf_{t\rightarrow\infty} f({\bf x}_t)
\end{equation}
for $\overline{\xx}\in\extspac{n}$,
where the infimum now is over all sequences $\seq{{\bf x}_t}$ in ${\R^n}$
converging to $\overline{\xx}$.
We will see soon that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$ is the minimum achievable
limit of $f({\bf x}_t)$ for any sequence $\seq{{\bf x}_t}$ in ${\R^n}$ that
converges to $\overline{\xx}$.
For instance, suppose
$f({\bf x})={\bf x}\cdot{\bf u}+b$ for all ${\bf x}\in{\R^n}$,
and for some ${\bf u}\in{\R^n}$ and $b\in{\mathbb{R}}$.
In this very simple case, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=\overline{\xx}\cdot{\bf u}+b$ since if
$\seq{{\bf x}_t}$ is any sequence in ${\R^n}$ that converges to $\overline{\xx}\in\extspac{n}$,
then ${\bf x}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})) so
$f({\bf x}_t)\rightarrow\overline{\xx}\cdot{\bf u}+b$ by continuity.
As a more involved example, suppose
\begin{equation} \label{eqn:recip-fcn-eg}
f({\bf x}) = f(x_1, x_2)
= \left\{ \begin{array}{cl}
\displaystyle{\frac{1}{x_1 x_2}}
& \mbox{if $x_1>0$ and $x_2>0$} \\[1em]
+\infty & \mbox{otherwise} \\
\end{array} \right.
\end{equation}
for ${\bf x}=\trans{[x_1,x_2]}\in{\mathbb{R}}^2$.
This function is convex, closed, proper and continuous everywhere.
Suppose $\beta\in{\mathbb{R}}$ and $\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\beta{\bf e}_2$
(where ${\bf e}_1$ and ${\bf e}_2$ are standard basis vectors).
If $\beta>0$,
then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=0$ since on any sequence $\seq{{\bf x}_t}$ converging
to $\overline{\xx}$, the first component ${\bf x}_t\cdot{\bf e}_1=x_{t1}$ must converge
to $\overline{\xx}\cdot{\bf e}_1=+\infty$,
while the second component ${\bf x}_t\cdot{\bf e}_2=x_{t2}$ must
converge to $\overline{\xx}\cdot{\bf e}_2=\beta>0$, implying that
$f({\bf x}_t)\rightarrow 0$.
If $\beta<0$, then a similar argument shows that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$.
And if $\beta=0$, so that
$\overline{\xx}=\limray{{\bf e}_1}$, then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$ is again equal to $0$,
although more care is now needed in finding a sequence that shows this:
For instance, if ${\bf x}_t=t^2{\bf e}_1 + (1/t){\bf e}_2$, then
$f({\bf x}_t)=1/t\rightarrow 0$.
This implies ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\leq 0$,
and since $f$ is nonnegative everywhere,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is as well, so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=0$.
For the remainder of this subsection,
we establish some basic properties of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\begin{proposition} \label{pr:h:1}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then the following hold for $f$'s extension, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$:
\begin{enumerate}
\item \label{pr:h:1a}
For all ${\bf x}\in{\R^n}$,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})=(\lsc f)({\bf x})\leq f({\bf x})$
(implying ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})=f({\bf x})$ if $f$ is already lower semicontinuous).
\item \label{pr:h:1b}
Let $U\subseteq\extspac{n}$ be a neighborhood of some point
$\overline{\xx}\in\extspac{n}$, and suppose
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<b$ for some $b\in{\mathbb{R}}$.
Then there exists a
point ${\bf x}\in U\cap {\R^n}$ with $f({\bf x})<b$.
\item \label{pr:h:1b2}
Let $U\subseteq\extspac{n}$ be a neighborhood of some point
$\overline{\xx}\in\extspac{n}$, and suppose
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})>b$ for some $b\in{\mathbb{R}}$.
Then there exists a
point ${\bf x}\in U\cap {\R^n}$ with $f({\bf x})>b$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:h:1a}):
The equality is immediate from definitions.
That $(\lsc f)({\bf x})\leq f({\bf x})$ follows by considering a sequence
$\seq{{\bf x}_t}$ in
which ${\bf x}_t={\bf x}$ for all $t$.
Part~(\ref{pr:h:1b}):
By definition of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<b$,
there exists a sequence $\seq{{\bf x}_t}$ in ${\R^n}$
converging to $\overline{\xx}$ such that
$\liminf f({\bf x}_t) < b$.
For some $t_0$, ${\bf x}_t$ must be in $U$ for all $t\geq t_0$.
For at least one of these points, we must have $f({\bf x}_t)<b$,
since
$\inf_{t\geq t_0} f({\bf x}_t) \leq \liminf f({\bf x}_t) < b$.
Part~(\ref{pr:h:1b2}):
Let $\seq{{\bf x}_t}$ be any sequence in ${\R^n}$ converging to $\overline{\xx}$.
Then $\liminf f({\bf x}_t)\geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})>b$,
which implies that $f({\bf x}_t)>b$ for infinitely many values of $t$.
Since $U$ is a neighborhood of $\overline{\xx}$,
${\bf x}_t\in U$ for all $t$ sufficiently large.
Thus, there must be some ${\bf x}_t$ with the claimed properties.
\qed
Because astral space is compact,
the minimum of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is always realized at some point
$\overline{\xx}\in\extspac{n}$, and is always equal to $\inf f$.
Thus, minimizing a function $f$ on ${\R^n}$ is equivalent to finding a
minimizer of its extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, which must always exist:
\begin{proposition} \label{pr:fext-min-exists}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then there exists a point $\overline{\xx}\in\extspac{n}$
for which ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=\inf f$.
Thus,
\[
\min_{\overline{\xx}\in\extspac{n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
=
\inf_{{\bf x}\in{\R^n}} f({\bf x})
=
\inf f.
\]
\end{proposition}
\proof
If $f\equiv +\infty$ then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}\equiv +\infty$ and the claim follows
trivially.
So we assume henceforth that $\inf f < +\infty$.
Every element $f({\bf x}_t)$ appearing in \eqref{eq:e:7} is at least
$\inf f$; therefore, $\inf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'} \geq \inf f$.
For each $t$,
let ${\bf x}_t\in{\R^n}$ be such that
\[ f({\bf x}_t)\leq \left\{ \begin{array}{cl}
\inf f + 1/t & \mbox{if $\inf f \in {\mathbb{R}}$}\\
-t & \mbox{if $\inf f = -\infty$.}
\end{array}
\right.
\]
Then $f({\bf x}_t)\rightarrow \inf f$.
Because $\extspac{n}$ is sequentially compact, the sequence $\seq{{\bf x}_t}$ has a
convergent subsequence which, by discarding all other elements, we can
take to be the entire sequence.
Let $\overline{\xx}$ be the resulting sequence's limit.
Then
\[ \inf f \leq \inf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'} \leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) \leq \lim f(x_t) = \inf f. \]
The first inequality was argued above.
The third inequality is by ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s definition.
This proves $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, whose minimum is $\inf f$.
\qed
We next show that the infimum appearing in \eqref{eq:e:7}
must be realized by some sequence for which
$f({\bf x}_t)$ converges to ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$:
\begin{proposition} \label{pr:d1}
Let $f:{\R^n}\rightarrow\overline{\R}$, and let $\overline{\xx}\in\extspac{n}$.
Then there exists a sequence $\seq{{\bf x}_t}$ in ${\R^n}$ such that
${\bf x}_t\rightarrow\overline{\xx}$ and $f({\bf x}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
\end{proposition}
\proof
Let $B_1, B_2, \ldots$ be a nested countable basis
for $\overline{\xx}$ (which exists by Corollary~\ref{cor:nest-count-basis}).
By Proposition~\ref{pr:h:1}(\ref{pr:h:1b}), for each $t$, there must
exist a point ${\bf x}_t\in B_t\cap {\R^n}$ for which
$f({\bf x}_t) \leq b_t$ where
\[ b_t = \left\{ \begin{array}{cl}
-t & \mbox{if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) = -\infty$}\\
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) + 1/t & \mbox{if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) \in {\mathbb{R}}$}\\
+\infty & \mbox{if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) = +\infty$.}
\end{array}
\right.
\]
The resulting sequence $\seq{{\bf x}_t}$ converges to $\overline{\xx}$
(by Corollary~\ref{cor:nest-count-basis}).
Thus,
\[
\liminf f({\bf x}_t) \geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}).
\]
On the other hand,
\[
\limsup f({\bf x}_t) \leq \limsup b_t = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}).
\]
Thus, $\lim f({\bf x}_t) = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
\qed
The function ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is always lower semicontinuous, as will be seen
shortly.
In addition, because $\extspac{n}$ is first countable, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is
{continuous} at some point $\overline{\xx}$ if and only if, for every sequence of
points $\seq{\overline{\xx}_t}$ in $\extspac{n}$, if $\overline{\xx}_t\rightarrow\overline{\xx}$, then
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
In fact, for ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ to be continuous in this sense, it is sufficient
if this holds only for sequences in ${\R^n}$.
(Thus, in the example of \eqref{eqn:recip-fcn-eg}, the arguments above
show that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at
$\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\beta{\bf e}_2$ if $\beta\neq 0$.)
This and other properties are shown in the next proposition:
\begin{proposition} \label{pr:wasthm:h:2}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:wasthm:h:2a}
Let $\seq{\overline{\xx}_t}$ be a sequence in $\extspac{n}$ converging to
$\overline{\xx}\in\extspac{n}$.
Then
$\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t) \geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
In other words, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is lower semicontinuous.
\item \label{pr:wasthm:h:2aa}
The extension of $f$ is the same as that of its lower
semicontinuous hull.
That is, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=\overline{(\lsc f)}$.
\item \label{pr:wasthm:h:2b}
Let $\overline{\xx}\in\extspac{n}$.
Suppose that for any sequence $\seq{{\bf x}_t}$ in ${\R^n}$, if
${\bf x}_t\rightarrow\overline{\xx}$ then $f({\bf x}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$.
\item \label{pr:wasthm:h:2c}
In $\extspac{n}$, the closures of the effective domains of $f$ and
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ are identical.
That is,
$\cldom{f}=\cldom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}$.
\end{enumerate}
\end{proposition}
\proof ~
Part~(\ref{pr:wasthm:h:2a}):
Suppose the claim is false.
Then there exists $b\in{\mathbb{R}}$ such that
$\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t) < b < {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Let $B_1, B_2, \ldots$ be a nested countable basis
for $\overline{\xx}$
(which must exist by Corollary~\ref{cor:nest-count-basis}).
For each $t$,
all but finitely many of the sequence elements $\overline{\xx}_s$
must be included in $B_t$
(since it is a neighborhood of $\overline{\xx}$).
On the other hand,
since $\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t) < b$, we must have
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_s)<b$ for infinitely many values of
$s$.
Therefore, there must exist some index
$s$ with $\overline{\xx}_{s}\in B_t$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_{s})<b$.
By
Proposition~\ref{pr:h:1}(\ref{pr:h:1b}), it follows that
there must exist
${\bf x}_t\in B_t\cap{\R^n}$ with $f({\bf x}_t)<b$.
By Corollary~\ref{cor:nest-count-basis}, the resulting sequence
$\seq{{\bf x}_t}$ converges to $\overline{\xx}$.
Therefore,
$\liminf f({\bf x}_t)\geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) > b$.
However, this is a contradiction since $f({\bf x}_t)<b$ for all $t$.
Part~(\ref{pr:wasthm:h:2aa}):
We need to show ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=\overline{h}$ where $h=\lsc f$.
Since $h\leq f$, $\overline{h}\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
For the reverse inequality, note first that
$h({\bf x})=(\lsc f)({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})$
for all ${\bf x}\in{\R^n}$,
by Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
Let $\overline{\xx}\in\extspac{n}$.
By Proposition~\ref{pr:d1}, there exists a sequence $\seq{{\bf x}_t}$ in
${\R^n}$ with ${\bf x}_t\rightarrow\overline{\xx}$ and
$h({\bf x}_t)\rightarrow\overline{h}(\overline{\xx})$.
Thus,
\[
\overline{h}(\overline{\xx})
=
\lim h({\bf x}_t)
=
\lim {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x}_t)
\geq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
\]
where the inequality follows from part~(\ref{pr:wasthm:h:2a}).
Part~(\ref{pr:wasthm:h:2b}):
Let $\seq{\overline{\xx}_t}$ be any sequence in $\extspac{n}$ converging to
$\overline{\xx}\in\extspac{n}$.
By part~(\ref{pr:wasthm:h:2a}),
$\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t) \geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
So, to prove continuity of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at $\overline{\xx}$, it remains only
to show
$ \limsup {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t) \leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) $.
Similar to
part~(\ref{pr:wasthm:h:2a}),
suppose this does not hold, so that
$\limsup {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t) > b > {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$
for some $b\in{\mathbb{R}}$.
Let $B_1, B_2, \ldots$ be a nested countable basis
for $\overline{\xx}$.
For each $t$,
there must exist some
$s$ with $\overline{\xx}_{s}\in B_t$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_{s})>b$
(since, as before,
all but finitely many of the sequence elements $\overline{\xx}_s$
must be included in $B_t$,
and since
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_s)>b$ for infinitely many values of
$s$).
Therefore, by
Proposition~\ref{pr:h:1}(\ref{pr:h:1b2}),
there exists
${\bf x}_t\in B_t\cap{\R^n}$ with $f({\bf x}_t)>b$.
By Corollary~\ref{cor:nest-count-basis}, the resulting sequence
$\seq{{\bf x}_t}$ converges to $\overline{\xx}$.
Therefore, by our assumption,
$f({\bf x}_t) \rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) < b$.
However, this is a contradiction since $f({\bf x}_t)>b$ for all $t$.
Part~(\ref{pr:wasthm:h:2c}):
By Proposition~\ref{pr:h:1}(\ref{pr:h:1a}),
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})\leq f({\bf x})$ for all ${\bf x}\in{\R^n}$.
Therefore,
$\dom f \subseteq \dom {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
implying
$\cldom{f} \subseteq \cldom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}$.
For the reverse inclusion, suppose $\overline{\xx}\in\cldom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}$.
Let $U$ be any neighborhood of $\overline{\xx}$.
Then $U\cap (\dom {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'})$ is nonempty, meaning there exists $\overline{{\bf z}}\in U$
with ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}})<+\infty$.
Let $b\in {\mathbb{R}}$ be such that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}})<b$.
Then, by
Proposition~\ref{pr:h:1}(\ref{pr:h:1b}),
there must exist ${\bf z}\in U\cap {\R^n}$ with
$f({\bf z})<b$ so that
${\bf z}\in U\cap (\dom f)$,
thereby proving $\overline{\xx}\in\cldom{f}$.
Thus,
$\cldom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}} \subseteq \cldom{f}$,
completing the proof.
\qed
\subsection{Reductions}
\label{sec:shadow}
We next begin the study of reductions, a core technique for
analyzing astral functions that will be used throughout the remainder
of the manuscript.
In Proposition~\ref{pr:h:6}, we saw how every infinite astral point $\overline{\xx}$ can
be decomposed into its dominant direction ${\bf v}$ and its
projection $\overline{\xx}^\perp$ perpendicular to ${\bf v}$, which has
lower astral rank.
As mentioned earlier, this decomposition will be the basis for proving
properties by induction on astral rank.
In a related manner, and in a way that will be similarly useful for
such proofs, we show next how we can define a kind of projection of
the function ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, effectively reducing the dimensionality of its
domain while preserving key properties.
Let $f:{\R^n}\rightarrow\overline{\R}$ be a convex function.
For ${\bf v}\in{\R^n}$, we define an associated function
$\fshad{\limray{\vv}}:{\R^n}\rightarrow\overline{\R}$ by the rule
\[
\fshad{\limray{\vv}}({\bf x}) = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
\]
for ${\bf x}\in{\R^n}$.
This is called
the \emph{reduction of $f$ at astron $\limray{{\bf v}}$}.
Because this reduction is at an astron, it is said to be
\emph{astronic};
more general reductions will be
introduced later in Section~\ref{sec:ent-closed-fcn}.
Let $g=\fshad{\limray{\vv}}$ be such a reduction.
This function is a constant in the positive and negative direction of
${\bf v}$ (that is,
$g({\bf x})=g({\bf x}+\lambda{\bf v})$ for all ${\bf x}\in{\R^n}$ and $\lambda\in{\mathbb{R}}$), which
means that the reduction $g$ can be regarded informally as a function
only over the space perpendicular to ${\bf v}$, even if it is formally
defined over all of ${\R^n}$.
In this sense, $f$ has been ``reduced'' in forming $g$.
For instance, in the example given in
\eqref{eqn:recip-fcn-eg}, suppose ${\bf v}={\bf e}_1$.
Then, for ${\bf x}\in{\mathbb{R}}^2$,
\begin{equation} \label{eqn:recip-fcn-eg:g}
g({\bf x})
=
g(x_1,x_2)
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
\left\{ \begin{array}{cl}
0 & \mbox{if $x_2\geq 0$} \\
+\infty & \mbox{otherwise.}
\end{array} \right.
\end{equation}
This follows from exactly the kind of arguments given earlier
in Section~\ref{sec:lsc:ext} in discussing this example.
Returning to the general case,
the vector ${\bf v}$ is usually (but not always) assumed to be
in the \emph{recession cone} of $f$, denoted $\resc{f}$.
This is the set of directions in which the function never increases:
\begin{equation} \label{eqn:resc-cone-def}
\resc{f} = \{{\bf v}\in{\R^n} : \forall {\bf x}\in{\R^n}, f({\bf x}+{\bf v})\leq f({\bf x})\}.
\end{equation}
In the example function just discussed, for instance,
the recession cone consists of all vectors ${\bf v}=\trans{[v_1,v_2]}$
with $v_1,v_2\geq 0$, that is, vectors in $\mathbb{R}_{\geq 0}^2$ where
$\mathbb{R}_{\geq 0}=[0,+\infty)$ is the set of all nonnegative reals.
Recall that a set $K\subseteq{\R^n}$ is a \emph{cone} if it is closed
under multiplication by positive scalars, that is,
if $\lambda{\bf x}\in K$ for
all ${\bf x}\in K$ and $\lambda>0$.
We say that a cone $K$ is \emph{pointed} if it includes the origin.
For a set $S\subseteq{\R^n}$, the \emph{conic hull of $S$} or
\emph{cone generated by $S$}, denoted $\cone S$, is the set of all
nonnegatively weighted combinations of
(finitely many) elements of $S$, or
equivalently, the smallest convex cone that includes $S\cup\{{\bf 0}\}$.
Later, in Section~\ref{sec:cones}, we will extend these
notions to astral space.
The next proposition states some facts about recession cones that are
quite standard; we include them for completeness and generality.
\begin{proposition} \label{pr:resc-cone-basic-props}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then $f$'s recession cone, $\resc{f}$, is a pointed convex cone.
If, in addition, $f$ is lower semicontinuous, then $\resc{f}$ is
closed in ${\R^n}$.
\end{proposition}
Note that the last line is saying that $\resc{f}$ is closed as a
subset of ${\R^n}$, not $\extspac{n}$.
\proof
That ${\bf 0}$ is in $\resc{f}$ is immediate from definitions.
\underline{Convex:}
Suppose ${\bf v}$ and ${\bf v}'$ are in $\resc{f}$, and that
$\lambda\in [0,1]$.
Let ${\bf w}=\lambda{\bf v} + (1-\lambda){\bf v}'$, which we aim to show is in
$\resc{f}$.
Let ${\bf x}\in{\R^n}$.
If $f({\bf x})=+\infty$ then $f({\bf x}+{\bf w})\leq f({\bf x})$.
Otherwise,
let $\beta\in{\mathbb{R}}$ be any number for which $f({\bf x})<\beta$.
Then $f({\bf x}+{\bf v})\leq f({\bf x})<\beta$ since ${\bf v}\in\resc{f}$, and
similarly
$f({\bf x}+{\bf v}')<\beta$.
So by Proposition~\ref{pr:rock:4.2},
\[
f({\bf x}+{\bf w})
=
f(\lambda({\bf x}+{\bf v})+(1-\lambda)({\bf x}+{\bf v}'))
<
\lambda \beta +(1-\lambda) \beta = \beta.
\]
Since this is true for all $\beta>f({\bf x})$, it follows that
$f({\bf x}+{\bf w})\leq f({\bf x})$.
Thus, $\resc{f}$ is convex.
\underline{Cone:}
Let ${\bf v}\in\resc{f}$.
By a simple induction argument on $n=1,2,\ldots$, $n{\bf v}$ must also be
in $\resc{f}$ since for $n>1$ and for all ${\bf x}\in{\R^n}$,
\[
f({\bf x}+n{\bf v})
=
f({\bf x}+(n-1){\bf v}+{\bf v})
\leq
f({\bf x}+(n-1){\bf v})
\leq
f({\bf x}),
\]
using the inductive hypothesis that
$(n-1){\bf v}\in\resc{f}$.
Now let $\lambda>0$, and let $n$ be any integer greater than
$\lambda$.
Since $\lambda{\bf v}$ is a convex combination of ${\bf 0}$ and $n{\bf v}$, both
of which are in $\resc{f}$, $\lambda{\bf v}$ must also be in $\resc{f}$,
since we already argued $\resc{f}$ is convex.
Therefore, $\resc{f}$ is a cone.
\underline{Closed:}
Assume $f$ is lower semicontinuous.
Let $\seq{{\bf v}_t}$ be any convergent sequence in $\resc{f}$, and
suppose its limit is ${\bf v}$.
Then for any ${\bf x}\in{\R^n}$,
\[
f({\bf x})
\geq
\liminf f({\bf x}+{\bf v}_t)
\geq
f({\bf x}+{\bf v})
\]
since ${\bf x}+{\bf v}_t\rightarrow{\bf x}+{\bf v}$ and $f$ is lower semicontinuous.
Thus, ${\bf v}\in\resc{f}$, so $\resc{f}$ is closed.
\qed
If ${\bf v}\in\resc{f}$, then the minimum of the reduction $g$
is the same as the minimum of $f$,
suggesting that $f$ can be minimized by first minimizing $g$ and
then adjusting the resulting solution appropriately.
Later, in Section~\ref{sec:minimizers},
we will develop ideas along these lines
which constructively characterize the minimizers of
$f$ by defining and recursively minimizing an astronic reduction.
Here is a statement of some simple properties of such reductions:
\begin{proposition} \label{pr:d2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Let ${\bf v}\in\resc{f}$,
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
\begin{enumerate}
\item \label{pr:d2:a}
For all ${\bf x}\in{\R^n}$, $g({\bf x}) = g({\xx^\bot})$, where ${\xx^\bot}$ is the
projection of ${\bf x}$ perpendicular to ${\bf v}$.
\item \label{pr:d2:b}
For all ${\bf x}\in{\R^n}$, $g({\bf x}) \leq f({\bf x})$.
\item \label{pr:d2:c}
$\inf g = \inf f$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:d2:a}) follows from
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\xx^\bot}$,
by Proposition~\ref{pr:h:5}(\ref{pr:h:5g}).
Part~(\ref{pr:d2:b}):
Since ${\bf v}$ is in the cone $\resc{f}$, for any $t>0$,
\[ f({\bf x}) \geq f({\bf x} + t {\bf v}). \]
Further, the sequence $\seq{{\bf x}+t{\bf v}}$
converges to $\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$
(by Proposition~\ref{pr:aff-cont} applied to the affine map
$\overline{{\bf z}}\mapsto{\bf x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$),
so
\[ f({\bf x}) \geq \liminf f({\bf x}+t{\bf v})
\geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}) = g({\bf x}).
\]
Part~(\ref{pr:d2:c}):
$g({\bf x})$ is the limit of a sequence of
values of $f$, by Proposition~\ref{pr:d1},
so $g({\bf x})\geq \inf f$ for all ${\bf x}$, and so
$\inf g \geq \inf f$.
On the other hand, by part~(\ref{pr:d2:b}),
$\inf g \leq \inf f$.
\qed
The reduction $g=\fshad{\limray{\vv}}$ has its own extension $\overline{g}$.
The next theorem shows how the values of $\overline{g}$ relate to those of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\begin{theorem} \label{thm:d4}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Let ${\bf v}\in\resc{f}$,
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Then for all $\overline{\xx}\in\extspac{n}$ (with projection $\xbar^{\bot}$
perpendicular to ${\bf v}$):
\begin{enumerate}
\item \label{thm:d4:a}
$\overline{g}(\xbar^{\bot}) = \overline{g}(\overline{\xx})$;
\item \label{thm:d4:b}
$\overline{g}(\overline{\xx})\leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$;
\item \label{thm:d4:c}
$\overline{g}(\overline{\xx})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})$.
\end{enumerate}
Summarizing, for all $\overline{\xx}\in\extspac{n}$,
\[
\inf g = \inf f
\leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot})
=\overline{g}(\xbar^{\bot})
= \overline{g}(\overline{\xx})
\leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}).
\]
\end{theorem}
\proof
~
Part~(\ref{thm:d4:a}):
By Proposition~\ref{pr:d1}, there exists a
sequence $\seq{{\bf x}_t}$ in ${\R^n}$ with ${\bf x}_t\rightarrow\overline{\xx}$ and
$g({\bf x}_t)\rightarrow\overline{g}(\overline{\xx})$.
Then
\[
\overline{g}(\overline{\xx})
=
\lim g({\bf x}_t)
=
\lim g({\xx_t^\bot})
\geq
\overline{g}(\xbar^{\bot}).
\]
The second equality is by Proposition~\ref{pr:d2}(\ref{pr:d2:a}),
and the inequality is because ${\xx_t^\bot}\rightarrow\xbar^{\bot}$
by Proposition~\ref{pr:h:5}(\ref{pr:h:5b}).
For the reverse inequality, by Proposition~\ref{pr:d1}, there exists a
sequence $\seq{{\bf x}'_t}$ in ${\R^n}$ with
${\bf x}'_t\rightarrow\xbar^{\bot}$ and
$g({\bf x}'_t)\rightarrow\overline{g}(\xbar^{\bot})$.
For each $t$, let $b_t=t+\norm{{\bf x}'_t}^2$, and let
${\bf x}_t={\bf x}'_t+b_t {\bf v}$.
Note that the values $b_t$ are growing to infinity much faster than
$\norm{{\bf x}'_t}$ (so that $b_t/\norm{{\bf x}'_t}\rightarrow+\infty$).
As a result, ${\bf x}_t\rightarrow\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}=\overline{\xx}$, so
\[
\overline{g}(\overline{\xx})
\leq
\liminf g({\bf x}_t)
=
\liminf g({\bf x}'_t)
=
\overline{g}(\xbar^{\bot})
\]
where the first equality is by Proposition~\ref{pr:d2}(\ref{pr:d2:a})
(since ${\xx_t^\bot}={\bf x}'^\bot_t$).
Part~(\ref{thm:d4:b}):
By Proposition~\ref{pr:d1},
there exists a sequence $\seq{{\bf x}_t}$ in ${\R^n}$ with
${\bf x}_t\rightarrow\overline{\xx}$ and
$f({\bf x}_t)\rightarrow{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Then
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
=
\lim f({\bf x}_t)
\geq
\liminf g({\bf x}_t)
\geq
\overline{g}(\overline{\xx}),
\]
where the first inequality is by
Proposition~\ref{pr:d2}(\ref{pr:d2:b}).
Part~(\ref{thm:d4:c}):
First,
\begin{eqnarray}
\overline{g}(\overline{\xx})
=
\overline{g}(\xbar^{\bot})
=
\overline{g}\paren{(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})^{\bot}}
&=&
\overline{g}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
\label{eq:d5:a}
\\
&\leq&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}).
\label{eq:d5:b}
\end{eqnarray}
\eqref{eq:d5:a} is by part~(\ref{thm:d4:a}), since
$(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})^{\bot} = \xbar^{\bot}$
(by Proposition~\ref{pr:h:5}(\ref{pr:h:5c})).
\eqref{eq:d5:b} is by part~(\ref{thm:d4:b}).
To show the reverse inequality, by Proposition~\ref{pr:d1},
there exists a sequence $\seq{{\bf x}_t}$
in ${\R^n}$ with ${\bf x}_t\rightarrow\overline{\xx}$ and
$g({\bf x}_t)\rightarrow\overline{g}(\overline{\xx})$.
For each $t$, let $\overline{{\bf y}}_t=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}_t$.
Then $\overline{{\bf y}}_t\rightarrow\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$
(by Proposition~\ref{pr:aff-cont} applied to the affine map
$\overline{{\bf z}}\mapsto\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$).
Thus,
\[
\overline{g}(\overline{\xx})
=
\lim g({\bf x}_t)
=
\lim {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x}_t)
=
\lim {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}_t)
\geq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx})
\]
where the last inequality uses Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2a}).
\qed
The next theorem proves that the reduction $g=\fshad{\limray{\vv}}$
is convex and lower
semicontinuous, assuming ${\bf v}$ is in $f$'s recession cone.
The theorem also relates $g$ to another function
${\tilde{g}}:{\R^n}\rightarrow\overline{\R}$ that is given by
\begin{equation} \label{eqn:gtil-defn}
{\tilde{g}}({\bf x})=\inf_{\lambda\in{\mathbb{R}}} f({\bf x}+\lambda {\bf v})
\end{equation}
for ${\bf x}\in{\R^n}$.
This function can be viewed as a kind of ``shadow'' of $f$ in the
positive and negative direction of ${\bf v}$.
Specifically, the theorem shows that $g$ is the
{lower semicontinuous hull} of ${\tilde{g}}$.
For instance,
for the example function given in \eqref{eqn:recip-fcn-eg} with
${\bf v}={\bf e}_1$,
\[
{\tilde{g}}({\bf x})
=
{\tilde{g}}(x_1,x_2)
=
\left\{ \begin{array}{cl}
0 & \mbox{if $x_2 > 0$} \\
+\infty & \mbox{otherwise,}
\end{array} \right.
\]
which differs from $g$, as given in \eqref{eqn:recip-fcn-eg:g}, only
when $x_2=0$.
Evidently, $g=\lsc\ {\tilde{g}}$ in this case, and in general.
\begin{theorem} \label{thm:a10-nunu}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Let ${\bf v}\in\resc{f}$,
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Also, let ${\tilde{g}}:{\R^n}\rightarrow\overline{\R}$ be defined by
\[ {\tilde{g}}({\bf x})=\inf_{\lambda\in{\mathbb{R}}} f({\bf x}+\lambda {\bf v}) \]
for ${\bf x}\in{\R^n}$.
Then ${\tilde{g}}$ is convex.
Furthermore, $g$ is the lower semicontinuous hull of ${\tilde{g}}$,
that is, $g=\lsc\ {\tilde{g}}$.
Therefore $g$ is convex and lower semicontinuous.
\end{theorem}
\proof
Since $f$ is convex, $f({\bf x}+\lambda{\bf v})$, as a function of
$({\bf x},\lambda)$, is also convex.
Therefore, ${\tilde{g}}$ is convex as well.
Both of these facts follow from \citet[Theorem 5.7]{ROC}.
Also, because ${\bf v}\in\resc{f}$, and because $\resc{f}$ is a cone,
$f({\bf x}+\lambda{\bf v})$ is nonincreasing as a function of $\lambda\in{\mathbb{R}}$,
for any fixed ${\bf x}\in{\R^n}$.
Therefore,
\begin{equation} \label{eq:j:4}
{\tilde{g}}({\bf x}) = \lim_{\lambda\rightarrow\infty} f({\bf x}+\lambda{\bf v}).
\end{equation}
Let $h=\lsc {\tilde{g}}$.
We will prove that $g=h$.
Let ${\bf x}\in{\R^n}$, and let $\seq{{\bf x}_t}$ be any sequence in ${\R^n}$ with
${\bf x}_t\rightarrow{\bf x}$.
For each $t$,
let $b_t\in{\mathbb{R}}$ be chosen to be so large that
$b_t > t$, and also so that
\[ f({\bf x}_t + b_t {\bf v}) < {\tilde{g}}({\bf x}_t) + \frac{1}{t}. \]
This is possible by \eqref{eq:j:4}.
Then ${\bf x}_t+b_t{\bf v}\rightarrow \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$
(since $b_t\rightarrow+\infty$ while ${\bf x}_t\rightarrow{\bf x}$), so
\begin{eqnarray*}
g({\bf x})
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\\
&\leq&
\liminf f({\bf x}_t + b_t{\bf v})
\\
&\leq&
\liminf
\paren{{\tilde{g}}({\bf x}_t) + \frac{1}{t}}
\\
&=&
\liminf {\tilde{g}}({\bf x}_t).
\end{eqnarray*}
Since this holds for every such sequence $\seq{{\bf x}_t}$ converging
to ${\bf x}$, it follows that $g({\bf x})\leq h({\bf x})$.
To prove the reverse inequality, let ${\bf x}\in{\R^n}$, and let ${\xx^\bot}$ be
its projection perpendicular to ${\bf v}$.
Then ${\bf x}={\xx^\bot}+b{\bf v}$ for some $b\in{\mathbb{R}}$.
By Proposition~\ref{pr:d1}, there exists a sequence $\seq{{\bf y}_t}$ in
${\R^n}$ such that
${\bf y}_t \rightarrow \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$
and
$f({\bf y}_t) \rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})$.
Further, by Proposition~\ref{pr:h:5},
${\yy_t^\bot} \rightarrow (\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})^{\bot} = {\xx^\bot}$.
Also, let ${\bf z}_t={\yy_t^\bot} + b{\bf v}$ so that
${\bf z}_t\rightarrow {\xx^\bot} + b{\bf v} = {\bf x}$.
Then
\begin{eqnarray*}
g({\bf x})
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\\
&=&
\lim f({\bf y}_t)
\\
&\geq&
\liminf {\tilde{g}}({\yy_t^\bot})
\\
&=&
\liminf {\tilde{g}}({\bf z}_t)
\\
&\geq&
h({\bf x}).
\end{eqnarray*}
Here we used $f({\bf y}_t) \geq {\tilde{g}}({\yy_t^\bot}) = {\tilde{g}}({\bf z}_t)$
by definition of ${\tilde{g}}$ (\eqref{eqn:gtil-defn}), and
since ${\bf y}_t$ and ${\bf z}_t$ both differ from ${\yy_t^\bot}$ by scalar
multiples of ${\bf v}$.
Since the lower semicontinuous hull of a convex function is also
convex, it follows
that $g$ is convex and lower semicontinuous
(cf. \citep[Page 52]{ROC} and \citep[pages 78--81]{HULL}).
\qed
\section{Conjugacy}
\label{sec:conjugacy}
Recall that the \emph{conjugate} of a convex function
$f:{\R^n}\rightarrow\overline{\R}$ is the function
\begin{equation} \label{eq:fstar-def}
{f^*}({\bf u}) = \sup_{{\bf x}\in{\R^n}} ({\bf x}\cdot{\bf u} - f({\bf x}))
\end{equation}
for ${\bf u}\in{\R^n}$, a centrally important notion in standard convex
analysis.
The conjugate ${f^*}$ encodes exactly those affine functions that are
majorized by $f$ in the sense that any
pair
$\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}$
is in the epigraph of ${f^*}$
if and only if
$v\geq {\bf x}\cdot{\bf u} - f({\bf x})$ for all ${\bf x}\in{\R^n}$,
that is,
if and only if
$f({\bf x})\geq {\bf x}\cdot{\bf u} - v$
so that $f$ majorizes the affine function
${\bf x} \mapsto {\bf x}\cdot{\bf u} - v$.
Moreover,
as is well known,
functions $f$ that are closed and convex are equal to their own double
conjugate, that is, the conjugate of ${f^*}$,
so that
$f=({f^*})^* = f^{**}$.
This means that $f$ can be represented as the supremum over all affine
functions that it majorizes, and furthermore,
that the original function $f$ can be fully reconstructed from the dual
representation afforded by ${f^*}$.
In this section, we will see how these very powerful ideas
can be extended to astral space.
\subsection{Conjugates and double conjugates}
\label{sec:conjugacy-def}
Let $F:\extspac{n}\rightarrow\overline{\R}$.
For now, we allow $F$ to be any function defined on the
astral space, although later we will focus especially on the case that
$F$ is the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of some convex function $f$.
How can the definition of the conjugate ${f^*}$ given in
\eqref{eq:fstar-def} be extended to a function $F$?
The most natural idea is to simply replace $f$ by $F$ and ${\bf x}$ by
$\overline{\xx}$ so that the expression inside the supremum
(now over $\overline{\xx}\in\extspac{n}$) becomes
$\overline{\xx}\cdot{\bf u} - F(\overline{\xx})$.
The problem, of course, is that
$\overline{\xx}\cdot{\bf u}$ and $F(\overline{\xx})$ might both be $+\infty$ (or both
$-\infty$) so that this expression is
undefined, being equal to the sum of $+\infty$ and $-\infty$.
To address this, we can re-express the conjugate in a way
that generalizes more easily.
In particular, since the epigraph of $f$ consists of all pairs
$\rpair{{\bf x}}{y}$ in ${\R^n}\times{\mathbb{R}}$ with $y\geq f({\bf x})$,
we can rewrite \eqref{eq:fstar-def} as
\begin{equation} \label{eq:fstar-mod-def}
{f^*}({\bf u}) = \sup_{\rpair{{\bf x}}{y}\in\epi f} ({\bf x}\cdot{\bf u} - y).
\end{equation}
This expression generalizes directly to
astral space by making the simple substitutions suggested earlier, and
specifically replacing ${\bf x}\cdot{\bf u}$ with its astral
analogue $\overline{\xx}\cdot{\bf u}$.
Thus,
for $F:\extspac{n}\rightarrow\overline{\R}$, we define the \emph{conjugate}
${F^*}:{\R^n}\rightarrow\overline{\R}$ to be
\begin{equation} \label{eq:Fstar-def}
{F^*}({\bf u}) = \sup_{\rpair{\overline{\xx}}{y}\in\epi F} (\overline{\xx}\cdot{\bf u} - y)
\end{equation}
for ${\bf u}\in{\R^n}$.
Although $\overline{\xx}\cdot{\bf u}$ might be infinite, $y$ is in ${\mathbb{R}}$ always, so
the earlier issue is entirely avoided.
The resulting conjugate function ${F^*}$ is always convex:
\begin{proposition} \label{pr:conj-is-convex}
Let $F:\extspac{n}\rightarrow\overline{\R}$.
Then its conjugate, ${F^*}$, is convex.
\end{proposition}
\proof
For any fixed $\overline{\xx}\in\extspac{n}$ and $y\in{\mathbb{R}}$,
$\overline{\xx}\cdot{\bf u}$, viewed as a function of ${\bf u}\in{\R^n}$,
is convex by Theorem~\ref{thm:h:5},
so $\overline{\xx}\cdot{\bf u}-y$ is convex as well.
Therefore, ${F^*}$ is convex since it is the pointwise supremum of
convex functions
\cite[Theorem~5.5]{ROC}.
\qed
The definition for ${F^*}$ given in \eqref{eq:Fstar-def} can be
rewritten using a form of addition that handles the sum of $+\infty$
and $-\infty$ differently than in leftward addition.
Specifically, the \emph{downward} sum of $\bar{x}, \bar{y}\in\overline{\R}$, denoted
$\bar{x} \plusd \bar{y}$, is the same as ordinary addition, except that if
\emph{either} $\bar{x}$ or $\bar{y}$ is $-\infty$, then their downward sum is also
defined to be $-\infty$.
Thus,
\begin{equation} \label{eq:down-add-def}
\bar{x} \plusd \bar{y} = \left\{ \begin{array}{cl}
-\infty & \mbox{if $\bar{x}=-\infty$ or $\bar{y}=-\infty$}\\
\bar{x} + \bar{y} & \mbox{otherwise.}
\end{array}\right.
\end{equation}
Downward addition is the same as ordinary addition
in all cases except when adding $-\infty$ and $+\infty$.
The next proposition summarizes this and other properties.
Part~(\ref{pr:plusd-props:d}) is especially useful for simplifying
expressions like \eqref{eq:Fstar-def}.
\begin{proposition} \label{pr:plusd-props}
Let $\bar{a},\bar{b},\bar{c}\in\overline{\R}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:plusd-props:a}
$\bar{a}\plusd \bar{b} = \bar{b}\plusd \bar{a}$.
\item \label{pr:plusd-props:b}
$(\bar{a}\plusd \bar{b}) \plusd \bar{c} = \bar{a}\plusd (\bar{b} \plusd \bar{c})$.
\item \label{pr:plusd-props:c}
If $\bar{a}$ and $\bar{b}$ are summable,
then
$\bar{a}\plusd \bar{b} = \bar{a}+\bar{b}$.
\item \label{pr:plusd-props:d}
$\sup \braces{\bar{a} - y : y\in{\mathbb{R}}, y\geq \bar{b}} = -\bar{b} \plusd \bar{a}$.
\item \label{pr:plusd-props:e}
$\bar{a} \geq \bar{b} \plusd \bar{c}$
if and only if
$-\bar{b} \geq -\bar{a} \plusd \bar{c}$.
\end{enumerate}
\end{proposition}
\proof
Part~(\ref{pr:plusd-props:d}) can be checked by considering
the separate cases that
$\bar{b}=+\infty$ (noting $\sup\emptyset = -\infty$);
$\bar{b}<+\infty$ and $\bar{a}\in\{-\infty,+\infty\}$;
and
$\bar{b}<+\infty$ and $\bar{a}\in{\mathbb{R}}$.
The other parts can be checked in a similar fashion.
\qed
Using Proposition~\ref{pr:plusd-props}(\ref{pr:plusd-props:d}),
we can rewrite \eqref{eq:Fstar-def} as
\begin{eqnarray}
{F^*}({\bf u})
&=& \sup_{\overline{\xx}\in\extspac{n}}
\brackets{
\sup \braces{\overline{\xx}\cdot{\bf u} - y : y\in{\mathbb{R}}, y\geq F(\overline{\xx})}
}
\nonumber
\\
&=& \sup_{\overline{\xx}\in\extspac{n}} \Parens{ - F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u} }.
\label{eq:Fstar-down-def}
\end{eqnarray}
(This same argument also shows more explicitly that
\eqref{eq:fstar-def} and \eqref{eq:fstar-mod-def} are equal.)
In this form, the conjugate ${F^*}$
(as well as the dual conjugate to be defined momentarily)
is the same as that defined by
\citet[Definition~8.2]{singer_book}
in his abstract treatment of convex analysis,
which can be instantiated to the astral setting
by letting his variables
$X$ and $W$ be equal to $\extspac{n}$ and ${\R^n}$, respectively,
and defining his function $\varphi$,
called the \emph{coupling function}, to be
$\varphi(\overline{\xx},{\bf u})=\overline{\xx}\cdot{\bf u}$ for $\overline{\xx}\in\extspac{n}$ and
${\bf u}\in{\R^n}$.
The equivalence with \eqref{eq:Fstar-def} is also mentioned by
\citet[Eq.~(8.34)]{singer_book}.
This general form of conjugate was originally presented by
\citet[Eq.~(14.7)]{moreau__convexity}.
As an example, consider the affine function
$F(\overline{\xx})=\overline{\xx}\cdot{\bf w}+b$, for
$\overline{\xx}\in\extspac{n}$, where ${\bf w}\in{\R^n}$ and $b\in{\mathbb{R}}$.
Then it can be checked that this function's conjugate is
\[
{F^*}({\bf u})
= \left\{ \begin{array}{cl}
-b & \mbox{if ${\bf u}={\bf w}$}\\
+\infty & \mbox{otherwise.}
\end{array} \right.
\]
This can be shown directly using
\eqref{eq:Fstar-down-def}.
Alternatively, since $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, where
$f({\bf x})={\bf x}\cdot{\bf w}+b$ is the corresponding function on ${\R^n}$,
we can apply
Proposition~\ref{pr:fextstar-is-fstar},
proved below, which implies in this case that the astral conjugate
${F^*}$ is the same as ${f^*}$, the standard conjugate of $f$.
As another example, for any $\overline{{\bf z}}\in\extspac{n}$, let us define the
one-point indicator function
$\indfa{\overline{{\bf z}}}:\extspac{n}\rightarrow\overline{\R}$
by
\begin{equation} \label{eq:indfa-zbar-defn}
\indfa{\overline{{\bf z}}}(\overline{\xx})
= \left\{ \begin{array}{cl}
0 & \mbox{if $\overline{\xx}=\overline{{\bf z}}$}\\
+\infty & \mbox{otherwise}
\end{array} \right.
\end{equation}
for $\overline{\xx}\in\extspac{n}$.
This function's conjugate is the function
$\indfastar{\overline{{\bf z}}}({\bf u})=\overline{{\bf z}}\cdot{\bf u}$, for ${\bf u}\in{\R^n}$;
that is, $\indfastar{\overline{{\bf z}}}=\ph{\overline{{\bf z}}}$, as defined in
\eqref{eq:ph-xbar-defn}.
Thus,
whereas the standard conjugate of any function
on ${\R^n}$
is always closed and convex,
this example shows that the astral conjugate
of a function
defined on $\extspac{n}$ might not be closed or even lower
semicontinuous (although it must be convex by
Proposition~\ref{pr:conj-is-convex}).
For example, if $n=1$ and $\bar{z}=+\infty$, then
$\indfastar{\bar{z}}(u)=\limray{u}$, for $u\in{\mathbb{R}}$,
which
is not proper or closed or lower semicontinuous.
Returning to our general development,
the conjugate operation $F\mapsto {F^*}$ maps a function defined on
$\extspac{n}$ to one defined on ${\R^n}$.
We next define a {dual conjugate} operation that maps in the
reverse direction, from functions on ${\R^n}$ to functions on
$\extspac{n}$.
In standard convex analysis, both a function $f$ and its conjugate
${f^*}$ are defined on ${\R^n}$ so that the same conjugate operation can
be used in either direction.
But in the astral setting, as is the case more generally in the
abstract setting of~\citet{singer_book}, a different dual conjugate is required.
This asymmetry exactly reflects that the core astral operation
$\overline{\xx}\cdot{\bf u}$ is defined with $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$
belonging to different spaces.
Let $\psi: {\R^n}\rightarrow\overline{\R}$ be any function on ${\R^n}$.
Of course, $\psi$ is just an ordinary function, but we use this
notation to emphasize that we think of $\psi$ as operating on
the dual variable ${\bf u}\in{\R^n}$; later, we will often take $\psi$ to
itself be the conjugate of some other function.
By direct analogy with the preceding definition of ${F^*}$,
we define the \emph{dual conjugate} of $\psi$ to be the function
$\psi^{\dualstar}:\extspac{n}\rightarrow\overline{\R}$
defined by
\begin{eqnarray}
\psi^{\dualstar}(\overline{\xx})
&=&
\sup_{\rpair{{\bf u}}{v}\in\epi \psi} (\overline{\xx}\cdot{\bf u} - v)
\label{eq:psistar-def}
\\
&=& \sup_{{\bf u}\in{\R^n}} \Parens{ - \psi({\bf u}) \plusd \overline{\xx}\cdot{\bf u} },
\label{eq:psistar-def:2}
\end{eqnarray}
for $\overline{\xx}\in\extspac{n}$
(again using
Proposition~\ref{pr:plusd-props}(\ref{pr:plusd-props:d})).
We use the notation $\psi^{\dualstar}$ rather than $\psi^{*}$ because
the latter denotes the standard conjugate of $\psi$
(so $\psi^{*}$ is a function defined on ${\R^n}$ while $\psi^{\dualstar}$ is
defined on $\extspac{n}$).
We will be especially interested in the \emph{double conjugate}
of a function $F:\extspac{n}\rightarrow\overline{\R}$,
that is, $F^{*\dualstar}=({F^*})^{\bar{*}}$,
the dual conjugate of the conjugate of $F$.
In standard convex analysis, the double conjugate $f^{**}$ for a
function $f:{\R^n}\rightarrow\overline{\R}$ is equal to its closure, $\cl f$, if
$f$ is convex (where $\cl f = \lsc f$ if $f>-\infty$, and
$\cl f\equiv-\infty$ otherwise).
Thus, as already discussed, if $f$ is closed and convex, then
$f=f^{**}$.
Furthermore, $f^{**}$ is in general
equal to the pointwise supremum over all
affine functions that are majorized by $f$.
An analogous result holds in the astral setting, as we show now.
For ${\bf u}\in{\R^n}$ and $v\in{\mathbb{R}}$, let
$\ell_{\uu,v}(\overline{\xx})=\overline{\xx}\cdot{\bf u}-v$, for $\overline{\xx}\in\extspac{n}$.
This is the extension of the affine
function ${\bf x}\mapsto{\bf x}\cdot{\bf u}-v$ on ${\R^n}$, and is also a
special case of the affine maps considered in
Section~\ref{sec:linear-maps}.
Then $F^{*\dualstar}$ is exactly the pointwise supremum over all such functions
that are majorized by $F$, as we show in the next
theorem, which is a special case of \citet[Theorem~8.5]{singer_book}.
\begin{theorem} \label{thm:fdub-sup-afffcns}
Let $F:\extspac{n}\rightarrow\overline{\R}$, and let
$\ell_{\uu,v}(\overline{\xx})=\overline{\xx}\cdot{\bf u}-v$, for all $\overline{\xx}\in\extspac{n}$,
and for ${\bf u}\in{\R^n}$ and $v\in{\mathbb{R}}$.
Then for $\overline{\xx}\in\extspac{n}$,
\begin{equation} \label{eq:thm:fdub-sup-afffcns:1}
F^{*\dualstar}(\overline{\xx}) = \sup\braces{\ell_{\uu,v}(\overline{\xx}) : {\bf u}\in{\R^n}, v\in{\mathbb{R}}, \ell_{\uu,v}\leq F}.
\end{equation}
Consequently, $F\geq F^{*\dualstar}$, and
furthermore, $F=F^{*\dualstar}$ if and only if $F$ is the pointwise supremum
of some collection of affine functions $\ell_{\uu,v}$.
\end{theorem}
\proof
Let ${\bf u}\in{\R^n}$ and $v\in{\mathbb{R}}$.
Then
$v\geq {F^*}({\bf u})$ if and only if
$\overline{\xx}\cdot{\bf u} - y \leq v$ for all $\rpair{\overline{\xx}}{y}\in \epi F$
(by \eqref{eq:Fstar-def}),
which in turn holds
if and only if
$\overline{\xx}\cdot{\bf u} - v \leq F(\overline{\xx})$ for all $\overline{\xx}\in\extspac{n}$.
In other words,
$\rpair{{\bf u}}{v}\in\epi {F^*}$ if and only if
$\ell_{\uu,v}\leq F$.
The result therefore follows directly from \eqref{eq:psistar-def} with
$\psi={F^*}$.
That $F\geq F^{*\dualstar}$ always is now immediate, as is the claim that if
$F=F^{*\dualstar}$ then $F$ is the pointwise supremum over a collection of
affine functions.
For the converse of the latter statement, suppose, for some set
$A\subseteq {\R^n}\times{\mathbb{R}}$, that
$F(\overline{\xx}) = \sup_{\rpair{{\bf u}}{v}\in A} \ell_{\uu,v}(\overline{\xx})$
for $\overline{\xx}\in\extspac{n}$.
If $\rpair{{\bf u}}{v}\in A$, then this implies $\ell_{\uu,v}\leq F$,
so that $\ell_{\uu,v}\leqF^{*\dualstar}$ by \eqref{eq:thm:fdub-sup-afffcns:1}.
Since this is true for every such pair in $A$, it follows that
$F\leqF^{*\dualstar}$.
\qed
We mainly focus on the double conjugates $F^{*\dualstar}$
(and later, $f^{*\dualstar}$, where $f$ is defined over ${\R^n}$).
But it is also possible to form a dual form of double conjugate
$\psi^{\dualstar*}=(\psi^{\bar{*}})^*$ from a function
$\psi:{\R^n}\rightarrow\overline{\R}$ by applying the conjugate operations in the
reverse order.
Analogous properties to those shown in
Theorem~\ref{thm:fdub-sup-afffcns} apply as well to $\psi^{\dualstar*}$.
We prove just one of these, which will be used at a later point.
Even though the proof is largely symmetric to the one in
Theorem~\ref{thm:fdub-sup-afffcns}, we include it here
for completeness:
\begin{theorem} \label{thm:psi-geq-psidub}
Let $\psi:{\R^n}\rightarrow\overline{\R}$.
Then $\psi\geq\psi^{\dualstar*}$.
\end{theorem}
\proof
Let ${\bf u}\in{\R^n}$ and suppose $v\geq \psi({\bf u})$, with $v\in{\mathbb{R}}$.
For any point $\rpair{\overline{\xx}}{y}\in\epi \psi^{\dualstar}$,
we have
\[
y
\geq
\psi^{\dualstar}(\overline{\xx})
\geq
\overline{\xx}\cdot{\bf u} - v,
\]
where the second inequality is by \eqref{eq:psistar-def}.
Thus, $v \geq \overline{\xx}\cdot{\bf u} - y$.
Since this holds for all
$\rpair{\overline{\xx}}{y}\in\epi \psi^{\dualstar}$,
it follows that
$v \geq \psi^{\dualstar*}({\bf u})$,
by \eqref{eq:Fstar-def}.
And since this holds for all $v\geq\psi({\bf u})$,
we conclude $\psi({\bf u})\geq \psi^{\dualstar*}({\bf u})$.
\qed
For example, for $\overline{{\bf z}}\in\extspac{n}$, the dual conjugate of the function
$\ph{\overline{{\bf z}}}$ defined in \eqref{eq:ph-xbar-defn}
is
\[
\phstar{\overline{{\bf z}}}(\overline{\xx})
=
\sup_{{\bf u}\in{\R^n}} \Parens{ - \overline{{\bf z}}\cdot{\bf u} \plusd \overline{\xx}\cdot{\bf u} }
=
\indfa{\overline{{\bf z}}}(\overline{\xx}),
\]
for $\overline{\xx}\in\extspac{n}$,
with $\indfa{\overline{{\bf z}}}$ as in
\eqref{eq:indfa-zbar-defn}.
To see this, note that $-\overline{{\bf z}}\cdot{\bf u}\plusd\overline{{\bf z}}\cdot{\bf u}$ is equal to
$0$ if $\overline{{\bf z}}\cdot{\bf u}\in{\mathbb{R}}$ (including if ${\bf u}={\bf 0}$), and otherwise
is equal to $-\infty$;
thus, $\phstar{\overline{{\bf z}}}(\overline{{\bf z}})=0$.
And if $\overline{\xx}\neq\overline{{\bf z}}$ then, for some ${\bf u}\in{\R^n}$,
$\overline{\xx}\cdot{\bf u}\neq\overline{{\bf z}}\cdot{\bf u}$
(by Proposition~\ref{pr:i:4});
as a result,
$- \overline{{\bf z}}\cdot(\lambda{\bf u}) \plusd \overline{\xx}\cdot(\lambda{\bf u})
= \lambda(\overline{\xx}\cdot{\bf u} - \overline{{\bf z}}\cdot{\bf u})$
can be made arbitrarily large for appropriate choices of $\lambda\in{\mathbb{R}}$.
Thus,
$\indfa{\overline{{\bf z}}}$ and $\ph{\overline{{\bf z}}}$ are conjugates of each other,
so
$\indfadub{\overline{{\bf z}}}=\indfa{\overline{{\bf z}}}$
and $\phdub{\overline{{\bf z}}}=\ph{\overline{{\bf z}}}$.
Furthermore, if $\overline{{\bf z}}$ is not in ${\R^n}$,
then $\ph{\overline{{\bf z}}}$'s standard conjugate is
$\phstars{\overline{{\bf z}}}\equiv+\infty$ (since it can be shown using
Proposition~\ref{pr:i:3}
that $\ph{\overline{{\bf z}}}({\bf u})=-\infty$ for some ${\bf u}\in{\R^n}$).
As a result, the standard conjugate operation
$\phstars{\overline{{\bf z}}}$ entirely erases the identity of $\overline{{\bf z}}$.
On the other hand, from the astral dual conjugate
$\phstar{\overline{{\bf z}}}$, we have just seen that
it is possible to reconstruct
$\ph{\overline{{\bf z}}}=\phdub{\overline{{\bf z}}}$, and so also to recover $\overline{{\bf z}}$.
In general,
this shows that the astral dual conjugate $\psi^{\dualstar}$ can
retain more information about the original function
$\psi:{\R^n}\rightarrow\overline{\R}$
than the standard conjugate $\psi^{*}$.
Next,
we turn to the particular case that $F$ is the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of
some function $f$ defined on ${\R^n}$.
In this case, the conjugate of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, denoted
${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}=({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'})^*$,
is the same as the
conjugate of $f$:
\begin{proposition} \label{pr:fextstar-is-fstar}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}={f^*}$.
\end{proposition}
\proof
If $\rpair{{\bf x}}{y}\in\epi f$, then $y\geq f({\bf x})\geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})$
(by Proposition~\ref{pr:h:1}(\ref{pr:h:1a})), so
$\rpair{{\bf x}}{y}\in\epi {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
That is, $\epi f \subseteq \epi{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Comparing \eqref{eq:fstar-mod-def}
and \eqref{eq:Fstar-def} (with $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$),
it therefore follows that ${f^*}\leq{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}$.
For the reverse inequality, let ${\bf u}\in{\R^n}$ and
let $\rpair{\overline{\xx}}{y}$ be any point in
$\epi {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ so that $y\geq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Let $\epsilon>0$.
By Proposition~\ref{pr:d1}, there exists a sequence $\seq{{\bf x}_t}$ in
${\R^n}$ which converges to $\overline{\xx}$ and with
$f({\bf x}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Then for sufficiently large $t$, $f({\bf x}_t)<y+\epsilon$; by discarding
all other elements, assume that this holds for all $t$.
Then $\rpair{{\bf x}_t}{y+\epsilon}\in\epi f$, so
\[
{\bf x}_t\cdot{\bf u} - y - \epsilon \leq {f^*}({\bf u})
\]
by \eqref{eq:fstar-mod-def}.
Since this holds for all $t$,
the limit of the left-hand side must also satisfy this inequality,
yielding
$ \overline{\xx}\cdot{\bf u} - y - \epsilon \leq {f^*}({\bf u})$.
Since this holds for all $\epsilon>0$, it follows that
$ \overline{\xx}\cdot{\bf u} - y \leq {f^*}({\bf u})$.
And since this holds for all $\rpair{\overline{\xx}}{y} \in \epi {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}({\bf u})\leq{f^*}({\bf u})$ by \eqref{eq:Fstar-def}
(again with $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$).
\qed
Applied to the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of a function
$f:{\R^n}\rightarrow\overline{\R}$,
this shows, when combined with \eqref{eq:psistar-def:2}, that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s
double conjugate is
\begin{equation} \label{eqn:fdub-form-lower-plus}
{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*\dualstar}(\overline{\xx})
=
f^{*\dualstar}(\overline{\xx})
=
\sup_{{\bf u}\in{\R^n}} (-{f^*}({\bf u}) \plusd \overline{\xx}\cdot{\bf u})
\end{equation}
for $\overline{\xx}\in\extspac{n}$.
This expression is very close in form to the standard double conjugate
$f^{**}$, and indeed shows that $f^{**}({\bf x})=f^{*\dualstar}({\bf x})$ for all ${\bf x}\in{\R^n}$.
Because leftward addition is so natural for the
astral setting, it will be convenient to further
rewrite this expression using leftward addition rather than
downward addition,
as we show now.
The next proposition provides sufficient conditions
(which will be satisfied if $\psi$ is convex and closed) for
the expression for $\psi^{\dualstar}$ given in \eqref{eq:psistar-def:2} to
continue to hold when downward addition is replaced by leftward
addition:
\begin{proposition} \label{pr:psi-with-plusl}
Let $\psi:{\R^n}\rightarrow\overline{\R}$, and assume that either
$\psi>-\infty$ or that $\psi({\bf 0})=-\infty$.
Then
\begin{equation} \label{eq:pr:psi-with-plusl:1}
\psi^{\dualstar}(\overline{\xx})
=
\sup_{{\bf u}\in{\R^n}} \Parens{ - \psi({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u} }.
\end{equation}
for $\overline{\xx}\in\extspac{n}$.
\end{proposition}
\proof
Note that $x\plusd y = x\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax y$ for all $x,y\in\overline{\R}$, except if
$x=+\infty$ and $y=-\infty$.
Therefore, if $\psi>-\infty$ then
$ - \psi({\bf u}) \plusd \overline{\xx}\cdot{\bf u} = - \psi({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u} $
for all ${\bf u}\in{\R^n}$ and $\overline{\xx}\in\extspac{n}$, proving the claim in this
case.
Otherwise, if $\psi({\bf 0})=-\infty$ then for all $\overline{\xx}\in\extspac{n}$,
\[+\infty = -\psi({\bf 0})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf 0} = -\psi({\bf 0})\plusd \overline{\xx}\cdot{\bf 0},\]
implying that both $\psi^{\dualstar}(\overline{\xx})$ and
the right-hand side of \eqref{eq:pr:psi-with-plusl:1} are equal to
$+\infty$, proving the claim in this case as well.
\qed
The next theorem summarizes results for ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ and its double
conjugate:
\begin{theorem} \label{thm:fext-dub-sum}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then
\begin{enumerate}
\item \label{thm:fext-dub-sum:a}
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}\geq{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*\dualstar} = f^{*\dualstar}$.
\item \label{thm:fext-dub-sum:b}
For all $\overline{\xx}\in\extspac{n}$,
\[
f^{*\dualstar}(\overline{\xx})
=
\sup_{{\bf u}\in{\R^n}} (-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u}).
\]
\end{enumerate}
\end{theorem}
\proof
~
Part~(\ref{thm:fext-dub-sum:a})
follows directly from
Theorem~\ref{thm:fdub-sup-afffcns} (applied to ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$) and
Proposition~\ref{pr:fextstar-is-fstar}.
Part~(\ref{thm:fext-dub-sum:b}):
If $f\equiv+\infty$ then ${f^*}\equiv-\infty$.
Otherwise, if $f({\bf x})<+\infty$ at some point ${\bf x}\in{\R^n}$, then for all
${\bf u}\in{\R^n}$,
${f^*}({\bf u})\geq {\bf x}\cdot{\bf u}-f({\bf x}) > -\infty$.
In either case, ${f^*}$ satisfies the conditions of
Proposition~\ref{pr:psi-with-plusl}, proving the claim.
\qed
Later, beginning in Section~\ref{sec:ent-closed-fcn},
we give necessary and sufficient conditions
(substantially more informative than those given in
Theorem~\ref{thm:fdub-sup-afffcns})
for when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
\subsection{Relation to reductions}
We next explore how conjugacy and astronic reductions relate to one
another, continuing our development of foundational tools and
techniques that will be applied later in the manuscript.
Before proceeding, we pause to recall a few notions from standard
convex analysis, especially as they relate to cones.
If $S\subseteq{\R^n}$ is convex, then $\cone S$ is the set of all
positive scalar multiples of elements of $S$, along with the origin
\cite[Corollary~2.6.3]{ROC}.
Of particular interest will be the cone generated by the effective
domain of the conjugate ${f^*}$.
Since ${f^*}$ is convex, its effective domain is as well, so
for any $f:{\R^n}\rightarrow\overline{\R}$,
\begin{equation} \label{eq:cone-dom-fstar}
\cone(\dom{{f^*}}) = \{{\bf 0}\}
\cup \{ \lambda{\bf u} : \lambda>0, {\bf u}\in \dom{{f^*}} \}.
\end{equation}
If $K\subseteq{\R^n}$ is a nonempty, convex cone,
its \emph{polar},
denoted $\polar{K}$, is the set of points whose inner product with every
point in $K$ is nonpositive, that is,
\begin{equation} \label{eqn:polar-def}
\polar{K} = \braces{{\bf v}\in{\R^n} : \forall {\bf u}\in K, {\bf v}\cdot{\bf u}\leq 0}.
\end{equation}
The set $\polar{K}$ is itself a nonempty, closed (in ${\R^n}$) pointed convex cone.
Also, $\polar{K}=\polar{(\cl K)}$ and $\dubpolar{K}=\cl K$
\cite[Theorem~14.1]{ROC}.
We will repeatedly make use of the next theorem from standard convex
analysis which states that
the recession cone of a closed, proper convex function
is polar to the cone generated by the effective
domain of its conjugate:
\begin{theorem} \label{thm:rescpol-is-con-dom-fstar}
Let $f:{\R^n}\rightarrow\overline{\R}$ be closed, proper and convex.
Then
\begin{equation} \label{eq:thm:rescpol-is-con-dom-fstar:1}
\rescpol{f} = \cl(\cone(\dom{{f^*}})).
\end{equation}
Consequently,
\begin{eqnarray}
\resc{f} &=& \polar{(\cone(\dom{{f^*}}))}
\label{eq:thm:rescpol-is-con-dom-fstar:2}
\\
&=&
\{{\bf v}\in{\R^n} : \forall {\bf u}\in\dom{{f^*}}, {\bf u}\cdot{\bf v}\leq 0\}.
\label{eq:thm:rescpol-is-con-dom-fstar:3}
\end{eqnarray}
\end{theorem}
\proof
\eqref{eq:thm:rescpol-is-con-dom-fstar:1} is extracted directly from
\citet[Theorem~14.2]{ROC}.
Thus, from the facts mentioned above, and since $\resc{f}$ is closed
in ${\R^n}$ (Proposition~\ref{pr:resc-cone-basic-props}),
\[
\resc{f}
=
\rescdubpol{f}
=
\polar{(\cl(\cone(\dom{{f^*}})))}
=
\polar{(\cone(\dom{{f^*}}))},
\]
proving \eqref{eq:thm:rescpol-is-con-dom-fstar:2}.
\eqref{eq:thm:rescpol-is-con-dom-fstar:3} then follows from
\eqref{eq:cone-dom-fstar} (since if ${\bf v}\cdot{\bf u}\leq 0$ for
all ${\bf u}\in \dom{{f^*}}$, then so also ${\bf v}\cdot(\lambda{\bf u})\leq 0$
for all $\lambda>0$).
\qed
Returning to reductions, the main topic of this subsection,
we start by relating the conjugate of a function $f$ to that of its
reduction $g=\fshad{\limray{\vv}}$, a relationship that turns out to be very simple:
\begin{theorem} \label{thm:e1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, proper and closed, and
let ${\bf v}\in\resc{f}$.
Let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$,
and let
\[ L=\{{\bf u}\in{\R^n} : {\bf u}\cdot{\bf v}=0\} \]
be the linear subspace perpendicular to ${\bf v}$.
Then ${f^*}\leq{g^*}$,
$\dom{{g^*}} \subseteq (\dom{{f^*}}) \cap L$,
and
$\cone(\dom{{g^*}}) \subseteq \cone(\dom{{f^*}}) \cap L$.
If, in addition, $g$ is proper (that is, if $g>-\infty$), then
\[
{g^*}({\bf u}) = \left\{ \begin{array}{cl}
{f^*}({\bf u}) & \mbox{if ${\bf u}\cdot{\bf v}=0$}\\
+\infty & \mbox{otherwise}
\end{array} \right.
\]
for all ${\bf u}\in{\R^n}$.
Consequently,
$\dom{{g^*}} = (\dom{{f^*}}) \cap L$
and
$\cone(\dom{{g^*}}) = \cone(\dom{{f^*}}) \cap L$.
\end{theorem}
\proof
First, by Proposition~\ref{pr:d2}(\ref{pr:d2:b}), $g\leq f$.
Therefore, by definition of conjugate (\eqref{eq:fstar-def}),
${f^*}\leq{g^*}$.
Since $f$ is proper and $g\leq f$, it cannot
be the case that $g\equiv+\infty$.
So if $g$ is improper, then there must exist some point ${\bf x}\in{\R^n}$ for
which $g({\bf x})=-\infty$.
This implies ${g^*}\equiv +\infty$, so in this case,
$\dom{{g^*}} = \emptyset \subseteq \dom{{f^*}} \cap L$,
and
$\cone(\dom{{g^*}}) = \{{\bf 0}\} \subseteq \cone(\dom{{f^*}}) \cap L$.
We therefore assume henceforth that $g$ is proper.
We claim that $\dom{{g^*}}\subseteq L$.
To see this, let ${\bf u}\in\dom{{g^*}}$.
For all $\lambda\in{\mathbb{R}}$,
by Proposition~\ref{pr:d2}(\ref{pr:d2:a}),
\[
g({\bf x}+\lambda{\bf v})
=
g(({\bf x}+\lambda{\bf v})^{\bot})
=
g({\xx^\bot})
=
g({\bf x}),
\]
for all ${\bf x}\in{\R^n}$,
since $({\bf x}+\lambda{\bf v})^\bot={\xx^\bot}$
(where ${\xx^\bot}$ is the projection of ${\bf x}$ perpendicular to ${\bf v}$).
Therefore, $\lambda{\bf v}\in\resc{g}$.
Applying
Theorem~\ref{thm:rescpol-is-con-dom-fstar}
to $g$, it follows that
$\lambda ({\bf u}\cdot{\bf v}) = {\bf u}\cdot(\lambda{\bf v})\leq 0$.
Since this holds for all $\lambda\in{\mathbb{R}}$, we must have
${\bf u}\cdot{\bf v}=0$ so ${\bf u}\in L$, as claimed.
Next, we claim that if ${\bf u}\in L$ then
${g^*}({\bf u})\leq{f^*}({\bf u})$.
This is trivial if ${f^*}({\bf u})=+\infty$, so assume henceforth that
${f^*}({\bf u})<+\infty$.
Let ${\bf x}\in{\R^n}$.
By Proposition~\ref{pr:d1}, there exists a sequence $\seq{{\bf y}_t}$ in
${\R^n}$ with ${\bf y}_t\rightarrow\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$ and
$f({\bf y}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=g({\bf x})$.
Then
\begin{eqnarray}
g({\bf x})
&=& {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\nonumber \\
&=& \lim f({\bf y}_t)
\nonumber\\
&\geq& \liminf
\brackets{{\bf y}_t\cdot{\bf u} - {f^*}({\bf u})}
\label{eqn:e1}\\
&=& \liminf
\brackets{{\yy_t^\bot}\cdot{\bf u} - {f^*}({\bf u})}
\label{eqn:e2}\\
&=& {\xx^\bot}\cdot{\bf u} - {f^*}({\bf u})
\label{eqn:e3}\\
&=& {\bf x}\cdot{\bf u} - {f^*}({\bf u}).
\label{eqn:e4}
\end{eqnarray}
\eqref{eqn:e1} is by definition of ${f^*}$.
\eqref{eqn:e2} and \eqref{eqn:e4} are because ${\bf u}\in L$.
\eqref{eqn:e3} is because
${\yy_t^\bot}\rightarrow (\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})^{\bot}={\xx^\bot}$,
by Proposition~\ref{pr:h:5}(\ref{pr:h:5b},\ref{pr:h:5c},\ref{pr:h:5e}).
Thus,
\[
{g^*}({\bf u}) = \sup_{{\bf x}\in{\R^n}} [{\bf x}\cdot{\bf u} - g({\bf x})]
\leq \sup_{{\bf x}\in{\R^n}} [{\bf x}\cdot{\bf u} - ({\bf x}\cdot{\bf u} - {f^*}({\bf u}))]
= {f^*}({\bf u})
\]
as claimed.
Combining, assuming $g$ is proper,
we have that ${g^*}({\bf u})=+\infty$ if
${\bf u}\not\in L$, and that
${f^*}({\bf u})={g^*}({\bf u})$ otherwise.
This immmediately implies
$\dom{{g^*}} = (\dom{{f^*}}) \cap L$.
It remains only to prove the final statement of the theorem in this
case.
Since $\dom{{g^*}}\subseteq \dom{{f^*}}$,
$\cone(\dom{{g^*}})\subseteq\cone(\dom{{f^*}})$.
And since $\dom{{g^*}}\subseteq L$, $\cone(\dom{{g^*}})\subseteq L$
since $L$ is a linear subspace (and using \eqref{eq:cone-dom-fstar}).
Thus, $\cone(\dom{{g^*}})\subseteq\cone(\dom{{f^*}})\cap L$.
For the reverse inclusion, let
${\bf w}\in\cone(\dom{{f^*}})\cap L$.
Clearly, ${\bf 0}$ is in $\cone(\dom{{g^*}})$ by its definition.
Otherwise, if ${\bf w}\neq{\bf 0}$, then since
${\bf w}\in \cone(\dom{{f^*}})$, using \eqref{eq:cone-dom-fstar},
we can write ${\bf w}=\lambda{\bf u}$ for some $\lambda>0$ and ${\bf u}\in\dom{{f^*}}$.
Since $L$ is a linear subspace, ${\bf u}={\bf w}/\lambda\in L$, implying
${\bf u}\in (\dom{{f^*}})\cap L = \dom{{g^*}}$.
Thus,
${\bf w}=\lambda{\bf u}\in\cone(\dom{{g^*}})$
(again by \eqref{eq:cone-dom-fstar}),
so
$\cone(\dom{{f^*}})\cap L\subseteq \cone(\dom{{g^*}})$.
\qed
\subsection{Barrier cone of a function}
\label{sec:slopes}
The effective domain of ${f^*}$ and its conic hull,
$\cone(\dom{{f^*}})$, play important roles in standard convex
analysis, as well as in our development of astral space,
as seen already, for instance, in
Theorems~\ref{thm:rescpol-is-con-dom-fstar} and~\ref{thm:e1}, and as
will become even more pronounced in later sections.
However, results such as the theorems just mentioned depend on
conditions; for instance, the conclusion in
Theorem~\ref{thm:e1} that
$\cone(\dom{{g^*}}) = \cone(\dom{{f^*}}) \cap L$
depended on both $f$ and $g$ being convex, closed and proper.
Nevertheless, we will see that such results can be generalized and
strengthened so that they can be shown to hold with almost no
assumptions beyond convexity.
To do so, rather than
$\cone(\dom{{f^*}})$,
we will need to re-state the results using a slightly larger set, as
we now discuss.
Recall that the \emph{barrier cone} of a nonempty set $S\subseteq{\R^n}$,
denoted $\barr S$,
is the set
\[
\barr S
=
\braces{ {\bf u}\in{\R^n} :
\sup_{{\bf x}\in S} {\bf x}\cdot{\bf u} < +\infty
},
\]
which can be viewed as the set of directions in
which the set $S$ is eventually bounded.
The barrier cone can also be expressed in terms of
the \emph{support function}
\begin{equation} \label{eq:e:2}
\indstar{S}({\bf u}) = \sup_{{\bf x}\in S} {\bf x}\cdot{\bf u},
\end{equation}
for ${\bf u}\in{\R^n}$,
which is the conjugate of the
\emph{indicator function}
\begin{equation} \label{eq:indf-defn}
\indf{S}({\bf x}) = \left\{ \begin{array}{cl}
0 & \mbox{if ${\bf x}\in S$} \\
+\infty & \mbox{otherwise}
\end{array} \right.
\end{equation}
for ${\bf x}\in{\R^n}$
(see \citet[Theorem~13.2]{ROC}).
With these definitions, we can write the barrier cone as
the effective domain of $\indstar{S}$:
\begin{equation} \label{eqn:bar-cone-defn}
\barr S
=
\braces{ {\bf u}\in{\R^n} :
\indstar{S}({\bf u}) < +\infty
}
=
\dom{\indstar{S}}.
\end{equation}
We will especially be interested in the
barrier cone of the epigraph of $f$ in ${\mathbb{R}}^{n+1}={\R^n}\times{\mathbb{R}}$.
Assuming for now that $f\not\equiv+\infty$, this is the set
\[
\barr(\epi f)
=
\braces{ \rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}} :
\sup_{\rpair{{\bf x}}{y}\in\epi f} ({\bf x}\cdot{\bf u} + yv) < +\infty
}.
\]
In fact,
the set $\cone(\dom{{f^*}})$ turns out to be closely related to
a subset of this set:
\begin{proposition} \label{pr:conedomfstar-and-barrepif}
Let $f:{\R^n}\rightarrow\overline{\R}$ and assume $f\not\equiv+\infty$.
Let ${\bf u}\in{\R^n}\setminus\{{\bf 0}\}$.
Then ${\bf u}\in\cone(\dom{f^*})$ if and only if
$\rpair{{\bf u}}{v}\in\barr(\epi f)$ for some $v<0$.
\end{proposition}
\proof
For all $\lambda>0$, we have
\begin{equation} \label{eqn:conedomisbarrier}
{f^*}(\lambda{\bf u})
=
\sup_{\rpair{{\bf x}}{y}\in\epi f} ({\bf x}\cdot(\lambda{\bf u}) - y)
=
\lambda \sup_{\rpair{{\bf x}}{y}\in\epi f} \Parens{{\bf x}\cdot{\bf u} - \frac{y}{\lambda}},
\end{equation}
using \eqref{eq:fstar-mod-def}.
Note that the rightmost expression is not $+\infty$ if and only if
$\rpair{{\bf u}}{-1/\lambda}\in\barr(\epi f)$,
while the leftmost expression is not $+\infty$ if and only if
$\lambda{\bf u}\in\dom{f^*}$.
Thus,
if ${\bf u}\in\cone(\dom{f^*})$ then
there exists $\lambda>0$ such that
$\lambda{\bf u}\in\dom{{f^*}}$
(by \eqref{eq:cone-dom-fstar}),
implying
$\rpair{{\bf u}}{-1/\lambda}\in\barr(\epi f)$
and proving the claim in one direction.
Conversely, if
$\rpair{{\bf u}}{v}\in\barr(\epi f)$ with $v<0$ then,
letting $\lambda=-1/v$,
\eqref{eqn:conedomisbarrier}
now implies
$\lambda{\bf u}\in\dom{{f^*}}$ so that
${\bf u}\in\cone(\dom{f^*})$.
\qed
Thus, except possibly for the origin,
the points in $\cone(\dom{{f^*}})$ can be characterized in
terms of a subset of $\barr(\epi f)$, namely, those elements of the
form $\rpair{{\bf u}}{v}$ with $v<0$.
Nevertheless,
for obtaining the most general results,
considering only this subset is not enough;
rather, we need to consider the \emph{entire} set
$\barr(\epi f)$, including crucially
those elements $\rpair{{\bf u}}{v}$ with $v=0$.
Accordingly, we define a set
called the \emph{barrier cone of $f$}, denoted $\slopes{f}$,
which will largely replace $\cone(\dom{{f^*}})$
in the development to follow.
For any function $f:{\R^n}\rightarrow\overline{\R}$
with $f\not\equiv+\infty$,
we define this set
to be the projection of $\barr(\epi f)$
into ${\R^n}$,
that is,
\[
\slopes{f}
=
P(\barr(\epi f))
=
\braces{ {\bf u}\in{\R^n} :
\exists v\in{\mathbb{R}},
\sup_{\rpair{{\bf x}}{y}\in\epi f} ({\bf x}\cdot{\bf u} + yv) < +\infty
},
\]
where $P$ is the projection map
$P({\bf u},v)={\bf u}$ for
$\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}$.
In the exceptional case that $f\equiv+\infty$, so that
$\epi{f}=\emptyset$, we define $f$'s barrier cone to be just the origin,
$\slopes{f} = \{{\bf 0}\}$.
This special definition is consistent with all other constant
functions
(whose barrier cones also consist of only the origin),
and will be convenient in what follows.
The pairs $\rpair{{\bf u}}{v}$ in
$\barr(\epi f)$ with $v=0$ are of particular interest, as they
do not generally correspond to vectors in $\cone(\dom{f^*})$.
We call the set of all such vectors ${\bf u}\in{\R^n}$
with $\rpair{{\bf u}}{0}\in\barr(\epi f)$
the \emph{vertical barrier cone of $f$}, denoted
$\vertsl{f}$, so called because
each vector ${\bf u}$ in this set defines a vertically-oriented halfspace
in ${\mathbb{R}}^{n+1}$ that includes all of $\epi f$.
This set can also be expressed as the barrier cone of $\dom f$.
Thus,
\begin{eqnarray}
\vertsl{f}
&=&
\{ {\bf u}\in{\R^n} : \rpair{{\bf u}}{0}\in\barr(\epi f) \}
\nonumber
\\
&=&
\braces{ {\bf u}\in{\R^n} :
\sup_{{\bf x}\in\dom{f}} {\bf x}\cdot{\bf u} < +\infty }
\nonumber
\\
&=&
\barr(\dom f) = \dom{\indstar{\dom{f}}}
\label{eqn:vertsl-defn}
\end{eqnarray}
using \eqref{eqn:bar-cone-defn}.
When $\vertsl{f}$ is combined with
$\cone(\dom{{f^*}})$, we obtain exactly the barrier cone of $f$
(assuming $f\not\equiv+\infty$), as shown in the next proposition.
Thus, $\cone(\dom{{f^*}})$ is always included in the barrier cone of
$f$, but this latter set might also include additional elements from
the vertical barrier cone.
\begin{proposition} \label{pr:slopes-equiv}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then
\[
\slopes{f} =
\left\{ \begin{array}{cl}
\{ {\bf 0} \} & \mbox{if $f\equiv+\infty$} \\
\cone(\dom{{f^*}}) \cup (\vertsl{f})
& \mbox{otherwise.}
\end{array}
\right.
\]
\end{proposition}
\proof
The claim holds by definition if $f\equiv+\infty$, so we assume
henceforth that $f\not\equiv+\infty$.
We note first that $\barr(\epi f)$ can only contain points
$\rpair{{\bf u}}{v}$ with $v\leq 0$.
To see this,
suppose to the contrary that
$\rpair{{\bf u}}{v}\in\barr(\epi f)$ with $v>0$.
Then there must exist $\beta\in{\mathbb{R}}$ such that
${\bf x}\cdot{\bf u}+yv\leq \beta$ for all
$\rpair{{\bf x}}{y}\in\epi{f}$.
Let $\rpair{{\bf x}}{y}$ be any point in $\epi{f}$
(which cannot be empty since $f\not\equiv+\infty$).
Then for all $y'\geq y$,
$\rpair{{\bf x}}{y'}$ must also be in $\epi{f}$,
implying ${\bf x}\cdot{\bf u}+ y' v \leq \beta$.
Since $v>0$ and $y'$ can be arbitrarily large, the
left-hand side of this inequality can also be arbitrarily large,
contradicting that it is bounded by $\beta$.
Next, we observe that
$\rpair{{\bf 0}}{0}\in\barr(\epi f)$
(since ${\bf x}\cdot{\bf 0}+y\cdot 0=0$ for all $\rpair{{\bf x}}{y}\in\epi{f}$),
implying that ${\bf 0}\in\vertsl{f}$.
Thus,
\begin{eqnarray*}
\slopes{f}
&=&
\braces{ {\bf u}\in{\R^n}\setminus\{{\bf 0}\} :
\exists v<0, \rpair{{\bf u}}{v}\in\barr(\epi f)
}
\\
&&
\cup
\braces{ {\bf u}\in{\R^n} : \rpair{{\bf u}}{0}\in\barr(\epi f) }
\\
&=&
(\cone(\dom{f^*})\setminus\{{\bf 0}\})
\cup
(\vertsl{f})
\\
&=&
\cone(\dom{f^*})
\cup
(\vertsl{f}).
\end{eqnarray*}
The first equality is
because $v\leq 0$ for all $\rpair{{\bf u}}{v}\in\barr(\epi f)$,
and the first and third equalities also use the observation that
${\bf 0}\in\vertsl{f}$.
The second equality follows from
Proposition~\ref{pr:conedomfstar-and-barrepif}
and the definition of $\vertsl{f}$.
\qed
Later, we will make use of the fact that
the barrier of a function is the same as that of its
lower semicontinuous hull:
\begin{proposition} \label{pr:slopes-same-lsc}
Let $f:{\R^n}\rightarrow\overline{\R}$.
Then $\slopes(\lsc f)=\slopes{f}$.
\end{proposition}
\proof
We assume $f\not\equiv+\infty$ since otherwise
$f$ is already lower semicontinuous.
As a general observation,
note that for any nonempty set $S\subseteq{\R^n}$,
and for all ${\bf u}\in{\R^n}$,
\[
\sup_{{\bf x}\in S} {\bf x}\cdot{\bf u}
=
\sup_{{\bf x}\in \cl{S}} {\bf x}\cdot{\bf u}
\]
by continuity of the map ${\bf x}\mapsto{\bf x}\cdot{\bf u}$.
Thus, the barrier cone of a set is the same as that of
its closure in ${\R^n}$; that is,
$\barr S = \barr(\cl{S})$.
In particular, this implies that
$\barr(\epi f)=\barr(\cl(\epi f))=\barr(\epi{(\lsc f)})$,
since the epigraph of the lower semicontinuous hull of $f$ is exactly
the closure of $f$'s epigraph in ${\mathbb{R}}^{n+1}$.
This proves the claim.
\qed
In general, $\slopes{f}$ need not be the same as
$\cone(\dom{{f^*}})$, even if $f$ is convex, closed and proper.
For example, in ${\mathbb{R}}^2$, let
\begin{equation} \label{eqn:ex:1}
f(x_1,x_2) = \left\{ \begin{array}{cl}
-x_1 & \mbox{if $x_2\geq 0$} \\
+\infty & \mbox{otherwise.}
\end{array} \right.
\end{equation}
This function is convex, closed and proper.
Its conjugate can be computed to be:
\begin{equation} \label{eqn:ex:1:conj}
{f^*}(u_1,u_2) = \left\{ \begin{array}{cl}
0 & \mbox{if $u_1=-1$ and $u_2\leq 0$} \\
+\infty & \mbox{otherwise.}
\end{array} \right.
\end{equation}
Thus,
\begin{equation} \label{eqn:ex:1:cone-dom-fstar}
\cone(\dom{{f^*}})=
\{{\bf 0}\} \cup
\{\trans{[u_1,u_2]}\in{\mathbb{R}}^2 : u_1<0 \mbox{ and } u_2\leq 0\},
\end{equation}
but it can be checked that
$\vertsl{f} = \{\trans{[u_1,u_2]}\in{\mathbb{R}}^2 : u_1=0 \mbox{ and } u_2\leq 0\}$
so that
$\slopes{f} = -\mathbb{R}_{\geq 0}^2$.
Later, in Section~\ref{sec:dual-char-ent-clos},
we will study when
$\slopes{f}=\cone(\dom{{f^*}})$, a property that we will see exactly
determines when a function's double conjugate $f^{*\dualstar}$ is the same as its
extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\subsection{Generalization via exponential composition}
\label{sec:exp-comp}
As discussed above, some of our results require certain
assumptions which limit their apparent generality.
Most often, the limitation is because of difficulties
arising when the function $f$ is not lower-bounded
(so that $\inf f = -\infty$).
Nevertheless,
there turns out to be a simple trick for applying and
generalizing such results to functions that might not satisfy such
limiting assumptions.
The idea is to replace the function $f$
by another function $f'$ that retains a
great many of $f$'s properties, but which is always lower-bounded.
We can then apply such results to the modified
function $f'$ to obtain a result for the original function $f$.
We now explain this technique in more detail.
First, let us define an extension of the exponential function, $e^x$,
to the extended domain $\overline{\R}$.
Specifically, for $\bar{x}\in\overline{\R}$, define
\[
{\overline{\exp}}(\bar{x}) = \left\{ \begin{array}{cl}
0 & \mbox{if $\bar{x}=-\infty$} \\
e^{\bar{x}} & \mbox{if $\bar{x}\in{\mathbb{R}}$} \\
+\infty & \mbox{if $\bar{x}=+\infty$.}
\end{array} \right.
\]
This function has many favorable properties:
It is continuous and strictly increasing over its entire domain, and
it is also convex.
Given a convex function $f$, we can compose it with ${\overline{\exp}}$ to derive
a new convex function $f'$, some of whose properties are given in
the next proposition.
\begin{proposition} \label{pr:j:2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, and
let $f'={\overline{\exp}}\circ f$, that is,
\[ f'({\bf x}) = {\overline{\exp}}(f({\bf x})) \]
for all ${\bf x}\in{\R^n}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:j:2a}
$f'$ is convex and $\inf f'\geq 0$.
\item \label{pr:j:2b}
For all ${\bf x},{\bf y}\in{\R^n}$,
$f'({\bf x})\leq f'({\bf y})$ if and only if $f({\bf x})\leq f({\bf y})$.
Therefore, $\resc{f'}=\resc{f}$.
\item \label{pr:j:2c}
For all $\overline{\xx}\in\extspac{n}$, $\fpext(\overline{\xx})={\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}))$.
That is, $\fpext = {\overline{\exp}}\circ{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\item \label{pr:j:2lsc}
For all ${\bf x}\in{\R^n}$,
$(\lsc f')({\bf x})={\overline{\exp}}((\lsc f)(x))$.
That is, $\lsc f' = {\overline{\exp}}\circ (\lsc f)$.
Therefore,
if $f$ is lower semicontinuous, then $f'$ is as well.
\item \label{pr:j:2d}
For all $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
$\fpext(\overline{\xx})\leq \fpext(\overline{{\bf y}})$ if and only if
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}})$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:j:2a}):
That $f'\geq 0$ is immediate.
Convexity of $f'$ almost follows from a general result like
\citet[Theorem~5.1]{ROC},
except for the fact that we allow $f$ to include
$-\infty$ in its range.
For completeness, we give a proof based on the characterization of
convexity given in Proposition~\ref{pr:rock:4.2}.
Let ${\bf x},{\bf y}\in{\R^n}$, and suppose $f'({\bf x})<a$ and $f'({\bf y})<b$
for some $a,b\in{\mathbb{R}}$.
Since $f'\geq 0$, these assumptions imply that $a,b>0$,
and furthermore, since ${\overline{\exp}}$ is strictly increasing,
they imply that $f({\bf x})<\ln a$ and $f({\bf y})<\ln b$.
Let $\lambda\in (0,1)$.
Then
\begin{eqnarray}
f'(\lambda {\bf x} + (1-\lambda) {\bf y})
&=&
{\overline{\exp}}(f(\lambda {\bf x} + (1-\lambda) {\bf y}))
\nonumber
\\
&<&
{\overline{\exp}}(\lambda \ln a + (1-\lambda) \ln b)
\label{eqn:j:2a}
\\
&\leq&
\lambda a + (1-\lambda) b.
\label{eqn:j:2b}
\end{eqnarray}
\eqref{eqn:j:2a} is because ${\overline{\exp}}$ is strictly increasing, and by
convexity of $f$ (using Proposition~\ref{pr:rock:4.2}).
\eqref{eqn:j:2b} is by convexity of $e^x$.
Therefore, $f'$ satisfies the condition of Proposition~\ref{pr:rock:4.2}
and is convex.
Part~(\ref{pr:j:2b}) follows from ${\overline{\exp}}$ being strictly
increasing, and because $\resc{f}$ is defined only in terms of the
(partial) ordering on ${\R^n}$ induced by $f$, which is identical for
$f'$.
Part~(\ref{pr:j:2c}):
By Proposition~\ref{pr:d1},
there exists a sequence $\seq{{\bf x}_t}$ in ${\R^n}$ converging to $\overline{\xx}$
with $f({\bf x}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
This implies
\[
{\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}))
=
{\overline{\exp}}(\lim f({\bf x}_t))
=
\lim {\overline{\exp}}(f({\bf x}_t))
=
\lim f'({\bf x}_t)
\geq
\fpext(\overline{\xx})
\]
by continuity of ${\overline{\exp}}$.
By a similar argument,
there exists a sequence $\seq{{\bf x}'_t}$ in ${\R^n}$ converging to $\overline{\xx}$
with $f'({\bf x}'_t)\rightarrow \fpext(\overline{\xx})$, implying
\[
\fpext(\overline{\xx})
=
\lim f'({\bf x}'_t)
=
\lim {\overline{\exp}}(f({\bf x}'_t))
=
{\overline{\exp}}(\lim f({\bf x}'_t))
\geq
{\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}))
\]
since ${\overline{\exp}}$ is continuous and strictly increasing.
Part~(\ref{pr:j:2lsc}):
For all ${\bf x}\in{\R^n}$,
\[
(\lsc f')({\bf x})
=
\fpext({\bf x})
=
{\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x}))
=
{\overline{\exp}}((\lsc f)({\bf x}))
\]
by
part~(\ref{pr:j:2c}) and
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
If $f$ is lower semicontinuous then
the right-hand side is equal to ${\overline{\exp}}(f({\bf x}))=f'({\bf x})$,
proving $f'$ is lower semicontinuous as well.
Part~(\ref{pr:j:2d}) follows from
part~(\ref{pr:j:2c}) since ${\overline{\exp}}$ is strictly increasing.
\qed
To apply this technique to results like
Theorems~\ref{thm:rescpol-is-con-dom-fstar} and~\ref{thm:e1},
we also need to relate the conjugate of $f'$ to that of $f$.
For this purpose, we can apply the following result of
\citet[Theorem~E.2.5.1]{HULL} which shows how to compute
the conjugate of a composite function of exactly the kind we are
considering.
We state the result in a slightly weaker form than in the original
using assumptions that have been simplified for our current purposes.
(See also \citet[Theorem 2]{hiriart_composition}.)
\begin{theorem} \label{thm:gen-conjugate-compose}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, closed and proper, and
let $g:{\mathbb{R}}\rightarrow{\mathbb{R}}$ be convex, closed, nondecreasing, and finite
everywhere, and set $g(+\infty)=+\infty$.
For ${\bf u}\in{\R^n}$ and $\alpha\in\mathbb{R}_{\geq 0}$, let
\[
\hlfcn{{\bf u}}{\alpha}
= \left\{ \begin{array}{cl}
\alpha {f^*}({\bf u}/\alpha)
& \mbox{if $\alpha>0$} \\
\indstar{\dom{f}}({\bf u})
& \mbox{if $\alpha=0$.}
\end{array} \right.
\]
Then for all ${\bf u}\in{\R^n}$,
\[
(g\circ f)^*({\bf u}) = \min_{\alpha\in\mathbb{R}_{\geq 0}} [\hlfcn{{\bf u}}{\alpha} + {g^*}(\alpha)].
\]
\end{theorem}
Using this theorem, it is straightforward to compute $\dom{{f'^*}}$,
which turns out to be exactly equal to $f$'s barrier cone, $\slopes{f}$
(except if $f\equiv+\infty$).
\begin{proposition} \label{pr:dom-fpstar}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and not identically $+\infty$,
and let $f'={\overline{\exp}}\circ f$.
Then
\[
\slopes{f}
=
\dom{{f'^*}}
=
\cone(\dom{{f'^*}}).
\]
\end{proposition}
\proof
We can assume without loss of generality that $f$ is lower
semicontinuous, since otherwise it suffices to prove the result for
its lower semicontinuous hull,
$\lsc f$.
This is because $\slopes(\lsc f)=\slopes{f}$ by
Proposition~\ref{pr:slopes-same-lsc}, and also,
by Proposition~\ref{pr:j:2}(\ref{pr:j:2lsc}),
$({\overline{\exp}}\circ(\lsc f))^*=(\lsc f')^*={f'^*}$.
We first prove
\begin{equation} \label{eqn:dom-fpstar:1}
\cone(\dom{{f^*}}) \cup (\vertsl{f})
=
\dom{{f'^*}},
\end{equation}
and thus that $\slopes{f} = \dom{{f'^*}}$
(by Proposition~\ref{pr:slopes-equiv}).
If $f({\bf x})=-\infty$ for some ${\bf x}\in{\R^n}$, then
$f$ is improper and lower semicontinuous, and therefore
equal to $-\infty$ or $+\infty$ everywhere
\citep[Corollary~7.2.1]{ROC}.
Thus, $f'$ is actually the indicator function for $\dom f$, that is,
$f' = \indf{\dom f}$, so
\[
\dom{{f'^*}} = \dom{\indstar{\dom f}} = \vertsl{f}
\]
by \eqref{eqn:vertsl-defn}.
Also, ${f^*}\equiv+\infty$ so $\cone(\dom{{f^*}})=\{{\bf 0}\}$,
and furthermore, ${\bf 0}\in\dom{{f'^*}}$ since $f'$ is lower-bounded.
Combining yields \eqref{eqn:dom-fpstar:1} in this case.
Otherwise, suppose $f>-\infty$, implying that $f$ is
closed and proper.
We apply Theorem~\ref{thm:gen-conjugate-compose}, setting
$g(x)=e^x$ for $x\in{\mathbb{R}}$
so that $g\circ f = f'$
(having set $g(+\infty)=+\infty$).
Let $\hlfcn{{\bf u}}{\alpha}$ be as defined in the theorem.
By a standard calculation, $\dom{{g^*}}=\mathbb{R}_{\geq 0}$
(specifically, ${g^*}(u) = u \ln u - u$ for $u\geq 0$, where, by
convention, $0 \ln 0 = 0$).
Therefore,
for all ${\bf u}\in{\R^n}$,
Theorem~\ref{thm:gen-conjugate-compose} implies that
${\bf u}\in\dom{{f'^*}}$ if and only if $\hlfcn{{\bf u}}{\alpha}<+\infty$ for
some $\alpha\geq 0$.
Furthermore, $\hlfcn{{\bf u}}{0}<+\infty$ if and only if
$\indstar{\dom{f}}({\bf u})<+\infty$, that is, if and only if
${\bf u}\in\vertsl{f}$
(by \eqref{eqn:vertsl-defn}).
And for $\alpha>0$,
$\hlfcn{{\bf u}}{\alpha}<+\infty$ if and only if
${f^*}({\bf u}/\alpha)<+\infty$, that is, if and only if
${\bf u}/\alpha\in \dom{{f^*}}$.
Therefore,
if ${\bf u}\neq{\bf 0}$ then
$\hlfcn{{\bf u}}{\alpha}<+\infty$ for some $\alpha>0$
if and only if
${\bf u}\in \cone(\dom{{f^*}})$.
Finally, if ${\bf u}={\bf 0}$, then ${\bf 0}\in\cone(\dom{f^*})$ and
also ${\bf 0}\in\dom{{f'^*}}$ since $f'$ is lower-bounded.
Combining yields \eqref{eqn:dom-fpstar:1} in this case as
well.
Thus, \eqref{eqn:dom-fpstar:1} holds in all cases.
Since the two sets on the left-hand side of
that equation are cones, their union is as well
(see Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:b})).
Thus, $\dom{{f'^*}}$, which is convex as the domain of a
convex function, is a convex cone that includes ${\bf 0}$
(being in $\cone(\dom{{f^*}})$), thereby implying
$\cone(\dom{{f'^*}})=\dom{{f'^*}}$ and completing the proposition.
\qed
\begin{proposition} \label{pr:barf-is-cone}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then $\slopes{f}$ is a pointed convex cone.
\end{proposition}
\proof
If $f\equiv+\infty$ then $\slopes{f}=\{{\bf 0}\}$, which is a pointed
convex cone.
Otherwise, let $f'={\overline{\exp}}\circ f$.
By Proposition~\ref{pr:dom-fpstar},
$\slopes{f}=\cone(\dom{{f'^*}})$, which is also a pointed convex
cone.
\qed
\subsection{Generalized results for reductions}
As a first application of this technique,
we can slightly generalize the
standard result in Theorem~\ref{thm:rescpol-is-con-dom-fstar}, putting
it in a form that will be more useful going forward:
\begin{corollary} \label{cor:rescpol-is-slopes}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then
\[ \resc{f} = \polar{(\slopes{f})} \]
and
\[ \rescpol{f} = \cl(\slopes{f}). \]
\end{corollary}
\proof
If $f\equiv+\infty$ then
$\slopes{f}=\{{\bf 0}\}$ so
$\resc{f}=\polar{(\slopes{f})}={\R^n}$.
Otherwise,
let $f'={\overline{\exp}}\circ f$.
Then
\[
\resc{f}
=
\resc{f'}
=
\polar{(\cone(\dom{{f'^*}}))}
=
\polar{(\slopes{f})}.
\]
The first equality is by
Proposition~\ref{pr:j:2}(\ref{pr:j:2b}).
The second is by Theorem~\ref{thm:rescpol-is-con-dom-fstar}
applied to $f'$
(which is convex and lower-bounded by
Proposition~\ref{pr:j:2}(\ref{pr:j:2a}), and so also closed and
proper).
The third is
by Proposition~\ref{pr:dom-fpstar}.
This proves the first claim of the corollary.
Taking the polar of both sides then yields the second claim.
\qed
In particular, combined with
Theorem~\ref{thm:rescpol-is-con-dom-fstar},
when $f$ is closed, proper and convex,
this shows that
$\slopes{f}$ and $\cone(\dom{{f^*}})$ have the same closures (in
${\R^n}$) and therefore can differ only in their relative boundaries.
As another application, we can immediately generalize the last part of
Theorem~\ref{thm:e1}, using fewer assumptions and with a cleaner
conclusion:
\begin{corollary} \label{cor:thm:e1:slopes}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous, and
let ${\bf v}\in\resc{f}$.
Let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Then
\[
\slopes{g} = (\slopes{f}) \cap L
\]
where
\[ L=\{{\bf u}\in{\R^n} : {\bf u}\cdot{\bf v}=0\} \]
is the the linear subspace perpendicular to ${\bf v}$.
\end{corollary}
\proof
If $f\equiv+\infty$, then $g\equiv+\infty$ so
$\slopes{g}=\slopes{f}=\{{\bf 0}\}$ and the result holds.
So we assume henceforth that $f\not\equiv+\infty$, which implies
$g\not\equiv+\infty$ (by Proposition~\ref{pr:d2}(\ref{pr:d2:c})).
Let $f'={\overline{\exp}}\circ f$.
Then $f'$ is convex, lower-bounded and lower semicontinuous
by Proposition~\ref{pr:j:2}(\ref{pr:j:2a},\ref{pr:j:2lsc}),
so $f'$ is also proper and closed.
In addition, let $g'=\fpshad{\limray{\vv}}$ be the reduction of $f'$ at
$\limray{{\bf v}}$.
Then for all ${\bf x}\in{\R^n}$,
\[
g'({\bf x})
=
\fpext(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}))
=
{\overline{\exp}}(g({\bf x})),
\]
with the second equality following from
Proposition~\ref{pr:j:2}(\ref{pr:j:2c}).
The reduction $g$ is convex and lower semicontinuous
(by Theorem~\ref{thm:a10-nunu}),
so $g'$ is convex, closed and proper by
the same reasoning as was used for $f'$.
Therefore,
\[
\slopes{g}
=
\dom{{g'^*}}
=
(\dom{{f'^*}}) \cap L
=
(\slopes{f}) \cap L,
\]
where the second equality is
by Theorem~\ref{thm:e1}, and the first and third equalities are
by Proposition~\ref{pr:dom-fpstar} (since neither $f$ nor $g$
are identically $+\infty$).
\qed
We had earlier studied basic properties of reductions $g=\fshad{\limray{\vv}}$,
for instance, in Theorem~\ref{thm:d4},
under the assumption that ${\bf v}$ is in $f$'s recession cone, $\resc{f}$.
Using Theorem~\ref{thm:fext-dub-sum} and the technique just developed,
we can now show that if ${\bf v}$ is \emph{not}
in $\resc{f}$, then $g$ is identically equal to $+\infty$:
\begin{theorem} \label{thm:i:4}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Let ${\bf v}\in{\R^n}$
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Assume ${\bf v}\not\in\resc{f}$.
Then
\[
\overline{g}(\overline{\xx})
={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
=+\infty
\]
for all $\overline{\xx}\in\extspac{n}$.
If, in addition, $f$ is closed, then
$f^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}) =+\infty$
for all $\overline{\xx}\in\extspac{n}$.
\end{theorem}
\proof
The cases $f\equiv +\infty$ or $f\equiv -\infty$
are both impossible since then we would have
$\resc{f}={\R^n}$, but ${\bf v}\not\in\resc{f}$.
Let us assume for the moment that $f$ is proper and closed.
Since ${\bf v}\not\in\resc{f}$, by
Theorem~\ref{thm:rescpol-is-con-dom-fstar},
there must exist a point
${\bf u}\in\dom{{f^*}}$ with ${\bf v}\cdot{\bf u}>0$.
Therefore, by
Theorem~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b}),
\begin{eqnarray*}
f^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
&\geq&
-{f^*}({\bf u})
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx})\cdot{\bf u}
\\
&=&
-{f^*}({\bf u})
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}}\cdot{\bf u}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u}
= +\infty.
\end{eqnarray*}
The last equality is because $-{f^*}({\bf u})>-\infty$
(since ${\bf u}\in\dom{{f^*}}$),
and $\limray{{\bf v}}\cdot{\bf u}=+\infty$
(since ${\bf v}\cdot{\bf u}>0$).
Thus, by
Theorem~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:a}),
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})\geqf^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})=+\infty$.
In particular, this shows that
$g({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=+\infty$
for all ${\bf x}\in{\R^n}$, and so
$\overline{g}\equiv +\infty$.
Returning to the general case,
suppose $f\not\equiv +\infty$ but $f$ is not
necessarily proper.
Let $f'={\overline{\exp}}\circ f$.
Then $f'$ is convex, lower-bounded and lower semicontinuous
by Proposition~\ref{pr:j:2}(\ref{pr:j:2a},\ref{pr:j:2lsc}),
so $f'$ is also proper and closed.
Also, $\resc{f'}=\resc{f}$
(by Proposition~\ref{pr:j:2}(\ref{pr:j:2b})),
so ${\bf v}\not\in\resc{f'}$.
Therefore,
for all $\overline{\xx}\in\extspac{n}$,
\[
{\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}))
=
\fpext(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})=+\infty,
\]
with the equalities following respectively from
Proposition~\ref{pr:j:2}(\ref{pr:j:2c})
and the argument above, applied to $f'$.
This implies
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})=+\infty$.
Thus, as above, $g\equiv+\infty$ so
$\overline{g}\equiv+\infty$.
\qed
As a summary, the next corollary combines Theorem~\ref{thm:i:4} with
Theorems~\ref{thm:d4} and~\ref{thm:a10-nunu}, making no assumption
about ${\bf v}$:
\begin{corollary} \label{cor:i:1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, let ${\bf v}\in{\R^n}$,
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Then the following hold:
\begin{enumerate}
\item \label{cor:i:1a}
$g$ is convex and lower semicontinuous.
\item \label{cor:i:1b}
$\overline{g}(\xbar^{\bot})=\overline{g}(\overline{\xx})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})$
for all $\overline{\xx}\in\extspac{n}$.
\item \label{cor:i:1c}
If ${\bf v}\in\resc{f}$, then $\overline{g}\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\end{enumerate}
\end{corollary}
\proof
Part~(\ref{cor:i:1c}) is exactly what is stated in
Theorem~\ref{thm:d4}(\ref{thm:d4:b}).
Assume for the remainder of the proof that $f$ is lower
semicontinuous.
This is without loss of generality since if it is not, we can replace
$f$ with $\lsc f$ and apply Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa}).
If ${\bf v}\in\resc{f}$, then parts~(\ref{cor:i:1a}) and~(\ref{cor:i:1b})
follow from Theorem~\ref{thm:d4}(\ref{thm:d4:a},\ref{thm:d4:c})
and Theorem~\ref{thm:a10-nunu}.
If ${\bf v}\not\in\resc{f}$, then
$\overline{g}(\overline{\xx})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})=+\infty$
for all $\overline{\xx}\in\extspac{n}$,
by Theorem~\ref{thm:i:4}.
This implies part~(\ref{cor:i:1b}).
This also implies that $g$ is the constant function $+\infty$, which is
convex and lower semicontinuous, proving part~(\ref{cor:i:1a}).
\qed
In a similar fashion, we can relate the double conjugate $g^{*\dualstar}$ of a
reduction $g=\fshad{\limray{\vv}}$ to that of the function $f$ from which it was
derived.
This fact will be used soon in characterizing when
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
\begin{theorem} \label{thm:i:5}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and closed, let ${\bf v}\in{\R^n}$,
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Then $g^{*\dualstar}(\overline{\xx})=f^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})$
for all $\overline{\xx}\in\extspac{n}$.
\end{theorem}
\proof
We proceed in cases.
Suppose first that $f$ is not proper.
In this case, since $f$ is convex, closed and improper, it must be either
identically $+\infty$ or
identically $-\infty$.
If $f\equiv +\infty$, then it can be checked that
$g\equiv +\infty$, ${f^*}={g^*}\equiv -\infty$
and $f^{*\dualstar}=g^{*\dualstar}\equiv +\infty$.
Similarly, if $f\equiv -\infty$ then
$f^{*\dualstar}=g^{*\dualstar}\equiv -\infty$.
Either way, the theorem's claim holds.
For the next case, suppose ${\bf v}\not\in\resc{f}$.
Then
Theorem~\ref{thm:i:4} implies that
$f^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}) =+\infty$
for all $\overline{\xx}\in\extspac{n}$,
and also that $g\equiv +\infty$, so $g^{*\dualstar}\equiv +\infty$.
Thus, the claim holds in this case as well.
We are left only with the case that $f$ is closed, convex and proper, and that
${\bf v}\in\resc{f}$, which we assume for the remainder of the proof.
We will argue that, for all ${\bf u}\in{\R^n}$,
\begin{equation} \label{eq:e:9}
-{g^*}({\bf u})
=
-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}}\cdot{\bf u}.
\end{equation}
This is sufficient to prove the theorem since it implies,
by Theorem~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b}), that
\begin{eqnarray*}
g^{*\dualstar}(\overline{\xx})
&=&
\sup_{{\bf u}\in{\R^n}} [- {g^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u} ]
\\
&=&
\sup_{{\bf u}\in{\R^n}} [- {f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u} ]
\\
&=&
\sup_{{\bf u}\in{\R^n}} [- {f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx})\cdot{\bf u} ]
\\
&=&
f^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx})
\end{eqnarray*}
for all $\overline{\xx}\in\extspac{n}$, as claimed.
Suppose first that $g$ is proper.
Let ${\bf u}\in{\R^n}$.
If ${\bf u}\cdot{\bf v}=0$ then by Theorem~\ref{thm:e1},
${g^*}({\bf u})={f^*}({\bf u})$ implying
\eqref{eq:e:9} in this case.
If ${\bf u}\cdot{\bf v}<0$, then ${g^*}({\bf u})=+\infty$ by
Theorem~\ref{thm:e1}, and ${f^*}({\bf u})>-\infty$ since $f$ is proper (so
that ${f^*}$ is as well).
These imply that both sides of \eqref{eq:e:9} are equal to $-\infty$
in this case.
And if ${\bf u}\cdot{\bf v}>0$, then ${\bf u}\not\in\dom{{f^*}}$
(by Theorem~\ref{thm:rescpol-is-con-dom-fstar},
since ${\bf v}\in\resc{f}$),
and so ${\bf u}\not\in\dom{{g^*}}$ by
Theorem~\ref{thm:e1}.
Therefore, in this case as well,
both sides of \eqref{eq:e:9} are equal to $-\infty$.
Now suppose that $g$ is not proper.
Then $g({\bf y})=-\infty$ for some ${\bf y}\in{\R^n}$, so
${g^*}\equiv +\infty$.
Further, by
Theorem~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:a}),
\[
f^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y})
=
g({\bf y})
=
-\infty.
\]
By
Theorem~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b}),
this implies, for all ${\bf u}\in{\R^n}$, that
\[
-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf y}\cdot{\bf u}
= -\infty.
\]
But since ${\bf y}\cdot{\bf u}\in{\mathbb{R}}$, this actually means
$
-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}}\cdot{\bf u}
= -\infty
$
for all ${\bf u}\in{\R^n}$, proving
\eqref{eq:e:9}.
\qed
\subsection{All reductions closed}
\label{sec:ent-closed-fcn}
In this subsection, we develop a general and precise characterization
for when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
As a step in that direction,
we first generalize the reductions introduced in
Section~\ref{sec:shadow} in a
straightforward way, allowing reductions now to be at any icon, not
just astrons.
Thus,
for a convex function $f:{\R^n}\rightarrow\overline{\R}$ and any
icon $\overline{\ee}\in{{\cal E}_n}$, we define the
\emph{reduction of $f$ at icon $\overline{\ee}$} to
be the function $\fshad{\ebar}:{\R^n}\rightarrow\overline{\R}$
defined by
\begin{equation} \label{eqn:mult-step-shad-def}
\fshad{\ebar}({\bf x}) = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
\end{equation}
for all ${\bf x}\in{\R^n}$;
such a reduction is said to be \emph{iconic}.
Clearly, a reduction $\fshad{\limray{\vv}}$ at astron $\limray{{\bf v}}$
is a special case of such an iconic reduction
in which $\overline{\ee}=\limray{{\bf v}}$.
When $\overline{\ee}={\bf 0}$, the resulting reduction at ${\bf 0}$ is
$\fshad{{\bf 0}}=\lsc f$,
the lower semicontinuous hull of $f$,
by
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
The function in \eqref{eqn:mult-step-shad-def}
captures the behavior of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ on the single galaxy
$\galax{\overline{\ee}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\R^n}$.
It can also be viewed as a kind of composition of multiple
astronic reductions, as will be seen explicitly in the course of
proving the next proposition.
It is the closedness of all reductions at
all icons in ${{\cal E}_n}$
that will characterize when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
First, we prove a few simple properties:
\begin{proposition} \label{pr:i:9}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, and
let $\overline{\ee}\in{{\cal E}_n}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:i:9a}
$\fshad{\ebar}$ is convex and lower semicontinuous.
\item \label{pr:i:9b}
$\fshadext{\ebar}(\overline{\xx}) = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})$
for all $\overline{\xx}\in\extspac{n}$.
\item \label{pr:i:9c}
Either $\fshadext{\ebar}\leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ or $\fshad{\ebar}\equiv +\infty$.
\end{enumerate}
\end{proposition}
\proof
By Proposition~\ref{pr:icon-equiv}, $\overline{\ee}={\bf V}{\bar{\bm{\omega}}}$ for some
matrix ${\bf V}=[{\bf v}_1,\ldots,{\bf v}_k]\in{\mathbb{R}}^{n\times k}$.
Let us define a sequence of functions $g_i:{\R^n}\rightarrow\overline{\R}$ by
$g_0 = \lsc f$ and
$g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$
for $i=1,\ldots,k$.
We claim, by induction on $i=0,\ldots,k$, that
\[
\overline{g}_i(\overline{\xx}) = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}([{\bf v}_1,\ldots,{\bf v}_i]{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
\]
for all $\overline{\xx}\in\extspac{n}$.
This is immediate when $i=0$ by Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
For the inductive step, using Corollary~\ref{cor:i:1}(\ref{cor:i:1b})
and our inductive hypothesis,
\begin{eqnarray*}
\overline{g}_i(\overline{\xx})
&=&
\overline{g}_{i-1}(\limray{{\bf v}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
\\
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}([{\bf v}_1,\ldots,{\bf v}_{i-1}]{\bar{\bm{\omega}}}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
\\
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}([{\bf v}_1,\ldots,{\bf v}_i]{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}).
\end{eqnarray*}
This shows also that each $g_i$ is convex and lower semicontinuous,
for $i=0,\ldots,k$, by repeated application of
Corollary~\ref{cor:i:1}(\ref{cor:i:1a}).
In particular, the foregoing shows that $\fshadext{\ebar}=\overline{g}_k$
and $\fshad{\ebar}=g_k$, and proves
parts~(\ref{pr:i:9a}) and~(\ref{pr:i:9b}).
For part~(\ref{pr:i:9c}), suppose first that
${\bf v}_i\not\in\resc{g_{i-1}}$ for some $i\in\{1,\ldots,k\}$.
Then $g_i\equiv+\infty$, by Theorem~\ref{thm:i:4}, so $g_j\equiv+\infty$
for $j\geq i$,
and in particular $\fshad{\ebar}=g_k\equiv+\infty$.
Otherwise,
${\bf v}_i\in\resc{g_{i-1}}$ for all $i\in\{1,\ldots,k\}$, and so,
by Corollary~\ref{cor:i:1}(\ref{cor:i:1c}),
\[
\fshadext{\ebar}(\overline{\xx})
=
\overline{g}_k(\overline{\xx})
\leq
\overline{g}_{k-1}(\overline{\xx})
\leq \cdots \leq
\overline{g}_0(\overline{\xx}) = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
\]
for all $\overline{\xx}\in\extspac{n}$.
\qed
So all of $f$'s iconic reductions are lower semicontinuous.
However, they are not necessarily closed.
The property of \emph{all} of the iconic reductions being closed turns
out to exactly characterize when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$, as we will see shortly.
Thus,
for a function $f:{\R^n}\rightarrow\overline{\R}$,
we say that \emph{$f$ has all reductions closed} if
{all} of its reductions are closed at {all} icons, that is,
if $\fshad{\ebar}$ is closed for every icon
$\overline{\ee}\in{{\cal E}_n}$.
Here are some useful facts about this property:
\begin{proposition} \label{pr:j:1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then the following hold:
\begin{enumerate}
\item \label{pr:j:1a}
Let $g=\fshad{\limray{\vv}}$
be the reduction of $g$ at $\limray{{\bf v}}$,
for some ${\bf v}\in{\R^n}$.
Then $\gshad{\overline{\ee}}=\fshad{\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}}$
for all $\overline{\ee}\in{{\cal E}_n}$.
Therefore,
if $f$ has all reductions closed, then $g$ does as well.
\item \label{pr:j:1b}
$f$ does \emph{not} have all reductions closed
if and only if there exists an
icon $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q},{\bf q}'\in{\R^n}$ such that
$\fshad{\ebar}({\bf q})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})=+\infty$
and
$\fshad{\ebar}({\bf q}')={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}')=-\infty$.
\item \label{pr:j:1c}
If $\inf f > -\infty$ then $f$ has all reductions closed.
\item \label{pr:j:1d}
If $f < +\infty$ then $f$ has all reductions closed.
\end{enumerate}
\end{proposition}
These last two facts show that $f$ has all reductions closed if it is either
bounded below, or if it is finite everywhere.
\proof
~
Part~(\ref{pr:j:1a}):
For all icons $\overline{\ee}\in{{\cal E}_n}$ and for all ${\bf x}\in{\R^n}$,
\[
\gshad{\overline{\ee}}({\bf x})
=
\overline{g}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
\fshad{\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}}({\bf x})
\]
by Corollary~\ref{cor:i:1}(\ref{cor:i:1b}).
Thus, $\gshad{\overline{\ee}}=\fshad{\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}}$.
Therefore, if $f$ has all reductions closed, then for all $\overline{\ee}\in{{\cal E}_n}$,
$\gshad{\overline{\ee}}=\fshad{\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}}$ is closed,
which implies that $g$ also has all reductions closed.
Part~(\ref{pr:j:1b}):
By definition, $f$ does not have all reductions closed if and only if
$\fshad{\ebar}$ is not closed for some $\overline{\ee}\in{{\cal E}_n}$.
Since $\fshad{\ebar}$ is lower semicontinuous (by
Proposition~\ref{pr:i:9}(\ref{pr:i:9a})),
it is not closed if and only if it is equal to $-\infty$ at some
point, and equal to $+\infty$ at some other point.
(See \citet{ROC}
Section~7, including Corollary~7.2.1, which shows
that a function that is lower semicontinuous but not closed must be
improper, is not finite anywhere, and is neither of the constant
functions $-\infty$ or $+\infty$.)
Part~(\ref{pr:j:1c}):
If $f \geq b$ for some $b > -\infty$,
then $\fshad{\ebar}({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})\geq b$ for all
$\overline{\ee}\in{{\cal E}_n}$ and ${\bf x}\in{\R^n}$.
Therefore, by part~(\ref{pr:j:1b}), $f$ must have all reductions closed.
Part~(\ref{pr:j:1d}):
If $f<+\infty$, then for all $\overline{\ee}\in{{\cal E}_n}$,
by Proposition~\ref{pr:i:9}(\ref{pr:i:9c}),
either $\fshad{\ebar}\equiv+\infty$ or
$\fshad{\ebar}({\bf x})\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})\leq f({\bf x})<+\infty$
for all ${\bf x}\in{\R^n}$.
In either case, the condition of part~(\ref{pr:j:1b}) is ruled out, and
therefore $f$ has all reductions closed.
\qed
We can now characterize precisely when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$:
\begin{theorem} \label{thm:dub-conj-new}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then $f$ has all reductions closed if and only if
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
\end{theorem}
\proof
Since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=\overline{(\lsc f)}$
(Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa})),
and since
${f^*}=(\lsc f)^*$, we can assume without loss of generality that $f$
is lower semicontinuous (replacing it with $\lsc f$ if it is not).
We first prove that if $f$ has all reductions closed then
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$ by induction on the astral rank of
$\overline{\xx}$.
More precisely, we show by induction on $k=0,\ldots,n$ that for every
lower semicontinuous, convex function $f$ with all reductions closed,
and for all $\overline{\xx}\in\extspac{n}$, if $\overline{\xx}$ has astral rank $k$ then
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$.
So suppose that $f$ is convex, lower semicontinuous, and
has all reductions closed.
In particular, this implies
that $\fshad{{\bf 0}}=\lsc f = f$ is closed.
Also, let $\overline{\xx}\in\extspac{n}$ have astral rank $k$.
In the base case that $k=0$, $\overline{\xx}$ must be some point
${\bf x}\in{\R^n}$.
Since $f$ is closed, $f^{**}=f$
\citep[Theorem~12.2]{ROC}, so
\begin{equation} \label{eq:thm:dub-conj-new:1}
f^{*\dualstar}({\bf x})=f^{**}({\bf x})=f({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x}),
\end{equation}
where the first equality follows from
\eqref{eqn:fdub-form-lower-plus}
(or
Theorem~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b})),
and the last equality
is by
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
For the inductive step when $k>0$, we can write
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}$ where ${\bf v}$ is $\overline{\xx}$'s dominant
direction, and $\xbar^{\bot}$ has astral rank $k-1$
(Proposition~\ref{pr:h:6}).
Let $g=\fshad{\limray{\vv}}$ be the reduction of $f$ at $\limray{{\bf v}}$.
Then $g$ is convex and lower semicontinuous
by
Corollary~\ref{cor:i:1}(\ref{cor:i:1a}),
and has all reductions closed by Proposition~\ref{pr:j:1}(\ref{pr:j:1a}).
Thus, we can apply our inductive hypothesis to $g$
at $\xbar^{\bot}$, yielding
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
=
\overline{g}(\xbar^{\bot})
=
g^{*\dualstar}(\xbar^{\bot})
=
f^{*\dualstar}(\overline{\xx}),
\]
where the equalities follow, respectively, from
Corollary~\ref{cor:i:1}(\ref{cor:i:1b}),
our inductive hypothesis,
and
Theorem~\ref{thm:i:5}.
This completes the induction.
Conversely, suppose now that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
Further, suppose by way of contradiction that $f$ does not have all
reductions closed.
Then by Proposition~\ref{pr:j:1}(\ref{pr:j:1b}), there exists an
icon $\overline{\ee}\in\extspac{n}$ and ${\bf q},{\bf q}'\in{\R^n}$ such that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})=+\infty$
and
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}')=-\infty$.
By our assumption, for any ${\bf x}\in{\R^n}$,
\begin{equation} \label{eq:i:8}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
f^{*\dualstar}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
\sup_{{\bf u}\in{\R^n}} (-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\ee}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x}\cdot{\bf u})
\end{equation}
by Theorem~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b}).
In particular, when ${\bf x}={\bf q}'$, the left-hand side is equal to
$-\infty$, so that the argument inside the supremum is
equal to $-\infty$ for all ${\bf u}\in{\R^n}$.
Since ${\bf q}'\cdot{\bf u}\in{\mathbb{R}}$, this means that
$-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\ee}\cdot{\bf u} = -\infty$
for all ${\bf u}\in{\R^n}$.
In turn, this implies, by \eqref{eq:i:8}, that actually
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=-\infty$ for all ${\bf x}\in{\R^n}$.
But this contradicts that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})=+\infty$.
Thus, $f$ has all reductions closed.
\qed
Combined with Proposition~\ref{pr:j:1}(\ref{pr:j:1c},\ref{pr:j:1d}),
we immediately obtain the following corollary:
\begin{corollary}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
If either $\inf f > -\infty$ or $f<+\infty$
then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
\end{corollary}
It is indeed possible that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}\neqf^{*\dualstar}$.
For example, consider the function $f$ in
\eqref{eqn:ex:1} whose conjugate is given in
\eqref{eqn:ex:1:conj}.
Let $\overline{\ee}=\limray{{\bf e}_1}$.
By direct calculation,
for all ${\bf x}\in{\mathbb{R}}^2$,
$f^{*\dualstar}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=-\infty$
while
\begin{equation} \label{eq:j:2}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}) = \left\{ \begin{array}{cl}
-\infty & \mbox{if $x_2\geq 0$} \\
+\infty & \mbox{otherwise.}
\end{array} \right.
\end{equation}
Thus, $+\infty={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})\neqf^{*\dualstar}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=-\infty$
if $x_2<0$.
\eqref{eq:j:2} also shows that the reduction
$\fshad{\ebar}$ is not closed;
thus, consistent with Theorem~\ref{thm:dub-conj-new}, $f$ does not
have all reductions closed.
In general, even if $f$ does not have all reductions closed, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
and $f^{*\dualstar}$ must agree at every
point in the closure of the effective domain of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ (which is the same as the closure of $\dom f$,
by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2c})), as well as
every point $\overline{\xx}\in\extspac{n}$ with $f^{*\dualstar}(\overline{\xx})>-\infty$.
This means that
at points $\overline{\xx}$ where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ and $f^{*\dualstar}$ differ, we can
say exactly what values each function will take, namely,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$ and $f^{*\dualstar}(\overline{\xx})=-\infty$,
as was the case in the preceding example.
\begin{theorem} \label{thm:fext-neq-fdub}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, and let $\overline{\xx}\in\extspac{n}$.
If $\overline{\xx}\in\cldom{f}$ or $f^{*\dualstar}(\overline{\xx})>-\infty$
then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$.
Consequently,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\neqf^{*\dualstar}(\overline{\xx})$ if and only if
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$ and $f^{*\dualstar}(\overline{\xx})=-\infty$.
\end{theorem}
\proof
The proof is similar to the first part of the proof of
Theorem~\ref{thm:dub-conj-new}.
By Propositions~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2c})
and~\ref{pr:fextstar-is-fstar},
$\cldom{f}=\cldom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}$ and ${f^*}={{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}$,
which means the theorem's claim can be stated entirely in terms of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Thus,
since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=\overline{(\lsc f)}$
(by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa})),
it suffices to prove the result for $\lsc f$.
Therefore, without loss of generality, we assume henceforth that $f$
is lower semicontinuous.
First, suppose that $f$ is not closed,
and that either
$\overline{\xx}\in\cldom{f}$ or $f^{*\dualstar}(\overline{\xx})>-\infty$.
Under these assumptions, we show that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$.
In this case,
as noted in the proof of Proposition~\ref{pr:j:1}(\ref{pr:j:1b}),
since $f$ is lower semicontinuous but not closed,
$f$ must be infinite everywhere so that $f({\bf x})\in\{-\infty,+\infty\}$
for all ${\bf x}\in{\R^n}$, and furthermore, $f\not\equiv+\infty$.
These facts imply ${f^*}\equiv+\infty$ so $f^{*\dualstar}\equiv-\infty$.
From our assumptions, this further implies that $\overline{\xx}\in\cldom{f}$,
which means there exists a sequence $\seq{{\bf x}_t}$ in $\dom{f}$ with
${\bf x}_t\rightarrow\overline{\xx}$.
For all $t$,
since $f({\bf x}_t)<+\infty$, the foregoing implies we must actually have
$f({\bf x}_t)=-\infty$, so
\[
-\infty
=
\liminf f({\bf x}_t)
\geq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
\]
by definition of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Therefore, as claimed,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=-\infty=f^{*\dualstar}(\overline{\xx})$ in this case.
As in the proof of
Theorem~\ref{thm:dub-conj-new}, we prove the general case by induction
on the astral rank of $\overline{\xx}$.
Specifically,
we prove by induction on $k=0,\ldots,n$ that for every
lower semicontinuous, convex function $f$,
and for all $\overline{\xx}\in\extspac{n}$,
if $\overline{\xx}$ has astral rank $k$,
and if either
$\overline{\xx}\in\cldom{f}$ or $f^{*\dualstar}(\overline{\xx})>-\infty$,
then
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$.
Let $f$ be such a function and let $\overline{\xx}$ have astral rank $k$
with
$\overline{\xx}\in\cldom{f}$ or $f^{*\dualstar}(\overline{\xx})>-\infty$.
We further assume that $f$ is closed, since the case that $f$ is not
closed was handled above.
In the base case that $k=0$, $\overline{\xx}$ is a point ${\bf x}\in{\R^n}$.
Since $f$ is closed,
\eqref{eq:thm:dub-conj-new:1} holds by the same reasoning as in the proof of
Theorem~\ref{thm:dub-conj-new},
proving ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$ in this case.
For the inductive step when $k>0$,
by Proposition~\ref{pr:h:6}, we can write
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}$ where ${\bf v}\in{\R^n}$ is $\overline{\xx}$'s
dominant direction, and $\xbar^{\bot}$, the projection of $\overline{\xx}$
perpendicular to ${\bf v}$, has astral rank $k-1$.
If ${\bf v}\not\in\resc{f}$, then Theorem~\ref{thm:i:4}
immediately implies that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty=f^{*\dualstar}(\overline{\xx})$ in this case.
Therefore, we assume henceforth that ${\bf v}\in\resc{f}$.
Let
$g=\fshad{\limray{\vv}}$ be the reduction of $f$ at $\limray{{\bf v}}$,
which is convex and lower semicontinuous
(Corollary~\ref{cor:i:1}(\ref{cor:i:1a})).
Then
$ {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) = \overline{g}(\xbar^{\bot})$
by Corollary~\ref{cor:i:1}(\ref{cor:i:1b}),
and
$ g^{*\dualstar}(\xbar^{\bot}) = f^{*\dualstar}(\overline{\xx}) $
by Theorem~\ref{thm:i:5}.
Therefore, to prove
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$, it suffices to show
$\overline{g}(\xbar^{\bot})=g^{*\dualstar}(\xbar^{\bot})$.
If $f^{*\dualstar}(\overline{\xx})>-\infty$ then
$ g^{*\dualstar}(\xbar^{\bot}) > -\infty $,
so that, by inductive hypthesis,
$\overline{g}(\xbar^{\bot})=g^{*\dualstar}(\xbar^{\bot})$.
Otherwise, we must have $\overline{\xx}\in\cldom{f}$,
meaning there exists a sequence $\seq{{\bf x}_t}$ in $\dom f$ with
${\bf x}_t\rightarrow\overline{\xx}$.
Thus, for each $t$,
by Proposition~\ref{pr:d2}(\ref{pr:d2:a},\ref{pr:d2:b}),
$g({\xx_t^\bot})\leq f({\bf x}_t)<+\infty$,
so ${\xx_t^\bot}\in\dom g$.
Since ${\xx_t^\bot}\rightarrow\xbar^{\bot}$
(by Proposition~\ref{pr:h:5}(\ref{pr:h:5b})),
this means $\xbar^{\bot}\in\cldom{g}$.
Therefore, by inductive hypthesis,
$\overline{g}(\xbar^{\bot})=g^{*\dualstar}(\xbar^{\bot})$
in this case as well.
This completes the induction and the proof.
\qed
\subsection{A dual characterization of all-reductions closedness}
\label{sec:dual-char-ent-clos}
We next give a dual characterization of when a function $f$ has
all reductions closed, and thus when ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
In Section~\ref{sec:slopes}, we defined $\slopes{f}$, the
barrier cone of $f$, which (except when $f\equiv+\infty$)
includes all of $\cone(\dom{{f^*}})$
as well as $f$'s vertical barrier cone, $\vertsl{f}$.
Assuming $f\not\equiv+\infty$,
we show now that $f$ has all reductions closed if and only if
$\slopes{f}=\cone(\dom{{f^*}})$, that is,
if and only if $f$'s vertical barrier cone is already entirely included
in $\cone(\dom{{f^*}})$.
Whereas all-reductions closedness would appear to be an astral property
involving the behavior of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ on every astral galaxy, this
shows that actually the property can be precisely characterized just in
terms of $f$ and its conjugate.
We prove each direction of the characterization separately.
\begin{theorem} \label{thm:slopescone-implies-entclosed}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, lower semicontinuous and not
identically $+\infty$.
If $\slopes{f}=\cone(\dom{{f^*}})$
then $f$ has all reductions closed.
\end{theorem}
\proof
We prove the theorem in the contrapositive:
Assuming $f$ does not have all reductions closed, we prove that
\begin{equation} \label{eq:thm:slopescone-implies-entclosed:1}
\slopes{f}\not\subseteq\cone(\dom{{f^*}}).
\end{equation}
By definition, $f$ does not have all reductions closed if and only if
one of its reductions $\fshad{\ebar}$ is not closed, for some
$\overline{\ee}\in{{\cal E}_n}$.
The proof is by induction on the astral rank of $\overline{\ee}$
(which, like that of all points in $\extspac{n}$, cannot exceed $n$).
More precisely, we prove the following by induction on $k=0,\ldots,n$:
For every function $f$ satisfying the conditions of the theorem's
statement, if $\fshad{\ebar}$ is not closed for some $\overline{\ee}\in{{\cal E}_n}$ whose
astral rank is at most $k$, then
\eqref{eq:thm:slopescone-implies-entclosed:1} holds.
In the base case that $k=0$,
${\bf 0}$ is the only icon of astral rank $0$
(Proposition~\ref{pr:i:8}(\ref{pr:i:8b})).
So suppose $f=\lsc f=\fshad{{\bf 0}}$ is not closed.
Since $f$ is lower semicontinuous but not closed,
we must have $f({\bf q})=-\infty$ at some point ${\bf q}\in{\R^n}$.
Therefore, ${f^*}\equiv+\infty$, so $\dom{{f^*}}=\emptyset$ and
$\cone(\dom{{f^*}})=\{{\bf 0}\}$.
Suppose, by way of contradiction, that
\eqref{eq:thm:slopescone-implies-entclosed:1}
does not hold.
Then $\slopes{f}=\{{\bf 0}\}$, implying $f$'s recession cone must be
$\resc{f}=\polar{(\slopes{f})}={\R^n}$,
by Corollary~\ref{cor:rescpol-is-slopes}.
In turn, this implies that $f$ is constant
(since $f({\bf x})\leq f({\bf y})$ for all ${\bf x},{\bf y}\in{\R^n}$), so that
$f\equiv-\infty$.
But this function is closed,
contradicting our assumption that $f$ is not closed.
Therefore,
\eqref{eq:thm:slopescone-implies-entclosed:1} holds.
For the inductive step, we assume $k>0$ and that the inductive
hypothesis holds for $k-1$.
Suppose $f$ (satisfying the conditions of the theorem) is such that
its reduction
$\fshad{\ebar}$ is not closed for some $\overline{\ee}\in{{\cal E}_n}$ with
astral rank $k$.
(We assume without loss of generality that the astral rank is exactly
$k$ since if it is less than $k$, then the claim follows by our
inductive hypothesis.)
We prove
\eqref{eq:thm:slopescone-implies-entclosed:1}
by considering several cases.
First, suppose
$f({\bf q})=-\infty$ for some ${\bf q}\in{\R^n}$.
If $f$ is closed, then we must have $f\equiv-\infty$ implying
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}\equiv-\infty$ and so $\fshad{\ebar}\equiv-\infty$,
contradicting that $\fshad{\ebar}$ is not closed.
Therefore, $f=\fshad{{\bf 0}}$ is not closed.
Since ${\bf 0}$ is an icon with astral rank $0\leq k-1$,
our inductive hypothesis then implies
\eqref{eq:thm:slopescone-implies-entclosed:1}
in this case.
We therefore assume henceforth that we are in the alternative case
that $f>-\infty$, implying $f$ is closed and proper.
Let ${\bf v}$ be $\overline{\ee}$'s dominant direction (which exists since $\overline{\ee}$
is infinite, being of positive astral rank).
Then $\overline{\ee}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\ebar^{\bot}$, where $\ebar^{\bot}$ is
the projection of $\overline{\ee}$ perpendicular to ${\bf v}$.
Let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Note that $\gshad{\ebar^{\bot}}=\fshad{\overline{\ee}}$ by
Proposition~\ref{pr:j:1}(\ref{pr:j:1a}).
Therefore, $\gshad{\ebar^{\bot}}$ is not closed.
This also implies that ${\bf v}\in\resc{f}$ since otherwise, we would have
$g\equiv+\infty$, by Theorem~\ref{thm:i:4}, so that
$\overline{g}\equiv+\infty$ and so
$\gshad{\ebar^{\bot}}\equiv+\infty$, contradicting that
$\gshad{\ebar^{\bot}}$ is not closed.
Thus, $g$ satisfies the conditions of the theorem
(by Proposition~\ref{pr:d2}(\ref{pr:d2:b}) and
Theorem~\ref{thm:a10-nunu}).
Since $\ebar^{\bot}$ has astral rank $k-1$
(Proposition~\ref{pr:h:6}),
we therefore can apply our
inductive hypothesis to $g$, which implies
there exists a point
${\bf u}\in(\slopes{g})\setminus\cone(\dom{{g^*}})$.
By Corollary~\ref{cor:thm:e1:slopes},
$ \slopes{g} = (\slopes{f}) \cap L $
where
$L=\{{\bf x}\in{\R^n} : {\bf x}\cdot{\bf v}=0\}$
is the linear subspace perpendicular to ${\bf v}$.
Therefore, ${\bf u}\in\slopes{f}$ and ${\bf u}\in L$.
We will complete the proof by showing that
${\bf u}\not\in\cone(\dom{{f^*}})$, proving
\eqref{eq:thm:slopescone-implies-entclosed:1}.
We consider separately the subcases that $g>-\infty$ or
$g\not> -\infty$.
Suppose first that $g>-\infty$.
In this case, Theorem~\ref{thm:e1} implies that
$\cone(\dom{{g^*}}) = \cone(\dom{{f^*}}) \cap L$.
Since ${\bf u}\in L$, but ${\bf u}\not\in\cone(\dom{{g^*}})$, this implies
that ${\bf u}\not\in\cone(\dom{{f^*}})$.
Otherwise, $g({\bf q})=-\infty$ for some ${\bf q}\in{\R^n}$.
Suppose, by way of contradiction, that
${\bf u}\in\cone(\dom{{f^*}})$.
Then $\lambda{\bf u}\in\dom{{f^*}}$ for some $\lambda>0$.
(See \eqref{eq:cone-dom-fstar}, noting that ${\bf u}\neq{\bf 0}$ since
${\bf u}\not\in\cone(\dom{{g^*}})$.)
Thus,
\begin{eqnarray*}
-\infty
=
g({\bf q})
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
\\
&\geq&
f^{*\dualstar}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
\\
&\geq&
-{f^*}(\lambda{\bf u})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})\cdot(\lambda{\bf u})
\\
&=&
-{f^*}(\lambda{\bf u})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf q}\cdot{\bf u}.
\end{eqnarray*}
The two inequalities are applications of
Theorem~\ref{thm:fext-dub-sum}.
The final equality is because ${\bf u}\in L$, implying ${\bf v}\cdot{\bf u}=0$.
However, this is a contradiction since ${f^*}(\lambda{\bf u})<+\infty$,
so that the final expression on the right cannot be $-\infty$.
Therefore, ${\bf u}\not\in\cone(\dom{{f^*}})$, completing the induction
and the proof.
\qed
Next, we prove the converse:
\begin{theorem} \label{thm:entclosed-implies-slopescone}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, lower semicontinuous and not
identically $+\infty$.
If $f$ does not have all reductions closed
then $\slopes{f}=\cone(\dom{{f^*}})$.
\end{theorem}
\proof
We prove the contrapositive, showing that if
$\slopes{f}\neq\cone(\dom{{f^*}})$ then
$f$ does not have all reductions closed.
Proof is by induction on the dimension of $\cone(\dom{{f^*}})$
(or equivalently, $\dom{{f^*}}$).
That is, we prove by induction on $k=0,\ldots,n$ that
for every function $f$ satisfying the conditions of the theorem's
statement, if $\cone(\dom{{f^*}})\subseteq L$ for some linear
subspace $L\subseteq{\R^n}$ of dimension at most $k$, and if
$\slopes{f}\neq\cone(\dom{{f^*}})$ then
$f$ does not have all reductions closed.
In all that follows, we assume that
$\slopes{f}\neq\cone(\dom{{f^*}})$,
which means there exists some point ${\bf u}$
in $(\slopes{f})\setminus\cone(\dom{{f^*}})$
(since $\cone(\dom{{f^*}})\subseteq\slopes{f}$ always,
by Proposition~\ref{pr:slopes-equiv}).
We also assume the other conditions of the induction
(that $f$ satisfies the conditions of the theorem, and that
$\cone(\dom{{f^*}})$ is included in some linear
subspace $L\subseteq{\R^n}$ of dimension at most $k$).
We begin with the base case that $k=0$.
Then
$\cone(\dom{{f^*}})\subseteq L =\{{\bf 0}\}$,
which means $\dom{{f^*}}$ is either $\emptyset$ or $\{{\bf 0}\}$.
If $\dom{{f^*}}=\{{\bf 0}\}$ then
$\inf f = -{f^*}({\bf 0}) > -\infty$, implying that $f$ is closed and
proper.
Therefore,
\[
\{{\bf 0}\}
=
\cl(\cone(\dom{{f^*}}))
=
\rescpol{f}
=
\cl(\slopes{f})
\]
by Theorem~\ref{thm:rescpol-is-con-dom-fstar}
and Corollary~\ref{cor:rescpol-is-slopes}.
However, this is only possible if $\slopes{f}=\{{\bf 0}\}$,
contradicting our assumption that
$\slopes{f}\neq\cone(\dom{{f^*}})$.
So this case is ruled out.
Otherwise, $\dom{{f^*}}=\emptyset$ so ${f^*}\equiv+\infty$
and $\cl f=f^{**}\equiv-\infty$.
Also,
by Proposition~\ref{pr:slopes-equiv},
since ${\bf u}$ is in $\slopes{f}$ but not
$\cone(\dom{{f^*}})$, it must be in $f$'s vertical barrier cone,
$\vertsl{f}$.
This means there exists $\beta\in{\mathbb{R}}$ such that for all ${\bf x}\in{\R^n}$, if
${\bf x}\cdot{\bf u}>\beta$ then $f({\bf x})=+\infty$.
Note that ${\bf u}\neq{\bf 0}$ since
${\bf u}\not\in\cone(\dom{{f^*}})=\{{\bf 0}\}$.
Thus, if we let ${\bf x}=\lambda{\bf u}$ where
$\lambda>\beta/\norm{{\bf u}}^2$ then ${\bf x}\cdot{\bf u}>\beta$ and
$f({\bf x})=+\infty$.
Since $f({\bf x})\neq (\cl f)({\bf x})$, this shows that
$f=\lsc f =\fshad{{\bf 0}}$ is
not closed, and
therefore $f$ does not have all reductions closed.
For the inductive step, we assume the conditions of
the induction hold with $k>0$.
If $f\not> -\infty$ then ${f^*}\equiv+\infty$ so
$\cone(\dom{{f^*}})=\{{\bf 0}\}$ and thus is included in a linear
subspace of dimension $0$.
Since $0\leq k-1$, this implies $f$ does not have all reductions
closed by inductive
hypothesis in this case.
Otherwise, if $f>-\infty$ then $f$ is closed and proper.
Therefore,
\[
{\bf u}
\in
\slopes{f}
\subseteq
\cl(\slopes{f})
=
\rescpol{f}
=
\cl(\cone(\dom{{f^*}}))
\]
by Theorem~\ref{thm:rescpol-is-con-dom-fstar}
and Corollary~\ref{cor:rescpol-is-slopes}.
On the other hand, ${\bf u}\not\in\cone(\dom{{f^*}})$.
Therefore, ${\bf u}$ is a relative boundary point of
$\cone(\dom{{f^*}})$.
(Recall that the \emph{relative boundary} of a convex set
$S\subseteq{\R^n}$ is the set
$(\cl S) \setminus (\ri S)$
consisting of all points that are in the closure of $S$
but not its relative interior.)
Before continuing, we pause to prove the following standard lemma
regarding convex cones and their relative boundary points:
\begin{lemma} \label{lem:rel-bnd-convcone}
Let $K\subseteq{\R^n}$ be a pointed convex cone, and
let ${\bf u}\in(\cl K) \setminus(\ri K)$ be a relative boundary point of
$K$.
Then there exists a vector ${\bf v}\in{\R^n}$ such that all of the
following hold:
\begin{enumerate}
\item
${\bf u}\cdot{\bf v} = 0$.
\item
${\bf x}\cdot{\bf v} \leq 0$ for all ${\bf x}\in K$.
That is, ${\bf v}\in \polar{K}$.
\item
${\bf y}\cdot{\bf v} < 0$ for some ${\bf y}\in K$.
\end{enumerate}
\end{lemma}
\proof
Because $K$ is convex and ${\bf u}$ is a relative boundary point, there
exists a linear function that is not constant
on $K$ and that achieves its maximum at ${\bf u}$.
(See \citet[Corollary~11.6.2]{ROC}.)
In other words, there exists ${\bf v}\in{\R^n}$ and $b\in{\mathbb{R}}$ such that
${\bf u}\cdot{\bf v}=b$ and ${\bf x}\cdot{\bf v}\leq b$ for all ${\bf x}\in K$,
with strict inequality for at least one point ${\bf y}$ in $K$.
We claim that $b=0$.
To see this, note first, since ${\bf 0}\in K$, that
$0={\bf 0}\cdot{\bf v}\leq b$, so $b\geq 0$.
On the other hand, for any $\epsilon>0$, since ${\bf u}\in\cl K$ and by
continuity of ${\bf x}\mapsto{\bf x}\cdot{\bf v}$,
there exists a point ${\bf z}\in K$
with ${\bf z}\cdot{\bf v} > b-\epsilon$.
Since $K$ is a cone, $2{\bf z}$ is also in $K$, implying
$2(b-\epsilon) < 2{\bf z}\cdot{\bf v} \leq b$, that is,
$b < 2\epsilon$.
Since this is true for all $\epsilon>0$, it follows that $b=0$.
Thus, ${\bf u}\cdot{\bf v}=0$.
Also, ${\bf x}\cdot{\bf v}\leq 0$ for all ${\bf x}\in K$, which means, by its
definition, that ${\bf v}$ is in $\polar{K}$.
In addition, there exists some point ${\bf y}\in K$ with ${\bf y}\cdot{\bf v}<0$.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Since ${\bf u}$ is a relative boundary point of the pointed convex cone
$\cone(\dom{{f^*}})$, Lemma~\ref{lem:rel-bnd-convcone} implies that
there exists ${\bf v}\in{\R^n}$ with ${\bf u}\cdot{\bf v}=0$ and ${\bf y}\cdot{\bf v}<0$ for
some ${\bf y}\in\cone(\dom{{f^*}})$.
Furthermore,
${\bf v}\in \polar{(\cone(\dom{{f^*}}))} = \resc{f}$
(by Theorem~\ref{thm:rescpol-is-con-dom-fstar}).
Let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Also, let
$M=\{{\bf x}\in{\R^n} : {\bf x}\cdot{\bf v}=0\}$
be the linear subspace perpendicular to ${\bf v}$,
and let $L'=L\cap M$.
Note that ${\bf y}\in\cone(\dom{{f^*}})\subseteq L$ but
${\bf y}\not\in M$ so ${\bf y}\not\in L'$.
Thus, $L'\subseteq L$ but $L'\neq L$ so $L'$ has dimension strictly
less than $L$.
By Corollary~\ref{cor:thm:e1:slopes},
$ \slopes{g} = (\slopes{f}) \cap M $.
As already noted, ${\bf u}\in M$, so
this implies
${\bf u}\in\slopes{g}$ since ${\bf u}\in\slopes{f}$.
Also,
$\cone(\dom{{g^*}}) \subseteq \cone(\dom{{f^*}}) \cap M$
by Theorem~\ref{thm:e1} which implies that
$\cone(\dom{{g^*}}) \subseteq L\cap M = L'$.
Further, since ${\bf u}\not\in\cone(\dom{{f^*}})$ this also implies
${\bf u}\not\in\cone(\dom{{g^*}})$.
Combining facts,
$g$ satisfies the conditions of the theorem
(by Proposition~\ref{pr:d2}(\ref{pr:d2:b}) and
Theorem~\ref{thm:a10-nunu}) with
$\slopes{g}\neq\cone(\dom{{g^*}})$
and $\cone(\dom{{g^*}})\subseteq L'$ where $L'$ has dimension at most
$k-1$.
Therefore, by inductive hypothesis, $g$ does not have all reductions
closed.
Therefore, $f$ also does not have all reductions closed, by
Proposition~\ref{pr:j:1}(\ref{pr:j:1a}).
This completes the induction and the proof.
\qed
Combined with Theorem~\ref{thm:dub-conj-new}, we thus have proved:
\begin{corollary} \label{cor:ent-clos-is-slopes-cone}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then the following are equivalent:
\begin{enumerate}
\item \label{cor:ent-clos-is-slopes-cone:a}
$f$ has all reductions closed.
\item \label{cor:ent-clos-is-slopes-cone:b}
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$.
\item \label{cor:ent-clos-is-slopes-cone:c}
Either $f\equiv+\infty$ or $\slopes{f}=\cone(\dom{{f^*}})$.
\end{enumerate}
\end{corollary}
\proof
By Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa}),
Proposition~\ref{pr:slopes-same-lsc},
and since
${f^*}=(\lsc f)^*$, we can assume without loss of generality that $f$
is lower semicontinuous (replacing it with $\lsc f$ if it is not).
The equivalence of
(\ref{cor:ent-clos-is-slopes-cone:a})
and
(\ref{cor:ent-clos-is-slopes-cone:b})
was proved in Theorem~\ref{thm:dub-conj-new}.
If $f\equiv+\infty$ then $f$ has all reductions closed.
Otherwise, when $f\not\equiv+\infty$,
Theorems~\ref{thm:slopescone-implies-entclosed}
and~\ref{thm:entclosed-implies-slopescone}
prove the equivalence of
(\ref{cor:ent-clos-is-slopes-cone:a})
and
(\ref{cor:ent-clos-is-slopes-cone:c}).
\qed
\section{Convex sets}
\label{sec:convexity}
We next study how the notion of convexity can be extended to the
astral space.
In ${\R^n}$, a set is convex if for every pair of points ${\bf x},{\bf y}$ in the
set, their convex combination
$\lambda{\bf x}+(1-\lambda){\bf y}$ is also in the set, for any
$\lambda\in [0,1]$.
Thus, as a natural first attempt, we might try to define
convexity in astral space in a similar way in terms of some kind
of convex combination
$\overline{{\bf z}}=\lambda\overline{\xx}+(1-\lambda)\overline{{\bf y}}$
of two points $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$.
We have not actually defined the ordinary sum of two astral points, as
in this expression, nor will we define it.
But if we could,
then for any ${\bf u}\in{\R^n}$, we would naturally want it
to be the case that
\[
\overline{{\bf z}}\cdot{\bf u} = \lambda \overline{\xx}\cdot{\bf u} +
(1-\lambda) \overline{{\bf y}}\cdot{\bf u}.
\]
The problem is that the right-hand side might easily involve
adding $+\infty$ and $-\infty$
(for instance, if $\overline{\xx}\cdot{\bf u}=+\infty$,
$\overline{{\bf y}}\cdot{\bf u}=-\infty$, and $\lambda\in (0,1)$),
which is not defined.
So this approach would seem to be doomed.
Here we present an alternative that avoids this difficulty.
\subsection{Defining convexity}
\label{sec:def-convexity}
We begin by noting that there is a very natural way to define
hyperplanes and halfspaces in astral space.
In ${\R^n}$, a closed halfspace is the set of points ${\bf x}$ for which
${\bf x}\cdot{\bf u}\leq a$
for some ${\bf u}\in{\R^n}$ and $a\in{\mathbb{R}}$.
We can immediately extend this to astral space, defining a
closed halfspace, denoted $\chs{\uu}{a}$, to be the set
\begin{equation} \label{eq:chsua-defn}
\chs{\uu}{a} = \{\overline{\xx}\in\extspac{n} : \overline{\xx}\cdot{\bf u} \leq a\}.
\end{equation}
In the same way, open halfspaces and hyperplanes in $\extspac{n}$ have
the form
\[
\{\overline{\xx}\in\extspac{n} : \overline{\xx}\cdot{\bf u}<a\}
\]
and
\[
\{\overline{\xx}\in\extspac{n} : \overline{\xx}\cdot{\bf u}=a\},
\]
respectively.
Open halfspaces are indeed open, being standard basis elements
(\eqref{eq:h:3a}), and closed halfspaces are closed, being the
complement of open halfspaces.
(As usual, these forms accomodate halfspaces defined by reverse
inequalities since $\overline{\xx}\cdot{\bf u}\geq a$ if and only if
$\overline{\xx}\cdot(-{\bf u})\leq -a$.)
Also, although we usually take $a$ to be in ${\mathbb{R}}$, as above, it is
sometimes useful to allow $a$ to be $\pm\infty$.
The intersection of a closed or open astral halfspace with ${\R^n}$ is
just the corresponding closed or open halfspace in ${\R^n}$; for instance,
\begin{equation} \label{eqn:chsua-cap-rn}
\chs{\uu}{a}\cap{\R^n} = \{{\bf x}\in{\R^n} : {\bf x}\cdot{\bf u} \leq a\}.
\end{equation}
The closure (in $\extspac{n}$) of a closed or open halfspace like this one in
${\R^n}$ is exactly the closed astral halfspace $\chs{\uu}{a}$.
(See Proposition~\ref{pr:halfspace-closure}.)
Similar facts hold for hyperplanes.
The starting point for standard convex analysis is the line segment
joining two points ${\bf x}$ and ${\bf y}$.
In ${\R^n}$, this is the set of all convex combinations of the two
points, a perspective that, as discussed above, does not immediately
generalize to astral space.
However, there is another way of thinking about the line segment
joining ${\bf x}$ and ${\bf y}$, namely, as the intersection of all
halfspaces that include both of the endpoints.
This interpretation generalizes directly to astral space.
Thus, having defined astral halfspaces,
for any two points $\overline{\xx}$ and $\overline{{\bf y}}$ in $\extspac{n}$,
we can define
$\lb{\overline{\xx}}{\overline{{\bf y}}}$, the \emph{segment joining} $\overline{\xx}$ and $\overline{{\bf y}}$,
to be the intersection of all halfspaces
containing $\overline{\xx}$ and $\overline{{\bf y}}$.
In more detail,
for any set $S\subseteq\extspac{n}$, we define $S$'s
\emph{outer convex hull} (or \emph{outer hull}, for short),
denoted $\ohull S$, to be the
intersection of all closed halfspaces that include $S$:
\[
\ohull S = \bigcap_{\scriptontop
{{\bf u}\in{\R^n}, a\in{\mathbb{R}}:}
{S\subseteq \chs{\uu}{a}}
} \chs{\uu}{a}.
\]
In these terms, the segment joining $\overline{\xx}$ and $\overline{{\bf y}}$ is exactly their
outer hull:
\[\lb{\overline{\xx}}{\overline{{\bf y}}} = \ohull\{\overline{\xx},\overline{{\bf y}}\}.\]
The outer hull of any set $S\subseteq\extspac{n}$ can in general be
written in a form that is often more useful, even in this case where
$S$ consists of only two points.
Since the closed halfspace $\chs{\uu}{a}$ consists of all points $\overline{{\bf s}}$
for which $\overline{{\bf s}}\cdot{\bf u}\leq a$,
$S\subseteq\chs{\uu}{a}$ if and only if
$\overline{{\bf s}}\cdot{\bf u}\leq a$ for all $\overline{{\bf s}}\in S$, that is,
if and only if
$\sup_{\overline{{\bf s}}\in S} \overline{{\bf s}}\cdot{\bf u} \leq a$.
Thus, $\ohull S$, which is the intersection of all such halfspaces,
consists exactly of those points $\overline{{\bf z}}\in\extspac{n}$ for which
$\overline{{\bf z}}\cdot{\bf u} \leq \sup_{\overline{{\bf s}}\in S} \overline{{\bf s}}\cdot{\bf u}$ for all
${\bf u}\in{\R^n}$.
We state this as a proposition:
\begin{proposition} \label{pr:ohull-simplify}
Let $S\subseteq\extspac{n}$.
Then
\[
\ohull S
=
\braces{\overline{{\bf z}}\in\extspac{n} : \forall {\bf u}\in{\R^n},
\overline{{\bf z}}\cdot{\bf u}\leq \sup_{\overline{{\bf s}}\in S} \overline{{\bf s}}\cdot{\bf u}}.
\]
\end{proposition}
In particular, from
Proposition~\ref{pr:ohull-simplify}, we see that
the segment joining $\overline{\xx}$ and $\overline{{\bf y}}$, which is equal
to $\ohull\{\overline{\xx},\overline{{\bf y}}\}$, is exactly the set
\begin{equation} \label{eqn:lb-def}
\lb{\overline{\xx}}{\overline{{\bf y}}}=
\braces{ \overline{{\bf z}}\in\extspac{n} :
\forall {\bf u}\in{\R^n},\, \overline{{\bf z}}\cdot{\bf u}
\leq\max\{\overline{\xx}\cdot{\bf u},\,\overline{{\bf y}}\cdot{\bf u}\} }.
\end{equation}
For example, in ${\mathbb{R}}^2$, let
$\overline{{\bf z}}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf e}_1$
(where ${\bf e}_1, {\bf e}_2$ are the standard basis vectors).
Then
\begin{equation} \label{eqn:eg-seg-zero-e2-plus-e1}
\lb{{\bf 0}}{\overline{{\bf z}}} =
\{ \lambda {\bf e}_2 : \lambda\geq 0 \}
\cup
\{ \limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf e}_1 : \lambda\in [0,1] \},
\end{equation}
since it can be argued that these are the only points that satisfy
\eqref{eqn:eg-seg-zero-e2-plus-e1}.
Alternatively, this follows directly
as a special case of Theorem~\ref{thm:lb-with-zero}, proved below.
As another example,
the segment joining $\limray{{\bf e}_1}$ and $\limray{{\bf e}_2}$
consists of all infinite points (that is,
in $\extspac{2}\setminus {\mathbb{R}}^2$) that can be
represented using only vectors in $\mathbb{R}_{\geq 0}^2$,
the upper-right quadrant of ${\mathbb{R}}^2$.
That is,
\[
\lb{\limray{{\bf e}_1}}{\limray{{\bf e}_2}}
= \braces{ \limray{{\bf v}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf v}_2} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q} \; :\;
{\bf v}_1,{\bf v}_2,{\bf q}\in \mathbb{R}_{\geq 0}^2
}
\; \setminus {\mathbb{R}}^2
\]
(using the fact that all points in $\extspac{2}$ have
astral rank at most 2).
This can be shown using
\eqref{eqn:eg-seg-zero-e2-plus-e1},
as before, or more easily
as a special case of Theorem~\ref{thm:conv-lmset-char} below.
So $\lb{\limray{{\bf e}_1}}{\limray{{\bf e}_2}}$
is entirely disjoint from ${\mathbb{R}}^2$ in this case.
As a last example, let ${\bf I}$ be the $n\times n$ identity matrix.
Then the segment joining the points
$-{\bf I} {\bar{\bm{\omega}}}$ and ${\bf I} {\bar{\bm{\omega}}}$
turns out to be all of $\extspac{n}$;
that is,
$ \lb{-{\bf I} {\bar{\bm{\omega}}}}{{\bf I} {\bar{\bm{\omega}}}}=\extspac{n} $.
To see this, let $\overline{{\bf z}}\in\extspac{n}$ and let ${\bf u}\in{\R^n}$.
If ${\bf u}\neq{\bf 0}$, then
${\bf I}{\bar{\bm{\omega}}}\cdot{\bf u}\in\{-\infty,+\infty\}$,
by
Proposition~\ref{pr:vtransu-zero},
so
\[
\overline{{\bf z}}\cdot{\bf u}\leq +\infty=\max\{-{\bf I}{\bar{\bm{\omega}}}\cdot{\bf u},\,{\bf I}{\bar{\bm{\omega}}}\cdot{\bf u}\}.
\]
Otherwise,
if ${\bf u}={\bf 0}$, then
$\overline{{\bf z}}\cdot{\bf 0}=0=\max\{-{\bf I}{\bar{\bm{\omega}}}\cdot{\bf 0},\,{\bf I}{\bar{\bm{\omega}}}\cdot{\bf 0}\}$.
Thus, in all cases, the inequality appearing in
\eqref{eqn:lb-def}
is satisfied, so
$\overline{{\bf z}}\in\lb{-{\bf I} {\bar{\bm{\omega}}}}{{\bf I} {\bar{\bm{\omega}}}}$.
This shows, in an extreme way, that the ``segment'' joining two
infinite astral
points can be very different from the standard, one-dimensional
segment joining points in ${\R^n}$.
Later,
in Section~\ref{sec:mono-passages}, we will explore a different
way of understanding segments (and so also convex sets) which we will
see are composed of structures that are,
in certain respects, more like ordinary line segments
in terms of their linear and unbending form.
Continuing our general development,
we are now ready to define convexity for subsets of astral space.
Extending the usual notion of convexity for subsets of ${\R^n}$,
we say that a set $S\subseteq\extspac{n}$ is \emph{astrally convex}
(or simply \emph{convex})
if for all $\overline{\xx},\overline{{\bf y}}\in S$, it is the case that the entire
segment joining them is also in $S$, that is,
\[ \lb{\overline{\xx}}{\overline{{\bf y}}} \subseteq S \;\; \mbox{ for all } \overline{\xx},\overline{{\bf y}}\in S. \]
Here are some simple properties:
\begin{proposition} \label{pr:e1}
~
\begin{enumerate}
\item \label{pr:e1:a}
If ${\bf x},{\bf y}\in{\R^n}$, then $\lb{{\bf x}}{{\bf y}}$ is the standard line segment
joining ${\bf x}$ and ${\bf y}$ in ${\R^n}$:
\[
\lb{{\bf x}}{{\bf y}} = \{(1-\lambda){\bf x} + \lambda{\bf y} :
\lambda\in [0,1]\}.
\]
Therefore, a subset of ${\R^n}$ is astrally convex if
and only if it is convex in the standard sense defined for ${\R^n}$.
\item \label{pr:e1:b}
The intersection of an arbitrary collection of astrally convex
sets is astrally convex.
\item \label{pr:e1:c}
Every hyperplane and closed or open halfspace in astral space is
astrally convex.
\item \label{pr:e1:d}
Every standard basis element of the topology on $\extspac{n}$
(as in \eqref{eq:h:3a})
is astrally convex.
Also, for any set $S\subseteq\extspac{n}$,
the outer hull, $\ohull S$, is astrally convex
and closed (in $\extspac{n}$).
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:e1:a}):
Suppose ${\bf x},{\bf y}\in{\R^n}$.
If ${\bf z}\in\lb{{\bf x}}{{\bf y}}$,
then for all ${\bf u}\in{\R^n}$, by \eqref{eqn:lb-def},
${\bf z}\cdot{\bf u}\leq \max\{{\bf x}\cdot{\bf u},{\bf y}\cdot{\bf u}\} < +\infty$,
implying that ${\bf z}\in{\R^n}$
(by Proposition~\ref{pr:i:3}).
Thus, $\lb{{\bf x}}{{\bf y}}\subseteq{\R^n}$, so
$\lb{{\bf x}}{{\bf y}}$ is included in some halfspace $\chs{\uu}{a}$ in $\extspac{n}$
if and only if it is included in the corresponding halfspace in ${\R^n}$
given in \eqref{eqn:chsua-cap-rn}.
It follows that $\lb{{\bf x}}{{\bf y}}$ is the intersection of all closed
halfspaces in ${\R^n}$ containing both ${\bf x}$ and ${\bf y}$.
By \citet[Corollary~11.5.1]{ROC}, this is exactly the convex hull of
$\{{\bf x},{\bf y}\}$, namely, the set of all their convex combinations
as given in the proposition.
This implies that, when restricted to subsets of ${\R^n}$, astral
convexity coincides exactly with the
standard definition of convexity in ${\R^n}$.
Part~(\ref{pr:e1:b}):
Let
\[ M = \bigcap_{\alpha\in{\cal I}} S_\alpha \]
where each $S_\alpha\subseteq\extspac{n}$ is convex, and ${\cal I}$ is an
arbitrary index set.
Let $\overline{\xx}$, $\overline{{\bf y}}$ be in $M$.
Then for all $\alpha\in{\cal I}$, $\overline{\xx},\overline{{\bf y}}\in S_\alpha$, so
$\lb{\overline{\xx}}{\overline{{\bf y}}}\subseteq S_\alpha$.
Since this holds for all $\alpha\in{\cal I}$,
$\lb{\overline{\xx}}{\overline{{\bf y}}}\subseteq M$, and $M$ is convex.
Part~(\ref{pr:e1:c}):
Suppose $\overline{\xx},\overline{{\bf y}}\in \chs{\uu}{a}$ and
let $\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$.
Then
\[ \overline{{\bf z}}\cdot{\bf u} \leq \max\{\overline{\xx}\cdot{\bf u},\overline{{\bf y}}\cdot{\bf u}\} \leq a. \]
The first inequality is from \eqref{eqn:lb-def},
since $\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$, and the
second is because $\overline{\xx},\overline{{\bf y}}\in \chs{\uu}{a}$.
Therefore $\overline{{\bf z}}\in \chs{\uu}{a}$, and so $\chs{\uu}{a}$ is convex.
The proof is the same for open halfspaces, but using strict
inequalities.
A hyperplane defined by $\overline{\xx}\cdot{\bf u}=a$ is convex by the above and
part~(\ref{pr:e1:b}) since it is equal to the
intersection of two closed halfspaces (since $\overline{\xx}\cdot{\bf u}= a$ if
and only if $\overline{\xx}\cdot{\bf u}\leq a$ and $\overline{\xx}\cdot(-{\bf u})\leq -a$).
Part~(\ref{pr:e1:d}):
It follows immediately from
parts~(\ref{pr:e1:b}) and~(\ref{pr:e1:c}) that an arbitrary
intersection of closed or open halfspaces is astrally convex.
Since a standard basis element, as in \eqref{eq:h:3a}, is an
intersection of finitely many open halfspaces, and since
$\ohull S$ is an intersection of closed halfspaces, the result follows.
\qed
We show next that if $S$ is any convex subset of ${\R^n}$, then its
astral closure $\overline{S}$ is also convex, and more specifically, is
exactly equal to the outer hull of $S$, the intersection of all
closed astral halfspaces that contain $S$.
\begin{theorem} \label{thm:e:6}
Let $S\subseteq{\R^n}$ be convex.
Then its astral closure, $\overline{S}$, is exactly equal to its outer
hull;
that is, $\overline{S}=\ohull S$.
Consequently, $\overline{S}$ is convex.
\end{theorem}
\proof
The result is immediate if $S=\emptyset$, so we assume henceforth that
$S$ is nonempty.
The set $\ohull S$ is closed (in $\extspac{n}$), being an intersection
of closed halfspaces.
Therefore, since $S\subseteq\ohull S$, its closure must also be in
$\ohull S$, that is,
$\overline{S}\subseteq \ohull S$.
For the reverse inclusion,
let $\indf{S}$ be the indicator function for $S$,
as defined in \eqref{eq:indf-defn}, which is convex since $S$ is
convex.
Let $\indstar{S}$ be its conjugate,
the support function for $S$,
as defined in \eqref{eq:e:2}.
Then by Proposition~\ref{pr:ohull-simplify}, we can express
$\ohull S$ in terms of $\indstar{S}$:
\begin{equation} \label{eq:e:4}
\ohull S
=
\braces{\overline{\xx}\in\extspac{n} : \forall {\bf u}\in{\R^n},
\overline{\xx}\cdot{\bf u}\leq \indstar{S}({\bf u}) }.
\end{equation}
Let $\overline{\indflet}_S$ be the lower semicontinuous extension
of $\indf{S}$ to astral space.
Note that $\indf{S}$ has all reductions closed by
Proposition~\ref{pr:j:1}(\ref{pr:j:1c}).
Therefore,
\begin{equation} \label{eq:e:3}
\overline{\indflet}_S(\overline{\xx}) =
\indflet^{*\dualstar}_S(\overline{\xx}) =
\sup_{{\bf u}\in{\R^n}} ( -\indstar{S}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u} )
\end{equation}
with the equalities following from Theorem~\ref{thm:dub-conj-new}
and Theorem~\ref{thm:fext-dub-sum},
respectively.
Suppose $\overline{\xx}\in \ohull S$.
For all ${\bf u}\in{\R^n}$,
we claim that the argument to the supremum in \eqref{eq:e:3} is never
positive.
This is immediate if $\indstar{S}({\bf u})=+\infty$.
If $\indstar{S}({\bf u})\in{\mathbb{R}}$, then since $\overline{\xx}\in \ohull S$,
by \eqref{eq:e:4},
$\overline{\xx}\cdot{\bf u}\leq\indstar{S}({\bf u})$,
and so
$-\indstar{S}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u} \leq 0$.
(The case $\indstar{S}({\bf u})=-\infty$ is impossible
since $S$ is not empty.)
Therefore, by \eqref{eq:e:3}, $\overline{\indflet}_S(\overline{\xx})\leq 0$.
By Proposition~\ref{pr:d1}, this means there exists a sequence
$\seq{{\bf x}_t}$ in ${\R^n}$ converging to $\overline{\xx}$ for which
$\indf{S}({\bf x}_t)\rightarrow \overline{\indflet}_S(\overline{\xx})\leq 0$.
This is only possible if $\indf{S}({\bf x}_t)=0$ for almost all $t$, in other
words, if ${\bf x}_t\in S$ for all but finitely many values of $t$.
But this means exactly that their limit $\overline{\xx}$ is in $\overline{S}$.
Therefore, $\ohull S\subseteq\overline{S}$,
completing the proof that $\overline{S}= \ohull S$.
That $\overline{S}$ is convex now follows from
Proposition~\ref{pr:e1}(\ref{pr:e1:d}).
\qed
In general,
Theorem~\ref{thm:e:6} does not hold for arbitrary convex sets in
$\extspac{n}$ (rather than in ${\R^n}$), as will be seen in
Section~\ref{sec:closure-convex-set}.
As a quick (and unsurprising) application, the closure of a closed or
open halfspace, whether in ${\R^n}$ or $\extspac{n}$, is equal to the
corresponding closed astral halfspace:
\begin{proposition} \label{pr:halfspace-closure}
Let $S = \{{\bf x}\in{\R^n} : {\bf x}\cdot{\bf u} < a\}$ where ${\bf u}\in{\R^n}$,
$a\in{\mathbb{R}}$, and ${\bf u}\neq{\bf 0}$.
Then $\overline{S} = \chs{\uu}{a}$ (as defined in \eqref{eq:chsua-defn}).
The same holds if $S$ is replaced by
the corresponding closed halfspace in ${\R^n}$ or
the corresponding open halfspace in $\extspac{n}$.
\end{proposition}
\proof
We assume without loss of generality that $\norm{{\bf u}}=1$
(dividing ${\bf u}$ and $a$ by $\norm{{\bf u}}$ if that is not the case).
To prove the proposition,
we compute $\ohull S$, the intersection of all closed halfspaces that
include $S$, which is the same as $\overline{S}$ by
Theorem~\ref{thm:e:6}.
Suppose $S\subseteq \chs{\ww}{b}$, for some ${\bf w}\in{\R^n}$ and $b\in{\mathbb{R}}$,
where again, we assume $\norm{{\bf w}}=1$.
(If ${\bf w}={\bf 0}$ then $\chs{\ww}{b}$ is either $\emptyset$ or $\extspac{n}$,
both cases that can be ignored.)
Let $c={\bf u}\cdot{\bf w}$ and ${\bf v}={\bf w}-c{\bf u}$ so that ${\bf w}=c{\bf u}+{\bf v}$ and
${\bf u}\cdot{\bf v}=0$.
Let ${\bf x}=(a-1){\bf u}+d{\bf v}$ for any $d\in{\mathbb{R}}$.
Then ${\bf x}\cdot{\bf u}=a-1<a$ so ${\bf x}\in S\subseteq \chs{\ww}{b}$,
implying ${\bf x}\cdot{\bf w}=c(a-1)+d\norm{{\bf v}}^2\leq b$.
Since this holds for all $d\in{\mathbb{R}}$, it must be that ${\bf v}={\bf 0}$.
Therefore, since $\norm{{\bf w}}=\norm{{\bf u}}=1$, ${\bf w}$ must be either
$-{\bf u}$ or ${\bf u}$.
We claim only the latter case is possible.
Otherwise,
if ${\bf w}=-{\bf u}$, then let ${\bf x}=(\min\{a,-b\}-1){\bf u}$
so that ${\bf x}\cdot{\bf u}\leq a-1 < a$, so ${\bf x}\in S$,
but ${\bf x}\cdot{\bf w}\geq b+1 > b$, so ${\bf x}\not\in\chs{\ww}{b}$, contradicting
$S\subseteq \chs{\ww}{b}$.
Thus, $S\subseteq\chs{\ww}{b}$ if and only if ${\bf w}={\bf u}$ and also
$b\geq \sup_{{\bf x}\in S} {\bf x}\cdot{\bf u} = a$.
Taking the intersection of all such halfspaces,
it therefore follows from Theorem~\ref{thm:e:6}
that $\overline{S}=\ohull S=\chs{\uu}{a}$, as claimed.
If
$U=\{{\bf x}\in{\R^n} : {\bf x}\cdot{\bf u}\leq a\}$
or
$U=\{\overline{\xx}\in\extspac{n} : \overline{\xx}\cdot{\bf u} < a\}$,
then $S\subseteq U\subseteq \chs{\uu}{a}$, implying, since $\chs{\uu}{a}$ is
closed, that
$\overline{S}\subseteq\overline{U}\subseteq\chs{\uu}{a}$.
Since $\overline{S}=\chs{\uu}{a}$, as just argued, this proves the final claims.
\qed
\subsection{Outer hull of finite sets}
\label{sec:simplices}
We next study the outer hull, $\ohull V$, of any finite set
$V=\{\overline{\xx}_1,\ldots,\overline{\xx}_m\}$ in $\extspac{n}$, providing
foundational characterizations of such sets that will be used
throughout the rest of this manuscript.
Similar to \eqref{eqn:lb-def},
from Proposition~\ref{pr:ohull-simplify} we can express $\ohull V$ as:
\begin{equation} \label{eq:e:5}
\ohull{V} =
\braces{\overline{{\bf z}}\in\extspac{n} :
\forall {\bf u}\in{\R^n},
\overline{{\bf z}}\cdot{\bf u}\leq \max\{\overline{\xx}_1\cdot{\bf u},\,\ldots,\,\overline{\xx}_m\cdot{\bf u}\}
}.
\end{equation}
As is generally true, $\ohull V$, being the intersection of closed
halfspaces, is itself closed and convex
(Proposition~\ref{pr:e1}(\ref{pr:e1:d})).
In standard convex analysis,
the convex hull of a finite set of points in ${\R^n}$ can be viewed in
two ways:
either as the intersection of all halfspaces containing the points, or
as the set of all convex combinations of the points.
Thus, there is both an ``outside'' and ``inside'' way of
characterizing convexity in this case.
The astral outer hull that we consider here has so far been described
in ``outside'' terms, as the intersection of closed halfspaces.
We next give an alternative ``inside'' description of this same set.
Specifically, we show that
$\ohull V$ can be characterized
in terms of sequences via a formulation
which says that a point is in the outer hull of $V$
if and only if it is the limit of points in ${\R^n}$ that are
themselves convex combinations of points converging to the
points in $V$.
More precisely:
\begin{theorem} \label{thm:e:7}
Let $V=\{\overline{\xx}_1,\ldots,\overline{\xx}_m\}\subseteq\extspac{n}$,
and let $\overline{{\bf z}}\in\extspac{n}$.
Then $\overline{{\bf z}}\in\ohull{V}$ if
and only if there exist sequences $\seq{{\bf x}_{it}}$
in ${\R^n}$ and
$\seq{\lambda_{it}}$ in ${\mathbb{R}}$, for $i=1,\ldots,m$,
such that:
\begin{itemize}
\item
For $i=1,\ldots,m$, ${\bf x}_{it}\rightarrow\overline{\xx}_i$ as
$t\rightarrow\infty$.
\item
For $i=1,\ldots,m$,
and for all $t$, $\lambda_{it}\geq 0$, and
$\sum_{i=1}^m \lambda_{it} = 1$.
\item
The sequence ${\bf z}_t=\sum_{i=1}^m \lambda_{it} {\bf x}_{it}$ converges to
$\overline{{\bf z}}$ as $t\rightarrow\infty$.
\end{itemize}
Additionally, the same statement holds if we further require that,
for each $i=1,\ldots,m$, the sequence $\seq{\lambda_{it}}$ converges to
some limit $\hat{\lambda}_i\in [0,1]$ as $t\rightarrow\infty$,
and that $\sum_{i=1}^m \hat{\lambda}_i = 1$.
\end{theorem}
\proof
Suppose first that there exist sequences of the form given in the
theorem.
Then for each ${\bf u}\in{\R^n}$,
\begin{eqnarray*}
\overline{{\bf z}}\cdot{\bf u}
&=& \lim {\bf z}_t\cdot{\bf u}
\\
&=& \lim \brackets{\sum_{i=1}^m \lambda_{it} {\bf x}_{it} \cdot{\bf u}}
\\
&\leq&
\lim \max\{{\bf x}_{1t}\cdot{\bf u},\, \ldots,\, {\bf x}_{mt}\cdot{\bf u}\}
\\
&=&
\max\{\overline{\xx}_1\cdot{\bf u},\, \ldots,\, \overline{\xx}_m\cdot{\bf u}\},
\end{eqnarray*}
where the last equality (as well as the fact that the last limit exists)
are by continuity of the function which
computes the maximum of $m$ numbers in $\overline{\R}$.
We also have used the assumptions that ${\bf z}_t\rightarrow\overline{{\bf z}}$ and
${\bf x}_{it}\rightarrow\overline{\xx}_i$
(implying
${\bf z}_t\cdot{\bf u}\rightarrow\overline{{\bf z}}\cdot{\bf u}$ and
${\bf x}_{it}\cdot{\bf u}\rightarrow\overline{\xx}_i\cdot{\bf u}$
by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
Since this holds for all ${\bf u}\in{\R^n}$, $\overline{{\bf z}}\in\ohull{V}$ by \eqref{eq:e:5}.
To prove the converse, assume that $\overline{{\bf z}}\in\ohull{V}$.
We will construct the needed sequences one time step $t$ at a time.
For this purpose, we first prove the following lemma showing that within
any neighborhood of $\overline{{\bf z}}$, there must exist a point that is a convex
combination (in ${\R^n}$) of points in any given collection of convex
real subsets that include the points in $V$ in their respective closures.
\begin{lemma} \label{lem:e:1}
Let $\overline{{\bf z}}\in\ohull{V}$, where $V=\{\overline{\xx}_1,\ldots,\overline{\xx}_m\}$.
Let $Z\subseteq\extspac{n}$ be any neighborhood of $\overline{{\bf z}}$.
For $i=1,\ldots,m$, let $X_i\subseteq{\R^n}$ be any convex subset of
${\R^n}$ that includes $\overline{\xx}_i$ in its closure, $\overline{X}_{i}$.
Then there exists ${\bf z}\in{\R^n}\cap Z$ and ${\bf x}_i\in X_i$,
for $i=1,\ldots,m$, such that ${\bf z}$ is a convex combination of
${\bf x}_1,\ldots,{\bf x}_m$, that is,
\[ {\bf z} = \sum_{i=1}^m \lambda_i {\bf x}_i \]
for some $\lambda_1,\ldots,\lambda_m\in [0,1]$ with
$\sum_{i=1}^m \lambda_i = 1$.
\end{lemma}
\proof
Let
\begin{equation} \label{eq:lem:e:1:1}
R= \bigcup_{i=1}^m X_i,
\end{equation}
and let $S$ be the convex hull of $R$, with closure
$\overline{S}$ in $\extspac{n}$.
Then, for $i=1,\ldots,m$,
$X_i\subseteq S$, so, by assumption,
$\overline{\xx}_i\in \overline{X}_{i}\subseteq\overline{S}$.
We further claim that $\overline{{\bf z}}$ is in $\overline{S}$.
This is because, by Theorem~\ref{thm:e:6}, $\overline{S}=\ohull S$, which implies,
for all ${\bf u}\in{\R^n}$, that
\[
\overline{{\bf z}}\cdot{\bf u}
\leq \max\{\overline{\xx}_1\cdot{\bf u},\,\ldots,\,\overline{\xx}_m\cdot{\bf u}\}
\leq \sup_{{\bf y}\in S} {\bf y}\cdot{\bf u}.
\]
The first inequality is from \eqref{eq:e:5},
since $\overline{{\bf z}}\in\ohull{V}$.
The second is because, for $i=1,\ldots,m$,
each $\overline{\xx}_i$ is in $\overline{S}$, which is equal to
$\ohull S$ by Theorem~\ref{thm:e:6}, implying that $\overline{\xx}_i\cdot{\bf u}$
is at most the expression on the right, by
Proposition~\ref{pr:ohull-simplify}.
Therefore, applying that theorem and proposition again,
it follows that $\overline{{\bf z}}$ is in $\overline{S}$.
Let ${\bf z}$ be any point in $Z\cap S$; this set cannot be empty because
$Z$ is a neighborhood of $\overline{{\bf z}}$, which is in the closure of $S$.
Since ${\bf z}$ is in $S$, the convex hull of $R$, the union given in
\eqref{eq:lem:e:1:1}, and since each $X_i$ is convex,
${\bf z}$ can be written as a convex combination of points, one in each
set $X_i$, as stated in the lemma.
(This last step follows directly from
\citet[Theorem~3.3]{ROC}.)
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
To complete the proof of Theorem~\ref{thm:e:7}, let
$Z_1, Z_2,\ldots$
be a nested countable basis for $\overline{{\bf z}}$ (which must exist by
Corollary~\ref{cor:nest-count-basis}).
Further, for each $i=1,\ldots,m$, let
$X_{i1}, X_{i2},\ldots$
be a nested, countable basis for $\overline{\xx}_i$.
Without loss of generality, we can choose each of these sets to be
standard basis elements,
which are therefore convex (by Proposition~\ref{pr:e1}(\ref{pr:e1:d})),
while still forming a nested countable basis.
For each $t=1,2,\ldots$,
let $X'_{it}=X_{it}\cap{\R^n}$, which is also convex by
Proposition~\ref{pr:e1}(\ref{pr:e1:a}, \ref{pr:e1:b}).
Also, $\overline{\xx}_i$ must be in the closure of $X'_{it}$ since
$\overline{\xx}_i\in X_{it} \subseteq \overline{X}_{it} = \overline{X'_{it}}$, by
Proposition~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:b}).
By Lemma~\ref{lem:e:1}, applied to $Z_t$ and the $X'_{it}$'s,
there exists ${\bf z}_t\in{\R^n}\cap Z_t$, ${\bf x}_{it}\in X'_{it}$,
$\lambda_{it}\in [0,1]$,
for $i=1,\ldots,m$, such that
\[ {\bf z}_t = \sum_{i=1}^m \lambda_{it} {\bf x}_{it} \]
with $\sum_{i=1}^m \lambda_{it} = 1$.
Then ${\bf z}_t\rightarrow \overline{{\bf z}}$
(by Corollary~\ref{cor:nest-count-basis}),
and likewise, ${\bf x}_{it}\rightarrow\overline{\xx}_i$ for each $i$.
This proves the entire theorem, except the final statement regarding
convergence of the $\lambda_{it}$'s.
To prove this, note that the sequence of vectors
$\trans{[\lambda_{1t},\ldots,\lambda_{mt}]}$
are in the compact set $[0,1]^m$, and therefore, there must exist a
convergent subsequence.
Discarding those $t$ outside
this subsequence yields a sequence that still satisfies all of the
properties stated in the theorem, and in addition, provides that
$\lambda_{it}\rightarrow\hat{\lambda}_i$, for some $\hat{\lambda}_i\in [0,1]$ with
$\sum_{i=1}^m \hat{\lambda}_i = 1$.
\qed
Since the segment joining two points is the same as the outer hull
of the two points, we immediately obtain the following corollary:
\begin{corollary} \label{cor:e:1}
Let $\overline{\xx},\,\overline{{\bf y}}\in\extspac{n}$,
and let $\overline{{\bf z}}\in\extspac{n}$.
Then $\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$ if
and only if there exist sequences $\seq{{\bf x}_t}$ and
$\seq{{\bf y}_t}$ in ${\R^n}$, and
$\seq{\lambda_t}$ in $[0,1]$
such that
${\bf x}_t\rightarrow\overline{\xx}$ and ${\bf y}_t\rightarrow\overline{{\bf y}}$,
and the sequence
\[ {\bf z}_t= (1-\lambda_t) {\bf x}_t + \lambda_t {\bf y}_t \]
converges to $\overline{{\bf z}}$.
Additionally, the same statement holds if we further require that
the sequence $\seq{\lambda_t}$ converges to a limit in $[0,1]$.
\end{corollary}
By combining Theorem~\ref{thm:e:7} with Carath\'{e}odory's theorem, we
obtain the following:
\begin{theorem} \label{thm:carath}
Suppose $\overline{{\bf z}}\in\ohull{V}$ for some finite set
$V\subseteq\extspac{n}$.
Then $\overline{{\bf z}}\in\ohull{V'}$ for some $V'\subseteq V$ with
$|V'|\leq n+1$.
\end{theorem}
\proof
Let $V=\{\overline{\xx}_1,\ldots,\overline{\xx}_m\}$.
Since $\overline{{\bf z}}\in\ohull{V}$, there exist sequences as given in
Theorem~\ref{thm:e:7}.
For each $t$,
let $I_t$ be the nonzero indices of the $\lambda_{it}$'s, that is,
\[ I_t = \braces{ i\in\{1,\ldots,m\} : \lambda_{it}>0 }. \]
By Carath\'{e}odory's theorem
\citep[Theorem~17.1]{ROC},
we can assume without loss of
generality that the $\lambda_{it}$'s have been chosen
in such a way that $|I_t|\leq n+1$, for all $t$.
Since there are only finitely many subsets of $m$ items, there must
exist some subset $I\subseteq\{1,\ldots,m\}$ for which
$I_t=I$ for infinitely many values of $t$ (implying $|I|\leq n+1$).
On the subsequence consisting of all such values of $t$,
all conditions of the theorem are satisfied with
each ${\bf z}_t$ a convex combination of only points ${\bf x}_{it}$ with
$i\in I$.
Applying Theorem~\ref{thm:e:7} to this subsequence then shows that
$\overline{{\bf z}}\in\ohull{V'}$ where
\[
V' = \braces{ \overline{\xx}_i : i\in I }.
\]
\qed
The following theorem shows that if $V$ is a finite
subset of some convex set,
then the outer hull of $V$ must also be entirely included
in that set.
This is useful, for instance, for characterizing the convex hull, as
will be seen shortly.
\begin{theorem} \label{thm:e:2}
Let $S\subseteq\extspac{n}$ be convex, and let $V\subseteq S$ be a
finite subset.
Then $\ohull{V}\subseteq S$.
\end{theorem}
\proof
Let $V=\{\overline{\xx}_1,\ldots,\overline{\xx}_m\}\subseteq S$.
Proof is by induction on $m=|V|$.
If $m=1$ and $\overline{{\bf z}}\in\ohull{V}$, then
by \eqref{eq:e:5},
for all ${\bf u}\in{\R^n}$,
$\overline{{\bf z}}\cdot{\bf u}\leq \overline{\xx}_1\cdot{\bf u}$, which further implies
$-\overline{{\bf z}}\cdot{\bf u} = \overline{{\bf z}}\cdot(-{\bf u})
\leq \overline{\xx}_1\cdot(-{\bf u}) = -\overline{\xx}_1\cdot{\bf u}$
(by Proposition~\ref{pr:i:2}).
Thus, $\overline{{\bf z}}\cdot{\bf u} = \overline{\xx}_1\cdot{\bf u}$ for all ${\bf u}\in{\R^n}$,
so $\overline{{\bf z}}=\overline{\xx}_1$
(by Proposition~\ref{pr:i:4}), which is in $S$.
Assume now that $m\geq 2$, and that the claim holds for $m-1$.
Let $\overline{{\bf z}}\in\ohull{V}$.
Let ${\bf z}_t$, ${\bf x}_{it}$ and $\lambda_{it}$ be sequences as given in
Theorem~\ref{thm:e:7} with $\lambda_{it}$ converging to
$\hat{\lambda}_i$.
Since no more than one of these limits $\hat{\lambda}_i$ can be equal to
$1$, assume without loss of generality that $\hat{\lambda}_m<1$.
Then for all $t$ sufficiently large, $\lambda_{mt}<1$; let us assume
that is the case for all $t$ (by discarding all others from the sequence).
Define a new sequence that is, for all $t$, a convex combination of
just the ${\bf x}_{it}$'s for $i=1,\ldots,m-1$:
\[
{\bf y}_t = \sum_{i=1}^{m-1} \frac{\lambda_{it}}{1-\lambda_{mt}} {\bf x}_{it}.
\]
Since $\extspac{n}$ is sequentially compact,
the ${\bf y}_t$'s must have a convergent subsequence.
By discarding all time steps $t$ that are not part of this
subsequence, let us assume that the entire sequence of ${\bf y}_t$'s
converge to some point $\overline{{\bf y}}\in\extspac{n}$.
Since all the conditions of Theorem~\ref{thm:e:7} are now satisfied
for $\overline{{\bf y}}$, it must be the case that
$\overline{{\bf y}}\in\ohull{\{\overline{\xx}_1,\ldots,\overline{\xx}_{m-1}\}}$,
and so $\overline{{\bf y}}\in S$ by inductive hypothesis.
Further,
${\bf z}_t = (1-\lambda_{mt}) {\bf y}_t + \lambda_{mt} {\bf x}_{mt}$,
which converges to $\overline{{\bf z}}$.
Thus, the conditions of Corollary~\ref{cor:e:1} are satisfied, and so
$\overline{{\bf z}}\in\lb{\overline{{\bf y}}}{\overline{\xx}_m}$.
Therefore, $\overline{{\bf z}}\in S$ since $\overline{{\bf y}}$ and $\overline{\xx}_m$ are.
\qed
\subsection{Convex hull}
\label{sec:convex-hull}
As is standard for sets in ${\R^n}$, we can define the (astral)
\emph{convex hull} of a set $S\subseteq\extspac{n}$, denoted $\conv{S}$,
to be the intersection of all convex sets in $\extspac{n}$ that contain
$S$.
Of course, this implies that $\conv{S}$ is convex, by
Proposition~\ref{pr:e1}(\ref{pr:e1:b}).
Thus, $\conv{S}$ is the {smallest} convex set that includes $S$.
Basic properties of standard convex hulls in ${\R^n}$ carry
over easily to astral convex hulls, as stated in the next proposition.
\begin{proposition} \label{pr:conhull-prop}
~
\begin{enumerate}
\item \label{pr:conhull-prop:aa}
If $S\subseteq U\subseteq\extspac{n}$
and $U$ is convex, then $\conv{S}\subseteq U$.
\item \label{pr:conhull-prop:a}
If $V\subseteq\extspac{n}$ is finite, then
$\ohull{V}=\conv{V}$.
\item \label{pr:conhull-prop:b}
If $S\subseteq U\subseteq\extspac{n}$ then
$\conv{S}\subseteq\conv{U}$.
\item \label{pr:conhull-prop:c}
If $S\subseteq\extspac{n}$ and $U\subseteq\conv{S}$, then
$\conv(S\cup U) = \conv{S}$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:conhull-prop:aa})
follows immediately from the definition of convex hull.
Part~(\ref{pr:conhull-prop:a}):
$\ohull{V}$ is convex and includes $V$.
Therefore, $\conv{V}\subseteq\ohull{V}$
by part~(\ref{pr:conhull-prop:aa}).
On the other hand, since $\conv{V}$ is convex and includes $V$,
$\ohull{V}\subseteq\conv{V}$ by Theorem~\ref{thm:e:2}.
Part~(\ref{pr:conhull-prop:b}):
Since $S\subseteq U\subseteq\conv{U}$, and since $\conv{U}$ is convex,
$\conv{S}\subseteq\conv{U}$
by part~(\ref{pr:conhull-prop:aa}).
Part~(\ref{pr:conhull-prop:c}):
$\conv{S}\subseteq\conv(S\cup U)$ by part~(\ref{pr:conhull-prop:b}).
And since $S\cup U\subseteq\conv{S}$ and $\conv{S}$ is convex,
$\conv(S\cup U)\subseteq\conv{S}$
by part~(\ref{pr:conhull-prop:aa}).
\qed
As in standard convex analysis,
the convex hull $\conv{V}$
of any finite set of points $V\subseteq\extspac{n}$ is
called the \emph{polytope formed by $V$}.
By Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:a}), this is
the same as its outer hull, $\ohull V$.
In general, however, the convex hull of an arbitrary set
$S\subseteq\extspac{n}$ need not be equal to its outer hull.
Indeed, $\conv{S}$ might not even be closed, but
even its closure need not be equal to $\ohull S$,
as will be seen in
Section~\ref{sec:closure-convex-set}.
Nevertheless,
we can always characterize the convex hull of any set in terms of the
polytopes formed by all its finite subsets:
\begin{theorem} \label{thm:convhull-of-simpices}
Let $S\subseteq\extspac{n}$.
Then its convex hull is equal to the union of all polytopes formed
by finite subsets of $S$, that is,
\begin{equation} \label{eq:e:6}
\conv{S}
= \bigcup_{\scriptontop{V\subseteq S:}{|V|<+\infty}} \ohull{V}
= \bigcup_{\scriptontop{V\subseteq S:}{|V|\leq n+1}} \ohull{V}.
\end{equation}
\end{theorem}
\proof
Let $U$ be the union appearing in the middle expression of
\eqref{eq:e:6}.
We aim to show that $\conv{S}=U$.
Since $\conv{S}$ is convex and includes $S$, by Theorem~\ref{thm:e:2},
it must also include $\ohull{V}$ for every finite $V\subseteq S$.
Thus, $\conv{S}\supseteq U$.
We claim that $U$ is convex.
To see this, suppose that $\overline{\xx},\,\overline{{\bf y}}\in U$, and that
$\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$.
To prove convexity, we need to show that $\overline{{\bf z}}\in U$.
Since $\overline{\xx}\in U$, it must be that
$\overline{\xx}\in\ohull{V}$ for some finite $V\subseteq S$.
Likewise,
$\overline{{\bf y}}\in\ohull{W}$ for some finite $W\subseteq S$.
Then for all ${\bf u}\in{\R^n}$,
\begin{eqnarray*}
\overline{{\bf z}}\cdot{\bf u}
&\leq& \max\{\overline{\xx}\cdot{\bf u},\; \overline{{\bf y}}\cdot{\bf u}\} \\
&\leq& \max\braces{
\max_{\overline{{\bf v}}\in V} \overline{{\bf v}}\cdot{\bf u},\;
\max_{\overline{{\bf w}}\in W} \overline{{\bf w}}\cdot{\bf u}
} \\
&=& \max_{\overline{{\bf s}}\in V\cup W} \overline{{\bf s}}\cdot{\bf u}.
\end{eqnarray*}
The first inequality is by \eqref{eqn:lb-def},
since $\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$, and the
second is by \eqref{eq:e:5},
since $\overline{\xx}\in\ohull{V}$ and
$\overline{{\bf y}}\in\ohull{W}$.
Thus,
$\overline{{\bf z}}\in\ohull{(V\cup W)}\subseteq U$,
again by \eqref{eq:e:5}.
Since $U$ is convex and includes $S$,
$\conv{S}\subseteq U$
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa}).
The second equality of \eqref{eq:e:6} follows from
Theorem~\ref{thm:carath}.
\qed
\section{Constructing and operating on convex sets}
\label{sec:convex-set-ops}
We next study some operations for constructing or manipulating convex sets.
\subsection{Leftward sum of convex sets}
We begin by showing that the leftward sum of a convex set in
$\extspac{n}$ with a convex set in ${\R^n}$ is also convex:
\begin{theorem} \label{thm:sum-convex-sets}
Let $X\subseteq\extspac{n}$ and $Y\subseteq{\R^n}$ both be convex.
Then $X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax Y$ is also convex.
\end{theorem}
\proof
Let $\overline{{\bf z}}$ and $\overline{{\bf z}}'$ be any points in $X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax Y$,
which thus can be written in the form
$\overline{{\bf z}}=\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y}$ and $\overline{{\bf z}}'=\overline{\xx}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y}'$ for some
$\overline{\xx},\overline{\xx}'\in X$ and ${\bf y},{\bf y}'\in Y$.
Let $\overline{{\bf w}}\in\lb{\overline{{\bf z}}}{\overline{{\bf z}}'}$.
To prove convexity, we aim to show that $\overline{{\bf w}}\in X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax Y$.
By Corollary~\ref{cor:e:1}, there exist sequences
$\seq{{\bf z}_t}$ and $\seq{{\bf z}'_t}$ in ${\R^n}$,
and $\seq{\lambda_t}$ in $[0,1]$ such that
${\bf z}_t\rightarrow\overline{{\bf z}}$,
${\bf z}'_t\rightarrow\overline{{\bf z}}'$,
$\lambda_t\rightarrow\hat{\lambda}$, for some $\hat{\lambda}\in [0,1]$,
and also
the sequence
${\bf w}_t=\lambda_t {\bf z}_t + (1-\lambda_t) {\bf z}'_t$
converges to $\overline{{\bf w}}$.
For each $t$, let
${\bf x}_t= {\bf z}_t - {\bf y}$ and
${\bf x}'_t= {\bf z}'_t - {\bf y}'$.
Then ${\bf x}_t\rightarrow\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (-{\bf y}) = \overline{\xx}$,
using Proposition~\ref{pr:aff-cont};
similarly, ${\bf x}'_t\rightarrow\overline{\xx}'$.
Let
$\tilde{\xx}_t=\lambda_t {\bf x}_t + (1-\lambda_t) {\bf x}'_t$.
Then the resulting sequence $\seq{\tilde{\xx}_t}$ has a convergent
subsequence; by discarding other elements, we can assume the entire
sequence converges to some point
$\overline{{\bf s}}$.
Then $\overline{{\bf s}}\in \lb{\overline{\xx}}{\overline{\xx}'}\subseteq X$
by Corollary~\ref{cor:e:1}, and since $X$ is convex.
Also, let
$\tilde{\yy}=\hat{\lambda} {\bf y} + (1-\hat{\lambda}) {\bf y}'$, which is
in $Y$, being convex.
By algebra,
${\bf w}_t=\tilde{\xx}_t + \tilde{\yy} + (\lambda_t - \hat{\lambda})({\bf y}-{\bf y}')$,
so ${\bf w}_t\rightarrow\overline{{\bf s}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\tilde{\yy}$,
since $\lambda_t\rightarrow\hat{\lambda}$ and
using Proposition~\ref{pr:aff-cont}.
Thus,
$\overline{{\bf w}}=\overline{{\bf s}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\tilde{\yy}\in X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax Y$, completing the proof.
\qed
If both $X$ and $Y$ are convex subsets of $\extspac{n}$ (rather than one
being a subset of ${\R^n}$), then Theorem~\ref{thm:sum-convex-sets} is no
longer true in general.
For instance, in ${\mathbb{R}}^2$,
suppose $X=\lb{{\bf 0}}{\limray{{\bf e}_1}}$ and
$Y=\lb{{\bf 0}}{\limray{{\bf e}_2}}$.
Using Corollary~\ref{cor:e:1}
(or Theorem~\ref{thm:lb-with-zero} below),
it can be checked that
$X=\{\lambda {\bf e}_1 : \lambda\geq 0\} \cup \{\limray{{\bf e}_1}\}$,
and similary for $Y$, so
$\limray{{\bf e}_1}$ and $\limray{{\bf e}_2}$ are in $X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax Y$,
but $\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$ is not.
On the other hand, this latter point is on the segment joining
$\limray{{\bf e}_1}$ and $\limray{{\bf e}_2}$, as can be seen from
Corollary~\ref{cor:e:1}
(or Theorem~\ref{thm:conv-lmset-char} below).
Therefore, $X\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax Y$ is not convex.
\subsection{Convexity under affine transformations}
We next consider affine transformations.
We begin by showing that the image of a polytope under an affine map is
equal to the polytope formed by the
images of the points that formed the original polytope.
This is useful, for instance, in showing that the image of a convex
set under an affine map is also convex.
\begin{theorem} \label{thm:e:9}
Let ${\bf A}\in{\mathbb{R}}^{m\times n}$, $\overline{\bb}\in\extspac{m}$, and
let $F:\extspac{n}\rightarrow\extspac{m}$ be the
affine map
$F(\overline{{\bf z}})=\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf A}\overline{{\bf z}}$
for all $\overline{{\bf z}}\in\extspac{n}$.
Then for any finite set
$V=\{\overline{\xx}_1,\ldots,\overline{\xx}_m\}\subseteq\extspac{n}$,
\[
\ohull{F(V)} = F(\ohull{V}).
\]
\end{theorem}
Thus, every point in $\ohull{V}$ is mapped, under $F$, to a point in
$\ohull{F(V)}$, the polytope formed by $F(V)$.
Furthermore, restricted to these sets, $F$ is surjective in
the sense that every point in $\ohull{F(V)}$ is the image of some
point in $\ohull{V}$.
We prove this theorem in a sequence of lemmas, beginning with one of
the inclusions that is needed:
\begin{lemma} \label{thm:e:9:lem:a}
Assume the setup of Theorem~\ref{thm:e:9}.
Then
\[ F(\ohull{V}) \subseteq \ohull{F(V)}. \]
\end{lemma}
\proof
For $\overline{{\bf z}}\in\ohull{V}$, we show that
$F(\overline{{\bf z}})\in\ohull{F(V)}$.
Let ${\bf u}\in{\R^n}$.
Then
\[
F(\overline{{\bf z}})\cdot{\bf u}
=
(\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{{\bf z}})\cdot{\bf u}
=
\overline{\bb}\cdot {\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf A}\overline{{\bf z}})\cdot{\bf u}
=
\overline{\bb}\cdot {\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot(\trans{{\bf A}} {\bf u}),
\]
with the second and third equalities following from
Propositions~\ref{pr:i:6} and~\ref{pr:mat-mult-def},
respectively.
This holds also for each $\overline{\xx}_i$, being in $V\subseteq\ohull V$.
Therefore,
\begin{eqnarray}
F(\overline{{\bf z}})\cdot{\bf u}
&=&
\overline{\bb}\cdot {\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot(\trans{{\bf A}} {\bf u})
\nonumber
\\
&\leq&
\max\{
\overline{\bb}\cdot {\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}_i\cdot(\trans{{\bf A}} {\bf u})
: i=1,\ldots,m
\}
\nonumber
\\
&=&
\max\{
F(\overline{\xx}_i)\cdot{\bf u}
: i=1,\ldots,m
\}.
\label{eq:g:5b}
\end{eqnarray}
The inequality can be seen by considering separately
the cases that $\overline{\bb}\cdot{\bf u}$ is finite or infinite:
If $\overline{\bb}\cdot{\bf u}\in\{-\infty,+\infty\}$, then both sides of this
inequality actually are equal to $\overline{\bb}\cdot{\bf u}$.
Otherwise, if $\overline{\bb}\cdot{\bf u}\in{\mathbb{R}}$, then the
inequality follows from \eqref{eq:e:5}
since $\overline{{\bf z}}\in\ohull{V}$.
Since \eqref{eq:g:5b} holds for all ${\bf u}\in{\R^n}$,
$F(\overline{{\bf z}})\in\ohull{F(V)}$ by \eqref{eq:e:5},
proving the lemma.
\qed
To prove the reverse inclusion, we first give lemmas for two
special cases; we then combine these to give the general result.
We begin with the special case that $\overline{\bb}={\bf b}$ is in ${\R^m}$:
\begin{lemma} \label{thm:e:9:lem:b}.
Assume the setup of Theorem~\ref{thm:e:9}, and suppose that
$F(\overline{{\bf z}})={\bf b}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{{\bf z}}$ for some matrix
${\bf A}\in{\mathbb{R}}^{m\times n}$ and point ${\bf b}\in{\R^m}$.
Then
\[ \ohull{F(V)} \subseteq F(\ohull{V}). \]
\end{lemma}
\proof
Let $\overline{{\bf z}}'\in\ohull{F(V)}$.
We prove that $\overline{{\bf z}}'\in F(\ohull{V})$ using the sequential
characterization and a similar proof technique to that given in
Theorem~\ref{thm:e:7}.
For each $i=1,\ldots,m$,
let $X_{i1}, X_{i2}, \ldots$ be a nested countable
basis in $\extspac{n}$ for $\overline{\xx}_i$,
and
let $Z'_1, Z'_2, \ldots$ be such a
basis in $\extspac{m}$ for $\overline{{\bf z}}'$;
these exist by Corollary~\ref{cor:nest-count-basis}.
Without loss of generality, we assume that each set
$X_{it}$ is a standard basis element, and therefore
convex
(Proposition~\ref{pr:e1}(\ref{pr:e1:d})).
For each $t$, let $X'_{it} = F(X_{it}\cap{\R^n})$.
Note that $X_{it}\cap{\R^n}$ is a convex subset of ${\R^n}$ (by
Proposition~\ref{pr:e1}(\ref{pr:e1:a},\ref{pr:e1:b})),
and so $X'_{it}$, its image under $F$, is
also convex, since here $F$ is restricted to ${\R^n}$ and thus is
acting as a standard affine map, ${\bf x}\mapsto {\bf A}{\bf x}+{\bf b}$.
Further, we claim that $F(\overline{\xx}_i)$ is in $\overline{X'_{it}}$, the closure of $X'_{it}$
in $\extspac{m}$.
This is because there exists a sequence $\seq{{\bf y}_s}$ in ${\R^n}$ converging to
$\overline{\xx}_i$ (Theorem~\ref{thm:i:1}(\ref{thm:i:1d})).
Almost all of the elements of this sequence are in $X_{it}$, being a
neighborhood of $\overline{\xx}_i$.
For all such elements, $F({\bf y}_s)\in X'_{it}$, and these converge
to $F(\overline{\xx}_i)$
since $F$ is continuous (Proposition~\ref{pr:aff-cont}).
Thus, $F(\overline{\xx}_i)\in\overline{X'_{it}}$.
With these preliminaries, we can apply Lemma~\ref{lem:e:1} to
$\overline{{\bf z}}'$, with neighborhood $Z'_t\subseteq\extspac{m}$,
and with the convex set $X'_{it}$
associated with each point $F(\overline{\xx}_i)$,
for $i=1,\ldots,m$.
This shows that there exists
${\bf z}'_t\in Z'_t\cap{\mathbb{R}}^m$,
${\bf x}_{it}\in X_{it}\cap{\R^n}$,
and $\lambda_{it}\in [0,1]$,
such that
\[
{\bf z}'_t = \sum_{i=1}^m \lambda_{it} F({\bf x}_{it})
= \sum_{i=1}^m \lambda_{it} ({\bf A}{\bf x}_{it}+{\bf b}),
\]
and $\sum_{i=1}^m \lambda_{it}=1$.
By Corollary~\ref{cor:nest-count-basis},
${\bf x}_{it}\rightarrow\overline{\xx}_i$
and
${\bf z}'_t\rightarrow\overline{{\bf z}}'$.
Let
\[
{\bf z}_t = \sum_{i=1}^m \lambda_{it} {\bf x}_{it}.
\]
This sequence has a convergent subsequence, being in $\extspac{n}$.
By discarding all other values of $t$, let us assume the entire
sequence converges to some point $\overline{{\bf z}}\in\extspac{n}$.
Then by Theorem~\ref{thm:e:7} applied to this sequence,
$\overline{{\bf z}}\in\ohull{V}$.
By algebra,
\[ {\bf z}'_t = {\bf A} {\bf z}_t + {\bf b} = F({\bf z}_t). \]
Thus, taking limits of both sides, and
by continuity (Proposition~\ref{pr:aff-cont}),
$\overline{{\bf z}}' = F(\overline{{\bf z}})$.
Therefore, $\overline{{\bf z}}'\in F(\ohull{V})$.
\qed
The next lemma handles the special case in which ${\bf A}$ is the identity
matrix and $\overline{\bb}$ is an astron $\limray{{\bf v}}$.
\begin{lemma} \label{thm:e:9:lem:c}
Assume the setup of Theorem~\ref{thm:e:9}, and suppose that
$F(\overline{{\bf z}})=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ for some
${\bf v}\in{\R^n}$.
Then
\[ \ohull{F(V)}) \subseteq F(\ohull{V}). \]
\end{lemma}
\proof
If ${\bf v}={\bf 0}$, the lemma is trivial.
So without loss of generality, we assume henceforth that $\norm{{\bf v}}=1$.
Let $\overline{{\bf z}}\in\ohull{F(V)}$.
We claim that ${\bf v}$ must be $\overline{{\bf z}}$'s dominant direction
(which also means $\overline{{\bf z}}\not\in{\R^n}$).
To see this, let ${\bf u}\in{\R^n}$.
By definition of $F$ and \eqref{eq:e:5},
\[
\overline{{\bf z}}\cdot{\bf u}
\leq
\max\{ \limray{{\bf v}}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}_i\cdot{\bf u}
: i = 1,\ldots,m
\}.
\]
Thus, if ${\bf v}\cdot{\bf u} < 0$, then the right-hand side of this
inequality is $-\infty$, implying that $\overline{{\bf z}}\cdot{\bf u}=-\infty$.
And if ${\bf v}\cdot{\bf u}>0$ then applying this argument to $-{\bf u}$ shows that
$\overline{{\bf z}}\cdot{\bf u}=+\infty$
(since ${\bf v}\cdot(-{\bf u})<0$ implies $\overline{{\bf z}}\cdot(-{\bf u})=-\infty$).
So by Theorem~\ref{thm:dom-dir} and
Proposition~\ref{pr:h:6},
$\overline{{\bf z}}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \zbar^{\bot}$
(where, as usual, $\xbar^{\bot}$ denotes projection of $\overline{\xx}$
perpendicular to ${\bf v}$).
As in Section~\ref{subsec:projections}, let
${\bf P}={\bf I} - {\bf v}\trans{{\bf v}}$,
where ${\bf I}$ is the $n\times n$ identity matrix,
and let $P:\extspac{n}\rightarrow\extspac{n}$ be the linear map
associated with ${\bf P}$ so that $P(\overline{\xx})={\bf P}\overline{\xx}=\xbar^{\bot}$
for all $\overline{\xx}\in\extspac{n}$.
Then we have
\begin{eqnarray}
\zbar^{\bot}
=
P(\overline{{\bf z}})
&\in&
P\paren{\ohull{\{\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}_i : i=1,\ldots,m\}}}
\nonumber
\\
&\subseteq&
\ohull{\{P(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}_i) : i=1,\ldots,m\}}
\label{eq:f:8a}
\\
&=&
\ohull{\{P(\overline{\xx}_i) : i=1,\ldots,m\}}
\label{eq:f:8b}
\\
&\subseteq&
P\paren{\ohull{\{\overline{\xx}_i : i=1,\ldots,m\}}}
\label{eq:f:8c}
\\
&=&
P({\ohull{V}}).
\end{eqnarray}
\eqref{eq:f:8a} is by Lemma~\ref{thm:e:9:lem:a}.
\eqref{eq:f:8b} is because
\[
P(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}_i)
=
(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}_i)^\bot
=
\xbar_i^{\bot}
=
P(\overline{\xx}_i)
\]
using
Proposition~\ref{pr:h:5}(\ref{pr:h:5c},~\ref{pr:h:5e}).
And \eqref{eq:f:8c} is by
Lemma~\ref{thm:e:9:lem:b}.
Therefore, there exists $\overline{{\bf y}}\in\ohull{V}$ such that
$P(\overline{{\bf z}}) = P(\overline{{\bf y}})$, that is, $\zbar^{\bot}=\ybar^{\bot}$.
This means
\[
\overline{{\bf z}} = \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\zbar^{\bot}
= \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\ybar^{\bot}
= \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}
= F(\overline{{\bf y}}),
\]
using Proposition~\ref{pr:h:5}(\ref{pr:h:5g}).
Thus,
$\overline{{\bf z}}\in F(\ohull{V})$, completing the proof.
\qed
\proof[Proof of Theorem~\ref{thm:e:9}]
We return now to the fully general set up of
Theorem~\ref{thm:e:9}.
That $F(\ohull{V}) \subseteq \ohull{F(V)}$ is proved in
Lemma~\ref{thm:e:9:lem:a}.
To complete the theorem, we prove that
$\ohull{F(V)} \subseteq F(\ohull{V})$
by induction on the astral rank $k$ of $\overline{\bb}$.
In the base case,
when $k=0$, $\overline{\bb}\in{\R^m}$ and the claimed
inclusion is proved by Lemma~\ref{thm:e:9:lem:b}.
For the inductive step when $k>0$, let ${\bf v}\in{\R^m}$ be $\overline{\bb}$'s dominant
direction so that $\overline{\bb}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\bbar^{\bot}$.
Let us define the affine maps
$ G(\overline{{\bf y}}) = \limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}} $
for $\overline{{\bf y}}\in\extspac{m}$,
and
$ H(\overline{{\bf z}}) = \bbar^{\bot}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{{\bf z}} $
for $\overline{{\bf z}}\in\extspac{n}$.
Then
\[
F(\overline{{\bf z}})
=
\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{{\bf z}}
=
\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(\bbar^{\bot}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{{\bf z}})
=
G(H(\overline{{\bf z}}))
\]
for all $\overline{{\bf z}}\in\extspac{n}$.
So
\begin{eqnarray*}
\ohull{F(V)}
&=&
\ohull{G(H(V))}
\nonumber
\\
&\subseteq&
G(\ohull{H(V)})
\label{eq:f:9a}
\\
&\subseteq&
G(H(\ohull{V}))
\label{eq:f:9b}
\\
&=&
F(\ohull{V}).
\nonumber
\end{eqnarray*}
The first inclusion is by Lemma~\ref{thm:e:9:lem:c}.
The second inclusion
is by our inductive hypothesis, since $\bbar^{\bot}$ has
rank $k-1$
(by Proposition~\ref{pr:h:6}).
This completes the induction and the proof.
\qed
Theorem~\ref{thm:e:9} does not hold in general if an astral
point $\overline{\bb}$ is added on the {right} rather than the left, that is,
for mappings of the form $\overline{\xx}\mapsto {\bf A}\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\bb}$.
For instance, in $\extspac{2}$, suppose
$F(\overline{\xx})=\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_2}$,
and
$V=\{{\bf 0}, \limray{{\bf e}_1}\}$,
where ${\bf e}_1$ and ${\bf e}_2$ are standard basis vectors.
Then using Theorem~\ref{thm:lb-with-zero} below,
it can be shown that
\[
F(\ohull{V}) =
\{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf e}_2} \}
\cup
\{ \limray{{\bf e}_2} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda {\bf e}_1 : \lambda \geq 0 \},
\]
which does not include $\overline{{\bf z}}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$.
On the other hand, $\overline{{\bf z}}$ \emph{is} included in
$\ohull{F(V)}$
(as can be shown, for instance, using Corollary~\ref{cor:e:1}).
As an immediate corollary of Theorem~\ref{thm:e:9}, the image of a
convex set under an affine map is also convex:
\begin{corollary} \label{cor:thm:e:9}
Let $F:\extspac{n}\rightarrow\extspac{m}$ be an affine map
(as in Theorem~\ref{thm:e:9}), and let $S$ be a convex subset of
$\extspac{n}$.
Then $F(S)$ is also convex.
\end{corollary}
\proof
Let $F(\overline{\xx})$ and $F(\overline{{\bf y}})$ be any two points of $F(S)$, where
$\overline{\xx},\overline{{\bf y}}\in S$.
Since the segment joining two points is just their outer hull,
we can apply Theorem~\ref{thm:e:9}, which yields
\[
\lb{F(\overline{\xx})}{F(\overline{{\bf y}})}
=
F(\lb{\overline{\xx}}{\overline{{\bf y}}})
\subseteq
F(S),
\]
with the inclusion following from the convexity of $S$.
Thus, $F(S)$ is convex.
\qed
The next corollary shows that Theorem~\ref{thm:e:9} holds for the
convex hull of arbitrary sets.
(Theorem~\ref{thm:e:9} is then a special case of the corollary in
which $S$ is finite.)
\begin{corollary} \label{cor:thm:e:9b}
Let $F:\extspac{n}\rightarrow\extspac{m}$ be an affine map
(as in Theorem~\ref{thm:e:9}), and let $S\subseteq\extspac{n}$.
Then
\[ \conv{F(S)} = F(\conv{S}). \]
\end{corollary}
\proof
Since $F(S)$ is included in $F(\conv{S})$, and since the latter set is
convex by Corollary~\ref{cor:thm:e:9},
we must have $\conv{F(S)}\subseteq F(\conv{S})$
(by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})).
For the reverse inclusion, suppose
$F(\overline{\xx})$ is any point in $F(\conv{S})$, where $\overline{\xx}\in\conv{S}$.
Then $\overline{\xx}\in\ohull{V}$ for some finite $V\subseteq S$,
by Theorem~\ref{thm:convhull-of-simpices}.
Thus,
\begin{eqnarray*}
F(\overline{\xx})
&\in&
F(\ohull{V})
\nonumber
\\
&=&
\ohull{F(V)}
\label{eq:ct-e9b-a}
\\
&\subseteq&
\conv{F(S)}.
\label{eq:ct-e9b-b}
\end{eqnarray*}
The equality is by Theorem~\ref{thm:e:9}.
The last line again uses
Theorem~\ref{thm:convhull-of-simpices}.
\qed
\subsection{The segment joining a point and the origin}
\label{sec:seg-join-point-origin}
As a useful and concrete illustration of earlier concepts,
we next explicitly work out $\lb{{\bf 0}}{\overline{\xx}}$,
the segment joining the origin and an arbitrary point
$\overline{\xx}\in\extspac{n}$.
We first prove a lemma that will be the basis for an inductive
construction of this set.
The lemma states that if $\overline{\xx}$ has dominant direction ${\bf v}$,
and thus, if $\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}'$ for some $\overline{\xx}'$, then
the points in $\lb{{\bf 0}}{\overline{\xx}}$ are of two types:
the finite points in ${\R^n}$, which are exactly all of the nonnegative
multiples of ${\bf v}$;
and the infinite points, which are exactly those points
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'$ where $\overline{{\bf z}}'$ is in the segment
joining ${\bf 0}$ and $\overline{\xx}'$.
\begin{lemma} \label{lem:lb-with-zero}
Suppose $\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}'$, for some ${\bf v}\in{\R^n}$ and
$\overline{\xx}'\in\extspac{n}$.
Then
\[
\lb{{\bf 0}}{\overline{\xx}}
=
\;
\{ \lambda {\bf v} : \lambda \geq 0 \}
\;\cup\;
\brackets{\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lb{{\bf 0}}{\overline{\xx}'}}.
\]
\end{lemma}
\proof
The lemma is trivial if ${\bf v}={\bf 0}$, so we assume henceforth, without
loss of generality, that $\norm{{\bf v}}=1$.
Let
\begin{eqnarray*}
D &=& \{ \lambda {\bf v} : \lambda \geq 0 \} \\
E &=& \limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lb{{\bf 0}}{\overline{\xx}'}.
\end{eqnarray*}
\setcounter{claimp}{0}
\medskip
\begin{claimp}
$D\cup E \subseteq \lb{{\bf 0}}{\overline{\xx}}$.
\end{claimp}
\proof
Let $\overline{{\bf z}}\in D\cup E$, and
let ${\bf u}\in{\R^n}$.
To show that $\overline{{\bf z}}\in\lb{{\bf 0}}{\overline{\xx}}$,
we aim to show that
\eqref{eqn:lb-def} is satisfied, that is, that
\begin{equation} \label{eq:g:6}
\overline{{\bf z}}\cdot{\bf u} \leq \max\{ 0,\, (\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}')\cdot{\bf u} \}.
\end{equation}
If ${\bf v}\cdot{\bf u}>0$, then the right-hand side of this inequality is
equal to $+\infty$, and thus is always satisfied.
We therefore assume for the
rest of the proof of this claim that ${\bf v}\cdot{\bf u}\leq 0$.
Also, the right-hand side of \eqref{eq:g:6}
is nonnegative, which means it
suffices, for proving the inequality, to show
$\overline{{\bf z}}\cdot{\bf u}\leq 0$.
If $\overline{{\bf z}}\in D$, so that $\overline{{\bf z}}=\lambda{\bf v}$ for some $\lambda\geq 0$,
then
$\overline{{\bf z}}\cdot{\bf u}=\lambda{\bf v}\cdot{\bf u}\leq 0$,
(since ${\bf v}\cdot{\bf u}\leq 0$),
completing the proof that $D\subseteq\lb{{\bf 0}}{\overline{\xx}}$.
Otherwise, suppose that $\overline{{\bf z}}\in E$, so
$\overline{{\bf z}}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}'$ for some
$\overline{{\bf z}}'\in\lb{{\bf 0}}{\overline{\xx}'}$.
If ${\bf v}\cdot{\bf u}<0$, then $\overline{{\bf z}}\cdot{\bf u}=-\infty$, so
\eqref{eq:g:6} is satisfied.
In the remaining case that ${\bf v}\cdot{\bf u}=0$, we have
\begin{eqnarray*}
\overline{{\bf z}}\cdot{\bf u}
&=&
(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}')\cdot{\bf u}
\\
&=&
\overline{{\bf z}}'\cdot{\bf u}
\\
&\leq&
\max\{ 0, \overline{\xx}'\cdot{\bf u} \}
\\
&=&
\max\{ 0, (\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}')\cdot{\bf u} \}.
\end{eqnarray*}
where the inequality holds because
$\overline{{\bf z}}'\in\lb{{\bf 0}}{\overline{\xx}'}$.
Thus, \eqref{eq:g:6} is satisfied in all cases, and so
$E\subseteq\lb{{\bf 0}}{\overline{\xx}}$.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
We next turn to proving the reverse inclusion.
For the rest of the proof,
let $\overline{{\bf z}}$ be any point in $\lb{{\bf 0}}{\overline{\xx}}$, which we aim to show
is in $D\cup E$.
\medskip
\begin{claimp}
If $\overline{{\bf z}}\in{\R^n}$ then $\overline{{\bf z}}\in D$.
Otherwise, if $\overline{{\bf z}}\not\in{\R^n}$, then $\overline{{\bf z}}$'s dominant direction
is ${\bf v}$.
\end{claimp}
\proof
By Corollary~\ref{cor:e:1}, there must exist sequences $\seq{{\bf x}_t}$
and $\seq{{\bf y}_t}$ in ${\R^n}$ with ${\bf x}_t\rightarrow\overline{\xx}$ and
${\bf y}_t\rightarrow{\bf 0}$, and also a sequence $\seq{\gamma_t}$ in
$[0,1]$ such that the sequece
\begin{equation} \label{eq:g:7a}
{\bf z}_t = \gamma_t {\bf x}_t + (1-\gamma_t) {\bf y}_t
\end{equation}
converges to $\overline{{\bf z}}$.
For each $t$,
let ${\bf d}_t = {\bf x}_t/\norm{{\bf x}_t}$,
let $\boldsymbol{\varepsilon}_t = (1-\gamma_t){\bf y}_t$,
and
let $s_t = \gamma_t \norm{{\bf x}_t}$.
(Since $\overline{\xx}\not\in{\R^n}$, we can discard all $t$ for which
${\bf x}_t={\bf 0}$, of which there can be at most finitely many.)
Then we can rewrite ${\bf z}_t$ as
\begin{equation} \label{eq:g:7c}
{\bf z}_t = s_t {\bf d}_t + \boldsymbol{\varepsilon}_t
\end{equation}
where $\norm{{\bf d}_t}=1$, $s_t\geq 0$,
$\boldsymbol{\varepsilon}_t\rightarrow {\bf 0}$,
and
${\bf d}_t\rightarrow{\bf v}$ (by Theorem~\ref{thm:dom-dir}).
Taking inner product with ${\bf u}\in{\R^n}$ on both sides of \eqref{eq:g:7a},
and then taking limits, yields
\begin{equation} \label{eq:g:7b}
\overline{{\bf z}}\cdot{\bf u}
=
\lim {\bf z}_t\cdot{\bf u}
=
\lim [ s_t {\bf d}_t\cdot{\bf u} + \boldsymbol{\varepsilon}_t\cdot{\bf u} ]
=
\lim s_t {\bf d}_t\cdot{\bf u},
\end{equation}
using Theorem~\ref{thm:i:1}(\ref{thm:i:1c}) in the first equality,
and continuity in the last equality.
When ${\bf u}={\bf v}$, this shows that $s_t\rightarrow\overline{{\bf z}}\cdot{\bf v}$ since
${\bf d}_t\cdot{\bf v}\rightarrow 1$.
This further implies
$\overline{{\bf z}}\cdot{\bf v}\geq 0$ since $s_t\geq 0$ for all $t$.
So if $\overline{{\bf z}}\cdot{\bf v}=\lim s_t$ is equal to some real number
$\lambda\geq 0$, then from \eqref{eq:g:7c}, $\overline{{\bf z}}=\lambda{\bf v}$, which
is in $D$.
Otherwise, if $s_t\rightarrow+\infty$, then \eqref{eq:g:7b} implies
that if ${\bf v}\cdot{\bf u}\neq 0$ then $\overline{{\bf z}}\cdot{\bf u}=\limray{{\bf v}}\cdot{\bf u}$,
since ${\bf d}_t\rightarrow{\bf v}$.
Thus, ${\bf v}$ is $\overline{{\bf z}}$'s dominant direction
(by Theorem~\ref{thm:dom-dir}).
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
The next claim will complete the proof:
\medskip
\begin{claimp}
If $\overline{{\bf z}}\not\in{\R^n}$, then $\overline{{\bf z}}\in E$.
\end{claimp}
\proof
As in the proof of Lemma~\ref{thm:e:9:lem:c},
let ${\bf P}={\bf I} - {\bf v}\trans{{\bf v}}$
and let $P:\extspac{n}\rightarrow\extspac{n}$ be the associated linear map
so that $P(\overline{{\bf y}})={\bf P}\overline{{\bf y}}=\ybar^{\bot}$
for all $\overline{{\bf y}}\in\extspac{n}$.
In particular, $\xbar^{\bot}=P(\overline{\xx})=P(\overline{\xx}')$
since $\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}'$
(using Proposition~\ref{pr:h:4} and ${\bf P}{\bf v}={\bf 0}$).
So by Theorem~\ref{thm:e:9}, applied twice,
\begin{eqnarray*}
P(\overline{{\bf z}})
\in
P(\lb{{\bf 0}}{\overline{\xx}})
&=&
\lb{P({\bf 0})}{P(\overline{\xx})}
\\
&=&
\lb{P({\bf 0})}{P(\overline{\xx}')}
\\
&=&
P(\lb{{\bf 0}}{\overline{\xx}'}).
\end{eqnarray*}
Therefore, there exists $\overline{{\bf y}}\in\lb{{\bf 0}}{\overline{\xx}'}$ such that
$P(\overline{{\bf z}})=P(\overline{{\bf y}})$, that is, $\zbar^{\bot}=\ybar^{\bot}$.
Thus,
\[
\overline{{\bf z}}
=
\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \zbar^{\bot}
=
\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \ybar^{\bot}
=
\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}}.
\]
The first equality is because ${\bf v}$ is $\overline{{\bf z}}$'s dominant direction, as
shown in the preceding claim, and by Proposition~\ref{pr:h:6}.
The last equality, which follows from
Proposition~\ref{pr:h:5}(\ref{pr:h:5g}),
shows that $\overline{{\bf z}}\in E$, as claimed.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Combining the three claims completes the proof.
\qed
Using Lemma~\ref{lem:lb-with-zero}, we can now state explicitly
exactly which points comprise the segment joining the origin and
any other point in $\extspac{n}$:
\begin{theorem} \label{thm:lb-with-zero}
Suppose $\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
where ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in{\R^n}$.
Then
\begin{eqnarray*}
\lb{{\bf 0}}{\overline{\xx}}
&=&
\braces{ \limrays{{\bf v}_1,\ldots,{\bf v}_{j-1}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf v}_j
: j\in\{1,\ldots,k\}, \lambda\geq 0 }
\\
&& \cup
\braces{ \limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf q}
: \lambda\in [0,1] }.
\end{eqnarray*}
\end{theorem}
Note that the representation of $\overline{\xx}$ given in the theorem need not
be canonical.
\proof
Proof is by induction on $k$.
In the base case that $k=0$,
$\overline{\xx}={\bf q}\in{\R^n}$ so
$\lb{{\bf 0}}{{\bf q}} = \{ \lambda{\bf q} : \lambda\in [0,1] \}$,
as claimed,
by standard convexity
(see Proposition~\ref{pr:e1}(\ref{pr:e1:a})).
For the inductive step, suppose $k>0$ and that the claim holds for $k-1$.
We can write $\overline{\xx}=\limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}$ where
$\overline{{\bf z}}=\limrays{{\bf v}_2,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$.
By inductive hypothesis, $\lb{{\bf 0}}{\overline{{\bf z}}}$ has the form given in
the theorem (modified according to $\overline{{\bf z}}$'s representation).
By Lemma~\ref{lem:lb-with-zero},
\[
\lb{{\bf 0}}{\overline{\xx}}
=
\{ \lambda {\bf v}_1 : \lambda \geq 0 \}
\cup
\paren{\limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lb{{\bf 0}}{\overline{{\bf z}}}}
\]
which, working through details, yields exactly the form given in the
theorem.
\qed
As an immediate corollary, we can also compute the segment
joining an arbitrary point in ${\R^n}$ and any other point in
$\extspac{n}$:
\begin{corollary} \label{cor:lb-with-finite}
Let ${\bf y}\in{\R^n}$, and
suppose $\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
where ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in{\R^n}$.
Then
\begin{eqnarray*}
\lb{{\bf y}}{\overline{\xx}}
&=&
\braces{ \limrays{{\bf v}_1,\ldots,{\bf v}_{j-1}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(\lambda{\bf v}_j+{\bf y})
: j\in\{1,\ldots,k\}, \lambda\geq 0 }
\\
&& \cup
\braces{ \limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
((1-\lambda){\bf y}+\lambda{\bf q})
: \lambda\in [0,1] }.
\end{eqnarray*}
\end{corollary}
\proof
Theorem~\ref{thm:e:9}, applied to the affine map
$\overline{{\bf z}}\mapsto{\bf y}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$, yields
\begin{equation} \label{eqn:cor:lb-with-finite:1}
\lb{{\bf y}}{\overline{\xx}}
=
{\bf y} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lb{{\bf 0}}{\overline{\xx} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (-{\bf y})}.
\end{equation}
The result now follows by evaluating the right-hand side using
Theorem~\ref{thm:lb-with-zero}.
\qed
Here is a simple but useful corollary of
Theorem~\ref{thm:lb-with-zero}:
\begin{corollary} \label{cor:d-in-lb-0-dplusx}
Let $\overline{\ee}\in{{\cal E}_n}$ and let $\overline{\xx}\in\extspac{n}$.
Then $\overline{\ee}\in\lb{{\bf 0}}{\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}}$.
\end{corollary}
\proof
By Proposition~\ref{pr:icon-equiv},
$\overline{\ee}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}$ for some
${\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$, while
$\overline{\xx}=\limrays{{\bf w}_1,\ldots,{\bf w}_\ell}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some
${\bf w}_1,\ldots,{\bf w}_\ell,{\bf q}\in{\R^n}$.
Since
\[
\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k;
{\bf w}_1,\ldots,{\bf w}_\ell}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q},
\]
the corollary follows as an immediate application of
Theorem~\ref{thm:lb-with-zero}.
\qed
\subsection{Interior of a convex set}
We prove next that the interior of a convex set in $\extspac{n}$ is also
convex.
For the proof,
recall that a convex set in ${\R^n}$ is \emph{polyhedral} if it is the
intersection of a finite collection of closed halfspaces.
\begin{theorem} \label{thm:interior-convex}
Let $S$ be a convex subset of $\extspac{n}$.
Then its interior, $\intr S$, is also convex.
\end{theorem}
\proof
We assume that $\intr S$ is not empty, since otherwise the claim holds
trivially.
Let $\overline{\xx},\,\overline{{\bf y}}\in \intr S$.
We aim to prove convexity by showing that the segment joining them,
$\lb{\overline{\xx}}{\overline{{\bf y}}}$, is entirely contained in $\intr S$.
\setcounter{claimp}{0}
\medskip
\begin{claimp} \label{cl:thm:interior-convex:1}
There exists a standard basis element $X\subseteq\extspac{n}$
that includes $\overline{\xx}$, and
whose closure, $\overline{X}$, is included in $S$.
\end{claimp}
\proof
Let $C=\extspac{n} \setminus (\intr S)$, which is closed, being the
complement of $\intr S$.
Furthermore, $\overline{\xx}\not\in C$.
Therefore, by
Proposition~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:c}),
there exist disjoint open sets $X$ and $V$ such that
$\overline{\xx}\in X$ and $C\subseteq V$.
Without loss of generality, we can assume $X$ is a standard basis
element (since otherwise we can replace it by a basis element
containing $\overline{\xx}$ and included in $X$).
Since $X$ and $V$ are disjoint,
$X$ is included in the closed set $\extspac{n} \setminus V$,
and therefore,
\[
\overline{X}
\subseteq \extspac{n} \setminus V
\subseteq \extspac{n}\setminus C = \intr S \subseteq S.
\]
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Let $X$ be as in Claim~\ref{cl:thm:interior-convex:1}, and let $Y$ be
a similar standard basis element for $\overline{{\bf y}}$
(so $\overline{{\bf y}}\in Y$ and $\overline{Y}\subseteq S$).
Next, let
$X'=\cl(X\cap{\R^n})$
and
$Y'=\cl(Y\cap{\R^n})$.
Since $X$ is a standard basis element (of the form given in
\eqref{eq:h:3a}),
$X\cap{\R^n}$ is an intersection of open halfspaces in ${\R^n}$, which
means that $X'$, its closure in ${\R^n}$, is an intersection of closed
halfspaces in ${\R^n}$; in other words, $X'$ is a polyhedral convex set,
as is $Y'$ by the same argument.
Furthermore, none of these sets can be empty since ${\R^n}$ is dense in
$\extspac{n}$.
Let
$R=\conv(X' \cup Y')$
be the convex hull of their union.
Since $X'$ and $Y'$ are polyhedral, $\cl R$,
the closure of $R$ in ${\R^n}$, is also polyhedral
\citep[by][Theorem 19.6]{ROC}.
Thus,
\begin{equation} \label{eqn:thm:interior-convex:1}
\cl R = \braces{ {\bf z}\in{\R^n} :
{\bf z}\cdot{\bf u}_i \leq b_i \mbox{~for~} i=1,\ldots,k }
\end{equation}
for some ${\bf u}_1,\ldots,{\bf u}_k\in{\R^n}$ and $b_1,\ldots,b_k\in{\mathbb{R}}$, and
some $k\geq 0$.
Furthermore, without loss of generality, we can assume
$\norm{{\bf u}_i}=1$ for $i=1,\ldots,k$
(since an inequality with ${\bf u}_i={\bf 0}$ can simply be discarded, and for all
others, we can divide both sides of the inequality by $\norm{{\bf u}_i}$).
The closure of $R$ in $\extspac{n}$ is exactly the convex hull of
$\overline{X}\cup\overline{Y}$:
\medskip
\begin{claimp} \label{cl:thm:interior-convex:2}
$\overline{R} = \conv(\overline{X}\cup\overline{Y})$.
\end{claimp}
\proof
By construction,
$X\cap{\R^n} \subseteq X' \subseteq R$.
Therefore, by
Proposition~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:b}),
$\overline{X} = \overline{X\cap{\R^n}}\subseteq \overline{R}$.
Likewise, $\overline{Y}\subseteq\overline{R}$.
Therefore, since $\overline{R}$ is convex
(by Theorem~\ref{thm:e:6}),
$\conv(\overline{X}\cup\overline{Y})\subseteq\overline{R}$
(by Propostion~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})).
For the reverse inclusion, suppose $\overline{{\bf z}}\in\overline{R}$, implying that
there exists a sequence $\seq{{\bf z}_t}$ in $R$ that converges to
$\overline{{\bf z}}$.
Then by $R$'s definition, and since $X'$ and $Y'$ are convex,
for each $t=1,2,\ldots$, we can write
\[
{\bf z}_t = \lambda_t {\bf x}_t + (1-\lambda_t) {\bf y}_t
\]
for some ${\bf x}_t\in X'$, ${\bf y}_t\in Y'$ and $\lambda_t\in [0,1]$
\citep[Theorem~3.3]{ROC}.
By sequential compactness,
the sequence $\seq{{\bf x}_t}$ must have a convergent subsequence; by
discarding all other elements, we can assume that the entire sequence
converges to some point $\overline{\xx}'$, which thereby must be in $\overline{X'}$.
Furthermore, $\overline{X'}=\overline{X\cap{\R^n}}=\overline{X}$
by Proposition~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:aa},~\ref{pr:closed-set-facts:b}).
Repeating this argument, we can take the sequence $\seq{{\bf y}_t}$ to
converge to some point $\overline{{\bf y}}'\in\overline{Y}$.
Therefore, applying Corollary~\ref{cor:e:1},
$\overline{{\bf z}}\in\lb{\overline{\xx}'}{\overline{{\bf y}}'}\subseteq\conv(\overline{X}\cup\overline{Y})$.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Let $Q$ be the set
\[
Q = \braces{ \overline{{\bf z}}\in\extspac{n} :
\overline{{\bf z}}\cdot{\bf u}_i < b_i \mbox{~for~} i=1,\ldots,k },
\]
which is the intersection of open halfspaces in $\extspac{n}$
corresponding to the closed halfspaces (in ${\R^n}$) whose intersection
defines $\cl R$ in \eqref{eqn:thm:interior-convex:1}.
This set is clearly open (and actually is a standard basis element).
\medskip
\begin{claimp} \label{cl:thm:interior-convex:3}
$Q\subseteq \overline{R}$.
\end{claimp}
\proof
Let $\overline{{\bf z}}\in Q$.
Then there exists a sequence $\seq{{\bf z}_t}$ in ${\R^n}$ that converges to
$\overline{{\bf z}}$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1d})).
Since $Q$ is open and therefore a neighborhood of $\overline{{\bf z}}$,
all but finitely many of the elements of this sequence must be in $Q$;
by discarding all others, assume that all elements of the sequence are
in $Q$.
Then for all $t$,
${\bf z}_t\in Q\cap{\R^n}\subseteq \cl R$,
implying $\overline{{\bf z}}\in \overline{\cl R} = \overline{R}$
(by Propostion~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:aa})).
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
\medskip
\begin{claimp} \label{cl:thm:interior-convex:4}
$\overline{\xx}\in Q$.
\end{claimp}
\proof
To prove the claim, we show $\overline{\xx}\cdot{\bf u}_i < b_i$ for
each $i\in\{1,\ldots,k\}$.
Let $H_i$ be the closed halfspace
\[ H_i=\{\overline{{\bf z}}\in\extspac{n} : \overline{{\bf z}}\cdot{\bf u}_i \leq b_i\}. \]
Then $R\subseteq \cl R \subseteq H_i$,
so $\overline{R}\subseteq H_i$ since $H_i$ is closed.
Therefore, using Claim~\ref{cl:thm:interior-convex:2},
$X \subseteq \overline{X} \subseteq \overline{R} \subseteq H_i$.
Since $\overline{\xx}\in X$, this means
$\overline{\xx}\cdot{\bf u}_i \leq b_i$.
It remains then only to show that this inequality must be strict.
Suppose, to the contrary, that
$\overline{\xx}\cdot{\bf u}_i = b_i$.
Let
\[
{H^c_i}
= \extspac{n}\setminus H_i
= \{\overline{{\bf z}}\in\extspac{n} : \overline{{\bf z}}\cdot{\bf u}_i > b_i \}
\]
be the open halfspace that is the complement of $H_i$.
Then our assumption means that $\overline{\xx}$ is in ${\overline{H^c_i}}$, the closure
of ${H^c_i}$ (by Proposition~\ref{pr:halfspace-closure}).
In particular, since $X$ is a neighborhood of $\overline{\xx}$, this implies
that $X$ intersects ${H^c_i}$.
But this is a contradiction since $X\subseteq H_i$, as argued above.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
By Claim~\ref{cl:thm:interior-convex:1}, $\overline{X}\subseteq S$, and
similarly $\overline{Y}\subseteq S$.
These imply,
by Claims~\ref{cl:thm:interior-convex:2}
and~\ref{cl:thm:interior-convex:3}
(as well as Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})),
that
$Q\subseteq \overline{R}\subseteq S$.
Since $Q$ is open, this further implies that
$Q$ is included in the interior of $S$.
By Claim~\ref{cl:thm:interior-convex:4},
$\overline{\xx}\in Q$, and by the same argument, $\overline{{\bf y}}\in Q$.
Since $Q$ is convex (by Proposition~\ref{pr:e1}(\ref{pr:e1:d})),
it follows that
$\lb{\overline{\xx}}{\overline{{\bf y}}}\subseteq Q \subseteq \intr S$,
completing the proof.
\qed
As a consequence, the convex hull of an open set is also open:
\begin{corollary} \label{cor:convhull-open}
Let $U\subseteq\extspac{n}$ be open.
Then its convex hull, $\conv{U}$, is also open.
\end{corollary}
\proof
Let $S=\conv{U}$.
Then $U\subseteq S$, implying, since $U$ is open, that
$U\subseteq \intr S$.
By Theorem~\ref{thm:interior-convex}, $\intr S$ is convex.
Therefore, $S=\conv{U}\subseteq \intr S$
(by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})).
Thus, $S= \intr S$ (since $\intr S\subseteq S$ always),
so $S$ is open.
\qed
\subsection{Closure of a convex set}
\label{sec:closure-convex-set}
In standard convex analysis, the closure of any convex set is also
convex.
In astral space, it was seen in Theorem~\ref{thm:e:6} that the closure
in $\extspac{n}$ of any convex set in ${\R^n}$ is also convex, and
furthermore, is equal to the outer hull of the set.
As we show in the next theorem, this does not turn out to be true
generally for arbitrary convex sets in $\extspac{n}$.
In other words, for $n\geq 2$, there exist sets in $\extspac{n}$
that are convex, but whose closures are not convex.
This also means that the closure of such a set
cannot be equal to its
outer hull, since the outer hull of any set is always convex.
\begin{theorem} \label{thm:closure-not-always-convex}
For $n\geq 2$, there exists a set $S\subseteq\extspac{n}$ that is
convex, but whose closure, $\overline{S}$, is not convex.
Consequently, its closure also is not equal to its outer hull;
that is, $\overline{S}\neq\ohull S$.
\end{theorem}
\proof
Let $n\geq 2$, and let ${\bf e}_1$ and ${\bf e}_2$ be the first two standard
basis vectors in ${\R^n}$.
For $\alpha\in{\mathbb{R}}$, let
\[R_{\alpha}=\lb{\alpha {\bf e}_1}{\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_2}},\]
the segment joining
$\alpha {\bf e}_1$ and $\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_2}$,
and let $S$ be their union over all $\alpha\in{\mathbb{R}}$:
\[ S = \bigcup_{\alpha\in{\mathbb{R}}} R_{\alpha}. \]
We will show that $S$ is convex, but that its closure, $\overline{S}$, is not
(implying $\overline{S}\neq\ohull S$ by Proposition~\ref{pr:e1}(\ref{pr:e1:d})).
First, for $\alpha\in{\mathbb{R}}$, we can compute $R_{\alpha}$ explicitly using
Corollary~\ref{cor:lb-with-finite} to be:
\begin{eqnarray}
R_{\alpha}
&=&
\{ \lambda {\bf e}_1 + \alpha {\bf e}_1 : \lambda\geq 0 \}
\nonumber
\\
&&
\cup \{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\lambda{\bf e}_2 + \alpha{\bf e}_1)
: \lambda\geq 0 \}
\nonumber
\\
&&
\cup \{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf e}_2} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda\alpha {\bf e}_1
: \lambda\in [0,1] \}
\nonumber
\\
&=&
\{ \lambda {\bf e}_1 : \lambda\geq \alpha \}
\cup \{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf e}_2
: \lambda\geq 0 \}
\cup \{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf e}_2} \},
\label{eq:thm:closure-not-always-convex:1}
\end{eqnarray}
with the simplifications in the second equality following from
Theorem~\ref{thm:i:2}.
Thus,
$S$ can now be seen to be
\begin{equation} \label{eq:thm:closure-not-always-convex:2}
S
=
\{ \lambda {\bf e}_1 : \lambda\in{\mathbb{R}} \}
\cup \{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf e}_2
: \lambda\geq 0 \}
\cup \{ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf e}_2} \}.
\end{equation}
To see that $S$ is convex, let $\overline{\xx}$ and $\overline{{\bf y}}$ be any two points
in $S$.
Then, by $S$'s definition, we must have $\overline{\xx}\inR_{\alpha}$ and
$\overline{{\bf y}}\inR_{\beta}$, for some $\alpha,\beta\in{\mathbb{R}}$.
Without loss of generality, assume $\alpha\leq\beta$.
Then
\eqref{eq:thm:closure-not-always-convex:1} implies that
$R_{\beta}\subseteqR_{\alpha}$, and thus that both $\overline{\xx}$ and $\overline{{\bf y}}$
are in $R_{\alpha}$.
The set $R_{\alpha}$ is convex since it is a segment
(by Proposition~\ref{pr:e1}(\ref{pr:e1:d}),
being the outer
hull of its two endpoints).
Therefore,
$\lb{\overline{\xx}}{\overline{{\bf y}}}\subseteqR_{\alpha}\subseteq S$,
proving that $S$ is convex.
We next show that $\overline{S}$ is not convex.
Let $\overline{\xx}=\limray{(-{\bf e}_1)}$ and
let $\overline{{\bf y}}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_2}$.
From \eqref{eq:thm:closure-not-always-convex:2}, we see that
both $\overline{\xx}$ and $\overline{{\bf y}}$ are in $\overline{S}$ since
$\overline{{\bf y}}\in S$ and $\overline{\xx}$ is the limit of the sequence
$\seq{-t {\bf e}_1}$, all of whose elements are in $S$.
Let ${\bf z}={\bf e}_2$.
To prove $\overline{S}$ is not convex, we will show that ${\bf z}$ is on the
segment joining $\overline{\xx}$ and $\overline{{\bf y}}$, but that ${\bf z}$ is not itself in
$\overline{S}$.
Indeed, ${\bf z}\not\in\overline{S}$ since, for instance, the open set
$\{ {\bf x} \in {\R^n} : {\bf x}\cdot{\bf e}_2 > 1/2 \}$
includes ${\bf z}$ but is entirely disjoint from $S$.
It remains to show that ${\bf z}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$.
To do so, we construct sequences satisfying the conditions of
Corollary~\ref{cor:e:1}.
Specifically, for each $t$, let
\begin{eqnarray*}
{\bf x}_t &=& -t {\bf e}_1 \\
{\bf y}_t &=& t(t-1) {\bf e}_1 + t {\bf e}_2 \\
{\bf z}_t &=& \paren{1-\frac{1}{t}} {\bf x}_t + \frac{1}{t} {\bf y}_t.
\end{eqnarray*}
Then ${\bf x}_t\rightarrow\overline{\xx}$ and ${\bf y}_t\rightarrow\overline{{\bf y}}$
(by Theorem~\ref{thm:seq-rep}).
Also, by algebra, ${\bf z}_t={\bf e}_2$, for all $t$, so
${\bf z}_t\rightarrow{\bf z}$.
Thus, as claimed, ${\bf z}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$
by Corollary~\ref{cor:e:1}, so $\overline{S}$ is not convex.
\qed
\section{Monotone passages}
\label{sec:mono-passages}
The definition of convexity given in Section~\ref{sec:def-convexity}
is in terms of the segment joining two points.
In both standard and astral
convex analysis, the segment joining two (distinct)
finite points in ${\R^n}$ is an ordinary line segment, a
one-dimensional set that is
homeomorphic to the interval $[0,1]$,
and whose elements can be linearly ordered by
distance from one endpoint.
Informally, such a segment provides a natural way of getting
from one endpoint to the other along a path that is straight and
continuous.
On the other hand, in Section~\ref{sec:def-convexity},
we saw that the segment
$\lb{-{\bf I} {\bar{\bm{\omega}}}}{{\bf I} {\bar{\bm{\omega}}}}$ is all of $\extspac{n}$, and
thus that the segment joining two infinite astral points
can be of a very different nature.
In this section, we give an alternative view of segments, and so also
of what it means for a set to be convex.
Analogous to the ordinary line segment joining two finite points, we
will see how any two astral points can be connected by a set
called a \emph{monotone passage set}, which, informally, provides a
linearly ordered, continuous ``route'' or ``passage''
for getting from one endpoint to the other.
Furthermore, such a route has a monotonicity property, described below,
that can be roughly interpreted as an analog of what it means for a
set in ${\R^n}$ to be ``straight'' or ``unbending.''
We will see that the segment joining two astral points $\overline{\xx}$ and
$\overline{{\bf y}}$ in $\extspac{n}$
consists exactly of all the monotone passage sets from $\overline{\xx}$ to
$\overline{{\bf y}}$.
As a result, a set $S\subseteq\extspac{n}$ is convex if and only if it
includes every monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$, for all
$\overline{\xx},\overline{{\bf y}}\in S$.
\subsection{Definition}
In more detail, let ${\bf x}$ and ${\bf y}$ be distinct points in ${\R^n}$,
and let $P=\lb{{\bf x}}{{\bf y}}$ be the line segment joining them.
As seen in Proposition~\ref{pr:e1}(\ref{pr:e1:a}),
this set consists of all points $\xi(\lambda)$, for
$\lambda\in [0,1]$, where
\[
\xi(\lambda)
=
(1-\lambda) {\bf x} + \lambda {\bf y}.
\]
This function $\xi:[0,1]\rightarrow P$
is a homeomorphism and defines a path along $P$
from $\xi(0)={\bf x}$ to $\xi(1)={\bf y}$.
Furthermore, it provides a natural linear ordering of $P$, since $[0,1]$
is linearly ordered.
Finally, we can consider the projection of this path in any direction
${\bf u}\in{\R^n}$, that is,
\begin{equation} \label{eqn:rho-cdot-u-monotone}
\xi(\lambda) \cdot {\bf u}
=
{\bf x}\cdot{\bf u} + \lambda ({\bf y}\cdot{\bf u} - {\bf x}\cdot{\bf u}),
\end{equation}
which we here regard as a function of $\lambda\in[0,1]$.
This function is monotonic in the sense of being either
nondecreasing (if ${\bf x}\cdot{\bf u}\leq{\bf y}\cdot{\bf u}$)
or nonincreasing (if ${\bf x}\cdot{\bf u}\geq{\bf y}\cdot{\bf u}$).
For a general set $P\subseteq{\R^n}$,
it can be shown that
there exists a homeomorphism
$\xi:[0,1]\rightarrow P$
that also satisfies this monotonicity property
(of $\lambda\mapsto\xi(\lambda)\cdot{\bf u}$ being monotonic,
for all ${\bf u}\in{\R^n}$)
if and only if $P$ is a line segment (with distinct endpoints).
These notions generalize to astral space, as will be seen in detail in
this section.
To do so, we will need to allow for more general linearly ordered sets
than $[0,1]$.
We recall some standard notions regarding ordered sets.
A nonempty set $L$ (with default order relation $\leq$) is a
\emph{partial order} if
it is reflexive ($\lambda\leq\lambda$),
antisymmetric
(if $\lambda\leq\mu$ and $\mu\leq\lambda$ then $\lambda=\mu$),
and transitive
(if $\lambda\leq\mu$ and $\mu\leq\nu$ then $\lambda\leq\nu$),
for all $\lambda,\mu,\nu\in L$.
The set is a \emph{linear order} if every pair of elements
$\lambda,\mu\in L$ is \emph{comparable}, meaning either
$\lambda\leq\mu$ or $\mu\leq\lambda$.
A \emph{chain} is a linearly ordered subset of the partial order $L$.
We use the symbols $<$, $\geq$, $>$ to have their usual meanings in
terms of $\leq$.
Let $L$ be a linear order.
Then $L$ is \emph{complete} if every
nonempty subset of $L$ that has an upper bound in $L$ also has a
\emph{least} upper bound in $L$.
A subset $M\subseteq L$ is \emph{dense in $L$} if
for all $\lambda,\mu\in L$, if $\lambda<\mu$ then there exists
$\nu\in M$ with $\lambda<\nu<\mu$.
We say that $L$ is \emph{dense} if it is dense in itself.
A linear order $L$ that is both dense and complete is called
a \emph{linear continuum}.
(Note, however, that some authors require
that a set $L$ consist of at least two elements to be considered
either dense or a linear continuum;
for mathematical convenience in what follows, we here allow
$L$ to be a singleton with regard to these definitions.)
Linear continua generalize the order properties of the real line,
and will take the place of $[0,1]$ in the definitions below.
To generalize the monotonicity property discussed above for ordinary
line segments,
it will be helpful to introduce a particular
partial ordering $\preceq$
of points in $\extspac{n}$ which, crucially,
is defined relative to two
endpoints $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$.
To make these endpoints explicit, we often add the phrase
``relative to $\overline{\xx},\overline{{\bf y}}$,'' but sometimes omit this when clear from
context.
The ordering is defined as follows:
for $\overline{{\bf w}},\overline{{\bf z}}\in\extspac{n}$, we write
\[
\overline{{\bf w}} \preceq \overline{{\bf z}} \mbox{\emph{~relative to~}} \overline{\xx},\overline{{\bf y}}
\]
if it is the case that
for all ${\bf u}\in{\R^n}$,
if $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$
then $\overline{{\bf w}}\cdot{\bf u}\leq\overline{{\bf z}}\cdot{\bf u}$.
Note that this condition, applied to $-{\bf u}$,
implies also that
if $\overline{\xx}\cdot{\bf u}\geq\overline{{\bf y}}\cdot{\bf u}$
then $\overline{{\bf w}}\cdot{\bf u}\geq\overline{{\bf z}}\cdot{\bf u}$.
Thus,
$\overline{{\bf w}} \preceq \overline{{\bf z}}$ relative to $\overline{\xx},\overline{{\bf y}}$
if and only if
the ordering of $\overline{{\bf w}}\cdot{\bf u}$ and $\overline{{\bf z}}\cdot{\bf u}$
(that is, the projections of $\overline{{\bf w}}$ and $\overline{{\bf z}}$
in direction ${\bf u}$)
is consistent
with the ordering of $\overline{\xx}\cdot{\bf u}$ and $\overline{{\bf y}}\cdot{\bf u}$
(the projections of $\overline{\xx}$ and $\overline{{\bf y}}$ in that same direction ${\bf u}$),
for all ${\bf u}\in{\R^n}$.
We call this the \emph{directional order relative to $\overline{\xx},\overline{{\bf y}}$}.
The directional-order relation is indeed a partial order:
\begin{proposition}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$.
The directional order $\preceq$ relative to $\overline{\xx},\overline{{\bf y}}$ is a partial
order.
\end{proposition}
\proof
Reflexivity and transitivity
are both straightforward.
To show antisymmetry, let $\overline{{\bf w}},\overline{{\bf z}}\in\extspac{n}$ and
suppose $\overline{{\bf w}}\preceq\overline{{\bf z}}$ and $\overline{{\bf z}}\preceq\overline{{\bf w}}$.
Let ${\bf u}\in{\R^n}$.
If $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$, then, by definition of
directional ordering,
$\overline{{\bf w}}\cdot{\bf u}\leq\overline{{\bf z}}\cdot{\bf u}$ and
$\overline{{\bf z}}\cdot{\bf u}\leq\overline{{\bf w}}\cdot{\bf u}$,
so
$\overline{{\bf w}}\cdot{\bf u}=\overline{{\bf z}}\cdot{\bf u}$.
Otherwise,
if $\overline{\xx}\cdot{\bf u}\geq\overline{{\bf y}}\cdot{\bf u}$, then the preceding argument,
applied to $-{\bf u}$, shows that
$\overline{{\bf w}}\cdot(-{\bf u})=\overline{{\bf z}}\cdot(-{\bf u})$,
and therefore
$\overline{{\bf w}}\cdot{\bf u}=\overline{{\bf z}}\cdot{\bf u}$ in this case as well.
Since this equality holds for all ${\bf u}\in{\R^n}$,
it follows that $\overline{{\bf w}}=\overline{{\bf z}}$
(by Proposition~\ref{pr:i:4}).
\qed
In terms of directional ordering
(relative to $\overline{\xx},\overline{{\bf y}}$),
\eqref{eqn:lb-def} can be rewritten to say that $\lb{\overline{\xx}}{\overline{{\bf y}}}$ is
exactly the closed interval in this ordering
consisting of all
points between $\overline{\xx}$ and $\overline{{\bf y}}$:
\begin{proposition} \label{pr:lb-def-by-part-order}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$.
Then
\[
\lb{\overline{\xx}}{\overline{{\bf y}}}
=
\braces{ \overline{{\bf z}}\in\extspac{n} :
\overline{\xx} \preceq \overline{{\bf z}} \preceq \overline{{\bf y}}
\mbox{~relative to~} \overline{\xx},\overline{{\bf y}} }.
\]
\end{proposition}
\proof
Throughout this proof, $\preceq$ is relative to $\overline{\xx},\overline{{\bf y}}$.
If
$\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$ and ${\bf u}\in{\R^n}$ with
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$
then
by \eqref{eqn:lb-def},
$\overline{{\bf z}}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$,
and also
$\overline{{\bf z}}\cdot(-{\bf u})\leq\overline{\xx}\cdot(-{\bf u})$,
so that
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf z}}\cdot{\bf u}$.
Thus,
$\overline{\xx} \preceq \overline{{\bf z}} \preceq \overline{{\bf y}}$.
Conversely, if
$\overline{\xx} \preceq \overline{{\bf z}} \preceq \overline{{\bf y}}$
and ${\bf u}\in{\R^n}$ then either
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf z}}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$
or
$\overline{{\bf y}}\cdot{\bf u}\leq\overline{{\bf z}}\cdot{\bf u}\leq\overline{\xx}\cdot{\bf u}$.
In either case,
$\overline{{\bf z}}\cdot{\bf u}\leq\max\{\overline{\xx}\cdot{\bf u},\,\overline{{\bf y}}\cdot{\bf u}\}$,
so
$\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$
by \eqref{eqn:lb-def}.
\qed
We say that a function
$\xi:L\rightarrow M$, where $L$ and $M$ are partial orders,
is \emph{order-preserving} if for all
$\lambda,\mu\in L$, if $\lambda\leq\mu$ then
$\xi(\lambda)\leq\xi(\mu)$.
The function $\xi$ is an \emph{order isomorphism} if $\xi$ is a
bijection and if it also holds that
$\lambda\leq\mu$ if and only if $\xi(\lambda)\leq\xi(\mu)$,
for all $\lambda,\mu\in L$.
If $\xi:L\rightarrow P$, where $L$ is a partial
order and $P\subseteq\extspac{n}$, we add the phrase
\emph{with range relative to $\overline{\xx},\overline{{\bf y}}$} to specify
that the range $P$ is taken to be ordered by
directional order relative to $\overline{\xx},\overline{{\bf y}}$
(although often this will be understood from context).
For instance, we will soon seek maps that are order-preserving with
range relative to $\overline{\xx},\overline{{\bf y}}$.
Likewise,
if $\xi:P\rightarrow L$, we use the phrase
\emph{with domain relative to $\overline{\xx},\overline{{\bf y}}$}
to specify that the domain $P$ is ordered by
directional order relative to $\overline{\xx},\overline{{\bf y}}$.
Note that these two phrases can be used together to specify ordering
of both the domain and the range.
In earlier discussing the segment joining points ${\bf x},{\bf y}\in{\R^n}$,
we noted that the function given in \eqref{eqn:rho-cdot-u-monotone} is
monotonic.
Indeed, that discussion showed more precisely that
$\xi$ is an order isomorphism between $[0,1]$ (in the usual order)
and $P$ as ordered directionally relative to ${\bf x},{\bf y}$, that is,
$\xi$ is an order isomorphism with range relative to ${\bf x},{\bf y}$.
We will use this formulation in generalizing to astral space.
Unless indicated otherwise, we generally
take subsets of $\extspac{n}$, such as
$P$, to be in the subspace topology, and we take
linearly ordered sets, such as $L$,
to be in the order topology,
whose subbasis elements are of the form
$\{ \lambda\in L : \lambda < \lambda_0 \}$
or
$\{ \lambda\in L : \lambda > \lambda_0 \}$,
for some $\lambda_0\in L$.
With these preliminaries, for $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
we can now define a
{monotone passage from $\overline{\xx}$ to $\overline{{\bf y}}$}
to be a continuous and surjective map $\xi:L\rightarrow P$,
where $P\subseteq\extspac{n}$
and where $L$ is some linear continuum whose minimum and maximum
elements map respectively to $\overline{\xx}$ and $\overline{{\bf y}}$, and whose ordering
is preserved in the directional ordering of $P$ relative to $\overline{\xx},\overline{{\bf y}}$.
More precisely, for a linear order $L$ and a subset
$P\subseteq\extspac{n}$, we say that the function
$\xi:L\rightarrow P$
is a \emph{monotone passage from $\overline{\xx}$ to $\overline{{\bf y}}$}
if all of the following hold:
\begin{enumerate}
\item \label{property:mono-pas:a}
$L$ is a linear continuum;
\item \label{property:mono-pas:b}
$\xi$ is continuous and surjective;
\item \label{property:mono-pas:c}
$L$ has a minimum element $\lambda_{\mathrm{min}}$ and a maximum element $\lambda_{\mathrm{max}}$;
furthermore, $\xi(\lambda_{\mathrm{min}})=\overline{\xx}$ and $\xi(\lambda_{\mathrm{max}})=\overline{{\bf y}}$;
\item \label{property:mono-pas:d}
$\xi$ is order-preserving with range relative to $\overline{\xx},\overline{{\bf y}}$.
\end{enumerate}
We say that a set $P\subseteq\extspac{n}$ is a
\emph{monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$}
if it is the range of some
monotone passage $\xi:L\rightarrow P$ from $\overline{\xx}$ to $\overline{{\bf y}}$,
for some ordered set $L$.
For example,
in ${\mathbb{R}}^2$ with $\overline{{\bf z}}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf e}_1$, the
segment $P=\lb{{\bf 0}}{\overline{{\bf z}}}$ is given in
\eqref{eqn:eg-seg-zero-e2-plus-e1}.
This segment is the range of the monotone passage
$\xi:L\rightarrow P$
where $L$ is the real interval $[0,2]$, and
\begin{equation} \label{eqn:mono-path-eg}
\xi(\lambda) =
\left\{ \begin{array}{cl}
{\displaystyle \frac{\lambda}{1-\lambda}}
\, {\bf e}_2
& \mbox{if $\lambda\in [0,1)$} \\[1.5em]
\limray{{\bf e}_2} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (\lambda-1) {\bf e}_1
& \mbox{if $\lambda\in [1,2]$.}
\end{array} \right.
\end{equation}
In the definition of monotone passage, the requirement that $\xi$
is order-preserving (property~(\ref{property:mono-pas:d}))
can be restated as follows.
For a function $\xi:L\rightarrow P$, where
$L$ is a linear order and
$P\subseteq\extspac{n}$, and for each ${\bf u}\in{\R^n}$, we define the
function
$\xi_{\uu}:L\rightarrow\overline{\R}$ by
\begin{equation} \label{eqn:mphomu-def}
\xi_{\uu}(\lambda) = \xi(\lambda)\cdot{\bf u}
\end{equation}
for $\lambda\in L$.
Then it follows directly from the definition of directional ordering
that
$\xi$ is order-preserving with range relative to $\overline{\xx},\overline{{\bf y}}$
if and only if
it is the case that
for all ${\bf u}\in{\R^n}$, if $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$
then
$\xi_{\uu}(\lambda)$ is nondecreasing in $\lambda\in L$.
We will often use this formulation in proving
property~(\ref{property:mono-pas:d}).
If $\xi:L\rightarrow P$ is a monotone passage from $\overline{\xx}$ to
$\overline{{\bf y}}$, then
the linear ordering of $L$ provides, in informal terms,
a directed route or passage along the set $P$
for getting from $\overline{\xx}$ to $\overline{{\bf y}}$.
As shown in the next proposition,
because $L$ is a linear continuum,
this passage is unbroken in the sense that the set $P$ must be
topologically connected,
thus connecting $\overline{\xx}$ and $\overline{{\bf y}}$
in $\extspac{n}$.
(Recall that a \emph{separation} of a topological space $X$ is
a pair of disjoint, nonempty open sets $U$ and $V$ whose union is all
of $X$.
The space is \emph{connected} if there does not exist a separation of
$X$.)
Furthermore,
for all ${\bf u}\in{\R^n}$, the function
$\xi_{\uu}$ (as in \eqref{eqn:mphomu-def})
is continuous and monotonic in $\lambda\in L$, and
includes in its image set all values
between $\overline{\xx}\cdot{\bf u}$ and $\overline{{\bf y}}\cdot{\bf u}$.
\begin{proposition} \label{pr:mono-pass-props}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
let $P\subseteq\extspac{n}$,
and
let $\xi:L\rightarrow P$ be a monotone passage from $\overline{\xx}$ to
$\overline{{\bf y}}$, for some ordered set $L$.
Then the following hold:
\begin{enumerate}
\item \label{pr:mono-pass-props:a}
$P$ is connected.
\item \label{pr:mono-pass-props:c}
For all ${\bf u}\in{\R^n}$,
let $\xi_{\uu}:L\rightarrow\overline{\R}$
be as defined in \eqref{eqn:mphomu-def}.
Then
$\xi_{\uu}$ is continuous.
Furthermore, if $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$,
then $\xi_{\uu}$ is nondecreasing, and $L$'s image under
$\xi_{\uu}$ is the entire interval
$\xi_{\uu}(L) = [\overline{\xx}\cdot{\bf u},\,\overline{{\bf y}}\cdot{\bf u}]$.
(Correspondingly,
if $\overline{\xx}\cdot{\bf u}\geq\overline{{\bf y}}\cdot{\bf u}$,
then $\xi_{\uu}$ is nonincreasing, and
$\xi_{\uu}(L) = [\overline{{\bf y}}\cdot{\bf u},\,\overline{\xx}\cdot{\bf u}]$.)
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:mono-pass-props:a}):
Since $L$ is a linear continuum, it is also connected,
so $P$ is connected as well since it is the image of $L$
under the continuous map $\xi$
\citep[Theorems~24.1 and~23.5]{munkres}.
Part~(\ref{pr:mono-pass-props:c}):
The function $\xi_{\uu}$ is continuous because it is the composition of $\xi$,
which is continuous, with
$\overline{{\bf z}}\mapsto\overline{{\bf z}}\cdot{\bf u}$,
which is also continuous
by
Theorem~\ref{thm:i:1}(\ref{thm:i:1c}).
For the rest of the proof,
assume $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$.
The arguments for the alternative case that
$\overline{\xx}\cdot{\bf u}\geq\overline{{\bf y}}\cdot{\bf u}$ are symmetric
(or can be derived from the present case
by replacing ${\bf u}$ with $-{\bf u}$).
As discussed above,
that $\xi_{\uu}$ is
nondecreasing
follows directly from $\xi$ being order-preserving, and the
definition of directional order.
For the last claim, because
$\xi_{\uu}$ is nondecreasing,
its minimum and maximum values are
$\xi_{\uu}(\lambda_{\mathrm{min}})=\overline{\xx}\cdot{\bf u}$ and
$\xi_{\uu}(\lambda_{\mathrm{max}})=\overline{{\bf y}}\cdot{\bf u}$, respectively.
Thus, $\xi_{\uu}(L)\subseteq [\overline{\xx}\cdot{\bf u},\,\overline{{\bf y}}\cdot{\bf u}]$.
Further, because $L$ is connected, the intermediate value theorem
\citep[Theorem~24.3]{munkres}
implies that for every
$\alpha \in [\overline{\xx}\cdot{\bf u},\,\overline{{\bf y}}\cdot{\bf u}]$
(that is, between $\xi_{\uu}(\lambda_{\mathrm{min}})$ and $\xi_{\uu}(\lambda_{\mathrm{max}})$),
there must exist $\lambda\in L$ with $\xi_{\uu}(\lambda)=\alpha$.
Thus, $\xi_{\uu}(L) = [\overline{\xx}\cdot{\bf u},\,\overline{{\bf y}}\cdot{\bf u}]$,
as claimed.
\qed
In our earlier discussion of line segments between points in ${\R^n}$,
the function $\xi$ given in
\eqref{eqn:rho-cdot-u-monotone}
defines a monotone passage from ${\bf x}$ to ${\bf y}$, but also has
two additional special properties:
first, the domain of $\xi$ in this case is $[0,1]$,
making it a topological path, a particularly natural special case;
and second, $\xi$ is not just continuous, surjective and
order-preserving, but is in fact a bijection defining both a
homeomorphism and an order isomorphism.
We define terminology for both properties.
In the first case,
we say that $\xi:L\rightarrow P$ is a
\emph{monotone path from $\overline{\xx}$ to $\overline{{\bf y}}$}
if $\xi$ is a monotone passage
from $\overline{\xx}$ to $\overline{{\bf y}}$,
and also $L$ is a closed real interval
(that is, a real interval $[a,b]$, for some $a,b\in{\mathbb{R}}$
with $a\leq b$).
In the second case above,
we say that a monotone passage $\xi:L\rightarrow P$
from $\overline{\xx}$ to $\overline{{\bf y}}$ is
\emph{strict} if, in addition to the other required properties,
$\xi$ is both a homeomorphism (and therefore bijective) and an
order isomorphism with range relative to $\overline{\xx},\overline{{\bf y}}$.
These additional properties mean that
$P$ is an exact copy of $L$ in terms of both its
topology and ordering (relative to $\overline{\xx},\overline{{\bf y}}$).
As with monotone passage sets,
we similarly say that a set $P\subseteq\extspac{n}$
is a (strict) monotone path/passage \emph{set}
from $\overline{\xx}$ to $\overline{{\bf y}}$
if it is the range of some
(strict) monotone path/passage $\xi:L\rightarrow P$
from $\overline{\xx}$ to $\overline{{\bf y}}$,
for some linearly ordered set $L$.
For example, the monotone passage given in
\eqref{eqn:mono-path-eg} is in fact a strict monotone path.
In Section~\ref{sec:mono-paths},
we will explore monotone paths in greater detail, and will see that,
as in this example,
every monotone passage set
between a finite point in ${\R^n}$ and an arbitrary astral point in
$\extspac{n}$ must actually be a monotone path set.
On the other hand, between infinite points, we will see that there can
exist monotone passage sets that are not monotone path sets.
Later,
in Corollary~\ref{cor:max-chain-iff-mono-pass},
we will also see that every monotone passage set
is in fact a strict monotone passage set, so
these two notions are actually equivalent.
\subsection{Maximal chains}
We turn next to proving the most central fact about monotone passages,
namely, that the segment joining any two astral points consists of all
the monotone passage sets from one endpoint to the other.
Thus, a point $\overline{{\bf z}}$ is in the segment joining $\overline{\xx}$ and $\overline{{\bf y}}$ if
and only if it is along some monotone passage from $\overline{\xx}$ to $\overline{{\bf y}}$.
We prove this by relating monotone passages to maximal chains of
elements in the directional order, thereby allowing us to make
statements about the existence of monotone passages using Zorn's
lemma.
For $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
we say that a set $P\subseteq\extspac{n}$
is a \emph{directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$}
(or simply a \emph{chain}, if clear from context)
if $P\subseteq\lb{\overline{\xx}}{\overline{{\bf y}}}$ and $P$ is a chain in the directional
order relative to $\overline{\xx},\overline{{\bf y}}$.
Such a chain $P$ is \emph{maximal} if there does not exist any other
directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$
that properly includes $P$.
We prove in this section that a set $P$ is
a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$
if and only if
it is a
maximal directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$.
Along the way, we also prove that either of these are also equivalent
to $P$ being
a \emph{strict} monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
As a first step, we show that every monotone passage set
is a linear continuum in the directional order (and therefore also a
chain):
\begin{proposition} \label{pr:mono-pass-lin-cont}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$, and let
$P$ be a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Then $P\subseteq\lb{\overline{\xx}}{\overline{{\bf y}}}$,
and
$P$ is a linear continuum
in the directional order relative to $\overline{\xx},\overline{{\bf y}}$.
\end{proposition}
\proof
By assumption,
there exists a monotone passage $\xi:L\rightarrow P$ from $\overline{\xx}$
to $\overline{{\bf y}}$, where
$L$ is some linear continuum with
minimum $\lambda_{\mathrm{min}}$ and maximum $\lambda_{\mathrm{max}}$.
Throughout this proof, $\preceq$ is relative to $\overline{\xx},\overline{{\bf y}}$.
We show first that $P$ is a directional chain from
$\overline{\xx}$ to $\overline{{\bf y}}$.
Let $\overline{{\bf z}}\in P$,
and let
$\lambda\in\xi^{-1}(\overline{{\bf z}})$
(which is nonempty since $\xi$ is surjective).
Then $\lambda_{\mathrm{min}}\leq\lambda\leq\lambda_{\mathrm{max}}$, implying
$\overline{\xx}\preceq\overline{{\bf z}}\preceq\overline{{\bf y}}$ (since $\xi$ is order-preserving),
so $\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$
(by Proposition~\ref{pr:lb-def-by-part-order}).
Thus, $P\subseteq\lb{\overline{\xx}}{\overline{{\bf y}}}$.
Now let $\overline{{\bf z}}'\in P$,
and let $\lambda'\in\xi^{-1}(\overline{{\bf z}}')$.
Then either $\lambda\leq\lambda'$ or $\lambda'\leq\lambda$,
implying, by order preservation, that either
$\overline{{\bf z}}\preceq\overline{{\bf z}}'$ or $\overline{{\bf z}}'\preceq\overline{{\bf z}}$.
Therefore, all pairs in $P$ are comparable (relative to
$\overline{\xx},\overline{{\bf y}}$),
so $P$ is linearly ordered by $\preceq$.
We show next that this ordering of $P$ is dense.
Suppose $\overline{{\bf z}},\overline{{\bf z}}'\in P$, and that $\overline{{\bf z}}\prec\overline{{\bf z}}'$
(so that $\overline{{\bf z}}\not\succeq\overline{{\bf z}}'$).
Then by definition of directional ordering,
there must exist ${\bf u}\in{\R^n}$ with
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$ and
$\overline{{\bf z}}\cdot{\bf u}<\overline{{\bf z}}'\cdot{\bf u}$, implying
$\overline{{\bf z}}\cdot{\bf u}<\alpha<\overline{{\bf z}}'\cdot{\bf u}$ for some $\alpha\in{\mathbb{R}}$.
Let $\xi_{\uu}:L\rightarrow \overline{\R}$ be as in
\eqref{eqn:mphomu-def}.
By
Proposition~\ref{pr:mono-pass-props}(\ref{pr:mono-pass-props:c}),
there exists $\mu\in L$ with
$\xi_{\uu}(\mu)=\alpha$.
Let $\overline{{\bf w}}=\xi(\mu)$.
Then $\overline{{\bf z}}\prec\overline{{\bf w}}$ (since otherwise, we would have
$\overline{{\bf z}}\cdot{\bf u}\geq\overline{{\bf w}}\cdot{\bf u}=\alpha$, a contradiction).
Likewise $\overline{{\bf w}}\prec\overline{{\bf z}}'$.
Thus, $P$ is dense in the directional order $\preceq$.
Finally, we show that $P$ is complete in this order.
Let $Q$ be a nonempty subset of $P$;
we aim to show that $Q$ has a least upper bound.
Let $M=\xi^{-1}(Q)$, which is nonempty (since $\xi$ is surjective)
and upper bounded by $\lambda_{\mathrm{max}}$.
Therefore, $M$ has a least upper bound $\mu$ in $L$,
since $L$ is complete.
We claim that $\xi(\mu)$ is a least upper bound on $Q$
in $P$.
Let $\overline{{\bf z}}\in Q$, and let $\lambda\in\xi^{-1}(\overline{{\bf z}})$.
Then $\lambda\in M$, so $\lambda\leq\mu$, implying
$\overline{{\bf z}}=\xi(\lambda)\preceq\xi(\mu)$ since $\xi$ is order-preserving.
Thus, $\xi(\mu)$ is an upper bound on $Q$.
To see that it is the least upper bound, suppose to the contrary that
there exists an upper bound $\overline{{\bf w}}\in P$ on $Q$ with
$\overline{{\bf w}}\prec\xi(\mu)$.
As shown above, $P$ is dense, so there also exists $\overline{{\bf w}}'\in P$ with
$\overline{{\bf w}}\prec\overline{{\bf w}}'\prec\xi(\mu)$.
Let
$\nu'\in\xi^{-1}(\overline{{\bf w}}')$.
Then for all $\lambda\in M$, $\xi(\lambda)\in Q$ so
\[\xi(\lambda)\preceq\overline{{\bf w}}\prec\overline{{\bf w}}'=\xi(\nu')\prec\xi(\mu),\]
implying $\lambda<\nu'<\mu$
(using order preservation in the contrapositive).
This shows that $\nu'$ is an upper bound on $M$ that is strictly less
than $\mu$, a contradiction.
\qed
Using this proposition,
we can now prove that every monotone passage set is a maximal chain.
\begin{theorem} \label{thm:mon-pas-is-max-chain}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $P\subseteq\extspac{n}$ be
a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Then $P$ is a maximal
directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$.
\end{theorem}
\proof
Throughout this proof, we understand $P$ to be ordered in
the directional order $\preceq$ relative to $\overline{\xx},\overline{{\bf y}}$.
In this order,
$P$ is a linear continuum,
by Proposition~\ref{pr:mono-pass-lin-cont},
and also is included in $\lb{\overline{\xx}}{\overline{{\bf y}}}$.
Thus, $P$ is a
directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$.
It remains then only to show that $P$, as a chain, is maximal.
Suppose it is not, and therefore, that there exists a point
$\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}\setminus P$
that is comparable
to every point in $P$.
Let
\begin{eqnarray*}
I &=& \{ \overline{{\bf z}}'\in P : \overline{{\bf z}}' \prec \overline{{\bf z}} \} \\
J &=& \{ \overline{{\bf z}}'\in P : \overline{{\bf z}}' \succ \overline{{\bf z}} \}.
\end{eqnarray*}
Then $\overline{\xx}\in I$
(by Proposition~\ref{pr:lb-def-by-part-order}), so $I$ is not empty, and is
upper-bounded by $\overline{{\bf y}}$.
Therefore, since $P$ is a linear continuum, $I$ has a least upper
bound $\overline{{\bf y}}'\in P$.
By a similar argument, $J$ has a greatest lower bound $\overline{\xx}'\in P$.
These must be different from $\overline{{\bf z}}$, which is not in $P$.
Also, $I\cup J=P$ since $P$ is a chain with
every element comparable to $\overline{{\bf z}}$.
We consider a few cases, deriving a contradiction in each one.
If $\overline{{\bf y}}' \prec \overline{{\bf z}}$, then
$\overline{{\bf y}}' \not\succeq \overline{{\bf z}}$ so there exists ${\bf u}\in{\R^n}$ such that
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$ and
$\overline{{\bf y}}'\cdot{\bf u} < \overline{{\bf z}}\cdot{\bf u}$,
so
$\overline{{\bf y}}'\cdot{\bf u} < \alpha < \overline{{\bf z}}\cdot{\bf u}$
for some $\alpha\in{\mathbb{R}}$.
Let
$I'=\{\overline{{\bf z}}'\in P : \overline{{\bf z}}'\cdot{\bf u} < \alpha\}$
and
$J'=\{\overline{{\bf z}}'\in P : \overline{{\bf z}}'\cdot{\bf u} > \alpha\}$,
which are disjoint and
both open in $P$, being standard basis elements of $\extspac{n}$
restricted to $P$.
If $\overline{{\bf z}}'\in I$ then
$\overline{{\bf z}}'\cdot{\bf u}\leq\overline{{\bf y}}'\cdot{\bf u}<\alpha$, since $\overline{{\bf y}}'$ is an
upper bound on $I$; thus, $I\subseteq I'$.
If $\overline{{\bf z}}'\in J$ then
$\overline{{\bf z}}'\cdot{\bf u}\geq\overline{{\bf z}}\cdot{\bf u}>\alpha$, by $J$'s definition;
thus, $J\subseteq J'$.
Thus, $I'$ and $J'$ are a separation of $P$, a contradiction since $P$
is connected
(Proposition~\ref{pr:mono-pass-props}(\ref{pr:mono-pass-props:a})).
By a symmetric argument, a contradiction can be derived if
$\overline{\xx}'\succ\overline{{\bf z}}$.
So we assume henceforth that $\overline{\xx}'\prec\overline{{\bf z}}\prec\overline{{\bf y}}'$.
Because $P$ is a linear continuum, there exists an element
$\overline{{\bf z}}'\in P$ with $\overline{\xx}'\prec\overline{{\bf z}}'\prec\overline{{\bf y}}'$.
If $\overline{{\bf z}}'\prec\overline{{\bf z}}$ then $\overline{{\bf z}}'$ is a lower bound on $J$,
contradicting that $\overline{\xx}'$ is the greatest lower bound on $J$.
A symmetric contradiction is reached if $\overline{{\bf z}}'\succ\overline{{\bf z}}$.
Having reached a contradiction in all cases, we conclude that $P$ is
maximal, as claimed.
\qed
As preliminary steps in proving the converse, the next proposition
establishes the continuity properties of an order-preserving bijection
from a
linearly ordered set to a subset of $\extspac{n}$, followed by a
proposition showing that every maximal chain (and thus, every monotone
passage set) is closed in $\extspac{n}$, and therefore compact.
\begin{proposition} \label{pr:mphom-inv-is-cont}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $\xi:L\rightarrow P$ be
a bijection that is
order-preserving with range relative to $\overline{\xx},\overline{{\bf y}}$,
for some linearly ordered set $L$
and some $P\subseteq\extspac{n}$.
Then $\xi$ is an
order isomorphism (with range relative to $\overline{\xx},\overline{{\bf y}}$),
and $\xi^{-1}$ is continuous.
If, in addition, $P$ is compact, then
$\xi$ is a homeomorphism.
\end{proposition}
\proof
The directional order $\preceq$ is understood to be
relative to $\overline{\xx},\overline{{\bf y}}$ throughout this proof.
To show $\xi$ is an order isomorphism,
let $\lambda,\mu\in L$.
If $\lambda\leq\mu$ then $\xi(\lambda)\preceq\xi(\mu)$, since
$\xi$ is order-preserving.
For the converse, suppoe $\lambda\not\leq\mu$.
Then $\lambda>\mu$,
implying $\xi(\lambda)\succeq\xi(\mu)$, since $\xi$ is
order-preserving,
and that $\xi(\lambda)\neq\xi(\mu)$,
since $\xi$ is a bijection.
Therefore,
$\xi(\lambda)\not\preceq\xi(\mu)$.
To prove continuity,
it suffices to show that the image of every subbasis element
$V\subseteq L$ is open in $P$
\citep[Section~18]{munkres}.
As such,
let $V=\{ \lambda\in L : \lambda < \lambda_0 \}$
be a subbasis element,
for some $\lambda_0\in L$.
(The case that $V$ is defined by the reverse inequality,
$\lambda > \lambda_0$, is entirely symmetric.)
Let $\overline{{\bf z}}_0=\xi(\lambda_0)$,
let $\overline{{\bf z}}$ be any point in $\xi(V)$,
and let $\lambda=\xi^{-1}(\overline{{\bf z}})$, implying
$\lambda\in V$, that is, $\lambda<\lambda_0$.
Since $\xi$ was shown to be an order isomorphism,
$\overline{{\bf z}}\not\succeq\overline{{\bf z}}_0$.
Therefore, there exists ${\bf u}\in{\R^n}$ such that
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$ and $\overline{{\bf z}}\cdot{\bf u}<\overline{{\bf z}}_0\cdot{\bf u}$,
so
$\overline{{\bf z}}\cdot{\bf u}<\alpha<\overline{{\bf z}}_0\cdot{\bf u}$
for some $\alpha\in{\mathbb{R}}$.
Let
$B = \{ \overline{{\bf w}} \in P : \overline{{\bf w}}\cdot{\bf u} < \alpha \}$,
which is open (in $P$), being a standard basis element of $\extspac{n}$
restricted to $P$;
further, $B$ includes $\overline{{\bf z}}$.
We claim $B\subseteq \xi(V)$.
To see this, suppose $\overline{{\bf w}}\not\in\xi(V)$.
Then $\xi^{-1}(\overline{{\bf w}})\geq\lambda_0$,
implying $\overline{{\bf w}}\succeq\overline{{\bf z}}_0$,
so
$\overline{{\bf w}}\cdot{\bf u}\geq\overline{{\bf z}}_0\cdot{\bf u}>\alpha$.
Therefore, $\overline{{\bf w}}\not\in B$.
Thus, every point $\overline{{\bf z}}$ in $\xi(V)$ has a neighborhood that is
included in $\xi(V)$, completing the proof that $\xi(V)$ is open
in $P$, and so that $\xi^{-1}$ is continuous.
Suppose now that $P$ is compact.
The linearly ordered set $L$, in the order topology,
is Hausdorff
\citep[Theorem~17.11]{munkres}.
Since, as just shown, $\xi^{-1}$ is a continuous bijection,
these facts together imply that
$\xi^{-1}:P\rightarrow L$ is a homeomorphism
\citep[Theorem~26.6]{munkres}.
Therefore, $\xi$ is a homeomorphism as well.
\qed
\begin{proposition} \label{pr:sbar-all-comp}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $P\subseteq\extspac{n}$ be a
directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$.
Let $\overline{{\bf z}}\in\extspac{n}$.
If $\overline{{\bf z}}\in\overline{P}$
then $\overline{{\bf z}}$ is comparable (relative to $\overline{\xx},\overline{{\bf y}}$)
to every point in $P$.
Consequently, if $P\subseteq\extspac{n}$ is a maximal
directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$
(and therefore also if $P$
is a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$) then
$P$ is closed in $\extspac{n}$.
\end{proposition}
\proof
We prove the contrapositive.
Suppose there exists a
point $\overline{{\bf z}}'\in P$ that is not comparable to $\overline{{\bf z}}$.
Then $\overline{{\bf z}}\not\preceq\overline{{\bf z}}'$, so there exists ${\bf u}\in{\R^n}$ such that
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$ and $\overline{{\bf z}}\cdot{\bf u} > \overline{{\bf z}}'\cdot{\bf u}$,
implying $\overline{{\bf z}}\cdot{\bf u} > \alpha > \overline{{\bf z}}'\cdot{\bf u}$ for some
$\alpha\in{\mathbb{R}}$.
Likewise,
$\overline{{\bf z}}\not\succeq\overline{{\bf z}}'$, so there exists ${\bf u}'\in{\R^n}$ and
$\alpha'\in{\mathbb{R}}$
such that
$\overline{\xx}\cdot{\bf u}'\leq\overline{{\bf y}}\cdot{\bf u}'$ and
$\overline{{\bf z}}\cdot{\bf u}' < \alpha' < \overline{{\bf z}}'\cdot{\bf u}'$.
Let
\[
V = \{\overline{{\bf w}}\in\extspac{n} :
\overline{{\bf w}}\cdot{\bf u} > \alpha
\mbox{~and~}
\overline{{\bf w}}\cdot{\bf u}' < \alpha'\}.
\]
Then $V$ is open (being a standard basis element), and
includes $\overline{{\bf z}}$.
We claim $V$ is disjoint from $P$, which will prove
that $\overline{{\bf z}}\not\in\overline{P}$.
Suppose $\overline{{\bf w}}\in P$.
Then either $\overline{{\bf w}}\preceq\overline{{\bf z}}'$ or $\overline{{\bf w}}\succeq\overline{{\bf z}}'$
(since $P$ is a chain).
If $\overline{{\bf w}}\preceq\overline{{\bf z}}'$ then
$\overline{{\bf w}}\cdot{\bf u}\leq\overline{{\bf z}}'\cdot{\bf u}<\alpha$.
And if
$\overline{{\bf w}}\succeq\overline{{\bf z}}'$ then
$\overline{{\bf w}}\cdot{\bf u}'\geq\overline{{\bf z}}'\cdot{\bf u}'>\alpha'$.
In either case, $\overline{{\bf w}}\not\in V$, as claimed.
For the last statement of the proposition,
if $P$ is not closed in $\extspac{n}$, then
there exists a point $\overline{{\bf z}}\in\overline{P}\setminus P$ which is comparable to
every point in $P$.
Therefore, $P\cup\{\overline{{\bf z}}\}$ is a directional chain that
is a proper superset of $P$, so $P$ cannot be maximal.
Thus, every maximal directional chain
(and so also every monotone passage set, by
Theorem~\ref{thm:mon-pas-is-max-chain})
is closed in $\extspac{n}$.
\qed
We now prove that every maximal directional chain
is a strict monotone passage set:
\begin{theorem} \label{thm:max-chain-is-mon-pas}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $P\subseteq\extspac{n}$ be
a maximal
directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$.
Then $P$ is
a strict monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
\end{theorem}
\proof
To show that $P$ is a strict monotone passage set, we need to construct a linear
continuum $L$ and bijection $\xi:L\rightarrow P$
satisfying all the required properties.
To do so, we simply let $L$ be equal to $P$, with order defined
to be the same as directional ordering $\preceq$ of elements of $P$
(with all directional ordering in this proof relative to $\overline{\xx},\overline{{\bf y}}$).
Then $L$ is linearly ordered since $P$ is a chain.
Note importantly that
$L$ and $P$ are identical as sets, but their topologies are
defined differently: $P$ has the subspace topology inherited from
$\extspac{n}$, while $L$ is in the order topology associated with
the order $\preceq$.
Since the elements of both $L$ and $P$ are astral points, we use the
same notation for both, such as $\overline{{\bf z}}$, rather than $\lambda$ as used
previously.
Likewise, we use $\preceq$ to denote order in $L$, rather than
$\leq$.
We further define $\xi:L\rightarrow P$ to simply be the identity
map, meaning $\xi(\overline{{\bf z}})=\overline{{\bf z}}$ for all $\overline{{\bf z}}\in L$.
This function is clearly (and trivially) a bijection, and also an
order isomorphism (with range relative to $\overline{\xx},\overline{{\bf y}}$).
By Proposition~\ref{pr:sbar-all-comp}, $P$, being a maximal chain,
is closed in $\extspac{n}$, and so also compact.
Therefore, $\xi$ is a homeomorphism by
Proposition~\ref{pr:mphom-inv-is-cont}.
By Proposition~\ref{pr:lb-def-by-part-order},
and since $P\subseteq\lb{\overline{\xx}}{\overline{{\bf y}}}$,
$\overline{\xx}\preceq\overline{{\bf z}}\preceq\overline{{\bf y}}$, for all $\overline{{\bf z}}\in P$.
As a result, $\overline{\xx}$ must be in $P$, being comparable to all elements
of $P$, since otherwise $P\cup\{\overline{\xx}\}$ would be a chain that
properly includes $P$, contradicting that $P$ is maximal.
Likewise, $\overline{{\bf y}}\in P$.
Since the ordering of $L$ is isomorphic to that of $P$,
this further shows that
$\overline{\xx}$ and $\overline{{\bf y}}$ are minimum and maximum elements in
$L$ with $\xi(\overline{\xx})=\overline{\xx}$ and $\xi(\overline{{\bf y}})=\overline{{\bf y}}$,
proving part~(\ref{property:mono-pas:c}) in the definition of
monotone passage.
It remains only to prove that $L$ is a linear continuum, which we show
in the next two claims:
\setcounter{claimp}{0}
\medskip
\begin{claimp}
$L$ is complete.
\end{claimp}
\proof
Let $M$ be a nonempty subset of $L$ (which is upper bounded
since $\overline{{\bf y}}$ is a maximum element in $L$).
We aim to show that $M$ has a least upper bound.
Let $J$ denote the set of all points in $L$ that upper bound $M$;
that is,
\begin{eqnarray*}
J
&=&
\{ \overline{{\bf z}}\in L : \forall \overline{{\bf w}}\in M,\, \overline{{\bf w}}\preceq\overline{{\bf z}} \}
\\
&=&
\bigcap_{\overline{{\bf w}}\in M}
\{ \overline{{\bf z}}\in L : \overline{{\bf w}}\preceq\overline{{\bf z}} \}.
\end{eqnarray*}
The sets appearing in the intersection in the last line are each
closed in $L$, being complements of subbasis elements; therefore, $J$
is closed in $L$ since it is the intersection of such sets.
Let $J'=\xi(J)$.
(Of course, $J$ and $J'$ are identical as sets, but they belong
to differently defined topological spaces.)
Then $J'$ is closed in $P$,
since $\xi^{-1}$ is continuous;
therefore, $J'$ is also compact,
since $P$ is compact.
As a result, the continuous function $\xi^{-1}$,
over the compact subspace $J'$,
must attain a minimum at some point $\overline{{\bf y}}'\in J'$
\citep[Theorem~27.4]{munkres}.
This means that
$\overline{{\bf y}}'=\xi^{-1}(\overline{{\bf y}}')\preceq \overline{{\bf z}}$ for all
$\overline{{\bf z}}\in J=\xi^{-1}(J')$, and also that
$\overline{{\bf y}}'$ is itself in $J$
(so that it is itself an upper bound on $M$).
Thus, $\overline{{\bf y}}'$ is a least upper bound on $M$.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
\medskip
\begin{claimp}
$L$ is dense.
\end{claimp}
\proof
Let $\overline{\xx}',\overline{{\bf y}}'\in L$ with $\overline{\xx}' \prec \overline{{\bf y}}'$.
We aim to show there exists $\overline{{\bf z}}\in L$ with
$\overline{\xx}' \prec \overline{{\bf z}} \prec \overline{{\bf y}}'$.
Suppose, by way of contradiction, that there does not exist any such point
in $L$.
Since $\overline{\xx}'\prec\overline{{\bf y}}'$, there exists ${\bf u}\in{\R^n}$ with
$\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$ and
$\overline{\xx}'\cdot{\bf u} < \overline{{\bf y}}'\cdot{\bf u}$.
Let $\alpha\in{\mathbb{R}}$ be such that
$\overline{\xx}'\cdot{\bf u} < \alpha < \overline{{\bf y}}'\cdot{\bf u}$.
We claim that there must exist a point
$\overline{{\bf z}}\in\lb{\overline{\xx}'}{\overline{{\bf y}}'}$ with
$\overline{{\bf z}}\cdot{\bf u}=\alpha$.
By Theorem~\ref{thm:i:1}(\ref{thm:i:1d}), there exist sequences
$\seq{{\bf x}_t}$ and $\seq{{\bf y}_t}$ in ${\R^n}$ such that
${\bf x}_t\rightarrow\overline{\xx}'$ and
${\bf y}_t\rightarrow\overline{{\bf y}}'$.
Since the open set
$\{\overline{{\bf w}}\in\extspac{n} : \overline{{\bf w}}\cdot{\bf u} < \alpha\}$
includes $\overline{\xx}'$, it also must include all but finitely many of the
points ${\bf x}_t$, since they converge to $\overline{\xx}'$.
By discarding all other sequence elements, we can assume that the
entire sequence is in this open set so that ${\bf x}_t\cdot{\bf u}<\alpha$ for
all $t$.
By a similar argument, we can assume that ${\bf y}_t\cdot{\bf u}>\alpha$ for
all $t$.
Thus, ${\bf x}_t\cdot{\bf u}<\alpha<{\bf y}_t\cdot{\bf u}$.
For each $t$, let
$ {\bf z}_t= (1-\lambda_t) {\bf x}_t + \lambda_t {\bf y}_t $
where
\[
\lambda_t
=
\frac{\alpha - {\bf x}_t\cdot{\bf u}}
{{\bf y}_t\cdot{\bf u} - {\bf x}_t\cdot{\bf u}},
\]
which is in $[0,1]$.
This choice ensures that
${\bf z}_t\cdot{\bf u} = \alpha$ for all $t$.
By sequential compactness,
the resulting sequence $\seq{{\bf z}_t}$ must have a convergent
subsequence.
By discarding all other elements, we can assume the entire sequence
converges to some point $\overline{{\bf z}}\in\extspac{n}$.
This point must be in $\lb{\overline{\xx}'}{\overline{{\bf y}}'}$ since all the conditions
of Corollary~\ref{cor:e:1}
have been satisfied.
Furthermore, $\alpha={\bf z}_t\cdot{\bf u}\rightarrow\overline{{\bf z}}\cdot{\bf u}$,
by construction and Theorem~\ref{thm:i:1}(\ref{thm:i:1c}),
so $\overline{{\bf z}}\cdot{\bf u}=\alpha$, proving the claim.
For all ${\bf u}'\in{\R^n}$, if $\overline{\xx}\cdot{\bf u}'\leq\overline{{\bf y}}\cdot{\bf u}'$ then
$\overline{\xx}'\cdot{\bf u}'\leq\overline{{\bf y}}'\cdot{\bf u}'$ (since $\overline{\xx}'\prec\overline{{\bf y}}'$)
so
$\overline{\xx}'\cdot{\bf u}'\leq\overline{{\bf z}}\cdot{\bf u}'\leq\overline{{\bf y}}'\cdot{\bf u}'$
by
Proposition~\ref{pr:lb-def-by-part-order}
since $\overline{{\bf z}}\in\lb{\overline{\xx}'}{\overline{{\bf y}}'}$.
This shows that
$\overline{\xx}'\preceq\overline{{\bf z}}\preceq\overline{{\bf y}}'$,
so actually
$\overline{\xx}'\prec\overline{{\bf z}}\prec\overline{{\bf y}}'$
since
$\overline{\xx}'\cdot{\bf u} < \overline{{\bf z}}\cdot{\bf u} < \overline{{\bf y}}'\cdot{\bf u}$.
It follows that $\overline{{\bf z}}\not\in P$,
since we assumed there is no point in $P$ that is strictly
between $\overline{\xx}'$ and $\overline{{\bf y}}'$.
We claim every point in $P$ is comparable to $\overline{{\bf z}}$.
This is because if $\overline{{\bf w}}\in P$,
then, by our initial assumption, it cannot be that
$\overline{\xx}'\prec\overline{{\bf w}}\prec\overline{{\bf y}}'$, so either
$\overline{{\bf w}}\preceq\overline{\xx}'\prec\overline{{\bf z}}$
or
$\overline{{\bf z}}\prec\overline{{\bf y}}'\preceq\overline{{\bf w}}$.
Thus, $P\cup\{\overline{{\bf z}}\}$ is a chain that is a proper superset of $P$,
contradicting that $P$ is a maximal chain.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
We conclude that $\xi$ is a strict monotone passage, having proved all
parts of its definition.
\qed
Combining yields the three-way equivalence of
monotone passage sets,
strict monotone passage sets, and
maximal chains:
\begin{corollary} \label{cor:max-chain-iff-mono-pass}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $P\subseteq\extspac{n}$.
Then the following are equivalent:
\begin{enumerate}
\item \label{cor:max-chain-iff-mono-pass:a}
$P$ is a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
\item \label{cor:max-chain-iff-mono-pass:b}
$P$ is a strict monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
\item \label{cor:max-chain-iff-mono-pass:c}
$P$ is a maximal directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$.
\end{enumerate}
\end{corollary}
\proof
That
(\ref{cor:max-chain-iff-mono-pass:a})
$\Rightarrow$
(\ref{cor:max-chain-iff-mono-pass:c})
and
(\ref{cor:max-chain-iff-mono-pass:c})
$\Rightarrow$
(\ref{cor:max-chain-iff-mono-pass:b})
follow respectively from
Theorems~\ref{thm:mon-pas-is-max-chain}
and~\ref{thm:max-chain-is-mon-pas}.
That
(\ref{cor:max-chain-iff-mono-pass:b})
$\Rightarrow$
(\ref{cor:max-chain-iff-mono-pass:a})
is immediate.
\qed
As an immediate corollary, the subspace topology on a monotone passage
set $P$ from $\overline{\xx}$ to $\overline{{\bf y}}$
is identical to the order topology on that same set (under the
directional order relative to $\overline{\xx},\overline{{\bf y}}$).
\begin{corollary} \label{cor:subspace-equals-order-topo}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and let $P$
be a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Then the set $P$ in the subspace topology (as a subspace of
$\extspac{n}$) is homeomorphic to the same set $P$ in the order
topology induced by directional order relative to $\overline{\xx},\overline{{\bf y}}$.
\end{corollary}
\proof
By Corollary~\ref{cor:max-chain-iff-mono-pass}, there exists a strict
monotone passage $\xi:L\rightarrow P$ for some linearly ordered set
$L$.
Since $\xi$ is an order isomorphism (with range relative to
$\overline{\xx},\overline{{\bf y}}$), the order topology on $P$ is
homeomorphic under $\xi^{-1}$ to the order
topology on $L$ under the bijection $\xi$.
Furthermore, since $\xi$ is a homeomorphism, the set $L$ in the
order topology is homeomorphic to $P$ in the subspace topology.
Composing homeomorphisms yields the corollary.
\qed
\subsection{Existence using Zorn's lemma}
We can use the characterization of monotone passage sets as maximal
chains to prove their existence by direct application of Zorn's lemma.
We use a version of Zorn's lemma specialized to families of sets
\citep[Section~10.2]{davey_priestley_02}.
We say that a family of sets ${\cal C}$ is an
\emph{inclusion-chain} if for all $P,P'\in{\cal C}$,
either $P\subseteq P'$ or $P'\subseteq P$.
Suppose that some family of sets ${\cal P}$ has the property that if
${\cal C}\subseteq{\cal P}$ is any nonempty inclusion-chain, then
\[ \bigcup_{P\in{\cal C}} P\in {\cal P}, \]
that is,
the union of all sets in the inclusion-chain ${\cal C}$
is also one of the sets in ${\cal P}$.
According to Zorn's lemma (specialized to this setting),
if ${\cal P}$ has this property, then there exists a set $P\in{\cal P}$ that
is \emph{maximal}, meaning it is not a proper subset of any set in
${\cal P}$.
(Both Zorn's lemma and Tychonoff's theorem, which we previously
invoked in Section~\ref{subsec:astral-pts-as-fcns},
require that we assume the axiom of choice.)
Using Zorn's lemma, we prove that any directional
chain $P_0$ from $\overline{\xx}$ to $\overline{{\bf y}}$ can be enlarged into a maximal chain
(and therefore a monotone passage set) while still
including the original ``seed'' chain.
For example, if we take $P_0=\{\overline{{\bf z}}\}$, where
$\overline{{\bf z}}$ is any point in $\lb{\overline{\xx}}{\overline{{\bf y}}}$,
this will prove that there exists a
monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$ that includes $\overline{{\bf z}}$.
\begin{theorem} \label{thm:exists-mon-pass-with-chain}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and let $P_0\subseteq\lb{\overline{\xx}}{\overline{{\bf y}}}$ be a
directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$.
Then there exists a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$
that includes $P_0$.
\end{theorem}
\proof
Let ${\cal P}$ consist of all directional chains $P$ from
$\overline{\xx}$ to $\overline{{\bf y}}$ that include $P_0$.
This family of sets is nonempty since, for instance, it includes
$P_0$.
We claim that this family has the property required by Zorn's lemma,
as described above.
To see this, let ${\cal C}\subseteq{\cal P}$ be a nonempty inclusion-chain in
${\cal P}$, and let
$U = \bigcup_{P\in{\cal C}} P$,
which we aim to show is in ${\cal P}$.
Since ${\cal C}$ is nonempty, it includes some set $P$ which, being in
${\cal P}$, must include $P_0$.
Thus, $P_0\subseteq U$.
Also, every set $P\in{\cal P}$ is included in $\lb{\overline{\xx}}{\overline{{\bf y}}}$, so
$U\subseteq\lb{\overline{\xx}}{\overline{{\bf y}}}$.
To show that $U$ is a directional chain,
let $\overline{{\bf w}},\overline{{\bf z}}\in U$, which we aim to show are comparable
(in the directional ordering relative to $\overline{\xx},\overline{{\bf y}}$).
By $U$'s definition,
there must exist sets $P$ and $P'$ in ${\cal C}$ with
$\overline{{\bf w}}\in P$
and
$\overline{{\bf z}}\in P'$.
Further, since ${\cal C}$ is an inclusion-chain, one of these sets must
be contained in the other.
Without loss of generality, assume $P'\subseteq P$.
Then $\overline{{\bf w}}$ and $\overline{{\bf z}}$ must both be in $P$,
implying that they are comparable
since $P$, being in ${\cal P}$,
is a directional chain.
Thus, $U$ is a directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$
that includes $P_0$, and therefore, $U\in{\cal P}$ as claimed.
As a result, Zorn's lemma now implies that there exists a maximal set
$P$ in ${\cal P}$.
By construction of ${\cal P}$,
this set includes $P_0$ and is a directional chain from
$\overline{\xx}$ to $\overline{{\bf y}}$ which must be maximal, since otherwise there would
be a set $P'$ in ${\cal P}$ that properly contains it.
Since $P$ is a maximal directional chain, it is also
a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$,
by Theorem~\ref{thm:max-chain-is-mon-pas}.
\qed
The next corollary summarizes some consequences of
Corollary~\ref{cor:max-chain-iff-mono-pass}
and
Theorem~\ref{thm:exists-mon-pass-with-chain},
many of which have already been discussed:
\begin{corollary} \label{cor:mon-pass-max-chain-conseq}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$.
\begin{enumerate}
\item \label{cor:mon-pass-max-chain-conseq:a}
There exists a monotone passage from $\overline{\xx}$ to $\overline{{\bf y}}$.
\item \label{cor:mon-pass-max-chain-conseq:b}
The segment $\lb{\overline{\xx}}{\overline{{\bf y}}}$ joining $\overline{\xx}$ and $\overline{{\bf y}}$ is
exactly equal to the union of all monotone passage sets from $\overline{\xx}$
to $\overline{{\bf y}}$.
That is, a point $\overline{{\bf z}}\in\extspac{n}$ is in $\lb{\overline{\xx}}{\overline{{\bf y}}}$
if and only if $\overline{{\bf z}}$ is included in some monotone passage set from
$\overline{\xx}$ to $\overline{{\bf y}}$.
\item \label{cor:mon-pass-max-chain-conseq:c}
A set $S\subseteq\extspac{n}$ is convex if and only if for all
$\overline{\xx}',\overline{{\bf y}}'\in S$, every monotone passage set from $\overline{\xx}'$ to
$\overline{{\bf y}}'$ is included in $S$.
\item \label{cor:mon-pass-max-chain-conseq:d}
Every nonempty convex set in $\extspac{n}$ is connected.
\item \label{cor:mon-pass-max-chain-conseq:e}
There exists only a single monotone passage set
from $\overline{\xx}$ to $\overline{{\bf y}}$
if and only if
the entire segment $\lb{\overline{\xx}}{\overline{{\bf y}}}$ is itself a monotone
passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
\end{enumerate}
\end{corollary}
\proof
~
Part~(\ref{cor:mon-pass-max-chain-conseq:a}):
This follows immediately from
Theorem~\ref{thm:exists-mon-pass-with-chain}
applied with $P_0=\emptyset$.
Part~(\ref{cor:mon-pass-max-chain-conseq:b}):
If $\overline{{\bf z}}$ is included in a monotone passage set $P$ from $\overline{\xx}$ to
$\overline{{\bf y}}$,
then $P$ is a (maximal)
{directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$},
by
Theorem~\ref{thm:mon-pas-is-max-chain},
so, by definition, $\overline{{\bf z}}\in P\subseteq \lb{\overline{\xx}}{\overline{{\bf y}}}$.
Conversely, if $\overline{{\bf z}}\in\lb{\overline{\xx}}{\overline{{\bf y}}}$,
then by Theorem~\ref{thm:exists-mon-pass-with-chain}
applied to $P_0=\{\overline{{\bf z}}\}$, there must exist
a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$
that includes $\overline{{\bf z}}$.
Part~(\ref{cor:mon-pass-max-chain-conseq:c}):
This follows immediately from
part~(\ref{cor:mon-pass-max-chain-conseq:b}),
and because, by definition of convexity, $S$ is convex if and only if
$\lb{\overline{\xx}'}{\overline{{\bf y}}'}\subseteq S$ for all $\overline{\xx}',\overline{{\bf y}}'\in S$.
Part~(\ref{cor:mon-pass-max-chain-conseq:d}):
Suppose to the contrary that some set $S\subseteq\extspac{n}$ is
nonempty, convex, but not connected.
Then there exist open sets $U$ and $V$ in $\extspac{n}$ such that
$S\subseteq U\cup V$,
and
$U\cap S$ and $V\cap S$ are disjoint and nonempty.
Let $\overline{\xx}'\in U\cap S$ and $\overline{{\bf y}}'\in V\cap S$, and let
$P$ be a monotone passage set from $\overline{\xx}'$ to $\overline{{\bf y}}'$, which exists by
part~(\ref{cor:mon-pass-max-chain-conseq:a}),
and is included in $S$ by
part~(\ref{cor:mon-pass-max-chain-conseq:c}).
Then
$P\subseteq S\subseteq U\cup V$,
and
$U\cap P$ and $V\cap P$ are disjoint
and nonempty, and so are a
separation of $P$, contradicting that $P$ is connected
by Proposition~\ref{pr:mono-pass-props}(\ref{pr:mono-pass-props:a}).
Part~(\ref{cor:mon-pass-max-chain-conseq:e}):
Suppose $\lb{\overline{\xx}}{\overline{{\bf y}}}$ is itself a monotone passage set
from $\overline{\xx}$ to $\overline{{\bf y}}$, and so is a
{directional chain from $\overline{\xx}$ to $\overline{{\bf y}}$},
by
Theorem~\ref{thm:mon-pas-is-max-chain}.
Then no proper subset of $\lb{\overline{\xx}}{\overline{{\bf y}}}$ can be a maximal
directional chain,
implying there can be no other monotone passage set from $\overline{\xx}$ to
$\overline{{\bf y}}$
(again by Theorem~\ref{thm:mon-pas-is-max-chain}).
Conversely,
using
part~(\ref{cor:mon-pass-max-chain-conseq:b}),
if $\lb{\overline{\xx}}{\overline{{\bf y}}}$ is not a monotone
passage set from $\overline{\xx}$ to $\overline{{\bf y}}$, there must nevertheless exist
such a monotone
passage set $P$, which must be a proper subset of $\lb{\overline{\xx}}{\overline{{\bf y}}}$.
Thus, there must exist a point
$\overline{{\bf z}}\in \lb{\overline{\xx}}{\overline{{\bf y}}}\setminus P$,
which
must be included in some other monotone passage set from $\overline{\xx}$ to
$\overline{{\bf y}}$ that is different from $P$.
Therefore, $P$ is not unique.
\qed
\subsection{Operating on monotone passages}
Next, we provide some tools for working with monotone passages.
To begin, we show that the image of a monotone
passage set under a map that is continuous and appropriately
order-preserving is also a monotone passage set; the same holds for
monotone paths.
\begin{proposition} \label{pr:mono-pass-map}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and let $P$ be a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Let $\rho:P\rightarrow\extspac{m}$ be continuous
and order-preserving
with domain relative to $\overline{\xx},\overline{{\bf y}}$,
and range relative to $\rho(\overline{\xx}),\rho(\overline{{\bf y}})$.
Then $\rho(P)$ is a monotone passage set from
$\rho(\overline{\xx})$ to $\rho(\overline{{\bf y}})$.
If, in addition,
$P$ is a monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$,
then $\rho(P)$ is a monotone path set from
$\rho(\overline{\xx})$ to $\rho(\overline{{\bf y}})$.
\end{proposition}
\proof
By assumption, there exists a monotone passage $\xi:L\rightarrow P$
from $\overline{\xx}$ to $\overline{{\bf y}}$, for some linearly ordered set $L$
with minimum $\lambda_{\mathrm{min}}$ and maximum $\lambda_{\mathrm{max}}$.
Let $P'=\rho(P)$, $\overline{\xx}'=\rho(\overline{\xx})$, and $\overline{{\bf y}}'=\rho(\overline{{\bf y}})$.
Also,
let $\xi':L\rightarrow P'$ be defined by
$\xi'(\lambda)=\rho(\xi(\lambda))$, for $\lambda\in L$.
Then $\xi'$ is continuous, since it is the composition of
continuous functions (with restricted range, which does not affect
continuity).
It is surjective, since $\xi$ is surjective.
It is order-preserving with range relative to $\overline{\xx}',\overline{{\bf y}}'$ since
for $\lambda,\mu\in L$,
if $\lambda\leq \mu$
then $\xi(\lambda)\preceq\xi(\mu)$ relative to $\overline{\xx},\overline{{\bf y}}$,
implying $\xi'(\lambda)\preceq\xi'(\mu)$ relative to $\overline{\xx}',\overline{{\bf y}}'$.
Finally, $\xi'(\lambda_{\mathrm{min}})=\overline{\xx}'$ and $\xi'(\lambda_{\mathrm{max}})=\overline{{\bf y}}'$.
Thus, $\xi'$ is a monotone passage from $\overline{\xx}'$ to $\overline{{\bf y}}'$.
If $P$ is a monotone path set, we can take $L$ to be a closed real
interval, yielding that $\xi'$ is a monotone path from
$\overline{\xx}'$ to $\overline{{\bf y}}'$.
\qed
Consequently, the image of a monotone passage set under an
affine map is also a monotone passage set
(likewise for monotone paths).
\begin{theorem} \label{thm:mono-pass-affine-map}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and let $P$ be a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Let ${\bf A}\in{\R^{m\times n}}$ and $\overline{\bb}\in\extspac{m}$,
and let $F:\extspac{n}\rightarrow\extspac{m}$
be defined by
$F(\overline{{\bf z}})=\overline{\bb}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf A}\overline{{\bf z}}$.
Then $F(P)$ is a monotone passage set from
$F(\overline{\xx})$ to $F(\overline{{\bf y}})$.
If, in addition,
$P$ is a monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$,
then $F(P)$ is a monotone path set from
$F(\overline{\xx})$ to $F(\overline{{\bf y}})$.
\end{theorem}
\proof
Let $\rho:P\rightarrow\extspac{m}$ be the restriction of $F$ to domain
$P$.
By Proposition~\ref{pr:aff-cont}, $F$ is continuous, so $\rho$ is as
well.
We need only show that
$\rho$ is order-preserving with domain relative to $\overline{\xx},\overline{{\bf y}}$ and
range relative to $F(\overline{\xx}),F(\overline{{\bf y}})$.
Once established, the theorem then follows by a direct application of
Proposition~\ref{pr:mono-pass-map} to $\rho$.
Suppose $\overline{{\bf z}},\overline{{\bf z}}'\in P$ with $\overline{{\bf z}}\preceq\overline{{\bf z}}'$
(relative to $\overline{\xx},\overline{{\bf y}}$).
We aim to show $F(\overline{{\bf z}})\preceq F(\overline{{\bf z}}')$
(relative to $F(\overline{\xx}),F(\overline{{\bf y}})$).
For this purpose,
let ${\bf u}\in{\R^m}$
be such that $F(\overline{\xx})\cdot{\bf u}\leq F(\overline{{\bf y}})\cdot{\bf u}$.
Our goal then is to show that
$F(\overline{{\bf z}})\cdot{\bf u}\leq F(\overline{{\bf z}}')\cdot{\bf u}$.
For all $\overline{{\bf w}}\in\extspac{n}$,
\begin{equation} \label{eqn:thm:mono-pass-affine-map:1}
F(\overline{{\bf w}})\cdot{\bf u}
=
\overline{\bb}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf A}\overline{{\bf w}})\cdot{\bf u}
=
\overline{\bb}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf w}}\cdot(\trans{{\bf A}}{\bf u})
\end{equation}
by Proposition~\ref{pr:mat-mult-def}.
In particular,
if $\overline{\bb}\cdot{\bf u}\in \{-\infty,+\infty\}$,
this implies
$F(\overline{{\bf z}})\cdot{\bf u}=\overline{\bb}\cdot{\bf u}=F(\overline{{\bf z}}')\cdot{\bf u}$.
Otherwise, $\overline{\bb}\cdot{\bf u}\in{\mathbb{R}}$.
In this case, \eqref{eqn:thm:mono-pass-affine-map:1} implies,
for all $\overline{{\bf w}},\overline{{\bf w}}'\in\extspac{n}$, that
$F(\overline{{\bf w}})\cdot{\bf u} \leq F(\overline{{\bf w}}')\cdot{\bf u}$
if and only if
$\overline{{\bf w}}\cdot(\trans{{\bf A}}{\bf u}) \leq \overline{{\bf w}}'\cdot(\trans{{\bf A}}{\bf u})$.
In particular, since
$F(\overline{\xx})\cdot{\bf u}\leq F(\overline{{\bf y}})\cdot{\bf u}$,
it now follows that
$\overline{\xx}\cdot(\trans{{\bf A}}{\bf u}) \leq \overline{{\bf y}}\cdot(\trans{{\bf A}}{\bf u})$.
Therefore,
$\overline{{\bf z}}\cdot(\trans{{\bf A}}{\bf u}) \leq \overline{{\bf z}}'\cdot(\trans{{\bf A}}{\bf u})$,
since
$\overline{{\bf z}}\preceq\overline{{\bf z}}'$ (relative to $\overline{\xx},\overline{{\bf y}}$),
so
$F(\overline{{\bf z}})\cdot{\bf u} \leq F(\overline{{\bf z}}')\cdot{\bf u}$,
completing the proof that $\rho$ is order-preserving.
\qed
We next consider subpassages, which are subsections of a monotone
passage set.
Formally,
let $P$ be a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$,
and let $\overline{\xx}',\overline{{\bf y}}'\in P$ be such that
$\overline{\xx}\preceq\overline{\xx}'\preceq\overline{{\bf y}}'\preceq\overline{{\bf y}}$
relative to $\overline{\xx},\overline{{\bf y}}$.
Then we define the \emph{subpassage of $P$ from $\overline{\xx}'$ to $\overline{{\bf y}}'$}
to be all points in $P$ between $\overline{\xx}'$ and $\overline{{\bf y}}'$ in the directional
order relative to $\overline{\xx},\overline{{\bf y}}$, that is, the interval
\[
\braces{ \overline{{\bf z}}\in\extspac{n} :
\overline{\xx}' \preceq \overline{{\bf z}} \preceq \overline{{\bf y}}'
\mbox{~relative to~} \overline{\xx},\overline{{\bf y}} }.
\]
Note that the endpoints of the interval are $\overline{\xx}'$ and $\overline{{\bf y}}'$, but
the ordering is relative to $\overline{\xx},\overline{{\bf y}}$.
As we show next, such a subpassage
is itself a monotone passage from $\overline{\xx}'$ to $\overline{{\bf y}}'$,
meaning a monotone passage set can be broken apart into
``smaller'' monotone passage sets.
\begin{theorem} \label{thm:subpassages}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$, and
let $P\subseteq\extspac{n}$ be a monotone passage set
from $\overline{\xx}$ to $\overline{{\bf y}}$.
Let $\overline{\xx}',\overline{{\bf y}}'\in P$ be such that
$\overline{\xx}\preceq\overline{\xx}'\preceq\overline{{\bf y}}'\preceq\overline{{\bf y}}$
relative to $\overline{\xx},\overline{{\bf y}}$.
Let $P'$ be the subpassage of $P$ from $\overline{\xx}'$ to $\overline{{\bf y}}'$.
Then $P'$ is a monotone passage set from $\overline{\xx}'$ to $\overline{{\bf y}}'$.
If, in addition, $P$ is a monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$,
then $P'$ is a monotone path set from $\overline{\xx}'$ to $\overline{{\bf y}}'$.
\end{theorem}
\proof
For clarity, throughout this proof,
we write $\preceq$ for directional order relative to $\overline{\xx},\overline{{\bf y}}$,
and $\preceq'$ for directional order relative to $\overline{\xx}',\overline{{\bf y}}'$.
We prove the theorem using Proposition~\ref{pr:mono-pass-map}.
To do so, let $F:P\rightarrow \extspac{n}$ be defined by
\[
F(\overline{{\bf z}})
=
\left\{ \begin{array}{cl}
\overline{\xx}' & \mbox{if $\overline{{\bf z}}\preceq\overline{\xx}'$} \\
\overline{{\bf z}} & \mbox{if $\overline{\xx}'\preceq\overline{{\bf z}}\preceq\overline{{\bf y}}'$} \\
\overline{{\bf y}}' & \mbox{if $\overline{{\bf y}}'\preceq\overline{{\bf z}}$.}
\end{array} \right.
\]
Thus, $F$ ``clamps'' $P$ between $\overline{\xx}'$ and $\overline{{\bf y}}'$.
Note that $F(P)=P'$, $F(\overline{\xx})=\overline{\xx}'$ and $F(\overline{{\bf y}})=\overline{{\bf y}}'$.
This function has three pieces:
Two of these pieces are
constant-valued functions, which are therefore continuous.
The third piece is the identity function
over a subspace of $P$ (in the same topology for domain and range),
and so is also continuous.
Moreover, these pieces are defined over intervals of $P$
that are each closed by
Corollary~\ref{cor:subspace-equals-order-topo},
being the complement of open sets in the order topology under $\preceq$.
Therefore, the piecewise function $F$ is continuous by a standard
pasting lemma
\citep[Theorem~18.3]{munkres}.
We next show that $F$ is order-preserving
with domain relative to $\overline{\xx},\overline{{\bf y}}$ and
with range relative to $\overline{\xx}',\overline{{\bf y}}'$.
Let $\overline{{\bf w}},\overline{{\bf z}}\in P$, and let
$\overline{{\bf w}}'=F(\overline{{\bf w}})$ and $\overline{{\bf z}}'=F(\overline{{\bf z}})$.
We suppose $\overline{{\bf w}}\preceq\overline{{\bf z}}$, which we aim to show implies
$\overline{{\bf w}}'\preceq'\overline{{\bf z}}'$ (that is, in the directional
order relative to $\overline{\xx}',\overline{{\bf y}}'$).
We note first that $\overline{{\bf w}}'\preceq\overline{{\bf z}}'$ (where, importantly, the
order here is relative to $\overline{\xx},\overline{{\bf y}}$).
This is because
if $\overline{{\bf w}}\preceq\overline{\xx}'$ then $\overline{{\bf w}}'=\overline{\xx}'\preceq\overline{{\bf z}}'$;
if $\overline{{\bf y}}'\preceq\overline{{\bf z}}$ then $\overline{{\bf w}}'\preceq\overline{{\bf y}}'=\overline{{\bf z}}'$;
otherwise,
$\overline{\xx}'\preceq\overline{{\bf w}}\preceq\overline{{\bf z}}\leq\overline{{\bf y}}'$ so
$\overline{{\bf w}}'=\overline{{\bf w}}\preceq\overline{{\bf z}}=\overline{{\bf z}}'$.
Thus, $\overline{\xx}'\preceq\overline{{\bf w}}'\preceq\overline{{\bf z}}'\preceq\overline{{\bf y}}'$.
To show $\overline{{\bf w}}'\preceq'\overline{{\bf z}}'$, by definition,
we need to prove that
for all ${\bf u}\in{\R^n}$,
if $\overline{\xx}'\cdot{\bf u}\leq\overline{{\bf y}}'\cdot{\bf u}$
then
$\overline{{\bf w}}'\cdot{\bf u}\leq\overline{{\bf z}}'\cdot{\bf u}$.
We prove this in the contrapositive.
Let ${\bf u}\in{\R^n}$ and suppose
$\overline{{\bf w}}'\cdot{\bf u}>\overline{{\bf z}}'\cdot{\bf u}$.
Combined with $\overline{{\bf w}}'\preceq\overline{{\bf z}}'$, this implies
$\overline{\xx}\cdot{\bf u}>\overline{{\bf y}}\cdot{\bf u}$.
Therefore,
\[
\overline{\xx}'\cdot{\bf u}
\geq
\overline{{\bf w}}'\cdot{\bf u}
>
\overline{{\bf z}}'\cdot{\bf u}
\geq
\overline{{\bf y}}'\cdot{\bf u},
\]
where the first and last inequalities are because
$\overline{\xx}'\preceq\overline{{\bf w}}'$ and $\overline{{\bf z}}'\preceq\overline{{\bf y}}'$.
Thus,
$\overline{\xx}'\cdot{\bf u}>\overline{{\bf y}}'\cdot{\bf u}$,
as needed,
so $\overline{{\bf w}}'\preceq'\overline{{\bf z}}'$ and $F$ is order-preserving.
Applying Proposition~\ref{pr:mono-pass-map} now shows that
$F(P)=P'$ is a monotone passage set
(or monotone path set, if $P$ is a monotone path set)
from $F(\overline{\xx})=\overline{\xx}'$ to $F(\overline{{\bf y}})=\overline{{\bf y}}'$.
\qed
Using Theorem~\ref{thm:subpassages},
we can show that the part of any monotone passage that passes through
${\R^n}$ must entirely lie along a line.
Combined with Theorem~\ref{thm:mono-pass-affine-map}, this implies
that the same holds for the image of any monotone passage
set under any affine map; that is, the intersection of that image with
${\R^n}$ must also lie along a line.
This is one of the ways in which monotone passage sets retain certain
linear properties, even when infinite astral points are involved.
\begin{proposition} \label{pr:mon-pass-collinear-in-rn}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $P\subseteq\extspac{n}$ be a monotone passage set
from $\overline{\xx}$ to $\overline{{\bf y}}$
(with $n\geq 1$).
Then there exists a line in ${\R^n}$ that includes
$P\cap{\R^n}$.
\end{proposition}
\proof
Let $P'=P\cap{\R^n}$.
As a first step, let ${\bf x},{\bf y},{\bf z}$ be any three points in
$P'$; we claim they must be collinear.
Because these points are included in $P$, which is a monotone passage
set,
they must be
comparable to one another in the directional order $\preceq$
(which is relative to
$\overline{\xx},\overline{{\bf y}}$ throughout this proof),
by Proposition~\ref{pr:mono-pass-lin-cont}.
With possible renaming of variables, we thus can assume
${\bf x}\preceq{\bf z}\preceq{\bf y}$.
By Theorem~\ref{thm:subpassages}, it follows that ${\bf z}$ belongs to a
monotone passage set from ${\bf x}$ to ${\bf y}$, which in turn is included in
$\lb{{\bf x}}{{\bf y}}$
(by Proposition~\ref{pr:mono-pass-lin-cont}).
Thus, ${\bf z}$ is on the ordinary line segment joining ${\bf x}$ and ${\bf y}$
(Proposition~\ref{pr:e1}(\ref{pr:e1:a})).
Therefore, the three points are collinear.
We now prove the proposition.
If $P'$ is empty or is a singleton, then the proposition is
trivially true.
Otherwise, let ${\bf x}$ and ${\bf y}$ be any two distinct points in $P'$.
Then every other point in $P'$ is collinear with ${\bf x}$ and ${\bf y}$,
as just argued,
and therefore is included in the line determined by the two points.
\qed
In the examples we have seen so far, there has existed just one
monotone passage set from one point to another.
Indeed, Theorem~\ref{thm:y-xbar-is-path} will later show that this is
always the case when at least one of the points is finite.
Nevertheless, using
Proposition~\ref{pr:mon-pass-collinear-in-rn},
we now can see that there sometimes must exist multiple
distinct monotone passage sets connecting two infinite points.
Indeed, whenever the segment $\lb{\overline{\xx}}{\overline{{\bf y}}}$ joining points
$\overline{\xx}$ and $\overline{{\bf y}}$ includes three points in ${\R^n}$ that are not
collinear,
Proposition~\ref{pr:mon-pass-collinear-in-rn}
implies that those points cannot all belong to the same monotone
passage set from $\overline{\xx}$ to $\overline{{\bf y}}$, and therefore there must exist
more than one
(by
Corollary~\ref{cor:mon-pass-max-chain-conseq}(\ref{cor:mon-pass-max-chain-conseq:b})).
For instance, we saw in Section~\ref{sec:def-convexity}
that
$\lb{-{\bf I} {\bar{\bm{\omega}}}}{{\bf I} {\bar{\bm{\omega}}}}=\extspac{n}\supseteq{\R^n}$,
and therefore, by this argument, there
must exist more than one monotone passage set
from $-{\bf I} {\bar{\bm{\omega}}}$ to ${\bf I} {\bar{\bm{\omega}}}$
(assuming $n\geq 2$).
Theorem~\ref{thm:subpassages} showed
how monotone passages can be broken apart.
Next, we show how they can be naturally pieced together:
\begin{theorem} \label{thm:combine-mono-pass}
Let $\overline{\xx},\overline{{\bf y}},\xbarj{0},\ldots,\xbarj{s}\in\extspac{n}$ with
\begin{equation} \label{eqn:thm:combine-mono-pass:1}
\overline{\xx}=\xbarj{0}\preceq\xbarj{1}\preceq\cdots\preceq\xbarj{s}=\overline{{\bf y}}
\end{equation}
relative to $\overline{\xx},\overline{{\bf y}}$.
For $j=1,\ldots,s$,
let $\Pj{j}\subseteq\extspac{n}$
be a monotone passage set from $\xbarj{j-1}$ to $\xbarj{j}$.
Let $P=\bigcup_{j=1}^s \Pj{j}$.
Then $P$ is a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
If, in addition, each $\Pj{j}$ is
a monotone path set from $\xbarj{j-1}$ to $\xbarj{j}$,
then $P$ is also a monotone path set.
\end{theorem}
\proof
For each $j=1,\ldots,s$,
there exists a monotone passage
$\mphomj{j}:\Lj{j}\rightarrow\Pj{j}$
from $\xbarj{j-1}$ to $\xbarj{j}$,
for some linearly ordered set $\Lj{j}$
with order denoted $\leq^{j}$, and
with minimum $\lcminj{j}$ and maximum $\lcmaxj{j}$.
Without loss of generality, we assume $\Lj{j}$ is not a singleton.
(Otherwise, we could replace $\Lj{j}$ with $[0,1]$, say,
and $\mphomj{j}$ with a constant function on $[0,1]$ that maps all points
to the single point in $\Pj{j}$.)
To construct a monotone passage for $P$, we will paste together the
sets $\Lj{j}$ in a natural way.
We use distinguished elements $\nuj{0},\ldots,\nuj{s}$ to denote the
boundaries between where one of the linearly ordered sets $\Lj{j}$ ends
and the next begins.
More specifically,
with possible renaming of the elements of the sets $\Lj{j}$, we assume
henceforth that the following hold:
First, $\lcminj{j}=\nuj{j-1}$ and $\lcmaxj{j}=\nuj{j}$ for $j=1,\ldots,s$.
Thus, the set $\Lj{j}$ has minimum $\nuj{j-1}$ and maximum $\nuj{j}$,
implying that the sets intersect at these points
(so $\nuj{j-1}$ is also the maximum of $\Lj{j-1}$,
and $\nuj{j}$ is also the minimum of $\Lj{j+1}$).
Other than these points of intersection, these sets are
entirely disjoint.
Thus, for $1\leq j<k\leq s$,
\[
\Lj{j}\cap\Lj{k}
=
\left\{ \begin{array}{cl}
\{\nuj{j}\} & \mbox{if $k=j+1$} \\
\emptyset & \mbox{otherwise.}
\end{array} \right.
\]
(Concretely, this renaming can be achieved, for instance,
by replacing the minimum and
maximum in each $\Lj{j}$ with the ``new'' elements
$\nuj{j-1}$ and $\nuj{j}$,
as above,
and replacing every other element $\lambda\in\Lj{j}$
with a pair $\rpair{j}{\lambda}$.)
Let $L=\bigcup_{j=1}^s \Lj{j}$.
We define a natural order on $L$.
Let $\lambda,\mu\in L$, and suppose
$j$ and $k$ are the least indices in $\{1,\ldots,s\}$
for which
$\lambda\in\Lj{j}$ and $\mu\in\Lj{k}$.
Then we define $\lambda\leq\mu$ if
$j<k$ or if $j=k$ and $\lambda\leq^{j}\mu$.
It is straightforward to check that this is a linear order.
Its minimum and maximum are $\lambda_{\mathrm{min}}=\nuj{0}$
and $\lambda_{\mathrm{max}}=\nuj{s}$, respectively.
We claim $L$ is also a linear continuum.
To see that it is dense, let
$\lambda,\mu\in L$ with $\lambda<\mu$,
and let $j$ and $k$ be as above.
If $j=k$ then we must have $\lambda<^{j}\mu$, so there exists
$\eta\in\Lj{j}$ with $\lambda<^{j}\eta<^{j}\mu$ (since $\Lj{j}$ is
dense), implying $\lambda<\eta<\mu$.
Otherwise,
if $j<k$ then $\lambda\leq^{j}\lcmaxj{j}=\nuj{j}$,
so $\lambda\leq\nuj{j}$,
and $\nuj{j}<\mu$ (otherwise, we would have $k\leq j$).
If $k=j+1$, then $\nuj{j},\mu\in\Lj{j+1}$, so there exists
$\eta\in\Lj{j+1}$ with
$\lambda\leq\nuj{j}<\eta<\mu$.
And if $k>j+1$, then $\nuj{j+1}\leq\mu$ so we can choose
$\eta\in\Lj{j+1}$ with
$\lambda\leq\nuj{j}<\eta<\nuj{j+1}\leq\mu$.
To see that $L$ is complete, let $D\subseteq L$ be nonempty.
Let $k\in\{1,\ldots,s\}$ be the largest index for which
$D'=D\cap \Lj{k}$ is nonempty.
Since $\Lj{k}$ is complete,
$D'$ has a least upper bound $\mu$ in $\Lj{k}$.
Then $\mu$ is an upper bound on $D$ in the $\leq$ order
since if $\lambda\in D'$, then $\lambda\leq\mu$,
and if $\lambda\in D\setminus D'$ then $\lambda\in \Lj{j}$ for some
$j<k$, again implying $\lambda\leq\mu$.
Further, a smaller upper bound on $D$ in the order on $L$ would also
be a smaller upper bound on $D'$, contradicting that $\mu$ is the
least upper bound on $D'$ in the $\leq^{k}$ order on $\Lj{k}$.
Thus, $\mu$ is a least upper bound on $D$ (in the $\leq$ order).
The functions $\mphomj{j}$ can now be pasted together straightforwardly
into a piecewise, composite function $\xi:L\rightarrow P$ by
letting
$\xi(\lambda)=\mphomj{j}(\lambda)$
if $\lambda\in\Lj{j}$,
for $j=1,\ldots,s$.
Note that at the boundary points $\nuj{j}$, $\xi$ has been
``defined twice'' since $\nuj{j}$ is both in $\Lj{j}$ and $\Lj{j+1}$.
Nevertheless, the two definitions are consistent since in one
definition,
$\xi(\nuj{j})=\mphomj{j}(\nuj{j})=\mphomj{j}(\lcmaxj{j})=\xbarj{j}$,
and in the other,
$\xi(\nuj{j})=\mphomj{j+1}(\nuj{j})=\mphomj{j+1}(\lcminj{j+1})=\xbarj{j}$.
This also shows that
$\xi(\lambda_{\mathrm{min}})=\xi(\nuj{0})=\xbarj{0}=\overline{\xx}$
and
$\xi(\lambda_{\mathrm{max}})=\xi(\nuj{s})=\xbarj{s}=\overline{{\bf y}}$.
The function $\xi$ is surjective, since each $\mphomj{j}$ is
surjective
(and since $L$ and $P$ are the unions of their respective domains and
ranges).
Also, each separate piece defining $\xi$ is continuous, since each
$\mphomj{j}$ is continuous, and since, by construction,
the order topology on $\Lj{j}$
is identical to the subspace topology it inherits as a subspace of $L$.
Further, each piece is defined over a closed interval of $L$, namely,
$\{ \lambda\in L : \nuj{j-1} \leq \lambda \leq \nuj{j} \}$,
which therefore is closed in the order topology on $L$.
Thus, the piecewise function $\xi$ is continuous by application of
a standard pasting lemma
\citep[Theorem~18.3]{munkres}.
Finally, we claim that $\xi$ is order-preserving.
To show this, let ${\bf u}\in{\R^n}$ with $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$.
Let $\xi_{\uu}$ be as in \eqref{eqn:mphomu-def},
and let $\mphomuj{j}$ be defined analogously in terms of $\mphomj{j}$.
We aim to show
$\xi_{\uu}(\lambda)$
is nondecreasing as a function of $\lambda\in L$.
Because $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$,
\eqref{eqn:thm:combine-mono-pass:1} implies
$\xbarj{j-1}\cdot{\bf u}\leq\xbarj{j}\cdot{\bf u}$
for $j=1,\ldots,s$.
Since $\mphomj{j}$ is a monotone passage from
$\xbarj{j-1}$ to $\xbarj{j}$,
this implies,
by Proposition~\ref{pr:mono-pass-props}(\ref{pr:mono-pass-props:c}),
that
$\mphomuj{j}(\lambda)$
is nondecreasing as a function of $\lambda\in \Lj{j}$.
Therefore,
$\xi_{\uu}$ is nondecreasing on this same interval $\Lj{j}$, that is,
for all $\lambda\in L$ with $\nuj{j-1}\leq\lambda\leq\nu{j}$.
Because these overlapping intervals cover all of $L$,
it follows that $\xi_{\uu}$ is nondecreasing over its entire domain.
We conclude that $\xi$ is a monotone passage from $\overline{\xx}$ to
$\overline{{\bf y}}$, completing the first part of the theorem.
If each $\Pj{j}$ is a monotone path, then each $\Lj{j}$ can be chosen to
be a closed real interval $[a_j,b_j]$ for some $a_j,b_j\in{\mathbb{R}}$ with
$a_j\leq b_j$.
Without loss of generality, we can choose $\Lj{j}=[j-1,j]$
(by, if necessary, replacing $\mphomj{j}$ with
$\lambda\mapsto\mphomj{j}(a_j + (\lambda-j+1)(b_j-a_j))$).
Then, the assumed requirements of the construction above
are satisfied with $\nuj{j}=j$, for $j=1,\ldots,s$,
and $L=[0,s]$.
The order defined above on $L$ is then identical to the usual order on
intervals of ${\mathbb{R}}$.
Thus, the resulting function $\xi:L\rightarrow P$ is in fact a
monotone path from $\overline{\xx}$ to $\overline{{\bf y}}$.
\qed
\subsection{Monotone paths}
\label{sec:mono-paths}
In this section, we focus specifically on monotone paths.
We show first that the equivalence between
monotone passage sets and strict monotone passage sets given in
Corollary~\ref{cor:max-chain-iff-mono-pass}
carries over more specifically to monotone path sets:
\begin{theorem}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $P\subseteq\extspac{n}$.
Then
$P$ is a monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$
if and only if
$P$ is a strict monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$.
\end{theorem}
\proof
The ``if'' part is immediate.
For the converse, suppose
$P$ is a monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Then there exists a monotone path
$\xi:L\rightarrow P$
where $L=[a,b]$ for some $a,b\in{\mathbb{R}}$ with $a\leq b$.
Without loss of generality, we assume $L=[0,1]$ since otherwise we can
replace $\xi$ with
$\lambda\mapsto \xi((1-\lambda)a + \lambda b)$,
for $\lambda\in [0,1]$, which also is a monotone path.
If $P$ is a single point, then that single point must be $\overline{\xx}=\overline{{\bf y}}$,
so $\xi':\{0\}\rightarrow P$, with $\xi'(0)=\overline{\xx}=\overline{{\bf y}}$,
is trivially a strict monotone path from $\overline{\xx}$ to $\overline{{\bf y}}$.
Therefore, we assume henceforth that $P$ is not a singleton.
Throughout this proof,
we understand $P$ to be ordered by directional order relative to
$\overline{\xx},\overline{{\bf y}}$.
In this order,
by Proposition~\ref{pr:mono-pass-lin-cont},
$P$ is a linear continuum
with minimum $\overline{\xx}$ and maximum $\overline{{\bf y}}$.
Since $P$ is dense with at least two elements, it is
infinite in cardinality.
We show next that $P$ has a countable dense subset, that is, a
countable subset that is dense in $P$.
This, together with $P$'s other properties,
will allow us to use general results to show that $P$ is itself
homeomorphic with $[0,1]$.
Specifically,
let
\[
D = \{ \xi(\lambda) : \lambda\in {\mathbb{Q}}\cap [0,1] \},
\]
which is clearly countable, and which
we argue now is dense in $P$.
Suppose $\overline{{\bf w}},\overline{{\bf z}}\in P$ with $\overline{{\bf w}}\prec\overline{{\bf z}}$.
Then because $P$ is dense,
there must exist $\overline{{\bf w}}',\overline{{\bf z}}'\in P$ with
$\overline{{\bf w}}\prec\overline{{\bf w}}'\prec\overline{{\bf z}}'\prec\overline{{\bf z}}$
(since there must exist $\overline{{\bf w}}'$ between $\overline{{\bf w}}$ and $\overline{{\bf z}}$,
and then also $\overline{{\bf z}}'$ between $\overline{{\bf w}}'$ and $\overline{{\bf z}}$).
Let $\lambda\in\xi^{-1}(\overline{{\bf w}}')$ and $\mu\in\xi^{-1}(\overline{{\bf z}}')$
(which cannot be empty since $\xi$ is surjective).
Then $\lambda,\mu\in [0,1]$,
and $\lambda<\mu$ since $\xi$ is order preserving.
Thus, there exists a rational number $\nu\in{\mathbb{Q}}$ with
$\lambda<\nu<\mu$,
implying $\xi(\nu)\in D$, and also that
\[
\overline{{\bf w}}
\prec
\overline{{\bf w}}'
=
\xi(\lambda)
\preceq
\xi(\nu)
\preceq
\xi(\mu)
=
\overline{{\bf z}}'
\prec
\overline{{\bf z}},
\]
since $\xi$ is order preserving, proving the claim.
Let $M=P\setminus \{\overline{\xx},\overline{{\bf y}}\}$, that is, $P$ with its
minimum and maximum elements removed.
Then $M$ is linearly ordered, since $P$ is.
Moreover, $M\cap D$ is a countable dense subset in $M$
since if $\overline{{\bf z}},\overline{{\bf z}}'\in M$ with $\overline{{\bf z}}\prec\overline{{\bf z}}'$ then there exists
$\overline{{\bf w}}\in D$ with $\overline{\xx}\prec\overline{{\bf z}}\prec\overline{{\bf w}}\prec\overline{{\bf z}}'\prec\overline{{\bf y}}$,
so $\overline{{\bf w}}\in D\cap M$.
Also, $M$ is complete since if $Q\subseteq M$ is nonempty, so that it
includes some element $\overline{{\bf z}}\in M$, and
upper bounded by some $\overline{{\bf z}}'\in M$, then $Q$ must have a least upper
bound $\overline{{\bf w}}$ in $P$
with $\overline{\xx}\prec\overline{{\bf z}}\preceq\overline{{\bf w}}\preceq\overline{{\bf z}}'\prec\overline{{\bf y}}$
so that actually $\overline{{\bf w}}\in M$.
Finally, $M$ has no minimum.
This is because, for all $\overline{{\bf z}}\in M$,
there must exist $\overline{{\bf w}}\in P$ with $\overline{\xx}\prec\overline{{\bf w}}\prec\overline{{\bf z}}$
(since $P$ is dense);
that is, $\overline{{\bf w}}\in M$ and strictly less than $\overline{{\bf z}}$.
Likewise, $M$ has no maximum.
Summarizing, $M$ is a complete linear order
with a countable dense subset,
and with no minimum or maximum.
Together, these properties imply that $M$ is order-isomorphic to the
real numbers.
That is, there exists a (bijective) order isomorphism
$\rho:{\mathbb{R}}\rightarrow M$
\citep[Theorem~4.5.7]{hrbacek_jech99}.
To complete the construction,
we scale $\rho$'s domain to $[0,1]$ while also adding back the
minimum and maximum elements in a natural way yielding the function
$\xi':[0,1]\rightarrow P$ given by
\[
\xi'(\lambda)
=
\left\{ \begin{array}{cl}
\overline{\xx} & \mbox{if $\lambda=0$} \\
\rho(\sigma(\lambda))
& \mbox{if $\lambda\in (0,1)$} \\
\overline{{\bf y}} & \mbox{if $\lambda=1$}
\end{array} \right.
\]
for $\lambda\in [0,1]$, where $\sigma$ is any continuous,
strictly increasing
function mapping $(0,1)$ bijectively to ${\mathbb{R}}$
(such as $\lambda\mapsto \ln(\lambda/(1-\lambda))$).
Since $\rho$ is an order isomorphism,
$\xi'$ is as well.
As a result, $\xi'$ is a homeomorphism by
Proposition~\ref{pr:mphom-inv-is-cont}
(and since $P$ is closed in $\extspac{n}$,
by Proposition~\ref{pr:sbar-all-comp}, and therefore compact).
Thus, $\xi'$ is a strict monotone path from $\overline{\xx}$ to $\overline{{\bf y}}$.
\qed
We previously discussed that the segment joining two finite points
${\bf x},{\bf y}\in{\R^n}$ is a single monotone path set.
In fact, this is true also for the segment $\lb{{\bf y}}{\overline{\xx}}$
joining any finite point
${\bf y}\in{\R^n}$ and any (possibly infinite) astral point
$\overline{\xx}\in\extspac{n}$; that segment consists of a single monotone path
set.
When ${\bf y}={\bf 0}$
and $\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$,
for some ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in{\R^n}$,
that monotone path can nearly be ``read off'' from the form of
$\lb{{\bf 0}}{\overline{\xx}}$ given in Theorem~\ref{thm:lb-with-zero}.
Informally, such a path begins at the origin, then passes along
the ray $\{\lambda {\bf v}_1 : \lambda\geq 0\}$
to the astron $\limray{{\bf v}_1}$,
then continues on to
$\{\limray {\bf v}_1 \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda {\bf v}_2 : \lambda\geq 0\}$,
and so on,
in this way passing sequentially
along each ``piece'' of the segment,
$\{ \limrays{{\bf v}_1,\ldots,{\bf v}_{j-1}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf v}_j : \lambda\geq 0 \}$,
for $j=1,\ldots,k$,
finally passing along the final piece,
$\{\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf q} : \lambda\in [0,1]\}$.
In the next theorem,
we give a formal proof using the tools we have been developing.
\begin{theorem} \label{thm:y-xbar-is-path}
Let ${\bf y}\in{\R^n}$ and $\overline{\xx}\in\extspac{n}$.
Then $\lb{{\bf y}}{\overline{\xx}}$ is a monotone path set from ${\bf y}$ to $\overline{\xx}$,
and moreover is the only monotone passage set from ${\bf y}$ to $\overline{\xx}$.
\end{theorem}
\proof
For now, we assume ${\bf y}={\bf 0}$, returning later to the more general case.
We prove that $\lb{{\bf 0}}{\overline{\xx}}$ is a monotone path set by induction
on the astral rank of $\overline{\xx}$.
Specifically, we prove by induction on
$k=0,\ldots,n$ that
for all $\overline{\xx}\in\extspac{n}$, if $\overline{\xx}$ has astral rank $k$, then
$\lb{{\bf 0}}{\overline{\xx}}$ is a monotone path set
from ${\bf 0}$ to $\overline{\xx}$.
In the base case that $k=0$, we have $\overline{\xx}={\bf q}\in{\R^n}$, and we can
choose $\xi:[0,1]\rightarrow \lb{{\bf 0}}{{\bf q}}$
with $\xi(\lambda)=\lambda {\bf q}$ for $\lambda\in [0,1]$.
This function is continuous and surjective
(using Proposition~\ref{pr:e1}(\ref{pr:e1:a}))
with $\xi(0)={\bf 0}$ and $\xi(1)={\bf q}$.
It is also order-preserving with range relative to ${\bf 0},{\bf q}$,
since if ${\bf q}\cdot{\bf u}\geq {\bf 0}\cdot{\bf u}=0$ then
$\xi_{\uu}(\lambda)=\xi(\lambda)\cdot{\bf u}=\lambda {\bf q}\cdot{\bf u}$
is nondecreasing in $\lambda$.
Thus, $\xi$ is a monotone path from ${\bf 0}$ to ${\bf q}$.
For the inductive step, suppose $\overline{\xx}\in\extspac{n}$ has astral rank
$k>0$, and that the claim holds for $k-1$.
Let ${\bf v}\in{\R^n}$ be $\overline{\xx}$'s dominant direction.
By Proposition~\ref{pr:h:6},
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}$, where $\xbar^{\bot}$ is
$\overline{\xx}$'s projection perpendicular to ${\bf v}$, which is of astral rank
$k-1$.
To prove that $\lb{{\bf 0}}{\overline{\xx}}$ is a monotone path set, we will
cobble together a few other monotone paths.
First, let
\[
\Pj{1}
=
\lb{{\bf 0}}{\limray{{\bf v}}}
=
\{ \lambda {\bf v} : \lambda \geq 0 \}
\;\cup\;
\{ \limray{{\bf v}} \},
\]
where the second equality follows from
Theorem~\ref{thm:lb-with-zero}.
We claim $\Pj{1}$ is a monotone path set
from ${\bf 0}$ to $\limray{{\bf v}}$.
To show this, we
define $\mphomj{1}:[0,1]\rightarrow \Pj{1}$
as
\[
\mphomj{1}(\lambda) =
\left\{ \begin{array}{cl}
{\displaystyle \frac{\lambda}{1-\lambda}}
\, {\bf v} & \mbox{if $\lambda\in [0,1)$} \\[1.5em]
\limray{{\bf v}}
& \mbox{if $\lambda=1$,}
\end{array} \right.
\]
which we argue is a monotone path.
This function is clearly surjective, with $\Pj{1}(0)={\bf 0}$ and
$\Pj{1}(1)=\limray{{\bf v}}$.
It is also continuous since
$\limray{{\bf v}}=\lim_{\tau\rightarrow+\infty} \tau {\bf v}$.
To show it is order-preserving with range relative to
${\bf 0},\limray{{\bf v}}$, let ${\bf u}\in{\R^n}$ with
$0={\bf 0}\cdot{\bf u}\leq\limray{{\bf v}}\cdot{\bf u}$, which implies
${\bf v}\cdot{\bf u}\geq 0$.
We then need to show the function
$\mphomuj{1}(\lambda)=\mphomj{1}(\lambda)\cdot{\bf u}$
is nondecreasing in
$\lambda\in [0,1]$.
If ${\bf v}\cdot{\bf u}=0$, then $\mphomuj{1}\equiv 0$, which is nondecreasing.
Otherwise, if ${\bf v}\cdot{\bf u}>0$, then $\mphomuj{1}$
increases monotonically from $0$ to $+\infty$.
Thus, $\mphomj{1}$ is a monotone path from ${\bf 0}$ to $\limray{{\bf v}}$,
completing the claim.
Next, let
\[
\Pj{2}
=
{\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lb{{\bf 0}}{\xbar^{\bot}}}.
\]
By inductive hypothesis, $\lb{{\bf 0}}{\xbar^{\bot}}$ is a monotone path
set from ${\bf 0}$ to $\xbar^{\bot}$.
Therefore, by
Theorem~\ref{thm:mono-pass-affine-map}, applied to the affine map
$\overline{{\bf z}}\mapsto\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$,
$\Pj{2}$ is a monotone path set from $\limray{{\bf v}}$
to
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}=\overline{\xx}$.
Let $P = \Pj{1} \cup \Pj{2}$.
Then
\begin{equation} \label{eqn:thm:y-xbar-is-path:3}
P
=
\;
\{ \lambda {\bf v} : \lambda \geq 0 \}
\;\cup\;
\brackets{\limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lb{{\bf 0}}{\xbar^{\bot}}}
=
\lb{{\bf 0}}{\overline{\xx}}
\end{equation}
where the first equality is from definitions
(and since $\limray{{\bf v}}\in \Pj{2}$),
and the second equality is
by Lemma~\ref{lem:lb-with-zero}.
\eqref{eqn:thm:y-xbar-is-path:3} implies
$\limray{{\bf v}}\in\lb{{\bf 0}}{\overline{\xx}}$, so
${\bf 0}\preceq\limray{{\bf v}}\preceq\overline{\xx}$
relative to ${\bf 0},\overline{\xx}$,
by Proposition~\ref{pr:lb-def-by-part-order}.
We have argued that
$\Pj{1}$ is a monotone path set from ${\bf 0}$ to $\limray{{\bf v}}$
and
$\Pj{2}$ is a monotone path set from $\limray{{\bf v}}$ to $\overline{\xx}$.
Therefore, we can apply
Theorem~\ref{thm:combine-mono-pass} yielding that
$P=\Pj{1}\cup\Pj{2}$ is
a monotone path set from ${\bf 0}$ to $\overline{\xx}$.
This completes the induction and the proof when ${\bf y}={\bf 0}$.
For general ${\bf y}$, not necessarily ${\bf 0}$, we
can, as in the proof of Corollary~\ref{cor:lb-with-finite},
write $\lb{{\bf y}}{\overline{\xx}}$ as in \eqref{eqn:cor:lb-with-finite:1}.
By the foregoing,
$\lb{{\bf 0}}{\overline{\xx} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (-{\bf y})}$ is a monotone path set
from ${\bf 0}$ to $\overline{\xx} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (-{\bf y})$.
Therefore, $\lb{{\bf y}}{\overline{\xx}}$ is a monotone path set
from ${\bf y}$ to $\overline{\xx}$
by application of
Theorem~\ref{thm:mono-pass-affine-map}
(using the affine map $\overline{{\bf z}}\mapsto{\bf y}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$).
Finally, since the entire segment $\lb{{\bf y}}{\overline{\xx}}$ is a monotone
path set from ${\bf y}$ to $\overline{\xx}$, there can be no other
monotone passage set from ${\bf y}$ to $\overline{\xx}$, by
Corollary~\ref{cor:mon-pass-max-chain-conseq}(\ref{cor:mon-pass-max-chain-conseq:e}).
\qed
As consequence, we can show that if a
monotone passage set $P$ from $\overline{\xx}$ to $\overline{{\bf y}}$
includes some finite point ${\bf q}\in{\R^n}$, then
actually $P$ must be a monotone path set.
Furthermore, $P$ is uniquely determined, meaning there can
exist no other monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$ that
includes ${\bf q}$.
Thus, the only possible monotone passage sets that are not monotone
path sets are those that are entirely disjoint from ${\R^n}$.
\begin{corollary} \label{cor:mon-pass-thru-rn}
Let $\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
and
let $P\subseteq\extspac{n}$
be a monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Suppose ${\bf q}\in P\cap{\R^n}$.
Then $P$ is a monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$.
Furthermore,
\[
P = \lb{\overline{\xx}}{{\bf q}} \cup \lb{{\bf q}}{\overline{{\bf y}}}.
\]
\end{corollary}
\proof
Let $\Pj{1}$ be the subpassage of $P$ from $\overline{\xx}$ to ${\bf q}$,
and
let $\Pj{2}$ be the subpassage of $P$ from ${\bf q}$ to $\overline{{\bf y}}$,
implying $P=\Pj{1}\cup\Pj{2}$.
These are both monotone passage sets, by
Theorem~\ref{thm:subpassages}.
Furthermore, because ${\bf q}\in{\R^n}$, these monotone passage sets are
actually monotone path sets with
$\Pj{1} = \lb{\overline{\xx}}{{\bf q}}$
and
$\Pj{2} = \lb{{\bf q}}{\overline{{\bf y}}}$,
by
Theorem~\ref{thm:y-xbar-is-path}.
Since ${\bf q}\in P$,
$\overline{\xx}\preceq{\bf q}\preceq\overline{{\bf y}}$ relative to $\overline{\xx},\overline{{\bf y}}$.
Thus, by Theorem~\ref{thm:combine-mono-pass},
$P$ is a monotone path set from $\overline{\xx}$ to $\overline{{\bf y}}$.
\qed
For example,
let ${\bf q}\in{\R^n}$, and let
$P$ be any monotone passage set from
$-{\bf I}{\bar{\bm{\omega}}}$ to ${\bf I}{\bar{\bm{\omega}}}$ that includes ${\bf q}$
(which must exist since, as earlier argued,
$\lb{-{\bf I}{\bar{\bm{\omega}}}}{{\bf I}{\bar{\bm{\omega}}}}=\extspac{n}$).
Then Corollary~\ref{cor:mon-pass-thru-rn},
combined with
Corollary~\ref{cor:lb-with-finite},
implies that the intersection of $P$ with ${\R^n}$ is exactly a line
through ${\bf q}$ and
parallel to ${\bf e}_1$ (the first standard basis vector), that is,
\[
P \cap {\R^n}
=
\{ {\bf q} + \lambda {\bf e}_1 : \lambda\in{\mathbb{R}} \}.
\]
We have seen several cases so far in which monotone paths exist, for
instance, when one of the endpoints is in ${\R^n}$.
However, we will see next that monotone paths do not necessarily exist
between every pair of astral points.
Indeed,
the next theorem shows that no monotone path can exist between two
infinite points
$\overline{\xx},\overline{{\bf y}}\in\extspac{n}$,
except possibly if their dominant directions are either the same or
opposites of one another:
\begin{theorem} \label{thm:no-mono-path-diff-domdir}
Let $n\geq 2$, and
let $\overline{\xx}_1,\overline{\xx}_2\in\extspac{n}\setminus{\R^n}$.
For $j=1,2$,
let ${\bf w}_j\in{\R^n}$ be $\overline{\xx}_j$'s dominant direction
so that $\overline{\xx}_j=\limray{{\bf w}_j}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}_j$
for some $\overline{{\bf y}}_j\in\extspac{n}$.
Assume ${\bf w}_1\neq{\bf w}_2$ and ${\bf w}_1\neq -{\bf w}_2$.
Then no monotone path exists from $\overline{\xx}_1$ to $\overline{\xx}_2$.
\end{theorem}
\proof
Let us assume, for now, that $n=2$,
${\bf w}_1={\bf e}_1$ and ${\bf w}_2={\bf e}_2$ (where ${\bf e}_1$ and ${\bf e}_2$ are the
standard basis vectors in ${\mathbb{R}}^2$).
We prove the result first in this special case.
We then return to the fully general case, which we will see can be
reduced to the special case.
We first show that all of the points in $\lb{\overline{\xx}_1}{\overline{\xx}_2}$
are infinite, and all have
dominant directions in $\mathbb{R}_{\geq 0}^2$.
Since this claim will be re-used later, we state it as a lemma.
\begin{lemma} \label{lem:thm:no-mono-path-diff-domdir:1}
For $j=1,2$,
let $\overline{\xx}_j=\limray{{\bf e}_j}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}_j$
where $\overline{{\bf y}}_j\in\extspac{2}$, and where
${\bf e}_1,{\bf e}_2$ are standard basis vectors in ${\mathbb{R}}^2$.
Then
\[
\lb{\overline{\xx}_1}{\overline{\xx}_2}
\subseteq
\bigcup_{{\bf v}\in\mathbb{R}_{\geq 0}^2} [\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\extspac{2}].
\]
\end{lemma}
\proof
Let $S=\lb{\overline{\xx}_1}{\overline{\xx}_2}$.
Suppose, contrary to the claim, that some finite point ${\bf q}$ exists in
$S\cap{\mathbb{R}}^2$.
Let ${\bf u}=-{\bf e}_1-{\bf e}_2$.
Then $\overline{\xx}_j\cdot{\bf u}=-\infty$, for $j=1,2$,
but ${\bf q}\cdot{\bf u}\in{\mathbb{R}}$.
Since ${\bf q}\in \lb{\overline{\xx}_1}{\overline{\xx}_2}$,
this contradicts \eqref{eqn:lb-def}.
Thus, every point in $S$ is infinite.
Let $\overline{{\bf z}}\in S$, and let ${\bf v}\in{\R^n}$ be its dominant direction so
that $\overline{{\bf z}}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$, for some $\overline{{\bf y}}\in\extspac{2}$.
Suppose, again contrary to the claim, that ${\bf v}\not\in\mathbb{R}_{\geq 0}^2$ so that
either ${\bf v}\cdot{\bf e}_1<0$ or ${\bf v}\cdot{\bf e}_2<0$.
Without loss of generality, assume the former.
Since $\norm{{\bf v}}=1$, this implies ${\bf v}\cdot{\bf e}_2<1$.
Letting ${\bf u}={\bf v}-{\bf e}_2$, it follows that
${\bf e}_1\cdot{\bf u}<0$,
${\bf e}_2\cdot{\bf u}<0$,
and
${\bf v}\cdot{\bf u}>0$.
Therefore,
$\overline{\xx}_j\cdot{\bf u}=-\infty$, for $j=1,2$,
but
$\overline{{\bf z}}\cdot{\bf u}=+\infty$.
Since $\overline{{\bf z}}\in\lb{\overline{\xx}_1}{\overline{\xx}_2}$,
this again contradicts \eqref{eqn:lb-def}.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Let $C=\{ {\bf v}\in\mathbb{R}_{> 0}^2 : \norm{{\bf v}}=1 \}$
be the quarter unit circle in the plane from ${\bf e}_1$ to ${\bf e}_2$ (excluding
the endpoints).
We claim next that every astron associated with points in $C$ must be
included in every monotone passage set from $\overline{\xx}_1$ to $\overline{\xx}_2$.
\setcounter{claimp}{0}
\medskip
\begin{claimp} \label{cl:thm:no-mono-path-diff-domdir:2}
Let $\xi:L\rightarrow P$
be a monotone passage from $\overline{\xx}_1$ to $\overline{\xx}_2$,
and let ${\bf v}\in C$.
Then $\limray{{\bf v}}\in P$.
\end{claimp}
\proof
We can write ${\bf v}$ in terms of its components as
${\bf v}=\trans{[v_1,v_2]}$, with $v_1,v_2>0$.
Let ${\bf u}=\trans{[-v_2, v_1]}$, which is perpendicular to ${\bf v}$.
Note that
$\overline{\xx}_1\cdot{\bf u}=\limray{{\bf e}_1}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}_1\cdot{\bf u}=-\infty$,
and similarly, $\overline{\xx}_2\cdot{\bf u}=+\infty$.
Therefore, by
Proposition~\ref{pr:mono-pass-props}(\ref{pr:mono-pass-props:c}),
there must be some point $\overline{{\bf z}}\in P$ with $\overline{{\bf z}}\cdot{\bf u}=0$.
We claim that this point $\overline{{\bf z}}$ must actually be $\limray{{\bf v}}$.
Let
$\overline{{\bf z}}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
be $\overline{{\bf z}}$'s canonical form.
By Lemma~\ref{lem:thm:no-mono-path-diff-domdir:1},
$\overline{{\bf z}}\not\in{\R^n}$ (since $P\subseteq \lb{\overline{\xx}_1}{\overline{\xx}_2}$,
by Proposition~\ref{pr:mono-pass-lin-cont});
therefore, $k\geq 1$.
Since $\overline{{\bf z}}\cdot{\bf u}=0$,
by Proposition~\ref{pr:vtransu-zero},
we must have
${\bf v}_i\cdot{\bf u}=0$, for $i=1,\ldots,k$,
and so also ${\bf q}\cdot{\bf u}=0$.
In ${\mathbb{R}}^2$,
by ${\bf u}$'s definition, these imply ${\bf q}$ and each ${\bf v}_i$ are scalar
multiples of ${\bf v}$;
thus, ${\bf q}=\beta{\bf v}$, for some $\beta\in{\mathbb{R}}$, and
each ${\bf v}_i$, being a unit vector, is either equal to ${\bf v}$ or $-{\bf v}$.
Since the ${\bf v}_i$'s are orthonormal, this implies $k=1$, and since
${\bf v}_1\cdot{\bf q}=0$, this further implies ${\bf q}={\bf 0}$.
Thus, $\overline{{\bf z}}$ is either equal to $\limray{{\bf v}}$ or $\limray{(-{\bf v})}$.
However, by
Lemma~\ref{lem:thm:no-mono-path-diff-domdir:1},
${\bf v}_1\in\mathbb{R}_{\geq 0}^2$, ruling out the latter possibility.
Therefore, $\overline{{\bf z}}=\limray{{\bf v}}$, proving the claim.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Suppose now, by way of contradiction, that there exists a monotone
path $\xi:L\rightarrow P$ from $\overline{\xx}_1$ to $\overline{\xx}_2$,
where $L$ is a closed real interval that, without loss of generality,
we take to be $[0,1]$.
As shown in
Theorem~\ref{thm:formerly-lem:h:1}
(with $n=2$),
for each ${\bf v}\in{\mathbb{R}}^2$, there exists an open set
${U_{\vv}}\subseteq\extspac{2}$ that includes
the astron $\limray{{\bf v}}$, but no other icons;
thus,
${U_{\vv}}\subseteq{\mathbb{R}}^2\cup [\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\mathbb{R}}^2]$.
For each ${\bf v}$ in the quarter unit circle $C$,
let ${U'_{\vv}}={U_{\vv}}\cap P$, which is open in the subspace topology of $P$,
and includes $\limray{{\bf v}}$
by claim~\ref{cl:thm:no-mono-path-diff-domdir:2}.
Because
$P\cap{\mathbb{R}}^2=\emptyset$
(by Lemma~\ref{lem:thm:no-mono-path-diff-domdir:1}),
every point ${U'_{\vv}}$ is infinite with iconic part $\limray{{\bf v}}$;
therefore, the sets ${U'_{\vv}}$, for ${\bf v}\in C$,
are disjoint from one another.
Thus, because $\xi$ is continuous and surjective, the pre-images
$\xi^{-1}({U'_{\vv}})$ are open and nonempty subsets of $[0,1]$.
Therefore, for each ${\bf v}\in C$, we can select a rational number
$r({\bf v})$ in ${\mathbb{Q}}\cap\xi^{-1}({U'_{\vv}})$.
The resulting function $r:C\rightarrow{\mathbb{Q}}$ is injective since
the sets $\xi^{-1}({U'_{\vv}})$, like the sets ${U'_{\vv}}$,
are disjoint from one another over ${\bf v}\in C$.
However, this is a contradiction since ${\mathbb{Q}}$ is countable, but $C$
is uncountable, so no such function can exist.
Having proved the theorem when $n=2$, ${\bf w}_1={\bf e}_1$ and ${\bf w}_2={\bf e}_2$,
we return now to the general case as in the theorem's statement.
Let ${\bf W}$ be the $n\times 2$ matrix ${\bf W}=[{\bf w}_1, {\bf w}_2]$.
Then $\trans{{\bf W}}{\bf W}$ is invertible since its determinant is
$1-({\bf w}_1\cdot{\bf w}_2)^2>0$, since ${\bf w}_1\neq{\bf w}_2$ and
${\bf w}_1\neq -{\bf w}_2$.
Let $\pseudinv{{\bf W}}=(\trans{{\bf W}}{\bf W})^{-1} \trans{{\bf W}}$
be ${\bf W}$'s \emph{pseudoinverse}.
Then $\pseudinv{{\bf W}}{\bf W}={\bf I}$, so that $\pseudinv{{\bf W}}{\bf w}_j={\bf e}_j$,
for $j=1,2$, where ${\bf e}_1,{\bf e}_2$, as before, are the standard basis
vectors in ${\mathbb{R}}^2$, and ${\bf I}$ is the $2\times 2$ identity matrix.
Let $F:\extspac{n}\rightarrow\extspac{2}$ be the function
$F(\overline{{\bf z}})=\pseudinv{{\bf W}} \overline{{\bf z}}$.
Suppose a monotone path set $P$ exists from $\overline{\xx}_1$ to $\overline{\xx}_2$.
Then, by Theorem~\ref{thm:mono-pass-affine-map},
$F(P)$ is a monotone path set from
$F(\overline{\xx}_1)=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\pseudinv{{\bf W}}\overline{{\bf y}}_1$
to
$F(\overline{\xx}_2)=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\pseudinv{{\bf W}}\overline{{\bf y}}_2$,
that is, in the special case considered above and shown not to be
possible.
Thus, in the general case as well,
no monotone path from $\overline{\xx}_1$ to $\overline{\xx}_2$ can exist.
\qed
Thus, no monotone path can exist between points satisfying the
conditions of Theorem~\ref{thm:no-mono-path-diff-domdir}.
On the other hand, we know from
Corollary~\ref{cor:mon-pass-max-chain-conseq}(\ref{cor:mon-pass-max-chain-conseq:a})
that there must nevertheless be a monotone passage connecting the two
points.
What then do such monotone passages look like?
For example, in $\extspac{2}$, there must exist a monotone passage
$\xi:L\rightarrow P$ from $\limray{{\bf e}_1}$ to $\limray{{\bf e}_2}$.
By claim~\ref{cl:thm:no-mono-path-diff-domdir:2} in the proof of
Theorem~\ref{thm:no-mono-path-diff-domdir}, $P$ must include every
astron $\limray{{\bf v}}$ for all ${\bf v}\in\mathbb{R}_{> 0}^2$.
As a first attempt at constructing such a monotone passage,
we might therefore consider simply running through all
these astrons.
More exactly, let
\begin{equation} \label{eqn:circf-def}
\circf_{\cfarga} = \left[\begin{array}{c}
\cos(\alpha \pi /2) \\
\sin(\alpha \pi /2)
\end{array}
\right]
\end{equation}
for $\alpha\in [0,1]$, which parameterizes all of the
points on the quarter unit circle in ${\mathbb{R}}^2$ from ${\bf e}_1$ to ${\bf e}_2$
(including the endpoints).
We might then try to construct a monotone passage
$\xi(\alpha)=\limray(\circf_{\cfarga})$,
for $\alpha\in [0,1]$,
which traces through all of the associated astrons.
Of course, this cannot actually
be a monotone passage from $\limray{{\bf e}_1}$ to
$\limray{{\bf e}_2}$ since we know from
Theorem~\ref{thm:no-mono-path-diff-domdir}
that no such monotone \emph{path} can exist.
The particular problem with this function is that it is not
continuous, and in fact, is discontinuous at \emph{every} point
(since each point $\limray{{\bf v}}$ in the range of $\xi$
is in its own open set ${U_{\vv}}$, as was
used in the proof of
Theorem~\ref{thm:no-mono-path-diff-domdir}, so that the pre-image of
that open set is a singleton in $[0,1]$, which is not open).
Thus,
a monotone passage from $\limray{{\bf e}_1}$ to $\limray{{\bf e}_2}$ must
include every astron, as above, but needs to also be
continuous by somehow managing to move
more smoothly from one astron to the next.
We show in the next proposition an example of how this can be done.
To slightly simplify the presentation, we consider finding a
monotone passage from
$\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{(-{\bf e}_2)}$
to
$\overline{{\bf y}}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{(-{\bf e}_1)}$.
(If desired,
a monotone passage from $\limray{{\bf e}_1}$ to $\limray{{\bf e}_2}$ can then
be extracted as the subpassage from $\limray{{\bf e}_1}$ to
$\limray{{\bf e}_2}$ using Theorem~\ref{thm:subpassages}.)
For each $\alpha\in [0,1]$,
let $\circf_{\cfarga}$ be as in \eqref{eqn:circf-def},
and let $\circfinv_{\cfarga}$ be a corresponding orthogonal unit
vector, namely,
\begin{equation} \label{eqn:circfinv-def}
\circfinv_{\cfarga}
=
\left[\begin{array}{r}
- \sin(\alpha \pi /2) \\
\cos(\alpha \pi /2)
\end{array}
\right].
\end{equation}
The monotone passage that we construct includes not only each astron
$\limray{\circf_{\cfarga}}$ as in the (unsuccessful) attempt above,
but also passes along a path through its entire galaxy (and its
closure), from
$\limray{\circf_{\cfarga}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{(-\circfinv_{\cfarga})}$
at one end to
$\limray{\circf_{\cfarga}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{\circfinv_{\cfarga}}$
at the other.
Such a path is followed separately for every
$\alpha\in [0,1]$; in this sense, the passage is,
informally, a ``path of paths.''
As such, we define it over a Cartesian product of closed real
intervals, specifically,
$L=[0,1]\times[0,1]$, with linear order over pairs
$\rpair{\alpha}{\beta}\in L$ defined
\emph{lexicographically}, meaning, in this case, that
$\rpair{\alpha}{\beta}\leq\rpair{\alpha'}{\beta'}$
if
$\alpha<\alpha'$ or
if $\alpha=\alpha'$ and $\beta\leq\beta'$.
\begin{proposition}
Let
$\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{(-{\bf e}_2)}$
and
$\overline{{\bf y}}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{(-{\bf e}_1)}$,
where ${\bf e}_1,{\bf e}_2$ are the standard basis vectors in ${\mathbb{R}}^2$.
Let
\[ P = \bigcup_{{\bf v}\in \mathbb{R}_{\geq 0}^2} [\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \extspac{2}], \]
and let
$L=[0,1]\times[0,1]$ be linearly ordered lexicographically.
Define $\xi:L\rightarrow P$ by
\[
\xi(\alpha,\beta)
=
\limray{\circf_{\cfarga}}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\left\{ \begin{array}{cl}
\limray{(-\circfinv_{\cfarga})} & \mbox{if $\beta=0$} \\
\sigma(\beta)\, \circfinv_{\cfarga} & \mbox{if $\beta\in (0,1)$} \\
\limray{\circfinv_{\cfarga}} & \mbox{if $\beta=1$,}
\end{array} \right.
\]
where $\circf_{\cfarga}$ and $\circfinv_{\cfarga}$ are as defined
in \eqref{eqn:circf-def} and \eqref{eqn:circfinv-def},
and where $\sigma(\beta)=\ln(\beta/(1-\beta))$
for $\sigma\in (0,1)$.
Then $\xi$ is a strict monotone passage from $\overline{\xx}$ to $\overline{{\bf y}}$.
Furthermore, $P=\lb{\overline{\xx}}{\overline{{\bf y}}}$, implying $P$ is the only
monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$.
\end{proposition}
\proof
We argue first that $\xi$ is bijective.
It can be checked that
the function $\sigma$ is strictly increasing
and
maps $(0,1)$ bijectively to ${\mathbb{R}}$.
Note that $\alpha\mapsto\limray{\circf_{\cfarga}}$ maps
$[0,1]$ bijectively
to the set of all astrons $\limray{{\bf v}}$ with ${\bf v}\in\mathbb{R}_{\geq 0}^2$.
For each $\alpha\in [0,1]$,
$\beta\mapsto\xi(\alpha,\beta)$ maps bijectively from
$[0,1]$ to
$\galcl{\limray{\circf_{\cfarga}}}=\limray{\circf_{\cfarga}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\extspac{2}$.
This can be seen by noting that every point
$\xi(\alpha,\beta)$, as defined in the proposition's statement, is
already in its unique canonical representation, which also makes it
straightforward to check that the mapping is surjective
onto all of $\galcl{\limray{\circf_{\cfarga}}}$
(since only two orthogonal directions are possible in $\extspac{2}$).
Thus, $\xi$ is a bijection from $L$ to $P$.
We next show that $\xi$ is order-preserving
with range relative to $\overline{\xx},\overline{{\bf y}}$.
Let ${\bf u}\in{\mathbb{R}}^2$ with $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$.
We aim to show that the function
$\xi_{\uu}(\alpha,\beta)=\xi(\alpha,\beta)\cdot{\bf u}$
is nondecreasing as a function of
$\rpair{\alpha}{\beta}\in L$
(ordered lexicographically).
We show this in cases, based on the signs of the two components of
${\bf u}=\trans{[u_1,u_2]}$.
If ${\bf u}={\bf 0}$, then $\xi_{\uu}\equiv 0$.
If $u_1>0$ and $u_2>0$, then
$\circf_{\cfarga}\cdot{\bf u}>0$ for all $\alpha\in [0,1]$,
so $\xi_{\uu}\equiv +\infty$.
Similarly, if $u_1<0$ and $u_2<0$, then
$\xi_{\uu}\equiv -\infty$.
If $u_1\geq 0$ and $u_2\leq 0$, but ${\bf u}\neq{\bf 0}$,
then
$\overline{\xx}\cdot{\bf u}=+\infty$ and $\overline{{\bf y}}\cdot{\bf u}=-\infty$,
contradicting that $\overline{\xx}\cdot{\bf u}\leq\overline{{\bf y}}\cdot{\bf u}$, so this case is
impossible.
We are left only with the case that
$u_1\leq 0$ and $u_2\geq 0$, but ${\bf u}\neq{\bf 0}$.
By normalizing ${\bf u}$ (which only changes $\xi_{\uu}$ by a positive
scalar), we can assume $\norm{{\bf u}}=1$.
Therefore, for some $\alpha_0\in[0,1]$, we must have
$\circfinv_{\cfarga_0}={\bf u}$.
It can be checked that, under our assumptions on ${\bf u}$,
\[
\alpha
\mapsto
\circf_{\cfarga}\cdot{\bf u}
=
u_1 \cos(\alpha \pi /2)
+
u_2 \sin(\alpha \pi /2)
\]
is strictly increasing as a function of $\alpha\in [0,1]$,
and is equal to zero if and only if $\alpha=\alpha_0$
(so that $\circf_{\cfarga_0}\cdot{\bf u}=\circf_{\cfarga_0}\cdot\circfinv_{\cfarga_0}=0$).
Therefore, we can compute $\xi_{\uu}$ explicitly as
\[
\xi_{\uu}(\alpha,\beta)
=
\left\{ \begin{array}{cl}
-\infty & \mbox{if $\alpha<\alpha_0$} \\
-\infty & \mbox{if $\alpha=\alpha_0$ and $\beta=0$} \\
\sigma(\beta) & \mbox{if $\alpha=\alpha_0$ and $\beta\in (0,1)$} \\
+\infty & \mbox{if $\alpha=\alpha_0$ and $\beta=1$} \\
+\infty & \mbox{if $\alpha>\alpha_0$,} \\
\end{array} \right.
\]
which is evidently nondecreasing in the lexicographic ordering of $L$.
Thus, $\xi$ is bijective and
order-preserving (with range relative to $\overline{\xx},\overline{{\bf y}}$).
It is known that $L$ is a linear continuum
\citep[Example~24.1]{munkres}.
Its minimum and maximum elements,
$\lambda_{\mathrm{min}}=\rpair{0}{0}$ and $\lambda_{\mathrm{max}}=\rpair{1}{1}$,
map to
$\xi(\lambda_{\mathrm{min}})=\overline{\xx}$
and
$\xi(\lambda_{\mathrm{max}})=\overline{{\bf y}}$.
By Proposition~\ref{pr:lb-def-by-part-order},
this implies that
$P=\xi(L)$ is included in $\lb{\overline{\xx}}{\overline{{\bf y}}}$.
On the other hand, $\lb{\overline{\xx}}{\overline{{\bf y}}}\subseteq P$
by Lemma~\ref{lem:thm:no-mono-path-diff-domdir:1}.
Thus, $P=\lb{\overline{\xx}}{\overline{{\bf y}}}$, which is closed in $\extspac{2}$
(Proposition~\ref{pr:e1}(\ref{pr:e1:d})),
and therefore compact.
Together, these facts imply,
by Proposition~\ref{pr:mphom-inv-is-cont},
that $\xi$ is a homeomorphism
and also is an
order isomorphism with range relative to $\overline{\xx},\overline{{\bf y}}$.
Thus, $\xi$ is a strict monotone passage.
That $P$ is the only monotone passage set from $\overline{\xx}$ to $\overline{{\bf y}}$ then
follows from
Corollary~\ref{cor:mon-pass-max-chain-conseq}(\ref{cor:mon-pass-max-chain-conseq:e}).
\qed
\section{Convex cones}
\label{sec:cones}
The standard notion of a cone extends immediately to astral space.
Cones will play an important role in studying properties of functions,
including how they are minimized and where they are continuous.
\subsection{Definition and basic properties}
\label{sec:cones-basic}
As in standard convex analysis, we say that a set in $\extspac{n}$
is a \emph{cone} if it is closed under
multiplication by any positive scalar; thus, $S\subseteq\extspac{n}$ is
a cone if $\lambda\overline{\xx}\in S$ for all $\overline{\xx}\in S$ and all
positive scalars $\lambda>0$.
In general, cones need not be convex, nor closed, nor must they
include the origin, even if closed;
in fact, an astral cone can be entirely disjoint
from ${\R^n}$.
Examples will be given shortly.
The next proposition provides several conditions that are sufficient
for a set to be a cone.
\begin{proposition} \label{pr:new:a:2}
~
\begin{enumerate}
\item \label{pr:new:a:2:a}
The intersection of an arbitrary collection of cones is a cone.
\item \label{pr:new:a:2:b}
The union of an arbitrary collection of cones is a cone.
\item \label{pr:new:a:2:c}
{
Every set of icons, that is, every subset of ${{\cal E}_n}$, is a cone.
}
\item \label{pr:new:a:2:d}
If $S\subseteq\extspac{n}$ is a cone, then
$S$'s convex hull, $\conv{S}$, is also a cone.
\item \label{pr:new:a:2:e}
If $S\subseteq\extspac{n}$ is a cone, then
$S$'s closure, $\overline{S}$, is also a cone.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:new:a:2:a}):
Let
\[ M = \bigcap_{\alpha\in{\cal I}} S_\alpha \]
where each $S_\alpha\subseteq\extspac{n}$ is a cone, and ${\cal I}$ is an
arbitrary index set.
If $\overline{\xx}\in M$ and $\lambda>0$, then for all $\alpha\in{\cal I}$,
$\overline{\xx}$ is in the cone $S_\alpha$ so $\lambda\overline{\xx}$ is also in $S_\alpha$,
and therefore in $M$.
Thus, $M$ is a cone.
Part~(\ref{pr:new:a:2:b}) is similar with $M$ replaced by a union
rather than intersection, and ``for all'' replaced by ``for some.''
Part~(\ref{pr:new:a:2:c}):
Let $S\subseteq{{\cal E}_n}$.
If $\overline{\xx}\in S$ and $\lambda>0$ then $\lambda\overline{\xx}\in S$ since
$\lambda\overline{\xx}=\overline{\xx}$ by Proposition~\ref{pr:i:8}(\ref{pr:i:8d}).
Part~(\ref{pr:new:a:2:d}):
Let $S$ be a cone.
Let $\lambda>0$ and
let $G:\extspac{n}\rightarrow\extspac{n}$ be the linear map associated
with multiplication by $\lambda$ so that
$G(\overline{\xx})=\lambda\overline{\xx}=(\lambda{\bf I})\overline{\xx}$
for all $\overline{\xx}\in\extspac{n}$, where ${\bf I}$ is the $n\times n$ identity matrix.
Let $\overline{\xx}$ be any point in $\conv{S}$.
Then
\[
\lambda\overline{\xx} = G(\overline{\xx})
\in
G(\conv{S})
=
\conv{G(S)}
\subseteq
\conv{S}.
\]
The second equality is by Corollary~\ref{cor:thm:e:9b}.
The last inclusion is by
Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})
since $\conv{S}$ is convex, and because
$G(S)\subseteq S\subseteq \conv{S}$
as a result of $S$ being a cone.
Thus, $\conv{S}$ is a cone.
Part~(\ref{pr:new:a:2:e}):
Let $S$ be a cone.
Suppose $\overline{\xx}\in\overline{S}$.
Then there exists a sequence $\seq{\overline{\xx}_t}$ in $S$ that converges to
$\overline{\xx}$.
Let $\lambda>0$ and let $G$ be as above.
By Proposition~\ref{pr:aff-cont}, $G$ is
continuous, so
$\lambda\overline{\xx}_t=G(\overline{\xx}_t) \rightarrow G(\overline{\xx})= \lambda\overline{\xx}$.
Since $S$ is a cone, $\lambda\overline{\xx}_t\in S$ for all $t$,
so $\lambda\overline{\xx}\in\overline{S}$.
Therefore, $\overline{S}$ is a cone.
\qed
As noted in Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:c}),
\emph{any} set of icons is a cone.
Thus, whereas in ${\R^n}$, the only finite cone is $\{{\bf 0}\}$, in the
astral space, there are a rich variety of cones that might be finite
or discrete.
Furthermore, the convex hull of sets of icons
(more specifically, of sets of astrons) will be a central element in
characterizing the continuity of astral functions.
By Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:d}),
such a set is necessarily a convex cone.
In general, as we show next, any closed convex cone in $\extspac{n}$ is
exactly equal to the convex hull of all its icons:
\begin{theorem} \label{thm:cl-conv-cone-convhull-corezeros}
Let $S$ be a closed convex cone in $\extspac{n}$.
Then $S$ is equal to the convex hull of all of its icons;
that is,
\[
S = \conv(S\cap{{\cal E}_n}).
\]
\end{theorem}
\proof
Since $S\cap{{\cal E}_n}\subseteq S$, and since $S$ is convex,
$\conv(S\cap{{\cal E}_n})\subseteq S$
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa}).
To prove the reverse inclusion, let $\overline{\xx}\in S$.
Then we can write $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where $\overline{\ee}\in{{\cal E}_n}$ is
an icon and ${\bf q}\in{\R^n}$
(Proposition~\ref{pr:i:8}(\ref{pr:i:8c})).
Since $S$ is a cone, the point
$t \overline{\xx} = \overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax t{\bf q}$
is also in $S$, for $t=1,2,\ldots$.
Since this sequence converges to $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}$, this shows
that $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}\in S$, since $S$ is closed.
(Convergence follows from $t{\bf q}\rightarrow\limray{{\bf q}}$ combined with
the continuity of the affine map
$\overline{{\bf z}}\mapsto\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$,
proved in Proposition~\ref{pr:aff-cont}.)
By similar arguments,
\[ \frac{1}{t} \overline{\xx} = \overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\frac{{\bf q}}{t} \]
is in $S$, for $t=1,2,\ldots$.
Since $(1/t)\overline{\xx} \rightarrow \overline{\ee}$, it follows that
$\overline{\ee}$ is in $S$ as well.
Thus, $\overline{\ee}$ and $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}$ are icons and in
$S$, and so are in $S\cap{{\cal E}_n}$.
By Theorem~\ref{thm:lb-with-zero}, ${\bf q}\in\lb{{\bf 0}}{\limray{{\bf q}}}$.
Further, considering the affine map $\overline{{\bf z}}\mapsto\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$,
we see that
\[
\overline{\xx}
=
\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}
\in
\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lb{{\bf 0}}{\limray{{\bf q}}}
=
\lb{\overline{\ee}}{\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}},
\]
by Theorem~\ref{thm:e:9}.
Since $\overline{\ee}$ and $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}$ are in
$S\cap{{\cal E}_n}$, and so in the convex set $\conv(S\cap{{\cal E}_n})$ as
well, it follows that $\overline{\xx}$, which is on the segment between those
points, is also in that set.
Therefore, $S\subseteq\conv(S\cap{{\cal E}_n})$.
\qed
For any $S\subseteq\extspac{n}$, we can generate
the {smallest cone containing $S$} by multiplying all elements of $S$
by all positive scalars, yielding the set
\[
\scone{S} =
\braces{ \lambda\overline{\xx} : \overline{\xx}\in S, \lambda > 0 },
\]
where $\mathbb{R}_{> 0}=(0,+\infty)$ is the set of all positive reals.
The set $\scone{S}$
is clearly a cone, and any cone containing $S$ must also
contain all positive-scalar multiples of the points in $S$, so any cone
containing $S$ must include $\scone{S}$.
Thus, $\scone{S}$ is the intersection of all cones that contain $S$.
Similarly,
for any $S\subseteq\extspac{n}$,
the {smallest convex cone containing $S$}, written $\scc{S}$,
is the intersection of all convex cones that contain $S$,
which must itself be a convex cone
(by
Propositions~\ref{pr:e1}(\ref{pr:e1:b})
and~\ref{pr:new:a:2}(\ref{pr:new:a:2:a})).
We call $\scc{S}$ the \emph{strict conic hull of $S$}, or the
\emph{convex cone strictly generated by $S$}.
For consistency with what is standard in ${\R^n}$ and what had previously
been introduced,
we also (re)define
the \emph{conic hull of $S$},
also called
the \emph{convex cone generated by $S$},
and denoted $\cone S$, to be the smallest convex cone containing $S$
\emph{as well as} the origin.
Thus,
\begin{equation} \label{eq:cone-def}
\cone S = \scc(S\cup \{{\bf 0}\}).
\end{equation}
When $S\subseteq{\R^n}$,
this definition and notation for $\cone S$
are entirely consistent with the standard ones discussed in
Section~\ref{sec:shadow}.
In this case, it can be shown that
$\cone S = (\scc{S}) \cup \{{\bf 0}\}$,
and that $\cone S$ is equal to the set of all nonnegatively weighted
combinations of zero or more elements of $S$.
By comparison,
the \emph{strict} conic hull, $\scc{S}$, for $S\subseteq{\R^n}$,
is equal to the set of all \emph{strictly positively}
weighted combinations of \emph{one or more} elements of $S$
\cite[Corollary~2.6.2]{ROC}.
Here are some simple properties:
\begin{proposition} \label{pr:cone-props}
~
\begin{enumerate}
\item \label{pr:cone-props:a}
If $S\subseteq U\subseteq\extspac{n}$
and $U$ is a convex cone, then $\scc{S}\subseteq U$.
\item \label{pr:cone-props:b}
If $S\subseteq U\subseteq\extspac{n}$ then
$\scc{S}\subseteq\scc{U}$.
\item \label{pr:cone-props:c}
For all $S\subseteq\extspac{n}$,
$\scc{S}=\conv(\scone{S})$.
\item \label{pr:cone-props:d}
If $S\subseteq\extspac{n}$ is a cone, then
$\scc{S}=\conv{S}$ and $\cone S=\conv(S\cup\{{\bf 0}\})$.
\end{enumerate}
\end{proposition}
\proof ~
Part~(\ref{pr:cone-props:a}) follows from definitions.
Part~(\ref{pr:cone-props:b}):
Proof is similar to that of
Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:b}).
Part~(\ref{pr:cone-props:c}):
The set $\conv(\scone{S})$ is a convex cone by
Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:d}).
Therefore,
$\scc{S}\subseteq\conv(\scone{S})$
by part~(\ref{pr:cone-props:a}).
On the other hand, $\scc{S}$ is a cone that includes $S$, and
so must include $\scone{S}$.
Therefore, since $\scc{S}$ is convex,
$\conv(\scone{S})\subseteq\scc{S}$
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa}).
Part~(\ref{pr:cone-props:d}):
Let $S$ be a cone.
Then $\scone{S}=S$, so
$\scc{S}=\conv{S}$ by part~(\ref{pr:cone-props:c}).
Applied to $S\cup\{{\bf 0}\}$, which is a cone by
Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:b}),
yields
$\cone S =\scc(S\cup\{{\bf 0}\})=\conv(S\cup\{{\bf 0}\})$.
\qed
\subsection{Convex cones generated by astrons}
\label{sec:cones:astrons}
As already noted, any set of icons forms a cone whose
convex hull is a convex cone (by
Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:c},~\ref{pr:new:a:2:d})).
In particular, this is true if, for any set $S\subseteq{\R^n}$, we form
the cone of all astrons associated with the points in $S$,
written $\lmset{S}$:
\[
\lmset{S} = \{\limray{{\bf v}} : {\bf v}\in S\}.
\]
We focus in this subsection on the convex hull of such a set,
$\conv(\lmset{S})$, a convex cone that will play a central role in
Section~\ref{sec:continuity} in precisely identifying exactly
where the lower semicontinuous extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of a convex function
$f:{\R^n}\rightarrow\overline{\R}$ is continuous.
We will give exact expressions for this set, and will also give
necessary and sufficient conditions for it to be closed, which will be
a key step in later characterizing precisely the continuity of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
To motivate intuition, let us suppose for the moment that the origin
is in $S$ (although the general results below do not require
this condition).
In this case, for any ${\bf v}\in S$, where ${\bf v}\neq{\bf 0}$, $\lmset{S}$
includes both ${\bf 0}$ and $\limray{{\bf v}}$, which means
the convex hull of $\lmset{S}$ must include the entire segment
joining these two points, $\lb{{\bf 0}}{\limray{{\bf v}}}$.
By Theorem~\ref{thm:lb-with-zero},
this segment includes the entire ray of all points
$\lambda{\bf v}$, for $\lambda\geq 0$.
Furthermore, $\conv(\lmset{S})$, being convex, must also include all
convex combinations of such points, and thus must include $\cone S$,
the entire convex cone (in ${\R^n}$) generated by the points in $S$.
Intuitively, then, there would seem to be a close connection between
$\conv(\lmset{S})$ and $\cone{S}$.
In fact, the
relationship between these two sets
can be stated with exact precision.
To explain this,
for any set $U\subseteq{\R^n}$, let us define $\repcl{U}$ to be
the set of all points in $\extspac{n}$ which have representations that
only involve vectors in $U$.
That is,
\[
\repcl{U} =
\braces{
\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q} :
{\bf q},{\bf v}_1,\ldots,{\bf v}_k,\in U, k\geq 0
}.
\]
(Note that these need \emph{not} be canonical representations.)
We call $\repcl{U}$ the \emph{representational closure} of $U$, since
it comprises exactly those astral points that can be represented using
only vectors in $U$.
For example, in this notation,
Corollary~\ref{cor:h:1} states exactly that
$\extspac{n}=\repcl{({\R^n})}$.
We can now state the main result of this subsection, an exact expression
for $\conv(\lmset{S})$ in terms of the representational closure of
$\cone{S}$.
In particular, if the origin is in the convex hull of $S$, then
$\conv(\lmset{S})$ is equal to $\repcl{(\cone{S})}$.
Otherwise, if the origin is not in $S$'s convex hull, then
$\conv(\lmset{S})$ consists of just the points in
$\repcl{(\cone{S})}$ that are outside ${\R^n}$.
\begin{theorem} \label{thm:conv-lmset-char}
Let $S\subseteq{\R^n}$.
If ${\bf 0}\in\conv{S}$, then
\[
\conv(\lmset{S})
= \repcl{(\cone{S})}.
\]
Otherwise, if ${\bf 0}\not\in\conv{S}$, then
\[
\conv(\lmset{S})
= \repcl{(\cone{S})} \setminus {\R^n}.
\]
\end{theorem}
We develop a proof of this theorem by proving a number of lemmas and
propositions regarding exactly which points are or are not in
$\conv(\lmset{S})$.
First, the astron associated with any point in $\scc{S}$ must also be
included in $\conv(\lmset{S})$.
Said differently, if ${\bf z}$ can be written as a positive linear
combination of points in $S$ (that is, a linear combination all of
whose coefficients are positive), then
$\limray{{\bf z}}\in\conv(\lmset{S})$.
\begin{proposition} \label{pr:scc-in-conv-lmset}
Let $S\subseteq{\R^n}$.
Suppose ${\bf z}\in\scc{S}$.
Then $\limray{{\bf z}}\in\conv(\lmset{S})$.
\end{proposition}
\proof
Since ${\bf z}\in\scc{S}$, it can be written in the form
\[
{\bf z} = c \sum_{i=1}^m \lambda_i {\bf v}_i
\]
for some $c>0$, ${\bf v}_1,\ldots,{\bf v}_m\in S$,
$\lambda_1,\ldots,\lambda_m\in [0,1]$,
$\sum_{i=1}^m \lambda_i = 1$.
(See \citet[Corollary~2.6.2]{ROC}.)
For $i=1,\ldots,m$, and
for $t=1,2,\ldots$, let us define
${\bf x}_{it} = c t {\bf v}_i$, and also define
\[
{\bf z}_t = \sum_{i=1}^m \lambda_i {\bf x}_{it} = t {\bf z}.
\]
Then for each $i=1,\ldots,m$,
${\bf x}_{it}\rightarrow \limray{{\bf v}_i}$,
and ${\bf z}_t\rightarrow\limray{{\bf z}}$.
Therefore, having met the conditions of Theorem~\ref{thm:e:7}, we
conclude that
\[
\limray{{\bf z}}\in\ohull{\{ \limray{{\bf v}_1},\ldots,\limray{{\bf v}_m}\} }
\subseteq \conv(\lmset{S})
\]
where the second inclusion is by Theorem~\ref{thm:e:2}.
\qed
Next, we show that $\conv(\lmset{S})$ must be included in
$\clcone{S}$, the topological closure of $\cone{S}$ in $\extspac{n}$:
\begin{proposition} \label{pr:a:3}
Let $S\subseteq{\R^n}$.
Then
\[ \conv(\lmset{S}) \subseteq \clcone{S}. \]
\end{proposition}
\proof
For all ${\bf v}\in S$,
since $\cone{S}$ is a cone that includes all of $S$,
$t{\bf v}$ must be in $\cone{S}$, for $t=1,2,\ldots$.
Therefore,
$\limray{{\bf v}}=\lim t{\bf v}$
is in $\clcone{S}$.
Thus, $\lmset{S}\subseteq\clcone{S}$.
Since $\clcone{S}$ is convex by Theorem~\ref{thm:e:6},
$\conv(\lmset{S})\subseteq\clcone{S}$
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa}).
\qed
We now come to a central lemma showing that, when the set $S$ is finite,
$\clcone{S}$, the \emph{topological} closure of $\cone{S}$,
is included in $\cone{S}$'s \emph{representational} closure.
Combined with Proposition~\ref{pr:a:3}, this will later allow us to
prove that $\conv(\lmset{S})\subseteq \repcl{(\cone{S})}$ for any set
$S\subseteq{\R^n}$.
\begin{lemma} \label{lem:conv-lmset-char:a}
Let $V$ be a finite subset of ${\R^n}$.
Then
$\clcone{V}\subseteq \repcl{(\cone V)}$.
\end{lemma}
\proof
If $V$ is empty, then
$\cone V = \{{\bf 0}\}$, so
$\clcone{V} = \{{\bf 0}\} = \repcl{(\cone V)}$.
For $V$ nonempty,
we prove the result by induction on astral rank.
More precisely,
we prove, by induction on $k=0,1,\ldots,n$, that for all
$\overline{\xx}\in\extspac{n}$ and for every finite, nonempty set $V\subseteq{\R^n}$,
if $\overline{\xx}$ has astral rank at most $k$ and if $\overline{\xx}\in\clcone{V}$
then $\overline{\xx}\in\repcl{(\cone V)}$.
Suppose $V=\{{\bf z}_1,\ldots,{\bf z}_m\}\subseteq{\R^n}$, $m\geq 1$.
In this case, the cone generated by $V$ is the set of all nonnegative
linear combinations of the points in $V$; that is,
\[
\cone V =
\braces{
\sum_{i=1}^m \alpha_i {\bf z}_i :
\alpha_1,\ldots,\alpha_m\geq 0
}.
\]
(See \citet[Corollary~2.6.2]{ROC} and following discussion.)
Moreover, $\cone V$ is closed in ${\R^n}$
\citep[by][Theorem~19.1]{ROC},
though not in $\extspac{n}$.
For the base case that $k=0$,
suppose $\overline{\xx}$ has astral rank zero and that $\overline{\xx}\in\clcone{V}$.
Then $\overline{\xx}$ is equal to some ${\bf q}\in{\R^n}$.
Since $\overline{\xx}\in\clcone{V}$, there exists a sequence $\seq{{\bf x}_t}$
in $\cone V$ converging to $\overline{\xx}={\bf q}$.
Thus, ${\bf q}\in \cone V$, since $\cone V$ is closed in ${\R^n}$.
Therefore, $\overline{\xx}\in\repcl{(\cone V)}$.
For the inductive step, suppose $\overline{\xx}$ has astral rank $k>0$ and
that $\overline{\xx}\in\clcone{V}$.
Furthermore, assume inductively
that the claim holds for all points with astral rank strictly less
than $k$.
Then
by Proposition~\ref{pr:h:6},
we can write
$\overline{\xx}=\limray{{\bf w}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\xbar^{\bot}$
where ${\bf w}$ is $\overline{\xx}$'s dominant direction, and
$\xbar^{\bot}$ is the
projection of $\overline{\xx}$ perpendicular to ${\bf w}$.
Since $\overline{\xx}\in\clcone{V}$, there exists a sequence $\seq{{\bf x}_t}$
in $\cone V$ that converges to $\overline{\xx}$.
By Theorem~\ref{thm:dom-dir},
\[
{\bf w} = \lim \frac{{\bf x}_t}{\norm{{\bf x}_t}}.
\]
(As usual, since $\norm{{\bf x}_t}\rightarrow+\infty$, we can discard
all elements of the sequence with ${\bf x}_t={\bf 0}$, of which there can be
at most finitely many.)
Since $\cone V$ is a cone,
${\bf x}_t/\norm{{\bf x}_t}\in\cone V$, for all $t$,
which implies, since $\cone V$ is closed in ${\R^n}$,
that ${\bf w}\in \cone V$.
For each $t$, we can write
${\bf x}_t = b_t {\bf w} + {\xx_t^\bot}$
where ${\xx_t^\bot}\cdot{\bf w}=0$ and $b_t={\bf x}_t\cdot{\bf w}\in{\mathbb{R}}$.
Since ${\bf x}_t\rightarrow\overline{\xx}$ and since ${\bf w}$ is $\overline{\xx}$'s dominant
direction,
${\bf x}_t\cdot{\bf w}\rightarrow \overline{\xx}\cdot{\bf w} = +\infty$
(by Theorems~\ref{thm:i:1}(\ref{thm:i:1c})
and~\ref{thm:dom-dir}).
Thus, $b_t$ must be positive for all but finitely many values of $t$;
by discarding these, we can assume that $b_t>0$ for all $t$.
Recall that each ${\bf x}_t$, being in $\cone V$, is a nonnegative linear
combination of ${\bf z}_1,\ldots,{\bf z}_m$, the points comprising $V$.
Therefore, each point
${\xx_t^\bot}={\bf x}_t-b_t{\bf w}$ is a nonnegative linear combination of the
points in the expanded set
\[
V' = V \cup \{ -{\bf w} \} = \{ {\bf z}_1,\ldots,{\bf z}_m,\; -{\bf w} \}.
\]
In other words, ${\xx_t^\bot}\in\cone V'$, for all $t$.
Furthermore, $\xbar^{\bot}$ must be in the astral closure of this cone,
$\clcone{V'}$,
since
${\xx_t^\bot}\rightarrow\xbar^{\bot}$
(by Proposition~\ref{pr:h:5}(\ref{pr:h:5b})).
Since $\xbar^{\bot}$ has strictly lower astral rank than $\overline{\xx}$
(by Proposition~\ref{pr:h:6}), we can apply our
inductive assumption which implies that
\[ \xbar^{\bot} = [{\bf v}'_1,\ldots,{\bf v}'_{k'}]{\bar{\bm{\omega}}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}' \]
for some ${\bf q}',{\bf v}'_1,\ldots,{\bf v}'_{k'}\in \cone V'$.
For $i=1,\ldots,k'$,
since ${\bf v}'_i\in \cone V'$,
we can write ${\bf v}'_i$ as a nonnegative linear combination over the
points in $V'$ so that
${\bf v}'_i={\bf v}_i - a_i {\bf w}$
where
\[ {\bf v}_i = \sum_{j=1}^m c_{ij} {\bf z}_j \]
for some $a_i\geq 0$, $c_{ij}\geq 0$, $j=1,\ldots,m$.
Note that ${\bf v}_i\in\cone V$.
Similarly we can write
${\bf q}'={\bf q}-b{\bf w}$ for some ${\bf q}\in \cone V$ and $b\geq 0$.
Thus,
\begin{eqnarray*}
\overline{\xx}
&=&
\limray{{\bf w}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \xbar^{\bot}
\\
&=&
\limray{{\bf w}}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
[{\bf v}'_1,\ldots,{\bf v}'_{k'}]{\bar{\bm{\omega}}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}'
\\
&=&
[{\bf w},\; {\bf v}'_1,\ldots,{\bf v}'_{k'}]{\bar{\bm{\omega}}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}'
\\
&=&
[{\bf w},\; {\bf v}_1,\ldots,{\bf v}_{k'}]{\bar{\bm{\omega}}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}.
\end{eqnarray*}
The last line follows from Proposition~\ref{pr:g:1} since
${\bf v}_i={\bf v}'_i+a_i{\bf w}$ for $i=1,\ldots,k'$
(relationships that can be expressed using an
appropriate positive upper triangular matrix), and since
${\bf q}'-{\bf q}=-b{\bf w}$ is in the appropriate column space.
Since ${\bf w}$, ${\bf q}$, and ${\bf v}_1,\ldots,{\bf v}_{k'}$ are all in
$\cone V$,
this shows that $\overline{\xx}\in\repcl{(\cone V)}$,
completing the induction and the proof.
\qed
When the origin is not included in the convex hull of $S$,
$\conv(\lmset{S})$ is entirely disjoint from ${\R^n}$:
\begin{lemma} \label{lem:conv-lmset-char:b}
Let $S\subseteq {\R^n}$.
If ${\bf 0}\not\in\conv{S}$ then
$\conv(\lmset{S})\cap{\R^n} = \emptyset$.
\end{lemma}
\proof
We prove the lemma in the contrapositive.
Suppose ${\bf q}\in\conv(\lmset{S})\cap{\R^n}$.
By Theorem~\ref{thm:convhull-of-simpices}, this implies that
\[
{\bf q}\in \ohull{\{\limray{{\bf v}_1},\ldots,\limray{{\bf v}_m}\}}
\]
for some ${\bf v}_1,\ldots,{\bf v}_m\in S$.
By \eqref{eq:e:5}, this means that for all
${\bf u}\in{\R^n}$,
\begin{equation} \label{eq:h:8a}
{\bf q}\cdot{\bf u} \leq \max\{\limray{{\bf v}_1}\cdot{\bf u},\,\ldots,\,\limray{{\bf v}_m}\cdot{\bf u}\}.
\end{equation}
This implies that
\begin{equation} \label{eq:h:8b}
0 \leq \max\{{{\bf v}_1}\cdot{\bf u},\,\ldots,\,{{\bf v}_m}\cdot{\bf u}\}
\end{equation}
since otherwise, if ${\bf v}_i\cdot{\bf u}<0$ for $i=1,\ldots,m$, then the
right-hand side of \eqref{eq:h:8a} would be equal to $-\infty$, which
is impossible since ${\bf q}\cdot{\bf u}\in{\mathbb{R}}$.
Let $U$ denote the unit sphere,
\[ U = \{ {\bf u}\in{\R^n} : \norm{{\bf u}} = 1 \}, \]
and let $P$ denote the set of nonnegative vectors
${\bf p}=\trans{[p_1,\ldots,p_n]}$ that sum to~1:
\[ P = \braces{ {\bf p}\in [0,1]^n : \sum_{i=1}^n p_i = 1 }. \]
Then from \eqref{eq:h:8b}, we have
\begin{eqnarray}
0
&\leq&
\min_{{\bf u}\in U} \max\{{\bf v}_i\cdot{\bf u} : i=1,\ldots,m\}
\label{eq:h:8c:1}
\\
&=&
\min_{{\bf u}\in U} \max_{{\bf p}\in P} \sum_{i=1}^m p_i {\bf v}_i\cdot{\bf u}
\label{eq:h:8c:2}
\\
&=&
\max_{{\bf p}\in P} \min_{{\bf u}\in U} \sum_{i=1}^m p_i {\bf v}_i\cdot{\bf u}
\label{eq:h:8c:3}
\\
&=&
\max_{{\bf p}\in P} \brackets{ - \norm{\sum_{i=1}^m p_i {\bf v}_i} }
\label{eq:h:8c:4}
\\
&=&
- \min_{{\bf p}\in P} { \norm{\sum_{i=1}^m p_i {\bf v}_i} }.
\label{eq:h:8c:5}
\end{eqnarray}
(We use $\min$ and $\max$ rather than $\inf$ and $\sup$ because
the functions being considered are all continuous and the sets $U$ and
$P$ are both compact.)
\eqref{eq:h:8c:1} is just stating that
\eqref{eq:h:8b} holds for all ${\bf u}\in U$ (since actually it holds for
all ${\bf u}\in{\R^n}$).
\eqref{eq:h:8c:2} is because the maximum over ${\bf p}\in P$ must be
realized when $p_i=1$ for that $i$ which maximizes ${\bf v}_i\cdot{\bf u}$.
\eqref{eq:h:8c:3} follows from \citet[Corollary~37.3.2]{ROC}.
\eqref{eq:h:8c:4} is because, for all ${\bf x}\in{\R^n}$,
\[ \min_{{\bf u}\in U} {\bf x}\cdot{\bf u} = -\norm{{\bf x}}. \]
Rewriting \eqref{eq:h:8c:5} yields
\[
\min_{{\bf p}\in P} { \norm{\sum_{i=1}^m p_i {\bf v}_i} } \leq 0.
\]
In other words, there exists ${\bf p}\in P$ with
$\sum_{i=1}^m p_i {\bf v}_i = {\bf 0}$, which means exactly that
${\bf 0}$ is a convex combination of the ${\bf v}_i$'s and thus of points in
$S$.
Therefore, ${\bf 0}\in \conv{S}$.
\qed
Finally, we are ready to prove Theorem~\ref{thm:conv-lmset-char}:
\proof[Proof of Theorem~\ref{thm:conv-lmset-char}]
We first show that $\conv(\lmset{S})$ is contained in the sets stated
in the theorem.
\setcounter{claimp}{0}
\medskip
\begin{claimp}
For all $S\subseteq {\R^n}$,
$\conv(\lmset{S}) \subseteq \repcl{(\cone{S})}$.
\end{claimp}
\proof
For any finite subset $V\subseteq S$,
\[
\ohull(\lmset{V})
=
\conv(\lmset{V})
\subseteq
\clcone{V}
\subseteq
\repcl{(\cone V)}
\subseteq
\repcl{(\cone{S})}
\]
where the equality and first two inclusions follow, respectively, from
Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:a}),
Proposition~\ref{pr:a:3},
and Lemma~\ref{lem:conv-lmset-char:a}.
The last inclusion is because $V\subseteq S$, a relation that is
preserved when generating cones and taking representational closure.
Since this holds for every finite $V\subseteq S$,
the claim now follows by applying
Theorem~\ref{thm:convhull-of-simpices}.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
From this claim, combined with Lemma~\ref{lem:conv-lmset-char:b}, it
also follows that if ${\bf 0}\not\in\conv{S}$ then
$\conv(\lmset{S}) \subseteq \repcl{(\cone{S})}\setminus {\R^n}$.
To complete the proof, we show that $\conv(\lmset{S})$ contains all of
the points stated in the theorem.
\medskip
\begin{claimp}
For all $S\subseteq {\R^n}$,
$\repcl{(\cone{S})} \setminus {\R^n}
\subseteq
\conv(\lmset{S})$.
\end{claimp}
\proof
Let $\overline{\xx}\in \repcl{(\cone{S})} \setminus {\R^n}$.
Then we can write
\[
\overline{{\bf z}} = \limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}
\]
for some ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in\cone{S}$.
Without loss of generality, we assume ${\bf v}_i\neq{\bf 0}$,
for $i=1,\ldots,k$.
Furthermore, since $\overline{{\bf z}}\not\in{\R^n}$, we must have $k>0$.
For $i=1,\ldots,k$,
let us define sequences
\[ {\bf x}_{it} = t^{k+1-i} {\bf v}_i, \]
and also ${\bf y}_t = t {\bf q}$.
Then ${\bf x}_{it}\rightarrow \limray{{\bf v}_i}$
and ${\bf y}_t\rightarrow \limray{{\bf q}}$.
Suppose first that ${\bf q}\neq{\bf 0}$.
In this case, we define
\[
{\bf z}_t = \sum_{i=1}^k \lambda_{it} {\bf x}_{it} + \gamma_t {\bf y}_t
\]
where
\[
\lambda_{it} = \frac{t-1}{tk}
\;\;\mbox{and}\;\;\;
\gamma_t = \frac{1}{t}.
\]
It can be checked that $\sum_{i=1}^k \lambda_{it} + \gamma_t = 1$,
and that ${\bf z}_t\rightarrow\overline{{\bf z}}$.
Thus, by Theorem~\ref{thm:e:7},
\begin{equation} \label{eq:h:8d}
\overline{{\bf z}} \in \ohull{\{\limray{{\bf v}_1},\,\ldots,\,\limray{{\bf v}_k},\,\limray{{\bf q}}\}}.
\end{equation}
For each $i=1,\ldots,k$,
${\bf v}_i$ is in $\cone{S}$, and because ${\bf v}_i\neq{\bf 0}$, ${\bf v}_i$
must actually be in $\scc{S}$.
Therefore, $\limray{{\bf v}_i}\in\conv(\lmset{S})$
by Proposition~\ref{pr:scc-in-conv-lmset}.
Similarly, $\limray{{\bf q}}\in\conv(\lmset{S})$.
Thus, \eqref{eq:h:8d} implies that
$\overline{{\bf z}}\in\conv(\lmset{S})$ by
Theorem~\ref{thm:e:2}.
The proof for when ${\bf q}={\bf 0}$ is similar.
In this case, we simply let
\[ {\bf z}_t = \frac{1}{k} \sum_{i=1}^k {\bf x}_{it} \]
(so $\lambda_{it}=1/k$ for $i=1,\ldots,k$, $t=1,2,\ldots$).
We again have that ${\bf z}_t\rightarrow\overline{{\bf z}}$, showing that
\[
\overline{{\bf z}} \in \ohull{\{\limray{{\bf v}_1},\,\ldots,\,\limray{{\bf v}_k}\}},
\]
and so
$\overline{{\bf z}}\in\conv(\lmset{S})$ by similar reasoning.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
This completes the proof for when ${\bf 0}\not\in\conv{S}$.
For the alternative case, it only remains to show that
$\cone{S}= \repcl{(\cone{S})} \cap {\R^n}$
is included in $\conv(\lmset{S})$.
\medskip
\begin{claimp}
For all $S\subseteq {\R^n}$,
if ${\bf 0}\in\conv{S}$ then
${\cone{S}}
\subseteq
\conv(\lmset{S})$.
\end{claimp}
\proof
Since ${\bf 0}$ is in $\conv{S}\subseteq\scc{S}$,
we must have ${\bf 0}\in\conv(\lmset{S})$ by
Proposition~\ref{pr:scc-in-conv-lmset}.
Let ${\bf q}\in\cone{S}$, and assume ${\bf q}\neq{\bf 0}$ (since we have
already shown that ${\bf 0}\in\conv(\lmset{S})$).
Since ${\bf q}\in(\cone{S})\setminus\{{\bf 0}\}\subseteq\scc{S}$,
$\limray{{\bf q}}\in\conv(\lmset{S})$,
again by Proposition~\ref{pr:scc-in-conv-lmset}.
From Theorem~\ref{thm:lb-with-zero},
${\bf q}$ is in $\lb{{\bf 0}}{\limray{{\bf q}}}$, the segment joining ${\bf 0}$
and $\limray{{\bf q}}$.
Therefore, ${\bf q}\in\conv(\lmset{S})$ by definition of convexity.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
This completes all the parts of Theorem~\ref{thm:conv-lmset-char}.
\qed
As a corollary of Theorem~\ref{thm:conv-lmset-char},
for any set $S\subseteq{\R^n}$ (whether or not the origin is in its convex hull),
the cone generated by $\lmset{S}$ is equal to
the representational closure of $\cone{S}$:
\begin{corollary}
\label{cor:cone-omegaS}
For any $S\subseteq{\R^n}$,
\[
\cone(\lmset{S}) = \repcl{(\cone{S})}.
\]
\end{corollary}
\proof
Let $S\subseteq{\R^n}$.
Then
\begin{eqnarray*}
\cone(\lmset{S})
&=&
\conv((\lmset{S})\cup\{{\bf 0}\})
\\
&=&
\conv(\lmset{(S\cup\{{\bf 0}\})})
\\
&=&
\repcl{\cone(S\cup\{{\bf 0}\})}
\\
&=&
\repcl{(\cone{S})},
\end{eqnarray*}
with the first and third equalities following from
Proposition~\ref{pr:cone-props}(\ref{pr:cone-props:d})
and
Theorem~\ref{thm:conv-lmset-char},
respectively.
\qed
As another corollary, for any convex cone $K\subseteq{\R^n}$ that includes the
origin, we have proved that $\repcl{K}$ is a convex cone that is
included in $K$'s topological closure:
\begin{corollary} \label{cor:a:1}
Let $K\subseteq{\R^n}$ be a pointed convex cone.
Then $\repcl{K}$ is a convex cone and
\[
K \subseteq \repcl{K}=\conv(\lmset{K})
\subseteq \overline{K}.
\]
\end{corollary}
\proof
Since $K$ is already a convex cone that includes ${\bf 0}$,
$\cone K = K$.
Thus, by Theorem~\ref{thm:conv-lmset-char},
$\repcl{K}=\conv(\lmset{K})$, which is a convex cone
(by Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:c},\ref{pr:new:a:2:d})).
That
$\conv(\lmset{K})\subseteq \overline{K}$ follows immediately from
Proposition~\ref{pr:a:3}.
That $K\subseteq\repcl{K}$ is immediate from the form of points in
$K$'s representational closure.
\qed
\subsection{Representational closure of polyhedral cones}
Corollary~\ref{cor:a:1} shows that
if $K\subseteq{\R^n}$ is a pointed convex cone,
then $\repcl{K}\subseteq\overline{K}$.
In fact, if $K$ is \emph{finitely generated}, that is, if $K=\cone V$
for some finite set $V\subseteq{\R^n}$, then
Lemma~\ref{lem:conv-lmset-char:a}, together with
Corollary~\ref{cor:a:1}, proves that $\repcl{K}=\overline{K}$, and thus, that
$\repcl{K}$ is closed.
As we show in this section, these properties of $K$ and its
representational closure turn out to be equivalent;
in other words, we will prove that $\repcl{K}$ is closed if and only
if $K$ is finitely generated.
This fact will be especially important in characterizing the
continuity of a function's extension, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
(see Section~\ref{subsec:charact:cont:duality}).
It is a standard fact that a convex set in ${\R^n}$ is finitely generated
if and only if it is \emph{polyhedral}, meaning that the set is the
intersection of finitely many closed halfspaces
\citep[Theorem~19.1]{ROC}.
Thus, when discussing convex sets in ${\R^n}$,
particularly convex cones, we can
use these terms interchangably.
{
Before proving the general result, we give an example
of a convex cone $K$ showing that $\repcl{K}$ need not be equal to
$\overline{K}$ in general.}
For this purpose, let us define the function $f:{\mathbb{R}}^2\rightarrow\overline{\R}$ by
\begin{equation} \label{eqn:curve-discont-notfinite-eg}
f(x_1,x_2) = \left\{ \begin{array}{cl}
x_1^2 / x_2 & \mbox{if $x_2 \geq |x_1|$ and $x_2>0$} \\
0 & \mbox{if $x_1 = x_2 = 0$} \\
+\infty & \mbox{otherwise.}
\end{array} \right.
\end{equation}
This function is convex, closed and proper.
Let $K\subseteq{\mathbb{R}}^3$ be the epigraph of $f$, that is, the set of all points
$\trans{[x_1,x_2,x_3]}\in{\mathbb{R}}^3$ with $x_3\geq f(x_1,x_2)$.
Since $f$ is convex, its epigraph $K$ is convex as well.
It also can be checked that $K$ is a cone, being closed under positive
scalar multiplication, and that ${\bf 0}\in K$.
Let $\overline{{\bf z}}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf e}_1$, where, as usual,
${\bf e}_1,{\bf e}_2,{\bf e}_3$ are standard basis vectors.
We claim that $\overline{{\bf z}}\in\overline{K}$.
This is because, for each $t=1,2,\ldots$, the point
${\bf z}_t=\trans{[1,\,t,\,1/t]}\in K$, so
$\overline{{\bf z}}=\lim {\bf z}_t$ is in $\overline{K}$.
On the other hand, $\overline{{\bf z}}\not\in\repcl{K}$.
To see this, observe first that if ${\bf x}\in K$ and $x_3=0$ then it must
also be the case that $x_1=0$.
Now suppose that $\overline{{\bf z}}\in\repcl{K}$ so that
$\overline{{\bf z}}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
for some ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in K$.
Note that $\overline{{\bf z}}\cdot{\bf e}_3=0$.
This implies that
${\bf v}_i\cdot{\bf e}_3=0$ for $i=1,\ldots,k$, and so also that
${\bf q}\cdot{\bf e}_3=0$
(by Proposition~\ref{pr:vtransu-zero}).
But then, by the preceding observation, it must also be the case that
${\bf q}\cdot{\bf e}_1=0$, and that
${\bf v}_i\cdot{\bf e}_1=0$ for $i=1,\ldots,k$.
These imply that $\overline{{\bf z}}\cdot{\bf e}_1=0$, a contradiction, since in fact,
$\overline{{\bf z}}\cdot{\bf e}_1=1$.
We conclude that $\repcl{K}\neq\overline{K}$ in this case.
We now return to the general case:
\begin{theorem} \label{thm:repcl-polyhedral-cone}
Let $K\subseteq{\R^n}$ be a pointed convex cone.
Then the following are equivalent:
\begin{enumerate}
\item \label{thm:repcl-polyhedral-cone:a}
$\repcl{K}$ is closed.
\item \label{thm:repcl-polyhedral-cone:b}
$\repcl{K}=\overline{K}$.
\item \label{thm:repcl-polyhedral-cone:c}
$K$ is finitely generated (or equivalently, polyhedral);
that is, $K=\cone V$ for some finite set $V\subseteq{\R^n}$.
\end{enumerate}
\end{theorem}
\proof
That
(\ref{thm:repcl-polyhedral-cone:b})
implies
(\ref{thm:repcl-polyhedral-cone:a})
is immediate.
Corollary~\ref{cor:a:1} shows that $\repcl{K}\subseteq\overline{K}$ in
general.
If $K$ is finitely generated, then
Lemma~\ref{lem:conv-lmset-char:a} proves $\overline{K}\subseteq\repcl{K}$,
thus proving
(\ref{thm:repcl-polyhedral-cone:c})
implies
(\ref{thm:repcl-polyhedral-cone:b}).
In the remainder of the proof, we show that
(\ref{thm:repcl-polyhedral-cone:a})
implies
(\ref{thm:repcl-polyhedral-cone:c}).
We assume henceforth that $\repcl{K}$ is closed, and therefore compact
being a closed subset of the compact space $\extspac{n}$.
To prove the result, we construct an open cover of $\repcl{K}$, which,
since $\repcl{K}$ is compact, must include a finite subcover.
From this, we show that a finite set of points can be extracted that
are sufficient to generate the cone $K$.
We make use of the open sets ${U_{\vv}}$ that were shown to exist in
Theorem~\ref{thm:formerly-lem:h:1}.
For ${\bf v}\in{\R^n}$, recall that the set ${U_{\vv}}$ includes the astron
$\limray{{\bf v}}$ but no other icons.
Moreover, ${U_{\vv}}$ is entirely contained in
${\R^n} \cup [\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\R^n}]$,
meaning all points in ${U_{\vv}}$ are either in ${\R^n}$ or have the form
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some ${\bf q}\in{\R^n}$.
Slightly overloading notation,
for any set $S\subseteq{\R^n}$, we further define ${U_{S}}$ to be the convex
hull of the union of all sets ${U_{\vv}}$ over ${\bf v}\in S$:
\[
{U_{S}} = \conv\paren{\bigcup_{{\bf v}\in S} {U_{\vv}}}.
\]
The parenthesized union is open (being the union of open sets),
implying, by Corollary~\ref{cor:convhull-open}, that its convex
hull, ${U_{S}}$, is also open, for all $S\subseteq{\R^n}$.
Using compactness, we show next that $\repcl{K}$ is included in one of
these sets ${U_{V}}$ for some {finite} set $V\subseteq K$.
First,
for all ${\bf v}\in{\R^n}$, $\limray{{\bf v}}\in{U_{\vv}}$;
therefore, for all $S\subseteq{\R^n}$,
$\lmset{S}\subseteq{U_{S}}$, and so
$\conv(\lmset{S})\subseteq{U_{S}}$
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})
since ${U_{S}}$ is convex.
As a result, we can cover $\repcl{K}$ using the collection of all
open sets ${U_{V}}$ for all finite subsets
$V\subseteq K$.
This is a cover because
\[
\repcl{K}
=
\conv(\lmset{K})
=
\bigcup_{\scriptontop{V\subseteq K:}{|V|<+\infty}} \conv(\lmset{V})
\subseteq
\bigcup_{\scriptontop{V\subseteq K:}{|V|<+\infty}} {U_{V}}.
\]
The equalities follow respectively from
Theorem~\ref{thm:conv-lmset-char}
and
Theorem~\ref{thm:convhull-of-simpices}
(and also
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:a})).
Since $\repcl{K}$ is compact, there exists a finite subcover, that is,
a finite collection of sets $V_1,\ldots,V_\ell$ where each $V_j$ is a
finite subset of $K$, for $j=1,\ldots,\ell$, and
such that
\begin{equation} \label{eqn:thm:repcl-polyhedral-cone:1}
\repcl{K} \subseteq \bigcup_{j=1}^{\ell} {U_{V_j}}.
\end{equation}
Let
\[ V=\{{\bf 0}\}\cup \bigcup_{j=1}^{\ell} V_j \]
be their union, along with the origin (added for later convenience),
which is also a finite subset of $K$.
Furthermore,
\eqref{eqn:thm:repcl-polyhedral-cone:1} implies
$\repcl{K}\subseteq {U_{V}}$ since, for $j=1,\ldots,\ell$, $V_j\subseteq V$,
implying ${U_{V_j}}\subseteq{U_{V}}$
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:b}).
Summarizing, $\repcl{K}\subseteq {U_{V}}$, where
${\bf 0}\in V\subseteq K$ and $|V|<+\infty$.
Let $\hat{K}=\cone V$ be the cone generated by $V$.
To complete the proof that $K$ is finitely generated, we show that
$K=\hat{K}$.
Actually,
$\hat{K}=\cone V=\scc V\subseteq K$
by Proposition~\ref{pr:cone-props}(\ref{pr:cone-props:a})
since $V$ is a subset of the convex cone $K$
(and since ${\bf 0}\in V$).
So it remains only to prove that $K\subseteq\hat{K}$.
Let ${\bf w}$ be any point in $K$, which
we aim to show is in $\hat{K}$.
Since ${\bf w}\in K$,
\[
\limray{{\bf w}}\in\repcl{K}\subseteq {U_{V}}
=
\conv\paren{\bigcup_{{\bf v}\in V} {U_{\vv}}}.
\]
Therefore, by
Theorem~\ref{thm:convhull-of-simpices},
$\limray{{\bf w}}$ is in the polytope formed by some finite set of points,
$\overline{\xx}_1,\ldots,\overline{\xx}_m$, where each $\overline{\xx}_j$ is
in $\bigcup_{{\bf v}\in V} {U_{\vv}}$.
From the form of points in ${U_{\vv}}$
(Theorem~\ref{thm:formerly-lem:h:1}(\ref{thm:formerly-lem:h:1:b})),
this means, for $j=1,\ldots,m$, that we can write
$\overline{\xx}_j=\limray{{\bf v}_j}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}_j$ for some ${\bf v}_j\in V$ and some
${\bf q}_j\in{\R^n}$.
(Note that this takes into account the possibility that $\overline{\xx}_j$
might be in ${\R^n}$ since in that case we can choose ${\bf v}_j={\bf 0}$,
which is in $V$.)
Thus,
\[
\limray{{\bf w}}
\in
\ohull{ \{ \overline{\xx}_1,\ldots,\overline{\xx}_m \} },
\]
implying, by \eqref{eq:e:5}, that
\begin{equation} \label{eqn:thm:repcl-polyhedral-cone:4}
\limray{{\bf w}}\cdot{\bf u}
\leq
\max\{ \overline{\xx}_1\cdot{\bf u},\ldots,\overline{\xx}_m\cdot{\bf u} \}
\end{equation}
for all ${\bf u}\in{\R^n}$.
We claim that ${\bf w}$ is in the polar
(defined in \eqref{eqn:polar-def})
of the polar of $\hat{K}$, that is, that ${\bf w}$ is in
$(\polar{\Khat})^{\circ}=\dubpolar{\Khat}$, which is the
same as $\hat{K}$.
Let ${\bf u}$ be any point in $\polar{\Khat}$.
Then for $j=1,\ldots,m$,
${\bf v}_j\in V \subseteq \cone V = \hat{K}$, so
${\bf v}_j\cdot{\bf u} \leq 0$.
Therefore,
\[
\overline{\xx}_j\cdot{\bf u}
=
\limray{{\bf v}_j}\cdot{\bf u}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
{\bf q}_j\cdot{\bf u}
< +\infty
\]
since $\limray{{\bf v}_j}\cdot{\bf u}\leq 0$ and ${\bf q}_j\cdot{\bf u}\in{\mathbb{R}}$.
Combined with \eqref{eqn:thm:repcl-polyhedral-cone:4}, this
means $\limray{{\bf w}}\cdot{\bf u}<+\infty$, and therefore
$\limray{{\bf w}}\cdot{\bf u}\leq 0$ (since $\limray{{\bf w}}$ is an icon)
so ${\bf w}\cdot{\bf u}\leq 0$.
Since this holds for all ${\bf u}\in\polar{\Khat}$, it follows that
${\bf w}\in\dubpolar{\Khat}$.
Furthermore,
because $\hat{K}$ is a finitely generated convex cone in ${\R^n}$, it
must be closed in ${\R^n}$;
therefore, $\dubpolar{\Khat}=\hat{K}$
\citep[Theorems~19.1 and~14.1]{ROC}.
Thus, $K=\hat{K}$, so $K$ is finitely generated.
\qed
\subsection{Astral polar cones}
\label{sec:ast-pol-cones}
We next study a natural extension of the standard polar cone
defined in \eqref{eqn:polar-def} to astral space.
We will see that such astral polar cones play a critical role both in
the minimization of convex functions and in their continuity
properties.
Specifically, for a convex cone $K\subseteq{\R^n}$, we define the
\emph{astral polar} of $K$, denoted $\apol{K}$, to be the set
\[
\apol{K} = \braces{\overline{\xx}\in\extspac{n} :
\forall {\bf u}\in K,\, \overline{\xx}\cdot{\bf u}\leq 0}.
\]
Here are some basic properties:
\begin{proposition} \label{pr:ast-pol-props}
Let $K\subseteq{\R^n}$ be a convex cone.
Then the following hold:
\begin{enumerate}
\item \label{pr:ast-pol-props:a}
$\polar{K}=\apol{K}\cap {\R^n}$.
\item \label{pr:ast-pol-props:b}
If $J$ is a convex cone with $J\subseteq K$, then
$\apol{K}\subseteq\apol{J}$.
\item \label{pr:ast-pol-props:c}
$\apol{K}$ is a closed (in $\extspac{n}$) convex cone that includes
the origin.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:ast-pol-props:a}) follows immediately from definitions.
Part~(\ref{pr:ast-pol-props:b}):
If $\overline{\xx}\in\apol{K}$ then $\overline{\xx}\cdot{\bf u}\leq 0$ for all ${\bf u}\in K$,
and therefore all ${\bf u}\in J$; thus, $\overline{\xx}\in\apol{J}$.
Part~(\ref{pr:ast-pol-props:c}):
The set $\apol{K}$ can be expressed as an intersection of closed
halfspaces, namely,
\[
\apol{K} = \bigcap_{{\bf u}\in K}
\braces{\overline{\xx}\in\extspac{n} : \overline{\xx}\cdot{\bf u}\leq 0}.
\]
Therefore, this set is closed and convex, by
Proposition~\ref{pr:e1}(\ref{pr:e1:b},\ref{pr:e1:c}).
Each closed halfspace is a cone
(since, for any $\lambda>0$,
$\overline{\xx}\cdot{\bf u}\leq 0$ if and only if $\lambda\overline{\xx}\cdot{\bf u}\leq 0$,
by Proposition~\ref{pr:i:2}), so $\apol{K}$ is also a cone
(by Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:a})).
That the origin is in $\apol{K}$ is immediate.
\qed
A basic property of the standard polar cone is that the polar of a
convex cone $K$ is the same as the polar of its closure, that is,
$\polar{K}=\polar{(\cl K)}$.
This property does not hold, in general, for the astral polar, meaning
that it is possible that
$\apol{K}$ differs from $\apol{(\cl K)}$.
For example, in ${\mathbb{R}}^2$, suppose $K$ is the open left halfplane:
$K = \{(u_1,u_2) \in {\mathbb{R}}^2 : u_1 < 0\}$.
The standard polar of this set is the ray
$\polar{K} = \{ \lambda {\bf e}_1 : \lambda \geq 0\}$.
Its astral polar, $\apol{K}$, includes $\polar{K}$, and also includes all
infinite points in $\extspac{2}$ whose dominant direction is ${\bf e}_1$.
That is because, for every such point $\overline{\xx}$,
$\overline{\xx}\cdot{\bf u}=\limray{{\bf e}_1}\cdot{\bf u}=-\infty$, for all ${\bf u}\in K$.
Thus,
\[
\apol{K} = \{ \lambda {\bf e}_1 : \lambda \geq 0\}
\cup [ \limray{{\bf e}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \extspac{2} ].
\]
On the other hand,
the astral polar $\apol{(\cl K)}$ of
$\cl K$, the closed left halfplane of ${\mathbb{R}}^2$, includes only
$\polar{K}=\polar{(\cl K)}$
together with the single infinite point $\limray{{\bf e}_1}$:
\[
\apol{(\cl K)}
=
\{ \lambda {\bf e}_1 : \lambda \geq 0\} \cup \{\limray{{\bf e}_1}\}.
\]
Note that this set is exactly the closure (in $\extspac{2}$) of
$\polar{K}$.
Indeed, this latter observation turns out to be general:
If $K\subseteq{\R^n}$ is a pointed convex cone, then
$K$ is closed (in ${\R^n}$) if and only if
its astral polar $\apol{K}$ is always the same as the closure
(in $\extspac{n}$) of its standard polar $\polar{K}$
(so that $\apol{K}=\overline{\Kpol}$).
This will be central
in characterizing the continuity of a
convex function's extension, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
in Section~\ref{subsec:charact:cont:duality}.
We first prove one direction of this result:
\begin{theorem} \label{thm:k-closed-implies-closed-apol}
Let $K\subseteq{\R^n}$ be a convex cone.
Then $\overline{\Kpol}\subseteq\apol{K}$.
If, in addition, $K$ is closed in ${\R^n}$, then
$\overline{\Kpol} = \apol{K}$.
\end{theorem}
\proof
By
Proposition~\ref{pr:ast-pol-props}(\ref{pr:ast-pol-props:a},\ref{pr:ast-pol-props:c}),
$\polar{K}=\apol{K}\cap{\R^n}\subseteq\apol{K}$
and
$\apol{K}$ is closed in $\extspac{n}$.
Therefore,
$\overline{\Kpol}\subseteq\apol{K}$.
For the rest of the proof, assume $K$ is closed in ${\R^n}$.
Under this assumption, we prove the reverse inclusion,
$\apol{K}\subseteq \overline{\Kpol}$.
By Theorem~\ref{thm:e:6} combined with
Proposition~\ref{pr:ohull-simplify}, $\overline{\Kpol}$ consists of all those
points $\overline{\xx}\in\extspac{n}$ for which
$\overline{\xx}\cdot{\bf u}\leq \indstar{\polar{K}}({\bf u})$,
where $\indstar{\polar{K}}$ is the support function for $\polar{K}$, as
defined in \eqref{eq:e:2}.
Further, $\indstar{\polar{K}}$ is the conjugate of the indicator function
$\indf{\polar{K}}$, given in \eqref{eq:indf-defn}.
It is known that the indicator functions of a closed convex cone and
its polar are conjugate to each other
\citep[Theorem~14.1]{ROC}.
Therefore, $\indstar{\polar{K}}=\indf{K}$, since $K$ is closed.
Now let $\overline{\xx}$ be any point in $\apol{K}$, which we aim to show is in
$\overline{\Kpol}$.
The preceding argument means that
to prove $\overline{\xx}\in\overline{\Kpol}$, it suffices to show that
$\overline{\xx}\cdot{\bf u}\leq\indf{K}({\bf u})$ for all ${\bf u}\in{\R^n}$.
Indeed,
if ${\bf u}\in K$ then
$\overline{\xx}\cdot{\bf u}\leq 0= \indf{K}({\bf u})$,
since $\overline{\xx}\in\apol{K}$.
Otherwise, if ${\bf u}\not\in K$, then
$\overline{\xx}\cdot{\bf u}\leq +\infty = \indf{K}({\bf u})$.
Thus, $\overline{\xx}\in\overline{\Kpol}$, proving $\apol{K}\subseteq\overline{\Kpol}$, and
completing the proof.
\qed
As a corollary, this implies that for any convex cone $K\subseteq{\R^n}$
(not necessarily closed), the astral polar of the standard polar
of $K$ is the same as the closure of $K$, that is,
$\polapol{K} = \apol{(\polar{K})} = \overline{K}$,
analogous to the standard result that
$\dubpolar{K}=\cl K$
\citep[Theorem~14.1]{ROC}.
\begin{corollary} \label{cor:kpolapol-is-kbar}
Let $K\subseteq{\R^n}$ be a convex cone.
Then $\polapol{K} = \overline{K}$.
\end{corollary}
\proof
The polar $\polar{K}$ is a closed convex cone, so applying
Theorem~\ref{thm:k-closed-implies-closed-apol}
to $\polar{K}$ yields
$\polapol{K}=\overline{\dubpolar{K}}=\overline{(\cl K)}=\overline{K}$
(using
Proposition~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:aa})).
\qed
As discussed above, for a pointed convex cone,
the condition of being closed in ${\R^n}$ is both necessary and
sufficient for its astral polar to be the closure of its standard
polar.
We now state this as a theorem:
\begin{theorem} \label{thm:k-closed-equals-closed-apol}
Let $K\subseteq{\R^n}$ be a pointed convex cone.
Then $\apol{K}=\overline{\Kpol}$ if and only if $K$ is closed in ${\R^n}$.
\end{theorem}
This theorem implies that if $K\subseteq{\R^n}$ is a convex cone, then
$\apol{(\cl K)}=\overline{\polar{(\cl K)}}=\overline{\Kpol}$, and therefore,
$\apol{K}=\apol{(\cl K)}$ if and only if $K$ is already closed in
${\R^n}$.
The proof of the theorem
will be by induction on the dimension of the cone $K$, and
as such, we will consider intersecting the cone with a linear
subspace, and how this affects its polar.
We prove two lemmas along these lines.
The first provides some basic, general facts:
\begin{lemma} \label{lem:kpol-lin-subs:1}
Let $K\subseteq{\R^n}$ be a pointed convex cone.
Let ${\bf v}\in\polar{K}$.
Also, let
\[ M = \{ {\bf u}\in{\R^n} : {\bf u}\cdot{\bf v} = 0 \} \]
be the linear subspace perpendicular to ${\bf v}$, and
let $J = \{\lambda {\bf v} : \lambda\in{\mathbb{R}} \}$.
Then
\begin{enumerate}
\item \label{lem:kpol-lin-subs:1:a}
$
\cl (K\cap M)
\subseteq
\polar{(\polar{K} + J)}
=
(\cl K) \cap M
$.
\item \label{lem:kpol-lin-subs:1:b}
$
\polar{(K\cap M)}
\supseteq
\cl(\polar{K} + J)
$.
\end{enumerate}
\end{lemma}
\proof
First,
$K\cap M\subseteq K$ so
$\cl(K\cap M)\subseteq \cl K$.
Likewise,
$\cl(K\cap M)\subseteq M$ since
$K\cap M\subseteq M$ and $M$ is closed.
Thus,
$\cl (K\cap M) \subseteq (\cl K) \cap M$.
Next,
\[\polar{(\polar{K} + J)} = \dubpolar{K}\cap \polar{J} = (\cl K) \cap M.\]
The first equality uses a general rule for the polar of the sum of
two convex cones
\citep[Corollary~16.4.2]{ROC}.
Also, $\polar{J}=M$
since a point ${\bf u}\in{\R^n}$ is in
$\polar{J}$ if and only if
${\bf u}\cdot(\lambda {\bf v})\leq 0$ for all $\lambda\in{\mathbb{R}}$,
which holds if and only if ${\bf u}\cdot{\bf v}=0$.
This proves part~(\ref{lem:kpol-lin-subs:1:a}).
Taking polars of both sides now yields
\begin{eqnarray*}
\polar{(K\cap M)}
&=&
\polar{(\cl(K\cap M))}
\\
&\supseteq&
\dubpolar{(\polar{K} + J)}
\\
&=&
\cl{(\polar{K} + J)},
\end{eqnarray*}
proving part~(\ref{lem:kpol-lin-subs:1:b}).
\qed
When $\apol{K}=\overline{\Kpol}$, we can prove that these inclusions hold with
equality, and can also prove that the same property holds for
$K\cap M$.
\begin{lemma} \label{lem:kpol-lin-subs:2}
Let $K$, ${\bf v}$, $J$ and $M$ be as defined in
Lemma~\ref{lem:kpol-lin-subs:1}.
In addition, assume that $\apol{K}=\overline{\Kpol}$.
Then the following hold:
\begin{enumerate}
\item \label{lem:kpol-lin-subs:2:a}
$\apol{(K\cap M)} = \overline{\polar{(K\cap M)}}$.
\item \label{lem:kpol-lin-subs:2:b}
$\cl (K\cap M) = \polar{(\polar{K} + J)} = (\cl K) \cap M$.
\item \label{lem:kpol-lin-subs:2:c}
$\polar{(K\cap M)} = \cl(\polar{K} + J)$.
\end{enumerate}
\end{lemma}
\proof
Let $K'=K\cap M$.
The key step in the proof is to show that
\begin{equation} \label{eqn:lem:kpol-lin-subs:2:1}
\apol{K'}\subseteq \overline{(\polar{K} + J)}.
\end{equation}
Once this is proved, all three parts of the lemma will follow
directly, as shown below.
Let $\overline{{\bf z}}\in\apol{K'}$, which we aim to show is in
$\overline{(\polar{K} + J)}$.
Let $\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$.
We claim that $\overline{\xx}\in\apol{K}$.
To see this, let ${\bf u}\in K$;
we aim to show $\overline{\xx}\cdot{\bf u}\leq 0$.
Then ${\bf v}\cdot{\bf u}\leq 0$ since ${\bf v}\in\polar{K}$.
If ${\bf v}\cdot{\bf u}<0$, then
$\limray{{\bf v}}\cdot{\bf u}=-\infty$ so
$\overline{\xx}\cdot{\bf u}=-\infty$.
Otherwise, if ${\bf v}\cdot{\bf u}=0$, then ${\bf u}\in M$, implying
${\bf u}\in K\cap M = K'$.
Also, $\limray{{\bf v}}\cdot{\bf u}=0$, so
$\overline{\xx}\cdot{\bf u}=\overline{{\bf z}}\cdot{\bf u}\leq 0$ since $\overline{{\bf z}}\in\apol{K'}$.
Thus, $\overline{\xx}\cdot{\bf u}\leq 0$ for all ${\bf u}\in K$.
Therefore, $\overline{\xx}\in\apol{K}=\overline{\Kpol}$, which means there exists a
sequence
$\seq{{\bf x}_t}$ in $\polar{K}$ such that ${\bf x}_t\rightarrow\overline{\xx}$.
For each $t$, we can write
${\bf x}_t = b_t {\bf v} + {\xx_t^\bot}$
for some $b_t\in{\mathbb{R}}$, where $\ybar^{\bot}$ denotes the projection of
$\overline{{\bf y}}$ perpendicular to ${\bf v}$.
Then ${\xx_t^\bot}\rightarrow\xbar^{\bot}$
(by Proposition~\ref{pr:h:5}(\ref{pr:h:5b})).
Also, ${\xx_t^\bot}={\bf x}_t - b_t {\bf v}$, which is in
$\polar{K}+J$, since ${\bf x}_t\in\polar{K}$.
Therefore,
$\xbar^{\bot}\in \overline{(\polar{K} + J)}$.
Furthermore, $\xbar^{\bot}=(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}})^\bot=\zbar^{\bot}$
(by
Proposition~\ref{pr:h:5}(\ref{pr:h:5c},\ref{pr:h:5e})).
Thus,
\begin{equation} \label{eqn:lem:kpol-lin-subs:2:2}
\zbar^{\bot}\in \overline{(\polar{K} + J)}= \polapol{(\polar{K} + J)},
\end{equation}
with the equality following from Corollary~\ref{cor:kpolapol-is-kbar}.
We claim $\overline{{\bf z}}$ is also in this set.
Let ${\bf u}\in\polar{(\polar{K} + J)}$.
Then
\[
\overline{{\bf z}}\cdot{\bf u}
=
\overline{{\bf z}}\cdot{\uu^\bot}
=
\zbar^{\bot}\cdot{\bf u}
\leq
0.
\]
The first equality is because,
by Lemma~\ref{lem:kpol-lin-subs:1}(\ref{lem:kpol-lin-subs:1:a}),
$\polar{(\polar{K} + J)}= (\cl K) \cap M\subseteq M$,
so ${\bf u}\in M$, implying ${\uu^\bot}={\bf u}$.
The second equality is by
Proposition~\ref{pr:h:5}(\ref{pr:h:5a}).
And the inequality is
by \eqref{eqn:lem:kpol-lin-subs:2:2}.
We conclude that
$ \overline{{\bf z}}\in \overline{(\polar{K} + J)}$,
proving \eqref{eqn:lem:kpol-lin-subs:2:1}.
We thus have that
\[
\apol{K'}
\subseteq
\overline{(\polar{K} + J)}
\subseteq
\overline{\polar{K'}}
\subseteq
\apol{K'}.
\]
The inclusions follow respectively from
\eqref{eqn:lem:kpol-lin-subs:2:1},
Lemma~\ref{lem:kpol-lin-subs:1}(\ref{lem:kpol-lin-subs:1:b}),
and
Theorem~\ref{thm:k-closed-implies-closed-apol}.
Therefore, these hold with equality, that is,
\[
\apol{K'}
=
\overline{(\polar{K} + J)}
=
\overline{\polar{K'}},
\]
proving part~(\ref{lem:kpol-lin-subs:2:a}).
Taking intersections with ${\R^n}$ now yields
\[
\polar{K'}
=
\apol{K'} \cap {\R^n}
=
\overline{(\polar{K} + J)} \cap {\R^n}
=
\cl{(\polar{K} + J)},
\]
with the first and third equalities following from
Proposition~\ref{pr:ast-pol-props}(\ref{pr:ast-pol-props:a})
and
Proposition~\ref{pr:closed-set-facts}(\ref{pr:closed-set-facts:a}).
This proves part~(\ref{lem:kpol-lin-subs:2:c}).
Finally, taking (standard) polars of both sides yields
\[
\cl K'
=
\dubpolar{K'}
=
\polar{(\cl(\polar{K}+J))}
=
\polar{(\polar{K}+J)},
\]
proving part~(\ref{lem:kpol-lin-subs:2:b}).
\qed
\proof[Proof of Theorem~\ref{thm:k-closed-equals-closed-apol}]
That $K$ being closed implies $\apol{K}=\overline{\Kpol}$ was proved in
Theorem~\ref{thm:k-closed-implies-closed-apol}.
We prove the converse
by induction on the dimension of $K$.
More precisely, we prove by induction on
$k=0,\ldots,n$ that for any pointed convex cone $K\subseteq{\R^n}$,
if $\apol{K}=\overline{\Kpol}$ and
if $K$ is included in some linear subspace
$L\subseteq{\R^n}$ of dimension at most $k$,
then $K$ is closed in ${\R^n}$.
In the base case that $k=0$,
if $K$ is included in a linear subspace $L$ of
dimension $0$, then we must have
$L=\{{\bf 0}\}$,
so $K=\{{\bf 0}\}$, which is closed.
For the inductive step, assume $k>0$ and that
$K$ is a pointed convex cone that is
included in some linear subspace $L$ of dimension at most $k$.
Let ${\bf u}$ be any point in $\cl K$, the closure of $K$ in ${\R^n}$;
we aim to show that ${\bf u}$ must be in $K$.
Furthermore, we can assume without loss of generality that ${\bf u}$ is on
the relative boundary of $K$,
that is, that ${\bf u}\in\cl K\setminus \relint{K}$,
since clearly $\relint{K}\subseteq K$.
Since ${\bf u}$ is a relative boundary point of the pointed convex cone $K$,
we can apply Lemma~\ref{lem:rel-bnd-convcone} which shows there
exists ${\bf v}\in{\R^n}$ with ${\bf u}\cdot{\bf v}=0$ and
${\bf x}\cdot{\bf v}\leq 0$ for all ${\bf x}\in K$, meaning
${\bf v}\in\polar{K}$.
In addition, there exists some point ${\bf y}\in K$ with ${\bf y}\cdot{\bf v}<0$.
Let $M=\{{\bf x}\in{\R^n} : {\bf x}\cdot{\bf v} = 0\}$,
and let
$K' = K\cap M$.
Since both $K$ and $M$ are pointed convex cones,
$K'$ is as well.
Also,
$K'\subseteq L'$ where
$L'=L\cap M$ is a linear subspace.
Since ${\bf y}\cdot{\bf v}<0$, ${\bf y}$ is not in $M$, and so also is not in
$L'$.
On the other hand,
${\bf y}\in K \subseteq L$.
Thus, $L'\subseteq L$ but $L'\neq L$; therefore,
$\dim{L'}<\dim{L}\leq k$.
Since $\apol{K}=\overline{\Kpol}$ and ${\bf v}\in\polar{K}$,
Lemma~\ref{lem:kpol-lin-subs:2}(\ref{lem:kpol-lin-subs:2:a})
can be applied, yielding
$\apol{K'}=\overline{\polar{K'}}$.
Since $\dim{L'}\leq k-1$, we can therefore apply our inductive
hypothesis, yielding that $K'$ is closed in ${\R^n}$.
Since ${\bf u}\cdot{\bf v}=0$ (so that ${\bf u}\in M$),
we now have that
\[
{\bf u}
\in
(\cl K) \cap M
=
\cl K'
=
K' = K\cap M \subseteq K.
\]
The first equality is by
Lemma~\ref{lem:kpol-lin-subs:2}(\ref{lem:kpol-lin-subs:2:b}).
The second is because $K'$ is closed in ${\R^n}$.
Thus, ${\bf u}\in K$, completing the induction and the proof.
\qed
\section{Convex functions}
\label{sec:conv:fct}
Having defined and studied convex sets in astral space, we next
consider how the notion of a convex function can be extended using
similar ideas.
Recall that a function $f:{\R^n}\rightarrow\overline{\R}$ is convex if its
epigraph, which is a subset of ${\R^n}\times{\mathbb{R}}={\R^{n+1}}$, is convex.
We would like to define what it means analogously for a function on
$\extspac{n}$ to be convex.
Let $F:\extspac{n}\rightarrow\overline{\R}$.
Its epigraph was previously defined in \eqref{eqn:epi-def} to be the
set of pairs
\[
\epi{F} = \braces{\rpair{\overline{\xx}}{y}\in\extspac{n}\times{\mathbb{R}} :
F(\overline{\xx}) \leq y}.
\]
This is a subset of $\extspac{n}\times{\mathbb{R}}$.
Informally, we would like to regard this latter space as a subset of
$\extspac{n+1}$, so that $\epi{F}$ also becomes a subset of
$\extspac{n+1}$ where convexity has already been defined.
It would then be natural to say that $F$ is a convex function if
its epigraph $\epi{F}$, as a subset of $\extspac{n+1}$, is
astrally convex.
At an intuitive level,
this is precisely the approach that we follow.
First, however, we need to overcome a modest technical hurdle, namely,
that $\extspac{n}\times{\mathbb{R}}$ need not be a subset of
$\extspac{n+1}$ according to our formal definitions.
To handle this, we embed
$\extspac{n}\times{\mathbb{R}}$ in $\extspac{n+1}$, that is, show that it is
homeomorphic with a subset of the larger space.
As usual,
we write pairs in $\extspac{n}\times{\mathbb{R}}$
as $\rpair{\overline{\xx}}{y}$.
If $\overline{\xx}={\bf x}$ happens to be in ${\R^n}$, we regard $\rpair{{\bf x}}{y}$ either as a
pair in ${\R^n}\times{\mathbb{R}}$ or as an equivalent vector in ${\R^{n+1}}$ (specifically,
a column vector, as is standard, despite how it is written).
Intuitively, $\extspac{n}\times{\mathbb{R}}$ should be identified with those
points $\overline{{\bf z}}$ in $\extspac{n+1}$ whose ``last coordinate''
(corresponding to $y$) is in ${\mathbb{R}}$.
We can extract that coordinate from $\overline{{\bf z}}$ by taking its inner
product with $\rpair{{\bf 0}}{1}$, the vector in ${\R^{n+1}}$ that is all zeros except
the last coordinate which is $1$.
Thus, the set of all such points is
\begin{equation} \label{eq:finclset-defn}
{{\cal M}_n} = \{ \overline{{\bf z}}\in \extspac{n+1} :
\overline{{\bf z}}\cdot \rpair{{\bf 0}}{1}\in {\mathbb{R}} \}.
\end{equation}
Indeed, $\extspac{n}\times{\mathbb{R}}$ and ${{\cal M}_n}$ are homeomorphic in a
natural way
that maps each point in ${\R^n}\times{\mathbb{R}}={\R^{n+1}}$ to itself, as we show in
the next theorem.
In what follows, ${\bf P}$ denotes the $n\times (n+1)$ matrix
whose first $n$ columns form the $n\times n$ identity matrix ${\bf I}$,
and whose last column is ${\bf 0}$, the all-zeros vector in ${\R^n}$.
Thus, in block form,
\begin{equation} \label{eqn:homat-def}
{\bf P}
=
\left[
\begin{array}{ccc|c}
& & & \\
~ & {\bf I} & ~ & {\bf 0} \\
& & & \\
\end{array}
\right].
\end{equation}
Note that ${\bf P} \rpair{{\bf x}}{y} = {\bf x}$
for all ${\bf x}\in{\R^n}$ and $y\in{\mathbb{R}}$.
As a result,
multiplying by ${\bf P}$ has the effect of extracting the first $n$
elements of a vector in ${\mathbb{R}}^{n+1}$.
\begin{theorem} \label{thm:homf}
Define
$\mu:\extspac{n}\times{\mathbb{R}}\rightarrow{{\cal M}_n}$ to be the function
\[
\mu(\overline{\xx},y) = \trans{{\bf P}} \overline{\xx} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \rpair{{\bf 0}}{y}
\]
for $\overline{\xx}\in\extspac{n}$, $y\in{\mathbb{R}}$.
Then $\mu$ has the following properties:
\begin{enumerate}
\item \label{thm:homf:a}
For all $\overline{\xx}\in\extspac{n}$, $y\in{\mathbb{R}}$,
\[
\mu(\overline{\xx},y) \cdot \rpair{{\bf u}}{v} = \overline{\xx}\cdot{\bf u} + y v
\]
for all $\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}={\R^{n+1}}$
(implying, in particular, that $\mu(\overline{\xx},y)$ is indeed in
${{\cal M}_n}$).
\item \label{thm:homf:aa}
$\mu$ is bijective with inverse
\begin{equation} \label{eqn:homfinv-def}
\mu^{-1}(\overline{{\bf z}}) = \rpair{{\bf P}\overline{{\bf z}}}{\;\;\overline{{\bf z}}\cdot\rpair{{\bf 0}}{1}}
\end{equation}
for $\overline{{\bf z}}\in\extspac{n+1}$.
\item \label{thm:homf:b}
$\mu$ is a homeomorphism (that is, both
$\mu$ and its inverse are continuous).
\item \label{thm:homf:c}
$\mu({\bf x},y)=\rpair{{\bf x}}{y}$ for all $\rpair{{\bf x}}{y}\in{\R^n}\times{\mathbb{R}}$.
\end{enumerate}
\end{theorem}
\proof
~
Part~(\ref{thm:homf:a}):
Let $\overline{\xx}\in\extspac{n}$, $y\in{\mathbb{R}}$, and
$\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}$.
Then
\begin{eqnarray*}
\mu(\overline{\xx},y) \cdot \rpair{{\bf u}}{v}
&=&
(\trans{{\bf P}} \overline{\xx}) \cdot \rpair{{\bf u}}{v}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
\rpair{{\bf 0}}{y} \cdot \rpair{{\bf u}}{v}
\\
&=&
\overline{\xx} \cdot ({\bf P} \rpair{{\bf u}}{v})
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
y v
\\
&=&
\overline{\xx} \cdot {\bf u}
+
y v.
\end{eqnarray*}
The first equality is by $\mu$'s definition and
Proposition~\ref{pr:i:6};
the second by Proposition~\ref{pr:mat-mult-def};
the third by ${\bf P}$'s definition (and since $y v\in{\mathbb{R}}$).
Taking $\rpair{{\bf u}}{v}=\rpair{{\bf 0}}{1}$ then shows that
$\mu(\overline{\xx},y)\in{{\cal M}_n}$.
Part~(\ref{thm:homf:aa}):
Let $\mu^{-1}$ be as given in \eqref{eqn:homfinv-def}.
Let $\overline{\xx}\in\extspac{n}$ and $y\in{\mathbb{R}}$.
We first show
$ \mu^{-1}(\mu(\overline{\xx},y)) = \rpair{\overline{\xx}}{y} $.
Let $\overline{{\bf z}}=\mu(\overline{\xx},y)$.
Then by $\mu$'s definition,
\[
{\bf P} \overline{{\bf z}}
=
{\bf P} \trans{{\bf P}} \overline{\xx}
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax
{\bf P} \rpair{{\bf 0}}{y}
=
\overline{\xx}
\]
since ${\bf P}\trans{{\bf P}}$ is the identity matrix
and
${\bf P} \rpair{{\bf 0}}{y} = {\bf 0}$.
That $\overline{{\bf z}}\cdot \rpair{{\bf 0}}{1}=y$ follows from
part~(\ref{thm:homf:a}).
Thus, $\mu^{-1}$ is a left inverse of $\mu$.
Next, let $\overline{{\bf z}}\in{{\cal M}_n}$,
and let
$\overline{{\bf z}}'=\mu(\mu^{-1}(\overline{{\bf z}}))$.
We aim to show $\overline{{\bf z}}'=\overline{{\bf z}}$.
Let $\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}$.
Then
\begin{eqnarray}
\overline{{\bf z}}'\cdot\rpair{{\bf u}}{v}
&=&
({\bf P}\overline{{\bf z}})\cdot{\bf u}
+
(\overline{{\bf z}}\cdot\rpair{{\bf 0}}{1}) v
\nonumber
\\
&=&
\overline{{\bf z}}\cdot(\trans{{\bf P}} {\bf u})
+
\overline{{\bf z}}\cdot\rpair{{\bf 0}}{v}
\nonumber
\\
&=&
\overline{{\bf z}}\cdot\rpair{{\bf u}}{0}
+
\overline{{\bf z}}\cdot\rpair{{\bf 0}}{v}
\nonumber
\\
&=&
\overline{{\bf z}}\cdot\rpair{{\bf u}}{v}.
\label{eqn:thm:homf:1}
\end{eqnarray}
The first equality is by
\eqref{eqn:homfinv-def}
combined with part~(\ref{thm:homf:a}),
noting that
$\overline{{\bf z}}\cdot\rpair{{\bf 0}}{1}\in{\mathbb{R}}$
since $\overline{{\bf z}}\in{{\cal M}_n}$.
The second equality is by
Propositions~\ref{pr:mat-mult-def}
and~\ref{pr:i:2}.
The third is a simple matrix calculation.
And the last is by Proposition~\ref{pr:i:1}
(since $\overline{{\bf z}}\cdot\rpair{{\bf 0}}{v}\in{\mathbb{R}}$).
Since \eqref{eqn:thm:homf:1} holds for all
$\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}$, this implies
$\overline{{\bf z}}'=\overline{{\bf z}}$
(by Proposition~\ref{pr:i:4}).
Thus, $\mu^{-1}$ is also a right inverse of $\mu$.
Therefore, $\mu$ is bijective with inverse
$\mu^{-1}$, as claimed.
Part~(\ref{thm:homf:b}):
The function $\mu$ is continuous since if
$\seq{\rpair{\overline{\xx}_t}{y_t}}$ is any sequence in $\extspac{n}\times{\mathbb{R}}$
that converges to $\rpair{\overline{\xx}}{y}\in\extspac{n}\times{\mathbb{R}}$,
then for all $\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}$,
\[
\mu(\overline{\xx}_t,y_t)\cdot\rpair{{\bf u}}{v}
=
\overline{\xx}_t\cdot{\bf u} + v y_t
\rightarrow
\overline{\xx}\cdot{\bf u} + v y
=
\mu(\overline{\xx},y)\cdot\rpair{{\bf u}}{v},
\]
by Theorem~\ref{thm:i:1}(\ref{thm:i:1c}) and
continuity
(and the equalities following from
part~(\ref{thm:homf:a})).
This implies
$\mu(\overline{\xx}_t,y_t)\rightarrow\mu(\overline{\xx},y)$,
again by Theorem~\ref{thm:i:1}(\ref{thm:i:1c}).
The function $\mu^{-1}$ is continuous by
Proposition~\ref{pr:aff-cont}
and
Theorem~\ref{thm:i:1}(\ref{thm:i:1c}).
Thus, $\mu$ is a homeomorphism.
Part~(\ref{thm:homf:c}):
Let ${\bf x}\in{\R^n}$, $y\in{\mathbb{R}}$, and $\overline{{\bf z}}=\mu({\bf x},y)$.
Then
$\overline{{\bf z}}\cdot\rpair{{\bf u}}{v}=\rpair{{\bf x}}{y}\cdot\rpair{{\bf u}}{v}$
for all $\rpair{{\bf u}}{v}\in{\R^n}\times{\mathbb{R}}$, by
part~(\ref{thm:homf:a}).
Combined with Proposition~\ref{pr:i:4},
this implies $\overline{{\bf z}}=\rpair{{\bf x}}{y}$.
(Alternatively, this could be proved directly from the definition of
$\mu$.)
\qed
For the rest of this section, $\mu$ denotes the function given in
Theorem~\ref{thm:homf}.
Thus, points and sets in $\extspac{n}\times{\mathbb{R}}$ can very much be
identified with their image under $\mu$ in
${{\cal M}_n}\subseteq\extspac{n+1}$.
In particular, we define a set $S\subseteq\extspac{n}\times{\mathbb{R}}$ to be
astrally convex if its image $\mu(S)$ is astrally convex in
$\extspac{n+1}$.
Further, we can now finally define a function
$F:\extspac{n}\rightarrow\overline{\R}$
to be \emph{astrally convex} (or simply, \emph{convex})
if its epigraph, $\epi{F}$, is astrally convex.
We will see soon that the main astral functions we have been
studying most closely are convex.
To that end, we next relate the closure
of the epigraph of a funtion $f:{\R^n}\rightarrow\overline{\R}$, to the epigraph
of its extentions ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
We write $\clgen{{\mathrm m}}{S}$ for the closure in $\extspac{n}\times{\mathbb{R}}$
of any set $S$ in that space.
(For a set
$S\subseteq\extspac{n+1}$, we continue to write $\overline{S}$ for the
closure of $S$ in $\extspac{n+1}$.)
The next proposition shows that
$\epi {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, the epigraph of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
is exactly equal to
$\epibar{f}$, the closure of $\epi{f}$ in $\extspac{n}\times{\mathbb{R}}$.
Furthermore, $\mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})$, its homeomorphic image in
$\extspac{n+1}$, is equal to
$\epi{f}$'s closure in $\extspac{n+1}$,
restricted to the set ${{\cal M}_n}$.
As a consequence, these imply that the closures of the epigraphs of
$f$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ are (homeomorphically) the same:
\begin{proposition} \label{pr:wasthm:e:3}
Let $f:{\R^n}\rightarrow\overline{\R}$, and let ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ be its
extension.
Then the following hold:
\begin{enumerate}
\item \label{pr:wasthm:e:3a}
$\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\epibar{f}$.
\item \label{pr:wasthm:e:3b}
$\mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) = \epibarbar{f} \cap {{\cal M}_n}$.
\item \label{pr:wasthm:e:3c}
$\overline{\mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})} = \epibarbar{f}$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:wasthm:e:3a}):
Suppose $\rpair{\overline{\xx}}{y}\in \epibar{f}$.
Then there exists a sequence $\seq{\rpair{{\bf x}_t}{y_t}}$
in ${\R^n}\times {\mathbb{R}}$
converging to $\rpair{\overline{\xx}}{y}$ with $f({\bf x}_t)\leq y_t$
for all $t$.
Thus,
\[
y = \lim y_t \geq \liminf f({\bf x}_t) \geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
\]
where the second inequality is by definition of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Thus, $\rpair{\overline{\xx}}{y}\in \epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, and
$\epibar{f} \subseteq \epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Now suppose $\rpair{\overline{\xx}}{y}\in \epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, meaning
$y\geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$ and $y\in{\mathbb{R}}$.
As in Proposition~\ref{pr:d1},
let $\seq{{\bf x}_t}$ in ${\R^n}$ be a sequence converging to $\overline{\xx}$ and for which
$f({\bf x}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\leq y<+\infty$, we discard those $t$ for
which $f({\bf x}_t)=+\infty$, of which there can only be finitely many.
Let $y_t=\max\{y,f({\bf x}_t)\} \in {\mathbb{R}}$ so that $\rpair{{\bf x}_t}{y_t}\in\epi{f}$.
Then $\rpair{{\bf x}_t}{y_t}\rightarrow \rpair{\overline{\xx}}{y}$, by continuity, and since
$y\geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Thus, $\rpair{\overline{\xx}}{y}\in\epibar{f}$, and
$\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} \subseteq \epibar{f}$.
Part~(\ref{pr:wasthm:e:3b}):
We relate $\epibar{f}$,
the closure of $\epi{f}$ in $\extspac{n}\times{\mathbb{R}}$,
to the closure of $\epi{f}$ in two other sets.
Specifically, since $\epi{f}$ is a subset of
${\R^{n+1}}\subseteq{{\cal M}_n}\subseteq\extspac{n+1}$,
we consider
$\epibarbar{f}$,
the closure of $\epi{f}$ in $\extspac{n+1}$,
and also
the closure of $\epi{f}$ in ${{\cal M}_n}$, which
we denote by $E$.
Because
$\epi{f}$ is in ${{\cal M}_n}$, a subspace of $\extspac{n+1}$,
its closure in ${{\cal M}_n}$ is equal to its closure in $\extspac{n+1}$
intersected with ${{\cal M}_n}$.
That is,
$E = \epibarbar{f} \cap {{\cal M}_n}$.
(See \citet[Theorem~17.4]{munkres}.)
On the other hand, $\epi{f}$ is a subset of ${\R^n}\times{\mathbb{R}}$
and so is equal to $\mu(\epi{f})$, its own image
under $\mu$,
by Theorem~\ref{thm:homf}(\ref{thm:homf:c}).
Thus, $E$ is the closure in ${{\cal M}_n}$ of
$\mu(\epi{f})=\epi f$.
Because $\mu$ is a homeomorphism, this implies that $E$,
the closure of the image of
$\epi{f}$ under $\mu$, is equal to the image of its closure in
$\extspac{n}\times{\mathbb{R}}$, namely,
$\mu(\epibar{f})$.
In other words,
$E=\mu(\epibar{f})$.
Combining the above observations with part~(\ref{pr:wasthm:e:3a})
(and since $\mu$ is a bijection) yields
\[
\mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})
=
\mu(\epibar{f})
=
E
=
\epibarbar{f} \cap {{\cal M}_n},
\]
as claimed.
Part~(\ref{pr:wasthm:e:3c}):
By Proposition~\ref{pr:h:1}(\ref{pr:h:1a}),
if $\rpair{{\bf x}}{y}\in\epi f$, then
$y\geq f({\bf x})\geq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})$, so
$\rpair{{\bf x}}{y}$ is also in $\epi {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Thus, $\epi f = \mu(\epi f) \subseteq \mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})$
(by Theorem~\ref{thm:homf}(\ref{thm:homf:aa},\ref{thm:homf:c})),
so
$\epibarbar{f} \subseteq \overline{\mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})}$.
For the reverse inclusion, part~(\ref{pr:wasthm:e:3b}) immediately
implies
$\mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \subseteq \epibarbar{f}$, yielding
$\overline{\mu(\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})} \subseteq \epibarbar{f}$
since $\epibarbar{f}$ is closed (in $\extspac{n+1}$).
\qed
Using Proposition~\ref{pr:wasthm:e:3},
we can now prove that the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of any convex function
$f$ is also convex:
\begin{theorem} \label{thm:fext-convex}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, its lower semicontinuous extension, is also convex.
\end{theorem}
\proof
Since $f$ is convex, its epigraph $\epi{f}$ is convex,
so $\epibarbar{f}$, its closure in $\extspac{n+1}$, is convex by
Theorem~\ref{thm:e:6}.
Furthermore, the set
\[
{{\cal M}_n} =
\{\overline{{\bf z}}\in\extspac{n+1} : \overline{{\bf z}}\cdot \rpair{{\bf 0}}{1} > -\infty \}
\cap
\{\overline{{\bf z}}\in\extspac{n+1} : \overline{{\bf z}}\cdot \rpair{{\bf 0}}{1} < +\infty \}
\]
is convex by
Proposition~\ref{pr:e1}(\ref{pr:e1:b},\ref{pr:e1:c}).
Thus, $\epibarbar{f} \cap {{\cal M}_n}$ is convex, implying
that $\epi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is convex
by Proposition~\ref{pr:wasthm:e:3}(\ref{pr:wasthm:e:3b})
and definition of convexity for sets
in $\extspac{n}\times{\mathbb{R}}$.
Therefore, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is also convex.
\qed
As is the case in standard convex analysis, the pointwise supremum of
any collection of convex functions is convex:
\begin{proposition} \label{pr:point-sup-is-convex}
Let $G_\alpha:\extspac{n}\rightarrow\overline{\R}$ be convex
for all $\alpha\in{\cal I}$, where
${\cal I}$ is any nonempty index set.
Let $F$ be their pointwise supremum, that is,
\[
F(\overline{\xx}) = \sup_{\alpha\in{\cal I}} G_\alpha(\overline{\xx})
\]
for $\overline{\xx}\in\extspac{n}$.
Then $F$ is convex.
\end{proposition}
\proof
The epigraph of $F$, $\epi{F}$, is exactly the intersection of the epigraphs of
the functions $G_\alpha$, that is,
\[
\epi{F} = \bigcap_{\alpha\in{\cal I}} \epi{G_\alpha}.
\]
This is because a pair $\rpair{\overline{\xx}}{y}$ is in $\epi{F}$, meaning
$y\geq F(\overline{\xx})$,
if and only if
$y\geq G_\alpha(\overline{\xx})$
for all $\alpha\in{\cal I}$,
that is, if and only if
$\rpair{\overline{\xx}}{y}$ is in $\epi{G_\alpha}$
for all $\alpha\in{\cal I}$.
Since $G_\alpha$ is convex, its epigraph, $\epi{G_\alpha}$, is convex,
for $\alpha\in{\cal I}$.
Thus, $\epi{F}$ is convex by
Proposition~\ref{pr:e1}(\ref{pr:e1:b}), and therefore $F$ is a convex function.
\qed
We had earlier shown (Proposition~\ref{pr:conj-is-convex})
that the conjugate ${F^*}$ of any function
$F:\extspac{n}\rightarrow\overline{\R}$ must always be convex.
We can now show that the dual conjugate $\psi^{\dualstar}$ of any function
$\psi:{\R^n}\rightarrow\overline{\R}$ must also always be convex.
In particular, this immediately implies that the double conjugate
$F^{*\dualstar}$ is always convex, as well as $f^{*\dualstar}$, for
any function $f:{\R^n}\rightarrow\overline{\R}$.
\begin{theorem}
Let $\psi:{\R^n}\rightarrow\overline{\R}$.
Then its dual conjugate, $\psi^{\dualstar}$, is convex.
\end{theorem}
\proof
For ${\bf u}\in{\R^n}$ and $v\in{\mathbb{R}}$, let us define the affine function
\[
h_{\uu,v}({\bf x}) = {\bf x}\cdot{\bf u} - v
\]
for ${\bf x}\in{\R^n}$.
This function is clearly convex, and it was argued earlier
(in Section ~\ref{sec:lsc:ext}) that its extension is
\[
\overline{h}_{\uu,v}(\overline{\xx}) = \overline{\xx}\cdot{\bf u} - v.
\]
By Theorem~\ref{thm:fext-convex}, $\overline{h}_{\uu,v}$ is therefore convex.
The dual conjugate $\psi^{\dualstar}$, defined in \eqref{eq:psistar-def}, can be written
\[
\psi^{\dualstar}(\overline{\xx}) = \sup_{\rpair{{\bf u}}{v}\in\epi \psi} \overline{h}_{\uu,v}(\overline{\xx}).
\]
Thus, $\psi^{\dualstar}$ is a pointwise supremum over convex functions and
therefore is convex by Proposition~\ref{pr:point-sup-is-convex}.
\qed
As an application, we can show that the effective domain and all
sublevel sets of an astrally convex function are also convex:
\begin{theorem} \label{thm:f:9}
Let $F:\extspac{n}\rightarrow\overline{\R}$ be convex.
Then $\dom F$,
the effective domain of $F$, is convex, as are all of $F$'s
sublevel sets.
\end{theorem}
\proof
A point $\overline{\xx}$ is in $\dom{F}$
if and only if,
for some $y\in{\mathbb{R}}$,
$\rpair{\overline{\xx}}{y}\in\epi{F}$,
or equivalently,
$\mu(\overline{\xx},y)\in\mu(\epi{F})$,
or
$\rpair{\overline{\xx}}{y}=\mu^{-1}(\overline{{\bf z}})$ for some
$\overline{{\bf z}}\in\mu(\epi{F})$.
The inverse of $\mu$ was given explicitly in \eqref{eqn:homfinv-def}
in terms of the matrix ${\bf P}$ given in \eqref{eqn:homat-def}.
Applying to the above yields
\[
\dom{F}
=
\{ {\bf P}\overline{{\bf z}} : \overline{{\bf z}}\in\mu(\epi{F}) \}
=
P(\mu(\epi{F})),
\]
where $P:\extspac{n+1}\rightarrow\extspac{n}$
is the linear map associated with ${\bf P}$:
$P(\overline{{\bf z}})={\bf P}\overline{{\bf z}}$ for $\overline{{\bf z}}\in\extspac{n+1}$.
Since $F$ is convex, its epigraph is convex, and so also, by
definition, is its image, $\mu(\epi{F})$.
Therefore, $\dom{F}$ is convex by
Corollary~\ref{cor:thm:e:9}.
Similarly, a point $\overline{\xx}\in\extspac{n}$ is in the sublevel set
associated with $y_0\in{\mathbb{R}}$, that is,
$F(\overline{\xx})\leq y_0$,
if and only if
$\rpair{\overline{\xx}}{y_0}\in\epi{F}$,
or
$\rpair{\overline{\xx}}{y_0}=\mu^{-1}(\overline{{\bf z}})$ for some $\overline{{\bf z}}\in\mu(\epi{F})$.
Again using \eqref{eqn:homfinv-def},
the foregoing means
we can write the sublevel set as
\begin{eqnarray}
\{ \overline{\xx}\in\extspac{n} : F(\overline{\xx}) \leq y_0 \}
&=&
\{ {\bf P}\overline{{\bf z}} : \overline{{\bf z}}\in\mu(\epi{F})\cap L \}
\nonumber
\\
&=&
P(\,\mu(\epi{F})\cap L\,),
\label{eqn:f:3}
\end{eqnarray}
where
\[
L = \{ \overline{{\bf z}}\in\extspac{n+1} : \overline{{\bf z}}\cdot \rpair{{\bf 0}}{1} = y_0 \}.
\]
As before, $\mu(\epi{F})$ is convex since $F$ is,
and $L$ and $\mu(\epi{F})\cap L$ are also convex by
Proposition~\ref{pr:e1}(\ref{pr:e1:b},\ref{pr:e1:c}).
Thus, the sublevel set in \eqref{eqn:f:3} is convex by
Corollary~\ref{cor:thm:e:9}.
\qed
\section{Minimizers and their structure}
\label{sec:minimizers}
We next study the general nature of astral points that
minimize the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of a convex function $f:{\R^n}\rightarrow\overline{\R}$.
Determining the astral \emph{points}
that minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is, to a large degree,
equivalent to finding or characterizing the \emph{sequences} that
minimize the original function
$f$ (see Proposition~\ref{pr:fext-min-exists}).
As seen in Proposition~\ref{pr:i:8}(\ref{pr:i:8c}),
every astral point $\overline{\xx}$ can be decomposed as
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some icon $\overline{\ee}\in{{\cal E}_n}$ and
finite ${\bf q}\in{\R^n}$.
In the same way, the problem of minimizing the function ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ decomposes
into the separate issues of how to minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ over the
choice of $\overline{\ee}$, and how to minimize it over ${\bf q}$,
both of which will be studied in detail.
In this section, we will see that if $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, then
its icon $\overline{\ee}$ must belong to a particular set called the astral
recession cone, which will be our starting point.
We study the properties of this set and the structure of its elements,
leading to a procedure that, in a sense described below, enumerates
all of the minima of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\subsection{Astral recession cone}
\label{sec:arescone}
The standard recession cone, as defined in \eqref{eqn:resc-cone-def},
is the set of directions in which a function (on ${\R^n}$)
is never increasing.
We begin by studying an extension of this notion to astral space,
which will be centrally important
to our understanding of minimizers, continuity, and more.
The \emph{astral recession cone} of a function
$F:\extspac{n}\rightarrow\overline{\R}$, denoted ${\arescone F}$, is the set
of points that, when leftwardly added to any other point, do not cause $F$
to increase:
\begin{equation} \label{eqn:aresconeF-def}
{\arescone F}
=
\braces{ \overline{{\bf z}}\in\extspac{n} :
\forall \overline{\xx}\in\extspac{n},\,
F(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}) \leq F(\overline{\xx}) }.
\end{equation}
We will especially be interested in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, the astral recession
cone of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ (which we sometimes refer to more loosely as the
astral recession cone of $f$ itself).
For example, it can be checked that
the standard recession cone of the function
$f$ given in \eqref{eqn:curve-discont-notfinite-eg}
is the closed ray
$\resc{f}=\{ \lambda {\bf e}_2 : \lambda \geq 0 \}$.
All of these points are also in ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s astral recession cone,
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
In addition, $\limray{{\bf e}_2}$, which is in the closure of this set, must be
in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
It can further be checked that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=0$
for all ${\bf x}\in{\mathbb{R}}^2$,
so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})=0\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$
for all $\overline{{\bf y}},\overline{\xx}\in\extspac{2}$,
which means any astral point of the form
$\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$ is in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
From general results to be proved below, these turn out to be all the
points in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
So in summary,
\[
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} = \{ \lambda {\bf e}_2 : \lambda \geq 0 \}
\cup
[\limray{{\bf e}_2} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \extspac{2}].
\]
This example points to the general fact that the astral recession cone
of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ must
include the standard recession cone of $f$,
as well as its (astral) closure, but can
include other points as well (such as $\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf e}_1$, in
this example).
The precise relationship between the standard and astral recession cones
will be key to understanding the continuity properties of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
The next proposition shows that the astral recession cone
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, as defined using \eqref{eqn:aresconeF-def},
is unchanged if instead stated only in terms of points in ${\R^n}$:
\begin{proposition} \label{pr:alt-arescone-def}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, and let $\overline{{\bf z}}\in\extspac{n}$.
Then $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ if and only if
for all ${\bf x}\in{\R^n}$,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}) \leq f({\bf x})$.
\end{proposition}
\proof
Suppose first that $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, and let
${\bf x}\in{\R^n}\subseteq\extspac{n}$.
Then from definitions and Proposition~\ref{pr:h:1}(\ref{pr:h:1a}),
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})\leq f({\bf x})$.
This proves the ``only if'' part of the proposition.
Conversely, suppose now that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}) \leq f({\bf x})$ for
all ${\bf x}\in{\R^n}$.
Let $\overline{\xx}\in\extspac{n}$.
By Proposition~\ref{pr:d1}, there exists a sequence $\seq{{\bf x}_t}$ in
${\R^n}$ converging to $\overline{\xx}$ and with
$f({\bf x}_t)\rightarrow{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Then
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
=
\lim f({\bf x}_t)
\geq
\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}_t)
\geq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}).
\]
The inequalities are, respectively,
by our assumption, and by
Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2a}), since
$\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}_t \rightarrow \overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$
(by Proposition~\ref{pr:aff-cont}).
Thus, $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
\qed
The astral recession cone, ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, when restricted to ${\R^n}$, is
exactly the standard recession cone, $\resc{f}$ (assuming $f$ is lower
semicontinuous):
\begin{proposition} \label{pr:f:1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then $({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{\R^n} = \resc{f}$.
\end{proposition}
\proof
The criterion in Proposition~\ref{pr:alt-arescone-def} when
$\overline{{\bf z}}\in{\R^n}$ is exactly the same as the definition of
$\resc{f}$ given in \eqref{eqn:resc-cone-def}
(using Proposition~\ref{pr:h:1}(\ref{pr:h:1a}) and our assumption that
$f$ is lower semicontinuous).
\qed
\subsection{A dual characterization}
\label{sec:astral-cone-dual}
We will often find it useful to rely on another fundamental
characterization of
the astral recession cone in terms of the function's dual
properties.
As seen in Theorem~\ref{thm:rescpol-is-con-dom-fstar},
the standard recession cone of a
closed, proper, convex function $f$ is the
(standard) polar of $\cone(\dom{{f^*}})$.
In the same way, the astral
recession cone of a function $f$'s extension
is equal to the astral polar of this same set,
assuming $f$ is convex, has all reductions closed,
and is not identically $+\infty$.
\begin{theorem} \label{thm:j:1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex with all reductions closed, and
assume $f\not\equiv+\infty$.
Then ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\apol{(\cone(\dom{\fstar}))}$.
\end{theorem}
Before proving this, we first give a lemma regarding the behavior of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ on points not in $\apol{(\cone(\dom{\fstar}))}$.
(This will later be strengthened in Theorem~\ref{thm:f:2}.)
\begin{lemma} \label{lem:f:2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, and
suppose $\overline{{\bf z}}\not\in\apol{(\cone(\dom{\fstar}))}$.
Then for all ${\bf q}\in{\R^n}$,
\[ {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}) \rightarrow +\infty \]
as $\lambda\rightarrow+\infty$.
\end{lemma}
\proof
Since $\overline{{\bf z}}\not\in\apol{(\cone(\dom{\fstar}))}$,
there exists ${\bf u}\in{\R^n}$ with ${f^*}({\bf u})<+\infty$
and $\overline{{\bf z}}\cdot{\bf u}>0$
(using \eqref{eq:cone-dom-fstar}).
By Theorem~\ref{thm:fext-dub-sum},
for all ${\bf q}\in{\R^n}$ and all $\lambda>0$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
\geq
f^{*\dualstar}(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
\geq
-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda\overline{{\bf z}}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}\cdot{\bf u}.
\]
If ${f^*}({\bf u})=-\infty$, then the right-hand side is equal to
$+\infty$ for all $\lambda$.
Otherwise, ${f^*}({\bf u})\in{\mathbb{R}}$,
${\bf q}\cdot{\bf u}\in{\mathbb{R}}$, and
$\overline{{\bf z}}\cdot{\bf u}>0$, so the right-hand side must
tend to $+\infty$ as $\lambda\rightarrow+\infty$.
\qed
\proof[Proof of Theorem~\ref{thm:j:1}]
We show first that ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}\subseteq\apol{(\cone(\dom{\fstar}))}$.
Since $f\not\equiv+\infty$,
there exists ${\bf y}\in{\R^n}$ with $f({\bf y})<+\infty$.
Let $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
We claim, by induction on $n$, that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(n\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf y})\leq f({\bf y})<+\infty$
for $n=0,1,2,\ldots$.
The base case that $n=0$ follows from
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
In the inductive case,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(n\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (n-1)\overline{{\bf z}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}((n-1)\overline{{\bf z}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y})
\leq
f({\bf y}).
\]
These steps follow, respectively, from
Proposition~\ref{pr:i:7}(\ref{pr:i:7c}),
the assumption that $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
and our inductive hypothesis.
As a result of this claim, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y})$ cannot converge to
$+\infty$ as $\lambda\rightarrow+\infty$.
Therefore, $\overline{{\bf z}}\in\apol{(\cone(\dom{\fstar}))}$ by Lemma~\ref{lem:f:2}.
For the reverse inclusion,
suppose $\overline{{\bf z}}\in\apol{(\cone(\dom{\fstar}))}$.
For all ${\bf u}\in{\R^n}$, we claim
\begin{equation} \label{eqn:h:1}
-{f^*}({\bf u})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\cdot{\bf u} \leq -{f^*}({\bf u}).
\end{equation}
This is immediate if ${f^*}({\bf u})\in\{-\infty,+\infty\}$.
Otherwise, if ${f^*}({\bf u})\in{\mathbb{R}}$ then
$\overline{{\bf z}}\cdot{\bf u}\leq 0$ since $\overline{{\bf z}}$ was assumed to be in
$\apol{(\cone(\dom{\fstar}))}$.
Since $f$ has all reductions closed,
Theorems~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b})
and~\ref{thm:dub-conj-new}, together with \eqref{eqn:h:1}, then
imply that
\begin{eqnarray*}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
=
f^{*\dualstar}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
&=&
\sup_{{\bf u}\in{\R^n}} (-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u})
\\
&\leq&
\sup_{{\bf u}\in{\R^n}} (-{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx}\cdot{\bf u})
\\
&=&
f^{*\dualstar}(\overline{\xx})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
\end{eqnarray*}
for all $\overline{\xx}\in\extspac{n}$.
Therefore, $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
\qed
If $f$ does not have all reductions closed,
then the sets ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ and $\apol{(\cone(\dom{\fstar}))}$ need not be equal.
For example, consider the function $f$ given in \eqref{eqn:ex:1}.
Let
$\overline{{\bf z}}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{(-{\bf e}_2)}$.
Then $\overline{{\bf z}}\in\apol{(\cone(\dom{\fstar}))}$ (since if ${f^*}({\bf u})\in{\mathbb{R}}$ then $u_1=-1$
implying $\overline{{\bf z}}\cdot{\bf u}=-\infty$).
But $\overline{{\bf z}}\not\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
(since, for instance,
$f({\bf 0})=0$ but
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf 0})=+\infty$).
We can nonetheless generalize Theorem~\ref{thm:j:1} so that it
holds for all convex functions,
even if they do not have all reductions closed,
using the technique developed in Section~\ref{sec:exp-comp},
with $\cone(\dom{{f^*}})$ replaced by $\slopes{f}$:
\begin{theorem} \label{th:ares-is-apolslopes}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\apol{(\slopes{f})}$.
\end{theorem}
\proof
If $f\equiv+\infty$ then
$\slopes{f}=\{{\bf 0}\}$ so
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\apol{(\slopes{f})}=\extspac{n}$.
Otherwise,
let $f'={\overline{\exp}}\circ f$.
Then
\[
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}
=
{\arescone \fpext}
=
\apol{(\cone(\dom{\fpstar}))}
=
\apol{(\slopes{f})}.
\]
The first equality is by
Proposition~\ref{pr:j:2}(\ref{pr:j:2d}).
The second is by Theorem~\ref{thm:j:1} applied to $f'$
(which is convex and lower-bounded
by Proposition~\ref{pr:j:2}(\ref{pr:j:2a}),
and therefore has all reductions closed by
Proposition~\ref{pr:j:1}(\ref{pr:j:1c})).
The third is
by Proposition~\ref{pr:dom-fpstar}.
\qed
This immediately implies that ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is always a closed,
convex cone, analogous to Proposition~\ref{pr:resc-cone-basic-props}:
\begin{corollary} \label{cor:thm:j:3}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then the astral recession cone ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ of its extension
is a closed (in
$\extspac{n}$) convex cone that includes the origin.
\end{corollary}
\proof
This follows immediately from
Theorem~\ref{th:ares-is-apolslopes}
combined with the fact that the astral polar of any set is a closed
convex cone
(Proposition~\ref{pr:ast-pol-props}(\ref{pr:ast-pol-props:c})).
\qed
We can now also prove a stronger and slightly expanded version of
Lemma~\ref{lem:f:2} in which we replace
$\apol{(\cone(\dom{\fstar}))}$ with
its superset ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\apol{(\slopes{f})}$:
\begin{theorem} \label{thm:f:2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, and
suppose $\overline{{\bf z}}\not\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Then for all ${\bf q}\in{\R^n}$,
\[ {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}) \rightarrow +\infty \]
as $\lambda\rightarrow+\infty$.
Furthermore, if $\overline{\ee}\in{{\cal E}_n}\setminus({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})$, then
\[ {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}) = +\infty \]
for all ${\bf q}\in{\R^n}$.
\end{theorem}
\proof
As in the proof of Theorem~\ref{th:ares-is-apolslopes},
let $f'={\overline{\exp}}\circ f$, implying,
by Proposition~\ref{pr:j:2}(\ref{pr:j:2d})
and Theorem~\ref{thm:j:1}, that
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}={\arescone \fpext}=\apol{(\cone(\dom{\fpstar}))}$.
Suppose $\overline{{\bf z}}\not\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, and let ${\bf q}\in{\R^n}$.
Then $\overline{{\bf z}}\not\in\apol{(\cone(\dom{\fpstar}))}$,
so by Lemma~\ref{lem:f:2},
$\fpext(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}) \rightarrow +\infty$
as $\lambda\rightarrow+\infty$.
By Proposition~\ref{pr:j:2}(\ref{pr:j:2c}) and continuity of ${\overline{\exp}}$,
this implies that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}) \rightarrow +\infty$
as well.
If $\overline{\ee}\in{{\cal E}_n}\setminus({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})$, then
$\lambda\overline{\ee}=\overline{\ee}$ for all $\lambda>0$
(Proposition~\ref{pr:i:8}(\ref{pr:i:8d})), implying
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
=
\lim_{\lambda\rightarrow+\infty} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
=
+\infty.
\]
\qed
In the contrapositive,
this theorem means that if there exists a single point ${\bf q}\in{\R^n}$ and some
$B\in{\mathbb{R}}$ for which ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}) \leq B$ for arbitrarily
large values of $\lambda>0$, then $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
In particular, if $\overline{\ee}$ is an icon
and there exists a single
${\bf q}\in{\R^n}$ for which ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})<+\infty$ then
$\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Consequently, any point in ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s
effective domain (including, specifically, all of its minimizers)
can be written in the form $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where
${\bf q}\in{\R^n}$ and $\overline{\ee}$ is an icon in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
\begin{corollary} \label{cor:a:4}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex,
let ${\bf q}\in{\R^n}$,
and let $\overline{\ee}\in{{\cal E}_n}$.
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})<+\infty$ then
$\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
\end{corollary}
We can now state precisely which points minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, namely,
those points of the form $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where $\overline{\ee}$ is an icon in
the astral recession cone, and ${\bf q}\in{\R^n}$ minimizes the reduction
$\fshad{\ebar}$
(defined in \eqref{eqn:mult-step-shad-def}).
Later, in Section~\ref{sec:univ-red-and-min},
we will develop a much more detailed
analysis of the minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, but this theorem provides a start:
\begin{theorem} \label{thm:arescone-fshadd-min}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Let $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
Then $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ if and only if $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
and ${\bf q}$ minimizes $\fshad{\ebar}$.
\end{theorem}
\proof
If $f\equiv+\infty$ then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}\equiv+\infty$, $\fshad{\ebar}\equiv+\infty$,
and ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\extspac{n}$, so the claim follows trivially.
Therefore, we assume $f\not\equiv+\infty$,
so $\min {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'} = \inf f < +\infty$
(by Proposition~\ref{pr:fext-min-exists}).
Suppose $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})<+\infty$, so $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
by Corollary~\ref{cor:a:4}.
If, contrary to the claim, ${\bf q}$ does not minimize $\fshad{\ebar}$,
then there exists ${\bf q}'\in{\R^n}$ with
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}')
=
\fshad{\ebar}({\bf q}')
<
\fshad{\ebar}({\bf q})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}),
\]
contradicting that $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Conversely, suppose $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ and ${\bf q}$ minimizes
$\fshad{\ebar}$.
Let $\beta$ be any number in ${\mathbb{R}}$ with $\beta > \inf f$,
and let ${\bf y}\in{\R^n}$ be such that $f({\bf y}) < \beta$.
Then
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
=
\fshad{\ebar}({\bf q})
\leq
\fshad{\ebar}({\bf y})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf y})
\leq
f({\bf y})
<
\beta.
\]
The first inequality is because ${\bf q}$ minimizes $\fshad{\ebar}$,
and the second is by Proposition~\ref{pr:alt-arescone-def}.
Since this holds for all $\beta>\inf f$,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=\inf f$, so $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\qed
\subsection{Constructing the astral recession cone}
\label{sec:astral-cone}
As just discussed,
every minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ must involve points in its astral recession
cone.
Thus, to minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ (as well as $f$), it will be helpful to
understand the structure of such points, and how to construct them.
We will see how this is done in this section using the methods
developed earlier based on projections and reductions.
Let ${\bf v}\in{\R^n}$, and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
We begin by showing how points in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
the astral recession cone of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
relate to points in ${\arescone \overline{g}}$,
the astral recession cone of $\overline{g}$.
\begin{theorem} \label{thm:f:3}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Let ${\bf v}\in{\R^n}$
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Then the following hold:
\begin{enumerate}
\item \label{thm:f:3a}
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} \subseteq {\arescone \overline{g}}$.
(Consequently, $\resc{f} \subseteq \resc{g}$.)
\item \label{thm:f:3b}
Suppose $\overline{{\bf y}}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ for some $\overline{{\bf z}}\in\extspac{n}$.
Then $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ if and only if
${\bf v}\in\resc{f}$ and
$\overline{{\bf z}}\in{\arescone \overline{g}}$.
\end{enumerate}
\end{theorem}
Part~(\ref{thm:f:3b}) of
this theorem provides a kind of recursive characterization of all of
the points comprising ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$:
The points in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ that are also in ${\R^n}$ are
exactly those in the standard recession cone $\resc{f}$,
by Proposition~\ref{pr:f:1}.
All of the other points in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ can be entirely enumerated
by considering each point
${\bf v}\in\resc{f}$, forming the reduction $g$ of $f$ at $\limray{{\bf v}}$,
finding $\overline{g}$'s astral recession cone ${\arescone \overline{g}}$, and then
adding $\limray{{\bf v}}$ to each element in ${\arescone \overline{g}}$.
Thus,
\[
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} = (\resc{f}) \cup
\bigcup_{{\bf v}\in\resc{f}}
\paren{\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \arescone{\fshadext{\limray{{\bf v}}}}}.
\]
Alternatively, we can think of part~(\ref{thm:f:3b}), together with
Proposition~\ref{pr:f:1}, as providing a test for determining if a
given point $\overline{{\bf y}}$ is in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$:
If $\overline{{\bf y}}$ is in ${\R^n}$, then it is in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ if and only if it
is in $\resc{f}$.
Otherwise, it is in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ if and only if
its dominant direction ${\bf v}$ is in $\resc{f}$ and its projection
$\ybar^{\bot}$ is in ${\arescone \overline{g}}$, as can be determined in a recursive
manner.
This characterization can also be interpreted in terms of sequences.
Suppose some sequence converges to a point $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
where $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$, and such that $f$ is
eventually bounded above on the sequence (as will be the case if $f$
is actually minimized by the sequence).
Then $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, by Corollary~\ref{cor:a:4}.
So Theorem~\ref{thm:f:3} tells us that, unless $\overline{\xx}\in{\R^n}$, the
sequence must have a dominant direction ${\bf v}$ in $\resc{f}$.
Moreover, we can project the sequence to the space perpendicular
to ${\bf v}$ and form the associated reduction $g$ of $f$ at $\limray{{\bf v}}$.
The projected sequence now must converge to $\ebar^{\bot}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\qq}^{\bot}$.
According to Theorem~\ref{thm:f:3}, $\ebar^{\bot}\in{\arescone \overline{g}}$, so
we can apply this same reasoning again to the projected sequence,
so that either the projected sequence converges to a point in
${\R^n}$, or its dominant direction is in $\resc{g}$.
Continuing in this fashion, we can effectively characterize all of the
dominant directions of the sequence.
\proof
~
Part~(\ref{thm:f:3a}):
If $f\equiv +\infty$, then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=\overline{g}\equiv +\infty$, so
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}={\arescone \overline{g}}=\extspac{n}$, implying the claim (trivially)
in this case.
If ${\bf v}\not\in\resc{f}$, then $\overline{g}\equiv +\infty$ by
Theorem~\ref{thm:i:4}, again implying ${\arescone \overline{g}}=\extspac{n}$ and
trivially yielding the claim.
So suppose ${\bf v}\in\resc{f}$ and that $f({\bf q})<+\infty$ for some
${\bf q}\in{\R^n}$.
Let $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Then for all $\lambda>0$,
\[
\overline{g}(\lambda \overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda \overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q})
\leq
f({\bf q})
< +\infty.
\]
The first inequality is by Corollary~\ref{cor:i:1}(\ref{cor:i:1c}).
The second inequality is
because $\lambda \overline{{\bf y}} \in {\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
since ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a cone (by Corollary~\ref{cor:thm:j:3}),
and Proposition~\ref{pr:alt-arescone-def}.
Therefore, by Theorem~\ref{thm:f:2}, $\overline{{\bf y}}\in{\arescone \overline{g}}$.
Having proved ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} \subseteq {\arescone \overline{g}}$,
it now follows, when combined with Proposition~\ref{pr:f:1}, that
\[
\resc{f}
=
({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{\R^n}
\subseteq
({\arescone \overline{g}})\cap{\R^n}
=
\resc{g}.
\]
Part~(\ref{thm:f:3b}):
Suppose first that ${\bf v}\in\resc{f}$ and that $\overline{{\bf z}}\in{\arescone \overline{g}}$.
Then for all $\overline{\xx}\in\extspac{n}$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx})
=
\overline{g}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\xx})
\leq
\overline{g}(\overline{\xx})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}).
\]
The second equality is by Corollary~\ref{cor:i:1}(\ref{cor:i:1b}).
The inequalities are because $\overline{{\bf z}}\in{\arescone \overline{g}}$,
and by Corollary~\ref{cor:i:1}(\ref{cor:i:1c})
(since ${\bf v}\in\resc{f}$).
Therefore, $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
For the converse, suppose for the rest of the proof that
$\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
We argue separately that ${\bf v}\in\resc{f}$ and $\overline{{\bf z}}\in{\arescone \overline{g}}$.
First, suppose by way of contradiction that ${\bf v}\not\in\resc{f}$.
Then $f\not\equiv +\infty$, since otherwise $\resc{f}={\R^n}$,
contradicting that ${\bf v}\not\in\resc{f}$.
So there exists a point ${\bf q}\in{\R^n}$ with $f({\bf q})<+\infty$.
On the other hand, since ${\bf v}\not\in\resc{f}$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q})
=
+\infty
\]
by Theorem~\ref{thm:i:4}.
But this is impossible since $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, which implies
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})\leq f({\bf q})$.
Thus, ${\bf v}\in\resc{f}$.
Next, from part~(\ref{thm:f:3a}), $\overline{{\bf y}}\in{\arescone \overline{g}}$.
Therefore, for all $\overline{\xx}\in\extspac{n}$,
\[
\overline{g}(\overline{\xx})
\geq
\overline{g}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
=
\overline{g}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})
=
\overline{g}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}).
\]
The last equality follows from
Corollary~\ref{cor:i:1}(\ref{cor:i:1b})
since
$
(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})^\bot
=
(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})^\bot
$
using Proposition~\ref{pr:h:5}(\ref{pr:h:5c},\ref{pr:h:5e}).
Thus $\overline{{\bf z}}\in{\arescone \overline{g}}$.
\qed
\subsection{Finding all minimizers}
\label{sec:find-all-min}
\begin{figure}[t]
\underline{Given:}
\begin{itemize}
\item
function $f:{\R^n}\rightarrow\overline{\R}$ that is convex and lower
semicontinuous
\item
test point $\overline{\xx} = \limrays{{\bf v}_1,\ldots,{\bf v}_k} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$
where ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in{\R^n}$
\end{itemize}
\smallskip
\underline{Return:} {\bf true}\ if $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$;
{\bf false}\ otherwise
\smallskip
\underline{Procedure:}
\begin{itemize}
\item if $k=0$ then
\begin{itemize}
\item
if ${\bf q}$ minimizes $f$ then return {\bf true}
\item
else return {\bf false}
\end{itemize}
\item else
\begin{itemize}
\item
if ${\bf v}_1\not\in \resc{f}$ then return {\bf false}
\item
else
\begin{itemize}
\item
let $g = \fshad{\limray{\vv}}$
\item
recursively test if $\limrays{{\bf v}_2,\ldots,{\bf v}_k} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$
minimizes $\overline{g}$, and return the result
\end{itemize}
\end{itemize}
\end{itemize}
\caption{A procedure for testing if a given point minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.}
\label{fig:min-test-proc}
\end{figure}
Combining the results developed above, we can now provide a procedure
for testing if a given astral point minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Such a procedure is shown in
Figure~\ref{fig:min-test-proc}.
The input is a function $f$ and an expicitly represented
test point $\overline{\xx} = \limrays{{\bf v}_1,\ldots,{\bf v}_k} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$.
The procedure determines if $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ using only
more basic primitives which operate on standard points and
functions over ${\R^n}$, specifically,
for testing if a point in ${\R^n}$ minimizes an ordinary
convex function $f:{\R^n}\rightarrow\overline{\R}$, and also for testing if a
vector in ${\R^n}$ is in the standard recession cone of such a function.
The operation and correctness of this procedure follow directly from
our development regarding minimizers and reductions:
If $k=0$ then $\overline{\xx}={\bf q}\in{\R^n}$, so $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ if and
only if ${\bf q}$ minimizes the standard function $f$
(which we have assumed is lower semicontinuous).
Otherwise, if $k>0$, then
$\overline{\xx}=\limray{{\bf v}_1} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf z}}$
where
$\overline{{\bf z}} = \limrays{{\bf v}_2,\ldots,{\bf v}_k} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$.
If ${\bf v}_1\not\in\resc{f}$, then
$\limrays{{\bf v}_1,\ldots,{\bf v}_k}$ cannot be in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by Theorem~\ref{thm:f:3}(\ref{thm:f:3b}), and therefore
$\overline{\xx}$ cannot minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, by
Theorem~\ref{thm:arescone-fshadd-min}.
Otherwise, with $g$ as defined in the figure,
if ${\bf v}_1\in\resc{f}$, then
$\overline{g}(\overline{{\bf z}})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$
and
$\min \overline{g} = \inf g = \inf f = \min {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
by Theorem~\ref{thm:d4}
and Proposition~\ref{pr:fext-min-exists}.
Therefore, $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
if and only if $\overline{{\bf z}}$ minimizes $\overline{g}$.
\begin{figure}[t]
\underline{Given:}
function $f:{\R^n}\rightarrow\overline{\R}$ that is convex and lower semicontinuous
\smallskip
\underline{Process:}
\begin{itemize}
\item $i \leftarrow 0$
\item $g_0 = f$
\item repeat \emph{at least} until $g_i$ has a finite minimizer
\begin{itemize}
\item $i \leftarrow i+1$
\item let ${\bf v}_i$ be \emph{any} point in
$\resc{g_{i-1}}$
\item $g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$
\end{itemize}
\item $k \leftarrow i$
\item $\overline{\ee}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}$
\item let ${\bf q}\in{\R^n}$ be \emph{any} finite minimizer of $g_k$
\item $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
\end{itemize}
\medskip
\underline{Properties:}
\begin{itemize}
\item $g_k = \fshad{\ebar}$
\item $\overline{\ee}\in ({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}$
\item ${\bf q}$ minimizes $\fshad{\ebar}$
\item $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
\end{itemize}
\caption{
A process for finding all astral minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.}
\label{fig:all-min-proc}
\end{figure}
Our study so far of the astral recession cone and the minimizers of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ also yields a general, iterative procedure
that, in a sense described below, can find {all} of the minimizers of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, as we present next.
By Theorem~\ref{thm:arescone-fshadd-min},
every minimizer $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ has an iconic part
$\overline{\ee}$ that is in the astral recession cone ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
and a finite part ${\bf q}$ that minimizes $\fshad{\ebar}$.
To find an icon in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by Theorem~\ref{thm:f:3}(\ref{thm:f:3b}),
we can first find a point ${\bf v}\in\resc{f}$, form the associated
reduction $g$ of $f$ at $\limray{{\bf v}}$,
and then repeat the process to find a point in
${\arescone \overline{g}}$, eventually choosing an appropriate time to stop.
More precisely, and using a bit more notation,
we initially let $g_0=f$.
On each iteration $i$, we find a vector ${\bf v}_i$ in $\resc{g_{i-1}}$, the
standard recession cone of $g_{i-1}$.
Then we define
$g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$
to be the next reduction,
in this way, ensuring that the resulting icon
$\overline{\ee}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}$
formed by the ${\bf v}_i$'s must be in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
We can continue this process until we manage to form a reduction
$g_k$ that has some finite minimizer ${\bf q}\in{\R^n}$.
By such a construction, $g_k$ actually is equal to the reduction
$\fshad{\ebar}$ at icon $\overline{\ee}$, so in fact, ${\bf q}$ minimizes $\fshad{\ebar}$,
which, combined with $\overline{\ee}$ being in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, ensures that the
point $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
We summarize this process in Figure~\ref{fig:all-min-proc}.
Although we describe the process in the form of an algorithm, we do
not literally mean to suggest that it be implemented on a computer, at
least not in this generality.
The point, rather, is to reveal the structure of the minimizers of a
function in astral space, and how that structure can be related to
standard notions from convex analysis.
Note, in particular, that the process is nondeterministic in the sense
that at various points, choices are made in way that is entirely
arbitrary.
This happens at three different points:
First, on each iteration of the main loop, an \emph{arbitrary} point
${\bf v}_i$ is selected from $\resc{g_{i-1}}$.
Second, this loop must iterate \emph{at least} until $g_i$ has a finite
minimizer, but can continue to iterate arbitrarily beyond that point.
Third, after terminating the loop, an \emph{arbitrary} finite
minimizer ${\bf q}$ of $g_k$ is selected.
Clearly, the point $\overline{\xx}$ that is eventually computed by the process
depends on these arbitrary choices.
Nevertheless, in all cases, the resulting
point $\overline{\xx}$ must be a minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Conversely, if $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, then it must be possible for
these arbitrary choices to be made in such a way that $\overline{\xx}$ is
produced
(while still respecting the constraints imposed at each step
of the process).
It is in this sense that the process computes \emph{all} of the
minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
When there exists such a sequence of choices that results in $\overline{\xx}$
as the final output or product of the computation, we say that $\overline{\xx}$
is a \emph{potential product} of the process.
Thus, we are claiming that a point $\overline{\xx}\in\extspac{n}$ minimizes
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ if and only if $\overline{\xx}$ is a potential product of the process.
This is shown formally by the next theorem, whose
condition~(\ref{thm:all-min-proc-correct:b})
captures exactly when the point $\overline{\xx}$ is a potential product.
\begin{theorem} \label{thm:all-min-proc-correct}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Let $\overline{\xx} = \limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where
${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in{\R^n}$.
Let $g_0=f$ and $g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$ for $i=1,\ldots,k$.
Then $g_k=\fshad{\ebar}$, and the following are equivalent:
\begin{enumerate}
\item \label{thm:all-min-proc-correct:a}
$\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\item \label{thm:all-min-proc-correct:b}
${\bf q}$ minimizes $g_k$
and
${\bf v}_i\in\resc{g_{i-1}}$ for $i=1,\ldots,k$.
\end{enumerate}
\end{theorem}
\proof
Let $\overline{\ee}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}$.
As preliminary steps, as in the proof of Proposition~\ref{pr:i:9}, a
straightforward induction on $i=0,\ldots,k$ shows that
$g_i({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limrays{{\bf v}_1,\ldots,{\bf v}_{i}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})$
for ${\bf x}\in{\R^n}$.
In particular, this shows that $g_k=\fshad{\ebar}$.
Also, for $i=0,\ldots,k$,
let $\overline{\ee}_i=\limrays{{\bf v}_{i+1},\ldots,{\bf v}_k}$.
Since $\overline{\ee}_{i-1}=\limray{{\bf v}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}_i$,
and by $g_i$'s definition,
Theorem~\ref{thm:f:3}(\ref{thm:f:3b}) implies that,
for $i=1,\ldots,k$,
$\overline{\ee}_{i-1}\in\aresconegsub{{i-1}}$
if and only if
${\bf v}_i\in\resc{g_{i-1}}$ and
$\overline{\ee}_{i}\in\aresconegsub{{i}}$.
\underline{(\ref{thm:all-min-proc-correct:a})
$\Rightarrow$
(\ref{thm:all-min-proc-correct:b})}:
Suppose $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Then Theorem~\ref{thm:arescone-fshadd-min} implies that
${\bf q}$ minimizes $\fshad{\ebar}=g_k$, and also that
$\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, or equivalently, that
$\overline{\ee}_0\in\aresconegsub{0}$.
From the preceding remarks,
it now follows by another straightforward induction that
${\bf v}_i\in\resc{g_{i-1}}$ and
$\overline{\ee}_{i}\in\aresconegsub{{i}}$
for $i=1,\ldots,k$.
\underline{(\ref{thm:all-min-proc-correct:b})
$\Rightarrow$
(\ref{thm:all-min-proc-correct:a})}:
Suppose
${\bf q}$ minimizes $g_k=\fshad{\ebar}$
and that
${\bf v}_i\in\resc{g_{i-1}}$
for $i=1,\ldots,k$.
Then by backwards induction on $i=0,\ldots,k$,
$\overline{\ee}_{i}\in\aresconegsub{{i}}$.
The base case, when $i=k$, holds because
$\overline{\ee}_k={\bf 0}\in\aresconegsub{k}$.
For the inductive step, when $i<k$,
$\overline{\ee}_{i+1}\in\aresconegsub{{i+1}}$ by inductive hypothesis, and
${\bf v}_{i+1}\in\resc{g_i}$ by assumption, so
the earlier remark implies
$\overline{\ee}_{i}\in\aresconegsub{{i}}$.
Thus, ${\bf q}$ minimizes $\fshad{\ebar}$ and
$\overline{\ee}=\overline{\ee}_0\in\aresconegsub{0}={\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Therefore, $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
by Theorem~\ref{thm:arescone-fshadd-min}.
\qed
Theorem~\ref{thm:all-min-proc-correct} shows that if this process
terminates, then the computed point $\overline{\xx}$ must minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
But what if the process never terminates?
Indeed, it is possible for the process to never terminate, or even to
reach the loop's (optional) termination condition.
For instance,
the same point ${\bf v}_i={\bf 0}$ (which is in every recession cone)
might be chosen on every iteration so
that the process never makes any progress toward a solution at all.
Also, superficially, it might seem plausible that the process could
make poor choices early on that make it impossible to eventually reach
a point at which the termination condition is satisfied.
We will address these issues later in
Section~\ref{sec:ensure-termination}.
\section{Universal reduction and universal minimizers}
\label{sec:univ-red-and-min}
For a convex function $f:{\R^n}\rightarrow\overline{\R}$,
we have seen so far that if $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
where $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$, then $\overline{\xx}$'s iconic part
$\overline{\ee}$ must be in ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s
astral recession cone (Theorem~\ref{thm:arescone-fshadd-min}).
In this section, we delve further into the structure of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s
minimizers, both their iconic and finite parts.
We will see that all finite parts ${\bf q}$ of all minimizers are exactly
captured as the (finite) minimizers of one
particular convex function called the universal reduction,
defined in a moment.
Furthermore, all of the minimizers of this function
are, in a sense explained below,
necessarily in a bounded region of ${\R^n}$, thereby
alleviating the consideration of minimizers at infinity.
Thus, the problem of finding the finite parts ${\bf q}$ of all minimizers
of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ can be reduced to minimization of a standard convex
function in the most favorable setting that finite minimizers exist
and only occur within some compact region.
Furthermore,
we will see that there exist choices for the icon $\overline{\ee}$ that
minimize
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})$ over choices for $\overline{\ee}$
\emph{simultaneously} for all ${\bf x}\in{\R^n}$.
We will discuss how to find such points, their properties, and how
they combine naturally with the universal reduction function just
described, yielding minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ that are, informally, the
most canonical or extreme of minimizers.
\subsection{The universal reduction}
\label{sec:fullshad}
For a convex function $f:{\R^n}\rightarrow\overline{\R}$, we begin by defining
the function briefly described above, called the
\emph{universal reduction} of $f$.
This function, written
$\fullshad{f}:{\R^n}\rightarrow \overline{\R}$,
computes the minimum possible value
of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ when some point ${\bf x}\in{\R^n}$ is combined with any icon
$\overline{\ee}\in{{\cal E}_n}$.
Thus,
\begin{equation} \label{eqn:fullshad-def}
\fullshad{f}({\bf x}) = \inf_{\overline{\ee}\in{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
= \inf_{\overline{\ee}\in{{\cal E}_n}} \fshad{\ebar}({\bf x})
\end{equation}
for ${\bf x}\in{\R^n}$.
In other words, $\fullshad{f}$ is the pointwise infimum of all
reductions $\fshad{\ebar}$ over all icons $\overline{\ee}\in{{\cal E}_n}$.
In this sense, $\fullshad{f}$ can be viewed as itself a reduction of
$f$ across the entire universe of astral space;
it is for this reason that it
is called the {universal reduction}.
Intuitively,
in minimizing $f$, this function ``washes out'' what is
possible by pursuing the trajectory of a sequence to infinity
beginning at ${\bf x}$ and
following the path defined by any icon $\overline{\ee}$.
Alternatively, $\fullshad{f}$ can be viewed informally as bringing in
$f$'s behavior at infinity to a compact region of ${\R^n}$.
The definition of $\fullshad{f}$ remains the same
if we consider adding points $\overline{\ee}$ that are instead in the astral
recession cone of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
whether or not restricted to those that are icons,
as stated in the next proposition.
Furthermore, in all cases, including
\eqref{eqn:fullshad-def}, the respective infima are always realized by
some point $\overline{\ee}$, which means we can state these expressions in
terms of minima rather than infima.
\begin{proposition} \label{pr:fullshad-equivs}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then for all ${\bf x}\in{\R^n}$,
\[
\fullshad{f}({\bf x})
= \min_{\overline{\ee}\in{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
= \min_{\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
= \min_{\overline{\ee}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}).
\]
In particular, this means that each of these minima is attained.
\end{proposition}
\proof
Let ${\bf x}\in{\R^n}$.
By Theorem~\ref{thm:f:2}, if $\overline{\ee}\in{{\cal E}_n}\setminus({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})$,
then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=+\infty$.
Therefore,
\begin{equation} \label{eqn:pr:fullshad-equivs:2}
\fullshad{f}({\bf x})
= \inf_{\overline{\ee}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\geq
\inf_{\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}),
\end{equation}
where the inequality is simply because
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a superset of $({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}$.
Let
\begin{equation} \label{eqn:pr:fullshad-equivs:3}
M = \inf_{\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}),
\end{equation}
and, for $t=1,2,\ldots$, let
\[ b_t = \left\{ \begin{array}{cl}
-t & \mbox{if $M = -\infty$}\\
M + 1/t & \mbox{if $M \in {\mathbb{R}}$}\\
+\infty & \mbox{if $M = +\infty$.}
\end{array}
\right.
\]
Let $\overline{{\bf y}}_t$ be any point in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ with
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}_t\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})\leq b_t$.
Then the sequence $\seq{\overline{{\bf y}}_t}$ has a convergent subsequence
(by sequential compactness);
discarding all other elements, we can assume the entire sequence
converges to some point $\overline{{\bf y}}\in\extspac{n}$.
Further, $\overline{{\bf y}}$ must be in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
since each $\overline{{\bf y}}_t$ is in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, which is closed
(by Corollary~\ref{cor:thm:j:3}).
Thus,
\begin{equation} \label{eqn:pr:fullshad-equivs:4}
M
=
\lim b_t
\geq
\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}_t\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\geq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}),
\end{equation}
where the second inequality is because
$\overline{{\bf y}}_t\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}\rightarrow\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$
(since the map $\overline{{\bf z}}\mapsto\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}={\bf x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ is
continuous, by Proposition~\ref{pr:aff-cont}),
and by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2a}).
We can write $\overline{{\bf y}}=\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some $\overline{\dd}\in{{\cal E}_n}$ and
${\bf q}\in{\R^n}$.
Then considering the sequence $t\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$, we have
\begin{eqnarray}
M
&\geq&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\nonumber
\\
&\geq&
\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}((t-1)\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\nonumber
\\
&=&
\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(t\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\nonumber
\\
&\geq&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}).
\label{eqn:pr:fullshad-equivs:1c}
\end{eqnarray}
The first inequality is \eqref{eqn:pr:fullshad-equivs:4}.
The second is because
$(t-1)\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, for $t\geq 1$, since
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a cone
(Corollary~\ref{cor:thm:j:3}).
The equality is by
Proposition~\ref{pr:i:7}(\ref{pr:i:7c}).
And the third inequality is
by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2a}), and because
\[ t\overline{{\bf y}}=\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax t{\bf q} \rightarrow \overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}} \]
using Proposition~\ref{pr:aff-cont} applied to the map
$\overline{{\bf z}}\mapsto\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$.
This further shows that $\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}$, which is iconic,
must also be in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, since that set is closed;
that is,
\begin{equation} \label{eqn:pr:fullshad-equivs:6}
\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}} \in ({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}.
\end{equation}
Combining now yields
\begin{eqnarray}
M
\geq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
&\geq&
\inf_{\overline{\ee}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\nonumber
\\
&=&
\inf_{\overline{\ee}\in{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\nonumber
\\
&=&
\fullshad{f}({\bf x})
\nonumber
\\
&\geq&
\inf_{\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
M.
\label{eqn:pr:fullshad-equivs:5}
\end{eqnarray}
The first two inequalities are from
\eqref{eqn:pr:fullshad-equivs:1c}
and
\eqref{eqn:pr:fullshad-equivs:6},
respectively.
The first two equalities and the last inequality are from
\eqref{eqn:fullshad-def} and \eqref{eqn:pr:fullshad-equivs:2}.
And the last equality is
\eqref{eqn:pr:fullshad-equivs:3}.
Thus, equality holds across all of
\eqref{eqn:pr:fullshad-equivs:5}.
Furthermore, this shows that
$\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf q}}$, which is in $({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}$,
realizes each of the three infima.
\qed
As noted above, the set of all minimizers of $\fullshad{f}$ is exactly
equal to the set of all finite parts of all minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$:
\begin{proposition} \label{pr:min-fullshad-is-finite-min}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex,
and let ${\bf q}\in{\R^n}$.
Then ${\bf q}$ minimizes $\fullshad{f}$ if and only if
there exists $\overline{\ee}\in{{\cal E}_n}$ such that
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Consequently, $\fullshad{f}$ attains its minimum.
\end{proposition}
\proof
Suppose first that ${\bf q}$ minimizes $\fullshad{f}$.
By Proposition~\ref{pr:fullshad-equivs}, there exists
$\overline{\ee}\in{{\cal E}_n}$ such that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})=\fullshad{f}({\bf q})$.
Let $\overline{\xx}'=\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$ be any astral point with iconic part
$\overline{\ee}'\in{{\cal E}_n}$ and finite part ${\bf q}'\in{\R^n}$.
Then
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
=
\fullshad{f}({\bf q})
\leq
\fullshad{f}({\bf q}')
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}')
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}').
\]
The first inequality is because ${\bf q}$ minimizes $\fullshad{f}$, and
the second is from $\fullshad{f}$'s definition
(\eqref{eqn:fullshad-def}).
Therefore, $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Conversely, suppose now
that $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, for some
$\overline{\ee}\in{{\cal E}_n}$.
Then for all ${\bf x}\in{\R^n}$,
\begin{equation} \label{eqn:pr:min-fullshad-is-finite-min:1}
\fullshad{f}({\bf q})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
=
\min {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}
\leq
\fullshad{f}({\bf x}).
\end{equation}
Both inequalities follow from
\eqref{eqn:fullshad-def},
and the equality is by assumption.
Therefore, ${\bf q}$ minimizes $\fullshad{f}$.
Finally,
by Proposition~\ref{pr:fext-min-exists},
such a minimizer $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ must exist,
for some $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
As just argued, this implies that ${\bf q}$ attains the minimum of
$\fullshad{f}$.
\qed
An important property of the universal reduction $\fullshad{f}$
is that it is
invariant to reducing at an astron.
In other words, if $g$ is a reduction
of $f$ at some astron, then $\fullshad{g}=\fullshad{f}$.
Because our approach to minimizing ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is based on such reductions,
this will be very useful since it will mean that, to find
$\fullshad{f}$, we can form a reduction $g$ of $f$ at some astron,
and instead focus on the possibly easier problem of finding $\fullshad{g}$.
\begin{theorem} \label{thm:f:5}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous,
let ${\bf v}\in\resc{f}$, and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Then the universal reductions of $f$ and $g$ are identical;
that is,
$\fullshad{g}=\fullshad{f}$.
\end{theorem}
\proof
Let ${\bf x}\in{\R^n}$.
Using Proposition~\ref{pr:fullshad-equivs},
we have
\begin{eqnarray*}
\fullshad{g}({\bf x})
&=&
\min_{\overline{\ee}\in{{\cal E}_n}} \overline{g}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\nonumber
\\
&=&
\min_{\overline{\ee}\in{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
\label{eq:f:3a}
\\
&\geq&
\min_{\overline{\ee}\in{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
\label{eq:f:3b}
\\
&=&
\fullshad{f}({\bf x}).
\nonumber
\end{eqnarray*}
The second equality
by Corollary~\ref{cor:i:1}(\ref{cor:i:1b}),
and the inequality
is because
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}$ is an icon since
$\overline{\ee}$ is
(by Proposition~\ref{pr:i:8}(\ref{pr:i:8-leftsum})).
On the other hand,
\[
\fullshad{f}({\bf x})
=
\min_{\overline{\ee}\in{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
\geq
\min_{\overline{\ee}\in{{\cal E}_n}} \overline{g}(\overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
=
\fullshad{g}({\bf x}),
\]
where the inequality is by
Corollary~\ref{cor:i:1}(\ref{cor:i:1c}).
\qed
We will see that the universal reduction $\fullshad{f}$
is effectively restricted to
a linear subspace of ${\R^n}$ in the sense that it is constant in all
directions that are perpendicular to it.
To explain this in more detail,
for any set $S\subseteq\extspac{n}$,
let us define ${S^{\bot}}$ to be the set of vectors ${\bf u}\in{\R^n}$
perpendicular to all the points in $S$, that is,
\begin{equation} \label{eqn:sperp-def}
{S^{\bot}} = \{{\bf u}\in{\R^n} : \forall \overline{\xx}\in S,\; \overline{\xx}\cdot{\bf u}=0\}.
\end{equation}
For a set $S\subseteq\extspac{n}$, we also write ${S^{\bot\bot}}$ for
$({S^{\bot}})^\bot$.
We pause in our development to state some straightforward
properties of these sets:
\begin{proposition} \label{pr:perp-props}
Let $S, U\subseteq\extspac{n}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:perp-props:a}
${S^{\bot}}$ is a linear subspace of ${\R^n}$.
\item \label{pr:perp-props:b}
If $S\subseteq U$ then ${U^{\bot}}\subseteq{S^{\bot}}$.
\item \label{pr:perp-props:c}
If $S\subseteq {\R^n}$ then ${S^{\bot\bot}}=\spn S$.
In particular, if $S$ is a linear subspace of ${\R^n}$, then
${S^{\bot\bot}}=S$.
\item \label{pr:perp-props:d}
If $S$ and $U$ are linear subspaces of ${\R^n}$ then
$S={U^{\bot}}$ if and only if ${S^{\bot}}=U$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:perp-props:a}):
Let us first consider the case that $S$ is a singleton, that is,
$S=\{\overline{\xx}\}$ for some $\overline{\xx}\in\extspac{n}$.
Let $L=\{\overline{\xx}\}^\bot$.
Then ${\bf 0}\in L$ since $\overline{\xx}\cdot{\bf 0}=0$.
And if ${\bf u},{\bf v}\in L$, then $\overline{\xx}\cdot{\bf u}=\overline{\xx}\cdot{\bf v}=0$ so
$\overline{\xx}\cdot({\bf u}+{\bf v})=0$ by Proposition~\ref{pr:i:1}; therefore,
${\bf u}+{\bf v}\in L$.
Finally, if ${\bf u}\in \{ \overline{\xx} \}^\bot$ and $\lambda\in{\mathbb{R}}$, then
$\overline{\xx}\cdot(\lambda{\bf u})=\lambda(\overline{\xx}\cdot{\bf u})=0$,
by Proposition~\ref{pr:i:2}, so $\lambda{\bf u}\in L$.
Thus, $L$ is a linear subspace.
For a general set $S\subseteq\extspac{n}$, it now follows that ${S^{\bot}}$
is a linear subspace since it can be expressed as an intersection of
linear subspaces, namely,
\[
{S^{\bot}} = \bigcap_{\overline{\xx}\in S} \{ \overline{\xx} \}^\bot.
\]
Part~(\ref{pr:perp-props:b}):
If ${\bf u}\in{U^{\bot}}$, then $\overline{\xx}\cdot{\bf u}=0$ for all $\overline{\xx}\in U$, and
therefore also for all $\overline{\xx}\in S$.
Thus, ${\bf u}\in{S^{\bot}}$.
Part~(\ref{pr:perp-props:c}):
If ${\bf x}\in S$, then
${\bf x}\cdot{\bf u}=0$ for all ${\bf u}\in{S^{\bot}}$, implying that
${\bf x}\in{S^{\bot\bot}}$.
Thus, $S \subseteq {S^{\bot\bot}}$.
Since ${S^{\bot\bot}}$ is a linear subspace
(by part~(\ref{pr:perp-props:a})), this further implies that
$\spn S \subseteq {S^{\bot\bot}}$.
For the reverse inclusion, suppose
${\bf x}\in{S^{\bot\bot}}$.
Then letting ${\bf y}$ be the projection of ${\bf x}$ onto $\spn S$, we can
write ${\bf x}={\bf y}+{\bf z}$ where ${\bf y}\in\spn S$ and ${\bf z}$ is orthogonal to
$\spn S$, implying ${\bf y}\cdot{\bf z}=0$ and ${\bf z}\in{S^{\bot}}$.
Thus, since ${\bf x}\in{S^{\bot\bot}}$,
\[
0={\bf x}\cdot{\bf z}={\bf y}\cdot{\bf z}+{\bf z}\cdot{\bf z}=\norm{{\bf z}}^2.
\]
So ${\bf z}={\bf 0}$ and ${\bf x}={\bf y}\in\spn S$.
Part~(\ref{pr:perp-props:d}):
If $S={U^{\bot}}$ then
${S^{\bot}}={U^{\bot\bot}}=U$ by
part~(\ref{pr:perp-props:c}).
Applying this fact with $S$ and $U$ reversed yields the converse.
\qed
Returning to the comments above,
in analyzing and minimizing $\fullshad{f}$,
we will see that we can safely focus our attention exclusively on
${(\aresconef)^{\bot}}$, the linear subspace consisting of just those
points that are perpendicular to all of the points in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
This is because $\fullshad{f}$ is constant in all directions
perpendicular to ${(\aresconef)^{\bot}}$, which means that in minimizing
$\fullshad{f}$, we can ignore points not in ${(\aresconef)^{\bot}}$.
Moreover, considering only this restricted domain, we will see that
$\fullshad{f}$'s sublevel sets are all bounded so that its minimizers must
all be within a bounded (and compact) region of ${\R^n}$.
By Proposition~\ref{pr:min-fullshad-is-finite-min},
these same comments apply to
the set of all finite parts of all minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, which is
identical to the set of all minimizers of $\fullshad{f}$.
As a simple example, suppose $f$ is the convex function
\begin{equation} \label{eq:simple-eg-exp-exp-sq}
f({\bf x}) = f(x_1,x_2,x_3) = e^{-x_1} + e^{-x_2} + (x_3 - 1)^2
\end{equation}
for ${\bf x}\in{\mathbb{R}}^3$.
The standard recession cone for this function is
\[
\resc{f} = \mathbb{R}_{\geq 0} \times \mathbb{R}_{\geq 0} \times \{0\}
= \{ \trans{[x_1,x_2,0]} : x_1,x_2\geq 0 \}.
\]
The astral recession cone turns out to be
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$, as can be checked using
Proposition~\ref{pr:f:1} and Theorem~\ref{thm:f:3}
(or alternatively using some of the general results that will be
proved later).
This implies that ${(\aresconef)^{\bot}}$ is the linear subspace
\[
{(\aresconef)^{\bot}} = \{0\} \times \{0\} \times {\mathbb{R}}
= \{ \trans{[0,0,x_3]} : x_3\in{\mathbb{R}} \}.
\]
The universal reduction of $f$ is simply
$\fullshad{f}({\bf x})=(x_3 - 1)^2$, which is effectively a function just of
points in ${(\aresconef)^{\bot}}$ in the sense that it is constant in all
directions perpendicular to ${(\aresconef)^{\bot}}$, that is, directions of the
form $\trans{[x_1,x_2,0]}$ where $x_1,x_2\in{\mathbb{R}}$; in other words,
$f(x_1,x_2,x_3)=f(0,0,x_3)$ for all $x_1,x_2,x_3\in{\mathbb{R}}$.
Further, it can be seen that $\fullshad{f}$ has bounded sublevel sets
when restricted to ${(\aresconef)^{\bot}}$.
The only minimizer of $\fullshad{f}$ in ${(\aresconef)^{\bot}}$ is
$\trans{[0,0,1]}$; so the entire set of minimizers of $\fullshad{f}$
(and therefore the set of all finite parts of
all minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$) is
\[
\trans{[0,0,1]} + {(\aresconef)^{\bot\bot}}
=
\{ \trans{[x_1,x_2,1]} : x_1, x_2 \in {\mathbb{R}} \}.
\]
Returning to the general case,
as a next step in proving the properties discussed above,
we show that, like $\fullshad{f}$,
${(\aresconef)^{\bot}}$ is invariant to reducing at an astron:
\begin{theorem} \label{thm:f:6}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous,
let ${\bf v}\in\resc{f}$, and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Then ${(\aresconeg)^{\bot}}={(\aresconef)^{\bot}}$.
\end{theorem}
\proof
By Theorem~\ref{thm:f:3},
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}\subseteq{\arescone \overline{g}}$,
which implies
${(\aresconeg)^{\bot}}\subseteq{(\aresconef)^{\bot}}$
by
Proposition~\ref{pr:perp-props}(\ref{pr:perp-props:b}).
To prove the reverse inclusion,
suppose ${\bf u}\in{(\aresconef)^{\bot}}$.
Let $\overline{{\bf y}}\in{\arescone \overline{g}}$.
Then $\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
by Theorem~\ref{thm:f:3}(\ref{thm:f:3b}), so
$\limray{{\bf v}}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\cdot{\bf u}=0$.
This can only be possible if ${\bf v}\cdot{\bf u}=0$
and $\overline{{\bf y}}\cdot{\bf u}=0$.
Thus, ${\bf u}\in{(\aresconeg)^{\bot}}$, proving
${(\aresconef)^{\bot}}\subseteq{(\aresconeg)^{\bot}}$.
\qed
Before continuing, we recall that the \emph{constancy space} of a
convex function $f:{\R^n}\rightarrow\overline{\R}$, which we denote $\conssp{f}$,
consists of those directions in which the value of $f$ remains constant:
\[
\conssp{f} =
\{{\bf v}\in{\R^n} : \forall {\bf x}\in{\R^n}, f({\bf x}+{\bf v}) = f({\bf x})\}.
\]
Plainly, the function $f$ is constant in the direction of
${\bf v}\in{\R^n}$ if and only if $f$ is neither increasing nor
decreasing in direction ${\bf v}$, that is, if and only if both
${\bf v}$ and $-{\bf v}$ are in the recession cone, $\resc{f}$.
Thus,
\begin{equation} \label{eq:cons-isresccap-resc}
\conssp{f}=(\resc{f}) \cap (-\resc{f}),
\end{equation}
implying, in particular, that
\begin{equation} \label{eq:conssp-in-resc}
{\bf 0}\in\conssp{f}\subseteq\resc{f}.
\end{equation}
Furthermore, $\conssp{f}$ is a linear subspace of ${\R^n}$
\citep[see][Theorem~2.7]{ROC}.
We say that a direction ${\bf v}$ is \emph{strictly recessive} if it is in
$(\resc{f})\setminus(\conssp{f})$, meaning that the function never
increases in direction ${\bf v}$, and is strictly decreasing in that
direction at at least one point.
\subsection{Constructing the universal reduction}
\label{sec:construct-univ-reduct}
As we show next,
the universal reduction $\fullshad{f}$
can be constructed using a process very
similar to the one in
Figure~\ref{fig:all-min-proc}, modified only in the termination
condition for the main loop.
Recall that that process constructs a sequence of reductions
$g_i$, each the reduction of $g_{i-1}$ at astron $\limray{{\bf v}_i}$,
for some ${\bf v}_i\in\resc{g_{i-1}}$.
Before reducing, $g_{i-1}$ is decreasing, or at least not increasing,
in the direction ${\bf v}_i$; after reducing, the new function $g_i$ is
constant in direction ${\bf v}_i$, by
Proposition~\ref{pr:d2}(\ref{pr:d2:a}).
In this way, directions in which the original function $f$ is
recessive are successively replaced by reductions that are
constant in those same directions.
At some point, this process might yield a reduction $g_k$ whose only
recessive directions are those in which the function is
constant, meaning $\resc{g_k}=\conssp{g_k}$.
When this happens, $g_k$ must increase to $+\infty$ in any
direction in which it is not a constant, implying that its minimizers are
all finite when all directions in its constancy space
are disregarded.
Indeed, at this point, $g_k$ is exactly $\fullshad{f}$.
\begin{figure}[t]
\underline{Given:}
function $f:{\R^n}\rightarrow\overline{\R}$ that is convex and lower semicontinuous
\smallskip
\underline{Process:}
\begin{itemize}
\item $i \leftarrow 0$
\item $g_0 = f$
\item repeat \emph{at least} until $\resc{g_i} = \conssp{g_i}$
\begin{itemize}
\item $i \leftarrow i+1$
\item let ${\bf v}_i$ be \emph{any} point in
$\resc{g_{i-1}}$
\item $g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$
\end{itemize}
\item $k \leftarrow i$
\item $\overline{\ee}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}$
\item let ${\bf q}\in{\R^n}$ be \emph{any} finite minimizer of $g_k$
\item $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
\end{itemize}
\medskip
\underline{Properties:}
\begin{itemize}
\item $g_k = \fullshad{f}$
\item $\overline{\ee}\in (\unimin{f})\cap{{\cal E}_n} \subseteq ({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}$
\item ${\bf q}$ minimizes $\fullshad{f}$
\item $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
\end{itemize}
\caption{A process for finding the universal reduction $\fullshad{f}$ and all
universal minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.}
\label{fig:min-proc}
\end{figure}
Thus, as shown in Figure~\ref{fig:min-proc},
to find the universal reduction $\fullshad{f}$, we use exactly the same
process as in
Figure~\ref{fig:all-min-proc}, except with a modified termination
condition for the main loop.
Previously, this loop could optionally terminate once the current
reduction $g_i$ has a finite minimizer.
Now, in the new process, the loop can terminate once
$\resc{g_i}=\conssp{g_i}$.
As before, the process is nondeterministic with a similar set of
choices that can be made arbitrarily.
We show later
(Corollary~\ref{cor:fig-cons-main-props}(\ref{cor:fig-cons-main-props:b}))
that if $\resc{g_k}=\conssp{g_k}$, then $g_k$ must have a finite
minimizer ${\bf q}$ as required by the process upon termination of the main
loop.
This also shows that if all the conditions of
Figure~\ref{fig:min-proc} are satisfied for some execution of the process,
then so will be those of Figure~\ref{fig:all-min-proc},
implying that properties proved for the
latter immediately carry over to the former.
In particular, this shows that in constructing $\fullshad{f}$,
the process of Figure~\ref{fig:min-proc}
also yields a point $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ that minimizes
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
This point's finite part, ${\bf q}\in{\R^n}$, is an arbitrary minimizer of
$\fullshad{f}$, which, as discussed above, could be selected by
considering only a compact region of ${\R^n}$.
Its iconic part, $\overline{\ee}\in{{\cal E}_n}$, is in
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, as was the case in Figure~\ref{fig:all-min-proc},
but also has an important property that will be
explored in detail in Section~\ref{subsec:univ-min}.
(In this regard, the figure mentions universal minimizers and
the set $\unimin{f}$, which will both be
introduced in Section~\ref{subsec:univ-min}, and so can be
disregarded for now.)
We proceed now to prove the claims made in the figure and the
preceding discussion.
We begin by showing that the termination condition
of the constancy space
being the same as the recession cone for a convex function $f$
is actually equivalent to the function being equal to its own
universal reduction.
In addition, if $f$ is closed, proper and has all reductions closed, then
these two conditions also are
equivalent to the domain of ${f^*}$ being entirely included in
${(\aresconef)^{\bot}}$.
A more general version of this condition is given shortly as a
corollary.
We prove these results first for a function $f$ by itself; we then apply these
results to get a more general statement regarding the process
in Figure~\ref{fig:min-proc}.
\begin{theorem} \label{thm:f:4}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, closed and proper,
with all reductions closed.
Then the following are equivalent:
\begin{enumerate}
\item \label{thm:f:4a}
$\resc{f}=\conssp{f}$.
\item \label{thm:f:4c}
$\fullshad{f} = f$.
\item \label{thm:f:4b}
$\dom{{f^*}}\subseteq{(\aresconef)^{\bot}}$.
\end{enumerate}
\end{theorem}
\proof
~
\underline{(\ref{thm:f:4a}) $\Rightarrow$ (\ref{thm:f:4b})}:
Assume $\resc{f}=\conssp{f}$.
To show $\dom{{f^*}}\subseteq{(\aresconef)^{\bot}}$,
we prove the following by induction on $k=0,\ldots,n$:
for all $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
and for all ${\bf u}\in\dom{{f^*}}$,
if the astral rank of $\overline{{\bf z}}$ is at most $k$ then
$\overline{{\bf z}}\cdot{\bf u}=0$.
In the base case that $k=0$, $\overline{{\bf z}}={\bf z}$ must be in ${\R^n}$ so
${\bf z}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{\R^n}=\resc{f}$ by Proposition~\ref{pr:f:1}.
Therefore, by Theorem~\ref{thm:rescpol-is-con-dom-fstar},
${\bf z}\cdot{\bf u}\leq 0$
for all ${\bf u}\in\dom{{f^*}}$.
Since $\resc{f}=\conssp{f}$, this applies to $-{\bf z}$ as well, implying
$(-{\bf z})\cdot{\bf u}\leq 0$, and therefore,
$\overline{{\bf z}}\cdot{\bf u}={\bf z}\cdot{\bf u}=0$, as claimed.
For the inductive case that $k>0$,
let $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ have astral rank $k$.
Let ${\bf v}$ be $\overline{{\bf z}}$'s dominant direction so that
$ \overline{{\bf z}} = \limray{{\bf v}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \zbar^{\bot}$
(by Proposition~\ref{pr:h:6}), where
$\zbar^{\bot}$ is $\overline{{\bf z}}$'s projection perpendicular to ${\bf v}$.
Let $g=\fshad{\limray{\vv}}$ be the reduction of $f$ at $\limray{{\bf v}}$.
Since $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
${\bf v}\in\resc{f}=\conssp{f}$ by Theorem~\ref{thm:f:3}(\ref{thm:f:3b}).
We claim $g=f$.
To see this,
let ${\tilde{g}}$ be the ``shadow'' function given in \eqref{eqn:gtil-defn}.
Then ${\tilde{g}}=f$, since ${\bf v}\in\conssp{f}$.
Therefore, $g=\lsc f = f$ by
Theorem~\ref{thm:a10-nunu} and since $f$ is closed.
Thus,
$\zbar^{\bot}\in{\arescone \overline{g}}={\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by a second application of Theorem~\ref{thm:f:3}(\ref{thm:f:3b}).
Let ${\bf u}\in\dom{{f^*}}$.
Since ${\bf v}\in\resc{f}\subseteq{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
(by Proposition~\ref{pr:f:1}), and since ${\bf v}$ has astral rank $0<k$,
${\bf v}\cdot{\bf u}=0$, by inductive hypothesis.
Also, $\zbar^{\bot}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ and has astral rank $k-1$
(by Proposition~\ref{pr:h:6}), so
$\zbar^{\bot}\cdot{\bf u}=0$, again by inductive hypothesis.
Therefore,
$\overline{{\bf z}}\cdot{\bf u}=\limray{{\bf v}}\cdot{\bf u}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\zbar^{\bot}\cdot{\bf u}=0$,
completing the induction and the proof.
\underline{(\ref{thm:f:4b}) $\Rightarrow$ (\ref{thm:f:4c})}:
Assume $\dom{{f^*}}\subseteq{(\aresconef)^{\bot}}$.
Let $\overline{\ee}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap {{\cal E}_n}$.
We claim first that
$-{f^*}({\bf u})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\ee}\cdot{\bf u} = -{f^*}({\bf u})$
for all ${\bf u}\in{\R^n}$.
This is immediate if ${f^*}({\bf u})\in\{-\infty,+\infty\}$.
Otherwise, if ${f^*}({\bf u})\in{\R^n}$, then
${\bf u}\in\dom{{f^*}}\subseteq{(\aresconef)^{\bot}}$, implying that
$\overline{\ee}\cdot{\bf u}=0$ since $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Therefore, by
Theorems~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b})
and~\ref{thm:dub-conj-new},
since $f$ has all reductions closed,
for all ${\bf x}\in{\R^n}$, and all $\overline{\ee}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}$,
\begin{eqnarray*}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
f^{*\dualstar}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
&=&
\sup_{{\bf u}\in{\R^n}}
\brackets{ -{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{\ee}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x}\cdot{\bf u} }
\\
&=&
\sup_{{\bf u}\in{\R^n}}
\brackets{ -{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x}\cdot{\bf u} }
\\
&=&
f^{*\dualstar}({\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})
=
f({\bf x})
\end{eqnarray*}
(using Proposition~\ref{pr:h:1}(\ref{pr:h:1a})).
Thus, by Proposition~\ref{pr:fullshad-equivs},
\[
f({\bf x})
=
\min_{\overline{\ee}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
\fullshad{f}({\bf x}).
\]
\qed
\underline{(\ref{thm:f:4c}) $\Rightarrow$ (\ref{thm:f:4a})}:
We prove this in the contrapositive.
Suppose $\resc{f}\neq\conssp{f}$.
Then there exists a strictly recessive direction
${\bf d}\in(\resc{f})\setminus(\conssp{f})$.
Since ${\bf d}\not\in\conssp{f}$, there exists some ${\bf x}\in{\R^n}$ for which
$f({\bf d}+{\bf x}) \neq f({\bf x})$, which implies, since ${\bf d}\in\resc{f}$, that
$f({\bf d}+{\bf x}) < f({\bf x})$.
By Proposition~\ref{pr:f:1}, ${\bf d}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, so
\[ \fullshad{f}({\bf x}) \leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf d}+{\bf x}) \leq f({\bf d}+{\bf x}) < f({\bf x}) \]
(with the first two equalities following from
Proposition~\ref{pr:fullshad-equivs} and
Proposition~\ref{pr:h:1}(\ref{pr:h:1a})).
This proves the result.
\qed
Using the technique developed in Section~\ref{sec:exp-comp}, we
immediately obtain a generalization of Theorem~\ref{thm:f:4} that
only requires that $f$ is convex and lower semicontinuous.
To obtain this generalization,
the dual condition~(\ref{thm:f:4b}) is replaced by the condition that
$\slopes{f}$ is entirely included in ${(\aresconef)^{\bot}}$.
\begin{corollary} \label{cor:thm:f:4:1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then the following are equivalent:
\begin{enumerate}
\item \label{cor:thm:f:4:1a}
$\resc{f}=\conssp{f}$.
\item \label{cor:thm:f:4:1c}
$\fullshad{f} = f$.
\item \label{cor:thm:f:4:1b}
$\slopes{f}\subseteq{(\aresconef)^{\bot}}$.
\end{enumerate}
\end{corollary}
\proof
If $f\equiv +\infty$,
then the corollary holds immediately since in this case,
$\resc{f}=\conssp{f}={\R^n}$,
$\slopes{f}=\{{\bf 0}\}={(\aresconef)^{\bot}}$,
and
$\fullshad{f}=f$.
We therefore assume henceforth that
$f\not\equiv +\infty$.
Let $f' = {\overline{\exp}}\circ f$.
Then
$f'$ is convex, lower-bounded and lower semicontinuous
by Proposition~\ref{pr:j:2}(\ref{pr:j:2a},\ref{pr:j:2lsc}),
proper and closed (since $f'\geq 0$ and
$f\not\equiv+\infty$),
and also has all reductions
closed (by Proposition~\ref{pr:j:1}(\ref{pr:j:1c})).
Therefore, the three conditions of Theorem~\ref{thm:f:4},
applied to $f'$, are equivalent to each other.
We show that each of these conditions is indivually equivalent to the
three conditions of the corollary.
First, $\resc{f'}=\resc{f}$,
by Proposition~\ref{pr:j:2}(\ref{pr:j:2b}), so
$\conssp{f'}=\conssp{f}$,
using \eqref{eq:cons-isresccap-resc};
thus,
$\resc{f'}=\conssp{f'}$
if and only if
$\resc{f}=\conssp{f}$.
Next, ${\arescone \fpext}={\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by Proposition~\ref{pr:j:2}(\ref{pr:j:2d}),
and $\dom{{f'^*}}=\slopes{f}$
by Proposition~\ref{pr:dom-fpstar}
(since $f\not\equiv+\infty$); thus,
$\dom{{f'^*}}\subseteq{(\aresconefp)^{\bot}}$
if and only if
$\slopes{f}\subseteq{(\aresconef)^{\bot}}$.
Finally, we claim that
$\fullshad{f'}={\overline{\exp}}\circ \fullshad{f}$; that is,
for all ${\bf x}\in{\R^n}$,
$\fullshad{f'}({\bf x})= {\overline{\exp}}(\fullshad{f}({\bf x}))$.
First, by Propostion~\ref{pr:fullshad-equivs}, there exists
$\overline{\ee}'\in{{\cal E}_n}$ realizing the minimum defining $\fullshad{f'}({\bf x})$.
Therefore,
\[
\fullshad{f'}({\bf x})
=
\fpext(\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}))
\geq
{\overline{\exp}}(\fullshad{f}({\bf x})),
\]
where the second equality is by
Proposition~\ref{pr:j:2}(\ref{pr:j:2c}),
and the inequality is by \eqref{eqn:fullshad-def} and
because ${\overline{\exp}}$ is
strictly increasing.
Similarly, again using Propostion~\ref{pr:fullshad-equivs},
there exists $\overline{\ee}\in{{\cal E}_n}$ realizing the minimum defining
$\fullshad{f}({\bf x})$, so
\[
{\overline{\exp}}(\fullshad{f}({\bf x}))
=
{\overline{\exp}}({}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}))
=
\fpext(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\geq
\fullshad{f'}({\bf x}).
\]
Thus,
$\fullshad{f'}=f'$
if and only if
$\fullshad{f}=f$
(since ${\overline{\exp}}$ is strictly increasing).
Combining now yields the corollary.
\qed
We can now apply these results more directly to the procedure outlined
above and in Figure~\ref{fig:min-proc}, thereby justifying our
termination condition.
In particular, once the termination criterion that
$\resc{g_k}=\conssp{g_k}$ has been reached, the next corollary shows
that $g_k=\fullshad{f}$.
\begin{corollary} \label{cor:f:4}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Let $g_0=f$, let ${\bf v}_i\in\resc{g_{i-1}}$, and let
$g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$
for $i=1,\ldots,k$.
Then the following are equivalent:
\begin{enumerate}
\item \label{cor:f:4a}
$\resc{g_k}=\conssp{g_k}$.
\item \label{cor:f:4c}
$g_k=\fullshad{f}$.
\item \label{cor:f:4b}
$\slopes{g_k}\subseteq{(\aresconef)^{\bot}}$.
\end{enumerate}
\end{corollary}
\proof
By Theorems~\ref{thm:f:5} and~\ref{thm:f:6}, applied repeatedly,
${(\aresconef)^{\bot}}=\perpresgsub{k}$, and
$\fullshad{f}={g^{\diamond}_k}$.
Furthermore, $g_k$ is lower semicontinuous, by
Corollary~\ref{cor:i:1}(\ref{cor:i:1a}).
With these facts, the corollary follows immediately from
Corollary~\ref{cor:thm:f:4:1}.
\qed
As discussed earlier, the universal reduction $\fullshad{f}$ must
realize its minimum at some point in a bounded region of ${\R^n}$.
More specifically, we now show that $\fullshad{f}$ is constant in all
directions perpendicular to the linear subspace ${(\aresconef)^{\bot}}$, which
means that, effectively, we can restrict attention only to
${(\aresconef)^{\bot}}$.
Furthermore, within ${(\aresconef)^{\bot}}$, all of the sublevel sets of
$\fullshad{f}$ are bounded and consequently compact,
which means that all minimizers must also be
in a compact region of ${\R^n}$ (indeed, in any nonempty sublevel set).
In Lemmas~\ref{lem:thm:f:4x:1} and~\ref{lem:thm:f:4x:2},
we prove these properties first under the restrictive assumption that
$\resc{f}=\conssp{f}$, that is, the termination
criterion used in Figure~\ref{fig:min-proc}.
These lemmas are stated in terms of $f$, but also are implicitly about
$\fullshad{f}$ since $f=\fullshad{f}$ when $\resc{f}=\conssp{f}$, by
Corollary~\ref{cor:thm:f:4:1}.
We then prove, in Theorem~\ref{thm:f:4x},
that the same properties therefore hold in general.
\begin{lemma} \label{lem:thm:f:4x:1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous, and
assume $\resc{f}=\conssp{f}$.
Then $f$ is constant exactly in those directions
that are perpendicular to ${(\aresconef)^{\bot}}$.
That is, $\conssp{f}={(\aresconef)^{\bot\bot}}$, and thus,
${(\aresconef)^{\bot}} = \consspperp{f} = \rescperp{f}$.
\end{lemma}
\proof
If $f\equiv+\infty$, then the result is immediate since
$\resc{f}=\conssp{f}={(\aresconef)^{\bot\bot}}={\R^n}$ in this case.
So we assume henceforth that $f\not\equiv+\infty$.
By \eqref{eq:conssp-in-resc} and Proposition~\ref{pr:f:1},
$\conssp{f}\subseteq\resc{f}\subseteq{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Applying
Proposition~\ref{pr:perp-props}(\ref{pr:perp-props:b})
then implies
${(\aresconef)^{\bot}}\subseteq\consspperp{f}$,
and so
\begin{equation} \label{eq:j:5}
\conssp{f}=\consspperperp{f}\subseteq{(\aresconef)^{\bot\bot}}
\end{equation}
by
Proposition~\ref{pr:perp-props}(\ref{pr:perp-props:b},\ref{pr:perp-props:c}).
To prove the reverse inclusion,
suppose first that $f$ is closed, proper, and has all reductions closed.
Let ${\bf d}\in{(\aresconef)^{\bot\bot}}$.
By Theorem~\ref{thm:f:4}, for ${\bf u}\in{\R^n}$, if ${f^*}({\bf u})\in{\mathbb{R}}$ then
${\bf u}\in{(\aresconef)^{\bot}}$, which implies that ${\bf d}\cdot{\bf u}=0$.
Thus, for all ${\bf u}\in{\R^n}$,
$-{f^*}({\bf u})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf d}\cdot{\bf u} = -{f^*}({\bf u})$.
Therefore, by
Theorems~\ref{thm:fext-dub-sum}(\ref{thm:fext-dub-sum:b})
and~\ref{thm:dub-conj-new},
and also Proposition~\ref{pr:h:1}(\ref{pr:h:1a}),
\begin{eqnarray*}
f({\bf d}+{\bf x})
&=&
\sup_{{\bf u}\in{\R^n}}
\brackets{ -{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf d}\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x}\cdot{\bf u} }
\\
&=&
\sup_{{\bf u}\in{\R^n}}
\brackets{ -{f^*}({\bf u}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x}\cdot{\bf u} }
\\
&=&
f({\bf x})
\end{eqnarray*}
for all ${\bf x}\in{\R^n}$.
In other words, ${\bf d}\in\conssp{f}$, proving
${(\aresconef)^{\bot\bot}}\subseteq\conssp{f}$
when $f$ has all reductions closed.
More generally, given $f$ that does not necessarily satisfy these
additional conditions, we
can use our usual trick of defining
$f'={\overline{\exp}}\circ f$,
similar to the proof of Corollary~\ref{cor:thm:f:4:1}.
Then
$f'$ is convex, lower-bounded, and lower semicontinuous
by Proposition~\ref{pr:j:2}(\ref{pr:j:2a},\ref{pr:j:2lsc}),
and so also is proper and closed,
and has
all reductions closed (by Proposition~\ref{pr:j:1}(\ref{pr:j:1c})).
So, by the above argument applied to $f'$,
\[
{(\aresconef)^{\bot\bot}}
=
{(\aresconefp)^{\bot\bot}}
\subseteq
\conssp{f'}
=
\conssp{f}
\]
by Proposition~\ref{pr:j:2}(\ref{pr:j:2b},\ref{pr:j:2d})
and \eqref{eq:cons-isresccap-resc}.
Thus, $\conssp{f}={(\aresconef)^{\bot\bot}}$, which is equivalent to the last
claim of the lemma by
Proposition~\ref{pr:perp-props}(\ref{pr:perp-props:d})
(since $\conssp{f}$ and ${(\aresconef)^{\bot}}$ are linear subspaces).
\qed
\begin{lemma} \label{lem:thm:f:4x:2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous, and
assume $\resc{f}=\conssp{f}$.
If restricted to ${(\aresconef)^{\bot}}$, all of $f$'s sublevel
sets are compact; that is, for all $\beta\in{\mathbb{R}}$, the set
\[ \{ {\bf x} \in {(\aresconef)^{\bot}} : f({\bf x}) \leq \beta \} \]
is a compact subset of ${\R^n}$.
\end{lemma}
To prove this lemma, we will first prove the more general result given
in Theorem~\ref{thm:lem:thm:f:4x:2}.
Note that this theorem only concerns concepts from standard convex
analysis, and could be proved using standard techniques
(indeed, closely related results are given, for instance, in
\citet[Section~8]{ROC}).
Here we instead give a more direct proof as an illustration of the
techniques developed in this manuscript.
\begin{theorem} \label{thm:lem:thm:f:4x:2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
If restricted to $\rescperp{f}$, all of $f$'s sublevel
sets are compact; that is, for all $\beta\in{\mathbb{R}}$, the set
\[ \{ {\bf x} \in \rescperp{f} : f({\bf x}) \leq \beta \} \]
is a compact subset of ${\R^n}$.
\end{theorem}
\proof
Let $\beta\in{\mathbb{R}}$, and let
\[ L = \{ {\bf x} \in \rescperp{f} : f({\bf x}) \leq \beta \}. \]
We claim first that $L$ is bounded.
Suppose not.
Then there exists a sequence
$\seq{{\bf x}_t}$ in $\rescperp{f}$ with
$f({\bf x}_t)\leq \beta$ for all $t$, and such that
$\norm{{\bf x}_t}\rightarrow+\infty$.
By sequential compactness
of $\extspac{n}$, there exists a subsequence of the
${\bf x}_t$'s that converges to some point $\overline{\xx}\in\extspac{n}$;
by discarding all other elements,
we can assume the entire sequence converges to
$\overline{\xx}$.
We can write $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some
$\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$
(by Proposition~\ref{pr:i:8}(\ref{pr:i:8c})).
Further, let ${\bf v}$ be the dominant direction of $\overline{\ee}$
so that $\overline{\ee}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\ebar^{\bot}$
(where $\ebar^{\bot}$ is $\overline{\ee}$'s projection perpendicular to ${\bf v}$,
by Proposition~\ref{pr:h:6}).
Then
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
\leq
\liminf f({\bf x}_t)
\leq \beta.
\]
Therefore, $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by Theorem~\ref{thm:f:2},
since $\overline{\ee}$ is an icon.
This further implies that ${\bf v}\in\resc{f}\subseteq{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by Theorem~\ref{thm:f:3}(\ref{thm:f:3b}) and
Proposition~\ref{pr:f:1}.
Thus, for all $t$,
${\bf x}_t\cdot{\bf v}=0$ since
${\bf x}_t\in\rescperp{f}$.
On the other hand,
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\ebar^{\bot}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$,
so ${\bf v}$ is also $\overline{\xx}$'s dominant direction.
Therefore, since ${\bf x}_t\rightarrow\overline{\xx}$,
\[
{\bf x}_t\cdot{\bf v} \rightarrow \overline{\xx}\cdot{\bf v} = +\infty,
\]
by Theorem~\ref{thm:dom-dir}.
But this is a contradiction, since ${\bf x}_t\cdot{\bf v}=0$ for all $t$.
Therefore, $L$ is bounded.
We claim next that $L$ is closed (where, throughout this discussion, ``closed''
means in ${\R^n}$, not $\extspac{n}$).
This is because $L$ is equal to the intersection of
two closed sets, specifically,
\[
L
=
\rescperp{f}\cap \{ {\bf x} \in {\R^n} : f({\bf x}) \leq \beta \}.
\]
These sets are closed because
$\rescperp{f}$ is a linear subspace
(by Proposition~\ref{pr:perp-props}(\ref{pr:perp-props:a}))
and therefore closed,
while the rightmost set is a sublevel set
which is closed since $f$ is
lower semicontinuous
\citep[Theorem~7.1]{ROC}.
Thus, $L$ is compact, being closed (in ${\R^n}$) and bounded.
\qed
\proof[Proof of Lemma~\ref{lem:thm:f:4x:2}]
The lemma is immediate from Theorem~\ref{thm:lem:thm:f:4x:2}
after noting,
by Lemma~\ref{lem:thm:f:4x:1},
that ${(\aresconef)^{\bot}} = \rescperp{f}$
since $\resc{f}=\conssp{f}$.
\qed
We now can prove in full generality the properties discussed above for
$\fullshad{f}$'s sublevel sets and constancy space.
\begin{theorem} \label{thm:f:4x}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then the following hold:
\begin{enumerate}
\item \label{thm:f:4xaa}
$\fullshad{f}$ is convex and lower semicontinuous.
\item \label{thm:f:4xa}
$\fullshad{f}$ is constant exactly in those directions
that are perpendicular to ${(\aresconef)^{\bot}}$.
That is, $\conssp{\fullshad{f}}={(\aresconef)^{\bot\bot}}$.
\item \label{thm:f:4xb}
If restricted to ${(\aresconef)^{\bot}}$, all of $\fullshad{f}$'s sublevel
sets are compact; that is, for all $\beta\in{\mathbb{R}}$, the set
\begin{equation} \label{eqn:thm:f:4x:1}
\{ {\bf x} \in {(\aresconef)^{\bot}} : \fullshad{f}({\bf x}) \leq \beta \}
\end{equation}
is a compact subset of ${\R^n}$.
\end{enumerate}
\end{theorem}
\proof
The theorem concerns only ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ and $\fullshad{f}$, and furthermore,
the latter is defined entirely in terms of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=\overline{(\lsc f)}$
(by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa})),
it therefore suffices to prove the theorem under the assumption that
$f$ is lower semicontinuous, replacing it with $\lsc f$ if it is not.
As such,
we prove the following statement by \emph{backwards} induction on
$\ell=0,1,\ldots,n+1$:
For all convex and lower semicontinuous functions
$f:{\R^n}\rightarrow\overline{\R}$,
if $\dim(\conssp{f})\geq\ell$ then
parts~(\ref{thm:f:4xaa}),~(\ref{thm:f:4xa}) and~(\ref{thm:f:4xb}) of the
theorem hold.
In the base case that $\ell=n+1$, the claimed statement is vacuously
true since $\dim(\conssp{f})\leq n$ always.
For the inductive step, assume $\ell\leq n$, and that the claim holds
for $\ell+1$.
Suppose $f$ is convex and lower semicontinuous, and that
$\dim(\conssp{f})\geq\ell$.
If $\resc{f}=\conssp{f}$, then
parts~(\ref{thm:f:4xa}) and~(\ref{thm:f:4xb}) of the theorem follow
immediately from
Lemmas~\ref{lem:thm:f:4x:1} and~\ref{lem:thm:f:4x:2}, after noting
that $f=\fullshad{f}$ by Corollary~\ref{cor:thm:f:4:1}.
This last fact also implies part~(\ref{thm:f:4xaa}) since $f$ is convex
and lower semicontinuous.
Otherwise, $\resc{f}\neq\conssp{f}$, so
there must exist ${\bf v}\in(\resc{f})\setminus(\conssp{f})$.
Let
$g=\fshad{\limray{\vv}}$
be the reduction at $\limray{{\bf v}}$.
We claim $\dim(\conssp{g}) > \dim(\conssp{f})$.
To see this,
by Theorem~\ref{thm:f:3}(\ref{thm:f:3a}),
$\resc{f} \subseteq \resc{g}$, implying,
by \eqref{eq:cons-isresccap-resc}, that
$ \conssp{f} \subseteq \conssp{g}$.
Also, by Proposition~\ref{pr:d2}(\ref{pr:d2:a}),
$g({\bf x}+\lambda{\bf v})=g({\bf x})$ for all ${\bf x}\in{\R^n}$ and all $\lambda\in{\mathbb{R}}$.
Therefore, ${\bf v}$ is in $\conssp{g}$, the constancy space of $g$, but
by assumption, ${\bf v}\not\in\conssp{f}$.
Combining, these show that $\conssp{f}$ is a proper subset of
$\conssp{g}$.
Since they are both linear subspaces, this implies
$\dim(\conssp{g}) > \dim(\conssp{f})$, as claimed.
Thus, $\dim(\conssp{g})\geq \dim(\conssp{f}) + 1 \geq \ell+1$, and so
our inductive hypothesis holds for $g$
(using Theorem~\ref{thm:a10-nunu}).
Also,
$\fullshad{g}=\fullshad{f}$ and
${(\aresconeg)^{\bot}}={(\aresconef)^{\bot}}$
by Theorems~\ref{thm:f:5} and~\ref{thm:f:6}, respectively.
Therefore,
\[
\conssp{\fullshad{f}}
=
\conssp{\fullshad{g}}
=
{(\aresconeg)^{\bot\bot}}
=
{(\aresconef)^{\bot\bot}}
\]
where the middle equality is by inductive hypothesis,
proving part~(\ref{thm:f:4xa}).
Similarly,
\[
\{ {\bf x} \in {(\aresconef)^{\bot}} : \fullshad{f}({\bf x}) \leq \beta \}
=
\{ {\bf x} \in {(\aresconeg)^{\bot}} : \fullshad{g}({\bf x}) \leq \beta \}
\]
and so is compact, by inductive hypothesis,
proving part~(\ref{thm:f:4xb}).
And part~(\ref{thm:f:4xaa}) follows from $\fullshad{f}=\fullshad{g}$,
by inductive hypothesis.
\qed
Thus, if $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
where $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$, then
$\overline{\ee}$ must be in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, by
Theorem~\ref{thm:arescone-fshadd-min}, and ${\bf q}$ must minimize
$\fullshad{f}$, by
Proposition~\ref{pr:min-fullshad-is-finite-min}.
Furthermore, as a consequence of Theorem~\ref{thm:f:4x},
${\bf q}$ can effectively be restricted to
a compact subset of
${(\aresconef)^{\bot}}$.
Regarding sequences, Theorem~\ref{thm:f:4x} implies that if a convex
function $f$ is minimized by some sequence, then
that sequence must
also minimize $\fullshad{f}$, as does also the projection of that
sequence onto the linear subspace ${(\aresconef)^{\bot}}$.
Further, that projected sequence cannot be unbounded.
\begin{proposition} \label{pr:proj-mins-fullshad}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Let $\seq{{\bf x}_t}$ be a sequence in ${\R^n}$, and for each $t$, let
${\bf q}_t$ be the projection of ${\bf x}_t$ onto the linear subspace
${(\aresconef)^{\bot}}$.
Assume $f({\bf x}_t)\rightarrow\inf f$.
Then
$\fullshad{f}({\bf x}_t)=\fullshad{f}({\bf q}_t)\rightarrow \min\fullshad{f}$.
Furthermore, the entire sequence $\seq{{\bf q}_t}$ is included in a
compact subset of ${\R^n}$.
\end{proposition}
\proof
If $f\equiv+\infty$, then $\fullshad{f}\equiv+\infty$ and
${(\aresconef)^{\bot}}=\{{\bf 0}\}$ so ${\bf q}_t={\bf 0}$ for all $t$, implying the
claim.
Therefore, we assume henceforth that $f\not\equiv+\infty$.
For all $t$,
\begin{equation} \label{eq:pr:proj-mins-fullshad:1}
\inf f
=
\min {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}
\leq
\min \fullshad{f}
\leq
\fullshad{f}({\bf q}_t)
=
\fullshad{f}({\bf x}_t)
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x}_t)
\leq
f({\bf x}_t).
\end{equation}
The first equality is by
Proposition~\ref{pr:fext-min-exists}.
The first inequality follows from the definition of $\fullshad{f}$
(\eqref{eqn:fullshad-def}).
The second equality follows from
Theorem~\ref{thm:f:4x}(\ref{thm:f:4xa})
(since ${\bf x}_t-{\bf q}_t$ is perpendicular to $\rescperp{f}$).
The third inequality follows also from the definition of
$\fullshad{f}$, since ${\bf 0}$ is an icon.
And the last inequality is from
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
Since $f({\bf x}_t)\rightarrow\inf f$,
\eqref{eq:pr:proj-mins-fullshad:1}
implies
$\fullshad{f}({\bf x}_t)=\fullshad{f}({\bf q}_t) \rightarrow \min \fullshad{f}$,
as claimed.
This also shows,
for any $\beta>\min\fullshad{f}$,
that all but finitely many of the ${\bf q}_t$'s are
in some sublevel set of $\fullshad{f}$, as in
\eqref{eqn:thm:f:4x:1}.
By Theorem~\ref{thm:f:4x}(\ref{thm:f:4xb}), every such sublevel set is
compact.
Therefore, there exists a (possibly larger) compact subset of ${\R^n}$
that includes
the entire sequence $\seq{{\bf q}_t}$.
\qed
\subsection{Ensuring termination}
\label{sec:ensure-termination}
As discussed in Section~\ref{sec:find-all-min},
the processes in
Figures~\ref{fig:all-min-proc} and~\ref{fig:min-proc}
might never terminate since, for instance, we might arbitrarily choose
${\bf v}_i={\bf 0}$ on every iteration of the main loop.
In general, if a vector ${\bf v}_i$ is chosen that is already in the span
of the preceeding vectors ${\bf v}_1,\ldots,{\bf v}_{i-1}$, then no progress
is made in the sense that the icon that is being constructed has not
changed; that is,
\[
\limrays{{\bf v}_1,\ldots,{\bf v}_i} = \limrays{{\bf v}_1,\ldots,{\bf v}_{i-1}}
\]
(by Proposition~\ref{pr:g:1}).
Thus, to ensure progress, we might insist that ${\bf v}_i$ be chosen
to be not only in the recession cone of $g_{i-1}$ but also
outside the span of ${\bf v}_1,\ldots,{\bf v}_{i-1}$ so that
\begin{equation} \label{eq:sensible-vi-dfn}
{\bf v}_i \in (\resc{g_{i-1}}) \setminus
\spnfin{{\bf v}_1,\ldots,{\bf v}_{i-1}}.
\end{equation}
We say such a choice of ${\bf v}_i$ is \emph{sensible}.
The next proposition shows that if no sensible choice is possible then
the termination condition of Figure~\ref{fig:min-proc} must have
already been reached:
\begin{proposition} \label{pr:no-sense-then-done}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Let $g_0=f$, let ${\bf v}_i\in\resc{g_{i-1}}$, and let
$g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$,
for $i=1,\ldots,k$.
Suppose
\[
\resc{g_k} \subseteq \spnfin{{\bf v}_1,\ldots,{\bf v}_k}.
\]
Then $\resc{g_k}=\conssp{g_k}$.
\end{proposition}
\proof
We claim by induction on $i=0,\ldots,k$ that
\[
{{\bf v}_1,\ldots,{\bf v}_i} \subseteq \conssp{g_i}.
\]
The base case that $i=0$ holds trivially.
For the inductive step, suppose $i>0$.
Then ${\bf v}_1,\ldots,{\bf v}_{i-1}\in\conssp{g_{i-1}}$
by inductive hypothesis.
Furthermore,
$\resc{g_{i-1}}\subseteq\resc{g_i}$
by Theorem~\ref{thm:f:3}(\ref{thm:f:3a}),
implying
$\conssp{g_{i-1}}\subseteq\conssp{g_i}$
by \eqref{eq:cons-isresccap-resc}.
In addition, for all ${\bf x}\in{\R^n}$,
$g_i({\bf x}+{\bf v}_i)=g_i({\bf x})$
by
Proposition~\ref{pr:d2}(\ref{pr:d2:a}), implying
${\bf v}_i\in\conssp{g_i}$.
Combining completes the induction.
Therefore,
\[
\conssp{g_k}
\subseteq
\resc{g_k}
\subseteq
\spnfin{{\bf v}_1,\ldots,{\bf v}_k}
\subseteq
\conssp{g_k}.
\]
These inclusions follow respectively from
\eqref{eq:cons-isresccap-resc},
by assumption,
and from the inductive claim proved above
(since $\conssp{g_k}$ is a linear subspace).
\qed
If, in the process of Figure~\ref{fig:min-proc},
each ${\bf v}_i$ is chosen sensibly, then the dimension of the space
spanned by the ${\bf v}_i$'s increases by one on each iteration of the
main loop.
Therefore, within $n$ iterations, no more sensible choices can be
possible, and therefore, by Proposition~\ref{pr:no-sense-then-done},
the termination condition must have been reached.
This shows that the process can always be run in a way that guarantees
termination within $n$ iterations.
Furthermore, it shows that the process cannot ``get stuck'' in the
sense that, no matter what preceeding choices have been made by the
processs, the ensuing choices of ${\bf v}_i$ can be made sensibly,
again ensuring termination within $n$ additional iterations.
The next corollary
summarizes some of the main properties of the construction in
Figure~\ref{fig:min-proc}.
In particular,
part~(\ref{cor:fig-cons-main-props:b})
shows that the termination condition of
Figure~\ref{fig:min-proc} implies that of
Figure~\ref{fig:all-min-proc},
as mentioned earlier.
Therefore, the comments above regarding termination apply to that
process as well.
\begin{corollary} \label{cor:fig-cons-main-props}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
There exists ${\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$, for some $k\geq 0$,
with ${\bf v}_i\in\resc{g_{i-1}}$, for $i=1,\ldots,k$,
and $\resc{g_k}=\conssp{g_k}$,
where $g_0=f$, and
$g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$.
Whenever these hold, the following are also true:
\begin{enumerate}
\item \label{cor:fig-cons-main-props:a}
For all ${\bf x}\in{\R^n}$,
$\fullshad{f}({\bf x})=g_k({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=\fshad{\ebar}({\bf x})$,
where $\overline{\ee}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}$.
\item \label{cor:fig-cons-main-props:b}
There exists some ${\bf q}\in{\R^n}$ that minimizes $g_k=\fullshad{f}$.
\item \label{cor:fig-cons-main-props:c}
The point $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\end{enumerate}
\end{corollary}
\proof
Suppose the process of Figure~\ref{fig:min-proc} is executed sensibly
so that, on each iteration, ${\bf v}_i$ is chosen to satisfy
\eqref{eq:sensible-vi-dfn}.
Then, as just discussed, within $n$ iterations the process must reach
a point at which no such choice is possible, implying, by
Proposition~\ref{pr:no-sense-then-done}, that the termination
condition of the main loop has been reached.
Upon termination, all of the claimed properties hold.
Part~(\ref{cor:fig-cons-main-props:a}):
Corollary~\ref{cor:f:4} proves that $g_k=\fullshad{f}$.
As in the proof of Theorem~\ref{thm:all-min-proc-correct},
\[ g_i({\bf x}) = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limrays{{\bf v}_1,\ldots,{\bf v}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}) \]
for all ${\bf x}\in{\R^n}$,
for $i=0,\ldots,k$.
Thus,
\[ g_k({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=\fshad{\ebar}({\bf x}), \]
for ${\bf x}\in{\R^n}$.
For the remainder of this proof only, for any point ${\bf x}\in{\R^n}$, let
${\xx^\bot}\in{(\aresconef)^{\bot}}$ denote ${\bf x}$'s projection onto the linear
subspace ${(\aresconef)^{\bot}}$.
By the nature of projection,
this implies ${\bf x}={\xx^\bot}+{\bf y}$ for some ${\bf y}\in{\R^n}$ which is
orthogonal to ${(\aresconef)^{\bot}}$, that is, some ${\bf y}\in{(\aresconef)^{\bot\bot}}$.
Since ${(\aresconef)^{\bot\bot}} = \conssp{\fullshad{f}}$
by Theorem~\ref{thm:f:4x}(\ref{thm:f:4xa}),
this means
$\fullshad{f}({\bf x})=\fullshad{f}({\xx^\bot})$
for all ${\bf x}\in{\R^n}$.
Part~(\ref{cor:fig-cons-main-props:b}):
If $\fullshad{f}\equiv +\infty$, then any point in ${\R^n}$ is a
minimizer.
Otherwise, for some $\beta\in{\mathbb{R}}$, the set
\[
L
=
\{ {\bf x} \in {(\aresconef)^{\bot}} : \fullshad{f}({\bf x}) \leq \beta \}
\]
is not empty (since if $\fullshad{f}({\bf x})\leq \beta$ then
$\fullshad{f}({\xx^\bot})\leq \beta$ as well)
and compact (by Theorem~\ref{thm:f:4x}(\ref{thm:f:4xb})).
Therefore,
since $\fullshad{f}=g_k$ is lower semicontinuous, it attains its minimum
over $L$ at some point ${\bf q}\in{(\aresconef)^{\bot}}$
\citep[Theorem~4.6.2.3]{bourbaki_general_topology_part1}.
Furthermore, ${\bf q}$ must actually minimize $\fullshad{f}$ over all of
${\R^n}$ since if ${\bf x}\in{\R^n}$, then
$\fullshad{f}({\bf x})=\fullshad{f}({\xx^\bot})\geq\fullshad{f}({\bf q})$.
Part~(\ref{cor:fig-cons-main-props:c}):
Having proved part~(\ref{cor:fig-cons-main-props:a}),
this follows directly from Theorem~\ref{thm:all-min-proc-correct},
all of whose conditions are satisfied.
\qed
\subsection{Universal minimizers}
\label{subsec:univ-min}
As seen in Corollary~\ref{cor:fig-cons-main-props},
the construction in Figure~\ref{fig:min-proc}
yields a minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of the form
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$, where ${\bf q}\in{\R^n}$ minimizes the universal reduction
$\fullshad{f}$, and where $\overline{\ee}$ has the property that
\begin{equation} \label{eqn:dbar-fullshad-min}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
\fullshad{f}({\bf x})
=
\min_{\overline{\ee}'\in{{\cal E}_n}} {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}'\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\end{equation}
for all ${\bf x}\in{\R^n}$.
That the minimum that appears here is realized was previously proved
in Proposition~\ref{pr:fullshad-equivs}.
In fact, \eqref{eqn:dbar-fullshad-min} is showing something much
stronger, namely, that $\overline{\ee}$ realizes that
minimum for \emph{all} ${\bf x}$ \emph{simultaneously}, which is fairly
remarkable.
Furthermore, the construction in Figure~\ref{fig:min-proc}
reveals that there is a whole set of
points with this same property, since the construction and proof hold
for a whole range of arbitrary choices, as previously discussed.
Here, we study some of the properties of that set.
We say that $\overline{{\bf z}}\in\extspac{n}$ is a
\emph{universal reducer} for a convex function $f$ if
\begin{equation} \label{eq:unimin-def}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})\leq\fullshad{f}({\bf x})
\end{equation}
for all ${\bf x}\in{\R^n}$.
Thus, adding $\overline{{\bf z}}$ to any ${\bf x}\in{\R^n}$ yields a point
where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is at least as small as the minimum that defines
$\fullshad{f}$ in \eqref{eqn:dbar-fullshad-min}
(or the other equivalent minima in
Proposition~\ref{pr:fullshad-equivs}).
We call such points ``universal'' because of their connection to
universal reductions, and also because they universally attain the
minimum in \eqref{eqn:dbar-fullshad-min} simultaneously for all
${\bf x}\in{\R^n}$.
We write $\unimin{f}$ for the set of all such universal reducers:
\[
\unimin{f}
= \braces{ \overline{{\bf z}}\in\extspac{n} : \forall {\bf x}\in{\R^n},\;
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})\leq\fullshad{f}({\bf x}) }.
\]
Suppose $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where $\overline{\ee}\in{{\cal E}_n}$ and
${\bf q}\in{\R^n}$.
Previously, we saw that if $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ then
$\overline{\ee}$ must be in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
(by Theorem~\ref{thm:arescone-fshadd-min}),
and ${\bf q}$ must be a minimizer of $\fullshad{f}$
(by Proposition~\ref{pr:min-fullshad-is-finite-min}).
The converse is false.
In other words, it is not the case that $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
for every choice of $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ and every ${\bf q}$ that minimizes
$\fullshad{f}$.
For instance, for the function $f$ in
\eqref{eq:simple-eg-exp-exp-sq},
${\bf q}=\trans{[0,0,1]}$ minimizes $\fullshad{f}$,
and
$\overline{\ee}=\limray{{\bf e}_1}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
but
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ does not minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ since
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=1>0=\inf f$.
Nevertheless,
as discussed above, Corollary~\ref{cor:fig-cons-main-props} shows that
the construction of Figure~\ref{fig:min-proc} yields a minimizer
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ of a particular form, namely, with
${\bf q}$ a finite minimizer of $\fullshad{f}$,
and icon $\overline{\ee}$ not only in ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, but also a universal
reducer, as shown in part~(\ref{cor:fig-cons-main-props:a}) of that
corollary.
We call such a point (where $\overline{\ee}\in(\unimin{f})\cap{{\cal E}_n}$
is a universal reducer, and ${\bf q}$ minimizes the universal reduction
$\fullshad{f}$) a \emph{universal minimizer} of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Every such point is indeed a minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
as follows from the next proposition.
Later, we will see that the process of
Figure~\ref{fig:min-proc} finds all of the universal minimizers
(and thereby all of the universal reducers as well).
\begin{proposition} \label{pr:unimin-to-global-min}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Suppose $\overline{{\bf z}}\in\unimin{f}$ and that ${\bf q}\in{\R^n}$ minimizes
$\fullshad{f}$.
Then $\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\end{proposition}
\proof
By Proposition~\ref{pr:min-fullshad-is-finite-min},
since ${\bf q}$ minimizes $\fullshad{f}$, there exists an icon
$\overline{\ee}\in{{\cal E}_n}$ such that $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Then
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
\leq
\fullshad{f}({\bf q})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
=
\min {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}
\]
since $\overline{{\bf z}}\in\unimin{f}$ and by \eqref{eqn:fullshad-def}.
Therefore $\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ also minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\qed
Not all minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ are universal minimizers.
For example, for the function $f$ in \eqref{eqn:recip-fcn-eg},
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is minimized by $\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf e}_2$,
but $\limray{{\bf e}_1}$ is not a universal reducer
(since, for instance,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(-{\bf e}_2))=+\infty$,
but
$\fullshad{f}\equiv 0$).
The next proposition shows that
universal reducers are themselves included in the astral recession
cone, which implies that the inequality in \eqref{eq:unimin-def}
actually is always an equality, for all ${\bf x}\in{\R^n}$.
This also shows that the universal reduction $\fullshad{f}$ of a
function $f$ is itself an iconic reduction.
In particular, applied to any icon
$\overline{\ee}\in{{\cal E}_n}$, the proposition shows that
$\fullshad{f}=\fshad{\ebar}$ (so that $\fullshad{f}$ is the reduction of $f$
at icon $\overline{\ee}$) if and only if $\overline{\ee}$ is a universal reducer.
\begin{proposition} \label{pr:new:thm:f:8a}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then the following hold:
\begin{enumerate}
\item \label{pr:new:thm:f:8a:a}
$\unimin{f}\subseteq{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
\item \label{pr:new:thm:f:8a:b}
Let $\overline{{\bf z}}\in\extspac{n}$.
Then
$\overline{{\bf z}}\in\unimin{f}$ if and only if
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=\fullshad{f}({\bf x})$ for all ${\bf x}\in{\R^n}$.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:new:thm:f:8a:a}):
Let $\overline{{\bf z}}\in\unimin{f}$.
For all ${\bf x}\in{\R^n}$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\fullshad{f}({\bf x})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})
\leq
f({\bf x}).
\]
The first inequality is because $\overline{{\bf z}}$ is a universal reducer.
The second is by definition of $\fullshad{f}$, and since
${\bf 0}\in{{\cal E}_n}$.
And the third is from
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
Therefore, $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ by
Proposition~\ref{pr:alt-arescone-def}.
Part~(\ref{pr:new:thm:f:8a:b}):
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=\fullshad{f}({\bf x})$
for all ${\bf x}\in{\R^n}$, then $\overline{{\bf z}}$ is a universal reducer,
by definition.
For the converse, if $\overline{{\bf z}}\in\unimin{f}$ then $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
as argued above, so
$\fullshad{f}({\bf x})\leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})$,
by Propostion~\ref{pr:fullshad-equivs}, for all ${\bf x}\in{\R^n}$.
Combining with \eqref{eq:unimin-def} completes the proof.
\qed
Let $\overline{{\bf z}}\in\extspac{n}$ which we write as
$\overline{{\bf z}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
If $\overline{{\bf z}}$ is a universal reducer for some function $f$,
then its iconic
part $\overline{\ee}$ must also be a universal reducer, while its real part
${\bf q}$ must be a point in the constancy space of $\fullshad{f}$.
The converse also holds.
Roughly, this means that
it is the iconic part of a point that
makes it a universal reducer, provided the finite part does no harm.
\begin{theorem} \label{thm:new:thm:f:8a}
Let $f:{\R^n}\rightarrow\extspac{n}$ be convex.
Suppose $\overline{{\bf z}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where
$\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
Then $\overline{{\bf z}}\in\unimin{f}$ if and only if $\overline{\ee}\in\unimin{f}$
and ${\bf q}\in\conssp{\fullshad{f}}$.
\end{theorem}
In terms of sets, the theorem is saying that:
\[
\unimin{f} =
((\unimin f) \cap {{\cal E}_n}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \conssp{\fullshad{f}}.
\]
\proof
Suppose first that
$\overline{\ee}\in\unimin{f}$ and ${\bf q}\in\conssp{\fullshad{f}}$.
Then using these two facts, for all ${\bf x}\in{\R^n}$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\fullshad{f}({\bf q}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
\fullshad{f}({\bf x}).
\]
Thus, $\overline{{\bf z}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\in\unimin{f}$, as claimed.
For the converse, suppose now that $\overline{{\bf z}}\in\unimin{f}$.
Then $\overline{{\bf z}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, by
Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:a}).
Let $h=\fshad{\ebar}$, that is,
$h({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})$
for ${\bf x}\in{\R^n}$,
which is convex (Proposition~\ref{pr:i:9}(\ref{pr:i:9a})).
We claim first that ${\bf q}$ is in
$\resc{h}$, the standard recession cone of $h$.
This is because, for all ${\bf x}\in{\R^n}$,
\begin{equation} \label{eqn:thm:new:thm:f:8a:1}
h({\bf x}+{\bf q})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\fullshad{f}({\bf x})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
h({\bf x}).
\end{equation}
The first and second inequality are
because $\overline{{\bf z}}$ is a universal reducer,
and from \eqref{eqn:fullshad-def}.
We next show that actually
\eqref{eqn:thm:new:thm:f:8a:1} holds with equalty.
Note that, since ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a cone
(Corollary~\ref{cor:thm:j:3}) that includes $\overline{{\bf z}}$, it also must include
$2\overline{{\bf z}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax 2{\bf q}$.
Similar to \eqref{eqn:thm:new:thm:f:8a:1}, for all ${\bf x}\in{\R^n}$,
\begin{eqnarray*}
h({\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf x} - {\bf q}))
\\
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf x} - {\bf q}))
\\
&\leq&
\fullshad{f}({\bf x}-{\bf q})
\\
&\leq&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax 2{\bf q} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf x}-{\bf q}))
\\
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax ({\bf x}+{\bf q}))
=
h({\bf x}+{\bf q}).
\end{eqnarray*}
The first and second inequalities are because $\overline{{\bf z}}$ is a
universal reducer, and from Proposition~\ref{pr:fullshad-equivs}
since $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax 2{\bf q} \in {\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
(We also have made liberal use of Propostion~\ref{pr:i:7}.)
Combined with \eqref{eqn:thm:new:thm:f:8a:1}, this shows that
\[
h({\bf x}+{\bf q})=h({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=\fullshad{f}({\bf x})
\]
for all ${\bf x}\in{\R^n}$.
This means that $\overline{\ee}\in\unimin{f}$, as claimed;
that $h=\fullshad{f}$;
and finally that
${\bf q}\in\conssp{h}=\conssp{\fullshad{f}}$, completing the
proof.
\qed
The set $\unimin{f}$ of all universal reducers is a closed, convex
cone (though possibly disjoint from ${\R^n}$, as will be seen in
Theorem~\ref{thm:new:f:8.1}(\ref{thm:new:f:8.1:b})).
\begin{theorem} \label{thm:new:thm:f:8cd}
Let $f:{\R^n}\rightarrow\extspac{n}$ be convex.
Then $\unimin{f}$ is a nonempty, closed, convex cone.
\end{theorem}
\proof
Being a universal reducer is actually a property of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, not $f$,
which means, by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa}), that
$\unimin{f}=\unimin(\lsc f)$.
Therefore, we can assume without loss of generality
that $f$ is lower semicontinuous, since if it is not, we can replace
$f$ with $\lsc f$.
\underline{Nonempty}:
As already discussed, Corollary~\ref{cor:fig-cons-main-props},
especially part~(\ref{cor:fig-cons-main-props:a}), shows that
$\unimin{f}$ is nonempty.
\underline{Closed}:
Let $\seq{\overline{{\bf y}}_t}$ be any sequence in $\unimin{f}$ which converges to
some point $\overline{{\bf y}}\in\extspac{n}$.
Then for all ${\bf x}\in{\R^n}$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\liminf {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}_t\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\fullshad{f}({\bf x}).
\]
The first inequality is by
Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2a}), and
because
$\overline{{\bf y}}_t\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}\rightarrow\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$
(by Proposition~\ref{pr:aff-cont}, applied to the affine map
$\overline{{\bf z}}\mapsto {\bf x}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}=\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$).
The second inequality is because each $\overline{{\bf y}}_t\in\unimin{f}$.
Therefore, $\overline{{\bf y}}\in\unimin{f}$.
\underline{Convex}:
For ${\bf z}\in{\R^n}$ and $\beta\in{\mathbb{R}}$, let us define the set
\[
L({\bf z},\beta) = \braces{ \overline{{\bf y}}\in\extspac{n} :
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}) \leq \beta }.
\]
We claim that this set is convex.
When ${\bf z}={\bf 0}$, $L({\bf 0},\beta)$ is just a usual sublevel set of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, which is convex
by Theorem~\ref{thm:f:9} and since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is convex
(Theorem~\ref{thm:fext-convex}).
Further, we can obtain $L({\bf z},\beta)$ from $L({\bf 0},\beta)$ simply by adding
$-{\bf z}$ to each point; that is,
\[ L({\bf z},\beta) = (-{\bf z}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax L({\bf 0},\beta). \]
This is because $\overline{{\bf y}}\in L({\bf z},\beta)$ if and only if
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\leq \beta$ and $\overline{\xx}=\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}$,
which in turn is true if and only if
$\overline{\xx}\in L({\bf 0},\beta)$ and $\overline{{\bf y}}=(-{\bf z})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$.
Thus, $L({\bf z},\beta)$ is obtained from $L({\bf 0},\beta)$ by applying the
affine map $\overline{\xx}\mapsto (-{\bf z})\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$
to the entire set.
Therefore, $L({\bf z},\beta)$ is convex by Corollary~\ref{cor:thm:e:9}.
The set of universal reducers consists of all points for which
\eqref{eq:unimin-def} holds, for all ${\bf x}\in{\R^n}$.
Therefore,
\[
\unimin{f} = \bigcap_{{\bf x}\in{\R^n}} L({\bf x},\fullshad{f}({\bf x})).
\]
Thus, $\unimin{f}$ is an intersection of convex sets, and so is
itself convex (by Proposition~\ref{pr:e1}(\ref{pr:e1:b})).
\underline{Cone}:
Let $\overline{{\bf y}}\in\unimin{f}$, and let $\lambda>0$.
Then $\overline{{\bf y}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some $\overline{\ee}\in{{\cal E}_n}$ and
${\bf q}\in{\R^n}$.
By Theorem~\ref{thm:new:thm:f:8a},
$\overline{\ee}\in\unimin{f}$ and ${\bf q}\in\conssp{\fullshad{f}}$.
Since $\conssp{\fullshad{f}}$ is a linear subspace, this implies that
$\lambda{\bf q}$ is in $\conssp{\fullshad{f}}$ as well.
Applying Theorem~\ref{thm:new:thm:f:8a} again now yields that
$\lambda\overline{{\bf y}} = \overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf q} \in \unimin{f}$.
Therefore, $\unimin{f}$ is a cone.
\qed
Universal reducers can be constructed in a recursive fashion very
similar to that seen for points in the astral recession cone in
Theorem~\ref{thm:f:3}.
Indeed, the recursive formulation is identical to what was seen in
that theorem.
The critical difference between the construction of points in
$\unimin{f}$ and ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ will come instead
in where these sets intersect ${\R^n}$.
As seen in Proposition~\ref{pr:f:1}, the astral recession cone
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ intersects ${\R^n}$ exactly at the standard recession cone
$\resc{f}$.
In contrast, $\unimin{f}$'s intersection with ${\R^n}$ depends on whether
the termination criterion used in the construction in
Figure~\ref{fig:min-proc} has been reached.
If that condition is satisfied, that is, if $\resc{f}=\conssp{f}$, then
$(\unimin{f})\cap{\R^n} = \resc{f}$ ($=\conssp{f}$), as was the case for
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
But if $\resc{f}\neq\conssp{f}$, then
$\unimin{f}$ is entirely disjoint from ${\R^n}$.
We prove these facts in the next two theorems.
\begin{theorem} \label{thm:new:f:8.1}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
\begin{enumerate}
\item \label{thm:new:f:8.1:a}
If $\resc{f}=\conssp{f}$, then $(\unimin{f})\cap{\R^n}=\resc{f}$
(implying, in particular, that ${\bf 0}\in\unimin{f}$ in this case).
\item \label{thm:new:f:8.1:b}
Otherwise, if $\resc{f}\neq\conssp{f}$,
then $(\unimin{f})\cap{\R^n}=\emptyset$.
\end{enumerate}
\end{theorem}
\proof
As a first observation, by
Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:b}),
${\bf 0}\in\unimin{f}$ if and only if
$f({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf 0}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})=\fullshad{f}({\bf x})$,
for all ${\bf x}\in{\R^n}$
(since $f$ is lower semicontinuous, and
using Proposition~\ref{pr:h:1}(\ref{pr:h:1a})).
Combined with Corollary~\ref{cor:thm:f:4:1}, this means
\begin{equation} \label{eqn:thm:new:f:8.1:b}
\resc{f}=\conssp{f}
\;\Leftrightarrow\;
f=\fullshad{f}
\;\Leftrightarrow\;
{\bf 0}\in\unimin{f}.
\end{equation}
Next, suppose ${\bf q}\in(\unimin{f})\cap{\R^n}$.
Then ${\bf 0}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\in\unimin{f}$, and since ${\bf 0}\in{{\cal E}_n}$ and
${\bf q}\in{\R^n}$, this implies that
${\bf 0}\in\unimin{f}$ and ${\bf q}\in\conssp{\fullshad{f}}$ by
Theorem~\ref{thm:new:thm:f:8a}.
This shows, in particular, that
\begin{equation} \label{eqn:thm:new:f:8.1:a}
(\unimin{f})\cap{\R^n} \subseteq \conssp{\fullshad{f}}.
\end{equation}
In addition, this shows that if $(\unimin{f})\cap{\R^n}\neq\emptyset$, then
${\bf 0}\in\unimin{f}$, and therefore $\resc{f}=\conssp{f}$, by
\eqref{eqn:thm:new:f:8.1:b}.
This proves the contrapositive of part~(\ref{thm:new:f:8.1:b}).
To prove part~(\ref{thm:new:f:8.1:a}),
suppose $\resc{f}=\conssp{f}$.
Then $f=\fullshad{f}$ and ${\bf 0}\in\unimin{f}$ by
\eqref{eqn:thm:new:f:8.1:b}.
Since ${\bf 0}\in(\unimin{f})\cap{{\cal E}_n}$,
Theorem~\ref{thm:new:thm:f:8a} then implies that
if ${\bf q}\in\conssp{\fullshad{f}}$ then
${\bf q}={\bf 0}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\in\unimin{f}$.
Thus,
$\conssp{\fullshad{f}}\subseteq (\unimin{f})\cap{\R^n}$.
When combined with \eqref{eqn:thm:new:f:8.1:a}, and since
$f=\fullshad{f}$, this now yields
\[
(\unimin{f})\cap{\R^n}
=
\conssp{\fullshad{f}}
=
\conssp{f}
=
\resc{f}.
\]
\qed
\begin{theorem} \label{thm:new:f:8.2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Let ${\bf v}\in{\R^n}$
and let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
Suppose $\overline{{\bf y}}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$ for some $\overline{{\bf z}}\in\extspac{n}$.
Then $\overline{{\bf y}}\in\unimin{f}$ if and only if
${\bf v}\in\resc{f}$ and
$\overline{{\bf z}}\in\unimin{g}$.
\end{theorem}
\proof
Suppose $\overline{{\bf y}}\in\unimin{f}$.
Then $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ by
Propostion~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:a}),
so ${\bf v}\in\resc{f}$ by Theorem~\ref{thm:f:3}(\ref{thm:f:3b}).
For all ${\bf x}\in{\R^n}$,
\[
\overline{g}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\fullshad{f}({\bf x})
=
\fullshad{g}({\bf x}).
\]
The first equality is by Theorem~\ref{thm:d4}(\ref{thm:d4:c}).
The inequality is because $\overline{{\bf y}}\in\unimin{f}$.
The last equality is by Theorem~\ref{thm:f:5}.
Thus, $\overline{{\bf z}}\in\unimin{g}$.
The converse is similar.
Suppose that ${\bf v}\in\resc{f}$ and $\overline{{\bf z}}\in\unimin{g}$.
Then for all ${\bf x}\in{\R^n}$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
\overline{g}(\overline{{\bf z}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\fullshad{g}({\bf x})
=
\fullshad{f}({\bf x}).
\]
The second equality is by Theorem~\ref{thm:d4}(\ref{thm:d4:c}).
The inequality is because $\overline{{\bf z}}\in\unimin{g}$.
The last equality is by Theorem~\ref{thm:f:5}.
Thus, $\overline{{\bf y}}\in\unimin{f}$.
\qed
As mentioned above, the potential products of the process of
Figure~\ref{fig:min-proc} are exactly the universal minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
as we show now.
In this sense, this process finds all universal minimizers, and so also
all universal reducers.
\begin{theorem} \label{thm:min-proc-all-can-min}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Let ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in{\R^n}$,
and let $\overline{\ee} = \limrays{{\bf v}_1,\ldots,{\bf v}_k}$.
Let $g_0=f$ and $g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$ for $i=1,\ldots,k$.
Then the following are equivalent:
\begin{enumerate}
\item \label{thm:min-proc-all-can-min:a}
$\overline{\ee}\in\unimin{f}$ and ${\bf q}$ minimizes $\fullshad{f}$;
that is, $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ is a universal minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\item \label{thm:min-proc-all-can-min:b}
$\resc{g_k}=\conssp{g_k}$;
${\bf q}$ minimizes $g_k$;
and
${\bf v}_i\in\resc{g_{i-1}}$ for $i=1,\ldots,k$.
\end{enumerate}
\end{theorem}
\proof
~
\underline{(\ref{thm:min-proc-all-can-min:a})
$\Rightarrow$
(\ref{thm:min-proc-all-can-min:b})}:
Suppose $\overline{\ee}\in\unimin{f}$ and
${\bf q}$ minimizes $\fullshad{f}$.
Then $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, by
Proposition~\ref{pr:unimin-to-global-min}.
Therefore,
by Theorem~\ref{thm:all-min-proc-correct},
${\bf q}$ minimizes $g_k$
and
${\bf v}_i\in\resc{g_{i-1}}$ for $i=1,\ldots,k$.
Also, $g_k=\fshad{\ebar}$,
as in the proof of Theorem~\ref{thm:all-min-proc-correct}.
Since $\overline{\ee}\in\unimin{f}$,
Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:b})
implies that $\fshad{\ebar}=\fullshad{f}$.
Thus, $g_k=\fullshad{f}$, which, combined with the foregoing, implies
$\resc{g_k}=\conssp{g_k}$ by Corollary~\ref{cor:f:4}.
\underline{(\ref{thm:min-proc-all-can-min:b})
$\Rightarrow$
(\ref{thm:min-proc-all-can-min:a})}:
Suppose
$\resc{g_k}=\conssp{g_k}$,
${\bf q}$ minimizes $g_k$,
and
${\bf v}_i\in\resc{g_{i-1}}$ for $i=1,\ldots,k$.
These conditions imply $g_k=\fshad{\ebar}=\fullshad{f}$,
by Corollary~\ref{cor:fig-cons-main-props}(\ref{cor:fig-cons-main-props:a}).
Therefore, $\overline{\ee}\in\unimin{f}$ by
Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:b}).
This also shows ${\bf q}$ minimizes $\fullshad{f}$ since it minimizes
$g_k$.
\qed
We saw in
Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:a})
that
$\unimin{f}\subseteq{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
In fact, there is a much more precise relationship that exists between
these two sets.
Specifically, the astral recession cone, ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, is exactly
the convex hull of the set $\unimin{f}$ of universal reducers,
adjoined with the origin, as stated in the next theorem.
The same also holds true if only the icons in $\unimin{f}$
are included.
Thus, the astral recession cone is effectively being
generated by the universal reducers (together with the origin).
The universal reducers are, in this sense, the
most extreme points of the astral recession cone.
\begin{theorem} \label{thm:res-convhull-unimin}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Then
\[
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} = \conv((\unimin{f}) \cup \{{\bf 0}\})
= \conv(((\unimin{f}) \cap {{\cal E}_n}) \cup \{{\bf 0}\}).
\]
\end{theorem}
This relationship could instead be stated as saying that
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is exactly the cone generated by $\unimin{f}$
(or $(\unimin{f})\cap{{\cal E}_n}$); that is,
\[
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}
=
\cone(\unimin{f})
=
\cone((\unimin{f}) \cap {{\cal E}_n}).
\]
The equivalence of this statement with
Theorem~\ref{thm:res-convhull-unimin}
follows immediately from
Proposition~\ref{pr:cone-props}(\ref{pr:cone-props:d})
since both $\unimin{f}$ and $(\unimin{f}) \cap {{\cal E}_n}$ are cones
(by Theorem~\ref{thm:new:thm:f:8cd} and
Proposition~\ref{pr:new:a:2}(\ref{pr:new:a:2:c})).
\proof
As a preliminary step, we claim that
$(\unimin{f})\cap{{\cal E}_n}$ cannot be empty.
Indeed,
Theorem~\ref{thm:new:thm:f:8cd} shows $\unimin{f}$ is a nonempty,
closed, convex cone.
Therefore, $\unimin{f}=\conv((\unimin{f})\cap{{\cal E}_n})$,
by Theorem~\ref{thm:cl-conv-cone-convhull-corezeros},
so if $(\unimin{f})\cap{{\cal E}_n}$ were empty, then $\unimin{f}$ would be
as well, a contradiction.
Let
$ U = \conv(((\unimin{f}) \cap {{\cal E}_n}) \cup \{{\bf 0}\}) $.
We argue first that
\begin{equation} \label{eqn:thm:res-convhull-unimin:a}
({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n} \subseteq U.
\end{equation}
Let $\overline{{\bf y}}\in({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}$; we aim to show that
$\overline{{\bf y}}\in U$.
Let $\overline{\ee}$ be any point in $(\unimin{f})\cap{{\cal E}_n}$, which
we just argued must exist.
Then $\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}$ is an icon
(by Proposition~\ref{pr:i:8}(\ref{pr:i:8-leftsum})),
and is also a universal reducer
since, for all ${\bf x}\in{\R^n}$,
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
\leq
\fullshad{f}({\bf x}).
\]
The first inequality is because $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
The second is because $\overline{\ee}\in\unimin{f}$.
Thus, $\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}$ is an iconic universal reducer, so
\[ \overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}\in(\unimin{f})\cap{{\cal E}_n}\subseteq U. \]
The origin is also in $U$, by its definition.
Therefore, since $U$ is convex, it must include
$\lb{{\bf 0}}{\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}}$, the segment joining the origin and
$\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}$.
Combined with
Corollary~\ref{cor:d-in-lb-0-dplusx}
(since $\overline{{\bf y}}$ is iconic),
this yields
\[
\overline{{\bf y}}\in\lb{{\bf 0}}{\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\ee}}\subseteq U,
\]
which proves \eqref{eqn:thm:res-convhull-unimin:a}.
We can now prove the theorem as follows:
\begin{eqnarray*}
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}
&=&
\conv(({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \cap {{\cal E}_n})
\label{eqn:thm:res-convhull-unimin:aa}
\\
&\subseteq&
U = \conv(((\unimin{f}) \cap {{\cal E}_n}) \cup \{{\bf 0}\})
\label{eqn:thm:res-convhull-unimin:b}
\\
&\subseteq&
\conv((\unimin{f}) \cup \{{\bf 0}\})
\label{eqn:thm:res-convhull-unimin:c}
\\
&\subseteq&
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}.
\label{eqn:thm:res-convhull-unimin:d}
\end{eqnarray*}
The first equality
is by Theorem~\ref{thm:cl-conv-cone-convhull-corezeros},
since ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a closed, convex cone
(by Corollary~\ref{cor:thm:j:3}).
The first inclusion
is by \eqref{eqn:thm:res-convhull-unimin:a}, and
by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})
since $U$ is convex.
The second inclusion
is by Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:b}).
The last inclusion
follows from
Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})
since ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a convex set that includes ${\bf 0}$ and also
includes $\unimin{f}$
(Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:a})).
\qed
\section{The structure of minimizers in some particular cases}
\label{sec:minimizers-examples}
We study next the nature of minimizers of an extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ in
some particular cases, focusing on the astral rank of minimizers, and
also a natural class of minimization problems commonly encountered in
statistics and machine learning.
\subsection{Minimizers can have maximum astral rank}
\label{sec:max-rank-minimizers}
When a convex function has no finite minimizer, but can only be
minimized by an unbounded sequence of points, it seems natural to
wonder if the function can nevertheless always be minimized by
following some straight and unwavering
ray to infinity in some direction.
For instance, the function $f$ in \eqref{eq:simple-eg-exp-exp-sq}
is minimized by a ray in the direction of
${\bf v}=\trans{[v_1,v_2,v_3]}$
so long as $v_1,v_2>0$ and $v_3=1$;
in other words, $f$ is minimized by a sequence $t{\bf v}$,
for $t=1,2,\ldots$, which means
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is minimized by $\limray{{\bf v}}$.
Can every convex function be minimized in this way?
Or, to minimize the function, might it be necessary to pursue a more
convoluted route to infinity?
In astral terms, we are asking
if the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
of a convex function can always be minimized by a point in $\extspac{n}$
whose astral rank is at most one.
In this section, we show that this is not always possible.
Quite on the contrary, we study an example of a convex function
$f:{\R^n}\rightarrow{\mathbb{R}}$ whose extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
can only be minimized by a point with astral rank $n$, the maximum
possible.
Thus, the function not only cannot be minimized by following a simple,
one-dimensional ray, but in fact, the only way to minimize the
function is by pursuing a trajectory involving all $n$ dimensions.
We also use this example to illustrate some of the earlier concepts
developed in the preceding sections.
Here is the function, where, as usual, we write $x_i$ for the $i$-th
component of a vector ${\bf x}\in{\R^n}$:
\begin{eqnarray}
f({\bf x})
&=&
\exp({x_2^2 - x_1}) + \exp({x_3^2 - x_2})
+ \cdots + \exp({x_n^2 - x_{n-1}}) + \exp({-x_n})
\nonumber
\\
&=&
\sum_{j=1}^n h_j({\bf x})
\label{eq:h:9}
\end{eqnarray}
where
\[
h_j({\bf x}) = \left\{ \begin{array}{ll}
\exp({x_{j+1}^2 - x_j}) & \mbox{if $j<n$} \\
\exp({- x_n}) & \mbox{if $j=n$.}
\end{array} \right.
\]
Each function $h_j$ is convex
(since $x_{j+1}^2 - x_j$ and $-x_n$ are both convex, implying $h_j$
is, by Proposition~\ref{pr:j:2}(\ref{pr:j:2a}));
therefore, $f$ is convex as well.
Clearly, $f$ and the $h_j$'s are all also continuous, closed, proper,
finite everywhere, and strictly positive everywhere.
Intuitively, to minimize $h_j$, we need $x_j$ to be growing to
$+\infty$ faster than $x_{j+1}^2$.
Thus, to minimize $f$, we need every variable $x_j$ to tend to
$+\infty$, but with $x_1$ growing much faster than $x_2$,
which is growing much faster than $x_3$,
and so on.
To see how this intuition is captured by our formal results, we first
analyze $\unimin{f}$, the set of universal reducers for this
function.
This set turns out to consist only of
the single point
\[
{\bf I} {\bar{\bm{\omega}}} = \limrays{{\bf e}_1,\ldots,{\bf e}_n}
= \limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\cdots\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_n}
\]
where ${\bf e}_i$ is the $i$-th standard basis vector, and
${\bf I}$ is the $n\times n$ identity matrix.
(We could also write this point simply as ${\bar{\bm{\omega}}}=\ommsub{n}$,
as in \eqref{eq:i:7}, but instead write it as above to make its form
more explicit.)
Consistent with the intuition suggested above,
this point also turns out to be the unique minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\begin{proposition} \label{pr:hi-rank-f-eg}
Let $f$ be as defined in \eqref{eq:h:9}.
Then $f$'s only universal reducer is ${\bf I}{\bar{\bm{\omega}}}$;
that is, $ \unimin{f} = \{ {\bf I}{\bar{\bm{\omega}}} \} $.
\end{proposition}
\proof
Let $\overline{\xx}$ be any point in $\unimin{f}$.
We will prove the proposition by showing this implies
$\overline{\xx}={\bf I}{\bar{\bm{\omega}}}$.
(That ${\bf I}{\bar{\bm{\omega}}}$ actually is in $\unimin{f}$ then follows from the
fact that $\unimin{f}$ cannot be empty, by
Theorem~\ref{thm:new:thm:f:8cd}.)
Let
$\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
be $\overline{\xx}$'s canonical representation
(see Theorem~\ref{pr:uniq-canon-rep}).
We define functions $g_0=f$ and
$g_i = \genshad{g_{i-1}}{\limray{\vv_i}}$ for $i=1,\ldots,k$,
and also
\[
{\hat{g}}_i({\bf x})
=
\sum_{j=i+1}^n h_j({\bf x})
\]
for $i=0,\ldots,n$ and ${\bf x}\in{\R^n}$.
Finally, let
\[
\overline{{\bf y}}_i
=
\limrays{{\bf v}_{i+1},\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}
\]
for $i=0,\ldots,k$.
Note that $\overline{{\bf y}}_0=\overline{\xx}$
and ${\hat{g}}_0=f=g_0$.
To show that $\overline{\xx}={\bf I}{\bar{\bm{\omega}}}$,
we prove the following by induction on $i=0,\ldots,n$:
\begin{enumerate}
\item \label{pr:hi-rank-f-eg:induc:a}
$k\geq i$.
\item \label{pr:hi-rank-f-eg:induc:b}
For all ${\bf x}\in{\R^n}$,
$g_i({\bf x}) = {\hat{g}}_i({\bf x})
= {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limrays{{\bf e}_1,\ldots,{\bf e}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})$.
\item \label{pr:hi-rank-f-eg:induc:c}
${\bf v}_j={\bf e}_j$ for all $j\leq i$.
\item \label{pr:hi-rank-f-eg:induc:d}
$\overline{{\bf y}}_i\in\unimin{g_i}$.
\end{enumerate}
Specifically, when $i=n$, items~(\ref{pr:hi-rank-f-eg:induc:a})
and~(\ref{pr:hi-rank-f-eg:induc:c}) will imply that
$\overline{\xx}={\bf I}{\bar{\bm{\omega}}}$.
The base case that $i=0$ follows immediately from definitions and
assumptions (also Proposition~\ref{pr:h:1}(\ref{pr:h:1a}) and
continuity of $f$).
For the inductive step, let $i\geq 1$ and
assume that the inductive hypothesis holds for $i-1$.
Then $k\geq i-1$,
$g_{i-1} = {\hat{g}}_{i-1}$,
$\overline{{\bf y}}_{i-1}\in\unimin{g_{i-1}}$,
and ${\bf v}_j={\bf e}_j$ for $j=1,\ldots,i-1$.
Note first that, for $\lambda\in{\mathbb{R}}$,
\[
g_{i-1}(\lambda{\bf e}_i)
=
{\hat{g}}_{i-1}(\lambda{\bf e}_i)
=
e^{-\lambda} + (n-i),
\]
which is strictly decreasing as a function of
$\lambda$, and so is bounded by
$g_{i-1}({\bf 0})<+\infty$ for $\lambda\geq 0$.
This shows that ${\bf e}_i\in(\aresconegsub{{i-1}})\cap{\R^n}=\resc{g_{i-1}}$,
by Theorem~\ref{thm:f:2} and Propostion~\ref{pr:f:1}
(or using results from standard convex analysis),
and also shows that $-{\bf e}_i\not\in\resc{g_{i-1}}$.
Specifically, this implies that
$\resc{g_{i-1}}\neq\conssp{g_{i-1}}$,
by \eqref{eq:cons-isresccap-resc}.
Therefore, $(\unimin{g_{i-1}})\cap{\R^n}=\emptyset$ by
Theorem~\ref{thm:new:f:8.1}(\ref{thm:new:f:8.1:b}).
Since $\overline{{\bf y}}_{i-1}\in\unimin{g_{i-1}}$,
this proves that $\overline{{\bf y}}_{i-1}\not\in{\R^n}$, so $k\geq i$,
proving item~(\ref{pr:hi-rank-f-eg:induc:a}).
Thus,
$\overline{{\bf y}}_{i-1}=\limray{{\bf v}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}_i$,
which further implies that
${\bf v}_i\in\resc{g_{i-1}}$
and
$\overline{{\bf y}}_i\in\unimin{g_i}$
by Theorem~\ref{thm:new:f:8.2},
proving item~(\ref{pr:hi-rank-f-eg:induc:d}).
The components of ${\bf v}_i$ are
${\bf v}_i=\trans{[v_{i1},\ldots,v_{in}]}$.
For all $j<i$, ${\bf v}_j={\bf e}_j$ is orthogonal to ${\bf v}_i$, since
$\overline{\xx}$ has been written in its canonical representation.
Thus, $v_{ij}=0$ for $j<i$.
We further claim that $v_{ij}=0$ for $j>i$ as well.
Suppose to the contrary that $v_{ij}\neq 0$ for some $j>i$.
Then for $\lambda\geq 0$,
\[
g_{i-1}({\bf 0})
\geq
g_{i-1}(\lambda{\bf v}_i)
=
{\hat{g}}_{i-1}(\lambda{\bf v}_i)
\geq
h_{j-1}(\lambda{\bf v}_i)
=
\exp(\lambda^2 v_{ij}^2 - \lambda v_{i,j-1}),
\]
where
the first inequality is because ${\bf v}_i$ is in $\resc{g_{i-1}}$
(which is a cone by Proposition~\ref{pr:resc-cone-basic-props}).
Note that the expression on the right tends to $+\infty$
as $\lambda\rightarrow+\infty$
since $v_{ij}\neq 0$.
But
this is a contradiction since $g_{i-1}({\bf 0})<+\infty$.
Thus, ${\bf v}_i$, which has unit length, must be either ${\bf e}_i$ or
$-{\bf e}_i$.
Since, as argued above, $-{\bf e}_i\not\in\resc{g_{i-1}}$, we conclude
that ${\bf v}_i={\bf e}_i$,
proving item~(\ref{pr:hi-rank-f-eg:induc:c}).
Finally, for all ${\bf x}\in{\R^n}$, we claim
\begin{equation} \label{eqn:pr:hi-rank-f-eg:1}
g_i({\bf x}) = \overline{g}_{i-1}(\limray{{\bf e}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
= {\hat{g}}_i({\bf x}).
\end{equation}
The first equality is by $g_i$'s definition.
To see the second equality, let $\overline{{\bf z}}=\limray{{\bf e}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}$.
By
Proposition~\ref{pr:d1},
there exists a sequence $\seq{{\bf z}_t}$ in ${\R^n}$ that converges to
$\overline{{\bf z}}$ and such that
$g_{i-1}({\bf z}_t)\rightarrow \overline{g}_{i-1}(\overline{{\bf z}})$.
Also,
$z_{tj}={\bf z}_t\cdot{\bf e}_j\rightarrow\overline{{\bf z}}\cdot{\bf e}_j$,
for $j=1,\ldots,n$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})),
implying $z_{ti}\rightarrow+\infty$ and
$z_{tj}\rightarrow x_j$ for $j\neq i$.
Therefore, by continuity,
$h_i({\bf z}_t)\rightarrow 0$ and
$h_j({\bf z}_t)\rightarrow h_j({\bf x})$ for $j>i$,
so that
${\hat{g}}_{i-1}({\bf z}_t)\rightarrow {\hat{g}}_i({\bf x})$.
Since $g_{i-1}({\bf z}_t)={\hat{g}}_{i-1}({\bf z}_t)$ for all $t$,
the limits of these two sequences must be equal;
thus, $\overline{g}_{i-1}(\overline{{\bf z}})={\hat{g}}_i({\bf x})$,
proving \eqref{eqn:pr:hi-rank-f-eg:1}.
In addition, using our inductive hypothesis
and Proposition~\ref{pr:i:9}(\ref{pr:i:9b}),
\[
\overline{g}_{i-1}(\limray{{\bf e}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limrays{{\bf e}_1,\ldots,{\bf e}_{i-1}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limrays{{\bf e}_1,\ldots,{\bf e}_i}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x}).
\]
Together with \eqref{eqn:pr:hi-rank-f-eg:1}, these
prove item~(\ref{pr:hi-rank-f-eg:induc:b}),
completing the induction.
This induction argument proves that
$k=n$ and ${\bf v}_i={\bf e}_i$ for $i=1,\ldots,n$.
This further shows ${\bf q}={\bf 0}$, being orthogonal
to ${\bf e}_1,\ldots,{\bf e}_n$, completing the proof.
\qed
Having determined $\unimin{f}$, we can now
use Theorem~\ref{thm:res-convhull-unimin} to
see that the astral recession cone is simply the segment
joining ${\bf 0}$ and ${\bf I}{\bar{\bm{\omega}}}$:
\[
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} = \conv{\{{\bf 0}, {\bf I}{\bar{\bm{\omega}}}\}}
= \lb{{\bf 0}}{{\bf I}{\bar{\bm{\omega}}}}.
\]
In particular, from Theorem~\ref{thm:lb-with-zero},
the icons in this set are exactly
\[
({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \cap {{\cal E}_n} =
\{ \limrays{{\bf e}_1,\ldots,{\bf e}_i} : i=0,\ldots,n \}.
\]
Further,
the proof of Proposition~\ref{pr:hi-rank-f-eg} shows, in
the notation of the proof, that
\begin{equation} \label{eq:fbar-ei-hj}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limrays{{\bf e}_1,\ldots,{\bf e}_i} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x})
=
g_i({\bf x})
=
{\hat{g}}_i({\bf x})
=
\sum_{j=i+1}^n h_j({\bf x}).
\end{equation}
Using these facts, we can now prove that ${\bf I}{\bar{\bm{\omega}}}$ is ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s unique
minimizer:
Let $\overline{\xx}\in\extspac{n}$, which we write as $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
for some $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
Suppose ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$.
Then $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, by Corollary~\ref{cor:a:4},
so, as just argued,
$\overline{\ee} = \limrays{{\bf e}_1,\ldots,{\bf e}_i}$
for some $i\in\{0,\ldots,n\}$.
If $i<n$, then \eqref{eq:fbar-ei-hj} shows that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}) = g_i({\bf q}) > 0$.
And if $i=n$, then $\overline{\xx}={\bf I}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}={\bf I}{\bar{\bm{\omega}}}$, and
$g_n\equiv 0$ so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=0$.
We have thus proved:
\begin{proposition} \label{pr:min-hi-rank-f-eg}
Let $f$ be as defined in \eqref{eq:h:9}.
Then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is uniquely minimized at ${\bf I}{\bar{\bm{\omega}}}$.
\end{proposition}
Thus, the function $f$ in \eqref{eq:h:9}
cannot be minimized by a sequence following a
straight line, nor even converging asymptotically to a line.
On the contrary, the function can only be minimized by a sequence that
expands unboundedly across all $n$ dimensions.
For example, $f$ is minimized by the sequence
\[
{\bf x}_t = \sum_{i=1}^n t^{3^{n-i}} {\bf e}_i,
\]
which converges to ${\bf I}{\bar{\bm{\omega}}}$.
However, $f$ need not be minimized by \emph{every} sequence
converging to this point (unless $n=1$);
for instance, the sequence
\[
{\bf x}'_t = \sum_{i=1}^n t^{n-i+1} {\bf e}_i
\]
converges to ${\bf I}{\bar{\bm{\omega}}}$, but $f({\bf x}'_t)\not\rightarrow 0$.
In other words, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at ${\bf I}{\bar{\bm{\omega}}}$.
Still, Proposition~\ref{pr:min-hi-rank-f-eg} does imply that convergence
to ${\bf I}{\bar{\bm{\omega}}}$ is a \emph{necessary} condition for a sequence to minimize
$f$, meaning $f$ cannot be minimized by any sequence that does
\emph{not} converge in $\extspac{n}$ to ${\bf I}{\bar{\bm{\omega}}}$.
\subsection{Sufficient conditions for rank one minimizers}
\label{subsec:rank-one-minimizers}
Propositions~\ref{pr:hi-rank-f-eg} and~\ref{pr:min-hi-rank-f-eg} show
that a convex function need not have a universal reducer (nor a minimizer)
of astral rank at most one.
For the function that was studied to prove this,
we also noted that its extension was not continuous.
In fact, there turns out to be a general connnection between
continuity and the existence of rank-one minimizers.
Among its consequences, we will see that this connection
implies that if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere
then there must exist a universal reducer (and therefore also a minimizer)
of astral rank at most one.
Continuity will be studied more closely in
Section~\ref{sec:continuity}.
Here, we make the link to rank-one universal reducers.
We first consider the relationship between the standard recession cone
$\resc{f}$ and the astral recession cone ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, where $f$ is
convex and lower semicontinuous.
We have seen, in Proposition~\ref{pr:f:1}, that ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ always
includes $\resc{f}$.
From this, we can conclude actually that ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ must always further
include $\rescbar{f}$, the closure of $\resc{f}$ in $\extspac{n}$, as
well as $\represc{f}$, the representational closure of $\resc{f}$:
\begin{proposition} \label{pr:repres-in-arescone}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then $\represc{f}$ is a convex cone, and furthermore,
\[
\resc{f}
\subseteq
\represc{f}
\subseteq
\rescbar{f}
\subseteq
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}.
\]
\end{proposition}
\proof
Proposition~\ref{pr:f:1} implies that
$\resc{f} \subseteq {\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Since ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is closed in $\extspac{n}$
(by Corollary~\ref{cor:thm:j:3}),
this further implies that this set must also include $\resc{f}$'s closure,
$\rescbar{f}$.
By Proposition~\ref{pr:resc-cone-basic-props},
$\resc{f}$ is a convex cone in ${\R^n}$ that includes ${\bf 0}$.
Therefore, Corollary~\ref{cor:a:1} immediately implies
that $\represc{f}$ is a convex cone, and that
$\resc{f} \subseteq \represc{f} \subseteq \rescbar{f}$,
completing the proof.
\qed
Later,
in Section~\ref{subsec:cond-for-cont},
we will see that $\represc{f}$ and its relationship to ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
play critically important roles in
characterizing exactly where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous.
When these sets are equal to each other,
so that ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$
(implying
$\represc{f} = \rescbar{f} = {\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
by Proposition~\ref{pr:repres-in-arescone}),
we say that $f$ is
\emph{recessive complete}.
When this property holds, $f$'s standard recession cone
$\resc{f}$ is, in a sense, sufficiently ``complete'' to entirely fill
or determine the astral recession cone ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
(via its representational closure, $\represc{f}$).
For example, consider
the function $f$ given in
\eqref{eqn:curve-discont-notfinite-eg}.
As was seen in Section~\ref{sec:arescone},
its standard recession cone
is $\resc{f} = \{ \lambda {\bf e}_2 : \lambda\geq 0\}$,
whose representational closure is
$\represc{f}=(\resc{f})\cup\{\limray{{\bf e}_2}\}$.
On the other hand, its astral recession
cone includes points that are outside and apparently quite unrelated
to this set, such as $\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf e}_1$
and $\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$.
Thus, $f$ is not recessive complete.
(An example of a natural class of functions that \emph{are}
all recessive complete is given
in Section~\ref{sec:emp-loss-min}.)
We will see that recessive completeness of a convex function $f$
is in general a necessary
condition for ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ to be continuous everywhere,
and furthermore that this condition is both necessary and sufficient
for continuity everywhere
if $f$ is also finite everywhere.
Details of this connection will be given in
Section~\ref{sec:continuity}.
For now, we focus just on the consequences of recessive completeness
for our study of minimizers.
In Theorem~\ref{thm:f:4x},
we saw how the domain of the universal reduction
$\fullshad{f}$ is effectively limited to the linear subspace
${(\aresconef)^{\bot}}$.
If $f$ is recessive complete,
the set ${(\aresconef)^{\bot}}$ is simply equal to
$\rescperp{f}$, the set of points perpendicular to all directions in
the standard recession cone, $\resc{f}$:
\begin{proposition} \label{pr:perpres-is-rescperp}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then ${(\aresconef)^{\bot}}\subseteq\rescperp{f}$.
If, in addition,
$f$ is recessive complete,
then
${(\aresconef)^{\bot}} = \rescperp{f}$.
\end{proposition}
\proof
That ${(\aresconef)^{\bot}}\subseteq\rescperp{f}$ follows immediately from
Proposition~\ref{pr:perp-props}(\ref{pr:perp-props:b})
since
$\resc{f} \subseteq {\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
(by Proposition~\ref{pr:f:1}).
Suppose now that ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$.
Let ${\bf u}\in\rescperp{f}$, and let
$\overline{\xx}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Since $\overline{\xx}$ is also in $\represc{f}$, we can write it in the form
$\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
where ${\bf v}_1,\ldots,{\bf v}_k,{\bf q}\in \resc{f}$.
Since ${\bf u}\in\rescperp{f}$, it is orthogonal to all points in
$\resc{f}$, so
${\bf q}\cdot{\bf u}=0$ and ${\bf v}_i\cdot{\bf u}=0$, for $i=1,\ldots,k$,
implying
$\overline{\xx}\cdot{\bf u}=0$.
Therefore, ${\bf u}\in{(\aresconef)^{\bot}}$, being perpendicular to every point in
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
so $\rescperp{f}\subseteq{(\aresconef)^{\bot}}$, completing the proof.
\qed
When $f$ is recessive complete
(and therefore, as remarked above, whenever ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous),
there must always exist a
universal reducer of astral rank at most one.
Specifically, as we show next,
if ${\bf v}$ is any point in the relative interior of
$f$'s standard recession cone, $\resc{f}$,
then the associated astron $\limray{{\bf v}}$ must be a universal reducer.
As a
consequence, $\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ must be a universal minimizer
of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
for every ${\bf q}\in{\R^n}$ that minimizes $\fullshad{f}$
(and thus, for all finite parts of all minimizers of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
by Proposition~\ref{pr:min-fullshad-is-finite-min}).
This is in dramatic contrast to the example of \eqref{eq:h:9} which had
no universal reducers and no minimizers of astral rank less than $n$.
\begin{theorem}\label{thm:unimin-can-be-rankone}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, lower semicontinuous,
and recessive complete.
Let ${\bf v}\in\ric{f}$, the relative interior of $f$'s
standard recession cone.
Then $\limray{{\bf v}}\in\unimin{f}$.
Consequently, if ${\bf q}\in{\R^n}$ minimizes
$\fullshad{f}$, then $\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ is a universal
minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
\end{theorem}
\proof
Let
$g=\fshad{\limray{\vv}}$
be the reduction of $f$ at $\limray{{\bf v}}$.
We will prove the result
using Corollary~\ref{cor:f:4}, specifically, by first showing that $g$
satisfies condition~(\ref{cor:f:4b}) of that corollary.
So suppose ${\bf u}\in\slopes{g}$;
we aim to show this implies that ${\bf u}\in{(\aresconef)^{\bot}}$.
Since ${\bf u}\in\slopes{g}$ and
${\bf v}\in\resc{f}$,
Corollary~\ref{cor:thm:e1:slopes} implies that
${\bf v}\cdot{\bf u}=0$ and ${\bf u}\in\slopes{f}$.
Let ${\bf w}$ be any point in $\resc{f}$.
This set is polar to
$\slopes{f}$ by Corollary~\ref{cor:rescpol-is-slopes};
therefore, ${\bf u}\cdot{\bf w}\leq 0$.
Also, because ${\bf v}\in\ric{f}$,
there must exist $\delta>0$ such that the point
${\bf v}+\delta({\bf v}-{\bf w})$
is in $\resc{f}$ as well;
this follows directly from \citet[Theorem~6.4]{ROC}.
Applying Corollary~\ref{cor:rescpol-is-slopes}
again to this point,
and since ${\bf u}\cdot{\bf v}=0$, yields
\[ -\delta {\bf u}\cdot{\bf w}
= {\bf u} \cdot [{\bf v}+\delta({\bf v}-{\bf w})]
\leq 0.
\]
Since $\delta>0$,
it follows that ${\bf u}\cdot{\bf w}=0$.
Thus, because this holds for all ${\bf w}\in\resc{f}$,
we conclude that ${\bf u}\in\rescperp{f}$.
Since ${(\aresconef)^{\bot}}=\rescperp{f}$,
by Proposition~\ref{pr:perpres-is-rescperp},
we have thus shown that
$\slopes{g}\subseteq{(\aresconef)^{\bot}}$.
In other words, we have shown that condition~(\ref{cor:f:4b}) of
Corollary~\ref{cor:f:4} is satisfied with $k=1$
and ${\bf v}_1={\bf v}$ (so that $g_0=f$ and $g_1=g$).
Therefore, by condition~(\ref{cor:f:4c}) of that corollary,
it now follows that
\[ {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})
= g({\bf x})
= \fullshad{f}({\bf x})
\]
for all ${\bf x}\in{\R^n}$.
In other words, $\limray{{\bf v}}$ is a universal reducer for $f$, as claimed.
(Therefore, if ${\bf q}\in{\R^n}$ minimizes $\fullshad{f}$, then
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ is a universal minimizer, by definition.)
\qed
\subsection{Empirical risk minimization}
\label{sec:emp-loss-min}
As another example, we next consider functions of the form
\begin{equation} \label{eqn:loss-sum-form}
f({\bf x}) = \sum_{i=1}^m \ell_i({\bf x}\cdot{\bf u}_i),
\end{equation}
for ${\bf x}\in{\R^n}$, where ${\bf u}_i\in{\R^n}$, and each function
$\ell_i:{\mathbb{R}}\rightarrow{\mathbb{R}}$ is convex, lower-bounded, and nondecreasing.
We focus on the minimizers of such functions, and especially how these
relate to concepts developed earlier.
Minimizing functions of the form given in \eqref{eqn:loss-sum-form}
is a fundamental problem
in machine learning and statistics.
Very briefly, in a typical setting, a learning algorithm might be
given random ``training examples'' $({\bf z}_i,y_i)$, for $i=1,\ldots,m$,
where ${\bf z}_i\in{\R^n}$ is an
``instance'' or ``pattern'' (such as an image or photograph, treated
as a vector of pixel intensities in ${\R^n}$),
and $y_i\in\{-1,+1\}$ is a ``label''
(that might indicate, for example, if the photograph is or is not of a
person's face).
The goal then is to find a rule for predicting if a new instance
${\bf z}\in{\R^n}$ should be labeled $-1$ or $+1$.
As an example, in logistic regression,
the learner finds a vector ${\bf w}\in{\R^n}$, based on the
training examples, and then predicts that a new instance ${\bf z}$ should
be labeled according to the sign of ${\bf w}\cdot{\bf z}$.
Specifically, ${\bf w}$ is chosen to minimize the ``logistic loss'' on the
training examples, that is,
\begin{equation} \label{eqn:logistic-reg-obj}
f({\bf w}) = \sum_{i=1}^m \ln\paren{1+\exp(-y_i {\bf w}\cdot{\bf z}_i)}.
\end{equation}
This kind of function, which is more generally called the
\emph{empirical risk}, has the same form as in
\eqref{eqn:loss-sum-form}
(with ${\bf x}={\bf w}$, ${\bf u}_i = -y_i {\bf z}_i$, and $\ell_i(z)=\ln(1+e^z)$).
Returning to the general case in \eqref{eqn:loss-sum-form},
for $i=1,\ldots,m$,
we will assume, without loss of generality, that $\inf \ell_i = 0$, and that $\ell_i$ is not constant
(in addition to the other assumptions mentioned above).
Since $\ell_i$ is nondecreasing,
these conditions imply that
$\lim_{x\rightarrow -\infty} \ell_i(x) = 0$ and
$\lim_{x\rightarrow +\infty} \ell_i(x) = +\infty$.
Each $\ell_i$ is convex and finite everywhere, and therefore
continuous everywhere
\citep[Corollary~10.1.1]{ROC};
the same is also true of $f$.
So $\ell_i$'s extension is
\begin{equation} \label{eqn:hard-core:1}
{\overline{\ell}}_i(\bar{x}) = \left\{ \begin{array}{cl}
0 & \mbox{if $\bar{x}=-\infty$} \\
\ell_i(\bar{x}) & \mbox{if $\bar{x}\in{\mathbb{R}}$} \\
+\infty & \mbox{if $\bar{x}=+\infty$}
\end{array} \right.
\end{equation}
for $\bar{x}\in\overline{\R}$.
The next proposition gives the form of $f$'s extension, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
as well as $f$'s astral and standard recession cones, and also shows
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere, implying
that $f$ is recessive complete.
\begin{proposition} \label{pr:hard-core:1}
Let $f:{\R^n}\rightarrow{\mathbb{R}}$ have the form given in
\eqref{eqn:loss-sum-form}, where, for $i=1,\ldots,m$,
${\bf u}_i\in{\R^n}$ and
$\ell_i:{\mathbb{R}}\rightarrow{\mathbb{R}}$ is convex, nondecreasing, not constant,
with $\inf \ell_i = 0$.
Then the following hold:
\begin{enumerate}
\item \label{pr:hard-core:1:a}
The lower semicontinuous extension of $f$ is
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) = \sum_{i=1}^m {\overline{\ell}}_i(\overline{\xx}\cdot{\bf u}_i),
\]
for $\overline{\xx}\in\extspac{n}$.
This function is continuous everywhere.
\item \label{pr:hard-core:1:b}
The astral and standard recession cones of $f$ are:
\begin{eqnarray*}
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} &=& \{ \overline{{\bf y}}\in\extspac{n} : \overline{{\bf y}}\cdot{\bf u}_i \leq 0
\mbox{ for } i=1,\ldots,m \} \\
\resc{f} &=& \{ {\bf y}\in{\R^n} : {\bf y}\cdot{\bf u}_i \leq 0 \mbox{ for } i=1,\ldots,m \}.
\end{eqnarray*}
Furthermore,
$f$ is recessive complete.
\end{enumerate}
\end{proposition}
\proof
~
Part~(\ref{pr:hard-core:1:a}):
Let $\overline{\xx}\in\extspac{n}$.
Let $\seq{{\bf x}_t}$ be any sequence in ${\R^n}$ that converges to $\overline{\xx}$.
Then, for $i=1,\ldots,m$,
${\bf x}_t\cdot{\bf u}_i \rightarrow \overline{\xx}\cdot{\bf u}_i$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
This implies
$\ell_i({\bf x}_t\cdot{\bf u}_i)\rightarrow {\overline{\ell}}_i(\overline{\xx}\cdot{\bf u}_i)$
by continuity of $\ell_i$, and so also of ${\overline{\ell}}_i$
(since $\ell_i$ is convex with effective domain all of ${\mathbb{R}}$,
and using
\citet[Corollary~10.1.1]{ROC}).
The result now follows by continuity of addition and since each
$\ell_i$ is lower-bounded.
Part~(\ref{pr:hard-core:1:b}):
Suppose $\overline{{\bf y}}\cdot{\bf u}_i\leq 0$ for $i=1,\ldots,m$.
Then because ${\overline{\ell}}_i$ is nondecreasing, for ${\bf x}\in{\R^n}$,
${\overline{\ell}}_i(\overline{{\bf y}}\cdot{\bf u}_i+{\bf x}\cdot{\bf u}_i) \leq {\overline{\ell}}_i({\bf x}\cdot{\bf u}_i)$
for all $i=1,\ldots,m$,
implying
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf x}) \leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x}) = f({\bf x})$.
Therefore, $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
by Proposition~\ref{pr:alt-arescone-def}.
For the converse, suppose for some $i\in\{1,\ldots,m\}$ that
$\overline{{\bf y}}\cdot{\bf u}_i > 0$.
If $\overline{{\bf y}}\cdot{\bf u}_i=+\infty$, then
${\overline{\ell}}_i(\overline{{\bf y}}\cdot{\bf u}_i)=+\infty$,
implying
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}})=+\infty>f({\bf 0})$,
so $\overline{{\bf y}}\not\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Otherwise, $\overline{{\bf y}}\cdot{\bf u}_i$ is a positive real number,
so as $\lambda\rightarrow+\infty$,
$\ell_i(\lambda\overline{{\bf y}}\cdot{\bf u}_i)\rightarrow+\infty$, implying
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf y}})\rightarrow+\infty$.
Thus, for some $\lambda>0$, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\lambda\overline{{\bf y}})>f({\bf 0})$
so $\lambda\overline{{\bf y}}\not\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, and therefore
$\overline{{\bf y}}\not\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ since ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a cone
(Corollary~\ref{cor:thm:j:3}).
The expression for $\resc{f}$ now follows immediately from
Proposition~\ref{pr:f:1}.
That $f$ is recessive complete follows from the fact that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
is continuous everywhere, by direct application of
Theorem~\ref{thm:cont-conds-finiteev} (to be proved later).
\qed
\citet{primal_dual_boosting} studied functions of the form we are considering,
and showed that the points ${\bf u}_1,\ldots,{\bf u}_m$ can be partitioned
into an \emph{easy set} and a \emph{hard core}.
Roughly speaking, when $f$ is minimized over ${\bf x}\in{\R^n}$,
if ${\bf u}_i$ is in the easy set, then the corresponding term
$\ell_i({\bf x}\cdot{\bf u}_i)$ is reduced to $0$, its minimum value, which is
achieved by driving ${\bf x}\cdot{\bf u}_i$ to $-\infty$.
This leaves the problem of minimizing the remaining terms in $f$,
those involving the hard core.
These terms cannot be driven to their minimum values; rather, for these,
${\bf x}\cdot{\bf u}_i$ must converge to some finite value.
Here, we revisit these notions, re-casting them in terms of central
concepts developed in this manuscript.
In astral terms,
${\bf u}_i$ is considered easy if there exists a point
$\overline{\xx}\in\extspac{n}$ with $\overline{\xx}\cdot{\bf u}_i=-\infty$ (so that
${\overline{\ell}}_i(\overline{\xx}\cdot{\bf u}_i)$ equals $0$, ${\overline{\ell}}_i$'s minimum value)
and for which ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$.
This is equivalent to there existing a sequence $\seq{{\bf x}_t}$ in ${\R^n}$ for
which ${\bf x}_t\cdot{\bf u}_i\rightarrow-\infty$,
without $f({\bf x}_t)$ becoming unboundedly large.
Otherwise, if there exists no such $\overline{\xx}$, then ${\bf u}_i$ is hard.
We can write any $\overline{\xx}\in\extspac{n}$ as $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some
$\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
The condition that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$ implies that
$\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by Corollary~\ref{cor:a:4}.
Further,
if $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ and
if it happens that ${\bf u}_i\in{(\aresconef)^{\bot}}$, then
$\overline{\ee}\cdot{\bf u}_i=0$, implying that $\overline{\xx}\cdot{\bf u}_i={\bf q}\cdot{\bf u}_i$,
which is in ${\mathbb{R}}$.
This shows that if ${\bf u}_i\in{(\aresconef)^{\bot}}$ then it must be hard in the
sense just described, since
for all $\overline{\xx}\in\extspac{n}$, either ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$ or
$\overline{\xx}\cdot{\bf u}_i>-\infty$.
This motivates our formal definition of the hard core, namely,
\begin{equation} \label{eq:hard-core:3}
\hardcore{f} = \{ i \in \{1,\ldots,m\} : {\bf u}_i\in\rescperp{f} \},
\end{equation}
where we have taken ${\bf u}_1,\ldots,{\bf u}_m$ to be
fixed and given to simplify notation.
Note also that $\rescperp{f}={(\aresconef)^{\bot}}$
since $f$ is recessive complete,
by Propositions~\ref{pr:perpres-is-rescperp}
and~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b}).
We have seen already that
the set $\unimin{f}$ of universal reducers
together with the universal reduction $\fullshad{f}$ are central
elements in the general theory we have developed for minimizing convex
functions.
Both of these can be expressed precisely in terms of the hard core, as
we show next:
\begin{theorem} \label{thm:hard-core:3}
Let $f:{\R^n}\rightarrow{\mathbb{R}}$ have the form given in
\eqref{eqn:loss-sum-form}, where, for $i=1,\ldots,m$,
${\bf u}_i\in{\R^n}$ and
$\ell_i:{\mathbb{R}}\rightarrow{\mathbb{R}}$ is convex, nondecreasing, not constant,
with $\inf \ell_i = 0$.
\begin{enumerate}
\item \label{thm:hard-core:3:a}
If $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ then $\overline{{\bf y}}\cdot{\bf u}_i=0$ for
all $i\in\hardcore{f}$.
\item \label{thm:hard-core:3:b}
Let $\overline{{\bf y}}\in\extspac{n}$.
Then $\overline{{\bf y}}\in\unimin{f}$ if and only if
\begin{equation} \label{eq:hard-core:2}
\overline{{\bf y}}\cdot{\bf u}_i = \left\{ \begin{array}{cl}
0 & \mbox{if $i\in\hardcore{f}$} \\
-\infty & \mbox{otherwise} \\
\end{array} \right.
\end{equation}
for $i=1,\ldots,m$.
\item \label{thm:hard-core:3:c}
Let $\overline{{\bf y}}\in\extspac{n}$, and
suppose, for some $i\in\hardcore{f}$, that
$\overline{{\bf y}}\cdot{\bf u}_i<0$.
Then there exists $j\in\hardcore{f}$ for which
$\overline{{\bf y}}\cdot{\bf u}_j>0$.
\item \label{thm:hard-core:3:d}
For ${\bf x}\in{\R^n}$,
\[ \fullshad{f}({\bf x}) = \sum_{i\in\hardcore{f}} \ell_i({\bf x}\cdot{\bf u}_i). \]
\end{enumerate}
\end{theorem}
\proof
~
Part~(\ref{thm:hard-core:3:a}):
If $i\in\hardcore{f}$ then ${\bf u}_i\in\rescperp{f}={(\aresconef)^{\bot}}$,
meaning exactly that $\overline{{\bf y}}\cdot{\bf u}_i=0$ if $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Part~(\ref{thm:hard-core:3:b}):
Suppose $\overline{{\bf y}}\in\unimin{f}$.
Then $\overline{{\bf y}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by
Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:a}),
so
$\overline{{\bf y}}\cdot{\bf u}_i\leq 0$ for $i=1,\ldots,m$
(by Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b})).
Specifically,
$\overline{{\bf y}}\cdot{\bf u}_i=0$ for $i\in \hardcore{f}$,
by part~(\ref{thm:hard-core:3:a}).
It remains then only to show that if $i\not\in\hardcore{f}$ then
$\overline{{\bf y}}\cdot{\bf u}_i=-\infty$, which would be implied by showing that
$\overline{{\bf y}}\cdot{\bf u}_i\not\in{\mathbb{R}}$.
Suppose then, by way of contradiction, that there exists
$j\not\in\hardcore{f}$ with $\overline{{\bf y}}\cdot{\bf u}_j\in{\mathbb{R}}$.
By definition,
since $j\not\in\hardcore{f}$,
there must exist ${\bf v}\in\resc{f}$ with ${\bf v}\cdot{\bf u}_j\neq 0$,
implying, by Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b}),
that actually ${\bf v}\cdot{\bf u}_j< 0$.
Let $\lambda\in{\mathbb{R}}$.
To derive a contradiction,
we compare function values at
$\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j$
and
$\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j$.
For $i=1,\ldots,m$, by application of
Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b}),
${\bf v}\cdot{\bf u}_i\leq 0$ and so
$\limray{{\bf v}}\cdot{\bf u}_i\leq 0$,
implying $\limray{{\bf v}}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Since ${\overline{\ell}}_i$ is nondecreasing,
this further shows that
\begin{eqnarray}
{\overline{\ell}}_i((\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_i)
&=&
{\overline{\ell}}_i(\limray{{\bf v}}\cdot{\bf u}_i
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_i)
\nonumber
\\
&\leq&
{\overline{\ell}}_i((\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_i).
\label{eqn:thm:hard-core:3:2}
\end{eqnarray}
In particular, when $i=j$,
\[
{\overline{\ell}}_j((\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_j)
=
{\overline{\ell}}_j(\limray{{\bf v}}\cdot{\bf u}_j
\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_j)
= 0
\]
since $\limray{{\bf v}}\cdot{\bf u}_j=-\infty$.
On the other hand,
\[
{\overline{\ell}}_j((\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_j)
=
{\overline{\ell}}_j(\overline{{\bf y}}\cdot{\bf u}_j \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf u}_j\cdot{\bf u}_j)
\rightarrow +\infty
\]
as $\lambda\rightarrow+\infty$,
since $\overline{{\bf y}}\cdot{\bf u}_j\in{\mathbb{R}}$ and ${\bf u}_j\neq{\bf 0}$ (since ${\bf v}\cdot{\bf u}_j<0$),
and since ${\overline{\ell}}_j$ is nondecreasing and not constant.
Thus, \eqref{eqn:thm:hard-core:3:2} holds for all
$i=1,\ldots,m$, and furthermore
the inequality is strict when $i=j$ and
when $\lambda$ is sufficiently large.
Therefore,
combining with Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:a}),
for $\lambda$ sufficiently large, we have shown that
\begin{eqnarray}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf u}_j)
&=&
\sum_{i=1}^m
{\overline{\ell}}_i((\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_i)
\nonumber
\\
&<&
\sum_{i=1}^m
{\overline{\ell}}_i((\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\lambda{\bf u}_j)\cdot{\bf u}_i)
\nonumber
\\
&=&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \lambda{\bf u}_j)
\leq
\fullshad{f}(\lambda{\bf u}_j),
\label{eq:hard-core:1}
\end{eqnarray}
where the final inequality is because $\overline{{\bf y}}\in\unimin{f}$.
On the other hand, since $\limray{{\bf v}}$ and $\overline{{\bf y}}$ are both in
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, their leftward sum $\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$ is as well
(since
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$
for $\overline{\xx}\in\extspac{n}$),
which means
\eqref{eq:hard-core:1} directly contradicts
Proposition~\ref{pr:fullshad-equivs}.
Thus, if $\overline{{\bf y}}\in\unimin{f}$, then
\eqref{eq:hard-core:2} is satisfied for $i=1,\ldots,m$.
For the converse, suppose
\eqref{eq:hard-core:2} is satisfied for $i=1,\ldots,m$.
Let $\overline{\ee}$ be any point in $\unimin{f}$ (which is nonempty by
Theorem~\ref{thm:new:thm:f:8cd}).
Then, as just argued, $\overline{\ee}$ satisfies \eqref{eq:hard-core:2} as
well, so $\overline{\ee}\cdot{\bf u}_i=\overline{{\bf y}}\cdot{\bf u}_i$
for $i=1,\ldots,m$.
So for all ${\bf x}\in{\R^n}$,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf x})\leq\fullshad{f}({\bf x})$
by
Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:a}).
Therefore, $\overline{{\bf y}}\in\unimin{f}$.
Part~(\ref{thm:hard-core:3:c}):
Let $\overline{\ee}$ be any point in $\unimin{f}$
(which exists by Theorem~\ref{thm:new:thm:f:8cd}), and let
$\overline{{\bf z}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$.
Then $\overline{{\bf z}}\cdot{\bf u}_i<0$ since
$\overline{\ee}\cdot{\bf u}_i=0$
by part~(\ref{thm:hard-core:3:b}), so
$\overline{{\bf z}}\not\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
by part~(\ref{thm:hard-core:3:a}).
Therefore, for some $j\in\{1,\ldots,m\}$,
$\overline{{\bf z}}\cdot{\bf u}_j>0$, by
Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b}).
Further, it must be that $j\in\hardcore{f}$ since
otherwise
part~(\ref{thm:hard-core:3:b}) would imply
$\overline{\ee}\cdot{\bf u}_j=-\infty$, so also
$\overline{{\bf z}}\cdot{\bf u}_j=-\infty$.
Thus,
$\overline{{\bf y}}\cdot{\bf u}_j>0$
since $\overline{\ee}\cdot{\bf u}_j=0$
by part~(\ref{thm:hard-core:3:b}).
Part~(\ref{thm:hard-core:3:d}):
This follows immediately from
Proposition~\ref{pr:new:thm:f:8a}(\ref{pr:new:thm:f:8a:b})
by plugging the result of part~(\ref{thm:hard-core:3:b})
into the expression given in
Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:a}).
\qed
If icon $\overline{\ee}$ is a universal reducer, and if finite ${\bf q}$
minimizes $\fullshad{f}$,
then $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ is a universal minimizer of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
In the current setting,
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:b})
shows that all universal reducers $\overline{\ee}$ are identical in terms of
the values of $\overline{\ee}\cdot{\bf u}_i$, for $i=1,\ldots,m$, as determined by
the hard core.
Every universal reducer has the effect of causing those terms
$i$ in \eqref{eqn:loss-sum-form} that are not in the hard core to vanish.
The remaining terms, those that are in the hard core, constitute
exactly the universal reduction $\fullshad{f}$.
As is generally the case (Theorem~\ref{thm:f:4x}),
all sublevel sets of this function are compact when restricted to
${(\aresconef)^{\bot}}=\rescperp{f}$, and the function is constant in all
directions orthogonal to this subspace.
We also remark that Theorem~\ref{thm:unimin-can-be-rankone},
combined with
Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b}),
shows that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
can always be minimized along a ray, specifically, at any point
$\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ if ${\bf v}\in\ric{f}$ and ${\bf q}$ minimizes
$\fullshad{f}$.
Notice that the sets we have been considering, namely, the standard
and astral recession cones, the set of universal reducers, as well
as the hard core, all depend exclusively on ${\bf u}_1,\ldots,{\bf u}_m$,
and are entirely {independent} of the specific functions $\ell_i$.
In other words, suppose we form a new function $f'$ as in
\eqref{eqn:loss-sum-form} with ${\bf u}_1,\ldots,{\bf u}_m$ unchanged, but
with each $\ell_i$ replaced by some other function $\ell'_i$ (though
still satisfying the same assumed properties).
Then the sets listed above are unchanged.
That is,
$\resc{f}=\resc{f'}$ and
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}={\arescone \fpext}$
by Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b});
$\hardcore{f}=\hardcore{f'}$ by the definition in
\eqref{eq:hard-core:3}; and
$\unimin{f}=\unimin{f'}$ by
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:b}).
Finally, as noted repeatedly, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is minimized at a point
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
if $\overline{\ee}\in(\unimin{f})\cap{{\cal E}_n}$ and ${\bf q}$ minimizes $\fullshad{f}$.
If, in addition to our preceding assumptions,
each $\ell_i$ is \emph{strictly} increasing, then these
conditions are not only sufficient but also necessary for $\overline{\xx}$ to
minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Furthermore, if each function $\ell_i$ is \emph{strictly} convex,
then $\fullshad{f}$
is uniquely minimized over the linear subspace
${(\aresconef)^{\bot}} = \rescperp{f}$:
\begin{proposition} \label{pr:hard-core:2}
Let $f:{\R^n}\rightarrow{\mathbb{R}}$ have the form given in
\eqref{eqn:loss-sum-form}, where, for $i=1,\ldots,m$,
${\bf u}_i\in{\R^n}$ and
$\ell_i:{\mathbb{R}}\rightarrow{\mathbb{R}}$ is convex and strictly increasing,
with $\inf \ell_i = 0$.
\begin{enumerate}
\item \label{pr:hard-core:2:a}
Let $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
Then $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ if and only if $\overline{\ee}\in\unimin{f}$
and ${\bf q}$ minimizes $\fullshad{f}$.
\item \label{pr:hard-core:2:b}
Suppose, in addition, that $\ell_i$ is strictly convex for
$i=1,\ldots,m$.
Then $\fullshad{f}$, if restricted to ${(\aresconef)^{\bot}} =
\rescperp{f}$, has a unique minimizer.
\end{enumerate}
\end{proposition}
\proof
Part~(\ref{pr:hard-core:2:a}):
The ``if'' direction follows from
Proposition~\ref{pr:unimin-to-global-min}.
For the converse, suppose $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Let $\overline{{\bf y}}$ be any point in $\unimin{f}$
(which exists by Theorem~\ref{thm:new:thm:f:8cd}).
Then for all ${\bf z}\in{\R^n}$,
\[
\fullshad{f}({\bf q})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
\leq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z})
\leq
\fullshad{f}({\bf z})
\]
where the inequalities are, respectively,
by \eqref{eqn:fullshad-def};
since $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$;
and
since $\overline{{\bf y}}\in\unimin{f}$.
Thus, ${\bf q}$ minimizes $\fullshad{f}$.
To show $\overline{\ee}\in\unimin{f}$, note
that $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
by Theorem~\ref{thm:arescone-fshadd-min}
(since $\overline{\xx}$ minimizes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$).
Therefore,
by Propositions~\ref{pr:icon-equiv}
and~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b}),
$\overline{\ee}\cdot{\bf u}_i\in\{-\infty,0\}$ for $i=1,\ldots,m$ since
$\overline{\ee}\in{{\cal E}_n}$.
Let $H'=\{ i \in \{1,\ldots,m\} : \overline{\ee}\cdot{\bf u}_i = 0\}$.
We claim that $H'=\hardcore{f}$.
The inclusion $\hardcore{f}\subseteq H'$ follows directly from
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:a}).
For the reverse inclusion,
suppose, by way of contradiction,
that there exists an index
$j$ in $H'\setminus\hardcore{f}$.
Then
\begin{eqnarray*}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q})
=
\sum_{i=1}^m {\overline{\ell}}_i(\overline{\ee}\cdot{\bf u}_i\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}\cdot{\bf u}_i)
&=&
\sum_{i\in H'} \ell_i({\bf q}\cdot{\bf u}_i)
\\
&>&
\sum_{i\in \hardcore{f}} \ell_i({\bf q}\cdot{\bf u}_i)
=
\fullshad{f}({\bf q})
\geq
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}).
\end{eqnarray*}
The first two equalities are from
Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:a})
and \eqref{eqn:hard-core:1}.
The first inequality is because $\ell_j({\bf q}\cdot{\bf u}_j)>{\overline{\ell}}_j(-\infty)=0$
since $\ell_j$ is strictly increasing
(and since
$\hardcore{f}\subseteq H'$ and
$j\in H'\setminus\hardcore{f}$).
The third equality is by
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:d}),
and the last inequality is because $\overline{{\bf y}}\in\unimin{f}$.
This is a contradiction since $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ minimizes
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
Thus, $H'=\hardcore{f}$ and therefore $\overline{\ee}\in\unimin{f}$ by
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:b}).
Part~(\ref{pr:hard-core:2:b}):
Suppose, by way of contradiction, that ${\bf q}$ and ${\bf q}'$ are distinct
minimizers of $\fullshad{f}$, both in $\rescperp{f}$.
We claim first that
${\bf q}\cdot{\bf u}_i\neq{\bf q}'\cdot{\bf u}_i$
for some $i\in\hardcore{f}$.
Suppose to the contrary that ${\bf d}\cdot{\bf u}_i=0$ for all
$i\in\hardcore{f}$, where ${\bf d}={\bf q}'-{\bf q}\neq{\bf 0}$.
Then because $\rescperp{f}$ is a linear subspace
(Proposition~\ref{pr:perp-props}(\ref{pr:perp-props:a})),
for all $\lambda\in{\mathbb{R}}$,
$\lambda{\bf d}\in\rescperp{f}$.
Furthermore, by
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:d}),
$\fullshad{f}(\lambda{\bf d})=\fullshad{f}({\bf 0})<+\infty$.
In other words, one of the sublevel sets of $\fullshad{f}$ in
$\rescperp{f}={(\aresconef)^{\bot}}$ includes the entire line
$\{ \lambda{\bf d} : \lambda\in{\mathbb{R}}\}$.
This, however, is a contradiction since all such sublevel sets are bounded,
by Theorem~\ref{thm:f:4x}(\ref{thm:f:4xb}).
So let $i\in\hardcore{f}$ be such that
${\bf q}\cdot{\bf u}_i\neq{\bf q}'\cdot{\bf u}_i$.
Let ${\bf z}=({\bf q}+{\bf q}')/2$.
Since each $\ell_j$ is convex,
$\ell_j({\bf z}\cdot{\bf u}_j)\leq(\ell_j({\bf q}\cdot{\bf u}_j)+\ell_j({\bf q}'\cdot{\bf u}_j))/2$.
And when $j=i$, by strict convexity of $\ell_i$, this inequality is strict.
Therefore, applying
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:d})
yields
$\fullshad{f}({\bf z}) < (\fullshad{f}({\bf q}) + \fullshad{f}({\bf q}'))/2$,
contradicting the assumption that ${\bf q}$ and ${\bf q}'$ both minimize
$\fullshad{f}$.
\qed
Finally, we mention some implications for minimizing sequences.
Let $\seq{{\bf x}_t}$ be a sequence in ${\R^n}$ that minimizes some function
$f$ satisfying the assumptions of
Proposition~\ref{pr:hard-core:2}.
Then it can be argued using
Theorem~\ref{thm:hard-core:3}(\ref{thm:hard-core:3:b})
and
Proposition~\ref{pr:hard-core:2}(\ref{pr:hard-core:2:a})
that
${\bf x}_t\cdot{\bf u}_i\rightarrow-\infty$ for all $i\not\in\hardcore{f}$.
Also, let ${\bf q}_t$ be the projection of ${\bf x}_t$ onto the linear
subspace $\rescperp{f}={(\aresconef)^{\bot}}$.
Then by
Proposition~\ref{pr:proj-mins-fullshad},
$\fullshad{f}({\bf x}_t)=\fullshad{f}({\bf q}_t) \rightarrow \min \fullshad{f}$,
and furthermore, the ${\bf q}_t$'s are all
in a compact region of $\rescperp{f}$.
Thus,
if $i\in\hardcore{f}$, then ${\bf u}_i\in\rescperp{f}$,
by $\hardcore{f}$'s definition,
implying ${\bf x}_t\cdot{\bf u}_i={\bf q}_t\cdot{\bf u}_i$ for all $t$
(since ${\bf x}_t-{\bf q}_t$ is orthogonal to $\rescperp{f}$).
Therefore, ${\bf x}_t\cdot{\bf u}_i$ remains always in some bounded interval of
${\mathbb{R}}$.
If, in addition, each $\ell_i$ is strictly convex, then
$\fullshad{f}$ has a unique minimum ${\bf q}$ in $\rescperp{f}$
(by
Proposition~\ref{pr:hard-core:2}(\ref{pr:hard-core:2:b}));
in this case, it can be further argued that ${\bf q}_t\rightarrow{\bf q}$,
so ${\bf x}_t\cdot{\bf u}_i\rightarrow{\bf q}\cdot{\bf u}_i$ for $i\in\hardcore{f}$.
\section{Continuity}
\label{sec:continuity}
As seen in earlier examples, the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of a convex function
$f:{\R^n}\rightarrow\overline{\R}$ may or may not be continuous at a particular
point, even if the function $f$ is continuous everywhere and
well-behaved in other ways.
Nevertheless,
in this section, we will characterize exactly the set of points where
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous in terms of properties of the original function
$f$.
We also will give precise necessary and sufficient conditions for
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ to be continuous everywhere.
{
We begin with some examples.
Consider first the function in
\eqref{eqn:recip-fcn-eg}.
We previously argued that if $\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\beta{\bf e}_2$
and $\beta\neq 0$, then on any sequence converging to $\overline{\xx}$, $f$ has
the same limit ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$;
in other words, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$.
However, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at $\overline{\xx}=\limray{{\bf e}_1}$:
We have already seen that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=0$, but on other sequences
converging to $\overline{\xx}$, $f$ need not converge to $0$.
For instance, for any constant $\lambda>0$, let ${\bf x}_t=\trans{[t/\lambda, 1/t]}$.
This sequence converges to $\overline{\xx}$, but
$f({\bf x}_t)=\lambda$ for all $t$.
Thus, on sequences converging to $\overline{\xx}$, $f$ can converge to any
nonnegative value.
As a second example, suppose in ${\mathbb{R}}^2$ that
\begin{equation} \label{eqn:curve-discont-finiteev-eg}
f(x_1,x_2)
=
\left\{ \begin{array}{cl}
{x_1^2}/{x_2} & \mbox{if $x_2 > |x_1|$} \\
2|x_1| - x_2 & \mbox{otherwise.} \\
\end{array} \right.
\end{equation}
This function is actually the same as the one in
\eqref{eqn:curve-discont-notfinite-eg} on the part of the domain where
$x_2\geq |x_1|$, and indeed,
the observations below apply to that function as well.
Here we have modified that function to yield the one in
\eqref{eqn:curve-discont-finiteev-eg} which is
not only convex, closed and proper, but also finite everywhere and
continuous everywhere.
Despite these properties,
$f$'s extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous, for instance, at
$\overline{\xx}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$.
For example, if ${\bf x}_t=\trans{[t,t^{3}]}$, then
${\bf x}_t\rightarrow\overline{\xx}$ and $f({\bf x}_t)=1/t\rightarrow 0$.
If instead ${\bf x}_t=\trans{[t,t^{3/2}]}$, then
${\bf x}_t\rightarrow\overline{\xx}$ but now
$f({\bf x}_t)=\sqrt{t}\rightarrow +\infty$.
Thus, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at $\overline{\xx}$.
Indeed, if ${\bf x}_t=\trans{[t,t^2/\lambda]}$, for any constant $\lambda>0$,
then ${\bf x}_t\rightarrow\overline{\xx}$ but $f({\bf x}_t)=\lambda$ for all
sufficiently large $t$.
Thus, on a sequence converging to $\overline{\xx}$, $f$ can converge to any
number in $[0,+\infty]$.
At an intuitive level,
these examples suggest two different ways in
which ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ can be discontinuous.
In the first example, the discontinuity seemed to arise as a result of
reaching the boundary between where the function $f$ is finite
(namely, all points with $x_1>0$ and $x_2>0$), and where it is
infinite, in other words, the boundary of $\dom{f}$.
On the other hand, in the second example, the function $f$ is finite
everywhere so there is no such boundary to its effective domain.
Instead, the discontinuity seemed to arise as a result of the
variety of ways in which we can follow a ``curved'' trajectory
reaching the same
astral point at infinity, but on which the function takes very
different values.
We will soon see that our characterization of continuity exactly
captures these two different kinds of discontinuity.
}
\subsection{Characterizing exactly where \texorpdfstring{${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$}{the extension of f} is continuous}
\label{subsec:charact:cont}
We turn now to the general case of a convex function
$f:{\R^n}\rightarrow\overline{\R}$.
Note first that
if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$, then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is necessarily continuous at $\overline{\xx}$
since $f$ must converge to $+\infty$ on every sequence converging to
$\overline{\xx}$.
Therefore, we focus on understanding continuity at points $\overline{\xx}$
where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})< +\infty$, that is, astral points in $\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Let $\contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ denote the set of all points $\overline{\xx}\in\extspac{n}$
where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous, and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$.
In this subsection, we will characterize exactly those points where
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous; in other words, we will determine the set
$\contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ exactly.
We will do this in two different ways.
On the one hand, we will see that $\contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is equal to the
interior of the effective domain of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, that is,
$\contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
This means that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere except for points
that are in $\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, but not its interior.
This provides a close analog to the continuity properties of standard
convex functions on ${\R^n}$.
In addition, we further characterize $\contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ in terms of the
original function $f$ itself.
In particular, we will see that this set consists exactly of all
points $\overline{\xx}\in\extspac{n}$ of a specific form
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
where ${\bf q}\in{\R^n}$ is in the interior of the effective domain (in
${\R^n}$) of $f$, and $\overline{\ee}\in{{\cal E}_n}$ is an icon in the representational
closure (in $\extspac{n}$) of $f$'s standard recession cone, $\resc{f}$;
that is, $\overline{\ee}\in\represc{f}$.
Thus,
\[
\contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} = (\represc{f}\cap{{\cal E}_n}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f}.
\]
We prove these results in a series of theorems.
We start by proving that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ must be continuous at every point of
the form just described.
\begin{theorem} \label{thm:g:2}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Suppose $\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where
$k\geq 0$, ${\bf v}_1,\ldots,{\bf v}_k\in\resc{f}$ and
${\bf q}\in\intdom{f}$.
Then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$ and
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$.
In other words,
\[
(\represc{f}\cap{{\cal E}_n}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f}
\subseteq \contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}.
\]
\end{theorem}
\proof
The main idea of the proof is to first construct a sequence
$\seq{{\hat{\bf x}}_t}$ in ${\R^n}$
converging to $\overline{\xx}$ whose limit can be easily determined, and to
then show, by comparison to this sequence, that every other sequence
$\seq{{\bf x}_t}$ in ${\R^n}$ converging to $\overline{\xx}$ must have the same limit.
We begin with a few preliminaries and some notation.
Without loss of generality,
by Proposition~\ref{pr:g:2}, we assume
${\bf v}_1,\ldots,{\bf v}_k$ are linearly independent
(though not necessarily orthonormal).
Also, let ${\bf V}=[{\bf v}_1,\ldots,{\bf v}_k]$.
For vectors ${\bf b},\,{\bf c}\in{\R^k}$, we write ${\bf b}\geq{\bf c}$ to mean
$b_i\geq c_i$ for every component $i\in\{1,\ldots,k\}$.
Note that if ${\bf b}\geq{\bf 0}$, then
${\bf V}{\bf b}=\sum_{i=1}^k b_i {\bf v}_i$
is in $\resc{f}$, being a convex cone
(Proposition~\ref{pr:resc-cone-basic-props}).
Let $L$ be the linear subspace perpendicular to
all of the ${\bf v}_i$'s:
\[
L = \{{\bf z}\in{\R^n} : \trans{{\bf V}}{\bf z} = {\bf 0} \}.
\]
Then because the ${\bf v}_i$'s are linearly independent,
we can write ${\bf q}$ uniquely in the form ${\bf q}={\bf V}{\hat{\bf c}}+{\hat{\bf q}}$
for some ${\hat{\bf q}}\in L$ and ${\hat{\bf c}}\in{\R^k}$.
Thus, $\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\hat{\bf q}}$
(by Proposition~\ref{pr:g:1}).
Since ${\bf q}\in\intdom{f}$, there exists an open set $U\subseteq{\R^n}$
that includes ${\bf 0}$ and
such that ${\bf q}+U\subseteq\dom{f}$.
To construct the sequence $\seq{{\hat{\bf x}}_t}$ mentioned above, we first define
the function $h:{\R^n}\rightarrow\overline{\R}$ given by
\[ h({\bf x})=\inf_{{\bf b}\in{\R^k}} f({\bf V}{\bf b}+{\bf x}) \]
for ${\bf x}\in{\R^n}$.
This function is convex by
\citet[Theorem~5.7]{ROC}.
Note that
\begin{equation} \label{eqn:h-prop:1}
h({\bf V}{\bf b}+{\bf x})=h({\bf x})\leq f({\bf x})
\end{equation}
for all ${\bf x}\in{\R^n}$ and ${\bf b}\in{\R^k}$.
In particular,
\[h({\hat{\bf q}})=h({\bf V}{\hat{\bf c}}+{\hat{\bf q}})=h({\bf q})\leq f({\bf q})<+\infty,\]
since ${\bf q}\in\dom{f}$.
Let $\seq{{\tilde{\bf b}_{t}}}$ be a sequence in ${\R^k}$ realizing successive
approximations of the
infimum in the definition of $h({\hat{\bf q}})$;
specifically, for each $t$, let ${\tilde{\bf b}_{t}}$ be such that
\[ f({\bf V} {\tilde{\bf b}_{t}}+{\hat{\bf q}}) <
\left\{ \begin{array}{cl}
h({\hat{\bf q}}) + \frac{1}{t} & \mbox{if $h({\hat{\bf q}})\in{\mathbb{R}}$} \\
-t & \mbox{if $h({\hat{\bf q}})=-\infty$.}
\end{array}
\right.
\]
Next, let us define a sequence $\seq{{\hat{\bf b}}_t}$ in ${\R^k}$
with the properties that:
\begin{itemize}
\item
${\hat{\bf b}}_t\geq {\tilde{\bf b}_{t}}$ for all $t$; and
\item
${\bf V}{\hat{\bf b}}_t \rightarrow {\bf V}{\bar{\bm{\omega}}}$.
\end{itemize}
Concretely, we can achieve both of these by defining each ${\hat{\bf b}}_t$
to have components
\[
{\hat{b}}_{t,i} = \left\{ \begin{array}{ll}
\max\braces{ {\tilde{b}}_{t,k},\; t } & \mbox{if $i=k$} \\
\max\braces{ {\tilde{b}}_{t,i},\; t\,{\hat{b}}_{t,i+1} }
& \mbox{otherwise,}
\end{array} \right.
\]
for $i=1,\ldots,k$.
Clearly, ${\hat{\bf b}}_t\geq{\tilde{\bf b}_{t}}$.
Furthermore,
${\hat{b}}_{t,k}\geq t$, and
${\hat{b}}_{t,i}\geq t{\hat{b}}_{t,i+1}$,
for $i=1,\ldots,k-1$, which implies each component $i$ is growing to
$+\infty$ much faster than the next, so that
${\bf V}{\hat{\bf b}}_t \rightarrow {\bf V}{\bar{\bm{\omega}}}$
(by Theorem~\ref{thm:seq-rep}).
Finally, let ${\hat{\bf x}}_t={\bf V} {\hat{\bf b}}_t + {\hat{\bf q}}$.
Then ${\hat{\bf x}}_t\rightarrow{\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\hat{\bf q}}=\overline{\xx}$
(by Proposition~\ref{pr:aff-cont}).
The next claim shows that $f({\hat{\bf x}}_t)$ converges to $h({\hat{\bf q}})$,
which will turn out to be the common limit of every sequence
converging to $\overline{\xx}$.
\setcounter{claimp}{0}
\medskip
\begin{claimp} \label{cl:a1}
$f({\hat{\bf x}}_t)\rightarrow h({\hat{\bf q}})$.
\end{claimp}
\proof
If $h({\hat{\bf q}})\in{\mathbb{R}}$, then for each $t$,
\begin{eqnarray*}
h({\hat{\bf q}})
&\leq&
f({\bf V} {\hat{\bf b}}_t+{\hat{\bf q}})
\nonumber
\\
&=&
f\paren{
{\bf V} {\tilde{\bf b}_{t}} + {\hat{\bf q}} + {\bf V} ({\hat{\bf b}}_t - {\tilde{\bf b}_{t}})
}
\nonumber
\\
&\leq&
f({\bf V} {\tilde{\bf b}_{t}} + {\hat{\bf q}})
\label{eq:a:2}
\\
&<&
h({\hat{\bf q}}) + \frac{1}{t},
\nonumber
\end{eqnarray*}
where the second inequality is because
${\bf V} ({\hat{\bf b}}_t - {\tilde{\bf b}_{t}}) \in \resc{f}$, since
${\hat{\bf b}}_t \geq {\tilde{\bf b}_{t}}$.
By a similar argument,
if $h({\hat{\bf q}})=-\infty$, then for each $t$,
\[ f({\bf V} {\hat{\bf b}}_t + {\hat{\bf q}})
\leq
f({\bf V} {\tilde{\bf b}_{t}} + {\hat{\bf q}})
< -t.
\]
Thus, in either case,
$f({\hat{\bf x}}_t)=f({\bf V}{\hat{\bf b}}_t + {\hat{\bf q}})\rightarrow h({\hat{\bf q}})$.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Now let $\seq{{\bf x}_t}$ be any sequence in ${\R^n}$ converging to $\overline{\xx}$.
To prove continuity at $\overline{\xx}$,
we aim to show that $f({\bf x}_t)$ has the same limit as $f({\hat{\bf x}}_t)$,
namely, $h({\hat{\bf q}})$.
By Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2b}),
it is sufficient to prove this only for
sequences in ${\R^n}$, rather than all of $\extspac{n}$.
By linear algebra,
for each $t$, we
can write ${\bf x}_t={\bf V}{\bf b}_t+{\bf q}_t$ for some ${\bf b}_t\in{\R^k}$, and some
${\bf q}_t\in L$.
\medskip
\begin{claimp} \label{cl:a2-and-3}
$b_{t,i}\rightarrow+\infty$, for $i=1,\ldots,k$.
Also, ${\bf q}_t\rightarrow{\hat{\bf q}}$.
\end{claimp}
\proof
Both of these facts are immediate from
Theorem~\ref{thm:seq-rep}.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
\medskip
\begin{claimp} \label{cl:a4}
$h$ is continuous at ${\hat{\bf q}}$.
\end{claimp}
\proof
For all ${\bf s}\in U$,
\[
h({\hat{\bf q}}+{\bf s})
=h({\bf V}{\hat{\bf c}} + {\hat{\bf q}}+{\bf s})
=h({\bf q} +{\bf s})
\leq f({\bf q}+{\bf s})<+\infty.
\]
The first equality and first inequality are both
from \eqref{eqn:h-prop:1}.
The last inequality is because ${\bf q}+U\subseteq\dom{f}$.
Thus, ${\hat{\bf q}}+U\subseteq\dom{h}$, so
${\hat{\bf q}}\in\intdom{h}$.
Since $h$ is convex, this implies that $h$ is continuous at ${\hat{\bf q}}$
\citep[by][Theorem~10.1]{ROC}.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
\medskip
\begin{claimp} \label{cl:a5}
$\liminf f({\bf x}_t) \geq h({\hat{\bf q}})$.
\end{claimp}
\proof
We have that
\begin{eqnarray*}
\liminf f({\bf x}_t)
&=&
\liminf f({\bf V} {\bf b}_t + {\bf q}_t)
\\
&\geq&
\liminf h({\bf q}_t)
\\
&=&
h({\hat{\bf q}}).
\end{eqnarray*}
The inequality is by \eqref{eqn:h-prop:1},
and the second equality is
by Claims~\ref{cl:a2-and-3} and~\ref{cl:a4}.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
\medskip
\begin{claimp} \label{cl:a6}
For all $t$ sufficiently large, $f$ is continuous at
${\hat{\bf x}}_t$.
\end{claimp}
\proof
Let us suppose that $t$ is so large that ${\hat{\bf b}}_t\geq{\hat{\bf c}}$, as must
happen eventually since every component of ${\hat{\bf b}}_t$ is growing to
$+\infty$
(by Claim~\ref{cl:a2-and-3}).
Then for all ${\bf s}\in U$,
\begin{eqnarray*}
f({\hat{\bf x}}_t + {\bf s})
&=&
f({\bf V} {\hat{\bf b}}_t + {\hat{\bf q}} + {\bf s})
\\
&=&
f({\bf V} {\hat{\bf c}} + {\hat{\bf q}} + {\bf s} + {\bf V}({\hat{\bf b}}_t-{\hat{\bf c}}))
\\
&=&
f({\bf q} + {\bf s} + {\bf V}({\hat{\bf b}}_t-{\hat{\bf c}}))
\\
&\leq&
f({\bf q} + {\bf s}) < +\infty.
\end{eqnarray*}
The first inequality is because
${\bf V}({\hat{\bf b}}_t-{\hat{\bf c}})\in\resc{f}$, since
${\hat{\bf b}}_t\geq{\hat{\bf c}}$,
and the last inequality is because ${\bf q}+U\subseteq\dom{f}$.
Thus, ${\hat{\bf x}}_t+U\subseteq\dom{f}$, so
${\hat{\bf x}}_t\in\intdom{f}$; therefore, $f$ is continuous at ${\hat{\bf x}}_t$
\citep[Theorem~10.1]{ROC}.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
\medskip
\begin{claimp} \label{cl:a7}
$\limsup f({\bf x}_t) \leq h({\hat{\bf q}})$.
\end{claimp}
\proof
Suppose first that $h({\hat{\bf q}})\in{\mathbb{R}}$.
Let $\epsilon>0$.
By Claims~\ref{cl:a1} and~\ref{cl:a6}, we can choose $t_0$ so that
$f({\hat{\bf x}}_{t_0}) < h({\hat{\bf q}}) + \epsilon$,
and so that $f$ is continuous at ${\hat{\bf x}}_{t_0}={\bf V} {\hat{\bf b}}_{t_0}+{\hat{\bf q}}$.
By continuity at this point, there exists $\delta>0$ such that for all
${\bf q}'\in{\R^n}$, if $\norm{{\bf q}'-{\hat{\bf q}}}<\delta$ then
\[ f({\bf V} {\hat{\bf b}}_{t_0}+{\bf q}') < f({\bf V} {\hat{\bf b}}_{t_0}+{\hat{\bf q}}) + \epsilon. \]
By Claim~\ref{cl:a2-and-3},
for all $t$ sufficiently large,
$\norm{{\bf q}_t - {\hat{\bf q}}} < \delta$,
and
${\bf b}_t\geq{\hat{\bf b}}_{t_0}$.
Combining facts, and since ${\bf V}({\bf b}_t-{\hat{\bf b}}_{t_0})\in\resc{f}$,
these imply, for all $t$ sufficiently large, that
\begin{eqnarray*}
f({\bf x}_t)
&=&
f({\bf V} {\bf b}_t + {\bf q}_t)
\\
&=&
f({\bf V} {\hat{\bf b}}_{t_0} + {\bf q}_t + {\bf V} ({\bf b}_t - {\hat{\bf b}}_{t_0}) )
\\
&\leq&
f({\bf V} {\hat{\bf b}}_{t_0} + {\bf q}_t)
\\
&<&
f({\bf V} {\hat{\bf b}}_{t_0} + {\hat{\bf q}}) + \epsilon
\\
&=&
f({\hat{\bf x}}_{t_0}) + \epsilon
\\
&<&
h({\hat{\bf q}}) + 2\epsilon.
\end{eqnarray*}
Since this holds for all $\epsilon>0$, this proves the claim
in this case.
The proof is similar when $h({\hat{\bf q}})=-\infty$.
In this case, let $\beta>0$, and
choose $t_0$ so that
$f({\hat{\bf x}}_{t_0}) < -\beta$
and also $f$ continuous at ${\hat{\bf x}}_{t_0}$
(by Claims~\ref{cl:a1} and~\ref{cl:a6}).
By continuity,
there exists $\delta>0$ so that $\norm{{\bf q}'-{\hat{\bf q}}}<\delta$ implies
\[ f({\bf V} {\hat{\bf b}}_{t_0} + {\bf q}') < f({\bf V} {\hat{\bf b}}_{t_0} + {\hat{\bf q}}) + \beta/2 < -\beta/2.\]
As before, for $t$ sufficiently large, these imply that
\[
f({\bf x}_t)
=
f({\bf V} {\bf b}_t + {\bf q}_t)
\leq
f({\bf V} {\hat{\bf b}}_{t_0} + {\bf q}_t)
<
-\beta/2.
\]
Since this holds for all $\beta>0$, this
proves the claim in this case as well.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Thus, by Claims~\ref{cl:a5} and~\ref{cl:a7},
$f({\bf x}_t)\rightarrow h({\hat{\bf q}})$ for any sequence ${\bf x}_t$ converging to
$\overline{\xx}$.
Therefore, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$, and
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=h({\hat{\bf q}})<+\infty$,
proving the theorem.
\qed
Next, we show that if $\overline{\xx}$ is in $\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
and if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$, then actually
$\overline{\xx}$ must be in the interior of $\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
\begin{theorem}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Suppose ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at some point $\overline{\xx}\in\extspac{n}$, and
that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$.
Then $\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
In other words,
\[ \contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} \subseteq \intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}. \]
\end{theorem}
\proof
Suppose, by way of contradiction, that $\overline{\xx}\not\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Let $B_1, B_2,\ldots$ be a nested countable basis
for $\overline{\xx}$
(which exists by Corollary~\ref{cor:nest-count-basis}).
Since we have assumed $\overline{\xx}\not\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
no neighborhood $B_t$ can be included in $\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Therefore, for each $t$, we can choose a point
$\overline{\xx}_t\in B_t \setminus (\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})$.
Then $\overline{\xx}_t\rightarrow\overline{\xx}$,
by Corollary~\ref{cor:nest-count-basis},
so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$ by continuity of
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at $\overline{\xx}$.
But this is a contradiction since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$, while
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t)=+\infty$ for all $t$.
\qed
Finally, we show that every point in $\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ must have the
form given in Theorem~\ref{thm:g:2}.
\begin{theorem}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then
\[ \intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} \subseteq
(\represc{f}\cap{{\cal E}_n}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f}.
\]
That is, if $\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, then
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some $\overline{\ee}\in\represc{f}\cap{{\cal E}_n}$
and some ${\bf q}\in\intdom{f}$.
\end{theorem}
\proof
Let $\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
By Corollary~\ref{cor:h:1},
we can write $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$,
where $\overline{\ee}=[{\bf v}_1,\ldots,{\bf v}_k]{\bar{\bm{\omega}}}$,
for some
${\bf v}_1,\ldots,{\bf v}_k\in{\R^n}$, $k\geq 0$,
and some ${\bf q}\in{\R^n}$.
Without loss of generality (by Proposition~\ref{pr:g:2}),
we assume ${\bf v}_1,\ldots,{\bf v}_k$ are linearly independent.
Since $\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, there exists a neighborhood $U$ of
$\overline{\xx}$ that is included in $\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
We prove the theorem in two parts:
\setcounter{claimp}{0}
\medskip
\begin{claimp}
There exists some ${\bf q}'\in\intdom{f}$ for which $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$.
\end{claimp}
\proof
We construct a sequence converging to $\overline{\xx}$
by first letting
\[
{\bf d}_t = \sum_{i=1}^{k} t^{k+1-i} {\bf v}_i,
\]
and then letting ${\bf x}_t={\bf d}_t+{\bf q}$, for $t=1,2,\ldots$.
Then ${\bf d}_t\rightarrow\overline{\ee}$ and
${\bf x}_t\rightarrow\overline{\xx}$
(provable using Theorem~\ref{thm:seq-rep}
and Proposition~\ref{pr:aff-cont}, respectively).
We claim there must exist some $t_0\geq 1$ and some $\epsilon>0$ for
which ${\cal B}({\bf x}_{t_0},\epsilon)\subseteq \dom{f}$
(where ${\cal B}(\cdot,\cdot)$ denotes an open ball in ${\R^n}$,
as in \eqref{eqn:open-ball-defn}).
Suppose this claim is false.
Then for each $t\geq 1$, there must exist a point ${\bf x}'_t$ that is in
${\cal B}({\bf x}_t,1/t)$ but not $\dom{f}$.
Consequently,
for all ${\bf u}\in{\R^n}$,
${\bf x}'_t\cdot{\bf u}$ and ${\bf x}_t\cdot{\bf u}$ must have the same limit
since
$|{\bf x}'_t\cdot{\bf u} - {\bf x}_t\cdot{\bf u}|\leq \norm{{\bf x}'_t-{\bf x}_t} \norm{{\bf u}} < \norm{{\bf u}}/t$.
Therefore, ${\bf x}'_t\rightarrow\overline{\xx}$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
Since $U$ is a neighborhood of $\overline{\xx}$, it follows that,
for $t$ sufficiently large,
${\bf x}'_t\in U\subseteq\dom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, implying
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x}'_t)=f({\bf x}'_t)<+\infty$
(by Proposition~\ref{pr:h:1}(\ref{pr:h:1a}),
since $f$ is lower semicontinuous).
But this is a contradiction since ${\bf x}'_t\not\in\dom{f}$ for all $t$.
So let $t_0\geq 1$ and $\epsilon>0$ be such that
${\cal B}({\bf x}_{t_0},\epsilon)\subseteq \dom{f}$, and let
${\bf q}'={\bf x}_{t_0} = {\bf d}_{t_0}+{\bf q}$.
Then ${\bf q}'\in\intdom{f}$.
And $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$, by Proposition~\ref{pr:g:1},
since ${\bf q}'-{\bf q}={\bf d}_{t_0}$ is in the span of ${\bf v}_1,\ldots,{\bf v}_k$.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
\medskip
\begin{claimp}
$\overline{\ee}\in\represc{f}$.
\end{claimp}
\proof
Let
$\overline{{\bf s}}_j = \limrays{{\bf v}_1,\ldots,{\bf v}_j}$
for $j=0,1,\ldots,k$; in particular, $\overline{\ee}=\overline{{\bf s}}_k$.
We prove by induction on $j=0,1,\ldots,k$ that
$\overline{{\bf s}}_j\in\represc{f}$, that is, that
$\overline{{\bf s}}_j=\limrays{{\bf w}_1,\ldots,{\bf w}_j}$
for some vectors ${\bf w}_1,\ldots,{\bf w}_j\in\resc{f}$.
When $j=k$, the claim is proved.
The base case, when $j=0$, holds vacuously.
For the inductive step, let $j\geq 1$, and assume
$\overline{{\bf s}}_{j-1}=\limrays{{\bf w}_1,\ldots,{\bf w}_{j-1}}$
for some ${\bf w}_1,\ldots,{\bf w}_{j-1}\in\resc{f}$.
Let
\[
\overline{{\bf y}} = \overline{{\bf s}}_j
= \overline{{\bf s}}_{j-1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \limray{{\bf v}_j}
= \limrays{{\bf w}_1,\ldots,{\bf w}_{j-1},{\bf v}_j},
\]
and let
$\overline{{\bf z}}=\limrays{{\bf v}_{j+1},\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$
so that $\overline{\xx}=\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$.
We construct several sequences converging to these points:
for $t=1,2,\ldots$, let
\begin{eqnarray*}
{\bf y}_t &=& \sum_{i=1}^{j-1} t^{k+1-i} {\bf w}_i + t^{k+1-j} {\bf v}_j
\\
{\bf z}_t &=& \sum_{i=j+1}^{k} t^{k+1-i} {\bf v}_i + {\bf q}
\\
{\bf x}_t &=& {\bf y}_t + {\bf z}_t.
\end{eqnarray*}
Also, let $\overline{{\bf y}}_t=\limray{{\bf y}_t}$, and
$\overline{\xx}_t=\overline{{\bf y}}_t\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}_t$.
Clearly, ${\bf x}_t\rightarrow\overline{\xx}$ and ${\bf y}_t\rightarrow\overline{{\bf y}}$
(by Theorem~\ref{thm:seq-rep}).
We claim furthermore that $\overline{\xx}_t\rightarrow\overline{\xx}$.
To see this, let ${\bf u}$ be any point in ${\R^n}$;
we aim to show that $\overline{\xx}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$.
First, if $\overline{{\bf y}}\cdot{\bf u}=+\infty$, then $\overline{\xx}\cdot{\bf u}=+\infty$.
Since ${\bf y}_t\cdot{\bf u}\rightarrow\overline{{\bf y}}\cdot{\bf u}$,
we must have ${\bf y}_t\cdot{\bf u}>0$ for $t$ sufficiently large, implying,
for all such $t$, that
$\overline{{\bf y}}_t\cdot{\bf u}=+\infty$ and so also $\overline{\xx}_t\cdot{\bf u}=+\infty$.
Thus, $\overline{\xx}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$ in this case.
The case $\overline{{\bf y}}\cdot{\bf u}=-\infty$ is handled similarly (or we can apply
the above argument to $-{\bf u}$).
The only remaining case is that $\overline{{\bf y}}\cdot{\bf u}=0$ (since $\overline{{\bf y}}$ is
an icon).
This is only possible if ${\bf v}_j=0$ and
${\bf w}_i\cdot{\bf u}=0$ for $i=1,\ldots,j-1$;
these imply, for all $t$, that ${\bf y}_t\cdot{\bf u}=0$ and so
$\overline{{\bf y}}_t\cdot{\bf u}=0$.
Thus,
\[
\overline{\xx}_t\cdot{\bf u}
=
\overline{{\bf y}}_t\cdot{\bf u} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf z}_t\cdot{\bf u}
=
{\bf z}_t\cdot{\bf u}
=
{\bf y}_t\cdot{\bf u} + {\bf z}_t\cdot{\bf u}
=
{\bf x}_t\cdot{\bf u} \rightarrow \overline{\xx}\cdot{\bf u}.
\]
We conclude $\overline{\xx}_t\rightarrow\overline{\xx}$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
Therefore,
for all sufficiently large $t$, $\overline{\xx}_t$ must be in the neighborhood
$U\subseteq\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
For the rest of the proof,
let $t$ be any such index so that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_{t})<+\infty$.
Then it must also be the case that ${\bf y}_t\in\resc{f}$, since
otherwise, we would have
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}_t)
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}}_t\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}_t)
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf y}_t}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}_t)
=
+\infty,
\]
with the last equality from Theorem~\ref{thm:i:4}.
Furthermore, by Proposition~\ref{pr:g:1},
\[
\overline{{\bf s}}_j = \overline{{\bf y}} = \limrays{{\bf w}_1,\ldots,{\bf w}_{j-1},{\bf y}_t}
\]
since, by ${\bf y}_t$'s definition,
${\bf y}_t - t^{k+1-j}{\bf v}_{j}$ is a linear combination of
${\bf w}_1,\ldots,{\bf w}_{j-1}$.
Setting ${\bf w}_j={\bf y}_t\in\resc{f}$ now completes the induction.
\leavevmode\unskip\penalty9999 \hbox{}\nobreak\hfill \quad\hbox{\ensuremath{\clqedsymbol}}
Combining the two claims proves the theorem.
\qed
Together, the last three theorems fully characterize exactly the
points where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous (and not $+\infty$):
\begin{corollary} \label{cor:cont-gen-char}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then for all $\overline{\xx}\in\extspac{n}$, the following are equivalent:
\begin{itemize}
\item
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$.
\item
$\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
\item
$\overline{\xx}=\limrays{{\bf v}_1,\ldots,{\bf v}_k}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some
${\bf q}\in\intdom{f}$ and some ${\bf v}_1,\ldots,{\bf v}_k\in\resc{f}$,
$k\geq 0$.
\end{itemize}
That is,
\[ \contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} = \intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} =
(\represc{f}\cap{{\cal E}_n}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f}
= \represc{f} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f}.
\]
\end{corollary}
\proof
All parts of this corollary were proved in the preceding
theorems, except for the final equality for which it suffices to show
\[\represc{f} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f} \subseteq (\represc{f}\cap{{\cal E}_n}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f}\]
(since the reverse inclusion is immediate).
Suppose $\overline{\xx}=\overline{{\bf y}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf z}$ where
$\overline{{\bf y}}\in\represc{f}$ and ${\bf z}\in\intdom{f}$.
Then $\overline{{\bf y}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some
$\overline{\ee}\in\represc{f}\cap{{\cal E}_n}$ and
${\bf q}\in\resc{f}$
(since $\overline{{\bf y}}$ can be represented using only vectors in $\resc{f}$).
Since ${\bf z}\in\intdom{f}$, there exists an open set $U\subseteq{\R^n}$
including ${\bf 0}$ such that ${\bf z}+U\subseteq\dom{f}$.
Since ${\bf q}\in\resc{f}$, for all ${\bf s}\in U$,
$f({\bf q}+{\bf z}+{\bf s})\leq f({\bf z}+{\bf s})<+\infty$,
so ${\bf q}+{\bf z}+U\subseteq\dom{f}$, implying ${\bf q}+{\bf z}\in\intdom{f}$.
Thus, $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax({\bf q}+{\bf z})\in (\represc{f}\cap{{\cal E}_n}) \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \intdom{f}$,
completing the proof.
\qed
{
Let $\overline{\xx}\in\extspac{n}$, and
suppose ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$.
We can write $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for some $\overline{\ee}\in{{\cal E}_n}$ and
${\bf q}\in{\R^n}$; furthermore, $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ by
Corollary~\ref{cor:a:4}.
Corollary~\ref{cor:cont-gen-char} makes explicit the precise
conditions under which ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is or is not continuous at
$\overline{\xx}$, namely
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$
if $\overline{\ee}\in\represc{f}$ and also ${\bf q}$ can be chosen
to be in $\intdom{f}$.
Otherwise, if
$\overline{\ee}\not\in\represc{f}$ or if there is no way of choosing ${\bf q}$ so
that $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ still holds and also ${\bf q}\in\intdom{f}$,
then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is discontinuous at $\overline{\xx}$.
These latter two conditions for there being a discontinuity at a point
$\overline{\xx}$ can be seen in the two examples given at the beginning of this
section.
In particular, we saw that the function in \eqref{eqn:recip-fcn-eg} is
discontinuous at the point $\overline{\xx}=\limray{{\bf e}_1}$.
We mentioned earlier that $\resc{f}=\mathbb{R}_{\geq 0}^2$ for this function, so
$\limray{{\bf e}_1}\in\represc{f}$, which means the first condition for
continuity is satisfied.
However, we can only write $\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ if
${\bf q}=\beta{\bf e}_1$ for some $\beta\in{\mathbb{R}}$.
Since no such point is in the effective domain of $f$, let alone its
interior, a discontinuity results at $\overline{\xx}$.
For the other example function in \eqref{eqn:curve-discont-finiteev-eg},
we saw that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at the point
$\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf 0}$ where
$\overline{\ee}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$.
In this case,
the function $f$ is finite everywhere, so all points in ${\mathbb{R}}^2$,
including the origin, are in the
interior of $\dom{f}={\mathbb{R}}^2$, thereby satisfying the second condition for
continuity.
However, for this function, it can be checked that the standard
recession cone is equal to
$\resc{f}=\{ \beta {\bf e}_2 : \beta \geq 0 \}$,
which implies that the only icons in $\represc{f}$ are
${\bf 0}$ and $\limray{{\bf e}_2}$.
In particular, this means
$\overline{\ee}\not\in\represc{f}$, yielding the
discontinuity at $\overline{\xx}$.
}
\subsection{Conditions for continuity}
\label{subsec:cond-for-cont}
We next explore general conditions for continuity, especially for
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ to be continuous everywhere, that is, at all points in
$\extspac{n}$.
We begin with the more direct implications of the characterization
given in Corollary~\ref{cor:cont-gen-char}.
As noted above, if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$, where $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$,
$\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$, then $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
(by Corollary~\ref{cor:a:4}).
And if $\overline{\ee}\not\in\represc{f}$, then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ cannot be continuous at
$\overline{\xx}$
(by Corollary~\ref{cor:cont-gen-char}).
Thus, for ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ to be continuous everywhere, it is necessary that
$({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}\subseteq\represc{f}$.
Actually, this latter condition is equivalent to
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ being equal to $\represc{f}$, as we show in the next
proposition.
When this simpler condition holds, that ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$, we
say that $f$ is recessive complete, as was
discussed previously in Section~\ref{subsec:rank-one-minimizers}
where the condition was shown to imply the existence of universal
minimizers with astral rank one.
\begin{proposition} \label{pr:r-equals-c-equiv-forms}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then the following are equivalent:
\begin{enumerate}
\item \label{pr:r-equals-c-equiv-forms:a}
$({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \cap {{\cal E}_n} = \represc{f} \cap {{\cal E}_n}$.
\item \label{pr:r-equals-c-equiv-forms:b}
$({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \cap {{\cal E}_n} \subseteq \represc{f}$.
\item \label{pr:r-equals-c-equiv-forms:c}
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}} = \represc{f}$.
(That is, $f$ is recessive complete.)
\end{enumerate}
\end{proposition}
\proof
That
(\ref{pr:r-equals-c-equiv-forms:a}) $\Rightarrow$
(\ref{pr:r-equals-c-equiv-forms:b}),
and
(\ref{pr:r-equals-c-equiv-forms:c}) $\Rightarrow$
(\ref{pr:r-equals-c-equiv-forms:a})
are both immediate.
To see
(\ref{pr:r-equals-c-equiv-forms:b}) $\Rightarrow$
(\ref{pr:r-equals-c-equiv-forms:c}),
suppose
$({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \cap {{\cal E}_n} \subseteq \represc{f}$.
By Proposition~\ref{pr:repres-in-arescone},
$\represc{f}\subseteq{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
On the other hand,
\[
{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}
=
\conv(({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \cap {{\cal E}_n})
\subseteq
\represc{f}.
\]
The equality follows from
Theorem~\ref{thm:cl-conv-cone-convhull-corezeros}
since ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ is a closed, convex cone
(Corollary~\ref{cor:thm:j:3}).
And the inclusion is by our assumption and since
$\represc{f}$ is convex
by Proposition~\ref{pr:repres-in-arescone}
(and also
Proposition~\ref{pr:conhull-prop}(\ref{pr:conhull-prop:aa})).
\qed
Expanding on the discussion above, we prove several direct
consequences of the characterization given in
Corollary~\ref{cor:cont-gen-char}.
First,
if $\overline{\ee}$ is an icon in
$({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}$ but not in $\represc{f}$, and if $\overline{\xx}$ is
any point in $\extspac{n}$, then the point $\overline{{\bf z}}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$
cannot be in $\contset{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$;
in other words, it is not possible both that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf z}})<+\infty$ and
that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{{\bf z}}$:
\begin{theorem} \label{thm:d5}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Suppose $\overline{\ee}\in (({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}})\cap{{\cal E}_n}) \setminus\represc{f}$.
Then for all $\overline{\xx}\in\extspac{n}$, either
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$
or
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$.
\end{theorem}
\proof
Let $\overline{\xx}\in\extspac{n}$.
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})=+\infty$ then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$ since
$\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
So assume ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx})<+\infty$ and suppose, by way of
contradiction, that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}$.
Then by Corollary~\ref{cor:cont-gen-char}, there exists
$\overline{\dd}\in\represc{f}\cap{{\cal E}_n}$ and ${\bf q}\in\intdom{f}$ such that
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}=\overline{\dd}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$.
This implies
$\overline{\dd}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{\xx}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax (-{\bf q})$,
and so $\overline{\ee}\in\lb{{\bf 0}}{\overline{\dd}}$ by
Corollary~\ref{cor:d-in-lb-0-dplusx}.
Since both ${\bf 0}$ and $\overline{\dd}$ are in $\represc{f}$, and since
$\represc{f}$ is convex
(by Proposition~\ref{pr:repres-in-arescone}),
it follows that $\overline{\ee}\in\represc{f}$, contradicting our initial
assumption.
\qed
We previously remarked that if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere then
$f$ is recessive complete.
Actually, we can make a somewhat stronger statement, namely, that if
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous \emph{at all of its minimizers}, then
$f$ is recessive complete.
Clearly, this implies the former assertion.
\begin{theorem} \label{thm:cont-at-min-implies-ares}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at all of its minimizers,
then
$f$ is recessive complete.
\end{theorem}
\proof
We assume $f\not\equiv +\infty$ since otherwise
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}=\extspac{n}$.
We prove the theorem in the contrapositive.
Suppose
$f$ is not recessive complete.
Then by Proposition~\ref{pr:r-equals-c-equiv-forms},
there exists a point
$\overline{\ee} \in (({\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}) \cap {{\cal E}_n}) \setminus \represc{f}$.
Let $\overline{{\bf y}}\in\extspac{n}$ be any point that minimizes
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, implying ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}})<+\infty$.
Furthermore, since $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}})\leq{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{{\bf y}})$, so
$\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$ must also minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$.
It now follows immediately from
Theorem~\ref{thm:d5}
that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at $\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf y}}$, which is one of
its minimizers.
\qed
So
recessive completeness
is a necessary condition for ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ to
be continuous everywhere, or even for it to be continuous at its minimizers.
When $f$ is convex and finite everywhere,
these conditions all turn out to be
equivalent.
In other words, in this case, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere if and
only if
$f$ is recessive complete.
Furthermore, and quite remarkably, if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at all its
minimizers, then it must actually be continuous everywhere.
Equivalently, if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is discontinuous anywhere, then it must be
discontinuous at one or more of its minimizers
(as was the case for the example in
\eqref{eqn:curve-discont-finiteev-eg}, which is finite everywhere).
\begin{theorem} \label{thm:cont-conds-finiteev}
Let $f:{\R^n}\rightarrow {\mathbb{R}}$ be convex.
Then the following are equivalent:
\begin{enumerate}
\item \label{thm:cont-conds-finiteev:a}
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere.
\item \label{thm:cont-conds-finiteev:b}
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at all its minimizers.
\item \label{thm:cont-conds-finiteev:c}
$f$ is recessive complete.
\end{enumerate}
\end{theorem}
\proof
Since $f$ is convex and finite everywhere, it is also continuous
everywhere
\citep[Corollary~10.1.1]{ROC}.
That
(\ref{thm:cont-conds-finiteev:a})
$\Rightarrow$
(\ref{thm:cont-conds-finiteev:b})
is immediate.
That
(\ref{thm:cont-conds-finiteev:b})
$\Rightarrow$
(\ref{thm:cont-conds-finiteev:c})
follows immediately from Theorem~\ref{thm:cont-at-min-implies-ares}.
To see
(\ref{thm:cont-conds-finiteev:c})
$\Rightarrow$
(\ref{thm:cont-conds-finiteev:a}),
suppose ${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$.
Let $\overline{\xx}\in\extspac{n}$, which we can write $\overline{\xx}=\overline{\ee}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ for
some $\overline{\ee}\in{{\cal E}_n}$ and ${\bf q}\in{\R^n}$.
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$, then $f$ converges to $+\infty$ on every
sequence converging to $\overline{\xx}$, so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$.
Otherwise, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$ so $\overline{\ee}\in{\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$
by Corollary~\ref{cor:a:4}.
Since $f$ is finite everywhere, $\dom{f}={\R^n}$ so
${\bf q}\in\intdom{f}$.
Therefore, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$ by
Corollary~\ref{cor:cont-gen-char}.
\qed
{
If $f$ is convex but not finite everywhere,
then it is possible that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is
continuous at all its minimizers, but not continuous elsewhere.
This is possible even if the function $f$ itself is continuous
everywhere.
For example, consider the following variation on the function given in
\eqref{eqn:recip-fcn-eg}:
\begin{equation} \label{eqn:recip-fcn-eg-mod}
f({\bf x}) = f(x_1, x_2)
= \left\{ \begin{array}{cl}
\displaystyle{\frac{1}{x_1 x_2}}
+ e^{- x_1}
+ e^{- x_2}
& \mbox{if $x_1>0$ and $x_2>0$}
\\[1.5em]
+\infty & \mbox{otherwise.} \\
\end{array} \right.
\end{equation}
It can be checked that
$f$ converges to zero, and is thereby minimized, on just
those sequences $\seq{{\bf x}_t}$ in ${\mathbb{R}}^2$ for which
${\bf x}_t\cdot{\bf e}_1\rightarrow+\infty$
and
${\bf x}_t\cdot{\bf e}_2\rightarrow+\infty$.
Thus, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is minimized, with ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=0$,
exactly at those points
$\overline{\xx}\in\extspac{2}$ for which
$\overline{\xx}\cdot{\bf e}_1=\overline{\xx}\cdot{\bf e}_2=+\infty$.
Moreover, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at all such points.
On the other hand, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at $\limray{{\bf e}_1}$ by
similar arguments to those given
following \eqref{eqn:recip-fcn-eg},
but this point is not a
minimizer since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\limray{{\bf e}_1})=1$.
}
\subsection{Dual characterization of continuity}
\label{subsec:charact:cont:duality}
As we show next,
the recessive completeness of
a convex function $f:{\R^n}\rightarrow\overline{\R}$,
which is so central to the continuity
of its extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, can be further characterized
in terms of very
simple and general dual properties,
specifically regarding its barrier cone,
$\slopes{f}$
(introduced in Section~\ref{sec:slopes}).
Specifically,
we will see that $f$ is recessive complete
if and only if $\slopes{f}$ is polyhedral, that is, if and
only if it is finitely generated.
Thus, this geometric property of the barrier cone $\slopes{f}$ entirely
determines recessive completeness, and thereby, at least in some cases
(such as when $f$ is finite everywhere), entirely determines if its
extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere.
Indeed, this characterization follows directly from what was proved
more generally for convex cones in Section~\ref{sec:cones}, applied
here to the standard and astral recession cones, which we
have seen can each be expressed respectively as the standard and astral
polars of the barrier cone $\slopes{f}$.
As a first such application,
we can characterize exactly when
$\represc{f}$ is closed and when
$\rescbar{f}={\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$:
\begin{theorem} \label{thm:dual-char-specs}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then:
\begin{enumerate}
\item \label{thm:dual-char-specs:a}
$\represc{f}=\rescbar{f}$ if and only if $\resc{f}$ is polyhedral.
\item \label{thm:dual-char-specs:b}
$\rescbar{f}={\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ if and only if $\slopes{f}$ is closed
in ${\R^n}$.
\end{enumerate}
\end{theorem}
\proof
~
Part~(\ref{thm:dual-char-specs:a}) is immediate from
Theorem~\ref{thm:repcl-polyhedral-cone} applied to $\resc{f}$.
Part~(\ref{thm:dual-char-specs:b}) follows from
Theorem~\ref{thm:k-closed-equals-closed-apol} applied to $\slopes{f}$
(which is a pointed convex cone,
by Proposition~\ref{pr:barf-is-cone}),
using
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\apol{(\slopes{f})}$ and
$\resc{f}=\polar{(\slopes{f})}$
by
Theorem~\ref{th:ares-is-apolslopes}
and
Corollary~\ref{cor:rescpol-is-slopes}.
\qed
Combining the two parts of this theorem, we can now prove our
characterization of recessive completeness:
\begin{theorem} \label{thm:dual-cond-char}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous.
Then the following are equivalent:
\begin{enumerate}
\item \label{thm:dual-cond-char:a}
$f$ is recessive complete.
\item \label{thm:dual-cond-char:b}
$\resc{f}$ is polyhedral and also $\slopes{f}$ is closed
in ${\R^n}$.
\item \label{thm:dual-cond-char:c}
$\slopes{f}$ is polyhedral.
\end{enumerate}
\end{theorem}
\proof
~
\underline{(\ref{thm:dual-cond-char:a})
$\Leftrightarrow$
(\ref{thm:dual-cond-char:b})}:
By Proposition~\ref{pr:repres-in-arescone},
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$ if and only if
$\represc{f}=\rescbar{f}$ and
$\rescbar{f}={\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Therefore,
Theorem~\ref{thm:dual-char-specs} immediately implies that
${\arescone {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}=\represc{f}$ if and only if
the conditions in (\ref{thm:dual-cond-char:b}) hold.
\underline{(\ref{thm:dual-cond-char:b})
$\Rightarrow$
(\ref{thm:dual-cond-char:c})}:
Suppose $\resc{f}$ is polyhedral and that $\slopes{f}$ is
closed in ${\R^n}$.
Then $\rescpol{f}$, the polar of $\resc{f}$, is also polyhedral
\citep[Corollary~19.2.2]{ROC},
and furthermore,
$
\rescpol{f} = \cl(\slopes{f}) = \slopes{f}
$
by Corollary~\ref{cor:rescpol-is-slopes},
and since $\slopes{f}$ is closed in ${\R^n}$.
\underline{(\ref{thm:dual-cond-char:c})
$\Rightarrow$
(\ref{thm:dual-cond-char:b})}:
Suppose $\slopes{f}$ is polyhedral.
Then it is closed in ${\R^n}$
\citep[Theorem~19.1]{ROC},
and its polar,
$\polar{(\slopes{f})}=\resc{f}$,
must also be polyhedral
(again using Corollary~\ref{cor:rescpol-is-slopes}
and \citet[Corollary~19.2.2]{ROC}).
\qed
The condition that $\resc{f}$ is polyhedral
(which implies that its polar
$\rescpol{f}=\cl(\slopes{f})$ must also be polyhedral)
is not
in itself sufficient for $f$ to be recessive complete.
For instance, as we have already seen, the recession cone $\resc{f}$ of
the function $f$ in
\eqref{eqn:curve-discont-finiteev-eg} is exactly the cone generated by
the singleton $\{{\bf e}_2\}$, and thus is polyhedral.
Nevertheless, we
have seen that this function is not continuous everywhere, implying
$f$ is not recessive complete
in this case, by
Theorem~\ref{thm:cont-conds-finiteev}.
Indeed, in more detail, it can be calculated that the effective domain
of ${f^*}$ is
\[
\dom{{f^*}} =
\braces{(u_1,u_2) \in {\mathbb{R}}^2 :
-1 \leq u_2 \leq -\frac{u_1^2}{4} }
\]
(and actually ${f^*}$ is the indicator function for this set).
It can then be further calculated that
$\slopes{f}$ (which is the same as
$\cone(\dom{{f^*}})$ in this case, by
Corollary~\ref{cor:ent-clos-is-slopes-cone}
since this function is finite everywhere and therefore has all
reductions closed by Proposition~\ref{pr:j:1}(\ref{pr:j:1d}))
is equal to the origin adjoined to the open lower
half-plane:
\[
\slopes{f} =
\cone(\dom{{f^*}}) = \{(0,0)\}
\cup \{ (u_1,u_2)\in{\mathbb{R}}^2 : u_2 < 0 \}.
\]
Consistent with Theorem~\ref{thm:dual-cond-char}, this set is not
polyhedral or closed, although its closure is polyhedral.
Theorem~\ref{thm:dual-cond-char} is not true in
general if in its statement we replace $\slopes{f}$ by
$\cone(\dom{{f^*}})$.
For example, for the function $f$ in \eqref{eqn:ex:1}, it can be
checked that $\resc{f}=\mathbb{R}_{\geq 0}^2$ and that
$f$ is recessive complete.
Nevertheless, as was seen earlier,
$\cone(\dom{{f^*}})$ for this function is the set given in
\eqref{eqn:ex:1:cone-dom-fstar},
which is not closed in ${\mathbb{R}}^2$.
(However, its closure is polyhedral, and furthermore,
$\slopes{f}=-\mathbb{R}_{\geq 0}^2$, which is polyhedral, consistent
with Theorem~\ref{thm:dual-cond-char}.)
So $\cone(\dom{{f^*}})$ being polyhedral is not necessary for
$f$ to be recessive complete.
On the other hand,
if $f$ is convex, closed and proper, then
$\cone(\dom{{f^*}})$ being polyhedral is always sufficient for
$f$ to be recessive complete.
This is because if $\cone(\dom{{f^*}})$ is polyhedral, then it is also closed, so
\[
\rescpol{f}
=
\cl(\cone(\dom{{f^*}}))
=
\cone(\dom{{f^*}})
\subseteq
\slopes{f}
\subseteq
\rescpol{f},
\]
where the first equality is
by Theorem~\ref{thm:rescpol-is-con-dom-fstar},
and the two inclusions are by
Proposition~\ref{pr:slopes-equiv}
and
Corollary~\ref{cor:rescpol-is-slopes},
respectively.
Thus,
$\slopes{f}$, being equal to
$\cone(\dom{{f^*}})$,
is also polyhedral, implying
$f$ is recessive complete
by
Theorem~\ref{thm:dual-cond-char}.
When $f$ has all reductions closed (including, for instance, when
$\inf f > -\infty$),
Corollary~\ref{cor:ent-clos-is-slopes-cone}
implies that $\slopes{f}=\cone(\dom{{f^*}})$, which means, in this
case, we can immediately replace $\slopes{f}$ by
$\cone(\dom{{f^*}})$
in Theorem~\ref{thm:dual-cond-char}
to obtain the following corollary:
\begin{corollary} \label{cor:thm:dual-cond-char:ent-closed}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and lower semicontinuous,
with all reductions closed.
Then the following are equivalent:
\begin{enumerate}
\item
$f$ is recessive complete.
\item
$\resc{f}$ is polyhedral and also $\cone(\dom{{f^*}})$ is closed
in ${\R^n}$.
\item
$\cone(\dom{{f^*}})$ is polyhedral.
\end{enumerate}
\end{corollary}
If $f$ is finite everywhere then it has all reductions closed
(Proposition~\ref{pr:j:1}(\ref{pr:j:1d})).
In this case, we can immediately combine
Corollary~\ref{cor:thm:dual-cond-char:ent-closed} with
Theorem~\ref{thm:cont-conds-finiteev} to obtain an expanded list of
conditions that are necessary and sufficient for
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ to be continuous everywhere:
\begin{corollary} \label{cor:thm:cont-conds-finiteev}
Let $f:{\R^n}\rightarrow {\mathbb{R}}$ be convex.
Then the following are equivalent:
\begin{enumerate}
\item \label{cor:thm:cont-conds-finiteev:a}
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere.
\item \label{cor:thm:cont-conds-finiteev:b}
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at all its minimizers.
\item \label{cor:thm:cont-conds-finiteev:c}
$f$ is recessive complete.
\item \label{cor:thm:cont-conds-finiteev:d}
$\resc{f}$ is polyhedral and also $\cone(\dom{{f^*}})$ is closed
in ${\R^n}$.
\item \label{cor:thm:cont-conds-finiteev:e}
$\cone(\dom{{f^*}})$ is polyhedral.
\end{enumerate}
\end{corollary}
For example, suppose $f$ is a function of the form given in
\eqref{eqn:loss-sum-form}
(under the same assumptions as in Proposition~\ref{pr:hard-core:1} and
throughout Section~\ref{sec:emp-loss-min}).
Then
Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:b}) shows that
$\resc{f}$ is evidently the polar of the cone generated by
$\{{\bf u}_1,\ldots,{\bf u}_m\}$.
Since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere
(Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:a})),
Corollary~\ref{cor:thm:cont-conds-finiteev} implies that
$\cone(\dom{{f^*}})$ is closed.
Thus,
\[
\slopes{f}
=
\cone(\dom{{f^*}})
=
\cl(\cone(\dom{{f^*}}))
=
\rescpol{f}
=
\cone \{{\bf u}_1,\ldots,{\bf u}_m\},
\]
which is finitely generated and therefore polyhedral.
(The form of $\cone(\dom{{f^*}})$ could also be derived more
directly using standard tools for calculating conjugates.)
\section{Subdifferentials and descent methods}
\label{sec:gradients}
We next develop a theory of subdifferentials for functions defined
over astral space, including connections to conjugacy and
optimality, as well as how they aid in analyzing gradient descent.
\subsection{Astral subgradients}
\label{sec:gradients-def}
For a convex function $f:{\R^n}\rightarrow\overline{\R}$, recall that a vector
${\bf u}\in{\R^n}$ is said to be a \emph{subgradient} of $f$ at ${\bf x}\in{\R^n}$
if
\begin{equation} \label{eqn:standard-subgrad-ineq}
f({\bf x}') \geq f({\bf x}) + ({\bf x}'-{\bf x})\cdot{\bf u}
\end{equation}
for all ${\bf x}'\in{\R^n}$, so that the affine function (in ${\bf x}'$) on the
right-hand side of this inequality is supporting $f$ at ${\bf x}$.
The \emph{subdifferential} of $f$ at ${\bf x}$, denoted
$\partial f({\bf x})$, is the set of all subgradients of $f$ at ${\bf x}$.
(Although these definitions are intended for convex functions, we also
apply them in what follows to general functions
$f:{\R^n}\rightarrow\overline{\R}$ that are not necessarily convex.)
If $f$ is convex, then its gradient, $\nabla f({\bf x})$,
if it exists, is always also a subgradient.
Subgradients are central to optimization since
${\bf 0}\in\partial f({\bf x})$ if and only if $f$ is
minimized at ${\bf x}$.
We show in this section how this important concept can be extended to
astral space in a way that captures subgradients ``at infinity,''
as well as tangent slopes that may be infinite.
For example, suppose, for $x\in{\mathbb{R}}$, that
\begin{equation} \label{eq:ex:x:2x}
f(x)
= \left\{ \begin{array}{cl}
\max\{x,2x\} & \mbox{if $x\leq 1$} \\
+\infty & \mbox{otherwise.} \\
\end{array} \right.
\end{equation}
Then the subdifferential of $f$ at $0$ is equal to the set $[1,2]$.
This is because, for $u\in [1,2]$, $f(x)\geq x u$
for all $x\in{\mathbb{R}}$, with equality at $x=0$.
Equivalently, in the plane ${\mathbb{R}}^2$, the epigraph of $f$ is entirely above
the line
$\{ \rpair{x}{y} : y= x u \}$,
which includes $\rpair{0}{f(0)}=\rpair{0}{0}$.
By similar reasoning,
\[
\partial f(x)
= \left\{ \begin{array}{cl}
\{1\} & \mbox{if $x < 0$} \\
{[1,2]} & \mbox{if $x = 0$} \\
\{2\} & \mbox{if $0 < x < 1$} \\
{[2,+\infty)} & \mbox{if $x =1$} \\
\emptyset & \mbox{if $x>1$.} \\
\end{array} \right.
\]
Note how the subgradients ``wrap around'' the graph of this function
at $x=1$.
The standard subdifferential is not defined at $-\infty$, but if it
were, we might reasonably expect that $1$ should be included as a
subgradient at this point since $f(x)\geq x$ for all $x\in{\mathbb{R}}$, with
equality holding asymptotically in the limit as $x\rightarrow-\infty$.
We might also expect $0$ to be a subgradient at $-\infty$ since $f$ is
minimized at this point.
We will see soon how both these intuitions are captured by the
definition we propose below.
In extending subdifferentials to astral space, we will need
two different definitions: the first maps astral
points in $\extspac{n}$ to subgradients which are real vectors in ${\R^n}$;
the other, which is a kind of dual, maps in the reverse
direction from ${\R^n}$ to subsets of $\extspac{n}$.
This asymmetry, previously encountered in the two forms of conjugate
developed in Section~\ref{sec:conjugacy},
is a consequence of the asymmetry of the two spaces we
are working over, with the key dot product operation
$\overline{\xx}\cdot{\bf u}$ being defined over $\overline{\xx}\in\extspac{n}$ but ${\bf u}\in{\R^n}$.
We begin by defining an astral subgradient on astral points in
$\extspac{n}$.
When considering
\eqref{eqn:standard-subgrad-ineq}
with standard convex functions,
only $f({\bf x}')$ and $f({\bf x})$ can be infinite, so there is no possibility
of adding $-\infty$ and $+\infty$ in this expression.
However, when extending to astral space,
other quantities,
particularly those involving inner products, may become infinite.
Furthermore, there is no operation for directly adding or subtracting
astral points analogous to the difference of vectors, ${\bf x}'-{\bf x}$, that
appears in
\eqref{eqn:standard-subgrad-ineq}.
As a result, it is not immediately clear how to generalize
the definition given in
\eqref{eqn:standard-subgrad-ineq} simply by
replacing each variable and function by its
astral counterpart.
Rather, we take an aproach that focuses on the function's epigraph.
To simplify this discussion, let us suppose momentarily that
$f({\bf x})\in{\mathbb{R}}$.
In this case,
as was seen in the example above,
the condition given in
\eqref{eqn:standard-subgrad-ineq}
means that,
for some $\beta\in{\mathbb{R}}$,
$f({\bf x}')\geq {\bf x}'\cdot{\bf u} - \beta$
for all ${\bf x}'\in{\R^n}$, with equality at ${\bf x}$
(so that $f({\bf x}) = {\bf x}\cdot{\bf u} - \beta$).
Equivalently,
for every point $\rpair{{\bf x}'}{y'}$ in
$f$'s epigraph,
$y'\geq {\bf x}'\cdot{\bf u} - \beta$,
so that
${\bf x}'\cdot{\bf u} - y'\leq \beta$.
Furthermore, equality holds at the point
$\rpair{{\bf x}}{f({\bf x})}$,
so that
$ {\bf x}\cdot{\bf u} - f({\bf x}) = \beta$.
Thus, when $f({\bf x})\in{\mathbb{R}}$,
${\bf u}\in\partial f({\bf x})$ if there exists $\beta\in{\mathbb{R}}$ such
that
\begin{enumerate}
\item
${\bf x}'\cdot{\bf u} - y' \leq \beta$
for all $\rpair{{\bf x}'}{y'}\in \epi f$; and
\item
${\bf x}\cdot{\bf u} - f({\bf x}) = \beta$.
\end{enumerate}
Restated in these terms, we can more readily extend subgradients
to astral space.
Let $F:\extspac{n}\rightarrow\overline{\R}$.
We aim to define what it means for some vector ${\bf u}\in{\R^n}$ to be a
subgradient of $F$ at some point $\overline{\xx}\in\extspac{n}$.
We begin by
requiring that there exist some $\beta\in\overline{\R}$ such that
$\overline{\xx}'\cdot{\bf u} - y' \leq \beta$
for all $\rpair{\overline{\xx}'}{y'}$ in $F$'s epigraph.
This is exactly the same as in the first condition above for standard
subgradients, except that we now allow $\beta$ to be $\pm\infty$.
To generalize the second condition,
it would be tempting to simply require
$\overline{\xx}\cdot{\bf u} - F(\overline{\xx}) = \beta$.
However, such an expression is problematic since it might result in
the undefined sum of $-\infty$ and $+\infty$.
Instead, we require that the second condition above hold in the limit for
some \emph{sequence} of pairs $\seq{\rpair{\overline{\xx}_t}{y_t}}$ in
$\epi F$; thus, we require that there exist such a sequence
which converges to the pair given by $\overline{\xx}$ and
its function value, $F(\overline{\xx})$ (so that
$\overline{\xx}_t\rightarrow\overline{\xx}$ and $y_t\rightarrow F(\overline{\xx})$),
and for which
$\overline{\xx}_t\cdot{\bf u} - y_t \rightarrow \beta$.
Thus, we say that
${\bf u}\in{\R^n}$ is an \emph{astral subgradient} of
$F:\extspac{n}\rightarrow\overline{\R}$ at
$\overline{\xx}\in\extspac{n}$ if there exists $\beta\in\overline{\R}$ such that:
\begin{enumerate}
\item \label{en:ast-sub-defn-cond-1}
$\overline{\xx}'\cdot{\bf u} - y' \leq \beta$
for all $\rpair{\overline{\xx}'}{y'} \in \epi F$; and
\item \label{en:ast-sub-defn-cond-2}
there exists a sequence $\seq{\rpair{\overline{\xx}_t}{y_t}}$ in
$\epi F$ such that
$\overline{\xx}_t\rightarrow \overline{\xx}$,
$y_t\rightarrow F(\overline{\xx})$,
and
$\overline{\xx}_t\cdot{\bf u} - y_t \rightarrow \beta$.
\end{enumerate}
We then define the \emph{astral subdifferential} of $F$ at
$\overline{\xx}\in\extspac{n}$, denoted $\asubdifF{\overline{\xx}}$,
to be the set of all such astral subgradients of $F$ at $\overline{\xx}$.
When $\beta\in{\mathbb{R}}$, condition~(\ref{en:ast-sub-defn-cond-1}) means that
$F(\overline{\xx}')\geq \overline{\xx}'\cdot{\bf u} - \beta$
for all $\overline{\xx}'\in\extspac{n}$, so that $F$ majorizes the given affine
function.
In particular, this implies
$y_t\geq F(\overline{\xx}_t)\geq \overline{\xx}_t\cdot{\bf u} - \beta$
for all $t$, while
condition~(\ref{en:ast-sub-defn-cond-2})
requires that
$y_t - (\overline{\xx}_t\cdot{\bf u} - \beta)\rightarrow 0$ so that, in this sense,
$F$ is asymptotically approaching this same affine function
as $\overline{\xx}_t\rightarrow\overline{\xx}$ and $y_t\rightarrow F(\overline{\xx})$.
Astral subgradients are meant to provide meaningful subgradients for
points at infinity.
For instance, let $f$ be the function in \eqref{eq:ex:x:2x} with
extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ (which is the same as $f$ with
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(-\infty)=-\infty$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(+\infty)=+\infty$).
At $\bar{x}=-\infty$,
we can see that $u=1$ is a subgradient of $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ according to the
definition above with $\beta=0$ and as witnessed by
the sequence $\seq{\rpair{x_t}{y_t}}$ with $x_t=y_t=-t$.
This same sequence also shows that each $u<1$ is a subgradient at
$-\infty$, but now with $\beta=+\infty$.
It can be checked that there are no other subgradients at $-\infty$.
Thus, $\partial {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(-\infty)=(-\infty,1]$, so that the astral
subgradients are seen to ``wrap around'' at $-\infty$, similar to the
behavior observed for standard subgradients at $x=0$ and $x=1$.
Note in particular that $0$ is a subgradient at $-\infty$,
consistent with ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ attaining its minimum at this point.
(It can also be checked that $\partial {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(+\infty)=\emptyset$.)
As another example, suppose
\begin{equation} \label{eq:ex:ln1plusexp}
f(x) = \ln(1+e^x)
\end{equation}
for $x\in{\mathbb{R}}$, and let ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ be the extension of $f$.
The (standard) subgradients of this function at points $x\in{\mathbb{R}}$ are
simply given by its derivative $f'$.
Consistent with this derivative tending to $1$ as
$x\rightarrow+\infty$, it can be checked that $u=1$ is a subgradient
of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at $\bar{x}=+\infty$ (with $\beta=0$ and using the sequence
$\seq{\rpair{x_t}{f(x_t)}}$ where $x_t=t$).
Indeed,
\[
\partial {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\bar{x})
= \left\{ \begin{array}{cl}
(-\infty,0] & \mbox{if $\bar{x} = -\infty$} \\
\{f'(\bar{x})\} & \mbox{if $\bar{x}\in{\mathbb{R}}$} \\
{[1,+\infty)} & \mbox{if $\bar{x} = +\infty$.}
\end{array} \right.
\]
In Section~\ref{sec:conv:fct},
we proved that the set of pairs
$\extspac{n}\times{\mathbb{R}}$ is homeomorphic to the set
${{\cal M}_n}\subseteq\extspac{n+1}$ given in \eqref{eq:finclset-defn},
with each point $\rpair{\overline{\xx}}{y}\in\extspac{n}\times{\mathbb{R}}$
mapped to $\mu(\overline{\xx},y)\in{{\cal M}_n}$
according to a unique function $\mu$, as specified
in Theorem~\ref{thm:homf}.
As a result, we can very much think of
$\extspac{n}\times{\mathbb{R}}$ as a subset of $\extspac{n+1}$.
To simplify notation, we therefore henceforth identify each point
$\rpair{\overline{\xx}}{y}$ with its homeomorphic image $\mu(\overline{\xx},y)$ so that,
when clear from context,
$\rpair{\overline{\xx}}{y}$ may denote either the given pair in
$\extspac{n}\times{\mathbb{R}}$ or the point
$\mu(\overline{\xx},y)\in{{\cal M}_n}\subseteq\extspac{n+1}$.
Note importantly that this convention only applies when $y$ is finite
(in ${\mathbb{R}}$, not $\overline{\R}$).
We also apply this simplification to subsets
$S\subseteq\extspac{n}\times{\mathbb{R}}$,
such as $\epi F$, writing simply $S$, when clear from context, to denote
its homeomorphic image, $\mu(S)\subseteq{{\cal M}_n}$.
(For instance, we can now write
$\extspac{n}\times{\mathbb{R}}\subseteq\extspac{n+1}$, which formally means that
$\mu(\extspac{n}\times{\mathbb{R}})\subseteq\extspac{n+1}$.)
As in Section~\ref{sec:conv:fct}, we define ${\bf P}$ to be the
$n\times (n+1)$ matrix given in
\eqref{eqn:homat-def}.
This matrix, along with
the basis vector $\rpair{{\bf 0}}{1}$, will be used frequently to
recover $\overline{\xx}$ and $y$ from a pair $\rpair{\overline{\xx}}{y}$ since,
by Theorem~\ref{thm:homf}(\ref{thm:homf:aa}),
${\bf P} \rpair{\overline{\xx}}{y} = \overline{\xx}$
and
$\rpair{\overline{\xx}}{y} \cdot \rpair{{\bf 0}}{1} = y$
for all $\overline{\xx}\in\extspac{n}$ and $y\in{\mathbb{R}}$.
As we show next,
in the definition of astral subgradient given above,
we can always take $\beta$ to be ${F^*}({\bf u})$, thereby making
condition~(\ref{en:ast-sub-defn-cond-1}) entirely redundant.
Further, the definition can be restated in terms of astral
points in $\extspac{n+1}$ rather than sequences.
In what follows, $\epi F$ is regarded as a subset of $\extspac{n+1}$, as
was discussed above, and $\epibarbar{F}$ likewise denotes its closure
in $\extspac{n+1}$.
\begin{proposition} \label{pr:equiv-ast-subdif-defn}
Let $F:\extspac{n}\rightarrow\overline{\R}$,
and let $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Then the following are equivalent:
\begin{enumerate}
\item \label{pr:equiv-ast-subdif-defn:a}
${\bf u}\in\partial F(\overline{\xx})$.
\item \label{pr:equiv-ast-subdif-defn:b}
There exists a sequence $\seq{\rpair{\overline{\xx}_t}{y_t}}$ in $\epi F$
such that
$\overline{\xx}_t\rightarrow\overline{\xx}$,
$y_t \rightarrow F(\overline{\xx})$,
and
$\overline{\xx}_t\cdot{\bf u} - y_t \rightarrow {F^*}({\bf u})$.
\item \label{pr:equiv-ast-subdif-defn:c}
There exists $\overline{{\bf z}}\in \epibarbar{F}$ such that
${\bf P} \overline{{\bf z}} = \overline{\xx}$,
$\overline{{\bf z}}\cdot \rpair{{\bf 0}}{1} = F(\overline{\xx})$,
and
$\overline{{\bf z}} \cdot \rpair{{\bf u}}{-1} = {F^*}({\bf u})$.
\item \label{pr:equiv-ast-subdif-defn:d}
$\overline{{\bf z}}'\cdot \rpair{{\bf u}}{-1}$, as a function of $\overline{{\bf z}}'$, is
maximized over $\epibarbar{F}$ by some $\overline{{\bf z}}\in\epibarbar{F}$ with
${\bf P}\overline{{\bf z}} = \overline{\xx}$
and
$\overline{{\bf z}}\cdot\rpair{{\bf 0}}{1} = F(\overline{\xx})$.
\end{enumerate}
\end{proposition}
\proof
~
\underline{
(\ref{pr:equiv-ast-subdif-defn:b})
$\Rightarrow$
(\ref{pr:equiv-ast-subdif-defn:a})}:
Suppose part~(\ref{pr:equiv-ast-subdif-defn:b}) holds.
Let $\beta={F^*}({\bf u})$.
Then $\overline{\xx}'\cdot{\bf u} - y' \leq {F^*}({\bf u})$
for all $\rpair{\overline{\xx}'}{y'}\in \epi F$
by \eqref{eq:Fstar-def}.
Thus, on the given sequence,
all conditions are satisfied for ${\bf u}$ to be an astral
subgradient at $\overline{\xx}$.
\underline{
(\ref{pr:equiv-ast-subdif-defn:c})
$\Rightarrow$
(\ref{pr:equiv-ast-subdif-defn:b})}:
Let $\overline{{\bf z}}$ be as specified in
part~(\ref{pr:equiv-ast-subdif-defn:c}).
Because $\overline{{\bf z}}\in\epibarbar{F}$, there exists a sequence
$\seq{\rpair{\overline{\xx}_t}{y_t}}$ in $\epi F$ that converges to $\overline{{\bf z}}$.
Then
\[
\overline{\xx}_t
=
{\bf P} \rpair{\overline{\xx}_t}{y_t}
\rightarrow
{\bf P} \overline{{\bf z}} = \overline{\xx}
\]
with convergence following from
Proposition~\ref{pr:aff-cont}
(and the first equality from
Theorem~\ref{thm:homf}(\ref{thm:homf:aa})).
Likewise,
\[
y_t
=
\rpair{\overline{\xx}_t}{y_t} \cdot \rpair{{\bf 0}}{1}
\rightarrow
\overline{{\bf z}} \cdot \rpair{{\bf 0}}{1} = F(\overline{\xx}),
\]
and
\[
\overline{\xx}_t \cdot {\bf u} - y_t
=
\rpair{\overline{\xx}_t}{y_t} \cdot \rpair{{\bf u}}{-1}
\rightarrow
\overline{{\bf z}} \cdot \rpair{{\bf u}}{-1} = {F^*}({\bf u})
\]
(using
Theorem~\ref{thm:i:1}(\ref{thm:i:1c})
and
Theorem~\ref{thm:homf}(\ref{thm:homf:a})).
Thus, the sequence
$\seq{\rpair{\overline{\xx}_t}{y_t}}$ satisfies
part~(\ref{pr:equiv-ast-subdif-defn:b}).
\underline{
(\ref{pr:equiv-ast-subdif-defn:d})
$\Rightarrow$
(\ref{pr:equiv-ast-subdif-defn:c})}:
Let $\overline{{\bf z}}$ be as specified in
part~(\ref{pr:equiv-ast-subdif-defn:d}).
Then
\begin{eqnarray*}
\overline{{\bf z}}\cdot \rpair{{\bf u}}{-1}
&=&
\max_{\overline{{\bf z}}'\in\epibarbar{F}} \overline{{\bf z}}'\cdot \rpair{{\bf u}}{-1}
\\
&=&
\sup_{\rpair{\overline{\xx}'}{y'}\in\epi F} \rpair{\overline{\xx}'}{y'}\cdot \rpair{{\bf u}}{-1}
\\
&=&
\sup_{\rpair{\overline{\xx}'}{y'}\in\epi F} [\overline{\xx}'\cdot{\bf u} - y']
= {F^*}({\bf u}).
\end{eqnarray*}
The four equalities hold respectively
by assumption;
by continuity of the function being maximized
(Theorem~\ref{thm:i:1}(\ref{thm:i:1c}));
by
Theorem~\ref{thm:homf}(\ref{thm:homf:a});
and
by \eqref{eq:Fstar-def}.
Thus, $\overline{{\bf z}}$ satisfies
part~(\ref{pr:equiv-ast-subdif-defn:c}).
\underline{
(\ref{pr:equiv-ast-subdif-defn:a})
$\Rightarrow$
(\ref{pr:equiv-ast-subdif-defn:d})}:
Let $\beta\in\overline{\R}$
and $\seq{\rpair{\overline{\xx}_t}{y_t}}$ in $\epi F$ be as in the definition
of astral subgradient.
Then this sequence must have a convergent subsequence; by discarding
all other elements, we can assume the entire
sequence converges to some point $\overline{{\bf z}}$, which must be in
$\epibarbar F$.
Similar to the previous arguments,
$\overline{\xx}_t = {\bf P} \rpair{\overline{\xx}_t}{y_t} \rightarrow {\bf P} \overline{{\bf z}}$,
so ${\bf P} \overline{{\bf z}} = \overline{\xx}$ since $\overline{\xx}_t\rightarrow\overline{\xx}$.
Likewise,
$y_t = \rpair{\overline{\xx}_t}{y_t} \cdot \rpair{{\bf 0}}{1} \rightarrow \overline{{\bf z}}\cdot\rpair{{\bf 0}}{1}$,
implying
$ \overline{{\bf z}}\cdot\rpair{{\bf 0}}{1} = F(\overline{\xx}) $,
and
$\overline{\xx}_t\cdot{\bf u} - y_t = \rpair{\overline{\xx}_t}{y_t} \cdot \rpair{{\bf u}}{-1} \rightarrow \overline{{\bf z}}\cdot\rpair{{\bf u}}{-1}$,
implying
$ \overline{{\bf z}}\cdot\rpair{{\bf u}}{-1} = \beta $.
We have that
\[
\rpair{\overline{\xx}'}{y'} \cdot \rpair{{\bf u}}{-1}
=
\overline{\xx}'\cdot{\bf u} - y' \leq \beta
\]
for $\rpair{\overline{\xx}'}{y'}\in\epi F$, so
$ \overline{{\bf z}}' \cdot \rpair{{\bf u}}{-1} \leq \beta $
for $\overline{{\bf z}}'\in\epibarbar{F}$.
Therefore,
$ \overline{{\bf z}}' \cdot \rpair{{\bf u}}{-1} $ is maximized when $\overline{{\bf z}}'=\overline{{\bf z}}$.
Thus, all conditions of part~(\ref{pr:equiv-ast-subdif-defn:d}) are
satisfied.
\qed
When $F$ is the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of a function $f:{\R^n}\rightarrow\overline{\R}$,
we can replace $\epibarbar{F}$ with $\epibarbar{f}$
and ${F^*}({\bf u})$ with ${f^*}({\bf u})$
in Proposition~\ref{pr:equiv-ast-subdif-defn}(\ref{pr:equiv-ast-subdif-defn:b},\ref{pr:equiv-ast-subdif-defn:c},\ref{pr:equiv-ast-subdif-defn:d}).
This is because
$\clepi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}=\epibarbar{f}$
by Proposition~\ref{pr:wasthm:e:3}(\ref{pr:wasthm:e:3c}),
and ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}={f^*}$
by Proposition~\ref{pr:fextstar-is-fstar}.
Thus, in particular, ${\bf u}\in\asubdiffext{\overline{\xx}}$
if and only if there exists a sequence
$\seq{\rpair{{\bf x}_t}{y_t}}$ in $\epi f$
with ${\bf x}_t\rightarrow\overline{\xx}$, $y_t \rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$,
and ${\bf x}_t\cdot{\bf u} - y_t \rightarrow {f^*}({\bf u})$.
In standard convex analysis, the Fenchel-Young inequality states that,
for any proper convex function $f$,
\begin{equation} \label{eqn:fenchel-stand}
f({\bf x}) + {f^*}({\bf u}) \geq {\bf x}\cdot{\bf u}
\end{equation}
for all ${\bf x}\in{\R^n}$ and all ${\bf u}\in{\R^n}$.
Furthermore, this holds with equality if and only if
${\bf u}\in\partial f({\bf x})$
\citep[Theorem 23.5]{ROC},
a very useful characterization of subgradients.
A version of
the Fenchel-Young inequality generalizes directly to functions
$F:\extspac{n}\rightarrow\overline{\R}$ over astral space
since \eqref{eq:Fstar-down-def}, which gives the form of the conjugate
${F^*}$,
immediately implies that
\begin{equation} \label{eqn:ast-fenchel}
{F^*}({\bf u})
\geq
- F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}
\end{equation}
for all $\overline{\xx}\in\extspac{n}$ and all ${\bf u}\in{\R^n}$.
The next theorem shows that if
\eqref{eqn:ast-fenchel} holds with equality, then ${\bf u}$ must be a
subgradient of $F$ at $\overline{\xx}$ (provided $\overline{\xx}\in\cldom{F}$).
Furthermore, if
$-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable
(so that the sum $-F(\overline{\xx})+\overline{\xx}\cdot{\bf u}$ is defined),
then the converse holds as well:
\begin{theorem} \label{thm:fenchel-implies-subgrad}
Let $F:\extspac{n}\rightarrow\overline{\R}$,
and let $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Consider the following statements:
\begin{enumerate}
\item \label{thm:fenchel-implies-subgrad:a}
${F^*}({\bf u}) = -F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}$
and
$\overline{\xx}\in\cldom{F}$.
\item \label{thm:fenchel-implies-subgrad:b}
${\bf u}\in\asubdifF{\overline{\xx}}$.
\end{enumerate}
Then
statement~(\ref{thm:fenchel-implies-subgrad:a})
implies
statement~(\ref{thm:fenchel-implies-subgrad:b}).
Furthermore, if
$-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable, then
the two statements~(\ref{thm:fenchel-implies-subgrad:a})
and~(\ref{thm:fenchel-implies-subgrad:b}) are equivalent.
\end{theorem}
\proof
~
\underline{
(\ref{thm:fenchel-implies-subgrad:a})
$\Rightarrow$
(\ref{thm:fenchel-implies-subgrad:b})}:
Suppose statement~(\ref{thm:fenchel-implies-subgrad:a}) holds.
Consider first the case that $F(\overline{\xx})=+\infty$, which implies, by our
assumption, that ${F^*}({\bf u})=-\infty$ so that
$\overline{\xx}'\cdot{\bf u}-y'=-\infty$
for all $\rpair{\overline{\xx}'}{y'}\in\epi F$
(by \eqref{eq:Fstar-def}).
Since $\overline{\xx}\in\cldom{F}$, there exists a sequence $\seq{\overline{\xx}_t}$ in
$\dom F$ that converges to $\overline{\xx}$.
For each $t$, let $y_t = \max\{t, F(\overline{\xx}_t)\}$,
so that $\rpair{\overline{\xx}_t}{y_t}\in\epi F$
(since $F(\overline{\xx}_t)<+\infty$),
implying
$\overline{\xx}_t\cdot{\bf u}-y_t=-\infty$.
Thus,
${\bf u}\in\partial F(\overline{\xx})$ by
Proposition~\ref{pr:equiv-ast-subdif-defn}
since
$\overline{\xx}_t\rightarrow\overline{\xx}$,
$y_t\rightarrow+\infty=F(\overline{\xx})$,
and
$\overline{\xx}_t\cdot{\bf u}-y_t\rightarrow-\infty={F^*}({\bf u})$.
Consider next the alternate case that $F(\overline{\xx})<+\infty$.
For each $t$,
let $\overline{\xx}_t=\overline{\xx}$ and $y_t=\max\{-t, F(\overline{\xx})\}$
so that $\rpair{\overline{\xx}_t}{y_t}\in\epi F$
with $\overline{\xx}_t\rightarrow\overline{\xx}$ and $y_t\rightarrow F(\overline{\xx})$.
If $\overline{\xx}\cdot{\bf u}=-\infty$ then
${F^*}({\bf u})=-\infty$ by the proposition's assumption, so
$\overline{\xx}_t\cdot{\bf u} - y_t = -\infty \rightarrow {F^*}({\bf u})$.
Otherwise,
\[
\overline{\xx}_t\cdot{\bf u} - y_t
\rightarrow
-F(\overline{\xx}) + \overline{\xx}\cdot{\bf u}
= {F^*}({\bf u})
\]
by continuity of addition (since $\overline{\xx}\cdot{\bf u}>-\infty$ and
$F(\overline{\xx})<+\infty$), and by our assumption.
In either case,
Proposition~\ref{pr:equiv-ast-subdif-defn} then implies
${\bf u}\in\partial F(\overline{\xx})$.
\underline{
(\ref{thm:fenchel-implies-subgrad:b})
[with conditions]
$\Rightarrow$
(\ref{thm:fenchel-implies-subgrad:a})}:
Suppose ${\bf u}\in\partial F(\overline{\xx})$
and that
$-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable.
Then by
Proposition~\ref{pr:equiv-ast-subdif-defn},
there exists a sequence $\seq{\rpair{\overline{\xx}_t}{y_t}}$ in $\epi F$
with
$\overline{\xx}_t\rightarrow\overline{\xx}$,
$y_t \rightarrow F(\overline{\xx})$,
and
$\overline{\xx}_t\cdot{\bf u} - y_t \rightarrow {F^*}({\bf u})$.
On the other hand,
\[
\overline{\xx}_t\cdot{\bf u} - y_t \rightarrow \overline{\xx}\cdot{\bf u} - F(\overline{\xx})
\]
by continuity of addition, since
$-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable
(and since $\overline{\xx}_t\cdot{\bf u}\rightarrow\overline{\xx}\cdot{\bf u}$
by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
Thus,
${F^*}({\bf u}) = \overline{\xx}\cdot{\bf u} - F(\overline{\xx})$
as claimed.
Also,
since each $\overline{\xx}_t\in\dom F$, their limit
$\overline{\xx}$ must be in $\cldom{F}$.
\qed
The (partial) converse proved in
Theorem~\ref{thm:fenchel-implies-subgrad}
does not hold in general
without the additional assumption that
$-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable.
In other words,
if $-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are not summable,
then it is possible
that ${\bf u}$ is a subgradient of $F$ at $\overline{\xx}$ but that
${F^*}({\bf u}) \neq - F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}$.
For instance, consider the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of the function $f$ given in
\eqref{eq:ex:ln1plusexp}, and let $u=1$ and $\bar{x}=+\infty$.
Then $u\in\partial{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\bar{x})$, as previously discussed, while
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\bar{x})=+\infty$ and ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}(u)={f^*}(u)=0$,
so that
${f^*}(u) = 0 \neq -\infty = - {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\bar{x}) \plusd \bar{x} u$.
The assumption that
$-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable
always holds when
$F(\overline{\xx})\in{\mathbb{R}}$ or
$\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$, including when $\overline{\xx}={\bf x}\in{\R^n}$ or when ${\bf u}={\bf 0}$.
Applied in this last case immediately yields that
${\bf 0}$ is a subgradient of $F$ at $\overline{\xx}$ if and only if
$\overline{\xx}$ minimizes $F$ (unless $F\equiv+\infty$):
\begin{proposition} \label{pr:asub-zero-is-min}
Let $F:\extspac{n}\rightarrow\overline{\R}$, with $F\not\equiv+\infty$.
Let $\overline{\xx}\in\extspac{n}$.
Then ${\bf 0}\in\asubdifF{\overline{\xx}}$ if and only if $\overline{\xx}$ minimizes
$F$.
\end{proposition}
\proof
\eqref{eq:Fstar-down-def} implies
\[
{F^*}({\bf 0})
=
\sup_{\overline{\xx}'\in\extspac{n}} [-F(\overline{\xx}')]
=
-\inf_{\overline{\xx}'\in\extspac{n}} F(\overline{\xx}').
\]
Therefore, $F(\overline{\xx})=-{F^*}({\bf 0})$ if and only if $\overline{\xx}$ minimizes
$F$.
So if ${\bf 0}\in\partial F(\overline{\xx})$, then
$F(\overline{\xx})=-{F^*}({\bf 0})$ by
Theorem~\ref{thm:fenchel-implies-subgrad}
implying that $\overline{\xx}$ minimizes $F$.
Conversely, if $\overline{\xx}$ minimizes $F$, then
$F(\overline{\xx})=-{F^*}({\bf 0})$ and also
$F(\overline{\xx})<+\infty$ since $F\not\equiv+\infty$.
Therefore, ${\bf 0}\in\partial F(\overline{\xx})$ by
Theorem~\ref{thm:fenchel-implies-subgrad}.
\qed
Astral subgradients generalize standard subgradients in the sense that
the standard subdifferential of a function
$f:{\R^n}\rightarrow\overline{\R}$ is the same as
the astral subdifferential of an essentially equivalent astral function
$F:\extspac{n}\rightarrow\overline{\R}$ that is the same as $f$ on ${\R^n}$ and
$+\infty$ everywhere else, so that the epigraphs of $f$ and $F$ are
identical.
In addition,
the astral subdifferential of $f$'s extension, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
is equal to the standard subdifferential of $f$
at all points in ${\R^n}$ where $f$ is lower semicontinuous
(which is empty everywhere else).
These statements hold always except if $f\equiv+\infty$,
in which case
$\partial f({\bf x})={\R^n}$ but
$\asubdiffext{{\bf x}}=\asubdifF{{\bf x}}=\emptyset$
for all ${\bf x}\in{\R^n}$.
\begin{proposition} \label{pr:asubdiffext-at-x-in-rn}
Let $f:{\R^n}\rightarrow\overline{\R}$, with $f\not\equiv+\infty$,
and let ${\bf x}\in{\R^n}$.
Then the following hold:
\begin{enumerate}
\item \label{pr:asubdiffext-at-x-in-rn:a}
Let $F:\extspac{n}\rightarrow\overline{\R}$ be defined by
\[
F(\overline{\xx})
=
\left\{ \begin{array}{cl}
f(\overline{\xx}) & \mbox{if $\overline{\xx}\in{\R^n}$} \\
+\infty & \mbox{otherwise.}
\end{array} \right.
\]
Then $\asubdifF{{\bf x}} = \partial f({\bf x})$.
\item \label{pr:asubdiffext-at-x-in-rn:b}
In addition,
\[
\partial f({\bf x})
=
\left\{ \begin{array}{cl}
\asubdiffext{{\bf x}} & \mbox{if $(\lsc f)({\bf x}) = f({\bf x})$} \\
\emptyset & \mbox{otherwise.}
\end{array} \right.
\]
Consequently, $\asubdiffext{{\bf x}} = \partial (\lsc f)({\bf x})$.
\end{enumerate}
\end{proposition}
\proof
Let ${\bf u}\in{\R^n}$.
As a preliminary step,
we observe that
by simple algebra,
the definition of standard subgradient given in
\eqref{eqn:standard-subgrad-ineq}
holds if and only if
${\bf x}\cdot{\bf u}-f({\bf x}) \geq {\bf x}'\cdot{\bf u}-f({\bf x}')$
for all ${\bf x}'\in{\R^n}$, that is, if and only if
\[
{\bf x}\cdot{\bf u}-f({\bf x})
=
\sup_{{\bf x}'\in{\R^n}} [{\bf x}'\cdot{\bf u}-f({\bf x}')].
\]
Since the term on the right is exactly ${f^*}({\bf u})$, this shows
that
${\bf u}\in\partial f({\bf x})$ if and only if
${f^*}({\bf u})={\bf x}\cdot{\bf u}-f({\bf x})$.
In particular, if $f({\bf x})=+\infty$ then
${f^*}({\bf u})\neq{\bf x}\cdot{\bf u}-f({\bf x})$
since ${f^*}>-\infty$ (by our assumption that
$f\not\equiv+\infty$), which in turn implies
${\bf u}\not\in\partial f({\bf x})$.
Part~(\ref{pr:asubdiffext-at-x-in-rn:a}):
We show that ${\bf u}\in\partial f({\bf x})$ if and only if
${\bf u}\in\asubdifF{{\bf x}}$.
By construction, $F({\bf x})=f({\bf x})$ and $\epi F = \epi f$,
implying ${F^*}={f^*}$, by comparison of
\eqref{eq:fstar-mod-def} and \eqref{eq:Fstar-def}.
Therefore,
${\bf x}\cdot{\bf u} - F({\bf x}) = {F^*}({\bf u})$
if and only if
${\bf x}\cdot{\bf u} - f({\bf x}) = {f^*}({\bf u})$.
If $F({\bf x})=f({\bf x})=+\infty$, then, as noted above,
${\bf x}\cdot{\bf u} - f({\bf x}) \neq {f^*}({\bf u})$ and
${\bf u}\not\in\partial f({\bf x})$,
implying
${\bf x}\cdot{\bf u} - F({\bf x}) \neq {F^*}({\bf u})$,
so that
${\bf u}\not\in\asubdifF{{\bf x}}$
by Theorem~\ref{thm:fenchel-implies-subgrad},
proving the claim in this case.
In the alternative case,
$F({\bf x})=f({\bf x})<+\infty$.
Since ${\bf u}\in\partial f({\bf x})$ if and only if
${\bf x}\cdot{\bf u} - f({\bf x}) = {f^*}({\bf u})$
(by the observation above), and since
${\bf u}\in\asubdifF{{\bf x}}$ if and only if
${\bf x}\cdot{\bf u} - F({\bf x}) = {F^*}({\bf u})$
by Theorem~\ref{thm:fenchel-implies-subgrad}
(since ${\bf x}\in\dom F$),
it follows that
${\bf u}\in\partial f({\bf x})$ if and only if
${\bf u}\in\asubdifF{{\bf x}}$, proving the claim in this case as well.
Part~(\ref{pr:asubdiffext-at-x-in-rn:b}):
To prove the result, we show that
${\bf u}\in \partial f({\bf x})$
if and only if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})=f({\bf x})$ and
${\bf u}\in\asubdiffext{{\bf x}}$.
(Recall that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})=(\lsc f)({\bf x})$, by
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).)
Suppose ${\bf u}\in\partial f({\bf x})$, which implies
$f({\bf x})<+\infty$, as noted above.
Then
\[
{\bf x}\cdot{\bf u} - {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})
\geq
{\bf x}\cdot{\bf u} - f({\bf x})
=
{f^*}({\bf u})
=
{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}({\bf u})
\geq
{\bf x}\cdot{\bf u} - {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x}).
\]
The first inequality is by
Proposition~\ref{pr:h:1}(\ref{pr:h:1a}).
The two equalities are by the observation above and
Proposition~\ref{pr:fextstar-is-fstar}, respectively.
The last inequality is by
\eqref{eqn:ast-fenchel} (applied to $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$).
Thus, $(\lsc f)({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})=f({\bf x})<+\infty$
(using ${\bf x}\cdot{\bf u}\in{\mathbb{R}}$), and
${\bf u}\in\asubdiffext{{\bf x}}$ by
Theorem~\ref{thm:fenchel-implies-subgrad}.
Conversely, suppose now that
$(\lsc f)({\bf x})={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})=f({\bf x})$
and that
${\bf u}\in\asubdiffext{{\bf x}}$.
Then
\[
{\bf x}\cdot{\bf u} - f({\bf x})
=
{\bf x}\cdot{\bf u} - {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}({\bf x})
=
{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}({\bf u})
=
{f^*}({\bf u})
\]
with the second equality following from
Theorem~\ref{thm:fenchel-implies-subgrad}
(since ${\bf u}\in\asubdiffext{{\bf x}}$ and ${\bf x}\cdot{\bf u}\in{\mathbb{R}}$),
and the third from
Proposition~\ref{pr:fextstar-is-fstar}.
Thus, ${\bf u}\in\partial f({\bf x})$ by the observation above.
Applied to $\lsc f$, which is lower semicontinuous everywhere,
this shows that
$\asubdiffext{{\bf x}} = \partial \overline{(\lsc f)}({\bf x})
= \partial(\lsc f)({\bf x})$
(by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa})).
\qed
\subsection{Astral dual subgradients}
\label{sec:astral-dual-subgrad}
We turn next to the dual subdifferential, which maps from ${\R^n}$,
regarded as a dual space, to subsets of astral space.
As such, let $\psi:{\R^n}\rightarrow\overline{\R}$ be any function, changing
notation to emphasize the switch to dual space, and because
later we will usually take $\psi$ to be the conjugate of some
function.
Replacing $f$ by $\psi$ and swapping variable names in
\eqref{eqn:standard-subgrad-ineq},
we can say that ${\bf x}\in{\R^n}$ is a usual, standard
subgradient of $\psi$ at ${\bf u}\in{\R^n}$ if
\begin{equation} \label{eqn:psi-subgrad:1}
\psi({\bf u}') \geq \psi({\bf u}) + {\bf x}\cdot({\bf u}'-{\bf u})
\end{equation}
for all ${\bf u}'\in{\R^n}$.
To extend this notion to astral space,
while avoiding the possibility of adding
$-\infty$ and $+\infty$, we again focus on epigraphs, specifically, on
$\epi \psi$.
\eqref{eqn:psi-subgrad:1} is equivalent to
\[
v' \geq \psi({\bf u}) + {\bf x}\cdot({\bf u}'-{\bf u})
\]
for all $\rpair{{\bf u}'}{v'}\in\epi \psi$, which in turn is equivalent to
\begin{equation} \label{eqn:psi-subgrad:1a}
-\psi({\bf u}) \geq {\bf x}\cdot({\bf u}'-{\bf u}) - v'
\end{equation}
by simple algebra.
In this form, the definition of standard subgradient immediately
and naturally generalizes to astral space since we can
simply replace ${\bf x}$ by an astral point $\overline{\xx}\in\extspac{n}$
in \eqref{eqn:psi-subgrad:1a}.
Thus, we say that $\overline{\xx}\in\extspac{n}$ is an
\emph{astral dual subgradient} of $\psi$ at ${\bf u}\in{\R^n}$ if
\begin{equation} \label{eqn:psi-subgrad:3a}
-\psi({\bf u}) \geq \overline{\xx}\cdot({\bf u}'-{\bf u}) - v'
\end{equation}
for all $\rpair{{\bf u}'}{v'}\in\epi \psi$.
We then define the
\emph{astral dual subdifferential} of $\psi$ at ${\bf u}\in{\R^n}$,
denoted $\adsubdifpsi{{\bf u}}$,
to be the set of all such astral dual subgradients
of $\psi$ at ${\bf u}$.
Equivalently,
$\overline{\xx}$ is an {astral dual subgradient} of $\psi$ at ${\bf u}$ if and only if
\begin{equation} \label{eqn:psi-subgrad:3}
-\psi({\bf u}) \geq -\psi({\bf u}') \plusd \overline{\xx}\cdot({\bf u}'-{\bf u})
\end{equation}
for all ${\bf u}'\in{\R^n}$, since,
by
Proposition~\ref{pr:plusd-props}(\ref{pr:plusd-props:d}),
the right-hand side of
\eqref{eqn:psi-subgrad:3} is equal to the supremum of
the right-hand side of \eqref{eqn:psi-subgrad:3a} over all
$v'\geq\psi({\bf u}')$.
Furthermore, by
Proposition~\ref{pr:plusd-props}(\ref{pr:plusd-props:e}),
\eqref{eqn:psi-subgrad:3}
is equivalent to
\begin{equation} \label{eqn:psi-subgrad:3-alt}
\psi({\bf u}') \geq \psi({\bf u}) \plusd \overline{\xx}\cdot({\bf u}'-{\bf u})
\end{equation}
for all ${\bf u}'\in{\R^n}$, which even more closely resembles the standard
definition given in
\eqref{eqn:psi-subgrad:1}.
We use the notation $\adsubdifplain{\psi}$ to distinguish the astral
dual subdifferential from the standard subdifferential
$\partial \psi$, since either operation can be applied to an ordinary
function $\psi$ over ${\R^n}$.
(In contrast, the astral subdifferential $\asubdifplain{F}$ is
only applied to functions $F$ over $\extspac{n}$.)
In addition to acting as a dual operation,
astral dual subdifferentials can be useful as a means of
capturing subgradients that may be infinite.
In other words, the astral primal subdifferential
$\asubdifF{\overline{\xx}}$ captures finite subgradients ${\bf u}\in{\R^n}$
of an astral function $F:\extspac{n}\rightarrow\overline{\R}$
at astral (and therefore potentially infinite) points
$\overline{\xx}\in\extspac{n}$,
while the astral dual subdifferential
$\adsubdifpsi{{\bf u}}$ captures astral (and so potentially infinite)
subgradients $\overline{\xx}\in\extspac{n}$
of a function $\psi:{\R^n}\rightarrow\overline{\R}$
at finite points ${\bf u}\in{\R^n}$.
For example, suppose
\[
\psi(u)
=
\left\{ \begin{array}{cl}
u \ln u & \mbox{if $u\geq 0$} \\
+\infty & \mbox{otherwise}
\end{array} \right.
\]
for $u\in{\mathbb{R}}$, (where, as usual, $0\ln 0 = 0$).
For $u > 0$,
the astral subgradients of this function are the same as its
derivative.
But for $u\leq 0$, the function has no standard subgradients, even at
$u=0$ which is in the function's effective domain.
Nevertheless, at $u=0$, $-\infty$ is an astral dual subgradient, as is
to be expected since the derivative of this function approaches
$-\infty$ as $u$ approaches $0$ from the right.
As a more interesting example, suppose
\[
\psi({\bf u})
=
\left\{ \begin{array}{cl}
-\sqrt{1-\norm{{\bf u}}^2} & \mbox{if $\norm{{\bf u}}\leq 1$} \\
+\infty & \mbox{otherwise}
\end{array} \right.
\]
for ${\bf u}\in{\R^n}$, whose graph in ${\R^{n+1}}$ is the bottom half of a
hypersphere.
Similar to the previous example,
the astral dual and standard subgradients at a point ${\bf u}$ with
$\norm{{\bf u}}<1$ are simply the gradients of the function.
However, at a point ${\bf u}$ with $\norm{{\bf u}}=1$,
which is still in its effective domain,
the function has no standard subgradients.
On the other hand,
it does have astral dual subgradients at such a point, including
the astron $\limray{{\bf u}}$, as well as every infinite point
$\overline{\xx}\in\extspac{n}$ with dominant direction ${\bf u}$;
that is, if $\norm{{\bf u}}=1$ then
\[
\adsubdifpsi{{\bf u}} = \limray{{\bf u}} \mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax \extspac{n}.
\]
Thus, the astral dual subgradients are not only infinite, but also
provide sensible directional information.
(The astral dual subdifferentials at points entirely outside the
effective domain of this and the preceding function are also nonempty,
as will be explored
in Section~\ref{sec:dual-subdiff-not-empty}.)
Returning to the general case,
for any function $\psi:{\R^n}\rightarrow\overline{\R}$,
\eqref{eqn:psi-subgrad:3} immediately implies that
${\bf u}\in{\R^n}$ minimizes $\psi$ if and only if
${\bf 0}$ is an astral dual subgradient of $\psi$ at ${\bf u}$ so that
${\bf 0}\in\adsubdifpsi{{\bf u}}$.
Also, astral dual subgradients generalize standard subgradients in the
sense that
the standard subdifferential of $\psi$ at ${\bf u}$
is exactly equal to the
finite points included in $\psi$'s astral dual subdifferential at ${\bf u}$:
\begin{proposition} \label{pr:adsubdif-int-rn}
Let $\psi:{\R^n}\rightarrow\overline{\R}$, and let ${\bf u}\in{\R^n}$.
Then $\partial\psi({\bf u}) = \adsubdifpsi{{\bf u}}\cap{\R^n}$.
\end{proposition}
\proof
When $\overline{\xx}={\bf x}\in{\R^n}$,
\eqref{eqn:psi-subgrad:1a} and
\eqref{eqn:psi-subgrad:3a} are equivalent
for all $\rpair{{\bf u}'}{v'}\in\epi \psi$.
Since the former defines standard subgradients and the latter defines
astral dual subgradients, this proves the claim.
\qed
\subsection{Conditions and relations among subgradients}
\label{sec:subgrad-inv}
In standard convex analysis, it is known that
$\partial f$ and $\partial {f^*}$ act as inverses of one another in
the sense that,
for ${\bf x}\in{\R^n}$ and ${\bf u}\in{\R^n}$,
if $f:{\R^n}\rightarrow\overline{\R}$ is proper and convex, and if
$(\cl f)({\bf x}) = f^{**}({\bf x})=f({\bf x})$ then
${\bf u}\in\partial f({\bf x})$ if and only if
${\bf x}\in\partial {f^*}({\bf u})$
\citep[Theorem 23.5]{ROC}.
We also discussed earlier that
${\bf u}\in\partial f({\bf x})$ if and only if
the Fenchel-Young inequality (\eqref{eqn:fenchel-stand}) holds with
equality.
Thus, under the conditions above, the following are equivalent:
\begin{enumerate}
\item
$f({\bf x})+{f^*}({\bf u}) = {\bf x}\cdot{\bf u}$.
\item
${\bf u}\in\partial f({\bf x})$.
\item
${\bf x}\in\partial {f^*}({\bf u})$.
\end{enumerate}
In this section, we explore in detail the analogous connections between
the astral subdifferential of a function $F:\extspac{n}\rightarrow\overline{\R}$
and the astral dual subdifferential of its dual ${F^*}$.
We also relate these to when the form of the Fenchel-Young
inequality given in \eqref{eqn:ast-fenchel} holds with equality, a
connection that was seen already in
Theorem~\ref{thm:fenchel-implies-subgrad}.
Thus, for a function $F:\extspac{n}\rightarrow\overline{\R}$ and points
$\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$,
we study how the following three conditions relate to one another:
\begin{enumerate}
\item
${F^*}({\bf u}) = -F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}$.
\item
${\bf u}\in\asubdifF{\overline{\xx}}$.
\item
$\overline{\xx}\in\adsubdifFstar{{\bf u}}$.
\end{enumerate}
It was seen in
Theorem~\ref{thm:fenchel-implies-subgrad}
that the first condition generally implies the second
(provided $\overline{\xx}\in\cldom{F}$),
and we will see shortly
in Theorem~\ref{thm:asubdif-implies-adsubdif}
that the second always implies the third.
Under appropriate summability conditions,
as in Theorem~\ref{thm:fenchel-implies-subgrad},
we will also see that the three conditions
are equivalent to one another.
Furthermore, in the centrally important case that $F$ is the
extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of a convex function $f:{\R^n}\rightarrow\overline{\R}$,
we prove below (Theorem~\ref{thm:adif-fext-inverses})
that
${\bf u}\in\asubdiffext{\overline{\xx}}$ if and only if
$\overline{\xx}\in\adsubdiffstar{{\bf u}}$,
for all ${\bf u}\in{\R^n}$ and $\overline{\xx}\in\cldom{f}$.
As a next step,
we show that
if ${\bf u}\in\asubdifF{\overline{\xx}}$ then
it always follows that
$\overline{\xx}\in\adsubdifFstar{{\bf u}}$ and also $\overline{\xx}\in\cldom{F}$.
In addition, if $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$ then we can further infer that
$F(\overline{\xx})=F^{*\dualstar}(\overline{\xx})$,
thereby generalizing a result from standard convex analysis that if
$\partial f({\bf x})$ is nonempty then $f({\bf x})=f^{**}({\bf x})$,
for $f:{\R^n}\rightarrow\overline{\R}$ proper and convex, and any ${\bf x}\in{\R^n}$
\citep[Corollary~23.5.2]{ROC}.
\begin{theorem} \label{thm:asubdif-implies-adsubdif}
Let $F:\extspac{n}\rightarrow\overline{\R}$, and let $\overline{\xx}\in\extspac{n}$ and
${\bf u}\in{\R^n}$.
Suppose ${\bf u}\in\asubdifF{\overline{\xx}}$.
Then all of the following hold:
\begin{enumerate}
\item \label{thm:asubdif-implies-adsubdif:a}
$\overline{\xx}\in\adsubdifFstar{{\bf u}}$.
\item \label{thm:asubdif-implies-adsubdif:b}
$\overline{\xx}\in\cldom{F}$.
\item \label{thm:asubdif-implies-adsubdif:c}
Either $F(\overline{\xx})=F^{*\dualstar}(\overline{\xx})$ or $\overline{\xx}\cdot{\bf u}\not\in{\mathbb{R}}$.
\end{enumerate}
\end{theorem}
\proof
Since ${\bf u}\in\asubdifF{\overline{\xx}}$,
Proposition~\ref{pr:equiv-ast-subdif-defn}(\ref{pr:equiv-ast-subdif-defn:b})
implies that there exists
a sequence $\seq{\rpair{\overline{\xx}_t}{y_t}}$ in $\epi F$
with
$\overline{\xx}_t\rightarrow\overline{\xx}$,
$y_t \rightarrow F(\overline{\xx})$,
and
\begin{equation} \label{eq:thm:asubdif-implies-adsubdif:2}
\overline{\xx}_t\cdot{\bf u} - y_t \rightarrow {F^*}({\bf u}).
\end{equation}
We use this same sequence for the entire proof.
Part~(\ref{thm:asubdif-implies-adsubdif:a}):
Using the equivalent form of the definition of astral dual subgradient
given in
\eqref{eqn:psi-subgrad:3-alt},
and setting ${\bf w}={\bf u}'-{\bf u}$, we can prove
$\overline{\xx}\in\adsubdifFstar{{\bf u}}$ by showing,
for all ${\bf w}\in{\R^n}$, that
\begin{equation} \label{eq:thm:asubdif-implies-adsubdif:1}
{F^*}({\bf u}+{\bf w})
\geq
{F^*}({\bf u}) \plusd \overline{\xx}\cdot{\bf w}.
\end{equation}
Let ${\bf w}\in{\R^n}$.
If either ${F^*}({\bf u})=-\infty$ or $\overline{\xx}\cdot{\bf w}=-\infty$, then
\eqref{eq:thm:asubdif-implies-adsubdif:1}
holds trivially since the right-hand side is $-\infty$ in either case.
Therefore, we assume henceforth that
${F^*}({\bf u})>-\infty$ and $\overline{\xx}\cdot{\bf w}>-\infty$.
From \eqref{eq:thm:asubdif-implies-adsubdif:2}
and since
\begin{equation} \label{eq:thm:asubdif-implies-adsubdif:3}
\overline{\xx}_t\cdot{\bf w} \rightarrow \overline{\xx}\cdot{\bf w}
\end{equation}
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})),
this implies that
$\overline{\xx}_t\cdot{\bf u}>-\infty$ and $\overline{\xx}_t\cdot{\bf w}>-\infty$
for $t$ sufficiently large; by discarding all other elements of the
sequence, we assume this holds for all $t$.
Thus, we have
\begin{eqnarray*}
{F^*}({\bf u}+{\bf w})
&\geq&
\overline{\xx}_t\cdot({\bf u}+{\bf w}) - y_t
\\
&=&
\overline{\xx}_t\cdot{\bf u} + \overline{\xx}_t\cdot{\bf w} - y_t
\\
&=&
(\overline{\xx}_t\cdot{\bf u} - y_t) + \overline{\xx}_t\cdot{\bf w}.
\end{eqnarray*}
The inequality is by \eqref{eq:Fstar-def}
(since $\rpair{\overline{\xx}_t}{y_t}\in\epi F$), and the first equality is by
Proposition~\ref{pr:i:1}
(since neither
$\overline{\xx}_t\cdot{\bf u}$ nor $\overline{\xx}_t\cdot{\bf w}$ can be $-\infty$).
Since this inequality holds for all $t$, it also must hold in the
limit.
Thus,
\[
{F^*}({\bf u}+{\bf w})
\geq
\lim [(\overline{\xx}_t\cdot{\bf u} - y_t) + \overline{\xx}_t\cdot{\bf w}]
=
{F^*}({\bf u}) + \overline{\xx}\cdot{\bf w}
\]
by
\eqref{eq:thm:asubdif-implies-adsubdif:2}
and
\eqref{eq:thm:asubdif-implies-adsubdif:3},
and by continuity of addition
(since neither
${F^*}({\bf u})$ nor $\overline{\xx}\cdot{\bf w}$ is $-\infty$).
This proves \eqref{eq:thm:asubdif-implies-adsubdif:1}.
Part~(\ref{thm:asubdif-implies-adsubdif:b}):
Each $\overline{\xx}_t\in\dom F$ and $\overline{\xx}_t\rightarrow\overline{\xx}$.
Therefore, $\overline{\xx}\in\cldom{F}$.
Part~(\ref{thm:asubdif-implies-adsubdif:c}):
Suppose $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$.
Then
Theorem~\ref{thm:fenchel-implies-subgrad}
implies
${F^*}({\bf u}) = \overline{\xx}\cdot{\bf u} - F(\overline{\xx})$,
and therefore
$F(\overline{\xx}) = \overline{\xx}\cdot{\bf u} - {F^*}({\bf u})$
since $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$.
Thus,
\[
F(\overline{\xx})
\geq
F^{*\dualstar}(\overline{\xx})
\geq
\overline{\xx}\cdot{\bf u} - {F^*}({\bf u})
=
F(\overline{\xx}),
\]
with the inequalities following respectively from
Theorem~\ref{thm:fdub-sup-afffcns}
and
\eqref{eq:psistar-def:2}
(applied to $\psi={F^*}$).
\qed
As discussed earlier, in standard convex analysis, it is known that if
$\partial f({\bf x})$ is nonempty
then $f({\bf x})=f^{**}({\bf x})$.
Theorem~\ref{thm:asubdif-implies-adsubdif} shows that if
there exists
${\bf u}\in\asubdifF{\overline{\xx}}$ for which it also holds that
$\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$
then
$F(\overline{\xx})=F^{*\dualstar}(\overline{\xx})$.
Without the additional condition that $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$,
the theorem would be false, in general.
In other words, it is possible that
${\bf u}\in\asubdifF{\overline{\xx}}$ but
$F(\overline{\xx})\neqF^{*\dualstar}(\overline{\xx})$.
For example, suppose
\[
F(\bar{x})
=
\left\{ \begin{array}{cl}
0 & \mbox{if $\bar{x}\in{\mathbb{R}}$} \\
+\infty & \mbox{otherwise}
\end{array} \right.
\]
for $\bar{x}\in\overline{\R}$.
Let $\bar{x}=+\infty$ and $u=1$.
Then it can be checked that ${F^*}(u)=+\infty$
and $F^{*\dualstar}\equiv 0$.
Let $x_t=t$ and $y_t=t/2$ for all $t$.
Then $x_t\rightarrow\bar{x}$,
$y_t\rightarrow F(\bar{x})=+\infty$,
and
$x_t u - y_t \rightarrow {F^*}(u) = +\infty$.
Thus, $u\in\asubdifF{\bar{x}}$,
but
$F(\bar{x})=+\infty\neq 0 =F^{*\dualstar}(\bar{x})$.
This counterexample works to make the same point if we re-define
$F(+\infty)$ to be equal to $1$ and also set $y_t=1$ for all $t$.
However, the
resulting function $F$ is not convex since $\rpair{+\infty}{1}$ and
$\rpair{0}{0}$ are both in $\epi F$, but $\rpair{+\infty}{0}$,
which is on the segment between these two points, is not.
Of course,
Theorem~\ref{thm:asubdif-implies-adsubdif}, in its present form, does
not assume convexity.
So far, we have taken as our starting point a function
$F:\extspac{n}\rightarrow\overline{\R}$, and have considered the conditions
discussed above in terms of its astral subdifferential
$\asubdifplain{F}$, its conjugate ${F^*}$, and the astral dual
subdifferential of that conjugate, $\adsubdifplain{{F^*}}$.
We next take a somewhat different approach in which, beginning
with a function $\psi:{\R^n}\rightarrow\overline{\R}$, we focus on its
astral dual subdifferential $\adsubdifplain{\psi}$, its astral dual
conjugate $\psi^{\dualstar}$, and the astral (primal) subdifferential of that
conjugate, $\asubdifplain{\psi^{\dualstar}}$.
The next theorem shows how conditions analogous to the ones we have
been considering regarding subgradients and the Fenchel-Young
inequality can be related to one another in this alternative formulation:
\begin{theorem} \label{thm:psi-subgrad-conds}
Let $\psi:{\R^n}\rightarrow\overline{\R}$, and
let $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Consider the following statements:
\begin{enumerate}
\item \label{thm:psi-subgrad-conds:a}
$\psi({\bf u}) = -\psi^{\dualstar}(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}$
and
$\overline{\xx}\in\cldom{\psi^{\dualstar}}$.
\item \label{thm:psi-subgrad-conds:b}
${\bf u}\in\asubdifpsistarb{\overline{\xx}}$
and
$\psi^{\dualstar*}({\bf u}) = \psi({\bf u})$.
\item \label{thm:psi-subgrad-conds:c}
$\overline{\xx}\in\adsubdifpsi{{\bf u}}$
and
$\overline{\xx}\in\cldom{\psi^{\dualstar}}$.
\end{enumerate}
Then
statement~(\ref{thm:psi-subgrad-conds:a})
implies
statement~(\ref{thm:psi-subgrad-conds:b}),
and
statement~(\ref{thm:psi-subgrad-conds:b})
implies
statement~(\ref{thm:psi-subgrad-conds:c}).
\end{theorem}
\proof
~
\underline{
(\ref{thm:psi-subgrad-conds:a})
$\Rightarrow$
(\ref{thm:psi-subgrad-conds:b})}:
Suppose statement~(\ref{thm:psi-subgrad-conds:a}) holds.
Then
\[
-\psi^{\dualstar}(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}
=
\psi({\bf u})
\geq
\psi^{\dualstar*}({\bf u})
\geq
-\psi^{\dualstar}(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}.
\]
The equality is by assumption.
The two inequalities follow from
Theorem~\ref{thm:psi-geq-psidub}
and
\eqref{eqn:ast-fenchel}.
Thus,
\[
\psi({\bf u})
=
\psi^{\dualstar*}({\bf u})
=
-\psi^{\dualstar}(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u},
\]
implying ${\bf u}\in\asubdifpsistarb{\overline{\xx}}$
by
Theorem~\ref{thm:fenchel-implies-subgrad}
(applied to $F=\psi^{\dualstar}$).
\underline{
(\ref{thm:psi-subgrad-conds:b})
$\Rightarrow$
(\ref{thm:psi-subgrad-conds:c})}:
Suppose statement~(\ref{thm:psi-subgrad-conds:b}) holds.
Then
Theorem~\ref{thm:asubdif-implies-adsubdif},
applied to $F=\psi^{\dualstar}$, implies that
$\overline{\xx}\in\cldom{\psi^{\dualstar}}$
and that
$\overline{\xx}\in\adsubdifpsidub{{\bf u}}$.
Therefore, using the form of the definition of astral dual subgradient
given in \eqref{eqn:psi-subgrad:3},
we have that for all ${\bf u}'\in{\R^n}$,
\[
-\psi({\bf u})
=
-\psi^{\dualstar*}({\bf u})
\geq
-\psi^{\dualstar*}({\bf u}') \plusd \overline{\xx}\cdot({\bf u}'-{\bf u})
\geq
-\psi({\bf u}') \plusd \overline{\xx}\cdot({\bf u}'-{\bf u}).
\]
The equality is by assumption, and the last inequality
is by Theorem~\ref{thm:psi-geq-psidub}.
Thus, $\overline{\xx}\in\adsubdifpsi{{\bf u}}$, as claimed.
\qed
When $-\psi^{\dualstar}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable,
the next theorem shows that
statement~(\ref{thm:psi-subgrad-conds:c})
in Theorem~\ref{thm:psi-subgrad-conds}
implies statement~(\ref{thm:psi-subgrad-conds:a}),
and therefore that all three statements appearing in that theorem are
equivalent to one another.
\begin{theorem} \label{thm:psi-subgrad-conds-part2}
Let $\psi:{\R^n}\rightarrow\overline{\R}$, and
let $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Suppose $-\psi^{\dualstar}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable,
and that
$\overline{\xx}\in\adsubdifpsi{{\bf u}}$.
Then
$\psi({\bf u}) = \overline{\xx}\cdot{\bf u} - \psi^{\dualstar}(\overline{\xx})$.
\end{theorem}
\proof
It suffices to prove
\begin{equation} \label{eq:thm:psi-subgrad-conds:1}
\psi({\bf u}) \leq \overline{\xx}\cdot{\bf u} - \psi^{\dualstar}(\overline{\xx})
\end{equation}
since, once proved, this will imply
\[
\psi({\bf u})
\geq
\psi^{\dualstar*}({\bf u})
\geq
\overline{\xx}\cdot{\bf u} - \psi^{\dualstar}(\overline{\xx})
\geq
\psi({\bf u}),
\]
with the first two inequalities following, as before,
from Theorem~\ref{thm:psi-geq-psidub}
and
\eqref{eqn:ast-fenchel}, thereby proving
the claim.
We aim therefore to prove
\eqref{eq:thm:psi-subgrad-conds:1}.
This inequality holds trivially if
$\psi({\bf u})=-\infty$
or
$\overline{\xx}\cdot{\bf u}=+\infty$
or
$\psi^{\dualstar}(\overline{\xx})=-\infty$.
Therefore, we assume henceforth that none of these conditions hold.
Let $\lambda\in{\mathbb{R}}$ be such that $\lambda<\psi^{\dualstar}(\overline{\xx})$.
(Such $\lambda$ must exist since $\psi^{\dualstar}(\overline{\xx})>-\infty$.)
By the definition of dual conjugate given in \eqref{eq:psistar-def},
there must exist $\rpair{{\bf u}'}{v'}\in\epi\psi$ with
$\overline{\xx}\cdot{\bf u}'-v'>\lambda$.
In particular, this implies $\overline{\xx}\cdot{\bf u}'>-\infty$.
Also, $\overline{\xx}\cdot(-{\bf u})>-\infty$ since $\overline{\xx}\cdot{\bf u}<+\infty$.
Thus,
\begin{eqnarray}
\lambda - \overline{\xx}\cdot{\bf u}
&=&
\lambda + \overline{\xx}\cdot(-{\bf u})
\nonumber
\\
&\leq&
(\overline{\xx}\cdot{\bf u}' - v') + \overline{\xx}\cdot(-{\bf u})
\nonumber
\\
&=&
\overline{\xx}\cdot({\bf u}'-{\bf u}) - v'
\nonumber
\\
&\leq&
-\psi({\bf u}).
\label{eq:thm:psi-subgrad-conds:2}
\end{eqnarray}
The second equality follows from
Proposition~\ref{pr:i:1}
since $\overline{\xx}\cdot{\bf u}'>-\infty$
and
$\overline{\xx}\cdot(-{\bf u})>-\infty$.
The last inequality uses our assumption that
$\overline{\xx}\in\adsubdifpsi{{\bf u}}$, together with the definition
of astral dual subgradient given in
\eqref{eqn:psi-subgrad:3a}.
Since $\psi({\bf u})>-\infty$,
\eqref{eq:thm:psi-subgrad-conds:2} implies that
$\overline{\xx}\cdot{\bf u}>-\infty$, and thus that $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$.
Therefore,
\eqref{eq:thm:psi-subgrad-conds:2} yields that
$\lambda\leq\overline{\xx}\cdot{\bf u}-\psi({\bf u})$.
Since this holds for all
$\lambda<\psi^{\dualstar}(\overline{\xx})$, it follows that
$\psi^{\dualstar}(\overline{\xx})\leq\overline{\xx}\cdot{\bf u}-\psi({\bf u})$, proving
\eqref{eq:thm:psi-subgrad-conds:1}
(since $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$), and completing the proof.
\qed
As noted above,
Theorems~\ref{thm:psi-subgrad-conds}
and~\ref{thm:psi-subgrad-conds-part2} together imply
the following equivalence:
\begin{corollary} \label{cor:psi-subgrad-conds}
Let $\psi:{\R^n}\rightarrow\overline{\R}$, and
let $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Assume the following hold:
\begin{itemize}
\item
$-\psi^{\dualstar}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable.
\item
Either $\psi({\bf u})>-\infty$ or $\overline{\xx}\in\cldom{\psi^{\dualstar}}$.
\end{itemize}
Then the following are equivalent:
\begin{enumerate}
\item \label{cor:psi-subgrad-conds:a}
$\psi({\bf u}) = \overline{\xx}\cdot{\bf u} - \psi^{\dualstar}(\overline{\xx})$.
\item \label{cor:psi-subgrad-conds:b}
${\bf u}\in\asubdifpsistarb{\overline{\xx}}$
and
$\psi^{\dualstar*}({\bf u}) = \psi({\bf u})$.
\item \label{cor:psi-subgrad-conds:c}
$\overline{\xx}\in\adsubdifpsi{{\bf u}}$.
\end{enumerate}
\end{corollary}
\proof
~
\underline{
(\ref{cor:psi-subgrad-conds:a})
$\Rightarrow$
(\ref{cor:psi-subgrad-conds:b})}:
Suppose
$\psi({\bf u}) = \overline{\xx}\cdot{\bf u} - \psi^{\dualstar}(\overline{\xx})$.
If $\psi({\bf u})>-\infty$, then this implies
$\psi^{\dualstar}(\overline{\xx})<+\infty$.
Therefore, our assumption that
either $\overline{\xx}\in\cldom{\psi^{\dualstar}}$ or $\psi({\bf u})>-\infty$
in fact always implies that
$\overline{\xx}\in\cldom{\psi^{\dualstar}}$.
The claim now follows immediately from
Theorem~\ref{thm:psi-subgrad-conds}.
\underline{
(\ref{cor:psi-subgrad-conds:b})
$\Rightarrow$
(\ref{cor:psi-subgrad-conds:c})}:
Immediate from Theorem~\ref{thm:psi-subgrad-conds}.
\underline{
(\ref{cor:psi-subgrad-conds:c})
$\Rightarrow$
(\ref{cor:psi-subgrad-conds:a})}:
Immediate from Theorem~\ref{thm:psi-subgrad-conds-part2}.
\qed
Returning to our earlier study of subgradients of a function
$F:\extspac{n}\rightarrow\overline{\R}$,
we can apply Theorem~\ref{thm:psi-subgrad-conds-part2} to
$\psi={F^*}$ to obtain an analogous result as corollary:
\begin{corollary} \label{cor:adsubdif-implies-fenchel}
Let $F:\extspac{n}\rightarrow\overline{\R}$, and let $\overline{\xx}\in\extspac{n}$ and
${\bf u}\in{\R^n}$.
Assume the following hold:
\begin{itemize}
\item
$\overline{\xx}\in\adsubdifFstar{{\bf u}}$.
\item
$-F^{*\dualstar}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable.
\item
$-F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u} = \overline{\xx}\cdot{\bf u} - F^{*\dualstar}(\overline{\xx})$.
\end{itemize}
Then
${F^*}({\bf u}) = -F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u}$.
\end{corollary}
\proof
Our assumptions and
Theorem~\ref{thm:psi-subgrad-conds-part2},
with $\psi={F^*}$, yield
\[
{F^*}({\bf u})
=
\overline{\xx}\cdot{\bf u} - F^{*\dualstar}(\overline{\xx})
=
-F(\overline{\xx}) \plusd \overline{\xx}\cdot{\bf u},
\]
as claimed.
\qed
For the case that $-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable
and that also $-F^{*\dualstar}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable,
we summarize the various results in the following corollary
which proves the equivalence of the conditions discussed above.
Note that these conditions are always satisfied when
$\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$, in which case the condition that
$\overline{\xx}\cdot{\bf u}-F(\overline{\xx})=\overline{\xx}\cdot{\bf u}-F^{*\dualstar}(\overline{\xx})$
(appearing below in part~(\ref{cor:asub-summable-summary:c}))
is equivalent to
the simpler closedness condition
$F(\overline{\xx})=F^{*\dualstar}(\overline{\xx})$.
Also, the assumption that
either ${F^*}({\bf u})>-\infty$ or $\overline{\xx}\in\cldom{F}$
is fairly minimal for such an equivalence to hold
(at least when
$\overline{\xx}\cdot{\bf u} - F(\overline{\xx}) = \overline{\xx}\cdot{\bf u} - F^{*\dualstar}(\overline{\xx})$)
since if
${F^*}({\bf u})=-\infty$ and $\overline{\xx}\not\in\cldom{F}$
then
${\bf u}\not\in\asubdifF{\overline{\xx}}$
(by Theorem~\ref{thm:asubdif-implies-adsubdif}),
but
$\overline{\xx}\in\adsubdifFstar{{\bf u}}$
(from its definition in \eqref{eqn:psi-subgrad:3a}).
Similar comments can be made about the analogous assumption appearing
in Corollary~\ref{cor:psi-subgrad-conds}.
\begin{corollary} \label{cor:asub-summable-summary}
Let $F:\extspac{n}\rightarrow\overline{\R}$, and let $\overline{\xx}\in\extspac{n}$ and
${\bf u}\in{\R^n}$.
Assume the following hold:
\begin{itemize}
\item
$-F(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable.
\item
$-F^{*\dualstar}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable.
\item
Either ${F^*}({\bf u})>-\infty$ or $\overline{\xx}\in\cldom{F}$.
\end{itemize}
Then the following are equivalent:
\begin{enumerate}
\item \label{cor:asub-summable-summary:a}
${F^*}({\bf u}) = \overline{\xx}\cdot{\bf u} - F(\overline{\xx})$.
\item \label{cor:asub-summable-summary:b}
${\bf u}\in\asubdifF{\overline{\xx}}$.
\item \label{cor:asub-summable-summary:c}
$\overline{\xx}\in\adsubdifFstar{{\bf u}}$
and
$\overline{\xx}\cdot{\bf u} - F(\overline{\xx}) = \overline{\xx}\cdot{\bf u} - F^{*\dualstar}(\overline{\xx})$.
\end{enumerate}
\end{corollary}
\proof
~
\underline{
(\ref{cor:asub-summable-summary:a})
$\Rightarrow$
(\ref{cor:asub-summable-summary:b})}:
Suppose ${F^*}({\bf u}) = \overline{\xx}\cdot{\bf u} - F(\overline{\xx})$.
Similar to the proof of Corollary~\ref{cor:psi-subgrad-conds},
if ${F^*}({\bf u})>-\infty$ then $F(\overline{\xx})<+\infty$.
Therefore, our assumption that
either ${F^*}({\bf u})>-\infty$ or $\overline{\xx}\in\cldom{F}$ actually implies
that $\overline{\xx}\in\cldom{F}$.
The claim then follows immediately from
Theorem~\ref{thm:fenchel-implies-subgrad}.
\underline{
(\ref{cor:asub-summable-summary:b})
$\Rightarrow$
(\ref{cor:asub-summable-summary:c})}:
This follows directly from
Theorem~\ref{thm:asubdif-implies-adsubdif}
since if
either $F(\overline{\xx})=F^{*\dualstar}(\overline{\xx})$ or if
$\overline{\xx}\cdot{\bf u}\in\{-\infty,+\infty\}$ then
we must have
$\overline{\xx}\cdot{\bf u} - F(\overline{\xx}) = \overline{\xx}\cdot{\bf u} - F^{*\dualstar}(\overline{\xx})$
(using our summability assumptions).
\underline{
(\ref{cor:asub-summable-summary:c})
$\Rightarrow$
(\ref{cor:asub-summable-summary:a})}:
Immediate from Corollary~\ref{cor:adsubdif-implies-fenchel}.
\qed
We turn next to when
$F$ is the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of a convex
function $f:{\R^n}\rightarrow\overline{\R}$, a primary focus of this manuscript.
In this case,
the astral subdifferential of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
and the astral dual subdifferential of ${f^*}$
are inverses over $\overline{\xx}\in\cldom{f}$
and ${\bf u}\in{\R^n}$ in the sense that
${\bf u}\in\asubdiffext{\overline{\xx}}$
if and only if
$\overline{\xx}\in\adsubdiffstar{{\bf u}}$ for all such pairs.
More precisely, we prove the following succinct and general
equivalence:
\begin{theorem} \label{thm:adif-fext-inverses}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex,
and let $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Then ${\bf u}\in\asubdiffext{\overline{\xx}}$
if and only if
$\overline{\xx}\in\adsubdiffstar{{\bf u}}$
and
$\overline{\xx}\in\cldom{f}$.
\end{theorem}
\proof
Note first that
$\cldom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}=\cldom{f}$
by Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2c}),
and ${{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}^{*}={f^*}$
by Proposition~\ref{pr:fextstar-is-fstar}.
We use these identities throughout the proof.
As a result,
Theorem~\ref{thm:asubdif-implies-adsubdif}
(with $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$)
immediately implies the ``only if'' part of the theorem's statement.
For the converse,
suppose for the rest of the proof that
$\overline{\xx}\in\adsubdiffstar{{\bf u}}$
and
$\overline{\xx}\in\cldom{f}$.
We aim to prove
${\bf u}\in\asubdiffext{\overline{\xx}}$.
From $\overline{\xx}\in\cldom{f}$, it follows
that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=f^{*\dualstar}(\overline{\xx})$
by
Theorem~\ref{thm:fext-neq-fdub},
and also that $f\not\equiv+\infty$, which in turn implies
${f^*}>-\infty$ (as can be seen from its definition,
\eqref{eq:fstar-def}).
Thus, if $-{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are summable,
then $-f^{*\dualstar}(\overline{\xx})$ and $\overline{\xx}\cdot{\bf u}$ are as well, so that
part~(\ref{cor:asub-summable-summary:c}) and
all the requisite conditions of
Corollary~\ref{cor:asub-summable-summary} are satisfied (with
$F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$), thereby implying
${\bf u}\in\asubdiffext{\overline{\xx}}$
in this case.
Therefore, we assume henceforth that $-{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$ and
$\overline{\xx}\cdot{\bf u}$ are \emph{not} summable.
Specifically, this means that
\begin{equation} \label{eq:thm:adif-fext-inverses:1}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
=\overline{\xx}\cdot{\bf u}\in\{-\infty,+\infty\}.
\end{equation}
We prove that
${\bf u}\in\asubdiffext{\overline{\xx}}$
by constructing a point $\overline{{\bf z}}\in\extspac{n+1}$ satisfying all the
conditions of
Proposition~\ref{pr:equiv-ast-subdif-defn}(\ref{pr:equiv-ast-subdif-defn:c})
(with $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$, and
where, as previously discussed, ${\bf P}$ is the matrix defined in
\eqref{eqn:homat-def}).
We begin by defining a matrix
${\bf A}\in{\R^{(n+1)\times (n+1)}}$ that will be used throughout the proof.
This matrix is identical to the
$(n+1)\times(n+1)$ identity matrix, except that the first $n$ entries
of the bottom row are equal to $\trans{{\bf u}}$.
Thus, in block form,
\[
{\bf A}
=
\left[
\begin{array}{ccc|c}
& & & \\
~ & {\bf I} & ~ & {\bf 0} \\
& & & \\
\hline
\rule{0pt}{2.5ex}
& \trans{{\bf u}} & & 1
\end{array}
\right]
\]
where ${\bf I}$ is the $n\times n$ identity matrix, and
${\bf 0}$, as usual, is the all-zeros vector in ${\R^n}$.
For $t=1,2,\ldots$,
let $y_t = \max\{-t, -{f^*}({\bf u})\}$.
Then $y_t\in{\mathbb{R}}$ (since ${f^*}>-\infty$),
$y_t\geq -{f^*}({\bf u})$,
and $y_t\rightarrow -{f^*}({\bf u})$.
Next, let $\overline{{\bf z}}_t={\bf A} \rpair{\overline{\xx}}{y_t}$, forming a sequence in
$\extspac{n+1}$.
By sequential compactness, this sequence must have a convergent
subsequence; by discarding all other elements, we can assume that the
entire sequece converges.
Let $\overline{{\bf z}}\in\extspac{n+1}$ be its limit; that is,
$\overline{{\bf z}}_t\rightarrow\overline{{\bf z}}$.
We proceed to prove that $\overline{{\bf z}}$ satisfies all the conditions of
Proposition~\ref{pr:equiv-ast-subdif-defn}(\ref{pr:equiv-ast-subdif-defn:c}).
First,
\[
{\bf P} \overline{{\bf z}}_t
=
{\bf P} {\bf A} \rpair{\overline{\xx}}{y_t}
=
{\bf P} \rpair{\overline{\xx}}{y_t}
=
\overline{\xx}.
\]
The second equality is because
${\bf P} {\bf A} = {\bf P}$, by a straightforward matrix calculation
(also using
Proposition~\ref{pr:h:4}(\ref{pr:h:4d})).
The last equality is by
Theorem~\ref{thm:homf}(\ref{thm:homf:aa}).
Since ${\bf P}\overline{{\bf z}}_t\rightarrow{\bf P}\overline{{\bf z}}$
(Proposition~\ref{pr:aff-cont}), it follows that
${\bf P}\overline{{\bf z}}=\overline{\xx}$.
For ${\bf w}\in{\R^n}$ and $a\in{\mathbb{R}}$,
we can compute $\overline{{\bf z}}_t \cdot \rpair{{\bf w}}{a}$, which
will be used repeatedly in the remainder of the proof:
\begin{eqnarray}
\overline{{\bf z}}_t \cdot \rpair{{\bf w}}{a}
&=&
\paren{{\bf A} \rpair{\overline{\xx}}{y_t}} \cdot \rpair{{\bf w}}{a}
\nonumber
\\
&=&
\rpair{\overline{\xx}}{y_t} \cdot
\paren{ {\trans{\amatu}} \rpair{{\bf w}}{a} }
\nonumber
\\
&=&
\rpair{\overline{\xx}}{y_t} \cdot \rpair{{\bf w} + a {\bf u}}{a}
\nonumber
\\
&=&
\overline{\xx}\cdot({\bf w} + a {\bf u}) + a y_t.
\label{eq:thm:adif-fext-inverses:4}
\end{eqnarray}
These equalities follow respectively from:
definition of $\overline{{\bf z}}_t$;
Proposition~\ref{pr:mat-mult-def};
straightforward matrix calculation;
and
Theorem~\ref{thm:homf}(\ref{thm:homf:a}).
In particular, setting ${\bf w}={\bf 0}$ and $a=1$,
\eqref{eq:thm:adif-fext-inverses:4}
shows that
\[
\overline{{\bf z}}_t \cdot \rpair{{\bf 0}}{1}
=
\overline{\xx}\cdot{\bf u} + y_t
=
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
\]
with the last equality following from
\eqref{eq:thm:adif-fext-inverses:1}.
Since
$\overline{{\bf z}}_t \cdot \rpair{{\bf 0}}{1} \rightarrow \overline{{\bf z}} \cdot \rpair{{\bf 0}}{1}$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})),
this implies that
$\overline{{\bf z}} \cdot \rpair{{\bf 0}}{1} = {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
Likewise, setting ${\bf w}={\bf u}$ and $a=-1$,
\eqref{eq:thm:adif-fext-inverses:4}
yields
\[
\overline{{\bf z}}_t \cdot \rpair{{\bf u}}{-1}
=
\overline{\xx}\cdot{\bf 0} - y_t = -y_t.
\]
Since
$y_t\rightarrow -{f^*}({\bf u})$,
and since
$\overline{{\bf z}}_t \cdot \rpair{{\bf u}}{-1} \rightarrow \overline{{\bf z}} \cdot \rpair{{\bf u}}{-1}$,
we conclude that
$\overline{{\bf z}} \cdot \rpair{{\bf u}}{-1} = {f^*}({\bf u})$.
Thus, $\overline{{\bf z}}$ satisfies all of the conditions of
Proposition~\ref{pr:equiv-ast-subdif-defn}(\ref{pr:equiv-ast-subdif-defn:c}),
except that it still remains to show that
$\overline{{\bf z}}\in\clepi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}$.
In fact,
$\clepi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}$ is the same as $\epibarbar{f}$
by Proposition~\ref{pr:wasthm:e:3}(\ref{pr:wasthm:e:3c})
(keeping in mind that we are now identifying points and sets in
$\extspac{n}\times{\mathbb{R}}$ with their homeomorphic images in
${{\cal M}_n}\subseteq\extspac{n+1}$).
Furthermore,
since $\epi f$ is a convex subset of ${\R^{n+1}}$ (since $f$ is convex),
its closure in $\extspac{n+1}$ is equal to its outer hull,
$\ohull(\epi f)$, by Theorem~\ref{thm:e:6}.
That is,
\[ \clepi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}=\epibarbar{f}=\ohull(\epi f). \]
Therefore, to complete the proof, it suffices to show that $\overline{{\bf z}}$ is
in $\ohull(\epi f)$.
To this end, we prove that every closed halfspace in $\extspac{n+1}$
that includes $\epi f$ must also include $\overline{{\bf z}}$.
Let $H$ be the closed halfspace
\[
H = \{ \overline{{\bf z}}' \in \extspac{n+1} :
\overline{{\bf z}}'\cdot\rpair{{\bf w}}{a} \leq b
\}
\]
for some ${\bf w}\in{\R^n}$ and $a, b\in{\mathbb{R}}$, and
assume $\epi f \subseteq H$.
Under this assumption, we aim to prove $\overline{{\bf z}}\in H$.
We proceed in cases based on the value of $a$.
Without loss of generality, we assume that $a\in\{-1,0,1\}$.
(Otherwise, if $a\neq 0$, we can divide ${\bf w}$, $a$ and $b$ by
$|a|$, resulting in the same halfspace, but now with $a\in\{-1,1\}$.)
Suppose first that $a=1$.
Actually, this case is impossible
under the assumption $\epi f\subseteq H$.
To see this,
let ${\bf x}'\in\dom f$.
Then for all $y'\in{\mathbb{R}}$ with $y'\geq f({\bf x}')$,
we have $\rpair{{\bf x}'}{y'}\in\epi f\subseteq H$,
which means
$ {\bf x}'\cdot{\bf w} + y' \leq b$.
However, this is a contradiction since this inequality cannot hold
for arbitrarily large values of $y'$
(with ${\bf x}'\cdot{\bf w}$ and $b$ in ${\mathbb{R}}$).
Thus, $a$ can only be in $\{-1,0\}$.
Suppose next that $a=0$.
Since $\overline{\xx}\in\cldom{f}$, there exists a sequence
$\seq{\rpair{{\bf x}'_t}{y'_t}}$ in $\epi f$
with ${\bf x}'_t\rightarrow\overline{\xx}$.
Since $\rpair{{\bf x}'_t}{y'_t}\in\epi f\subseteq H$,
we have
\begin{equation} \label{eq:thm:adif-fext-inverses:3}
{\bf x}'_t\cdot{\bf w}
=
\rpair{{\bf x}'_t}{y'_t}\cdot\rpair{{\bf w}}{0}
\leq b.
\end{equation}
Thus,
\[
\overline{{\bf z}}\cdot\rpair{{\bf w}}{0}
=
\overline{{\bf z}}\cdot(\trans{{\bf P}} {\bf w})
=
({\bf P} \overline{{\bf z}}) \cdot {\bf w}
=
\overline{\xx}\cdot {\bf w}
\leq
b.
\]
The equalities follow respectively from
a simple matrix calculation;
Proposition~\ref{pr:mat-mult-def};
and our argument above establishing that ${\bf P}\overline{{\bf z}}=\overline{\xx}$.
The inequality follows from
\eqref{eq:thm:adif-fext-inverses:3}
since
${\bf x}'_t\cdot{\bf w} \rightarrow \overline{\xx}\cdot{\bf w}$
(by Theorem~\ref{thm:i:1}(\ref{thm:i:1c})).
This proves $\overline{{\bf z}}\in H$ in this case.
Finally, suppose $a=-1$.
In this case, we have that for all
$\rpair{{\bf x}'}{y'}\in\epi f$,
\[
{\bf x}'\cdot{\bf w} - y'
=
\rpair{{\bf x}'}{y'} \cdot \rpair{{\bf w}}{-1}
\leq
b.
\]
Since this holds for all
$\rpair{{\bf x}'}{y'}\in\epi f$,
it follows from \eqref{eq:fstar-mod-def} that
${f^*}({\bf w})\leq b$;
in other words,
$\rpair{{\bf w}}{b}\in\epi {f^*}$.
For each $t$, we now have that
\begin{eqnarray*}
\overline{{\bf z}}_t \cdot \rpair{{\bf w}}{-1}
&=&
\overline{\xx}\cdot({\bf w}-{\bf u}) - y_t
\\
&=&
(\overline{\xx}\cdot({\bf w}-{\bf u}) - b) - y_t + b
\\
&\leq&
-{f^*}({\bf u}) - y_t + b
\\
&\leq&
b.
\end{eqnarray*}
The first equality is a direct application of
\eqref{eq:thm:adif-fext-inverses:4}
(with $a=-1$).
The first inequality follows from
the definition of astral dual subgradient
(\eqref{eqn:psi-subgrad:3a}) since
$\overline{\xx}\in\adsubdiffstar{{\bf u}}$, and since, as just shown,
$\rpair{{\bf w}}{b}\in\epi {f^*}$.
The last inequality is because $y_t\geq -{f^*}({\bf u})$ by construction.
Thus, each $\overline{{\bf z}}_t\in H$, implying their limit, $\overline{{\bf z}}$, is also in
$H$, being a closed halfspace.
Therefore, $\overline{{\bf z}}\in\ohull(\epi f)=\clepi{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}}$,
completing the proof.
\qed
Theorem~\ref{thm:adif-fext-inverses} is no longer true in general
if the requirement that $\overline{\xx}\in\cldom{f}$ is removed.
In other words, there exist cases in which
$\overline{\xx}\in\adsubdiffstar{{\bf u}}$
but
${\bf u}\not\in\asubdiffext{\overline{\xx}}$.
For example, consider the convex function $f$ given in
\eqref{eqn:ex:1} whose conjugate is given in \eqref{eqn:ex:1:conj},
and let $\overline{\xx}=\limray{{\bf e}_1}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax(-{\bf e}_2)$ and
${\bf u}={\bf 0}$.
Then
$\overline{\xx}\in\adsubdiffstar{{\bf u}}$,
as can be seen
by checking \eqref{eqn:psi-subgrad:3},
noting that if ${f^*}({\bf u}')<+\infty$ then
${\bf u}'\cdot{\bf e}_1=-1<0$ so that $\overline{\xx}\cdot({\bf u}'-{\bf u})=-\infty$.
On the other hand,
$\overline{\xx}\not\in\cldom{f}$
(since the open set
$\{\overline{\xx}'\in\extspac{n} : \overline{\xx}'\cdot{\bf e}_2 < 0\}$
includes $\overline{\xx}$ but is disjoint from $\dom f$).
Therefore,
${\bf u}\not\in\asubdiffext{\overline{\xx}}$
by Theorem~\ref{thm:adif-fext-inverses}.
(Alternatively, we can see ${\bf u}={\bf 0}$ is not in
$\asubdiffext{\overline{\xx}}$
using Proposition~\ref{pr:asub-zero-is-min}
since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not minimized at $\overline{\xx}$.)
\subsection{Dual subdifferentials are never empty}
\label{sec:dual-subdiff-not-empty}
In standard convex analysis, a convex function has a nonempty
subdifferential at every point in the relative interior of its
effective domain
\citep[Theorem~23.4]{ROC}.
Nevertheless,
it is possible for the function to have no
subgradients at other points, as indeed will be the case at all points
outside its effective domain, and possibly at some or all of its
relative boundary points.
In contrast, for any convex function $\psi:{\R^n}\rightarrow\overline{\R}$,
the astral dual subdifferential is nonempty at \emph{every}
point:
\begin{theorem} \label{thm:adsubdiff-nonempty}
Let $\psi:{\R^n}\rightarrow\overline{\R}$ be convex, and let ${\bf u}\in{\R^n}$.
Then $\psi$ has an astral dual subgradient at ${\bf u}$;
that is, $\adsubdifpsi{{\bf u}}\neq\emptyset$.
\end{theorem}
\proof
We first prove the theorem when ${\bf u}={\bf 0}$.
We then show how the general case follows easily from this special
case.
Proof is by induction on the dimension of the effective domain of
$\psi$.
More precisely, we show by induction on $k=0,\ldots,n$ that for every
convex function $\psi:{\R^n}\rightarrow\overline{\R}$, if $\dom\psi$ is included
in some linear subspace $L\subseteq{\R^n}$ of dimension at most $k$, then
$\adsubdifpsi{{\bf 0}}\neq\emptyset$; that is, there exists
$\overline{\xx}\in\extspac{n}$ such that
\begin{equation} \label{eqn:thm:adsubdiff-nonempty:1}
-\psi({\bf 0})
\geq
-\psi({\bf w}) \plusd \overline{\xx}\cdot{\bf w}
\end{equation}
for all ${\bf w}\in{\R^n}$
(using the form of the definition of astral dual subdifferential given
in \eqref{eqn:psi-subgrad:3}).
In the base case that $k=0$, suppose
$\psi:{\R^n}\rightarrow\overline{\R}$ is convex and that
$\dom\psi \subseteq \{{\bf 0}\}$.
Then $\psi({\bf w})=+\infty$ for all ${\bf w}\in{\R^n}\setminus\{{\bf 0}\}$,
so \eqref{eqn:thm:adsubdiff-nonempty:1} holds, for all ${\bf w}\in{\R^n}$,
with $\overline{\xx}={\bf 0}$.
Note that this includes the case $\psi\equiv+\infty$.
For the inductive step with $k>0$, suppose the claim holds for $k-1$,
and let
$\psi:{\R^n}\rightarrow\overline{\R}$ be convex.
Suppose $\dom\psi \subseteq L$ where $L\subseteq{\R^n}$ is a linear
subspace with dimension $k$.
We consider a few cases:
First, if $\psi\equiv +\infty$, then the claim holds, as argued as
part of the base case.
Next, consider the case that ${\bf 0}\in\ri(\dom \psi)$.
If $\psi$ is improper, then this implies $\psi({\bf 0})=-\infty$
\citep[Theorem~7.2]{ROC},
and so \eqref{eqn:thm:adsubdiff-nonempty:1} holds, for all
${\bf w}\in{\R^n}$, with $\overline{\xx}={\bf 0}$ (or really any choice of $\overline{\xx}$).
Otherwise, if $\psi$ is proper, then it must have a standard
subgradient at ${\bf 0}$, being in $\ri(\dom \psi)$
\citep[Theorem~23.4]{ROC}.
Thus, there exists ${\bf q}\in{\R^n}$ such that
$\psi({\bf w}) \geq \psi({\bf 0}) + {\bf q}\cdot{\bf w}$
for all ${\bf w}\in{\R^n}$, which implies
\eqref{eqn:thm:adsubdiff-nonempty:1} with $\overline{\xx}={\bf q}$.
In the final case, suppose
${\bf 0}\not\in\ri(\dom \psi)$ and that $\psi\not\equiv+\infty$.
Then $\dom\psi$ is nonempty and convex, and its relative interior is
disjoint from $\ri \{{\bf 0}\} = \{{\bf 0}\}$, which means there exists a
hyperplane that properly separates $\{{\bf 0}\}$ from $\dom\psi$
\cite[Theorem~11.3]{ROC}.
That is, there exists ${\bf v}\in{\R^n}$ and $b\in{\mathbb{R}}$ such that
${\bf w}\cdot{\bf v} \leq b$ for all ${\bf w}\in\dom\psi$
and
$0 = {\bf 0}\cdot{\bf v} \geq b$.
Further, since the separation is proper,
one of these inequalities is strict for at least one point.
If the former inequality is strict at some point, then
there exists ${\bf y}\in \dom\psi$ with
${\bf y}\cdot{\bf v} < b\leq 0$.
Otherwise, if the latter inequality is strict, then
$b<0$, which implies ${\bf w}\cdot{\bf v}\leq b<0$ for all ${\bf w}\in\dom\psi$.
(And as already noted, $\dom\psi$ is nonempty.)
Thus, in either case, we conclude that
${\bf w}\cdot{\bf v} \leq 0$ for all ${\bf w}\in\dom\psi$,
and that there exists at least one point ${\bf y}\in\dom\psi$ with
${\bf y}\cdot{\bf v} < 0$.
Let us define the function $\rho:{\R^n}\rightarrow\overline{\R}$ as
\[
\rho({\bf w})
=
\left\{ \begin{array}{cl}
\psi({\bf w}) & \mbox{if ${\bf w}\cdot{\bf v} = 0$} \\
+\infty & \mbox{otherwise.}
\end{array} \right.
\]
This function is convex
since
\[
\epi \rho = (\epi \psi)
\cap \{ \rpair{{\bf w}}{y} \in {\R^n}\times{\mathbb{R}} : {\bf w}\cdot{\bf v}=0 \},
\]
implying its epigraph is an intersection of two convex sets, and is
therefore convex.
Let $M=\{{\bf w}\in{\R^n} : {\bf w}\cdot{\bf v}=0\}$, and let $L'=L\cap M$.
Clearly,
$\dom\rho\subseteq\dom\psi\subseteq L$ and
$\dom\rho\subseteq M$,
so $\dom\rho\subseteq L'$.
However, ${\bf y}\in\dom\psi\subseteq L$ but
${\bf y}\not\in M$ (since ${\bf y}\cdot{\bf v}<0$), so
$L'\subseteq L$ but $L'\neq L$, and therefore $L'$ is a linear
subspace of dimension strictly less than $L$.
Since $\dim L' < k$, we can apply our inductive hypothesis which
implies that there exists $\overline{{\bf z}}\in\extspac{n}$ such that
\[
-\rho({\bf 0})
\geq
-\rho({\bf w}) \plusd \overline{{\bf z}}\cdot{\bf w}
\]
for all ${\bf w}\in{\R^n}$.
Let $\overline{\xx}=\limray{{\bf v}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}$, which
we now argue satisfies \eqref{eqn:thm:adsubdiff-nonempty:1}.
Let ${\bf w}\in{\R^n}$.
If ${\bf w}\cdot{\bf v}=0$ then $\rho({\bf 0})=\psi({\bf 0})$ and
$\rho({\bf w})=\psi({\bf w})$.
Further, in this case,
$\overline{\xx}\cdot{\bf w}=\limray{{\bf v}}\cdot{\bf w}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\cdot{\bf w} = \overline{{\bf z}}\cdot{\bf w}$.
Therefore,
\[
-\psi({\bf 0})
=
-\rho({\bf 0})
\geq
-\rho({\bf w})\plusd \overline{{\bf z}}\cdot{\bf w}
=
-\psi({\bf w})\plusd \overline{\xx}\cdot{\bf w}.
\]
If ${\bf w}\cdot{\bf v}<0$ then
$\overline{\xx}\cdot{\bf w}=\limray{{\bf v}}\cdot{\bf w}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\overline{{\bf z}}\cdot{\bf w} = -\infty$,
implying
\eqref{eqn:thm:adsubdiff-nonempty:1}.
And if ${\bf w}\cdot{\bf v}>0$ then ${\bf w}\not\in\dom\psi$, so
$\psi({\bf w})=+\infty$, again implying
\eqref{eqn:thm:adsubdiff-nonempty:1}.
This completes the induction and the proof for the case ${\bf u}={\bf 0}$.
In the general case, suppose ${\bf u}\in{\R^n}$, not necessarily the origin.
Let $\psi'({\bf w})=\psi({\bf w}+{\bf u})$ for all ${\bf w}\in{\R^n}$, which is convex.
Then there exists $\overline{\xx}\in\extspac{n}$ satisfying
\eqref{eqn:thm:adsubdiff-nonempty:1} for $\psi'$, as argued above.
Thus,
\[
-\psi({\bf u})
=
-\psi'({\bf 0})
\geq
-\psi'({\bf w}')\plusd\overline{\xx}\cdot{\bf w}'
=
-\psi({\bf w}'+{\bf u})\plusd\overline{\xx}\cdot{\bf w}'
\]
for all ${\bf w}'\in{\R^n}$.
Replacing ${\bf w}'$ by ${\bf w}-{\bf u}$ yields
\[
-\psi({\bf u})
\geq
-\psi({\bf w})\plusd\overline{\xx}\cdot({\bf w}-{\bf u})
\]
for all ${\bf w}\in{\R^n}$,
which is the same as \eqref{eqn:psi-subgrad:3},
thereby proving that $\overline{\xx}$ is in $\adsubdifpsi{{\bf u}}$,
which
therefore is nonempty.
\qed
For any convex function $\psi:{\R^n}\rightarrow\overline{\R}$,
Theorem~\ref{thm:adsubdiff-nonempty} shows there exists an
astral dual subgradient $\overline{\xx}\in\adsubdifpsi{{\bf 0}}$.
Combined with
Proposition~\ref{pr:asub-zero-is-min}
and
Corollary~\ref{cor:psi-subgrad-conds},
this implies that $\overline{\xx}$ minimizes $\psi^{\dualstar}$,
assuming $\psi({\bf 0})>-\infty$.
Thus, every function $\psi^{\dualstar}$ that is an astral dual conjugate of
some convex function $\psi:{\R^n}\rightarrow\overline{\R}$
with $\psi({\bf 0})>-\infty$
must have a minimizer
in $\extspac{n}$.
As a special case,
if $\psi={f^*}$ for any function $f:{\R^n}\rightarrow\overline{\R}$
(with $f\not\equiv+\infty$), this implies
$f^{*\dualstar}$ must have a minimizer.
If $f$ has all reductions closed, then ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}=f^{*\dualstar}$, so ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ must have a
minimizer as well; of course, this was already known (and with fewer
assumptions) from
Proposition~\ref{pr:fext-min-exists}.
As noted earlier,
if ${\bf u}\not\in\dom\psi$, then $\psi$ has no standard subgradient at
${\bf u}$, but nevertheless has an astral dual subgradient, by
Theorem~\ref{thm:adsubdiff-nonempty}.
To get a sense about such subgradients,
for an arbitrary convex set $S\subseteq{\R^n}$,
consider the case that $\psi=\indf{S}$
is the indicator function on $S$ defined in \eqref{eq:indf-defn},
and suppose ${\bf u}\not\in S$.
Then
Theorem~\ref{thm:adsubdiff-nonempty}
ensures that there exists an astral dual subgradient
$\overline{\xx}\in\extspac{n}$ of $\psi$ at ${\bf u}$, which implies, by
\eqref{eqn:psi-subgrad:3},
that, for all ${\bf w}\in S$,
\begin{equation} \label{eqn:subdiff-for-ind-fcn}
-\infty
=
-\psi({\bf u})
\geq
-\psi({\bf w})\plusd \overline{\xx}\cdot({\bf w}-{\bf u})
=
\overline{\xx}\cdot({\bf w}-{\bf u}).
\end{equation}
We can write
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax {\bf q}$ for some matrix
${\bf V}\in{\mathbb{R}}^{n\times k}$
and ${\bf q}\in{\R^n}$.
Then \eqref{eqn:subdiff-for-ind-fcn}
implies $\overline{\xx}\cdot({\bf w}-{\bf u})=-\infty$ for ${\bf w}\in S$, and thus that
${\bf V} {\bar{\bm{\omega}}}\cdot({\bf w}-{\bf u}) = -\infty$
(by Proposition~\ref{pr:vtransu-zero}).
This latter condition can be expressed in terms of
{lexicographic ordering}.
Specifically, for any two vectors ${\bf a},{\bf b}\in{\R^k}$,
we say that ${\bf a}$ is
\emph{lexicographically less than}
${\bf b}$, written ${\bf a} <_L {\bf b}$,
if ${\bf a}\neq{\bf b}$ and if, in the first entry where they differ, ${\bf a}$ is
less than ${\bf b}$;
that is, if
for some $j\in\{1,\ldots,k\}$,
$a_j < b_j$ and
$a_i = b_i$ for $i=1,\ldots,j-1$.
Then it can be checked that the condition above, that
${\bf V} {\bar{\bm{\omega}}}\cdot({\bf w}-{\bf u}) = -\infty$,
holds if and only if
$\trans{{\bf V}}({\bf w}-{\bf u}) <_L {\bf 0}$,
or equivalently,
$\trans{{\bf V}}{\bf w} <_L \trans{{\bf V}}{\bf u}$.
In this sense, since this holds for all ${\bf w}\in S$,
the set $S$ is
\emph{lexicographically separated} from the point ${\bf u}$.
Thus, for a convex indicator function $\psi$, any astral dual
subgradient at any point ${\bf u}\not\in\dom\psi$
lexicographically separates ${\bf u}$ from $\dom\psi$,
in the sense just described.
(Actually, a similar argument shows that this is true for any
proper, convex function $\psi:{\R^n}\rightarrow\overline{\R}$,
not just indicator functions.)
Therefore,
Theorem~\ref{thm:adsubdiff-nonempty} shows as a corollary that
every convex set $S\subseteq{\R^n}$ can be
lexicographically separated from any point ${\bf u}\not\in S$.
This is a special case of the more general results on
lexicographic separation proved by
\citet{lexicographic_separation}.
\subsection{Subdifferentials and continuity}
\label{sec:subdif:cont}
A convex function $f:{\R^n}\rightarrow\overline{\R}$ is minimized at a
point ${\bf x}\in{\R^n}$ if and only if ${\bf 0}$ is a subgradient of $f$ at
${\bf x}$, and in particular, if the gradient $\nabla f({\bf x})$ exists and is
equal to ${\bf 0}$.
Therefore, to minimize $f$ numerically, it is natural to construct a
sequence $\seq{{\bf x}_t}$ in ${\R^n}$ whose gradients converge to ${\bf 0}$,
that is, for which $\nabla f({\bf x}_t)\rightarrow{\bf 0}$.
If this is possible, and if the sequence converges to some point
${\bf x}\in{\R^n}$, then indeed, ${\bf x}$ must minimize $f$;
that is, $f({\bf x}_t)\rightarrow \inf f$,
assuming $f$ is closed and proper
\citep[Theorem~24.4]{ROC}.
Moreover, even if the sequence $\seq{{\bf x}_t}$ does not converge but
nevertheless remains within a bounded region of ${\R^n}$, then an
argument can again be made that $f({\bf x}_t)\rightarrow \inf f$.
Since driving the gradient of a function to ${\bf 0}$ seems so closely
connected to minimization, especially for convex functions,
one might expect that it should also be effective as a means of
minimizing the function --- for any sequence, not just
bounded sequences.
In other words, for a convex function
$f:{\R^n}\rightarrow\overline{\R}$ and sequence $\seq{{\bf x}_t}$ in ${\R^n}$, we might
expect that if $\nabla f({\bf x}_t)\rightarrow{\bf 0}$ then
$f({\bf x}_t)\rightarrow \inf f$.
However, this is false in general, even for a convex function with
many favorable properties.
As an example, let $f$ be the function defined in
\eqref{eqn:curve-discont-finiteev-eg}.
This function is convex,
closed, proper, continuous everywhere, finite everywhere, and
nonnegative everywhere.
It is also continuously differentiable everywhere except along the ray
$\{\trans{[0,x_2]} : x_2\leq 0\}$,
a part of the space that is
far from the sequences we will be considering.
Let
${\bf x}_t=\trans{[t,\,t^{3/2}]}$, for $t=1,2,\ldots$.
Then it can be checked that $f$'s gradient at each ${\bf x}_t$ is
$ \nabla f({\bf x}_t) = \trans{[ 2/\sqrt{t},\, -1/t]} $,
so $\nabla f({\bf x}_t)\rightarrow{\bf 0}$.
Nevertheless, as seen earlier,
$f({\bf x}_t)=\sqrt{t}\rightarrow +\infty$.
Thus, the gradients are converging to ${\bf 0}$, but the function values
are becoming infinite, as far away from $\inf f = 0$ as is possible.
It is no coincidence that this same function $f$ was used earlier
in Section~\ref{sec:continuity}
as an example of a function whose extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is discontinuous.
The same sequence was also used in that discussion
where it was seen that
${\bf x}_t\rightarrow \overline{\xx}=\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$,
and that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is discontinuous at $\overline{\xx}$.
Indeed, there is a close connection between continuity in the astral
space and convergence of gradients:
As will be seen below, in general,
for a convex function $f:{\R^n}\rightarrow\overline{\R}$,
if $\seq{{\bf x}_t}$ is a sequence in ${\R^n}$ converging to a point
$\overline{\xx}\in\extspac{n}$ where ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous
(and also with ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$),
and if $\nabla f({\bf x}_t)\rightarrow{\bf 0}$ then
$f({\bf x}_t)\rightarrow \inf f$.
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous at $\overline{\xx}$, then this statement need not
hold, as just seen in the preceding example.
We focus particularly on convergence of subgradients.
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex, and suppose
$\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ are sequences in ${\R^n}$
with each ${\bf u}_t$ a subgradient of $f$ at ${\bf x}_t$
so that ${\bf u}_t\in\partial f({\bf x}_t)$.
Suppose also that ${\bf x}_t\rightarrow\overline{\xx}$ and
${\bf u}_t\rightarrow{\bf u}$ for some $\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
We seek conditions that guarantee ${\bf u}\in\asubdiffext{\overline{\xx}}$
so that ${\bf u}$, the limit of the subgradients ${\bf u}_t$,
will itself be a subgradient of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at $\overline{\xx}$,
the limit of the points ${\bf x}_t$.
Indeed, from standard convex analysis, this is known to be the case if
$\overline{\xx}={\bf x}\in{\R^n}$, provided $f$ is closed and proper
(via \citet[Theorem~24.4]{ROC} and
Proposition~\ref{pr:asubdiffext-at-x-in-rn}).
In general, the sequence of values ${\bf x}_t\cdot{\bf u}_t$ need \emph{not}
converge to $\overline{\xx}\cdot{\bf u}$.
(For example, for the function $f$ in
\eqref{eqn:curve-discont-finiteev-eg} and sequence $\seq{{\bf x}_t}$
discussed above, setting ${\bf u}_t=\nabla f({\bf x}_t)$,
it can be checked that ${\bf x}_t\cdot{\bf u}_t=\sqrt{t}\rightarrow+\infty$,
but ${\bf u}=\lim {\bf u}_t={\bf 0}$ so $\overline{\xx}\cdot{\bf u}=0$.)
The convergence properties of the sequence $\seq{{\bf x}_t\cdot{\bf u}_t}$
turn out to be closely connected to the continuity of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at
$\overline{\xx}$, as we show in the next few theorems.
We show first that the closer ${\bf x}_t\cdot{\bf u}_t$ comes to
$\overline{\xx}\cdot{\bf u}$, the closer will the function values $f({\bf x}_t)$ get to
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$:
\begin{theorem} \label{thm:limsup-beta-bnd}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Let $\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ be sequences in ${\R^n}$
with each ${\bf u}_t\in\partial f({\bf x}_t)$, and with
${\bf x}_t\rightarrow\overline{\xx}$ and ${\bf u}_t\rightarrow{\bf u}$ for some
$\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Assume $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$ and
suppose $\limsup {\bf x}_t\cdot{\bf u}_t \leq \beta$
for some $\beta\in{\mathbb{R}}$.
Then
\begin{eqnarray}
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
&\leq&
\liminf f({\bf x}_t)
\label{thm:limsup-beta-bnd:s1}
\\
&\leq&
\limsup f({\bf x}_t)
\nonumber
\\
&\leq&
-{f^*}({\bf u}) + \beta
\label{thm:limsup-beta-bnd:s3}
\\
&\leq&
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) + \beta - \overline{\xx}\cdot{\bf u}.
\label{thm:limsup-beta-bnd:s4}
\end{eqnarray}
\end{theorem}
\proof
\eqref{thm:limsup-beta-bnd:s1} is immediate from
\eqref{eq:e:7}
since ${\bf x}_t\rightarrow\overline{\xx}$.
\eqref{thm:limsup-beta-bnd:s4} follows from
\eqref{eqn:ast-fenchel} (applied to $F={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
and using Proposition~\ref{pr:fextstar-is-fstar}),
and since $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$.
It remains to prove \eqref{thm:limsup-beta-bnd:s3}.
For all $t$, ${\bf u}_t\in\partial f({\bf x}_t)$ implies that
$-f({\bf x}_t)+{\bf x}_t\cdot{\bf u}_t={f^*}({\bf u}_t)$.
(This follows from \citet[Theorem~23.5]{ROC}, which is only stated for convex
functions that are also proper, but the same proof shows that
${\bf u}\in\partial f({\bf x})$ if and only if
$-f({\bf x})+{\bf x}\cdot{\bf u}={f^*}({\bf u})$
for any convex function $f:{\R^n}\rightarrow\overline{\R}$
and for all ${\bf x},{\bf u}\in{\R^n}$.)
Let $\epsilon>0$.
Then for all $t$ sufficiently large,
${\bf x}_t\cdot{\bf u}_t\leq\beta+\epsilon$.
Together, these facts imply
\begin{eqnarray*}
\limsup f({\bf x}_t)
&=&
\limsup (-{f^*}({\bf u}_t) + {\bf x}_t\cdot{\bf u}_t)
\\
&\leq&
\limsup -{f^*}({\bf u}_t) + \beta+\epsilon
\\
&\leq&
-{f^*}({\bf u}) + \beta+\epsilon.
\end{eqnarray*}
The last line is because
the function ${f^*}$, being a conjugate, is closed and therefore
lower semicontinuous so that
$\liminf {f^*}({\bf u}_t)\geq {f^*}({\bf u})$
since ${\bf u}_t\rightarrow{\bf u}$.
Since this holds for all $\epsilon>0$, this proves
\eqref{thm:limsup-beta-bnd:s3}.
\qed
Taking $\beta=\overline{\xx}\cdot{\bf u}$, Theorem~\ref{thm:limsup-beta-bnd}
immediately implies that, under the same assumptions,
if $\limsup {\bf x}_t\cdot{\bf u}_t \leq \overline{\xx}\cdot{\bf u}$
then
${\bf u}$ must be an astral subgradient of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at $\overline{\xx}$, and also
the sequence of values $f({\bf x}_t)$ must converge to ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$:
\begin{corollary} \label{cor:xuleqxbaru-then-subgrad}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
Let $\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ be sequences in ${\R^n}$
with each ${\bf u}_t\in\partial f({\bf x}_t)$, and with
${\bf x}_t\rightarrow\overline{\xx}$ and ${\bf u}_t\rightarrow{\bf u}$ for some
$\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Assume $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$ and
that $\limsup {\bf x}_t\cdot{\bf u}_t \leq \overline{\xx}\cdot{\bf u}$.
Then
$f({\bf x}_t)\rightarrow {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$
and
${\bf u}\in\asubdiffext{\overline{\xx}}$.
\end{corollary}
\proof
Given the stated assumptions,
we can immediately apply Theorem~\ref{thm:limsup-beta-bnd}
with $\beta=\overline{\xx}\cdot{\bf u}$, yielding
\[
{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})
=
\lim f({\bf x}_t)
=
-{f^*}({\bf u}) + \overline{\xx}\cdot{\bf u}.
\]
That ${\bf u}\in\asubdiffext{\overline{\xx}}$ now follows directly from
Theorem~\ref{thm:fenchel-implies-subgrad},
using $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$.
\qed
On the other hand,
Theorem~\ref{thm:limsup-beta-bnd} also provides
asymptotic lower bounds on the sequence ${\bf x}_t\cdot{\bf u}_t$,
showing that
$\liminf {\bf x}_t\cdot{\bf u}_t\geq\overline{\xx}\cdot{\bf u}$ if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})>-\infty$,
and that ${\bf x}_t\cdot{\bf u}_t\rightarrow+\infty$ if
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$:
\begin{corollary} \label{cor:xugeqxbaru}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and not identically $+\infty$.
Let $\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ be sequences in ${\R^n}$
with each ${\bf u}_t\in\partial f({\bf x}_t)$, and with
${\bf x}_t\rightarrow\overline{\xx}$ and ${\bf u}_t\rightarrow{\bf u}$ for some
$\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Assume $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})>-\infty$.
Then $\liminf {\bf x}_t\cdot{\bf u}_t \geq \overline{\xx}\cdot{\bf u}$.
If, in addition, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$, then
${\bf x}_t\cdot{\bf u}_t\rightarrow+\infty$.
\end{corollary}
\proof
We consider first the case that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\in{\mathbb{R}}$.
Suppose, by way of contradiction,
that $\liminf {\bf x}_t\cdot{\bf u}_t < \overline{\xx}\cdot{\bf u}$.
Then there exists $\epsilon>0$ and infinitely many values of $t$ for
which
${\bf x}_t\cdot{\bf u}_t \leq \overline{\xx}\cdot{\bf u} - \epsilon$.
By discarding all other sequence elements, we can assume that this
holds for all values of $t$.
We therefore can apply
Theorem~\ref{thm:limsup-beta-bnd}
with $\beta=\overline{\xx}\cdot{\bf u} - \epsilon$.
However, this yields
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) \leq {}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) - \epsilon$,
an obvious contradiction.
Therefore,
$\liminf {\bf x}_t\cdot{\bf u}_t \geq \overline{\xx}\cdot{\bf u}$
if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\in{\mathbb{R}}$.
Next,
consider the case that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$.
Suppose ${\bf x}_t\cdot{\bf u}_t \not\rightarrow +\infty$.
Then there exists $\beta\in{\mathbb{R}}$ such that
${\bf x}_t\cdot{\bf u}_t\leq\beta$ for
infinitely many values of $t$.
As before, we can discard all other sequence elements, so that this
holds for all values of $t$.
We can again apply
Theorem~\ref{thm:limsup-beta-bnd}
with this choice of $\beta$, yielding
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx}) \leq -{f^*}({\bf u}) + \beta$.
Since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$, this implies ${f^*}({\bf u})=-\infty$.
But this is a contradiction since we assumed $f\not\equiv+\infty$,
implying ${f^*}>-\infty$.
Therefore, $\lim {\bf x}_t\cdot{\bf u}_t = +\infty \geq \overline{\xx}\cdot{\bf u}$
if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=+\infty$.
\qed
If ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=-\infty$,
Corollary~\ref{cor:xugeqxbaru} need not hold;
in other words, it need not be the case that
$\liminf {\bf x}_t\cdot{\bf u}_t \geq \overline{\xx}\cdot{\bf u}$.
For example, suppose
\[
f(x)
=
\left\{ \begin{array}{cl}
-\ln x & \mbox{if $x > 0$} \\
+\infty & \mbox{otherwise,}
\end{array} \right.
\]
for $x\in{\mathbb{R}}$.
Let $x_t=t$, and $u_t=f'(x_t)=-1/t$, where $f'$ is the first
derivative of $f$.
Then $x_t \rightarrow +\infty$ and $u_t \rightarrow 0$.
Also, $x_t u_t = -1$ for all $t$, so
$x_t u_t \rightarrow -1 < 0 = (+\infty)\cdot 0$.
Corollary~\ref{cor:xuleqxbaru-then-subgrad} shows that, given our
other assumptions, to prove
${\bf u}$ is an astral subgradient of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at $\overline{\xx}$,
it suffices to show
$\limsup {\bf x}_t\cdot{\bf u}_t \leq \overline{\xx}\cdot{\bf u}$.
Indeed,
this will be the case if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$
and if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$, as we show next.
Thus, continuity in astral space provides a sufficient condition for a
sequence of subgradients to converge to an astral subgradient.
Our earlier counterexample shows that this need not be true in general
without continuity.
\begin{theorem} \label{thm:cont-subgrad-converg}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and proper.
Let $\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ be sequences in ${\R^n}$
with each ${\bf u}_t\in\partial f({\bf x}_t)$, and with
${\bf x}_t\rightarrow\overline{\xx}$ and ${\bf u}_t\rightarrow{\bf u}$ for some
$\overline{\xx}\in\extspac{n}$ and ${\bf u}\in{\R^n}$.
Assume $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$,
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$, and that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$.
Then $\limsup {\bf x}_t\cdot{\bf u}_t \leq \overline{\xx}\cdot{\bf u}$,
and therefore
${\bf u}\in\asubdiffext{\overline{\xx}}$.
\end{theorem}
\proof
Since $f$ is proper, $f({\bf x}_t)>-\infty$ for all $t$.
Further,
since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$, and since
$f({\bf x}_t)\rightarrow{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$ (by continuity of ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ at
$\overline{\xx}$),
we can have $f({\bf x}_t)=+\infty$ for at most finitely many values of
$t$.
Discarding these, we assume henceforth that
$f({\bf x}_t)\in{\mathbb{R}}$ for all $t$.
Let $h=\lsc f$, which implies $h=\cl f$, since $f$ is convex and
proper.
Also, $\overline{h}={}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ by
Proposition~\ref{pr:wasthm:h:2}(\ref{pr:wasthm:h:2aa}), so in
particular,
$\overline{h}(\overline{\xx})<+\infty$ and $\overline{h}$ is continuous at $\overline{\xx}$.
By Corollary~\ref{cor:cont-gen-char}, we therefore
can write $\overline{\xx}$ as
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}'$,
where ${\bf V}=[{\bf v}_1,\ldots,{\bf v}_k]$,
for some ${\bf v}_1,\ldots,{\bf v}_k\in\resc{h}$
and some ${\bf q}'\in{\R^n}$.
Without loss of generality, we assume
${\bf v}_1,\ldots,{\bf v}_k$ are linearly independent, by
Proposition~\ref{pr:g:2}.
Also, by letting ${\bf q}\in{\R^n}$ be the projection of ${\bf q}'$ onto the linear
subspace orthogonal to the ${\bf v}_i$'s, we can rewrite
$\overline{\xx}={\bf V}{\bar{\bm{\omega}}}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax{\bf q}$ where $\trans{{\bf V}}{\bf q}={\bf 0}$,
using Theorem~\ref{thm:i:2}.
By linear algebra, for each $t$, we can write
${\bf x}_t={\bf V} {\bf b}_t + {\bf q}_t$
for some ${\bf b}_t\in{\R^k}$ and ${\bf q}_t\in{\R^n}$ with
$\trans{{\bf V}}{\bf q}_t = {\bf 0}$.
Since ${\bf x}_t\rightarrow\overline{\xx}$, by
Theorem~\ref{thm:seq-rep},
$b_{t,i}\rightarrow+\infty$ for $i=1,\ldots,k$.
Therefore, by discarding at most finitely many elements from the
sequence, we can assume $b_{t,i}>0$ for all $t$ and all $i$.
Theorem~\ref{thm:seq-rep} also implies ${\bf q}_t\rightarrow{\bf q}$.
For each $t$, because ${\bf u}_t\in\partial f({\bf x}_t)$, it also holds that
$h({\bf x}_t)=f({\bf x}_t)$ and that ${\bf u}_t\in\partial h({\bf x}_t)$
\citep[Corollary~23.5.2]{ROC}.
For all ${\bf v}$ in $h$'s recession cone,
we claim that ${\bf v}\cdot{\bf u}_t\leq 0$
for all $t$.
To see this, let ${\bf v}\in\resc{h}$.
Then for each $t$,
\[
h({\bf x}_t)
\geq
h({\bf x}_t + {\bf v})
\geq
h({\bf x}_t) + {\bf v}\cdot{\bf u}_t
\]
where the first inequality is because ${\bf v}\in\resc{h}$, and
the second is because ${\bf u}_t\in\partial h({\bf x}_t)$.
Thus, ${\bf v}\cdot{\bf u}_t\leq 0$ since $h({\bf x}_t)=f({\bf x}_t)\in{\mathbb{R}}$.
In particular, for each $t$, this implies that
\[
{\bf x}_t\cdot{\bf u}_t
=
\sum_{i=1}^k b_{t,i} {\bf v}_i\cdot{\bf u}_t + {\bf q}_t\cdot{\bf u}_t
\leq
{\bf q}_t\cdot{\bf u}_t
\]
since ${\bf v}_i\in\resc{h}$ and $b_{t,i}>0$ for $i=1,\ldots,k$.
Since ${\bf q}_t\cdot{\bf u}_t\rightarrow{\bf q}\cdot{\bf u}$,
it follows that
\[
\limsup {\bf x}_t\cdot{\bf u}_t \leq {\bf q}\cdot{\bf u} = \overline{\xx}\cdot{\bf u},
\]
where the equality is from
Proposition~\ref{pr:vtransu-zero}, since $\overline{\xx}\cdot{\bf u}\in{\mathbb{R}}$.
That ${\bf u}\in\asubdiffext{\overline{\xx}}$ now follows immediately from
Corollary~\ref{cor:xuleqxbaru-then-subgrad}.
\qed
In particular, when ${\bf u}={\bf 0}$,
Theorem~\ref{thm:cont-subgrad-converg} shows that
for a convex, proper function $f:{\R^n}\rightarrow\overline{\R}$,
if ${\bf x}_t\rightarrow\overline{\xx}$ and
$\nabla f({\bf x}_t)\rightarrow{\bf 0}$
(or more generally, ${\bf u}_t\rightarrow{\bf 0}$ where
${\bf u}_t\in\partial f({\bf x}_t)$)
then $\overline{\xx}$ must minimize ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$,
and also $f({\bf x}_t)\rightarrow\inf f$.
Importantly,
this assumes ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$
and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$, or equivalently
(by Corollary~\ref{cor:cont-gen-char}), that
$\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
As a corollary, we can prove that even if the sequence $\seq{{\bf x}_t}$
does not converge, if the gradients or subgradients are converging to
zero (so that $\nabla f({\bf x}_t)\rightarrow{\bf 0}$ or
${\bf u}_t\rightarrow{\bf 0}$),
then the function values must be converging to the minimum of $f$,
so that $f({\bf x}_t)\rightarrow\inf f$.
For this, we need to assume that any convergent subsequence has a
limit in $\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$;
for instance, this will be the case if
$\limsup f({\bf x}_t) < +\infty$
and
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere.
As discussed earlier,
such convergence to $f$'s minimum
on a sequence whose gradients are converging to ${\bf 0}$
cannot in general be guaranteed
without assuming continuity.
\begin{theorem} \label{thm:subdiff-min-cont}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and proper.
Let $\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ be sequences in ${\R^n}$
with each ${\bf u}_t\in\partial f({\bf x}_t)$.
Assume ${\bf u}_t\rightarrow{\bf 0}$,
and also that the limit of every convergent subsequence of
$\seq{{\bf x}_t}$ is in $\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$
(as will be the case, for instance, if
$\limsup f({\bf x}_t)<+\infty$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere).
Then $f({\bf x}_t)\rightarrow \inf f$.
\end{theorem}
\proof
Suppose, contrary to the theorem's conclusion, that
$f({\bf x}_t)\not\rightarrow\inf f$.
Then there exists $\beta\in{\mathbb{R}}$ with $\beta>\inf f$ such that
$f({\bf x}_t)\geq \beta$ for infinitely many values of $t$.
By discarding all other sequence elements, we can assume that this
holds for all $t$.
Since $\extspac{n}$ is sequentially compact, the sequence $\seq{{\bf x}_t}$
must have a convergent subsequence which, by assumption, has a limit
$\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
By again discarding all elements not in this subsequence, we can
assume ${\bf x}_t\rightarrow\overline{\xx}$.
Since $\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$, by
Corollary~\ref{cor:cont-gen-char}, ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})<+\infty$ and ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$
is continuous at $\overline{\xx}$.
Applying Theorem~\ref{thm:cont-subgrad-converg} with
${\bf u}={\bf 0}$ therefore yields that
${\bf 0}\in\asubdiffext{\overline{\xx}}$, and thus that
${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})=\inf f$
(by Proposition~\ref{pr:asub-zero-is-min}).
Further, since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at $\overline{\xx}$,
$f({\bf x}_t)\rightarrow{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})$.
However, this is a contradiction since $f({\bf x}_t)\geq\beta>\inf f$
for all $t$.
We conclude that $f({\bf x}_t)\rightarrow \inf f$.
Finally, for an arbitrary sequence $\seq{{\bf x}_t}$ in ${\R^n}$,
suppose $\limsup f({\bf x}_t)<+\infty$ and that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is
continuous everywhere.
We argue, under these assumptions,
that the limit of every convergent subsequence of
$\seq{{\bf x}_t}$ is in $\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$.
Since $\limsup f({\bf x}_t)<+\infty$, there exists $b\in{\mathbb{R}}$ such
that $f({\bf x}_t)\leq b$ for all $t$ sufficiently large.
Therefore, if $\overline{\xx}\in\extspac{n}$ is the limit of some convergent
subsequence, then we must have ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}(\overline{\xx})\leq b<+\infty$.
Further, since ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere, it must be continuous
at $\overline{\xx}$, implying $\overline{\xx}\in\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ by
Corollary~\ref{cor:cont-gen-char}.
\qed
\subsection{Convergence of iterative methods}
\label{sec:iterative}
The preceding results can be applied to prove the convergence of
iterative methods for minimizing a function, as we now briefly
illustrate.
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex.
We consider methods that compute a sequence of iterates $\seq{{\bf x}_t}$
in ${\R^n}$ with the purpose of asymptotically minimizing $f$.
A classic example is \emph{gradient descent} in which ${\bf x}_1\in{\R^n}$ is
arbitrary, and each successive iterate is defined by
\begin{equation} \label{eqn:grad-desc-defn}
{\bf x}_{t+1} = {\bf x}_t - \eta_t \nabla f({\bf x}_t)
\end{equation}
for some step size $\eta_t > 0$.
Although certainly an important example, our aim is to develop
techniques that are broadly applicable well beyond gradient descent.
In analyzing the convergence of such iterative methods, it is very
common to assume that $f$ has a finite minimizer in ${\R^n}$,
and often also that we are effectively searching for a minimizer over
only a compact subset of ${\R^n}$.
(See, for example,
\citet[Chapters~9,~10,~11]{boyd_vandenberghe}.)
Depending on the problem setting, such assumptions may or may not be
reasonable.
A primary purpose of the current work, of course,
has been to develop a foundation that overcomes such difficulties and
that can be
applied without relying on such assumptions.
Indeed, as we have seen,
astral space is itself compact, and
the extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ of any convex function $f:{\R^n}\rightarrow\overline{\R}$
always has a minimizer that is attained
at some astral point in $\extspac{n}$.
Before exploring how astral methods can be used to prove general
convergence results, we first give examples in the next theorem
of how standard (that is, non-astral) techniques can sometimes be
applied to prove specialized convergence for particular algorithms, in
this case, subgradient descent, a generalization of gradient descent
in which
${\bf x}_{t+1} = {\bf x}_t - \eta_t {\bf u}_t$
where $\eta_t>0$ and ${\bf u}_t$ is any subgradient of $f$ at ${\bf x}_t$.
Part~(\ref{thm:grad-desc-converges:a}) of the theorem,
which is taken from
\citep[Lemma 2]{ziwei_poly_tail_losses},
proves convergence to the function's minimum
assuming a particular lower bound on how much $f({\bf x}_t)$
decreases on each iteration, an assumption that will be discussed
further below.
Under a different condition, which does not require per-iteration
progress,
part~(\ref{thm:grad-desc-converges:b})
proves that a weighted average of the iterates must
minimize $f$ via a similar proof.
This second part contrasts with standard results in the literature,
such as \citep{nesterov_book,Zhang04solvinglarge},
which typically require that the iterates ${\bf x}_t$ remain bounded.
\begin{theorem} \label{thm:grad-desc-converges}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and proper.
Let $\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ be sequences in ${\R^n}$
with $f({\bf x}_t)\in{\mathbb{R}}$,
${\bf u}_t\in\partial f({\bf x}_t)$,
and
${\bf x}_{t+1} = {\bf x}_t - \eta_t {\bf u}_t$,
for $t=1,2,\ldots$,
where $\eta_t>0$.
Assume
$\sum_{t=1}^\infty \eta_t = +\infty$.
Then the following hold:
\begin{enumerate}
\item \label{thm:grad-desc-converges:a}
Suppose
\begin{equation} \label{eqn:thm:grad-desc-converges:1}
f({\bf x}_{t+1}) \leq f({\bf x}_t) - \frac{\eta_t}{2} \norm{{\bf u}_t}^2
\end{equation}
for all $t$.
Then $f({\bf x}_t) \rightarrow \inf f$.
\item \label{thm:grad-desc-converges:b}
Suppose instead that
$\sum_{t=1}^\infty \eta_t^2 \|{\bf u}_t\|^2 < +\infty$.
For each $t$, let
\[
{\hat{\bf x}}_t
=
\frac{\sum_{s=1}^t \eta_s {\bf x}_s}{\sum_{s=1}^t \eta_s}.
\]
Then $f({\hat{\bf x}}_t) \rightarrow \inf f$.
\end{enumerate}
\end{theorem}
\proof
Both parts of the theorem rely on the following observation.
For any ${\bf z}\in{\R^n}$, and for all $t$,
\begin{eqnarray*}
\|{\bf x}_{t+1} - {\bf z}\|^2
&=&
\norm{({\bf x}_{t} - {\bf z}) - \eta_t {\bf u}_t}^2
\\
&=&
\|{\bf x}_{t} - {\bf z}\|^2
- 2 \eta_t {\bf u}_t\cdot ({\bf x}_t - {\bf z}) + \eta_t^2 \norm{{\bf u}_t}^2
\\
&\leq&
\|{\bf x}_{t} - {\bf z}\|^2
- 2\eta_t \Parens{f({\bf x}_t) - f({\bf z})}
+ \eta_t^2 \norm{{\bf u}_t}^2,
\end{eqnarray*}
with the last line following from
${\bf u}_t\in\partial f({\bf x}_t)$.
Applying this inequality repeatedly and rearranging then yields
\begin{eqnarray}
2\sum_{s=1}^{t} \eta_{s} ( f({\bf x}_{s}) - f({\bf z}) )
&\leq&
\norm{{\bf x}_{t+1} - {\bf z}}^2
+
2\sum_{s=1}^{t} \eta_{s} ( f({\bf x}_{s}) - f({\bf z}) )
\nonumber
\\
&\leq&
\norm{{\bf x}_1 - {\bf z}}^2
+
\sum_{s=1}^{t} \eta_{s}^2 \norm{{\bf u}_{s}}^2.
\label{eq:sgd:magic}
\end{eqnarray}
The proof now considers the two cases separately.
\underline{Part~(\ref{thm:grad-desc-converges:a})}:
Suppose, by way of contradiction, that
$f({\bf x}_t)\not\rightarrow \inf f$.
\eqref{eqn:thm:grad-desc-converges:1} implies
that the sequence of values $f({\bf x}_t)$ is nonincreasing, which means
that they must have a limit, which is also equal to their infimum.
Let $\gamma=\lim f({\bf x}_t)=\inf f({\bf x}_t)$.
By our assumption, $\gamma>\inf f$, which also implies $\gamma\in{\mathbb{R}}$.
Thus, there exists a point ${\bf z}\in{\R^n}$ with
$\gamma > f({\bf z}) \geq \inf f$
(and with $f({\bf z})\in{\mathbb{R}}$ since $f$ is proper).
Thus, for all $t$,
\begin{eqnarray*}
2\paren{\sum_{s=1}^t \eta_{s}} (\gamma - f({\bf z}))
&\leq&
2 \sum_{s=1}^t \eta_{s} (f({\bf x}_{s+1}) - f({\bf z}))
\\
&\leq&
2 \sum_{s=1}^t \eta_{s} (f({\bf x}_{s}) - f({\bf z}))
-
\sum_{s=1}^t \eta_{s}^2 \norm{{\bf u}_{s}}^2
\\
&\leq&
\norm{{\bf x}_1 - {\bf z}}^2.
\end{eqnarray*}
The first inequality is because $\gamma=\inf f({\bf x}_t)$.
The second and third inequalities are by
\eqref{eqn:thm:grad-desc-converges:1}
and
\eqref{eq:sgd:magic}, respectively.
The left-hand side of this inequality is converging to $+\infty$
as $t\rightarrow+\infty$, since $\gamma>f({\bf z})$
and since $\sum_{t=1}^\infty \eta_t = +\infty$.
But this is a contradiction since
the right-hand side is constant and finite.
\underline{Part~(\ref{thm:grad-desc-converges:b})}:
Similar to the last argument,
suppose again, by way of contradiction, that
$f({\hat{\bf x}}_t)\not\rightarrow \inf f$.
Then for some $\gamma>\inf f$, $\gamma\in{\mathbb{R}}$,
and some infinite set of indices $S\subseteq\{1,2,\ldots\}$,
we must have
$f({\hat{\bf x}}_t)\geq \gamma$ for all $t\in S$.
Further, there exists ${\bf z}\in{\R^n}$ such that
$\gamma>f({\bf z})\geq \inf f$,
implying $f({\bf z})\in{\mathbb{R}}$
(since $f$ is proper).
Note that, because $f$ is convex,
\begin{equation} \label{eqn:thm:grad-desc-converges:3}
f({\hat{\bf x}}_t)
\leq
\frac{\sum_{s=1}^t \eta_{s} f({\bf x}_{s})}{\sum_{s=1}^t \eta_{s}}.
\end{equation}
Thus, for all $t\in S$,
\begin{eqnarray*}
2\paren{\sum_{s=1}^t \eta_{s}} (\gamma - f({\bf z}))
&\leq&
2\paren{\sum_{s=1}^t \eta_{s}} (f({\hat{\bf x}}_t) - f({\bf z}))
\\
&\leq&
2 \sum_{s=1}^t \eta_{s} (f({\bf x}_{s}) - f({\bf z}))
\\
&\leq&
\norm{{\bf x}_1 - {\bf z}}^2
+
\sum_{s=1}^{\infty} \eta_{s}^2 \norm{{\bf u}_{s}}^2.
\end{eqnarray*}
The first inequality is because $t\in S$.
The second and third inequalities are from
\eqref{eqn:thm:grad-desc-converges:3}
and
\eqref{eq:sgd:magic}, respectively.
As before,
our assumptions imply that the left-hand side can be made
arbitrarily large, since $S$ is infinite.
But this is a contradiction since the right-hand side is
finite and constant.
\qed
Theorem~\ref{thm:grad-desc-converges}(\ref{thm:grad-desc-converges:a})
proves convergence assuming
a lower bound on how much $f({\bf x}_t)$ decreases on each iteration
as a function of its gradient, an approach
that
will henceforth be our main focus.
As an example of when this is possible,
suppose $f$ is \emph{smooth}, meaning
\begin{equation} \label{eqn:grad-bnd-smooth}
f({\bf x}')
\leq
f({\bf x}) + \nabla f({\bf x})\cdot({\bf x}'-{\bf x})
+ \frac{\beta}{2} \norm{{\bf x}'-{\bf x}}^2
\end{equation}
for all ${\bf x},\,{\bf x}'\in{\R^n}$, for some constant $\beta>0$
(and assuming $f$ is differentiable).
Then if ${\bf x}_{t+1}$ is computed as in \eqref{eqn:grad-desc-defn} with
$\eta_t=1/\beta$, then this smoothness condition implies
\[
f({\bf x}_{t+1})
\leq
f({\bf x}_t)
- \frac{1}{2\beta} \norm{\nabla f({\bf x}_t)}^2.
\]
Thus, this update is guaranteed to decrease
the function values of the iterates (from $f({\bf x}_t)$ to $f({\bf x}_{t+1})$)
by at least a constant times $\norm{\nabla f({\bf x}_t)}^2$.
Once established, such a guarantee of progress can sometimes
be sufficient to
ensure $f({\bf x}_t)\rightarrow \inf f$, as was just seen in
Theorem~\ref{thm:grad-desc-converges}(\ref{thm:grad-desc-converges:a}).
Intuitively, if $\nabla f({\bf x}_t)$ is getting close to ${\bf 0}$, then we
should be approaching $f$'s minimum; on the other hand, as long as
$\norm{\nabla f({\bf x}_t)}$ remains large, we are assured of
significant progress (in reducing $f({\bf x}_t)$) on each iteration.
As such, we might expect that a progress guarantee of this kind
should suffice to ensure the convergence of a broad family of methods,
not just (sub)gradient descent.
Nevertheless, as will be seen shortly, although these intuitions seem
superficially reasonable, it is not always the case that such a
guarantee is sufficient to ensure convergence to the
function's minimum.
For the remainder of this subsection, we apply astral methods
to study in substantially greater generality
when such convergence is assured for an arbitrary iterative method,
given such a lower bound on per-iteration progress in terms
of gradients or subgradients.
In particular,
we will see that continuity in astral space is sufficient to
ensure such convergence.
We will also see that such a result is not possible, in general,
when astral continuity is not assumed.
We thus focus on generalizing the approach taken in
part~(\ref{thm:grad-desc-converges:a}) of
Theorem~\ref{thm:grad-desc-converges}
using astral methods, leaving the generalization of the approach
in
part~(\ref{thm:grad-desc-converges:b})
for future work.
Let $\seq{{\bf x}_t}$ be any sequence in ${\R^n}$ (not necessarily computed
using gradient descent), and let
$\seq{{\bf u}_t}$ in ${\R^n}$ be a corresponding sequence of subgradients
so that ${\bf u}_t\in\partial f({\bf x}_t)$ for all $t$.
Generalizing the kind of progress bounds considered above
(such as \eqref{eqn:thm:grad-desc-converges:1}),
we suppose
that
\begin{equation} \label{eqn:aux-fcn-prog-bnd}
f({\bf x}_{t+1})
\leq
f({\bf x}_t) - \alpha_t h({\bf u}_t)
\end{equation}
for some $\alpha_t\geq 0$, and some function
$h:{\R^n}\rightarrow\mathbb{R}_{\geq 0}$.
We assume $h$ satisfies the properties that $h({\bf 0})=0$ and, for all
$\epsilon>0$,
\[
\inf \{ h({\bf u}) : {\bf u}\in{\R^n}, \norm{{\bf u}}\geq\epsilon \} > 0.
\]
We call such a function an \emph{adequate auxiliary function}.
Intuitively, if $h({\bf u})$ is small, these properties force ${\bf u}$ to be
close to ${\bf 0}$.
For example,
\eqref{eqn:grad-bnd-smooth} satisfies
\eqref{eqn:aux-fcn-prog-bnd}
with $\alpha_t=1/(2\beta)$ and $h({\bf u})=\norm{{\bf u}}^2$, which clearly
is an adequate auxiliary function.
In general, if $h$ is continuous, strictly positive except at ${\bf 0}$,
and radially nondecreasing (meaning $h(\lambda{\bf u})\geq h({\bf u})$ for all
$\lambda\geq 1$ and all ${\bf u}\in{\R^n}$), then it must be
an adequate auxiliary function
(by compactness of
$\{ {\bf u}\in{\R^n} : \norm{{\bf u}}=\epsilon \}$
for $\epsilon>0$).
Given a bound as in \eqref{eqn:aux-fcn-prog-bnd},
if ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous
(either everywhere, or just at the limits of all
convergent subsequences of iterates), then
the next theorem shows how we can use our
previous results to prove convergence to $f$'s minimum,
without requiring $f$ to have a finite minimizer, nor the sequence of
iterates $\seq{{\bf x}_t}$ to remain bounded.
Unlike
Theorem~\ref{thm:grad-desc-converges},
this theorem can be applied to any sequence $\seq{{\bf x}_t}$, regardless
of how it is computed or constructed (provided, of course, that it
satisfies the stated conditions).
Furthermore, the theorem relies on
an assumed progress bound of a much weaker and more general form.
\begin{theorem}\label{thm:fact:gd}
Let $f:{\R^n}\rightarrow\overline{\R}$ be convex and proper.
Let $\seq{{\bf x}_t}$ and $\seq{{\bf u}_t}$ be sequences in ${\R^n}$
with each ${\bf u}_t\in\partial f({\bf x}_t)$.
Assume that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at the limit of every convergent
subsequence of $\seq{{\bf x}_t}$.
Also assume $f({\bf x}_1)<+\infty$ and that
\[f({\bf x}_{t+1}) \leq f({\bf x}_t) - \alpha_t h({\bf u}_t)\]
for each $t$, where $\alpha_t\geq 0$ and
$\sum_{t=1}^\infty \alpha_t = +\infty$,
and where
$h:{\R^n}\rightarrow\mathbb{R}_{\geq 0}$ is
an adequate auxiliary function.
Then $f({\bf x}_t)\rightarrow \inf f$.
\end{theorem}
\proof
Suppose first that $\liminf h({\bf u}_t) = 0$.
Then in this case,
there exists a subsequence $\seq{{\bf x}_{s(t)}}$, with indices
$s(1)<s(2)<\cdots$, such that $h({\bf u}_{s(t)})\rightarrow 0$.
We claim further that ${\bf u}_{s(t)}\rightarrow{\bf 0}$.
Suppose otherwise.
Then for some $\epsilon>0$, $\norm{{\bf u}_{s(t)}}\geq\epsilon$ for
infinitely many values of $t$.
Let
$ \delta = \inf \{ h({\bf u}) : {\bf u}\in{\R^n}, \norm{{\bf u}}\geq\epsilon \} $.
Then $\delta>0$, since $h$ is an adequate auxiliary function.
Thus, $h({\bf u}_{s(t)})\geq\delta>0$ for infinitely many values of $t$.
But this contradicts that $h({\bf u}_{s(t)})\rightarrow 0$.
Since ${\bf u}_{s(t)}\rightarrow{\bf 0}$,
we can apply Theorem~\ref{thm:subdiff-min-cont} to the
extracted subsequence, noting that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous at the limit
of all of its convergent subsequences
(by assumption), and that $\sup f({\bf x}_t)<+\infty$ since
$f({\bf x}_1)<+\infty$ and since the function
values $f({\bf x}_t)$ are nonincreasing
(from \eqref{eqn:aux-fcn-prog-bnd},
since $\alpha_t\geq 0$ and $h\geq 0$);
together, these imply that the limit of every convergent subsequence
is in $\intdom{{}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}}$ by
Corollary~\ref{cor:cont-gen-char}.
Theorem~\ref{thm:subdiff-min-cont} thus
yields $f({\bf x}_{s(t)})\rightarrow \inf f$.
Furthermore, this shows the entire sequence, which is nonincreasing,
converges to $f$'s minimum as well,
so that $f({\bf x}_t)\rightarrow\inf f$.
In the alternative case, suppose
$\liminf h({\bf u}_t) > 0$.
Then there exists a positive integer $t_0$ and $\epsilon>0$ such that
$h({\bf u}_t)\geq\epsilon$ for all $t\geq t_0$.
Summing \eqref{eqn:aux-fcn-prog-bnd} yields, for $t > t_0$,
\[
f({\bf x}_t)
\leq
f({\bf x}_1)
-
\sum_{s=1}^{t-1} \alpha_{s} h({\bf u}_{s})
\leq
f({\bf x}_1)
-
\epsilon \sum_{s=t_0}^{t-1} \alpha_{s}.
\]
As $t\rightarrow+\infty$, the sum on the right converges to $+\infty$
(by assumption, even disregarding finitely many terms), implying
$f({\bf x}_t)\rightarrow -\infty$.
Thus, $\inf f=-\infty$ and
$f({\bf x}_t)\rightarrow\inf f$ in this case as well.
\qed
If we drop the assumption regarding ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$'s continuity, then
the convergence proved in Theorem~\ref{thm:fact:gd} can no longer be
assured, in general, even given a progress bound like the one in
\eqref{eqn:aux-fcn-prog-bnd}
with the $\alpha_t$'s all equal to a positive constant,
and even when $h({\bf u})=\norm{{\bf u}}^2$, the
most standard case.
For example, consider again the function $f$
given in \eqref{eqn:curve-discont-finiteev-eg}.
Let ${\bf x}_t=\trans{[t+1,\, t(t+1)]}$.
Then $f({\bf x}_t)=(t+1)/t$
and
$\nabla f({\bf x}_t)=\trans{[2/t,\, -1/t^2]}$.
It can be checked that, for $t\geq 1$,
\[
f({\bf x}_{t+1}) - f({\bf x}_t)
=
-\frac{1}{t(t+1)}
\leq
-\frac{1}{10}
\cdot
\frac{4t^2+1}{t^4}
=
-\frac{1}{10}
\norm{\nabla f({\bf x}_t)}^2.
\]
In other words,
\eqref{eqn:aux-fcn-prog-bnd} is satisfied with $\alpha_t=1/10$, for
all $t$, and $h({\bf u})=\norm{{\bf u}}^2$ (and with ${\bf u}_t=\nabla f({\bf x}_t)$).
Thus, all of the conditions of Theorem~\ref{thm:fact:gd} are
satisfied, except that ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is not continuous everywhere, nor at
the limit of this particular sequence
(namely, $\limray{{\bf e}_2}\mathpalette\v@rplusl\relax}\newcommand{\plusd}{\mathpalette\v@rplusd\relax}\newcommand{\plusu}{\mathpalette\v@rplusu\relax\limray{{\bf e}_1}$).
And indeed, the theorem's conclusion is false in this case since
$f({\bf x}_t)\rightarrow 1 > 0 = \inf f$.
This shows that Theorem~\ref{thm:fact:gd} is false
if continuity is not assumed.
Nevertheless, this does not rule out the possibility that particular
algorithms might be effective at minimizing a convex function, even
without a continuity assumption;
indeed, this was shown to be true of subgradient descent in
Theorem~\ref{thm:grad-desc-converges}(\ref{thm:grad-desc-converges:a}).
Before providing concrete consequences of Theorem~\ref{thm:fact:gd},
we pause to discuss related literature.
An intermediate between the guarantees of
Theorems~\ref{thm:grad-desc-converges} and~\ref{thm:fact:gd}
is the classical \emph{Zoutendijk condition} \citep[Eq. (3.14)]{nocedal_wright}:
for a broad family of descent methods whose steps merely have a positive inner product with the
negative gradient direction, applied to functions which are smooth as in
\Cref{eqn:grad-bnd-smooth},
the sequence of {gradients} must converge to zero, that is,
$\nabla f({\bf x}_t) \to 0$.
However, in the example that was just discussed,
the function $f$ constructed in \eqref{eqn:curve-discont-finiteev-eg}
has a sequence $({\bf x}_t)$ satisfying these progress conditions,
but also $\nabla f({\bf x}_t) \to 0$ and $f({\bf x}_t) \to 1 > 0 = \inf f$;
thus,
the classical Zoutendijk analysis is insufficient to recover
the conclusion of \Cref{thm:fact:gd}.
More recent work has established conditions under which gradient descent converges to minimizers
under smoothness even without convexity
\citep{jason_minimizers}; however, this work assumes the iterates
${\bf x}_t$ are bounded and that a finite minimizer exists in ${\R^n}$,
unlike \Cref{thm:fact:gd} which makes no such assumptions.
A natural follow-up to \Cref{thm:fact:gd} is whether it is possible to establish
convergence of the entire sequence of iterates $({\bf x}_t)$ in
$\extspac{n}$,
not just subsequences.
This question will not be settled here and is moreover left open by the existing literature,
even in well-studied special cases.
In particular, the promising literature on \emph{implicit regularization}
of standard descent methods can show that if the function being minimized satisfies certain structural conditions,
then {coordinate-descent} iterates (defined below) lie in a certain cone finer than
the recession cone
\citep{boosting_margin,zhang_yu_boosting,mjt_margins},
and furthermore that gradient-descent iterates
have a convergent dominant direction \citep{nati_logistic,riskparam}.
However, these guarantees and their associated proofs alone are insufficient
to establish convergence over $\extspac{n}$.
We return now to consequences of Theorem~\ref{thm:fact:gd},
which we show can be
applied to prove the effectiveness of a number of standard algorithms,
even when no finite minimizer exists.
\emph{Coordinate descent} is one such method in which, at each
iteration, just one of the coordinates of ${\bf x}_t$ is chosen and
updated, not modifying any of the others; thus,
${\bf x}_{t+1}={\bf x}_t + \eta_t {\bf e}_{i_t}$ for some basis vector ${\bf e}_{i_t}$ and
some $\eta_t\in{\mathbb{R}}$.
In a \emph{gradient-based} version of coordinate descent, $i_t$ is
chosen to be the largest coordinate of the gradient $\nabla f({\bf x}_t)$.
In a \emph{fully greedy} version, both $i_t$ and $\eta_t$ are chosen to
effect the maximum possible decrease in the function value, that is,
to minimize $f({\bf x}_t+\eta{\bf e}_i)$ among all choices of
$i\in\{1,\ldots,n\}$ and $\eta\in{\mathbb{R}}$.
Many other variations are possible.
\emph{Steepest descent}
\citep[Section~9.4]{boyd_vandenberghe}
is a technique generalizing both gradient
descent and one form of coordinate descent.
In this method, for some $p\geq 1$, on each iteration $t$,
${\bf v}_t\in{\R^n}$ is chosen to
maximize $\nabla f({\bf x}_t)\cdot{\bf v}_t$ subject to $\normp{{\bf v}_t}\leq 1$,
and then subtracted from ${\bf x}_t$ after scaling by some step size
$\eta_t>0$;
thus,
${\bf x}_{t+1} = {\bf x}_t - \eta_t {\bf v}_t$.
(Here, $\normp{{\bf x}}$ denotes the $\ell_p$-norm of ${\bf x}\in{\R^n}$.)
When $p=2$, this is gradient descent;
when $p=1$, it is (gradient-based) coordinate descent.
If $f$ is smooth so that \eqref{eqn:grad-bnd-smooth} is satisfied,
then for all $p\geq 1$,
steepest descent satisfies \eqref{eqn:aux-fcn-prog-bnd} and thus
Theorem~\ref{thm:fact:gd} can be applied to prove its convergence
(given the other conditions of that theorem).
To see this, recall that there exists a constant $C_p>0$ for which
$\norms{{\bf x}}\leq C_p \normp{{\bf x}}$ for all ${\bf x}\in{\R^n}$.
Let $\beta'=C_p^2 \beta$ and let $q\geq 1$ be such that
$1/p + 1/q = 1$ (allowing either $p$ or $q$ to be $+\infty$).
Then we have:
\begin{eqnarray*}
f({\bf x}_{t+1})
&\leq&
f({\bf x}_t)
+ \nabla f({\bf x}_t)\cdot\Parens{{\bf x}_{t+1} - {\bf x}_t}
+ \frac{\beta}{2} \norms{{\bf x}_{t+1} - {\bf x}_{t}}^2
\\
&\leq&
f({\bf x}_t)
+ \nabla f({\bf x}_t)\cdot\Parens{{\bf x}_{t+1} - {\bf x}_t}
+ \frac{\beta'}{2} \normp{{\bf x}_{t+1} - {\bf x}_{t}}^2
\\
&=&
f({\bf x}_t)
- \eta_t \nabla f({\bf x}_t)\cdot{\bf v}_t
+ \frac{\beta'\eta_t^2}{2}
\\
&=&
f({\bf x}_t)
- \eta_t \normq{\nabla f({\bf x}_t)}
+ \frac{\beta'\eta_t^2}{2}
\\
&=&
f({\bf x}_t) - \frac{1}{2\beta'} \normq{\nabla f({\bf x}_t)}^2.
\end{eqnarray*}
The first equality is because ${\bf x}_{t+1}-{\bf x}_t=-\eta_t {\bf v}_t$
and $\normp{{\bf v}_t}=1$.
The second equality is because, for all ${\bf z}\in{\R^n}$,
\[ \max_{{\bf v}:\normp{{\bf v}}\leq 1} {\bf z}\cdot{\bf v} = \normq{{\bf z}}. \]
And the last equality holds if we set
$\eta_t = \normq{\nabla f({\bf x}_t)} / \beta'$.
Thus, \eqref{eqn:aux-fcn-prog-bnd} is satisfied with
$\alpha_t = 1/(2\beta')$
and
$h({\bf u}) = \normq{{\bf u}}^2$, which is an adequate auxiliary function.
In particular,
for smooth convex functions $f$ as above, gradient-based coordinate
descent effectively minimizes $f$, for appropriate step sizes.
This also shows that the same holds for fully-greedy coordinate
descent since this version makes at least as much progress
on each iteration at descreasing the function values as the
gradient-based version.
As a consequence, any of the algorithms just discussed can be applied
to a range of commonly-encountered convex optimization problems which,
in general, might have no finite minimizers.
For instance, logistic regression, as seen in
\eqref{eqn:logistic-reg-obj},
is based on minimization of a function of the form
\[
f({\bf x}) = \sum_{i=1}^m \ln(1+\exp({\bf x}\cdot{\bf u}_i))
\]
for all ${\bf x}\in{\R^n}$, and
for some vectors ${\bf u}_1,\ldots,{\bf u}_m\in{\R^n}$.
This function is convex and proper, but its minimizers might well be at infinity.
Furthermore, its extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere
(by Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:a})).
It can be shown that this function is smooth, for instance,
by explicitly computing the Hessian $\nabla^2 f({\bf x})$ for any
${\bf x}\in{\R^n}$, each of whose components can be bounded by some constant
that depends only on
${\bf u}_1,\ldots,{\bf u}_m$, which in turn implies
\eqref{eqn:grad-bnd-smooth}, for an appropriate $\beta>0$,
using Taylor's theorem.
Thus, any of the methods discussed above can be applied to minimize
this function.
As another example, for ${\bf x}\in{\R^n}$, let
\[
f({\bf x}) = {\sum_{i=1}^m \exp({\bf x}\cdot{\bf u}_i) }
\]
for some vectors ${\bf u}_1,\ldots,{\bf u}_m\in{\R^n}$, and let
$g({\bf x})=\ln f({\bf x})$.
Both functions $f$ and $g$ are convex and proper.
The extension ${}\mkern4.5mu\overline{\mkern-4.5mu f\mkern1.5mu}\mkern-1.5mu}\newcommand{\fpext}{\overline{f'}$ is continuous everywhere
(by Proposition~\ref{pr:hard-core:1}(\ref{pr:hard-core:1:a})),
which also implies that $\overline{g}$ is continuous everywhere.
The function $g$ can be shown to be smooth, similar to the sketch
above.
Since $\log$ is strictly increasing, minimizing $g$ is equivalent
to minimizing $f$.
Thus, either $f$ or $g$ can be minimized by applying any of the
methods above to $g$ (even if $\inf g = -\infty$).
Moreover, whether applied to $f$ or $g$,
fully-greedy coordinate descent is identical in its
updates (as are the other methods,
for appropriate choices of step size).
In particular, the AdaBoost
algorithm~\citep{schapire_freund_book_final}
can be
viewed as minimizing an \emph{exponential loss} function of exactly
the same form as $f$ using fully-greedy coordinate descent;
therefore, the arguments above prove that AdaBoost effectively
minimizes exponential loss
(as had previously been proved by \citet{collins_schapire_singer_adaboost_bregman} using more specialized techniques based on Bregman distances).
As a last example, let $U$ be a finite subset of ${\R^n}$.
For any ${\bf x}\in{\R^n}$, we can define a natural probability distribution
$q_{\xx}$ over $U$ of the form
\[
q_{\xx}({\bf u}) = \frac{\exp({\bf x}\cdot{\bf u})}
{\sum_{{\bf w}\in U} \exp({\bf x}\cdot{\bf w})}.
\]
Given examples ${\bf u}_1,\ldots,{\bf u}_m\in U$, we can then attempt to
estimate the distribution generating these examples by minimizing the
\emph{negative log likelihood}
\[
f({\bf x}) = -\frac{1}{m} \sum_{i=1}^m \ln q_{\xx}({\bf u}_i).
\]
Similar to the preceding examples,
this function is convex, proper, smooth and can be shown to have an extension
that is continuous everywhere.
Therefore, once again, any of the methods above can be applied to
minimize it, even if none of its minimizers are finite.
\bibliographystyle{plainnat}
\clearpage
\phantomsection
\addcontentsline{toc}{section}{\refname}
|
\section{\textbf{Introduction}}
The dark energy(DE) which is responsible for accelerated expansion \cite{Riess98,Perl99} of the Universe and whose origin is still under investigation, fills about $70\%$ energy content of the cosmos \cite{Peebles87, Ratra87, Wetterich87}. The cosmological constant $\Lambda$ is the most basic approach to the DE problem \cite{Padmanabhan02}. Recent observations motivate an interaction between dark matter (DM) and DE sectors \cite{Amendola99,Zimdahl01, Chimento03, delCampo06, Wang16, Pavon05, Sheykhi11}. Such an interaction also shows that their development is intertwined. As a consequence, the overall energy momentum conservation rule is decomposed as
\begin{equation}
\dot{\rho_m}+3H\rho_m=Q, \label{3.1}
\end{equation}
\begin{equation}
\dot{\rho_d}+3H(1+w_d)\rho_d=-Q, \label{3.2}
\end{equation}
where $\rho_m,\; \rho_d$ and $w_d \left(\equiv \dfrac{P_d}{\rho_d}\right)$ are DM energy density, DE density and equation of state (EoS) parameter respectively with DE pressure $P_d$. $Q>0$ indicates transfer of energy from DM to DE whereas $Q<0$ corresponds for DE to DM transfer.\cite{Wang16} provides detailed reference about interacting DE models. The observational data support $\Lambda-$CDM model \cite{Telkamp18}, but fail to answer cosmic coincidence and fine tuning problems.
The Agegraphic Dark Energy (ADE) model is a different way of answering the problems above. The ADE model is based on quantum mechanics' uncertainty principle \cite{Cai07}. The space time $t$ uncertainty is given by $\delta t=\beta t_p^{\frac{2}{3}}t^{\frac{1}{3}}$ for Minkowskian space-time with reduced Plank time $t_p$ and dimensionless constant $\beta$ of order unity.
Altering the entropy of the system modifies the corresponding ADE model which signifies the importance of the entropy relation \cite{Cai07, Wei08, Wei09}. Various ADE models are studied in \cite{Wei08a,Sheykhi09, Zadeh20, Sharma20}.
The basic criteria to construct holographic dark energy (HDE) model connects entropy of the system with it's own radius. The Bekenstein-Hawking entropy, which is derived from a black hole and a cosmic application of the Boltzmann-Gibbs entropy, is the most popular. Kaniadakis postulated the Kaniadakis entropy \cite{Kaniadakis02} as a one-parameter generalisation of the Boltzmann-Gibbs entropy, that has been studied in the literature \cite{Sharma21}. The basic idea behind HDE model is the relation $\rho_dL^4\leq S$, where $L$ represents the longest distance and $S$, the entropy relation used in a black hole of radius $L$ \cite{Li04}. For standard Bekenstein-Hawking entropy, the entropy-area relationship is given by $S_{BH}\propto\dfrac{A}{4G}$ with Newton's constant $G$. The Kaniadakis entropy with dimensionless parameter $K$ is defined as $S_K=\dfrac{1}{K}sinh(KS_BH)$ with $\lim_{K\rightarrow0}S_K=S_{BH}$. Which can be rewritten as $ S_K=S_{BH}+\dfrac{K^2}{6}S_{BH}^3+\mathrm{O(K^4)}$.
And hence the inequality $\rho_dL^4\leq S$ leads to \cite{Drepanou21, Pandey21}
\begin{equation}
\rho_d=\dfrac{3c^2M_p^2}{L^2}+K^2M_p^6L^2. \label{3.5}
\end{equation}
In this manuscript we will use (\ref{3.5}) to propose Kaniadakis agegraphic dark energy (KADE) model by employing universe age as IR cutoff and investigate the consequences on the Universe's development.
The manuscript is organized as follows: Next section is dedicated to the formulation of proposed KADE model for which the universe age is used as the IR cutoff to analyze the universe evolutionary behavior. The investigation is carried out by considering non interaction as well as interaction among the DE and DM sectors of the Universe. We will study the non-interacting KADE model in section 3. One full section is dedicated to study the interaction among two sectors. The final section will summarize the results, conclusion and future possibilities of the model.
\section{\textbf{Formulation of KADE Model}}
Age of the Universe is defined as
\begin{equation}
T=\int_0^{a(t)} \mathrm{d}t=\int_0^{a(t)}\dfrac{\mathrm{d}a(t)}{Ha(t)}, \label{3.6}
\end{equation}
with scale factor $a(t)$ and the Hubble parameter $H$, related by $H=\frac{\dot{a}(t)}{a(t)}$. By considering (\ref{3.6}) as IR cutoff, the energy density of KADE defined by (\ref{3.5}) can be written as
\begin{equation}
\rho_d=\dfrac{3c^2M_p^2}{T^2}+K^2M_p^6T^2. \label{3.7}
\end{equation}
The standard ADE model \cite{Cai07} can be recovered for $K=0$. The metric for flat FRW universe is considered as $\mathrm{d}s^2 =-\mathrm{d}t^2 +\delta_{ij} \mathrm{d}x^i \mathrm{d}x^j a^2(t)$. Assuming the FRW universe to be filled by a pressureless fluid with density $\rho_m$ and KADE density $\rho_d$, the first Friedmann equation is expressed as
\begin{equation}
H^2=\dfrac{1}{3M_p^2}\left(\rho_m+\rho_d\right). \label{3.8}
\end{equation}
Equation (\ref{3.8}) can be rewritten as
\begin{equation}
\Omega_d+\Omega_m=1, \label{3.9}
\end{equation}
where $\Omega_m=\dfrac{\rho_m}{3M_p^2H^2},\quad \Omega_d=\dfrac{\rho_d}{3M_p^2H^2}$.\\
Because we want to demonstrate the model's potential, we only chose specific system parameters values, such as $K, c$ and $b^2$, that produce diverse behaviors. We choose $\Omega_{d_0}= 0.70$ and $H(a = 1) = 67.9$ as current values of these parameters.
\section{Non-interacting KADE model}
In this part, we will develop the formulation of KADE model by considering the independence of DE and DM sectors, i.e. $Q=0$. The expression for EoS parameter is obtained by the time derivative of equation (\ref{3.7}) and using (\ref{3.2}), i.e.
\begin{equation}
w_d=-1-\dfrac{2}{3HT}\left(\dfrac{K^2M_p^4T^4-3c^2}{K^2M_p^4T^4+3c^2}\right), \label{3.10}
\end{equation}
where
\begin{equation}
T=\sqrt{\dfrac{3H^2 \Omega_d-\sqrt{9H^4\Omega_d^2-12c^2K^2M_p^4}}{2K^2M_p^4}}. \label{3.11}
\end{equation}
Differentiating (\ref{3.8}) w.r.t. time, using (\ref{3.1}) and (\ref{3.2}) we get
\begin{equation}
\dfrac{\dot{H}}{H^2}=-\dfrac{3}{2}\left(1-\Omega_d\right)+\dfrac{\Omega_d}{HT}\left(\dfrac{K^2M_p^4T^4-3c^2}{K^2M_p^4T^4+3c^2}\right). \label{3.12}
\end{equation}
From (\ref{3.12}), the expression for deceleration parameter $q$ is given by
\begin{equation}
q=-1-\dfrac{\dot{H}}{H^2}=\dfrac{1}{2}(1-3\Omega_d)-\dfrac{\Omega_d}{HT}\left(\dfrac{K^2M_p^4T^4-3c^2}{K^2M_p^4T^4+3c^2}\right). \label{3.13}
\end{equation}
The cosmic time derivative of the KADE density parameter is given by
\begin{equation}
\dot{\Omega_d}=2H\left(1+q \right))\Omega_d+\dfrac{2\Omega_d}{T}\left(\dfrac{K^2M_p^4T^4-3c^2}{K^2M_p^4T^4+3c^2}\right). \label{3.14}
\end{equation}
In order to analyze the stability of the non-interacting KADE model, the squared sound speed $v_s^2$ expressed by
\begin{equation}v_s^2=\dfrac{\mathrm{d}P_d}{\mathrm{d}\rho_d}=\dfrac{\dot{P_d}}{\dot{\rho_d}}=\dfrac{\rho_d}{\dot{\rho_d}}\dot{\omega_d}+\omega_d, \label{3.15}
\end{equation}
is studied.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_wd_k}
\caption{\begin{small}
Behavior of $w_d$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.1}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_wd_c}
\caption{\begin{small}
Behavior of $w_d$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.2}
\end{figure}
Figures \ref{P3.1} and \ref{P3.2} depict the EoS parameter behavior for redshift $z$. In figure \ref{P3.1}, $c=3$ is fixed and $K$ made to vary from 200 to 600. Whereas, figure \ref{P3.2} is plotted for fixed $K=200$ and varying $c$ from 2.4 to 3.2. In both cases, the model reflects a pure quintessence behavior. All the above plots meet in $w_d=-1$, the $\Lambda-$CDM model in far future.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_q_k}
\caption{\begin{small}
Behavior of $q$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.3}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_q_c}
\caption{\begin{small}
Behavior of $q$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.4}
\end{figure}
The deceleration parameter behavior in figures \ref{P3.3} and \ref{P3.4} clearly says the universe to transit from decelerated to accelerated phase at $z\approx0.6$. The current value of deceleration parameter for considered values of $K$ and $c$ is found to be in the vicinity of $-0.4$.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_Omd_k}
\caption{\begin{small}
Behavior of $\Omega_d$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.5}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_Omd_c}
\caption{\begin{small}
Behavior of $\Omega_d$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.6}
\end{figure}
The behavior of KADE density parameter $\Omega_d$ is shown in figures \ref{P3.5} and \ref{P3.6}. For figure \ref{P3.5}, $K$ is allowed to vary from 200 to 600 by fixing $c(=3)$, whereas \ref{P3.6} is plotted based on varying $c$ from 2.4 to 3.2 and fixed $K(=200)$. Both the figures clearly depict the evolutionary behavior of the universe by telling matter dominated in the past to fully KADE dominated in future.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_vs2_k}
\caption{\begin{small}
Behavior of $v_s^2$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.7}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{01_KADE_NI_z_vs2_c}
\caption{\begin{small}
Behavior of $v_s^2$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.8}
\end{figure}
The squared sound speed plot in figures \ref{P3.7} and \ref{P3.8} shows the model to be stable for future cosmic behaviors.
\section{Interacting KADE model}
In this section we have considered the interaction between the KADE and the DM sectors coupled by $Q$, defined as
\begin{equation}
Q=3b^2H(\rho_m+\rho_d) \label{3.15a}
\end{equation}
\cite{Pavon05, Sadeghi13, Honarvaryan15, Zadeh18, Sharma:2019bgp}. The expressions for EoS parameter, deceleration parameter and the KADE density parameter are obtained under such interaction. The stability of the method is analyzed using squared speed of sound. Using time derivative of (\ref{3.7}), making use of (\ref{3.15a}), equation (\ref{3.2}) leads to the expression for EoS parameter
\begin{equation}
w_d=-1-\dfrac{b^2}{\Omega_d}-\dfrac{2}{3HT}\left(\dfrac{K^2M_p^4T^4-3c^2}{K^2M_p^4T^4+3c^2}\right) \label{3.16}
\end{equation}
where $T$ is defined by equation (\ref{3.11}).
From (\ref{3.1}) and (\ref{3.2}) with equation (\ref{3.15a}) and the time derivative of (\ref{3.8}) we get the expressions for deceleration parameter
\begin{equation}
q=\dfrac{1}{2}\left(1-3b^2-3\Omega_d\right)-\dfrac{\Omega_d}{HT}\left(\dfrac{K^2M_p^4T^4-3c^2}{K^2M_p^4T^4+3c^2}\right). \label{3.17}
\end{equation}
On simple adjustment in equation (\ref{3.17}) gives KADE density parameter as
\begin{equation}
\dot{\Omega_d}=2H(1+q)\Omega_d+\dfrac{2\Omega_d}{T}\left(\dfrac{K^2M_p^4T^4-3c^2}{K^2M_p^4T^4+3c^2}\right). \label{3.18}
\end{equation}
The figures \ref{P3.9} to \ref{P3.11} show the EoS parameter plot against redshift $z$. Clearly, the model describes the universe to evolve from phantom zone. It enters the quintessence zone by crossing the divide line $(w_d=-1)$ and rests the same till present. In far future, again, it enters the phantom zone.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_wd_k}
\caption{\begin{small}
Behavior of $w_d$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.9}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_wd_c}
\caption{\begin{small}
Behavior of $w_d$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.10}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_wd_b2}
\caption{\begin{small}
Behavior of $w_d$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $b^2$.
\end{small}}
\label{P3.11}
\end{figure}
Figures \ref{P3.12} and \ref{P3.13} are plotted for fixed $b^2$. We observe the universe to enter the accelerated phase some where in $0.6<z<0.8$ and the current value of deceleration parameter $q(z=0)=q_0$ lies in the vicinity of $-0.4$ $(\geq -0.4)$. But, while varying $b^2$ in figure \ref{P3.14} the $z$ interval become wider. $b^2=0.01$ and 0.03 favors the range $0.6<z<0.8$ but values beyond 0.03 indicate the universe to transit for $z>1$. Such a variation in $b^2$ indicate $q_0$ to be $\leq -0.4$.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_q_k}
\caption{\begin{small}
Behavior of $q$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.12}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_q_c}
\caption{\begin{small}
Behavior of $q$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.13}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_q_b2}
\caption{\begin{small}
Behavior of $q$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $b^2$.
\end{small}}
\label{P3.14}
\end{figure}
Figures \ref{P3.15} and \ref{P3.16} says the universe to be dominated by matter sector in the past i.e. for $z \gtrsim 0.5$ which lies in line with \cite{Frieman08} but increasing values of $b^2$ in figure \ref{P3.17} violate $z \gtrsim 0.5$. All the three figures indicate the universe to be fully dominated by KADE in far future.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_Omd_k}
\caption{\begin{small}
Behavior of $\Omega_d$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.15}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_Omd_c}
\caption{\begin{small}
Behavior of $\Omega_d$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.16}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_Omd_b2}
\caption{\begin{small}
Behavior of $\Omega_d$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $b^2$.
\end{small}}
\label{P3.17}
\end{figure}
The $v_s^2$ behavior in figures \ref{P3.18} and \ref{P3.20}, which is based on variation either in $K$ or $b^2$, reflects the model to be stable in future predictions. While varying $c$ indicates the model to be unstable as evidenced in figure \ref{P3.19}.
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_vs2_k}
\caption{\begin{small}
Behavior of $v_s^2$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $K$.
\end{small}}
\label{P3.18}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_vs2_c}
\caption{\begin{small}
Behavior of $v_s^2$ vs $z$ for $H_0=67.8, \quad M_p^2=1$ and prescribed values of $c$.
\end{small}}
\label{P3.19}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=16cm,height=8cm, angle=0]{02_KADE_I_z_vs2_b2}
\caption{\begin{small}
Behavior of $v_s^2$ vs $z$ for $ H_0=67.8, \quad M_p^2=1$ and prescribed values of $b^2$.
\end{small}}
\label{P3.20}
\end{figure}
\section{Conclusion}
In this study, age of the universe is considered as the IR cut-off and the entropy role is played by Kaniadakis entropy. The evolutionary behavior of the universe is studied by considering the matter and energy sectors to independent as well as interacting. If we consider the two sectors to be fully independent, the KADE model becomes a pure quintessence model as evidenced from figure \ref{P3.1} and \ref{P3.2}. On considering the interaction among the two sectors, figures \ref{P3.9} to \ref{P3.11} shows that the universe evolved from phantom zone and enters to quintessence region by crossing the divide line $(w_d=-1)$ and rests there at present but again enters to phantom zone in far future. Hence the KADE model behaves phantom like in past and far future, whereas quintessence like at present. This reflects the richness of the KADE model. A close look at figures \ref{P3.3}, \ref{P3.4}, \ref{P3.12} and \ref{P3.13} indicates the transition of the universe to accelerated phase happen in $0.6 \leq z \leq 0.8$ irrespective of the interaction between the two sectors and the current $q$ value is found to be greater than $-0.4$. But, while varying $b^2$ in figure \ref{P3.14} the value of $q_0$ lies $\lesssim -0.4$ and the interval of transition also become wider. The universe, ruled by matter sector in the past $(z \gtrsim 0.5)$ and completely dominated by KADE in future, is evidenced from figures \ref{P3.5}, \ref{P3.6} and \ref{P3.15}, \ref{P3.16} and hence the domination by KADE happens in the neighborhood of $z=0.5$. The variation in $b^2$ in figure \ref{P3.17} favors matter domination in the past and KADE domination in the future but creates a distinction for the matter to KADE domination. Hence the condition of $z \approx 0.5$ is no more valid for the domination. The squared speed of sound is plotted in figure 7, 8, 18 and 20 shows the stability of the KADE model in far future for both interacting and non-interacting. Figure 19 shows the instability of the interacting case while varying the value of $c$. Hence we conclude the KADE model will be suitable for the study of the future universe. The parameter estimation with observational data is the future study of this work.
|
\section*{Acknowledgments}
This work is mainly supported by Beijing Academy of Artificial Intelligence (BAAI) as well as supported in part by the Fundamental Research Funds for the Central Universities under Grant N2216013, National Science and Technology Major Project (J2019-IV-0002-0069), the Natural Science Foundation of China under Grant 62006129 and the College Scientific Research Project of Inner Mongolia under Grant NJZY21263.
\section{Methodology}
This section describes our Conditional Autoencoder (ConAE).
\subsection{Preliminary of Dense Retrieval}\label{model:preliminary}
Given a query $q$ and a document collection $D=\{d_1,\dots,d_j,\dots,d_n\}$, dense retrievers~\cite{xiong2020dense,xiong2020approximate,karpukhin2020dense} employ pre-trained language models~\cite{devlin2019bert,liu2019roberta}, to encode $q$ and $d$ as $K$-dimensional embeddings, $h_q$ and $h_d$.
Then we can calculate the retrieval score $f(q,d)$ of $q$ and $d$ with dot product $f(h_q,h_d)=h_q \cdot h_d$ and contrastively train encoders by maximizing the probability $P(d^+|q, \{d^+\} \cup D^-)$ of the relevant document $d^+$~\cite{karpukhin2020dense,xiong2020dense,xiong2020approximate}:
\begin{equation}\small
\label{eq:prob}
P(d^+|q, \{d^+\} \cup D^-) = \frac{e^{f(h_q,h_{d^+})}}{e^{f(h_q,h_{d^+})} + \sum\limits_{d^-\in D^-}{e^{f(h_q,h_{d^-})}}},
\end{equation}
where $d^-$ is the document from the irrelevant document set $D^-$.
\subsection{Dimension Compression with ConAE}\label{model:autoencoder}
In this subsection, we introduce ConAE to compress the high-dimensional embeddings $h_q$ and $h_d$ of both queries and documents to the low-dimensional embeddings $h_q^e$ and $h_d^e$.
\textbf{Encoder.} We first get the initial dense representations for query $q$ and document $d$ through dense retrievers, such as ANCE~\cite{xiong2020approximate}. Then these high-dimensional embeddings can be compressed to low dimensional ones:
\begin{equation}
\small
h_q^e = \text{Linear}_q(h_q); h_d^e = \text{Linear}_d(h_d),
\end{equation}
where $h_q^e$ and $h_d^e$ are $L$-dimensional embeddings of $q$ and $d$, which are respectively encoded by two different linear layers, $\text{Linear}_q$ and $\text{Linear}_d$. The dimension $L$ can be much lower than the dimension of initial representations, such as 256, 128 and 64.
Then we use KL divergence to regulate encoded embeddings to mimic the initial embedding distributions of queries and their top-ranked documents:
\begin{equation}\label{eq:kl}
\small
L_{KL} = \sum\limits_{q}{\sum_{d} P(d|q, D_{\text{top}}) \cdot \log \frac{P(d|q, D_{\text{top}})}{P_e(d|q, D_{\text{top}} )}},
\end{equation}
where $P_e(d|q, D_{\text{top}})$ is calculated with E.q.~\ref{eq:prob}, using the encoded embeddings $h_q^e$ and $h_d^e$.
\textbf{Decoder.} The decoder module maps the encoded embeddings $h_q^e$ and $h_d^e$ into the original embedding space and aligns them with $h_q$ and $h_d$, aiming at optimizing encoder modules to maximally maintain ranking features from $h_q$ and $h_d$.
Firstly, we use one linear layer and project $h_q^e$ and $h_d^e$ to $K$-dimensional embeddings, $\hat{h}_q$ and $\hat{h}_d$:
\begin{equation}\small
\hat{h}_q = \text{Linear}(h_q^e);
\hat{h}_d = \text{Linear}(h_q^e).
\end{equation}
Then we respectively train the decoded embeddings $\hat{h}_q$ and $\hat{h}_d$ with the original frozen document and query embeddings of ANCE to recover ranking features from vanilla ANCE and better align them with $h_q$ and $h_d$ in the original embedding space.
The first loss $L_{q}$ is used to optimize the generated query representation $\hat{h}_{q}$:
\begin{equation}\small
L_{q}= \sum\limits_{q}{1 + \tanh f(\hat{h}_{q}, h_{d^-}) - \tanh f(\hat{h}_{q}, h_{d^+})},
\end{equation}
and we can also optimize the generated document representation $\hat{h}_{d}$ with the second loss function $L_{d}$:
\begin{equation}\small
L_{d}= \sum\limits_{q}{1 + \tanh f(h_{q}, \hat{h}_{d^-}) - \tanh f(h_{q}, \hat{h}_{d^+})},
\end{equation}
\textbf{Training Loss.} Finally, we train our conditional autoencoder model with the following loss $L$:
\begin{equation}
\small
L= L_{KL} + \lambda L_{q} + \lambda L_{d},
\end{equation}
where $\lambda$ is a hyper-parameter to weight the autoencoder losses.
\section{Evaluation Result}
Three experiments are conducted in this section to study the effectiveness of ConAE in reducing embedding dimensions for dense retrieval.
\subsection{Overall Performance}\label{sec:overall}
The performance of different dimension reduction models is shown in Table~\ref{tab:overall}. PCA, CE and ConAE are based on ANCE (teacher), which freeze the teacher model and only optimize the dimension projection layers. ANCE starts from CE and continuously tunes all parameters in the model.
Compared with PCA and CE~\cite{ma2021simple}, ConAE achieves the best performance on almost of datasets, which shows its effectiveness in compressing dense retrieval embeddings. ConAE can achieve comparable performance with ANCE (teacher) using 128-dimensional embeddings to build the document index on MS MARCO, which reduces the retrieval latency (from 17.152 ms to 3.942 ms per query) and saves the index storage (from 26G to 4.3G) significantly. It demonstrates that ConAE is effective to eliminate the redundancy of the embeddings learned by dense retrievers.
Among all baselines, PCA shows significantly worse ranking performance on MS MARCO, indicating that embedding dimensions of dense retrievers are usually nonorthogonal. ConAE achieves more than 11\% improvements than CE and much better performance on TREC-COVID, demonstrating the ranking effectiveness and generalization ability of the compressed embeddings encoded by ConAE. ANCE can further improve the retrieval performance of CE by
adapting the teacher model to the low-dimensional version~\cite{zhou2021meta}.
\subsection{Ablation Study}\label{sec:ablation}
This subsection conducts ablation studies in Table~\ref{tab:ablation} to investigate the roles of modules in ConAE.
The source of the effectiveness of ConAE mainly derives from our KL module, which is used to directly optimize the encoder module of ConAE (E.q.~\ref{eq:kl}). The KL module learns the distributions of low-dimensional embeddings of queries and documents from ANCE (teacher) by mimicking its document ranking probability. When only optimizing the embedding projection layers, KL shows much better retrieval performance than CE (Sec.~\ref{sec:overall}), which thrives from optimizing the embedding encoders with more fine-grained ranking signals provided by ANCE (teacher). Besides, the autoencoder module can further improve the ranking performance with compressed embeddings, verifying our claim that the compressed embeddings should maintain the maximal information and recover the ranking features of high-dimensional embeddings.
\subsection{Embedding Visualization with ConAE}
Finally, we randomly sample one case from MS MARCO and visualize the embedding space of query and documents in Figure~\ref{fig:tsne}.
We first employ t-SNE~\cite{van2008visualizing} to visualize the embedding spaces of ANCE (teacher) and ConAE. Compared with ANCE (teacher), the reduced embeddings usually conduct a more uniform space to better retain ranking features from teacher models. As shown in Figure~\ref{fig:tsne:128}, ConAE-128 conducts a more meaningful visualization: the related query-document pair is closer and the other documents are distributed around the golden document according to their relevance to the query. The visualization of ANCE (teacher) is slightly distorted and different from expectation, which mainly due to its redundancy. The redundant features usually mislead t-SNE to overfit these ranking features, thus reducing the embedding dimension of dense retriever to 128 for eliminating redundant features provides a possible way to visualize the embedding space of dense retrievers using t-SNE. Besides, ConAE-64 shows decreased retrieval performance than ConAE-128 (Sec.~\ref{sec:overall})
which derives from that ConAE-64 loses some ranking features with the limited embedding dimensions.
The other way to visualize the embedding space is using ConAE (w/o Decoder) to project the embedding to a 2-dimensional coordinate, which uses a uniform function to map embeddings and focuses on maintaining primary features to mimic the relevance score distribution of documents. As shown in Figure~\ref{fig:tsne:2}, the distributions of documents are distinguishable and provide an intuitive way to analyze the ranking-oriented document distribution. In addition, the query is usually far away from the documents. The main reasons lie that the relevance scores are calculated by dot product and the embedding norms are also meaningful to distinguish the relevant documents.
\section{Conclusion}
This paper presents ConAE to reduce embedding dimensions of dense retrievers. Our experiments show the effectiveness of ConAE by achieving comparable performance with teacher model, significantly reducing the index storage, accelerating the searching process and visualizing the embedding space more intuitively and effectively.
\section{Introduction}
As the first stage of numerous multi-stage IR and NLP tasks~\cite{nogueira2019document,chen2017reading,thorne2018fact}, dense retrievers~\cite{xiong2020approximate} have shown lots of advances in conducting semantic searching and avoiding the vocabulary mismatch problem~\cite{robertson2009probabilistic}. Dense retrievers usually encode queries and documents as high-dimensional embeddings, which are necessary to guarantee the retrieval effectiveness during training~\cite{ma2021simple,reimers2021curse} but exhaust the memory to store the index and lead to longer retrieval latency~\cite{indyk1998approximate,meiser1993point}.
The research of building efficient dense retrieval systems has been stimulated recently~\cite{effecientqa}. To reduce the dimensions of document embeddings, existing work reserves the principle dimensions or compresses query and document embeddings for more efficient retrievers ~\cite{yang2021designing,ma2021simple}.
There are two challenges in compressing dense retriever embeddings: The compressed embeddings should share a similar distribution with the original embeddings, making the low-dimensional embedding space uniform and the document embeddings distinguishable; All the compressed embeddings should have the ability to maintain the maximal information for matching related queries and documents during retrieval, which helps better align the related query-document pairs.
This paper proposes Conditional Autoencoder (ConAE), which aims to build efficient dense retrieval systems by reducing the embedding dimensions of queries and documents. ConAE first encodes high-dimensional embeddings into a low-dimensional embedding space and then generates embeddings that can be aligned to related queries or documents in the original embedding space. In addition, ConAE designs a conditional loss to regularize the low-dimensional embedding space to keep a more similar embedding distribution with original embeddings. Our experiments show that ConAE is effective to compress the high-dimensional embeddings and avoid redundant ranking features by achieving comparable retrieval performance with vanilla dense retrievers and better visualizing the embedding space with t-SNE.
\section{Experimental Methodology}
This section describes the datasets, evaluation metrics, baselines and implementation details in our experiments.
\textbf{Dataset.} Four datasets are used to evaluate retrieval performance, MS MARCO (Passage Ranking)~\cite{bajaj2016ms}, NQ~\cite{kwiatkowski2019natural}, TREC DL~\cite{craswelloverview} and TREC-COVID~\cite{roberts2020trec}.
In our experiments, we use MS MARCO for training and evaluate on MS MARCO (Dev), TREC DL and TREC-COVID. We randomly sample 50,000 queries from the training set of MS MARCO as the development set.
\textbf{Evaluation Metrics.} NDCG@10 is used as the evaluation metric on three datasets, MS MARCO, TREC DL and TREC-COVID. MS MARCO also uses MRR@10 as the primary evaluation metric~\cite{bajaj2016ms}. For the NQ dataset, the hit accuracy on top20 and top10 is used to evaluate retrieval performance~\cite{karpukhin2020dense}.
\textbf{Baselines.} In our experiments, we compare ConAE with two baselines from the previous work~\cite{ma2021simple}, Principle Component Analysis (PCA) and CE. PCA reduces the embedding dimension by retaining the principle dimensions that can keep most of the variance within the original representation. CE model uses two linear layers $W_q$ and $W_d$ without biases to transform dense representations of queries and documents into lower embeddings~\cite{ma2021simple}. We also start from CE models and continuously train the whole model to implement our ANCE models to conduct embeddings with the same reduced dimensions.
\textbf{Implementation Details.}
The rest describes our implementation details.
All embedding dimension reduction models base on one of the best dense retrievers ANCE~\cite{xiong2020approximate} and build document index with exact match (flat index) and is implemented by FAISS~\cite{johnson2019billion}. For each query, we sample 7 negative documents to contrastively train CE and ANCE, while using 1 negative document to train ConAE. During training models, we set the batch size to 2 and accumulate step to 8 for ANCE, while the batch size and accumulate step are 128 and 1 for other models. All models are implemented with PyTorch and tuned with Adam optimizer. The learning rates of ANCE and other models are set to $2e-6$ and 0.001, respectively. $\lambda$ of ConAE is set to 0.1.
\section{Related Work}
Dense retrievers use BERT-Siamese architecture to encode queries and documents to conduct an embedding space for retrieval~\cite{karpukhin2020dense,xiong2020dense,xiong2020approximate,lewis2020pre,zhan2020learning,li2021more,Yu2021FewShotCD}. To learn an effective embedding space, dense retrievers are forced to maintain high-dimensional embeddings during training.
The most direct way to reduce the dimension of embeddings is that retaining part of dimensions of the high-dimensional embeddings~\cite{yang2021designing,ma2021simple}. Some work uses the first 128 dimensions to encode both questions and documents~\cite{yang2021designing}, while using PCA can retain the primary dimensions to recover most information from the raw embeddings~\cite{ma2021simple}. Previous supervised models~\cite{ma2021simple} use neural networks to compress the high-dimensional embeddings as lower-dimensional ones. They provide a better dimension reduction way than unsupervised models by avoiding missing too much information. To optimize the encoders, some work~\cite{ma2021simple} continuously trains dense retrievers with the contrastive training objection~\cite{karpukhin2020dense,xiong2020approximate}.
\section{Appendices}
\subsection{Data Statistics}
In our experiments, we use four datasets to evaluate retrieval performance, MS MARCO (Passage Ranking)~\cite{bajaj2016ms}, NQ~\cite{kwiatkowski2019natural}, TREC DL~\cite{craswelloverview} and TREC-COVID~\cite{roberts2020trec}. These datasets come from different retrieval scenarios and all data statistics are shown in Table~\ref{tab:dataset}.
\input{table/dataset}
\subsection{Retrieval Performance with HNSW}
\input{table/hnsw}
Besides exact searching, we also show the retrieval results in Table~\ref{tab:hnsw}, which are implemented by an ANN retrieval method, Hierarchical Navigable Small World (HNSW). Using HNSW, the retrieval efficiency can be further improved, especially for high-dimensional embeddings. ConAE keeps its advanced retrieval performance again with less than 1ms retrieval latency.
|
\section{An Application}
|
\section{Introduction}
Observations have revealed the widespread existence of long ($\sim$ kpc) and
coherent gas filament in the Milky Way disk {\citep{Li2013, Goodman2014,
2014A&A...568A..73R, Wang2015,2016ApJS..226....9W,2021ApJ...921L..42V}}, which has been reproduced in simulations {
\citep[e.g.][]{2006MNRAS.371.1663D,2015MNRAS.447.3390D,2016MNRAS.455.3640S,2020MNRAS.492.1594S}}. In recent years, data from the Gaia mission
\citep{2016A&A...595A...1G} have enabled one to map the spatial distributions of
gas in the Solar vicinity with good accuracy. These observations have confirmed
that these long filaments are basic units through which gas in the Milky Way
disk is organized \citep{2019A&A...625A.135L, Chen2019}. Among these objects, a
particularly interesting one is the kpc-scale Radcliffe Wave discovered by
\citet{2020Natur.578..237A}.
Before Gaia, it was widely believed that the ``Gould Belt"
\citep{1874AmJSA...8..325G}, see also \citet[][and references
therein]{2014Ap.....57..583B} is a ring-like arrangement of molecular gas. It
is only until recently people realized that this structure belongs to a 2.7
kpc-sized, spatially coherent filamentary structure \citep{2020Natur.578..237A}.
In addition to this, \citet{2020Natur.578..237A} have identified an undulation
mode in the $z$ direction with a period of around $ 2 \,\rm kpc$, the origin
of such an undulating pattern remains undetermined.
The coherent nature of the object can be further confirmed using radial velocity measurement available for the whole structure \citep{2020Natur.578..237A}.
In this paper, using proper motion data provided by the Gaia mission, we study
the velocity structure of the Young Stellar Objects (YSOs) associated with
clouds in the Radcliffe Wave and report the discovery and subsequent
analyses of a new, { coherent, wave-like} undulation pattern in the velocity
space.
\begin{figure*}
\centering
\includegraphics[width = 0.8 \textwidth]{fig_fl.pdf}
\caption { \label{fig:xy} A top-down view of the the Milky Way disk in the Solar vicinity. The Milky Way center is at ($X,~Y$) = $(0,~0)$\,kpc, The Sun sits at $(-8.34, 0)$\,kpc. The blue circles are the catalogue clouds taken from \citet{2020MNRAS.493..351C}. The "+" signs represent the locations of Class~I/II YSOs selected from \citet{2016MNRAS.458.3479M}, where YSOs associated with the Radcliffe Wave are colored in red and the rest in yellow.}
\end{figure*}
\section{Data}
We focus on the kinematic structure measured
along the vertical $z$ direction. The proper motions of
the YSOs that are associated with the Radcliffe Wave are adopted in the current work. YSOs are stars that were newly
born within clouds and they inherit the motion gas from which they originate. We use their velocities to study the kinematic structure of the gas they are associated with. We note that although the YSOs inhere the velocity of the gas from which they were born. Because of turbulence, velocity of the gas we observe now can deviate from the velocity measured at the time when the YSOs are born. As an estimation, the YSOs we use have an age of $\lesssim 2\rm Myr$ \citep[][and references therein]{2015ApJS..220...11D}.
Assuming that the turbulence in the cloud follows the relation as described by \citet{1981MNRAS.194..809L} where $\sigma_{\rm v}/1 {\rm\,km/ s} \approx (l/1\,\rm pc)^{0.38}$, the crossing time is a function of scale where $t / 1 \,{\rm Myr} = (l / 1 \,{\rm pc})^{0.62}$. Thus, an age difference of 2 Myr translates to velocity dispersion of $1.5\,\rm km\,s^{-1}$, which, to put in simple terms, is the estimated velocity difference between
the individual YSOs and the gas they are associated with.
\subsection{YSO selection}
The YSOs are selected from the work of \citet{2016MNRAS.458.3479M}.
Based on the near-infrared photometry from Two Micron
All-Sky Survey (2MASS; \citealt{Skrutskie2006}) and Wide-field Infrared Survey
Explorer (WISE; \citealt{Wright2010}) and the 353\,GHz R2.01 Planck dust
opacity map (\citealt{Planck2014}), \citet{2016MNRAS.458.3479M} presented an
all-sky probabilistic catalog of YSOs using different machine learning
techniques. We use the sources from the
\citet{2016MNRAS.458.3479M} catalogue that are classified as Class I or Class
II YSOs, which are the youngest stars that still locate near their birthplaces
and have similar motions as their parent gases to study the kinematic structure
of the Radcliffe Wave. Distances and proper motions of these YSOs are obtained by cross-matching with the Gaia Data Release 2 (Gaia DR2; \citealt{Gaia2018}). We further removed sources whose Gaia DR2 parallax
uncertainties are larger than 20\,per\,cent. This yields 15\,499 Class~I/II
YSO candidates.
Fig.~\ref{fig:xy} plots the spatial distribution of the YSOs in the Galactic disk, where locations of molecular clouds from \citet{2020MNRAS.493..351C} are overlaid. The selection method can be found in the supplementary material.
\section{Analysis}
To study the undulation pattern, we adopt the same coordinate system as
used by \citet{2020Natur.578..237A}, where we define a new axis $X'$. The
starting point of this axis anchors at the CMa cloud, which locates at $(X =
-9.3$\,kpc and $Y = -1$\,kpc), and it follows along the filament. Along this axis, in Fig.~\ref{fig:variations}, we plot the spatial variation of $z$ and $v_z$. As a amplitude of the undulation decreases relatively fast \citep{2020Natur.578..237A}, we focus on its first cycle.
\subsection{Spatial variations of $z$ and $v_z$}
From Fig,
\ref{fig:variations}, one can identify the sinusoidal modulation in
the $z$ direction as reported by \citet{2020Natur.578..237A}. In addition to this, a
new yet similar feature can also be seen in the $v_z$ plot where the period of the
undulation appears to be similar to the previous one.
\begin{figure}
\centering
\includegraphics[width = 0.49 \textwidth]{fig_z.pdf}\\
\includegraphics[width = 0.49 \textwidth]{fig_vz.pdf}\\
\caption{Undulation structures of the Radcliffe Wave defined
along the $X'$ axis. The $X'$ starts from $(X=-7.2$\,kpc and $Y=-1$\,kpc) and follows along
the filament. We focus on the unshaded region, where $0 < X' < 2 \,\rm kpc$, which corresponds to the first cycle of the undulation.
{\it Top panels:} the spatial undulation. The uppermost panel
plots the spatial displacement from the Galactic disk mid-plane as a function
of $X'$, where the positions of the YSOs are represented using blue ``+" signs,
from which we constructed a smoothed profile indicated using the red solid
line. The second panel plots the result from our Ensemble Empirical Mode
Decomposition (EEMD; \citealt{journals/aada/WuH09}) decomposition, where
different curves represents different Intrinsic Mode Function (IMF) components. The $n=2$ IMF, indicated by
the green solid line, has the largest amplitude and it corresponds to the
undulation reported in \citep{2020Natur.578..237A}. {\it Bottom panels:}
undulation in the velocity space. Third panel: $v_z$ as a function of $X'$, where the
velocity of the individual YSOs are presented as blue `+" signs, and the red line is the smoothed curve.
Fourth panel: results from the EEMD decomposition. Curves of different colors represent different IMFs. The $n=2$ IMF, indicated by
the green solid line, has the largest amplitude and we propose that this is the undulation signal in the velocity space.
\label{fig:variations}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width = 0.52 \textwidth]{data_phi.pdf}\\
\caption{\label{fig:phase} Synchronization of the spatial and the velocity undulations. The blue line in the top panel
represents the $n$=2 IMF of the spatial profile (thus the spatial undulation), and the yellow line in the top panel represents the $n$=2 IMF of the velocity ($v_z$) profile (velocity undulation).
{\it Bottom panel}: phase difference between the spatial undulation and that of the velocity undulation. They appear to be synchronized within $0{\rm\,kpc}<X'<2 {\rm\,kpc}$.
}
\end{figure}
We use the program
\texttt{smoothfit}\footnote{\url{https://pypi.org/project/smoothfit/}.} to
produce smoothed curves which describe the change of $z$ and $v_z$ as a
function of the location $X'$. In this approach, the smoothing is achieved through a
constrained optimization approach. This allows us to average out the measurement errors of the individual YSOs and focus on the overall structure.
The smoothed curves are presented in Fig. \ref{fig:variations}.
\subsection{Decomposition of the $z$ and $v_z$ measurements}
We decompose the $z$ and $v_z$ profiles into components of different { spatial}
frequencies using a \texttt{python} version of Ensemble Empirical Mode
Decomposition (EEMD; \citealt{journals/aada/WuH09}). Being an extension to the
Empirical Mode Decomposition (EMD; \citealt{1998RSPSA.454..903H}), EEMD is an empirical
method that is designed to decompose a signal into components of different
frequencies, where each component is called an Intrinsic Mode Function (IMF).
These components (IMFs) form a complete and nearly orthogonal basis for the
original signal. Compared to methods such as the wavelet transform, EEMD have a significant advantage in that its performance is robust for nonlinear and non-stationary data. In this paper, we performed EEMD decompositions to both the $z$ and $v_z$ profiles. The results are presented in Fig. \ref{fig:variations}.
From the EEMD results,
we identify the $n=2$ IMF from the $z$ profile as the mode that corresponds to the undulation seen in \citet{2020Natur.578..237A}, based on the fact that it is the most prominent IMF, as well as the fact that it has a period of around $\approx 1.5 \,\rm kpc$, which is { very similar to} the period reported in \citet{2020Natur.578..237A} who used a damped sinusoidal function to fit the spatial structure. In this paper, we call this feature the ``spatial undulation".
The fact that we are able to recover this mode from our analysis provides a good validation to our approach.
Apart from this, the $n= 0, 1$ IMFs represent some low-spatial-frequency components with lower spatial frequencies, whereas the $n\gtrsim 3$ IMF components represent some higher-order {variations}.
From the decomposition of the $v_z$ profile, we find that the $n=2$ IMF also appears to be the most prominent mode. Similar to the spatial undulation, it also has a period of $\approx 1.5 \,\rm kpc$. In this paper, we called this pattern the ``velocity undulation". From the plots, we measure an amplitude of $75\,\rm pc$ for the spatial undulation and an amplitude of $6\,\rm km\,s^{-1}$ for the velocity undulation. Both the spatial and the velocity undulation appear to be damped, i.e. their amplitudes tend to decrease with increasing $X'$.
The fact that the spatial undulation and velocity undulation has a similar frequency, as well as the fact that both are damped in a similar way strongly imply that they are physically related.
We note that the parameters we derived for the undulation is similar yet not
identical to the values derived by \citet{2020Natur.578..237A}: we
find an amplitude of $H \approx 130\,\rm pc$ and a period of $p \approx 1.5\rm \,kpc$, whereas
\citet{2020Natur.578..237A} found $H \approx 160$ pc and $p \approx 2 \, \rm kpc$. The difference is
likely to be caused by the differences in model and method:
The \citet{2020Natur.578..237A} model assumed that the structure is aligned with the disk
mid-plane, whereas in our model we do not impose such a condition and have found that the wave is tilted, e.g.
the $X'=0$ side of the wave is tilted downwards. {According to our model, a wave crest can be found at $X'=0$, and this is different from \citet{2020Natur.578..237A}. }
The inclusion of the new tiled component
is the main reason why our analyses yield a smaller amplitude and a
smaller period. Nevertheless, we believe that the true period is likely to be
around our value of 1.5\,kpc as this is the common period shared by the spatial
and the velocity undulation. The statistical error and systematics of our analyses
should also be discussed: In the first cycle, our dataset has a YSO density of about
400 per kpc. Therefore, the crest of the wave, which is about 0.3 kpc long, is
sampled by around 120 YSOs, and a typical, 50 pc-sized cloud should be sampled by
$\sim$ 20 YSOs. Assuming a typical distance of $\sim$ 1 kpc, the GAIA distance error is $\sim$ 100 pc per YSO, and we expect a statistical distance error of $\sim$ 22 pc per cloud. Since we are
only interested in studying the pattern in the $z$ direction, this only amounts to $22\,{\rm pc} / 1\,{\rm kpc} = 2\%$ error.
Our sample
YSOs have a median parallax error of 0.17 mas/yr, which corresponds to an
velocity error of 0.8 $\rm km\,s^{-1}$ for individual YSOs at 1 kpc, and a statistical error of
0.2 $\rm km\,s^{-1}$ per cloud. The velocity measurement
also contain other errors: Due to the lack of radial velocities, what we measured is not strictly $v_z$, but
rather the velocity projected onto an equi-distance sphere from the earth.
This velocity error caused by projection can be as large as 20\,per\,cent (Sec.~\ref{sec:vz:comp}).
Other factors such as intrinsic thickness and intrinsic velocity dispersion can also be important. In our formalism, these are contained in the n= 1, 2 IMFs of the EEMD results. This is true both for the decomposition of the spatial undulation signal and the velocity one. From these, we estimate an error of about 6 pc for the spatial undulation signal and an error of around 1 $\rm km s^{-1}$ for the velocity undulation signal, for the first cycle.
\subsection{Determining the phase difference}
We use Hilbert transform \citep[e.g.][]{Hilbert} to study phase difference between the spatial and the velocity undulation. The Hilbert transform is a linear transform typically applied to real-valued signals to derive the real-time evolution of the phase and the amplitude.
The results are present in Fig. \ref{fig:phase}. Within $0\,{\rm kpc}<X'<1.5\,{\rm kpc}$, the spatial and velocity undulation have a phase difference of $\approx 2/3 \pi$. They start to de-synchronize from $X' = 1.5\,{\rm kpc}$. However, this might be due to the fact that the error of proper motion has increased and the fact that the filament is sparely sampled by YSOs for $X' > 1.5$\,kpc.
\section{Discussion}
\subsection{Phase shift}
The most striking discovery from our research is the coherent, wave-like undulation in the velocity space. The spatial frequency of the velocity undulation is similar to that of the spatial undulation, and these undulations exhibit a phase difference of $\approx 2/3 \pi$. This value places our target between being traveling-wave-like
(of phase difference $\pm\,\pi/2$) and being standing-wave-like ($\pm \pi$). This synchronized pattern provides insights into the nature of the structure.
\subsection{Overall Amplitude}
We estimate the full extent within which the oscillation occurs. We assume that the oscillation has a total height of $H_{\rm tot}$, and the velocity of the filament comes from the kinetic energy released when it falls from $H = H_{\rm tot}$ to its current height $H_0$.
We use the following equation to estimate th full amplitude (see online supplementary material for more details):
\begin{equation}
H_{\rm tot} \approx \Big{(} \frac{v_0^2}{4 \pi G \rho_*} + H_0^2 \Big{)}^{1/2} \approx 130 \pm 20 \,\rm pc ,
\end{equation}
where $\rho_*$ is the mass density from stars and dark matter, $v_0$ is the amplitude of the velocity undulation and $H_0$ is the amplitude of the spatial undulation.
We use $v_{\rm 0} = 6\pm 1\,\rm km\,s^{-1}$, $H_0 = 70\pm 6\,\rm pc$ and
$\rho_* = 0.05\,\rm M_{\odot}\,\rm pc^{-3}$ \citep{2003A&A...409..523R}. The oscillation should have an amplitude
of $130\,\rm pc$, meaning that the whole structure is probably oscillating
within a vertical range of around $260\,\rm pc$. We can also estimate the maximum velocity the gas would reach when the it passes through the Galactic disk, which is
\begin{equation}
v_{\rm max} = \sqrt{4 \pi G \rho_* H_0^2} \approx 7\pm 0.6 \,\rm km\,s^{-1} \,.
\end{equation}
The Radcliffe Wave has an estimated height of 130\,pc (or a spatial
extent of 260\,pc). Thus oscillation is not confined to the thick molecular disk
(FWHM = 120\.pc; \citealt{2015ARA&A..53..583H}).
\subsection{Spatial inhomogeneity and Possible Origin}
In line with previous studies, our results suggest that the structure of the
Radcliffe Wave is spatially inhomogeneous: not only do the
undulation appears to be damped such that the amplitude decreases with
increasing $X'$, the Radcliffe
Wave is also tilted, e.g. the spatial offset measured with respect to the $b=0$ plane
decreases by around 100 pc as $X'$ increases from 0 to 1.5 kpc. Should the Gould
Belt Radcliffe Wave be created by a perturbation, the source should be near $X' =0$.
We combine these undulation
patterns to determine the nature of the undulation pattern. If the undulation signature evolves in a way that is similar to that of the travelling wave, we can determine
the direction of the travel. At $X'=$ 1 kpc, seen from
Fig. \ref{fig:phase}, the wave has reached maximum velocity, and are moving
upwards. The corresponding spatial pattern then implies that the wave would be
propagating leftwards. This scenario is unlikely as the wave is damped when
measured from the left to the right. Should the undulation pattern be caused by
a perturbation occurring at $X'=0$, we expect the wave to travel towards the
right.
Because of this, the standing wave {might be} a slightly favoured
scenario. We note, however that as the phase shift changes as a function of the
location, the exact nature of the undulation remains ambiguous.
Recent data from the GAIA
satellite have revealed many phase-space patterns, which can be explained by passages of dwarf galaxies \citep{2018Natur.561..360A,2019MNRAS.489.4962K}. These dwarf galaxies are abundant \citep{2019ARA&A..57..375S}, thus these passages should occur regularly. The undulation may be caused by one of these passages.
\section{Conclusion}
We perform a study of the kinematic structure of the newly-discovered Radcliffe Wave -- a 2.7-kpc long wave-like arrangement of molecular gas. We propose to use the velocities of YSOs, which are young accreting stars that were newly born within molecular clouds, to trace their kinematic structure.
Using our new data,
we identify a similar pattern from the velocity along the Galactic $z$ direction, which might share a common origin as the undulation signature. We call our new coherent, wave-like velocity pattern the ``velocity undulation".
\citet{2020Natur.578..237A} identified an undulation of a period of $\approx 2\,\rm kpc$ and an amplitude of 160 pc from the spatial arrangement of gas. Because of the differences in the models, our analyses yield slightly different results: The left side of the wave is tilted upwards. The wave has a period of 1.5 kpc, a height of $H = 70\pm 6 \,\rm pc$, and a velocity of $7\pm 1\,\rm km/s$.
The velocity undulation is damped similarly to that of the spatial undulation. The spatial and the velocity undulation share an almost identical period of $\approx 1.5\,\rm kpc$, and they exhibit a phase difference of $\approx 2/3 \,\pi$, which places our objects in the middle of being standing-wave-like and traveling-wave-like. We note at $X' = 1 \,\rm kpc$, the gas is sitting close to the disk midplane and is moving upwards (with a positive $v_z$), consistent with the structure behaving either like a standing wave, or a wave traveling towards the left.
From a joint analysis of the spatial and velocity undulation, we estimate that
the wave is possibly oscillating around the Galactic disk mid-plane with an amplitude of $\approx 130\,\pm 20 \,\rm pc$ (e.g. the gas should oscillate within $z = \pm 130\,\rm pc$). The scale on which the oscillation occurs is larger than the mean thickness of the thin molecular disk in the solar vicinity ($H = 73\,\rm pc$), indicating that the wave is moderately perturbed. This, combined with the spatial inhomogeneity suggests that the undulation seen in the Radcliffe Wave can be caused by some perturbations e.g. the passage of a dwarf galaxy.
\section*{Acknowledgements}
This work is partially supported by National Natural Science Foundation of China grants No.\,W820301904, 12173034 and 11833006,
and Yunnan University grant No.~C176220100007 and No.~C176220100028. We acknowledge the science research
grants from the China Manned Space Project with NO.\,CMS-CSST-2021-A09, CMS-CSST-2021-A08 and CMS-CSST-2021-B03.
This work presents results from the European Space Agency (ESA) space mission Gaia. Gaia data are being processed by the Gaia Data Processing and Analysis Consortium (DPAC). Funding for the DPAC is provided by national institutions, in particular the institutions participating in the Gaia MultiLateral Agreement (MLA). The Gaia mission website is https://www.cosmos.esa.int/gaia. The Gaia archive website is https://archives.esac.esa.int/gaia.
\section*{Data availability}
The paper makes use of publicly-available data presented in \citet{2016MNRAS.458.3479M,Gaia2018}.
\bibliographystyle{mnras}
|
\section{Abstract}
\textbf{ Quantum cascade laser (QCL) frequency combs~\cite{hugi2012midinfrared} have revolutionized mid-infrared (MIR) spectroscopy by their high brightness and fast temporal resolution~\cite{klocke2018singleshot}, and are a promising technology for fully-integrated and cost-effective sensors~\cite{schwarz2014monolithically}.
As for other integrated comb sources such as micro-combs~\cite{kippenberg2018dissipative} and interband cascade laser~\cite{bagheri218passively}, QCLs have a comb spacing of several GHz, which is adequate for measurements of wide absorbing structures, typically found in liquid or solid samples~\cite{klocke2018singleshot}.
However, high-resolution gas-phase spectra require spectral interleaving and frequency calibration~\cite{kuse2021frequencyscanned,gianella2022frequency,lepere2022midinfrared}.
We developed a frequency calibration scheme for fast interleaved measurements with combs featuring multi-GHz spacing.
We then demonstrate dual-comb spectroscopy with 600~kHz accuracy in single-shot 54-ms measurements over 40~cm$^{-1}$ using two QCLs at 7.8~$\mathrm{\mu}$m.
This work is an important contribution towards fast fingerprinting of complex molecular mixtures in the MIR~\cite{gianella2020highresolution}. Moreover, the calibration scheme could be used with micro-combs for spectroscopy and ranging, both in comb-swept~\cite{kuse2021frequencyscanned,riemensberger2020massively} and comb-calibrated setups~\cite{delhaye2009frequency}.
}
\section{Main}
Over the last two decades, dual-comb spectroscopy~\cite{coddington2016dualcomb} has developed into a powerful tool for applications requiring fast temporal resolution, high spectral accuracy, and broad spectral coverage. It has enabled difficult or previously impossible measurements, such as multi-species trace gas detection in open-path~\cite{giorgetta2021openpath}, monitoring of irreversible processes with micro-second resolution~\cite{klocke2018singleshot}, hyperspectral 3D imaging~\cite{vicentini2021dualcomb}, and fast determination of isotope ratios of multiple species~\cite{parriaux2022isotope}. Moreover, photonic chip integrated comb sources such as micro-combs~\cite{kippenberg2018dissipative}, quantum cascade lasers (QCL)~\cite{hugi2012midinfrared}, and interband cascade lasers~\cite{bagheri218passively}, are potential game-changers, by enabling mass-producible integrated gas sensors for demanding fields, such as health, security, environment, and industrial process monitoring.
However, these integrated comb sources, due to their small size, have large (compared to fiber mode-locked lasers) repetition rates, $f_\mathrm{rep}$, on the order of 10 GHz or higher. The large $f_\mathrm{rep}$, while allowing a high temporal resolution, leads to a sparse spectrum, where narrow molecular absorption are potentially lost between the comb teeth.
This limitation is not unique to dual-comb spectroscopy or to integrated comb sources, and has been addressed by interleaving many (up to many thousands) offset spectra, where $f_\mathrm{rep}$ and/or the offset frequency, $f_\mathrm{o}$, are tuned in steps~\cite{villares2014dualcomb, muraviev2020broadband, luo2020longwave, hjalten2021optical,lepere2022midinfrared} or continuously~\cite{gianella2020highresolution, kuse2021frequencyscanned}.
Step-tuning of referenced combs typically consists in stepping the repetition rate by 10s of Hz with dwell times on the order of seconds or more adapted to the measurement. Thus, the full interleaved spectrum is acquired in 10s or 100s of seconds and the frequency accuracy is that of the static comb~\cite{muraviev2020broadband,hjalten2021optical}.
We note that step-tuning was also used with free running mid-infrared (MIR) QCLs, where a full spectrum was obtained in 1200~s with an accuracy better than 12 MHz~\cite{lepere2022midinfrared}.
Continuous sweeps yield faster measurements with higher spectral point density. However, they require an adapted frequency calibration scheme.
The endless frequency comb was proposed by Benkler \textit{et al.}~\cite{benkler2013endless}, but it is not convenient in the MIR as it requires electro-optic components and pulsed emission.
Alternatively, the use of an unbalanced Mach-Zehnder interferometer provided frequency scales with 10-MHz level precision for NIR micro-combs~\cite{kuse2021frequencyscanned} and MIR QCLs~\cite{gianella2022frequency}, with sweep times of 50~$\mathrm{\mu s}$ and 30~ms respectively.
In this work, we present a method to calibrate fast continuous sweeps of multi-GHz frequency combs with sub-MHz accuracy, which can be applied independently of the wavelength range.
We thus demonstrate a QCL-based MIR dual comb spectrometer with a unique combination of fast acquisition, high frequency accuracy, and high spectral point density.
As a proof of concept, we measure 43 transitions of the $\nu_1$ ro-vibrational band of $\mathrm{N_2O}$ and 64 transitions belonging to the $\nu_4$ band of $\mathrm{CH_4}$.
With the former measurement, we characterize the accuracy of the computed frequency scale and assess the precision of the measurement at varying scanning speeds. The latter measurement serves to improve the literature accuracy of the transition frequencies of $\mathrm{CH_4}$, independently of two recent works~\cite{lepere2022midinfrared,germann2022methane}.
Rapid scanning and high accuracy are critical features for fingerprinting of complex molecular mixtures and parallel frequency modulated lidar~\cite{riemensberger2020massively}. Accordingly, we trust that our enhanced method for frequency axis calibration is a highly valuable asset for ranging and MIR sensing.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{Figs/Figure1}
\caption{\textbf{Experimental setup and schematic for accurate frequency scale.} (a) Experimental setup featuring the distributed feedback (\texttt{DFB}) laser serving as frequency reference, the gapless dual comb spectrometer, and the setups to measure the repetition rate and the optical frequency during an acquisition. \texttt{MPC}: multi-pass cell, \texttt{AMP}: amplifier, \texttt{BPF}, \texttt{LPF}: band-pass (low-pass) filter, \texttt{DBM}: double-balanced mixer, \texttt{LOGAMP}: logarithmic amplifier, \texttt{EQ}: equalizer, \texttt{COMB GEN}: comb generator, \texttt{WFG}: waveform generator, \texttt{DAQ}: data acquisition unit, \texttt{DS, DN, DR, DF}: photodetector (see main text). (b) Optical spectra of \texttt{COMB1}, \texttt{COMB2}, the position of the reference DFB laser, and the multi-heterodyne beat spectrum. (c) Schematic of the referencing scheme of the optical frequencies of \texttt{COMB1} while they are tuned across more than one free-spectral range ($f_\mathrm{rep}$), relying on the use of a synthetic comb. (d) Absorption spectrum of low pressure methane acquired in $54\,\mathrm{ms}$ by our spectrometer. The top panel is digitally filtered for visibility while the lower panel features two absorption lines at the original resolution of $660\,\mathrm{kHz}$.}
\label{fig:setup}
\end{figure*}
Our dual-comb spectrometer is composed of two QCL frequency combs~\cite{gianella2020highresolution} (\texttt{COMB1} and \texttt{COMB2}) is schematized in Fig.~\ref{fig:setup}(a).
\texttt{COMB1} is the interrogating comb which probes the sample contained in a multi-pass cell, and \texttt{COMB2} is the local oscillator.
The repetition frequency of the combs, $f_\mathrm{rep}\approx9.88\,\mathrm{GHz}$, is too large to sample Doppler-broadened gas-phase spectra of small molecules.
Therefore, we tune the driving current to scan the comb lines across $f_\mathrm{rep}$ and interleave spectra to reduce the spectral point spacing by 4 orders of magnitude.
The local oscillator comb is tuned synchronously to keep the multi-heterodyne spectrum within the detection bandwidth of the system [see Fig.~\ref{fig:setup}(b)].
The interferograms measured on the photodetectors \texttt{DS} and \texttt{DN} are divided into \emph{slices} of length $\tau=13.1\,\mathrm{\mu s}$, which are sufficiently long to resolve the beat notes (beat note spacing, $\Delta f_\mathrm{rep}\approx 4.5\,\mathrm{MHz}$).
A lower limit for the spectral resolution, $\delta\nu=\dot{\nu}\cdot(\tau/3.77)$, can be estimated from the slice length, $\tau$, and the tuning rate, $\dot{\nu}$.
The factor $3.77$ stems from the employed flattop apodization.
For the tuning rates (scan duration) $\dot{\nu}=380\,\mathrm{GHz/s}$~$(27\,\mathrm{ms})$, $190\,\mathrm{GHz/s}$~$(54\,\mathrm{ms})$, $95\,\mathrm{GHz/s}$~$(107\,\mathrm{ms})$, and $47\,\mathrm{GHz/s}$~$(215\,\mathrm{ms})$ (see below) we find, respectively, $\delta\nu=1.3\,\mathrm{MHz},\,660\,\mathrm{kHz},\,330\,\mathrm{kHz},\,160\,\mathrm{kHz}$.
For example, the spectrometer can acquire spectra spanning $40\,\mathrm{cm^{-1}}$ in $54\,\mathrm{ms}$ at a point spacing and resolution of $0.66\,\mathrm{MHz}$, allowing measurement of low-pressure methane [see Fig.~\ref{fig:setup}(d)].
The noise-equivalent absorbance and noise-equivalent absorption coefficient depend on the signal-to-noise ratio (SNR) of the corresponding beat note [see Fig.~\ref{fig:setup}(b)].
For strong beat notes, they are of the order of $3\times10^{-4}\,\mathrm{Hz^{-1/2}}$ and $8\times10^{-8}\,\mathrm{cm^{-1}Hz^{-1/2}}$, respectively, given the absorption path length of $36\,\mathrm{m}$ in the multi-pass cell.
The distributed feedback QCL (\texttt{DFB}) is locked to a molecular transition ($\mathrm{N_2O}$, fundamental $\nu_1$ band, P(14)) and acts as optical frequency reference with frequency $\nu_\mathrm{ref}$~\cite{komagata2022absolute}.
The beat frequency, $f_b$, between the reference laser and \texttt{COMB1} is detected on a $1.2\,\mathrm{GHz}-$bandwidth detector (\texttt{DF}).
The frequency, $\nu_\mu(t)$, of all lines, $\mu=0,\pm 1,\pm 2,...$, of the interrogating comb over the time of the scan is computed from the measured repetition rate, $f_\mathrm{rep}(t)$, and beat frequency, $f_b(t)$, at each instant of the sweep, and is given by
\begin{equation}
\nu_\mu(t)=\nu_\mathrm{ref}+f_b(t)+\mu\cdot f_\mathrm{rep}(t).
\label{eq:freq-axis}
\end{equation}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{Figs/Fig2_markers.pdf}
\caption{\textbf{Assessment of $f_b$ and $f_\mathrm{rep}$ over a gapless measurement (sweep range $\geq f_\mathrm{rep}$).} (a) Filtered marker signal as a function of time (solid lines) and peaks localized by the algorithm (filled circles). The mean position of two peaks gives the marker position shown by a star, plus or x. (b) Zoom over 3 pairs of peaks. (c) Frequency of the markers as a function of time. (d) Frequency of the markers minus the linear trend and polynomial fitting of the comb line frequency. The first 4 pairs of marker originate from the beat with another comb line ($\mu=1$ instead of $\mu=0$). (e) Residuals from the fit from (d). (f) Measured $f_\mathrm{rep}$ using two methods as a function of time and fit. The inset shows the a zoom over a smaller time interval. (g,h) Residuals of $f_\mathrm{rep}$ measured through the fast processor (g), which is limited by the integration time, or measured with high resolution (H.R.) (h). The right plot shows a zoom over a smaller interval. The fit using the processor data is accurate for the H.R. data within 2.4 kHz. }
\label{fig:marker}
\end{figure*}
The measurement of $f_b(t)$ is based on the open-loop solution proposed in 2009 by Del'Haye \textit{et al.}~\cite{delhaye2009frequency} with a few notable changes.
Here, the reference laser is fixed while the comb is swept by approximately $f_\mathrm{rep}$.
Therefore, a single pair of peaks is detected through RF path 1 [see Fig.~\ref{fig:setup}(a)].
These are observed at the instants when the band-pass filter \texttt{BPF1} transmits the beat to the rectifying logarithmic amplifier (\texttt{LOGAMP}).
The recorded signal on channel 1 of \texttt{DAQ2} is shown in Fig.~\ref{fig:marker}(a,b) as \emph{Signal 1}.
We implement a second RF path [RF path 2, see Fig.~\ref{fig:setup}(a)], where we mix the beat with a $f_\mathrm{RF}=475\,\mathrm{MHz}$-spaced harmonic RF comb.
The RF comb provides a synthetic grid of frequencies centered on the reference laser frequency, $\nu_\mathrm{ref}$, and having a spacing of $f_\mathrm{RF}$ [see Fig.~\ref{fig:setup}(c)], whereby the time of passage of $\nu_0$ across each grid line is measured.
Hence, the determination of $f_b(t)$ (or $\nu_0(t)$) is based on the measurement of the times at which $f_b$ equals multiples of $f_\mathrm{RF}$.
These markers allow to map the frequency of $f_b(t)$ (or $\nu_0(t)$) versus time, as shown in Fig.~\ref{fig:marker}(c).
The marker on channel 1 allows identification of $\nu_0(t_0)$ modulo $f_\mathrm{rep}$.
The $5\,\mathrm{GHz}$ RF bandwidth required to measure $f_b$ throughout the sweep, corresponding to $f_\mathrm{rep}/2$, poses a challenge, as the specified bandwidth of the photodetector is only $1.2\,\mathrm{GHz}$.
We thus used logarithmic amplifiers to increase the marker detection dynamic range and gain equalizers (\texttt{EQ}) to reduce signal strength imbalances.
Moreover, in order to further balance the signal strength and optimize the RF mixers, the signals were split below and above $3.8\,\mathrm{GHz}$ [not shown in Fig.~\ref{fig:setup}(c)]. Further information can be found in Methods and Supplementary Note 1.
We plot the markers after removal of the linear trend in Fig.~\ref{fig:marker}(d). The first 4 sets of markers originate from the beating with a higher comb line $\mu=1$ [see Fig.~\ref{fig:setup}(c)], which was unavoidable to accommodate a sweep about 10 \% larger than $f_\mathrm{rep}$.
We fit the markers with a 10th order polynomial using a two comb line model separated by $f_\mathrm{rep}(t)$.
The algorithm is robust against spurious peaks, which produce markers that can be discarded if they don't fit within a certain tolerance region.
The residuals of the fit are shown in Fig.~\ref{fig:marker}(e). The standard deviation of the fit residuals is found to be $241\,\mathrm{kHz}$.
The repetition rate, detected by the horn antenna, is down-mixed to $\sim20\,\mathrm{MHz}$, digitized on \texttt{DAQ1}, channel 2, and processed together with the interferogram [see Fig.~\ref{fig:setup}(a)].
The resolution is limited to $76\,\mathrm{kHz}$ due to the finite length ($\tau=13.1\,\mathrm{\mu s}$) of the slices [see Fig.~\ref{fig:marker}(f)].
We fit these data using a 10th order polynomial.
The residuals are limited by the resolution as shown in Fig.~\ref{fig:marker}(g,h).
On a control measurement, we recorded the raw digitized data from \texttt{DAQ1}, channel 2, and computed $f_\mathrm{rep}(t)$ with higher resolution.
The residuals between the fit from the low resolution data and the high resolution data are shown in Fig.~\ref{fig:marker}(h), showing correspondence within $2.4\,\mathrm{kHz}$.
The high-resolution data feature oscillations of $f_\mathrm{rep}$ due to back-reflection from the photodetectors.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{Figs/Figure3.png}
\caption{\textbf{Spectra and retrieved line centers in a $54\,\mathrm{ms}$ acquisition} (a) Single shot spectra of the $\nu_1$ fundamental band of nitrous oxide with $660\,\mathrm{kHz}$ spectral point spacing, acquired in 54-ms and comparison with simulation. (b) Retrieved line centers compared with literature reference with 1-$\sigma$ error bars. (c-f) Zoom of (e) on 4 different lines illustrating stronger/weaker absorbance and high/low SNR, and residuals of the Voigt fit shown in red.}
\label{fig:spectrum}
\end{figure*}
Figure~\ref{fig:spectrum}(a) shows a spectrum of the fundamental $\nu_1$ band of nitrous oxide acquired in a single sweep of $54\,\mathrm{ms}$.
The bimodal comb amplitude distribution gives rise to a variable SNR across the spectrum, as can be seen in Figs.~\ref{fig:spectrum}(c-f). We used a Voigt fit to retrieve the line centers of the molecular transitions.
We benchmark the frequency accuracy of our spectrometer by comparing the computed line center frequencies to those obtained by two other groups with 100-kHz level accuracy~\cite{alsaif2018high, hjalten2021optical}.
Based on multiple sets of measurements over the course of a few weeks, we estimate that we retrieve, in a single-shot acquisition, 40 line centers (spanning $30.4\,\mathrm{cm^{-1}}$) with a 1-$\sigma$ accuracy better than $2\,\mathrm{MHz}$, among which 23 lines (spanning $17.1\,\mathrm{cm^{-1}}$) are retrieved with an accuracy better than $0.6\,\mathrm{MHz}$, as can be seen in Fig.~\ref{fig:spectrum}(b).
This estimation includes the uncertainties of the reference laser frequency $\nu_\mathrm{ref}$ ($180\,\mathrm{kHz}$ after correction of a systematic bias of $(180\pm100)\,\mathrm{kHz}$), $f_\mathrm{rep}$ ($2.6\,\mathrm{kHz}$ after the correction of a systematic bias of $(2.9\pm0.7)\,\mathrm{kHz}$), the fit of the markers as in Fig.~\ref{fig:marker}(e) ($240\,\mathrm{kHz}$), and the standard deviation of the retrieved line center frequencies over 12 consecutive 54-ms measurements (below $300\,\mathrm{kHz}$ for the lines with the best SNR, see Fig.~\ref{fig:scanning_speed}).
The retrieved line centers and the uncertainties are consistent with the literature values.
The frequency accuracy of the spectrometer allows us to retrieve the line centers of $\mathrm{CH_4}$ with a higher accuracy than the current literature~\cite{gordon2022hitran2020}. Our results are given in Table~\ref{table:methane}. For this measurement, we averaged the retrieved line centers over 15 measurements.
\begin{table}
\centering
\begin{tabular}{|c|c|}\hline%
\bfseries Transition & \bfseries Line center\\
(J',C',$\alpha$') - (J'',C'',$\alpha$'') & (cm$^{-1}$)\\\hline
\csvreader[
head to column names,
separator=semicolon,
late after line=\\,
late after last line=\\\hline
]{transitionExport2.csv}{}{%
\Transition & \FrequencyMeas
}%
\end{tabular}
\caption{Measured line list of the methane $\nu_4$ band. The upper (lower) levels are labeled with a single (double) prime and are denoted according to the Hitran~\cite{gordon2022hitran2020} notation.}
\label{table:methane}
\end{table}
\begin{figure*}
\centering
\includegraphics{Figs/Figure4.png}
\caption{\textbf{Frequency precision for varying scanning speed.} Frequency error and statistical spread of the retrieved line centers at 4 different scan speeds, decreasing from left to right by factors of 2. The precision increases with the scan time.}
\label{fig:scanning_speed}
\end{figure*}
The scanning speed of the spectrometer can be varied to measure spectra in between $27\,\mathrm{ms}$ to $216\,\mathrm{ms}$.
We estimate an accumulated chirp~\cite{riemensberger2020massively} rate of $44\,\mathrm{THz/s}$ to $5.6\,\mathrm{THz/s}$, considering 120 lines which sweep in parallel.
The sweep becomes discontinuous at faster rates.
Slower tuning is possible, but requires a larger buffer in the acquisition unit.
Figure~\ref{fig:scanning_speed} displays the statistics in the form of box plots of the retrieved $\mathrm{N_2O}$ line centers with respect to literature reference for varying scanning speeds.
Scan speeds up to $380\,\mathrm{GHz/s}$ do not induce noticeable offsets in the retrieved line centers, within the uncertainties.
This shows that delays due to the RF scheme are negligible.
The measurement at 380~GHz/s was performed at a later date. Potential delays at faster scanning speeds could be assessed and corrected from a reference measurement.
We have demonstrated a fast and accurate mid-infrared dual-comb spectrometer.
The combs were emitted by quantum cascade lasers, which can be tuned over their free spectral range ($f_\mathrm{rep}$) in tens of milliseconds to provide high-resolution spectra of broad molecular bands.
We calibrated the frequency axis by measuring the beat between the interrogating comb and a reference laser and by acquiring the comb's repetition rate.
Moreover, we have enhanced the marker method~\cite{delhaye2009frequency} using a synthetic comb, to process the fast chirp of the beat between the comb and the reference laser. This method is compatible with other integrated comb technologies featuring multi-GHz repetition frequencies such as micro-resonator combs~\cite{kippenberg2018dissipative}, and could also be applied to comb-calibrated spectroscopy~\cite{delhaye2009frequency}.
Thus, our agile spectrometer provides unique performances in terms of acquisition speed and frequency accuracy compared to other broadband spectrometers with small ($<10\,\mathrm{MHz}$) spectral point spacing~\cite{delhaye2009frequency, nishiyama2013high, alsaif2018high, gotti2020comblocked, hjalten2021optical}.
Our spectrometer is a promising candidate towards an on-chip gas sensor that is highly sensitive and selective to many relevant molecules, allowing real-time monitoring of complex gas mixtures at Hz-level refresh rates.
Finally, we believe that the spectrometer sensitivity, bandwidth, and power-efficiency will improve thanks to ongoing development in quantum cascade laser frequency combs, such as mutual stabilization~\cite{komagata2021coherentlyaveraged, hillbrand2022synchronization}, dispersion management~\cite{villares2016dispersion}, and full spectrometer integration~\cite{schwarz2014monolithically}.
\section{Methods}
\subsection{Experimental setup}
The two laser sources were InGaAs/AlInAs on InP-based dual-stack QCLs~\cite{jouy2017dual,gianella2020highresolution}. They were $4.5\,\mathrm{mm}$ long ($f_\mathrm{rep}\approx9.88\,\mathrm{GHz}$, $\Delta f_\mathrm{rep}\approx4.5\,\mathrm{MHz}$) and had both end facets uncoated.
The combs emitted several hundred mW of optical power over $\sim50\,\mathrm{cm^{-1}}$ in the $1300\,\mathrm{cm^{-1}}$ range [see Fig.~\ref{fig:setup}(b)].
They were operated at constant temperature of $1.9^{\circ}\mathrm{C}$ and $2.2^{\circ}\mathrm{C}$, respectively.
The combs were electrically driven by a pair of QubeCL drivers (ppqSense).
The bias currents for the two combs were $1532\,\mathrm{mA}$ and $1226\,\mathrm{mA}$, respectively.
The currents were modulated by asymmetric triangular waveforms (\texttt{WFG1} and \texttt{WFG2} are the two outputs of an ArbStudio 1102, LeCroy waveform generator) with the currents initially decreasing at constant rate during a time $T=30\,\mathrm{ms}$, $60\,\mathrm{ms}$, $120\,\mathrm{ms}$, or $240\,\mathrm{ms}$ and returning to the initial value over a shorter time $0.25\cdot T$.
The peak-to-peak current modulation amplitude was approximately $90\,\mathrm{mA}$ for both lasers.
This required a modification of the QubeCL drivers which, by default, do not allow such large current modulations.
The beam splitters were custom-made to obtain the appropriate power ratios at each photodetector. They were designed for reflecting 1\%, 10\% or 50\% at an angle of incidence of 45° for s-polarized light from 1220-1370$\,\mathrm{cm}^{-1}$. They were produced by ion beam sputtering (IBS) on 5 mm thick wedged (0.5 deg) $\mathrm{CaF_2}$ substrates with an anti-reflection coating on the back side to minimize undesired interference effects.
Two $1\,\mathrm{GHz}-$bandwidth MIR photodetectors were employed to measure the sample (\texttt{DS}) and to normalize laser intensity and phase noise (\texttt{DN}).
The detector outputs were sampled at $2.5\,\mathrm{GSa/s}$ with 12 bit resolution by the data acquisition unit \texttt{DAQ1} (ADQ32, Teledyne), generating $10\,\mathrm{Gbyte/s}$ of raw data for the duration of the sweep.
The data acquisition was triggered at $t=0.07\cdot T$ after the start of the (decreasing) current ramp and continued for a duration of $27\,\mathrm{ms}$, $54\,\mathrm{ms}$, $107\,\mathrm{ms}$, or $215\,\mathrm{ms}$, which corresponds to $2^{26}$, $2^{27}$, $2^{28}$, or $2^{29}$ samples at $2.5\,\mathrm{GSa/s}$.
The clocks of \texttt{DAQ1} (Fig.~\ref{fig:setup}(a)) and of the synthesizer generating the $475\,\mathrm{MHz}$ fundamental for the synthetic comb generator were referenced to a GPS-disciplined clock (GPSDO, Leo Bodnar).
A custom software (C++ and CUDA) was used to process the two interferograms.
The program is capable of fully processing (from raw data to absorbance values) 25 $27\,\mathrm{ms}$-long sweeps per second, corresponding to a raw data throughput of $6.75\,\mathrm{Gbyte/s}$, or 67.5\% of the maximum raw data input rate.
The interferograms were first divided into \emph{slices} of length $2^{15}=32,768$ samples, corresponding to $13.1\,\mathrm{\mu s}$.
The beat note amplitudes were computed for each slice by fast Fourier transform, from which the sample's absorbance was determined.
A distributed feedback QCL (\texttt{DFB}, Alpes Lasers) emitting a continuous wave (cw) beam was locked to a molecular transition ($\mathrm{N_2O}$, fundamental $\nu_1$ band, P(14)) by means of a wavelength modulation scheme and acted as optical frequency reference with frequency $\nu_\mathrm{ref}$.
\subsection{Measurement of the markers and retrieval of the frequency axis}
The heterodyne beat between the DFB reference laser and \texttt{COMB1} is detected on the fast photodetector (UHSM-10.6 PV-4TE-10.6-0.05-butt, Vigo System).
Through RF path 1, two pulses are transmitted in rapid succession when $f_b(t)$ passes near zero and are amplified by a logarithmic amplifier (ZX47-60LN-S+, Minicircuits).
The timings of the two peaks, $t_0^{\pm}$, fulfill $f_b(t_0^{\pm})=\pm f_\mathrm{bp,1}$, where $f_\mathrm{bp,1}=25\,\mathrm{MHz}$ is the center frequency of \texttt{BPF1}.
We assume linear chirping of the beat between $t_0^-$ and $t_0^+$, so that $f_b(t_0)=0$, where $t_0$ is the mean of $t_0^+$ and $t_0^-$.
With Eq.~\ref{eq:freq-axis} we find $\nu_0(t_0)=\nu_\mathrm{ref}$.
As for RF path 2, $f_b(t)$ is mixed with an RF comb with spacing $f_\mathrm{RF}$. Therefore, each time the \emph{DFB-comb} beat, $f_b(t)$, nears a multiple of $f_\mathrm{RF}$, two double pulses are recorded on channels 2 and 3 of \texttt{DAQ2} (\emph{Signal 2} and \emph{Signal 3} in Fig.~\ref{fig:marker}(a,b)).
The timings, $t_n^{\pm}$ of the two pulses of the $n$th ($n=\pm1, \pm2,...$) double pulse fulfill $f_b(t_n^{\pm}) = n\cdot f_\mathrm{RF} \pm f_\mathrm{bp,i}$, where $f_\mathrm{bp,2}=21\,\mathrm{MHz}, f_\mathrm{bp,3}=45\,\mathrm{MHz}$ are the center frequencies of \texttt{BPF2} and \texttt{BPF3}, respectively.
For the mean times, $t_n=(t_n^-+t_n^+)/2$, we find, as before, $f_b(t_n)=n\cdot f_\mathrm{RF}$ and with Eq.~\ref{eq:freq-axis}, $\nu_0(t_n)=\nu_\mathrm{ref}+n\cdot f_\mathrm{RF}$. More details concerning the processing of the markers can be found in Supplementary Figure 1 and Supplementary Note 1.
As post-processing, we first smooth the marker signals with a low-pass filter.
Then we detect the peaks which match specific criteria in terms of prominence, width, and proximity to other lines using a peak-finding algorithm (find\_peaks, scipy).
If two peaks have the appropriate time delay, we establish that $f_b$ was at a multiple integer $n$ of $f_\mathrm{RF}$ at the mean time of the two peaks.
The integer $n$ is guessed from linear interpolation using marker 1 and an estimated chirp.
The discontinuity from the change of comb line can be easily recognized and the markers originating from comb line $\mu=1$ can be used to determine the frequency of comb line $\mu=0$ by knowledge of $f_\mathrm{rep}(t)$.
We fit the markers with polynomials of increasing orders up to 10, keeping the coefficients of the previous order as initial guess and providing the initial guess for the new coefficient from the residuals of the previous fit.
A larger weight is given to marker 2 as \texttt{BPF2} is narrower than the other band-pass filters.
After the fits with the $4^\mathrm{th}$ and $8^\mathrm{th}$ order polynomial, markers outside a tolerance of $50\,\mathrm{MHz}$ and $10\,\mathrm{MHz}$ are discarded, respectively.
This allows the removal of most spurious peaks, which do not align in the plot Fig.~\ref{fig:marker}(d).
\subsection{Measurement of the repetition rate}
In QCL frequency combs, the beating of all co-existing modes in the laser produces a measurable voltage modulation on the laser electrodes with frequency equal to $f_\mathrm{rep}$~\cite{piccardo2018timedependent}.
The laser module features a $50\,\mathrm{\Omega}$ transmission line on a printed circuit board that connects an SMA connector at the back of the housing to a point very close (within $1\,\mathrm{mm}$) to the back facet of the laser.
From there, a short bond wire establishes an electrical connection to the laser's top electrode.
This pathway can be used for RF injection or to measure $f_\mathrm{rep}$.
However, connecting cables and other circuitry to the SMA port tends to increase laser phase noise.
It is observable as a broadening of the beat notes.
Therefore, a horn antenna (\texttt{ANT}, PowerLOG 40040, Aaronia) was pointed at the open SMA connector. The detected $f_\mathrm{rep}$ was then amplified and down-mixed to $\sim20\,\mathrm{MHz}$ to enable its measurement on one channel of \texttt{DAQ1}.
\subsection{Gas sample preparation and measurement}
We dynamically diluted $\mathrm{N_2O}$ gas with nitrogen to a concentration of about of $450\,\mathrm{ppmv}$ and flowed the diluted mixture through the multi-pass cell (\texttt{MPC}).
We then closed off the cell and pumped down to $1.5\,\mathrm{mbar}$, which provided sufficiently strong absorption lines and sufficiently low pressure to be reasonably close to the Doppler-broadened limit and, more importantly, to minimize the pressure-induced shift on the measured transitions.
Multiple sweeps were taken but not averaged together.
The time delay between sweeps was several seconds, to ensure proper synchronization of the two data sets acquired by the two DAQs, which were read out by independent software.
By integrating both DAQs in the same processing routine, the acquisition rate (sweeps per unit time) could be increased significantly (about 12 $54\,\mathrm{ms}$-long sweeps per second).
\section*{Acknowledgments}
We thank Prof. Jérôme Faist and Dr. Mathieu Bertrand at ETH for providing the QCL combs and lending the horn antenna used in this work. We also thank Dr. Stéphane Schilt for fruitful discussions concerning the marker scheme.
\section*{References}
|
\section{Cheats and how you can used them}
To come clean.
\begin{itemize}
\item
{\em Up to} $1.1\ 10^{12}$ GPop/s is from one run.
The median $0.65\ 10^{12}$ GPop/s is more typical
\cite[Tab.~3]{Langdon:2022:ALJ}.
\item
We used 16 threads on one node of a shared cluster
which supported 48 threads.
(The cluster job scheduler used the other 32 threads
for a wide mix of jobs during our runs.)
\item
The operating system manages the individual CPU core clocks
and is free to reduce them below 3.0GHz.
\item
This is the mean performance across the whole run
and so includes everything,
including gathering statistics.
(In this and another run
we turned off gathering some unused population statistics,
such as calculating the height of the trees.)
\item
{\em Equivalent of}.
This is the one that matters.
The headline figure is based on assuming each GP opcode
is interpreted on each test case.
We really do use every test case every time but the big
result is to avoid repeated work,
Section~\ref{sec:avoidwork}.
\hspace{1ex}
Each generation is created from its parent generation.
We have evaluated all of the parents.
Their children are only marginally different.
We do our best to inherit their fitness evaluation
within the constraints of not using more memory.
\end{itemize}
\section{Background}
My goal is summarise recent improvement in genetic programming
implementations in the hope that the ideas used may be beneficial
elsewhere.
Section~\ref{sec:small_mem_ga} describes how to reduce
memory usage by more than 50\%,
whilst
Section~\ref{sec:par} is concerned with exploiting parallel hardware.
In converged populations of enormous trees
incremental fitness evaluation~\cite{langdon:2021:EuroGP}
(Section~\ref{sec:avoidwork})
proved so successful at reducing fitness evaluation run time
that crossover became the dominant cost,
so Section~\ref{sec:fast_xo} describes ways
to reduce the cost of crossover.
The success of incremental evaluation also lead to
consideration of information flows within GP trees
and more widely in hierarchical structures,
inspiring failed disruption propagation, FDP~\cite{Petke:2021:FSE-IVR},
and the realisation that deep nested structures will be
resistant to evolution and,
as discussed in Section~\ref{sec:FDP},
the suggestion that to evolve large complex artefacts,
they will have to be composed of many shallow
sub-modules~\cite{langdon:2021:sigevolution}
(see Figure~\ref{fig:lung_like},
page~\pageref{fig:lung_like}).
Details can be found in the published papers and C++~code.
\subsection{Faster Genetic Programming}
Genetic Programming~\cite{koza:book},
like many Artificial Intelligence techniques based on search,
can be expensive in terms of computer resources.
Indeed Koza in the first GP book~\cite{koza:book}
describes ways to speed it up~\cite{poli08:fieldguide}.
Until recently (Section~\ref{sec:fast_xo}),
as is common with evolutionary computing in general,
almost the whole GP runtime is taken by
the time to do fitness evaluation
and so speed up techniques have concentrated upon
faster fitness functions.
\subsection{Faster Fitness}
\label{sec:fasterfit}
Almost all GP (as we do) interprets the evolved code.
However, particularly when GP individuals are run many times,
e.g.\ because there are many test cases,
it can be beneficial not to interpret the code
but to compile it
and run the compiled binary code
(e.g.\ Harding~\cite{eurogp07:harding},
Gregor and Spalek~\cite{Gregor:2016:ELEKTRO}).
However running a traditional compiler is expensive
and
\label{p.Fukunaga}
Fukunaga et al.~\cite{fukunaga:1998:gchpGP} and
Vasicek and Slany~\cite{vasicek:2012:EuroGP}
have advocated generating
machine code inside the GP system
and executing that.
Indeed Nordin~\cite{Nordin:1995:tcp,nordin:thesis}
evolved executable machine code directly,
leading to Discipulus~\cite{francone:manual}
which evolved Intel x86 binary machine code~%
\cite{foster:2001:discipulus}.
\subsection{Better Fitness}
Fundamentally the goal of the fitness function is not
to measure how well an evolved individual works
or solves a problem but to guide search
and essentially fitness and selection boils
down to a single bit:
does this individual get a child or not.
The evolutionary computing community,
puts a huge amount of effort into the fitness function,
whose output is often corrupted by the addition of
stochastic selection noise.
This single (noisy) bit is often extremely expensive.
Many techniques have been advanced to reduce the overall cost
of the fitness function and/or provide better results.
Sometimes the goal is to reduce costs but
(also) it may be to drive evolution in a different direction.
Recognising that the child/no child decision is often made
using tournament selection,
Teller and Andre~\cite{Teller:1997:acnfc} devised a scheme
which aimed at using statistical tests whereby,
assuming a random sample,
only enough tests are run by the fitness function to
reliably tell which of a candidate set of parents
would win the tournament if all the tests were run.
That is, their ``rational allocation of trials'' (RAT)
aims (although with some noise)
to drive evolution as before but with considerably
less effort.
Treating individuals with zero children as pure overhead to the GP system,
Tackett proposed ``soft brood selection''
\cite{Tackett:1994:broodGP},
which aims to dispose of likely unfit children
without going to the expense of growing them all to adulthood.
When evolving agents~\cite{langdon:1998:antspace}
or programs~\cite{Langdon:2012:mendel,Petke:gisurvey},
an individual which enters a never ending loop will have low fitness
but high fitness evaluation cost.
To reduce fitness effort,
it is common in such cases to impose a timeout
and cut short fitness evaluation
when the imposed time limit is reached.
However the chosen limit is application specific.
(This also has the unfortunate effect
that when different limits are chosen for the same ``benchmark''
\cite{langdon:1998:antspace}
subtle difference arise
between supposedly the same system.)
To try and reduce the impact of arbitrary time limits,
Maxwell proposed a ``coroutine'' model~\cite{icec94:maxwell}
in which (like a conventional operating system time sharing system),
each member of the population gets a quantum of computer time
to run fitness tests.
Where selection,
i.e.~child/no child, is not clear cut,
aspiring individuals may get additional quanta to try and assemble more fitness points.
Again the goal is to avoid ``wasting'' fitness effort,
but accepting the introduced noise may drive evolution in another direction.
Gathercole~\cite{ga94aGathercole,gathercole:thesis}
proposed schemes for dynamically changing which
fitness tests are used as evolution proceeds.
Gathercole's DSS scheme was used commercially
\cite{francone:manual}.
Spector's Lexicase~%
\cite{Spector:2012:GECCOcompA}
can be thought of having elements of both RAT and DSS\@.
Again it runs only a varying subset of the available fitness test cases
but its primary motivation is to
evolve a better solution, rather than faster fitness evaluation.
Some of these schemes have elements of
coevolution~\cite{Reynolds:1994:tag}
about them.
In which the fitness test cases are thought of as a separate
population and the primary goal is better evolution
rather than reduced fitness effort.
\subsection{\mbox{Distributed Populations: Slower Convergence}}
The implementation changes to GPquick
(Sections~\ref{sec:small_mem_ga} to~\ref{sec:avoidwork})
allow the study of the evolution of far bigger
genetic programming systems
within the constraints of the available computers.
The implementation takes pains to sequester
the use of pseudo random operations,
and synchronise parallel operations,
so that the results are identical
to those before the improvement were made
and experiments can be repeated.
By relaxing the exact reproducibility constraint you can
ease the implementation of
exciting stochastic effects;
such as studying immigration/emigration
between demes and island sub-populations.
These may not only easily allow operation over
multiple computers (Section~\ref{sec:par})
but also delay convergence~%
\cite{langdon:book},
\cite{DBLP:journals/heuristics/AffenzellerW04}.
\section{\mbox{Generational $\le$\! Steady State Memory}}
\label{sec:small_mem_ga}
Although introduced for the benefit of genetic programming
populations with large trees,
these algorithms apply to any evolutionary algorithm~(EA) with
a fixed population size~$M$.
For example, they might be useful for
very large traveling salesman problems (TSP),
such as Drori et al.'s proposed billion city
Galaxy TSP benchmark~\cite{drori2018galaxytsp}.
Long ago
Koza et al.~\cite[pages 1044-1045]{koza:gp3}
showed that a generational evolutionary algorithm need
take no more memory than a steady state EA\@.
Essentially at the start of each generation
during crossover and mutation they maintain multiple
queues of parents according to the number of children
they have yet to create.
As each crossover or mutation is performed,
the parents' number of unborn children are reduced
and when necessary the parents are moved between queues.
Parents with no yet to be created children
are deleted immediately,
allowing their buffers to be used by new children.
The queue of parents with only one unborn child
is dealt with first.
If it is empty, the queue holding parents
with two children waiting to be created is processed.
Of course processing the first item in this queue,
may result in moving individuals
to the first queue.
If so, it will again be processed first.
Koza et al.\ allow two parent crossover to create two children.
Their queues can be slightly simplified if each
genetic operation only generates one child
\cite{langdon:2020:small_mem_ga_2pages}.
The essential idea remains;
process those individuals which will free resources first.
I have not done this,
but the individual parents within each queue
could be ordered to process
those that will free most memory first.
(Note the current implementation uses un-ordered queues
so that each operation takes constant time
and the overhead is more-or-less the same as before.)
Alternatively,
to balance the threads
in a multi-threaded environment,
you might want
to start processing operations that will take the longest time first.
Unlike incremental fitness evaluation~\cite{langdon:2021:EuroGP},
the time taken by crossover or mutation can be accurately estimated
in advance from the size of the crossover fragments to be moved
and the computer's memory bandwidth.
Like my earlier implementations of GPquick~\cite{singleton:byte},
there are many
generational EA systems which
for a population of $M$~individuals
simply create the new generation whilst keeping the old
one and thus need twice as much memory (i.e.~$2M$ buffers).
Using a single threaded version of Koza et al.\ scheme
and two parent crossover
a single threaded implementations needs at most $M+2$
buffers.
As my trees became bigger,
it became important to move GPquick from using
$2M$ buffers.
Indeed as fitness evaluation got faster
\cite{langdon:2019:gpquick,langdon:2020:cec}
it became worthwhile running crossover
(like fitness evaluation)
in parallel.
With synchronisation locks the child birthing
queues can be processed in parallel.
Notice the order they are processed in now depends
upon the exact thread timings and will typically vary between runs.
However all the tasks are processed and the results
are the same.
GPquick generates one child per crossover.
The parallel Linux pthreads~($t$) version needs at most
$M+2t$ buffers
\cite{langdon:2020:small_mem_ga_2pages,langdon:2020:small_mem_ga}.
If low fitness individuals are discarded,
with two parent crossover,
in practise
the memory can be reduce
to about $0.65M$.
With some virtual memory systems,
it may be possible to allocate $M+2t$ buffers
but in most generations actually only use two thirds of the memory.
Reducing memory usage below $M$ buffers
has become possible
due to
innovations introduced to reduce the cost of
memory buffer copies in crossover.
Such as
evaluating fitness before crossover
(Section~\ref{sec:fitness_first})
and extracting and separately storing
the small subtrees to be donated to children
in the next generation
\cite{Langdon:2021:GECCO,Langdon:2021:GPTP}
(also Section~\ref{sec:fatherless}).
\subsection{GPquick uses Flatten Linear GP Trees}
For efficiency~\cite{kinnear:keith},
GPquick~\cite{singleton:byte} flattens each GP tree
into a linear array of opcodes.
Reproduction, mutation, crossover and fitness evaluation
all use this.
For work on parallel graphics accelerators
(GPUs)~\cite{langdon:2008:eurogp}
(see Section~\ref{sec:gpu}),
the normal depth first tree order can be replaced
by reverse polish notation (RPN),
which processes in left to right order the RPN expression,
explicitly pushing intermediate results on to a stack,
rather than recursively processing the tree (with an implicit stack).
In both cases the depth of the stack used is the tree height.
Again reproduction, mutation, crossover and fitness evaluation
all use the linear buffer,
avoiding the need for data conversion between the CPU and GPU\@.
To avoid heap fragmentation GPquick
allocates a user defined number of fixed sized byte arrays
to hold the GP population at the start of the run.
Usually the flatten evolved tree does not fill the whole of its
buffer,
in which case the contents of the rest of the buffer is left
undefined.
For applications with many test cases,
GPquick has a FASTEVAL compile time option.
If FASTEVAL is enabled,
a single pass through each GP tree is inserted before fitness
evaluation,
which replaces each byte sized opcode by a pointer
(occupying 8 bytes)
to the function inside the fitness evaluation code responsible
for evaluating the opcode.
For every test case,
this avoids a table look up to translate each opcode to
a function pointer.
However, for simplicity and to save memory,
I did not use FASTEVAL\@.
Notice the FASTEVAL pass is analogous to
the translation from GP tree to
binary machine code advocated by
Alex Fukunaga and others mentioned above
in Section~\ref{sec:fasterfit}.
\section{Parallel Evaluation}
\label{sec:par}
It has long been recognised that GP in particular
and evolutionary computing in general,
are well able to take advantage of parallel computing,
e.g., \cite{Fraser:1994:inkbiro},
\cite{juille:1995:fgSIMD,pollack:1996:aigp2},
\cite{Koza:1995:pGPnt},\cite{bennett:1999:BPCSPHPD},
\cite{PDPTA96b},
\cite{bolis:2001:EuroGP,Maitre:2013:ecgpu},
\cite{DBLP:conf/ppsn/ArenasCEJGPPS02},
\cite{Cahon:2004:JoH,DBLP:journals/heuristics/AffenzellerW04},
\cite{harding_genetic_2008,langdon:2008:eurogp,langdon:2008:SC},
\cite{FernandezdeVega:pdci},
\cite{langdon_2012_sigevolution},
\cite{Gong:2015:ASC},
\cite{Burlacu:2020:GECCOcomp},
\cite{LaCava:2021:NeurIPS}.
\subsection{Parallel Threads and Multi-Core CPUs}
\label{sec:pthreads}
The previous section has already mentioned
the use of Linux pthreads to perform crossover in parallel.
In fact multi-threading was implemented in GPquick
to speed up fitness evaluation
\cite{langdon:2019:gpquick}.
With a population that is bigger than the number of cores
simply giving the next operation
(e.g.\ child to create
or fitness to evaluate)
to each thread as it finishes its previous task
(with appropriate synchronisation locks)
works well and usually shares the work equitably between
the available compute cores.
If the tasks are very unequal,
this simple way of handing out work may be problematic.
For example,
with incremental evaluation
and huge trees,
it was sometimes the case that one fitness evaluation
took far longer than the others,
causing the whole population to wait
for one fitness evaluation to complete.
I have not investigated better ways to share tasks between threads,
but an easy alternative might be simply to run
more independent jobs on your computer than it has cores.
With our Centos version of Linux,
the operating system overhead of rescheduling work on a core is low.
Therefore, if you have more than one EA task
and sufficient memory,
allowing the operating system to keep the cores
busy could be an efficient and easy to implement solution.
\subsection{Parallel Fitness Evaluations using
SIMD Vector Instruction CPUs}
\label{sec:avx}
Riccardo Poli pointed out in bit-wise problems
the basic hardware is capable of doing 32 or 64 bit operations
simultaneously
\cite{poli:1999:aigp3,poli:2000:22par}.
Indeed I used a 64~bit Intel computer to do all
64 fitness cases in parallel for
the first long term evolution experiment (LTEE)
\cite{Langdon:2017:GECCO}
running GP on a binary tree version of Koza's~\cite{koza:book}
6-multiplexor problem
for 100\,000 generations%
\footnote{
Rich Lenski's team \cite{Lenski:2015:PRoySocB}
have been running microbiology laboratory evolutionary experiments
on the bacteria E.~coli since 1988~\cite{lenski1988},
which have now reached
73\,500 generations~\cite{Kokko:2021:rspb}.
}.
Some current Intel hardware supports 512 bit operations.
I am unaware of any genetic programming
work using AVX-512 to process 512 Boolean test cases in parallel.
However it appears
Hrbacek and Sekanina~\cite{Hrbacek:2014:GECCO}'s
impressive work
with cartesian genetic programming
using 256 bit SSE instructions could be readily
extended to the newer 512 AVX.
In addition to bit-wise operations, the AVX vector instructions
can simultaneously perform 16 floating point calculations.
I wrote an AVX interpreter for GPquick
so that instead of processing each tree once per test case
it passed through each tree only once and
instead executed each opcode on all the test cases.
The first AVX version used the traditional depth first ordering
which requires each node within the tree to recursively
call the subtrees below it
and store intermediate results on a stack.
For efficiency an external stack
(of depth equal to the height of the tree)
is used.
GPquick requires the user to specify a maximum size of the trees
before it starts.
From the maximum size an estimate can be made of the maximum
height of any evolved tree
and used to set the size of the stack
when GPquick starts.
Vasicek and Slany \cite{vasicek:2012:EuroGP},
like Hrbacek and Sekanina~\cite{Hrbacek:2014:GECCO},
used 256 bit SSE instructions to allow
cartesian genetic programming's fitness function to process
up to eight floating point numbers simultaneously.
Again it would seem possible to extend this to the newer 512 AVX
now available on some Intel servers
as well as Intel specialist plug-in GPU like accelerators.
De~Melo et al.~\cite{DeMelo:2020:evoapplications}
also used AVX parallel hardware and
claim an impressive median peak performance of 7.5 billion
genetic programming operations per second in a single CPU core,
however this is for their interpreter alone
and does not include the rest of their GP system.
\subsection{Alternative Parallel Approaches}
\subsubsection{Hybrid Computing, Compute Accelerators: GPU and TPU}
\label{sec:gpu}
Many flavours of hybrid architecture are possible.
Here we mean a non-X86 or Intel Xeon boards,
which may be housed inside the computer
and certainly within a few feet or it.
Although experiments continue with FPGA's~\cite{tgordon},
running computation on graphics cards (GPUs) and domestic gaming boards,
usually manufactured by nVidia
\cite{langdon:2008:eurogp},\cite{langdon:2013:ecgpu},
and known as general purpose computing on GPUs,
GPGPU~\cite{owens_2007_cgf},
continues to be popular.
However more recently,
especially for training deep neural networks,
Google's TPU
hardware accelerators have become popular.
Although perhaps for Pandemic related reasons,
there is perceived to be a worldwide shortage of silicon chips
and the steep increase in the GPU price performance curve
has soften,
it is expected that the impressive raw performance of
GPU and other accelerators will continue to rise.
Indeed already GPUs with multiple gigabytes of on board memory and
thousand of fully capable computer cores
are in use.
Although traditionally it is tricky to code GPGPU to get the best of the hardware:
evolutionary computing,
indeed genetic programming
\cite{langdon_2012_sigevolution},
have been shown to be fully capable of exploiting GPUs.
\subsubsection{Data Center Networks: MPI}
It is clear that CPU clock speeds are stuck at around 3~GHz
and no great speed up can be expected in that direction.
However compute power has continued to increase
by allowing greater use of parallel hardware.
Sections~\ref{sec:pthreads}
and~\ref{sec:avx}
have discussed multi-core and
SIMD vector instructions.
Further increases in both can be expected.
However the dominant growth in computer power
has been in loosely coupled
clusters of compute servers.
Indeed, today it is not uncommon for
universities to run their own clusters of thousands of cores.
And academics have not been slow in exploiting
this architecture,
originally called Beowulf computing
\cite{bennett:1999:BPCSPHPD},
and nowadays referred to as cluster, grid~\cite{Wagner:2004:SS}
or cloud computing.
Indeed many use commercial off campus cloud computing,
e.g.\
Amazon AWS EC2,
Google Cloud and Microsoft Azure.
Although many protocols exist for sharing data
between peer networks,
MPI%
\footnote{%
It seems that MPI has replaced the earlier PVM,
e.g., as used in the distributed evolutionary algorithm
ParadisEO~\cite{Cahon:2004:JoH}.}
is popular in the Linux world
when interconnecting compute nodes within the same
data centre.
\subsubsection{Global Networks: HTTP \& P2P}
It has been known for a long time that
it is possible to distribute evolutionary computing
globally~\cite{chong:1999:jDGPi,Laredo:2014:GPEM}.
Often this is done by distributing both population and
fitness evaluation,
for example Laredo et al.'s
EvAg~\cite{Laredo:2009:GPEM}.
However perhaps the most famous example is
SETI@home~\cite{Anderson:2002:CACM},
which uses otherwise idle compute time donated by
volunteers world wide.
\begin{figure}
\includegraphics[scale=0.299287]{pic/crossover}
\caption{\label{fig:crossover}
Original GPquick
subtree crossover
requires three \mbox{memcpy} buffer copies:
1)~root segment of donating parent (mother, red/brown) is copied to offspring buffer.
2)~subtree from second parent (father, blue/black) is copied to offspring.
3)~tail (brown) of $1^{\rm st}$ parent copied to child.
}
\end{figure}
\begin{figure}
\flushleft
\includegraphics[scale=0.299287]{pic/memmove}
\caption{\label{fig:crossover2}
In place subtree crossover uses only one GPquick buffer.
The offspring is the last child of the $1^{\rm st}$ parent (mother)
and reuses its buffer.
Another optimisation is that
the only part of the $2^{\rm nd}$ parent (father) to be kept is the
subtree to be inserted (black).
It is small enough to be stored on the heap.
In 71\% of children the
subtree to be remove (white) and to be inserted (black)
are different sizes,
and so memmove is used to
shuffle the second part of the mother's buffer (brown) up or down.
Finally the father subtree overwrites the buffer.
The unused part of the buffer (right end) is not cleared
and so may contain junk from earlier generations.
}
\end{figure}
\section{Faster Crossover}
\label{sec:fast_xo}
GPquick uses a very compact representation
for its GP trees.
This only requires one byte per tree node
\cite[pp~297]{kinnear:keith}.
Its compactness allows both very large populations
(e.g.\ 5~million trees~\cite{langdon:2008:SC})
and very large trees,
(e.g.\ two billion nodes~\cite{Langdon:2022:ALJ}).
It was only when fitness evaluation was sped up dramatically
that the cost of crossover became noticeable.
When the evolved trees become very large,
reproduction, e.g.\ crossover and mutation,
becomes expensive in GPquick
due to the volume of data to be copied,
see Figure~\ref{fig:crossover}.
GPquick now uses several ways to speed up crossover:
\begin{itemize}
\item
Run in parallel thread (see Section~\ref{sec:pthreads} above)
\item
Run fitness evaluation before genetic manipulation.
Thus children of fit parents who have
low fitness, or are just unlucky,
and so have no children themselves
need never be created
\cite{Langdon:2021:GECCO,Langdon:2021:GPTP}.
\item ``Fatherless'' crossover
\cite{Langdon:2021:GECCO,Langdon:2021:GPTP}
(see also Section~\ref{sec:fatherless}).
\hspace{1ex}
To allow exact reproduction,
GPquick performs
stochastic activities, such as choosing crossover points,
in a single master thread
before per generation parallel activities.
Since the crossover fragments are known,
it becomes easy to copy them all before
any genetic operations are done.
This means:
\begin{itemize}
\item Parents (``Fathers'') who only donate subtrees can be deleted
immediately.
Indeed only parents which donate their root node to at least one child
need be saved and processed by the crossover queues
(Section~\ref{sec:small_mem_ga}).
This saves full length buffers
and potentially reduces actual memory requirement below $M$~buffers.
\item
The copied subtree fragments are small enough to keep on the heap
and so potentially allow better use of the CPU's L2 and L3 caches.
\item
The crossover/mutation queues (Section~\ref{sec:small_mem_ga})
also make it easy to spot
when a parent has only one child left to be born.
In such cases, the new child is not allocated
a new buffer but simply reuses its parent's buffer.
(This is known as in place crossover.
Contrast Figure~\ref{fig:crossover2}
with Figure~\ref{fig:crossover},
both from~\cite{Langdon:2021:GECCO}.)
This will usually require about half the existing buffer to be
shuffled
but this is usually faster than copying almost the whole buffer
to a new location
\cite[Sec.~5]{Langdon:2021:GPTP}.
\end{itemize}
\end{itemize}
\noindent
GNU glibc-2.19's memmove drops into X86 assembler
to move opcodes in the buffer.
It uses a combination of
{\tt rep~movsb} and
{\tt rep~movsl}
without the need to take a copy of the data.
The need to move the data up or down within the buffer arises when
the subtree removed by crossover is not the same size
as the subtree to be inserted
(Figure~\ref{fig:crossover2}).
Most crossover fragments are small
and so usually the buffer need be shuffled by two or perhaps four bytes.
With large trees,
memmove is on average 14\% faster than memcpy
\cite[Sec.~5]{Langdon:2021:GPTP}.
This may be because most of its
{\tt movsl} source and destinations are in the same L1 cache line
whereas with memcpy they reference different lines.
\section{Avoiding Fitness Evaluation Work}
\label{sec:avoidwork}
The best way to optimise anything is not to do the existing work better
but to reduce the amount of work to be done.
Section~\ref{sec:gpinc}
describes in fitness converged populations how to avoid evaluating
more than 99\% of opcodes
using a method based on incremental fitness evaluation
\cite{langdon:2021:EuroGP}.
\subsection{Population as a directed acyclic graph}
Simon Handley~\cite{Handley:1994:DAGpcp}
had a very elegant GP representation
in which the whole evolving population of genetic programming trees
was represented by a directed acyclic graph (DAG).
Each individual was represented by an entry point into the common code.
Common subtrees were held only once in the DAG\@.
When fit individuals were selected for crossover,
they exchanged existing subtrees.
The DAG kept track of this,
thus compactly representing the whole population.
See also Nic McPhee et al.'s Sutherland~\cite{mcphee:1998:sutherland}.
In the absence of side effects,
when the same fitness test cases are used in the next generation,
the evaluation of the each subtree is unchanged in the next generation
and the DAG can be used to cache subtree evaluations.
On many non-image processing tasks,
it is feasibly to cache every subtree evaluation.
Meaning after each genetic operation,
only the sub-evaluations from the crossover or mutation
site to the tree's root node need be re-evaluated.
That is,
fitness evaluation time
no longer scales as the size of the trees
but as their depth.
However...
Caching partial evaluations of the whole population
would defeat GPquick's compact opcode representation
(Section~\ref{sec:fast_xo})
and if trees are allowed to grow without restriction
it would not be feasible.
\subsection{Bottom Up Fitness Evaluation}
\label{sec:gpinc}
What to do?
As is common in evolutionary computing,
we used the same test cases every generation.
The huge evolved trees were converging and thus
burning huge amounts of CPU time
to calculate the same fitness values over and over again.
In the {\bf absence of side effects} the GP trees are pure functional expressions
and thus can be evaluated in any order.
I.e.\ they do not have to be evaluated
in the traditional root first recursive top down way.
Instead incremental fitness evaluation~\cite{langdon:2021:EuroGP},
evaluates each new child from the change point outwards.
Remember the only change between parent and child
is here.
All the rest of the code is identical
and, unless the code below it has changed,
it will evaluate identically both in the parent and its offspring.
With incremental evaluation,
GPquick
re-evaluates the ``mother''
(the parent who donates the root node to the child)
as it evaluates her child,
see Figure~\ref{fig:incfit}.
Firstly notice in almost all the tree
the two evaluations are identical,
so the worst case overhead is modest.
Secondly the only place they can be different,
is on the chain of function calls from
the changed code to root node.
If at any point along that chain,
the evaluation of the new code
is equal to that in its parent,
then the two evaluation will
remain identical from that point to the root node.
If they are identical on all test cases,
the child's fitness will be identical
to that of its ``mother''.
Thus if at any point on the route up the tree
from the crossover or mutation site
to the root node
both evaluations on all the test cases
are identical,
the child's fitness is known to be
that of its parent and we can immediately stop
fitness evaluation.
(Figure~\ref{fig:david2_colour},
page~\pageref{fig:david2_colour},
shows fitness disruption
caused by a crossover fading a way in a small tree.)
In large converged populations
the savings can be two or three orders of magnitude.
In fact so fast,
that the cost of crossover becomes dominant.
\begin{figure}
\centerline{\includegraphics[scale=0.149884]{pic/inceval_evosoft2022}}
\caption{\label{fig:incfit}
With fitness first,
fitness is evaluated using only the parents,
i.e., before the child is created by crossover.
Assuming no side effects,
the subtree to be inserted (black) is evaluated on all test cases
and values are transferred to evaluation of mother at the location of
the subtree to be removed (white).
Using incremental evaluation~\protect\cite{langdon:2021:EuroGP}
differences between the original
code (white subtree) and the new, are propagated up the $1^{\rm st}$
parent (mother) until either all differences are zero
or we reach the root node.
Based on~\protect\cite{Langdon:2021:GECCO}.
}
\end{figure}
\begin{figure*}
\centerline{\includegraphics[width=\textwidth]{david2_colour}}
\caption{\label{fig:david2_colour}
The impact of a crossover (red subtree) dissipates
the further away from the crossover site we measure it.
The change in fitness test values is shown with coloured nodes.
The size of the coloured nodes shows the number of disrupted test cases.
The colour shows, on a log scale, the average difference
in evaluation on the remaining disrupted nodes.
Brightest yellow shows smallest non-zero difference
\mbox{(RMS $3.1\ 10^{-10}$)}.
Eventually this crossover has no fitness impact at all.
White nodes are identical before and after the crossover.
Dark grey nodes show the part of the tree which does not have to be evaluated.
From~\protect\cite{langdon:2021:ieeeblog}.
}
\end{figure*}
\subsection{\mbox{Fitness Selection Before Crossover/Mutation}}
\label{sec:fitness_first}
\label{sec:fatherless}
Notice that bottom up evaluation
only requires the child's new subtree,
e.g.\
the subtree it will inherit from is ``father''.
The rest of the code is identical to
the parent it inherits its root node from
(its ``mother'').
Thus it is easy to implement
evaluating the child's fitness
before the child has been created
(mentioned in Section~\ref{sec:fast_xo} above).
Given everyone in the population's fitness,
it is possible to perform selection
and discover who will mate with whom
and who, even by just bad luck,
will not mate.
In every evolutionary algorithm,
if crossover or mutation is going to be expensive
and we know that the individual is not
going to pass its genetic material on,
we do not have to create it.
Early in an EA run,
i.e.\ before convergence,
the saving can be large.
For example, with tournament selection
and mutation only,
we can save creating on average
$1-1/({\rm tournament\ size})$
of the children.
E.g.\ in a generational EA
with a diverse population of fitness values,
no elitism, only mutation,
and with tournaments of size~7,
on average
86\% of the next population will not themselves
have children and so can be safely ignored.
Even with 100\% two parent crossover and
a fully converged population,
$1-e^{2}$ $=14\%$ of children will not themselves have children
and again need not be created.
\section{Deep Expressions Hide Errors}
\label{sec:FDP}
\subsection{Information Theory}
Information theory suggests what we have
described for floating point~\cite{langdon:GPEM:gpconv,Langdon:2022:ALJ}
and now integer genetic programming~\cite{langdon:2022:GECCOcomp}
will hold where there aren't side effects to carry
information long distances through program code
or GP trees
(see Figure~\ref{fig:david2_colour},
page~\pageref{fig:david2_colour}).
Essentially the argument is:
for a mutation, change, bug, transient error, etc.,
to have an effect it must give rise to a change
of state
and that change of state must propagate from the error
to the program's output~%
\cite{Petke:2021:FSE-IVR}.
However,
without side effects,
the change must pass through a series of nested digital computations
between it and the program's output.
Typically each computation has fewer output states
than it has inputs states.
(In deterministic programs it can have the same
number of output states as it has input states
but it cannot have more.)
As we saw in Section~\ref{sec:gpinc}
with bottom up evaluation of trees,
if at any point,
the disrupted execution creates the same state
as the original correct execution,
the twin executions remain
locked together,
so that when execution reaches
any output,
the correct and erroneous calculation,
generate identical output.
The error has disappeared%
\footnote{%
Notice failed disruption propagation (FDP) holds in all digital systems,
not just genetic programming.
Any disruption in any deterministic system will tend not
to propagate as it passes through
intermediate operations.
}.
Indeed so too has any fitness signal
from the mutation.
Without fitness to guide it,
evolution becomes an undirected random walk.
\subsection{Better Testing May Only Help a Little}
For GECCO 2022~\cite{langdon:2022:GECCO2}
we showed if fitness tests are independent of each other
and equally effective,
then the effectiveness of the whole test set
increases only slowly,
O$(\log n)$,
with increasing number $n$ of fitness tests.
If the tests not equal,
the test set is dominated by the most effective tests
and the weaker ones contribute little.
Of course if we have insight into the code being
tested we may be able to make much better intelligent
choices of test cases values.
\subsection{Will Side Effects Help?}
In a recent issue of the
ACM Special Interest Group on Genetic and Evolutionary Computation's
newsletter~\cite{langdon:2021:sigevolution},
I argue that to evolve large complex artifacts of any sort,
we must give up on any hope of evolving large deep structures
but instead limit their effective depth
by evolving them close to their environment.
Perhaps adapting an open structure
(see Figure~\ref{fig:lung_like},
page~\pageref{fig:lung_like}),
where a flood tide can bring in new inputs
and and out going ebb tide can move the output of shallow
structures both to the environment and to other parts
of our environmentally embedded evolving mycelium
like program or artifact.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{random_graph_merge.png}}
\caption{
Proposed
lung like open complex evolving system
composed of 1300 separate shallow GP programs.
These compute element are placed
side-by-side to form an open structure.
The gaps promote short cut side effects between functions'
input and outputs and the environment.
From~\cite{langdon:2021:sigevolution}.
\label{fig:lung_like}
}
\end{figure}
\section{Conclusions}
We have shown a number of speed up techniques
which could be implemented in other evolutionary algorithms.
The C++ code used by GPquick is available
or you may prefer to make your own implementation
of these ideas.
As a few of the citations have shown,
the ideas of distributed parallel running
are widely known in evolutionary algorithms.
However GPU and TPU implementations
are at present thinner on the ground.
Replicating SETI@home's success
with evolutionary search might also be worth another look.
Similarly if you are interested in using EAs
on problems with a population of large chromosomes,
some of the ideas for reducing
memory consumption or
speeding up crossover/mutation
might be useful.
Alternatively if you are hoping to run evolution
on tiny ``mote'', embeded, internet of things (IOT)
or ``edge computing'' devices,
these ideas for shoehorning genetic programming~(GP)
into the available memory
may also be interesting.
The vast bulk of GP uses static fitness
test and primitives without side effects
and thus the GPquick implementation
I have sketched could be of interest.
Perhaps the most important result,
due to its universality,
was unlooked for:
as structures get deeper it becomes
harder to measure the impact of changes.
Although good choices of test cases can help,
simply increasing the number $n$ of test cases
can perhaps only make a slow O($\log(n)$) improvement.
The impossibility of deeply probing
applies to programming in general
not just to evolutionary algorithms.
Notice the prevalence of unit testing
shows this is somewhat tacitly accepted in Software Engineering,
because unit testing places
the test oracle, the check for correct output,
close to the error.
In the case of SBSE, EAs, Artificial Life, etc.,
concealing deep errors, bugs, mutations, etc.,
means it is hard to see if search is proceeding.
Without a fitness delta,
we cannot hope to evolve anything deep.
\subsection*{Acknowledgements}
I would like to thank
the
\href{https://dev.heuristiclab.com/trac.fcgi/wiki/EvoSoft}
{EvoSoft 2022} reviewers,
Dagstuhl Seminars
\href{https://www.dagstuhl.de/en/program/calendar/semhp/?semnr=17191}
{17191} on the theory of randomized heuristics and
\href{https://www.dagstuhl.de/en/program/calendar/semhp/?semnr=18052}
{18052} on
Genetic Improvement of Software~\cite{gi_dagstuhl_2018},
for inspiring conversations
and
the Meta OOPS project.
C++ code is available in
\url{http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/gp-code/GPinc.tar.gz}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
Persistent homology \cite{EdelsZomo, EdelsHarer} is a parameterized version of homology, with the parameter arising from a filtration of a simplicial complex. One of its fundamental features as compared to homology is stability \cite{EdelsMoro}. In particular, small perturbations of filtration functions induce small perturbations to the barcode and to the lifespans of homology elements. However, critical simplices and homology representatives of bars in a barcode of persistent homology cannot change continuously with arbitrary perturbations of filtration functions.
As simplices keep appearing in a filtration, non-trivial homology classes are appearing and terminating. A simplex terminating a homology class (i.e., identifying the class with the trivial class) is called a \textbf{terminal} critical simplex.
A homology representative of a bounded bar (equivalently, a finite point in a persistence diagram) is obtained from the corresponding terminal simplex. As such, the terminal simplices allow us to localize and determine a geometric manifestation of bars. While finding a homology representative of a bar is by itself challenging \cite{Chao}, the fact that small perturbations of a filtration function may yield completely different terminal simplices results in unstable representatives. There has been an attempt to circumvent this issue in practice \cite{Bub}.
In this paper we study the region of parameter $\varepsilon$, for which the terminal simplex of a persistent homology class $[\alpha]$ of an injective filtration function $f$ is \textbf{rigid} (i.e., constant) through $\varepsilon$-perturbations of $f$. Let $[a,b)$ be the lifespan interval of an $n$-dimensional homology class $[\alpha]. $ Our \textbf{main results} are the following (under suitable assumptions).
\begin{enumerate}
\item Theorem \ref{ThmConseq1}: The terminal simplex of $[\alpha]$ is rigid for $\varepsilon$-perturbations of $f$ if
\begin{itemize}
\item no class of $H_{n+1}$ is born on $(b,b+2\varepsilon]$ and
\item no class of $H_n$ terminates on $[b-2\varepsilon,b)$.
\end{itemize}
\item Theorem \ref{ThmBarcodeFinal}: A version of Theorem \ref{ThmConseq1} for significant bars in the barcode. One of the main \textbf{advantages} of this result is that rigidity can be deduced solely from the barcode, without looking at the underlying filtration.
\end{enumerate}
Theorem \ref{ThmConseq1} describes the two ways in which non-rigidity of a terminal simplex may occur: either via sequentially critical pair of simplices or independently critical pair of simplices. As a result, we are not only able to provide bounds on the region of rigidity, but also to locate a simplex ($\Delta_2$ in Theorem \ref{ThmConseq1}) appearing as a new terminal simplex for $[\alpha]$ in the region of non-rigidity. For a demonstration see Figure \ref{FigMainIntro1g} and Figure \ref{FigMainIntro2g}.
\input{figmainintro1.tex}
\input{figmainintro2.tex}
On the other hand, our results contribute a new interpretation of the structure of persistence diagrams. So far, persistence diagrams have been known to encode topological information of the space at small scales \cite{Haus, Lat, ZV3}, intrinsic combinatorial structure of filtrations (such as Rips complexes) \cite{AA, Ad5}, proximity of spaces via the stability result \cite{ZVCounterex}, geometric properties of spaces \cite{ZV2}, shortest homology basis \cite{ZV}, spaces of contraction \cite{ZVCont}, filling radius \cite{Memoli}, curvature \cite{Bubenik2020}, width of homology class \cite{ACos}, and more. Our results imply that, to a degree, persistence diagrams encode rigidity of terminal simplices.
\section{Preliminaries}
We first introduce the setup of persistent homology, see \cite{EdelsHarer} for details. Throughout this paper we assume that $K$ is a finite simplicial complex and $f\colon K \to \mathbb{R}$ is a filtration function on $K$ (if $\sigma$ is a face of $\tau$ then $f(\sigma)\leq f(\tau)$ for all $\sigma, \tau\in K$). We also assume that $f$ is injective. As such $f$ encodes an order on the simplices of $K$. For example, in the original persistent homology algorithm \cite{EdelsZomo} such an order is used to arrange simplices in the boundary matrix.
Given two injective filtration functions $f$ and $g$, we define the distance between them as
$$||f-g||_\infty = \max_{\sigma\in K}|f(\sigma)-g(\sigma)|.$$
Function $g$ corresponds to some permutation of the ordering of simplices encoded by $f$.
The sublevel sets of $f$ are subcomplexes of $K$ defined for all $r\in\mathbb{R}$ as the pre-images $K_r^f = f^{-1}((-\infty, r])$. We can also define $K^f_\infty = f^{-1}((-\infty, \infty)) = K$. The notation $K^f$ denotes the standard sublevel filtration of $K$ obtained through $f$, i.e., the collection of subcomplexes $\{K^f_r\}_{r\in \mathbb{R}}$ along with the natural inclusions
$$
\iota_{q,r} \colon K^f_q \to K^f_r
$$
for all $q\leq r$.
All homology groups are assumed to be with coefficients in a fixed field $\mathbb{F}$ and therefore not mentioned in the notation for homology.
Given a non-trivial homology element $[\alpha]\in H_n(K^f_r)$ for some $r$, we define:
\begin{description}
\item [birth] $a\in \mathbb{R}$ of $[\alpha]$ as the infimum of levels $q \leq r$, for which there exists $[\alpha_q]\in H_n(K^f_q)$ such that $\iota^*_{q,r} [\alpha_q]=[\alpha]$. We say that $[\alpha]$ is born at $a$.
\item [termination scale] $b\in \mathbb{R} \cup \{\infty\}$ of $[\alpha]$ as the infimum of levels $q \geq r$, for which $\iota^*_{r,q} [\alpha_q]=0$, or $\infty$ if such levels do not exist. We say that $[\alpha]$ is terminated at $b$.
\end{description}
Note that the termination scale is defined differently than death in \cite{EdelsHarer}.
Given our setup of sublevel complexes defined through preimages of closed intervals, the infima in the definition of birth and termination scale are always attained if finite.
Since $f$ is injective, at most one simplex is added at each level $r$. Consequently each simplex either gives birth to a non-trivial homology class or terminates a non-trivial homology class.
We next state the stability theorem and introduce the corresponding notation, see \cite{Bauer} for details.
Applying homology $H_n$ as a functor to a filtration in our setup we obtain a persistence module, i.e., a collection of vector spaces $\{H_n(K^f_r)\}_{r\in \mathbb{R}}$ along with the inclusions induced linear maps
$$\iota^*_{q,r} \colon H_n(K^f_q) \to H_n(K^f_r)$$
for all $q\leq r$. Each persistence module obtained in this setting decomposes as a finite direct sum of interval modules $\mathbb{F}_{[a_i,b_i)}$, where persistence module $\mathbb{F}_{[a,b)}$ for $a<b$ is a collection of vector spaces $\{V_r\}_{r\in \mathbb{R}}$, with:
\begin{itemize}
\item $V_r=\mathbb{F}$ for $r\in [a,b)$,
\item $V_r=0$ for $r\notin [a,b)$ and
\item the bonding linear maps $V_r \to V_{r'}$ being identities for parameters $r<r'$ from $[a,b)$.
\end{itemize}
Note that the type of endpoints of intervals (closed on the left, open on the right) is a consequence of our setup of a filtration. The collection of intervals $[a_i,b_i)$ is called a barcode.
Given injective filtration functions $f$ and $g$ and $n\in \{0,1,\ldots\}$, assume
$$
\{H_n(K^f_r)\}_{r\in \mathbb{R}} = \bigoplus_{i\in I} \mathbb{F}_{[a_i,b_i)}
$$
and
$$
\{H_n(K^g_r)\}_{r\in \mathbb{R}} = \bigoplus_{j\in J} \mathbb{F}_{[a'_j,b'_j)}
$$
are decompositions into interval modules.
If $||f-g||_\infty \leq \varepsilon$, the stability theorem states that the bottleneck distance between the barcodes is at most $\varepsilon$, i.e., there exists
a bijection $\varphi \colon I' \to J'$ for subsets $I' \subseteq I$ and $J' \subseteq J$ such that:
\begin{itemize}
\item $|a_i - a'_{\varphi(i)}| \leq \varepsilon$ and $|b_i - b'_{\varphi(i)}| \leq \varepsilon$ for all $i\in I'$,
\item $|a_i-b_i | \leq 2\varepsilon$ for all $i\in I\setminus I'$ and
\item $|a'_j-b'_j | \leq 2\varepsilon$ for all $j\in J\setminus J'$.
\end{itemize}
\subsection{Filtration manipulation}
The following propositions explain local adjustments to filtration functions that result in a predetermined permutation of a collection of simplices.
\begin{proposition}
\label{switch}
Let $\sigma_1$ and $\sigma_2$ be two $n$-dimensional simplices in a finite simplicial complex $K$ and let $f$ be an injective filtration function on $K$. Assume that for some $\varepsilon > 0$ we have
$$f(\sigma_1) < f(\sigma_2)< f(\sigma_1)+2\varepsilon.$$
Then there exists an injective filtration function $g$ on $K$, such that $||f-g||_\infty \leq \varepsilon$ and $g(\sigma_2) < g(\sigma_1)$.
\end{proposition}
\begin{proof}
Without loss of generality we can slightly decrease $\varepsilon$ so that the assumptions of the proposition still hold and that
\begin{equation}
\label{injectiveg}
\varepsilon \notin \{|f(\sigma)- f(\tau)| \;;\; \sigma, \tau \in K\} \textrm{ and }2\varepsilon \notin \{|f(\sigma)- f(\tau)| \;;\; \sigma, \tau \in K\}.
\end{equation}
Let $U=\mathord{\mathrm{Upper}}(\sigma_1)\subseteq K$ denote the subset of all simplices in $K$ that contain $\sigma_1$ as a face (the upper set of $\sigma_1$ in the Hasse diagram of $K$), and let $L=\mathord{\mathrm{Lower}}(\sigma_2)\subseteq K$ denote the subset of all faces of $\sigma_2$ (the lower set of $\sigma_2$ in the Hasse diagram of $K$). Note that $\sigma_1\in U$ and $\sigma_2\in L$.
If $\mu\in U$, set $g(\mu) = f(\mu)+\varepsilon$. If $\mu\in L$, set $g(\mu) = f(\mu)-\varepsilon$. For all other simplices $\mu$ let $g(\mu) = f(\mu)$.
Since $\sigma_1$ and $\sigma_2$ are two distinct simplices of the same dimension, we have $U\cap L = \emptyset$, so $g$ is well-defined. It is also obvious that $|g(\mu)-f(\mu)|\leq\varepsilon$ for all $\mu\in K$, so $||g-f||_\infty\leq\varepsilon$. Finally, if the value at a simplex $\mu$ has decreased (or increased) by $\varepsilon$, then the same was true for all the faces (or cofaces) of $\mu$, so $g$ is a filtration function.
The function $g$ defined this way is injective by the condition (\ref{injectiveg}).
\end{proof}
Given a collection of $k$ simplices of dimension $n$ and an arbitrary permutation $\pi\in S_k$, we do not have to switch the order one pair at a time but can instead mix them up all at once. Corollary \ref{switchmany} shows one way of doing this.
\begin{corollary}
\label{switchmany}
Let $\sigma_1,\ldots,\sigma_k$ be a selection of $n$-dimensional simplices in a finite simplicial complex $K$, $\pi\in S_k$ an arbitrary permutation of the indices $\{1,2,\ldots,k\}$ and $f$ an injective filtration function on $K$. Assume that for some $\varepsilon > 0$ we have
$$f(\sigma_1) < f(\sigma_2)< \ldots < f(\sigma_k) < f(\sigma_1)+2\varepsilon.$$
Then there exists an injective filtration function $g$ on $K$, such that $||f-g||_\infty \leq \varepsilon$ and
$$g(\sigma_{\pi(1)}) < g(\sigma_{\pi(2)})< \ldots < g(\sigma_{\pi(k)}).$$
\end{corollary}
\begin{proof}
If we write $a = f(\sigma_k)-\varepsilon$ and $b = f(\sigma_1)+\varepsilon$, then $(a,b\,]$ is an interval of length $\delta < 2\varepsilon$. Divide this interval into $k+1$ equal pieces of length $\frac{\delta}{k+1}$ and define
\begin{eqnarray*}
g(\sigma_{\pi(1)}) &=& a +\frac{\delta}{k+1},\\
g(\sigma_{\pi(2)}) &=& a +\frac{2\delta}{k+1},\\
&\vdots & \\
g(\sigma_{\pi(k-1)}) &=& a +\frac{(k-1)\delta}{k+1},\\
g(\sigma_{\pi(k)}) &=& a +\frac{k\delta}{k+1}.
\end{eqnarray*}
It is not difficult to show that
$|g(\sigma_{i})-f(\sigma_{i})| \leq \varepsilon - \frac{\delta}{k+1}$.
Define
\begin{eqnarray*}
I^+ &=& \{i\in\{1,2,\ldots,k\}\;;\; g(\sigma_i)-f(\sigma_i) > 0\},\\
I^- &=& \{i\in\{1,2,\ldots,k\}\;;\; g(\sigma_i)-f(\sigma_i) < 0\},\\
\varepsilon^+ &=& \max\{g(\sigma_i)-f(\sigma_i)\;;\; i\in I^+\},\\
\varepsilon^- &=& \max\{|g(\sigma_i)-f(\sigma_i)|\;;\; i\in I^-\},\\
\end{eqnarray*}
\vskip-1.2cm
$$U = \bigcup_{i\in I^+}\mathord{\mathrm{Upper}}(\sigma_i)\quad\textrm{and}\quad L = \bigcup_{i\in I^-}\mathord{\mathrm{Lower}}(\sigma_i).$$
It is easy to see that $\varepsilon^+ <\varepsilon$ and $\varepsilon^- <\varepsilon$.
For all $\mu\in U\setminus\{\sigma_1,\ldots,\sigma_k\}$ let $g(\mu) = f(\mu)+\varepsilon^+$. For all $\mu\in L\setminus\{\sigma_1,\ldots,\sigma_k\}$ let $g(\mu) = f(\mu)-\varepsilon^-$. Finally, for all other $\mu\in K$ set $g(\mu) = f(\mu)$. Then $g$ is a filtration function with all the desired properties except (perhaps) injectivity, but it can be made injective either with a small perturbation of values (upholding the property of being a filtration function), or with a minor decrease in $\varepsilon$ at the beginning (as was done in the proof of Proposition \ref{switch}).
\end{proof}
\section{Rigidity for homology classes}
As before, let $f$ be an injective filtration function on a simplicial complex $K$ and let $n$ be a positive integer. Assume that an $n$-cycle $\alpha$ is created when its last $n$-dimensional simplex is added at level $a$, and that the homology class $[\a]\in H_n(K_a^f)$ is born at $a$ and terminated at $b$.
Choose $\varepsilon<\frac{b-a}{2}$. Then for each $\varepsilon$-perturbation $g$ of $f$ the class $[\a]$ exists and is non-trivial in $H_n(K_r^g)$ at least for $r\in [a+\varepsilon, b-\varepsilon)$.
For every injective filtration function $g$ at distance at most $\varepsilon<\frac{b-a}{2}$ from $f$ let $\Delta_{g,\a}$ denote the $(n+1)$-simplex in $K^g$ that terminates $[\a]$. Define
$$\Sigma_{\varepsilon} = \{\Delta_{g,\a}\;;\; \textrm{$g$ an injective $\varepsilon$-perturbation of $f$}\}.$$
Note that $g(\Delta_{g,\a})\in [b-\varepsilon, b+\varepsilon]$ and $f(\Delta_{g,\a})\in [b-2\varepsilon,b+2\varepsilon]$. We say that $[\alpha]$ is {\em $\varepsilon$-terminally-rigid} if $|\Sigma_{\varepsilon}|=1$.
Since $f$ is injective and $K$ is finite, we can define
$$\rho = \min\{|f(\sigma)-f(\tau)|\;;\;\sigma \neq \tau\in K\}>0,$$
the {\em injectivity radius} of $f$.
The aim of this paper is to study algebraic effects of terminal non-rigidity. Homology class $[\alpha]$ as defined above is $\varepsilon$-terminally-rigid for small $\varepsilon$, certainly for $\varepsilon<\frac{\rho}{2}$ since $\frac{\rho}{2}$-perturbations of $f$ retain the order of simplices appearing in the filtration of $K$. For larger $\varepsilon$ the class $[\alpha]$ is typically not terminally-rigid. We intend to focus on the region of $\varepsilon$ in which the initial form of non-rigidity occurs.
Let $d = \dim(K)$ and let $\sigma_j^{(i)}$ for $j=1,\ldots,k_i$ be the simplices of $K$ of dimension $i$. Then $f$ determines a linear ordering on the set
$$\{\sigma_1^{(0)},\ldots,\sigma_{k_0}^{(0)},\ldots,\sigma_1^{(d)},\ldots,\sigma_{k_{d}}^{(d)}\}$$
of all simplices of $K$. We can encode this ordering as a permutation $\overline{\pi}^f\in S_{k_0+k_1+\ldots+k_d}$ of the pre-determined ordering of simplices given above. A different injective filtration function $h$ defines a potentially different ordering of the simplices of $K$, corresponding to a potentially different permutation $\overline{\pi}^h$. Note that not all permutations in the symmetric group $S_{k_0+k_1+\ldots+k_d}$ correspond to filtration functions.
If we limit ourselves to only the simplices of a single dimension, however, all possible permutations of those simplices can be realized by Corollary \ref{switchmany}. Our filtration function $f$ determines a permutation $\pi_{n+1}^f\in S_{k_{n+1}}$, defined by the induced linear ordering of the set
$$\{\sigma_1^{(n+1)},\ldots,\sigma_{k_{n+1}}^{(n+1)}\}$$
of $(n+1)$-dimensional simplices of $K$. We will assume that the simplices of $K$ have been ordered in such a way that $\pi_{n+1}^f$ is the identity permutation. To unburden the notation we will write $\pi_{n+1}^f = \pi^f$ and $m=k_{n+1}$ from now on.
Any injective filtration functions that correspond to the same permutation generate the same boundary matrices in the classical matrix reduction algorithm for persistent homology, although the labels (function values) of simplices generally differ.
Also recall that for each $\varepsilon\in D=(0,\frac{b-a}{2}]$ and each injective filtration function $g$ with $||f-g||_\infty \leq \varepsilon$ the class $[\alpha]$ has a non-trivial lifespan in $K^g$.
Now, let us consider
\begin{itemize}
\item the function $\varepsilon \mapsto \Sigma_\varepsilon$ defined on $D$ that returns the collection of all critical simplices of $\varepsilon$-perturbations of $f$ that terminate $[\alpha]$ and
\item the function $\varepsilon \mapsto \Pi_\varepsilon$ defined on $D$ that returns the collection of all permutations corresponding to $\varepsilon$-perturbations of $f$.
\end{itemize}
The functions $\varepsilon\mapsto|\Sigma_\varepsilon|$ and $\varepsilon\mapsto|\Pi_\varepsilon|$ defined on $D$
\begin{enumerate}
\item are increasing,
\item have values in the discrete sets $\{1,2,\ldots, m\}$ and $\{1,2,\ldots, m!\}$, respectively, and
\item attain the value of $1$ for small $\varepsilon$ (at least for all $\varepsilon<\frac{\rho}{2}$).
\end{enumerate}
As a result, both functions are increasing step functions on $D$ as each of them partitions $D$ into finitely many intervals, such that the function is constant on each interval of that partition. It is also apparent that a change in $\Sigma_\varepsilon$ may occur at some $\varepsilon$ only if $\Pi_\varepsilon$ also changes at the same parameter. The next lemma shows that $|\Sigma_\varepsilon|$ and $|\Pi_\varepsilon|$ are lower semi-continuous (the intervals of the two partitions of $D$ are open on the left and closed on the right).
\begin{lemma}
\label{LemmaUSC}
For each $t\in D$ there exists $\delta>0$ such that $\Sigma_\varepsilon$ and $\Pi_\varepsilon$ are constant on $(t-\delta,t]$.
\end{lemma}
\begin{proof}
Choose finitely many injective filtration functions generating $\Pi_t$. Each of them can be brought a bit closer to $f$ because of injectivity.
\end{proof}
We next discuss potential points of discontinuity of $|\Sigma_\varepsilon|$ and $|\Pi_\varepsilon|$. The following lemma shows that they always correspond to exactly half the distance between two (not necessarily consecutive) values of $f$.
\begin{lemma}
\label{LemmaDiscontinuity}
Suppose $t\in D$ is a point of discontinuity of $\Pi_\varepsilon$. Then there exist $i\neq j$ such that $2t=|f(\sigma_i)-f(\sigma_j)|$.
\end{lemma}
\begin{proof}
As a consequence of Lemma \ref{LemmaUSC} there exists a $\delta>0$ such that $\Pi_\varepsilon$ is constant on $(t,t+\delta]$. Since $t$ is the point of discontinuity, there exists a permutation $\nu \in \Pi_{t+\delta}\setminus \Pi_t$. For each $N\in \mathbb{N}$ choose an injective filtration function $g_N$ inducing $\nu$ and satisfying $||f-g_N||_\infty \leq t+ 1/N$. Without loss of generality we may assume that $\{g_N(\sigma_i)\}_{N\in \mathbb{N}}$ converges for each $i$ and define $f_\infty$ as the limiting function (if any of them do not converge we can choose a converging subsequence). Note that while $f_\infty$ is a filtration function, it cannot be injective, because it would have corresponded to the premutation $\nu$ which is not in $\Pi_t$ while $||f-f_\infty||_\infty \leq t$ by definition.
Let us try fixing the values of $f_\infty$ to make it injective and see where exactly that fails.
Assume that $q$ is a value attained by $f_\infty$ at more than one simplex. Let $A_q$ denote the collection of such simplices. If there exist two distinct $\sigma_i, \sigma_j$ such that $q = f(\sigma_i)-t = f(\sigma_j)+t$, then we have found two simplices that satisfy the desired condition. If not, there are two possible reasons for that.
\begin{itemize}
\item If there is no $i$ such that $q=f(\sigma_i)-t$, then the $f$-values of all simplices in $A_q$ lie on $[q-t,q+t)$. In this case the $f_\infty$ values of all the simplices in $A_q$ can be slightly decreased by Corollary \ref{switchmany} so that:
\begin{itemize}
\item $f_\infty(\sigma)\neq f_\infty(\sigma')$ for all $\sigma\neq\sigma'$ in $A_q$,
\item the relative position of each simplex of $A_q$ to the simplices outside of $A_q$ is the same as in $\nu$,
\item simplices of $A_q$ appear in the same order as in $\nu$ and
\item the resulting filtration function $f'_\infty$ satisfies $||f-f'_\infty||_\infty \leq t$.
\end{itemize}
Since this would make $f_\infty$ injective, it cannot happen for all $q$ and there must be a simplex $\sigma_i$ such that $q = f(\sigma_i)-t$.
\item The case where there is no $j$ such that $q = f(\sigma_j)+t$ and the $f$-values of simplices from $A_q$ all lie on $(q-t,q+t]$ can be handled similarly by a small local increase of the values of $f_\infty$.
\end{itemize}
We can conclude that there exist $i\neq j$ such that $q = f(\sigma_i)-t = f(\sigma_j)+t$.
\end{proof}
\begin{definition}
An injective filtration function $f$ is called \textbf{generic}, if
$$|f(\sigma_i)-f(\sigma_j)|=|f(\sigma_{i'})-f(\sigma_{j'})|$$
implies
$\{i,j\}=\{i',j'\}$ whenever $i\neq j$ and $i' \neq j'$.
\end{definition}
\begin{proposition}
\label{PropStep2}
Let $f$ be a generic injective filtration function. Let $t_0=\max\{x\in D; |\Sigma_x|=1\}$ and $\Sigma_{t_0}=\{\Delta_1\}$. By Lemma \ref{LemmaDiscontinuity} there exist indices $i,j$ satisfying $q=f(\sigma_{j})-t_0=f(\sigma_{i})+t_0$.
Choose $t>t_0$ such that $|\Sigma_{t}|$ is as small as possible, i.e., $|\Sigma_{t}|=\lim_{x \searrow t_0} |\Sigma_x|$.
Then:
\begin{enumerate}
\item[(1)] $\Delta_1\in \{\sigma_{i}, \sigma_{j}\}$.
\item[(2)] $\Delta_1$ is a maximal simplex in $K^f_{f(\Delta_1)+2s}$ for all $s<t_0$.
\item[(3)] $\Delta_1=\sigma_{j}$ implies $\{\sigma_{i}, \sigma_{j}\} \subseteq \Sigma_{t}$ and $\sigma_{i}$ terminates a class in $H_*(K^f)$. In this case we say simplices $\sigma_i$ and $\sigma_j$ are \textbf{sequentially critical}, see Figure \ref{example5figure2}.
\item[(4)] $\Delta_1=\sigma_{i}$ implies $\{\sigma_{i}, \sigma_{j}\} \subseteq \Sigma_{t}$ and $\sigma_{j}$ creates a class in $H_*(K^f)$. In this case we say simplices $\sigma_i$ and $\sigma_j$ are \textbf{independently critical}, see Figure \ref{example5figure1}.
\end{enumerate}
\end{proposition}
\begin{proof}
Without loss of generality we may choose $t$ so that no element of the form $\frac{1}{2}|f(\sigma_i)-f(\sigma_j)|$ lies on $(t_0,t]$.
We will further develop the limiting argument presented in Lemma \ref{LemmaDiscontinuity}. Choose $\nu \in \Pi_{t}\setminus \Pi_{t_0}$. For each $N\in \mathbb{N}$ choose a filtration function $g_N$ inducing $\nu$ and satisfying $||f-g_N||_\infty \leq t_0+1/N$. Without loss of generality we may assume $\{g_N(\sigma_i)\}_{N\in \mathbb{N}}$ converges for each $i$ and define $f_\infty$ as the limiting function. Using local modifications as in the proof of Lemma \ref{LemmaDiscontinuity} we may assume that $f_\infty^{-1}(x)$ contains at most one simplex for each $x \neq q$. Define $A_q=f_\infty^{-1}(q)$. Choose an open interval $H$ around $q$ such that
$$H \subseteq \left(\bigcap_{\sigma\in \{\sigma_i, \sigma_j\}} (f(\sigma)-t,f(\sigma)+t)\right)\ \cap \ \left(\bigcap_{\sigma\in A_q\setminus\{\sigma_{i}, \sigma_{j}\}} (f(\sigma)-t_0,f(\sigma)+t_0)\right)$$
and $H \cap \mathrm{im}(f_\infty)=\{q\}$.
Redefining $f_\infty$ on $A_q$ by any injective assignment of values in $H$ respecting dimension (i.e., faces of a simplex are assigned smaller values than the simplex) we obtain an injective filtration function at a distance at most $t$ from $f$. As a result, any dimension-respecting permutation of elements in $A_q$, nested between other simplices as determined by $f_\infty$, determines a permutation in $\Pi_{t}$. We will call such a permutation a {\em local $A_q$-perturbation} of $\nu$.
On the other hand, redefining $f_\infty$ on $A_q$ by any injective assignment of values in $H$ respecting dimension such that $f_\infty(\sigma_{i}) < q < f_\infty(\sigma_{j})$, we obtain an injective filtration function at distance at most $t_0$ from $f$. As a result, any dimension respecting permutation of elements in $A_q$ nested between other simplices as determined by $f_\infty$, in which $\sigma_{i}$ appears before $\sigma_{j}$, determines a permutation in $\Pi_{t_0}$. We will call such a permutation a {\em restricted local $A_q$-perturbation} of $\nu$. Roughly speaking, the difference between $\Pi_{t_0}$ and $\Pi_{t}$ is that the later may swap $\sigma_{i}$ and $\sigma_{j}$. Observe also that $f(A_q) \subset [f(\sigma_i), f(\sigma_j)]$.
Recall that $\Sigma_{t_0}=\{\Delta_1\}$ and fix a permutation $\nu\in\Pi_{t} \setminus \Pi_{t_0}$ (and the corresponding adjusted limit $f_\infty$ of filtration functions) such that the corresponding critical simplex terminating $\alpha$ is $\Delta_2\neq \Delta_1$. Observe that $\nu$ swaps $\sigma_{i}, \sigma_{j}$, i.e., $\sigma_{i}$ appears after $\sigma_{j}$ in this permutation.
\begin{enumerate}
\item[(1)] We will now show that $\Delta_1\in \{\sigma_{i}, \sigma_{j}\}.$ Assume that $\Delta_1\notin \{\sigma_{i}, \sigma_{j}\}.$
\begin{enumerate}
\item Then $\Delta_1$ may appear as the first of the simplices of $A_q$ in some restricted local $A_q$ perturbation of $\nu$. In particular, $\Delta_1$ and the simplices appearing before $A_q$ contain a nullhomology of $[\alpha]$.
\item On the other hand, $\Delta_1$ may appear as the last of the simplices of $A_q$ in some restricted local $A_q$ perturbation of $\nu$. In particular, $A_q\setminus\{\Delta_1\}$ and the simplices appearing before $A_q$ do not contain a nullhomology of $[\alpha]$.
\end{enumerate}
As a result, $\Sigma_{t}$ can only contain $\Delta_1$ as in permutation $\nu$, class $[\alpha]$ becomes trivial by the time $\Delta_1$ is added (by (a)) but not before (by (b)). This is a contradiction. This proves our claim.
\item[(2)] If $\Delta_1$ was a facet (i.e., a codimension $1$ face) of a simplex $\widehat \Delta$ in $K^f_{f(\Delta_1)+2s}$, then by Corollary \ref{switchmany} we could construct an injective filtration function $g$ of $K$ with $||f-g||_\infty < t_0$ in which the boundary of $\widehat \Delta$ would appear before $\Delta_1$, while the simplices in $K^f$ appearing before $\Delta_1$ would also appear before $\Delta_1$ in $K^g$. This would mean that $\Delta_1$ would not be the simplex terminating $[\alpha]$ in $K^g$ as it could be replaced by the same-dimensional simplices of $\partial \widehat \Delta$. This contradicts the fact that $\Sigma_{t_0}=\{\Delta_1\}$.
\item[(3)]
Let $\Delta_1=\sigma_{j}$.
\begin{enumerate}
\item There is a restricted local $A_q$ perturbation of $\nu$ in which $\sigma_{j}$ appears as the last simplex. As $\Sigma_{t_0}=\{\sigma_{j}\}$, the simplices of $A_q\setminus\{\sigma_{j}\}$ and all the simplices appearing before them do not contain a nulhomology of $\alpha$, i.e., each nulhomology induced by a local $A_q$-perturbation of $\nu$ contains $\sigma_{j}$.
\item On the other hand, there is a restricted local $A_q$ perturbation of $\nu$ in which $\sigma_{j}$ appears as the second simplex, right after $\sigma_{i}$. This means that $\sigma_{i}, \sigma_{j}$ and the simplices appearing before $A_q$ contain a nulhomology of $\alpha$.
\item There is a local $A_q$-perturbation of $\nu$ in which $\sigma_{i}$ appears as the second simplex, right after $\sigma_{j}$. By (2) the terminal simplex of $\alpha$ in this perturbation is either $\sigma_i$ or $\sigma_j$. If $\sigma_{j}$ was the terminal simplex of $\alpha$ in this perturbation, $ \sigma_{j}$ and the simplices appearing before $A_q$ would contain a nulhomology of $\alpha$.
Consequently (a) would imply the terminal simplex of $\alpha$ in any local $A_q$ perturbation of $\nu$ is $\sigma_j$ implying $|\Sigma_{t}|=1,$ a contradiction. According to (b), $\sigma_{i}$ appears as the terminal simplex for this permutation implying $\{\sigma_{i}, \sigma_{j}\} \subseteq \Sigma_{t}$.
\end{enumerate}
If $\sigma_{i}$ was a birth simplex, the term $\partial \sigma_{i}$ could have been replaced by a combination of boundaries of simplices appearing before $A_q$. We would thus transform the nulhomology mentioned in (c), which consists of terms containing $\sigma_{i}, \sigma_{j}$ and simplices appearing before $A_q$, into a nulhomology consisting of terms $\sigma_{j}$ and simplices appearing before $A_q$. Such a nulhomology does not exist, as was mentioned in (c), so $\sigma_{i}$ must be a terminal simplex in $K^f$.
\item[(4)] Let $\Delta_1=\sigma_{i}$.
\begin{enumerate}
\item There is a restricted local $A_q$-perturbation of $\nu$ in which $\sigma_{j}$ appears as the last simplex and $\sigma_{i}$ appears as the second simplex from the last. As $\Sigma_{t_0}=\{\sigma_{i}\}$, the simplices of $A_q\setminus\{\sigma_{i}, \sigma_{j}\}$ and all the simplices appearing before them do not contain a nulhomology of $\alpha$, i.e., each nulhomology induced by a local $A_q$-perturbation of $\nu$ contains $\sigma_{j}$ or $\sigma_{i}$.
\item On the other hand, there is a restricted local $A_q$-perturbation of $\nu$ in which $\sigma_{i}$ appears as the first simplex. This means that $\sigma_{i}$ and the simplices appearing before $A_q$ contain a nulhomology of $\alpha$.
\item There is a local $A_q$-perturbation of $\nu$ in which $\sigma_{i}$ appears as the last simplex and $\sigma_{j}$ appears as the second simplex from the last. By (a) the critical the terminal simplex of $\alpha$ in this perturbation is either $\sigma_i$ or $\sigma_j$. If $\sigma_{i}$ was the terminal simplex of $\alpha$ in this perturbation, then simplices of $A_q\setminus\{\sigma_{i}\}$ and all the simplices appearing before them do not contain a nulhomology of $\alpha$. By (b) this would mean $|\Sigma_{t}|=1,$ a contradiction. Thus $\sigma_{j}$ appears as the terminal simplex for this permutation implying $\{\sigma_{i}, \sigma_{j}\} \subseteq \Sigma_{t}$.
\end{enumerate}
It remains to prove that $\sigma_{j}$ is a birth simplex in $K^f$.
By (b) we have
$$
\alpha = \partial \mu_1 \sigma_i +\partial \sum_{k} \lambda_k \tau_k,
$$
with $\mu_1 \neq 0$ and $\tau_k$ being simplices appearing before $A_q$. On the other hand, (c) implies
$$
\alpha = \partial \mu_2 \sigma_j +\partial \sum_{l} \lambda'_l \tau'_l,
$$
with $\mu_2 \neq 0$ and each $\tau'_l$ being a simplex from $A_q$ or appearing before $A_q$. Subtracting the equations we obtain
$$
\partial \mu_2 \sigma_j =\partial\Big(- \sum_{l} \lambda'_l \tau'_l+ \mu_1 \sigma_i + \sum_{k} \lambda_k \tau_k\Big).
$$
Recall that $f(A_q) \subset [f(\sigma_i), f(\sigma_j)]$ and thus the $f$-values of simplices $\tau_k$, $\tau'_l$ and $\sigma_i$ are below $f(\sigma_j)$.
The last equality thus implies $\sigma_{j}$ is a birth simplex in $K^f$.
\end{enumerate}
\end{proof}
\input{example5figure2.tex}
\input{example5figure1.tex}
\input{example6.tex}
\subsection{Sufficient conditions for rigidity of homology classes}
\begin{theorem}
\label{ThmConseq1}
Given an injective filtration function $f$ on a simplicial complex $K$ and a homology class $[\alpha]\in H_n(K^f_a)$ born at $a$ and terminating at $b$, let:
\begin{itemize}
\item $R_u =\{|f(\tau)-b|;\; \tau \textrm{ a birth $n$-simplex with } f(\tau)>b\}$ or $\infty$ if undefined;
\item $R_l =\{|f(\tau)-b|;\; \tau \textrm{ a terminal $n$-simplex with } f(\tau)<b\}$ or $\infty$ if undefined.
\end{itemize}
Then $[\alpha]$ is $\varepsilon$-terminally-rigid for $\varepsilon=\frac{1}{2}\min\{b-a, R_u, R_l\}.$
\end{theorem}
\begin{proof}
For a generic $f$ the conclusion follows from Proposition \ref{PropStep2}.
Assume $f$ is not generic. For a positive $\delta < \max\{\rho/2,\varepsilon/2 \} $ choose a generic function $f_\delta$ such that $||f-f_\delta || \leq \delta$. Quantities $b-a, R_u, R_l$ for this new function are at most $2\delta$ smaller than the original quantities for $f$ and hence $[\alpha]$ in $K^{f_\delta}$ is $(\varepsilon-\delta)$-terminally-rigid.
As any $(\varepsilon-2\delta)$-perturbation of $f$ is also an $(\varepsilon-\delta)$-perturbation of $f_\delta$, we conclude $f$ is $(\varepsilon-2\delta)$-terminally-rigid. As $\delta$ may be arbitrarily small, Lemma \ref{LemmaUSC} concludes the proof.
\end{proof}
\section{Rigidity for barcodes}
In the previous sections we assumed $\varepsilon< (b-a)/2$, which ensures that any $\varepsilon$-perturbation $g$ of $f$ still contains a non-trivial homology class $[\alpha]$ in $K^g$. Our main results so far described the change of the terminal simplex of $[\alpha]$ with $\varepsilon$-perturbations. The situation is a bit more problematic with barcodes. While the barcodes induced by $f$ and $g$ are at the bottleneck distance at most $\varepsilon$, there is no natural way to define the underlying matching \cite{Bauer}. In particular, if $[\alpha]$ represents a bar, the bar matched to it by the isometry theorem may not be represented by $[\alpha]$. On a similar note, given a homology class $[\alpha]$ that is born at $a$ and terminates at $b$, there may be no bar of the form $[a,b)$, see Remark \ref{RemBarcodeIssue}.
\begin{remark}
\label{RemBarcodeIssue}
Let $f$ be an injective filtration function on a simplicial complex $K$ and let $n$ be a positive integer. Assume that for some $a,b\in\mathbb{R}$ an $n$-cycle $\alpha$ is created, so that the corresponding homology class $[\alpha]\in H_n(K_r^f)$ is born at $a$ and terminates at $b$. The creation of $\alpha$ causes the birth of a bar in the persistence diagram. A while later, $\alpha$ might become homologuous to an older cycle $\alpha'$ at a time $c\in (a,b)$, at which point the bar born at $a$ would be terminated according to the elder rule. The class $[\alpha]$ becomes trivial in $H_n(K^f_{b})$ however, so there must exist a bar in the persistence diagram that begins at or before $a$ and dies exactly at $b$. In other words, the starting point of such a bar might be ambiguous, but the moment of its termination is certain.
\end{remark}
This leads us to focus on a setting in which the same homology class determines the endpoint of a designated bar and of its matched bar arising from $\varepsilon$-perturbation.
\begin{theorem}
\label{ThmBarcodeFinal}
Let $f$ be an injective filtration function on a simplicial complex $K$. Assume $[a,b)$ is a bar of $\{H_n(K^f_r)\}_{r\in \mathbb{R}}$ represented by $[\alpha]\in H_n(K^f_a)$ (a homology class born at $a$ and terminating at $b$). Choose $\varepsilon < (b-a)/4$ and assume that for all other bars $[a_i,b_i)$ of $\{H_n(K^f_r)\}_{r\in \mathbb{R}}$ either $a_i > a+2\varepsilon$ or $b_i < b-2\varepsilon$. Let $g$ be an injective filtration function satisfying $||f-g||_\infty \leq \varepsilon$, and assume the induced matching matches the bar $[a,b)$ of $\{H_n(K^f_r)\}_{r\in \mathbb{R}}$ to a bar $[a',b')$ of $\{H_n(K^g_r)\}_{r\in \mathbb{R}}$. Then
\begin{enumerate}
\item[(1)] $[\alpha]$ as a homology class in $\{H_n(K^g_r)\}_{r\in \mathbb{R}}$ terminates at $b'$ and
\item[(2)] the simplex in $K^f$ terminating the bar $[a,b)$ is the same as the simplex in $K^g$ terminating the bar $[a',b')$ if $\varepsilon \leq \frac{1}{2} \min \{R_u,R_l\},$ where:
\begin{itemize}
\item $R_u =\{|f(\tau)-b|;\; \tau \textrm{ a birth $n$-simplex with } f(\tau)>b\}$ or $\infty$ if undefined;
\item $R_l =\{|f(\tau)-b|;\; \tau \textrm{ a terminal $n$-simplex with } f(\tau)<b\}$ or $\infty$ if undefined.
\end{itemize}
\end{enumerate}
\end{theorem}
\begin{proof}
Let
$$\{H_n(K^g_r)\}_{r\in \mathbb{R}} = \bigoplus_{j\in J} \mathbb{F}_{[a'_j,b'_j)}$$
be the decomposition into interval modules, indexed so that $[a',b')$ corresponds to $[a'_0,b'_0)$. Our assumption on the barcode of $ \{H_n(K^f_r)\}_{r\in \mathbb{R}}$ and the stability theorem imply the following:
if $a + \varepsilon \in [a'_j,b'_j)$ then either $j=0$ or $b'_j <b - \varepsilon$. For each $j$ let $[\alpha_j]$ be the homology class corresponding to the bar $[a'_j,b'_j)$, i.e., $[\alpha_j]$ is born at $a'_j$ and terminates at $b'_j$.
The homology class
$$[\beta] = \sum_{j\in J} \lambda_j [\alpha_j]\in H_n(K^g_{a + \varepsilon})$$
terminates either at:
\begin{itemize}
\item $b'$, if $\lambda_0 \neq 0$, or
\item before $b - \varepsilon$, if $\lambda_0 = 0$.
\end{itemize}
We now apply this observation to $[\alpha]$. By stability theorem it represents an element of $H_n(K^g_{a + \varepsilon})$ and does not terminate before $b - \varepsilon$, so it terminates at $b'$ and this concludes the proof of (1).
Bars $[a,b)$ of $\{H_n(K^f_r)\}_{r\in \mathbb{R}}$ and $[a',b')$ of $\{H_n(K^g_r)\}_{r\in \mathbb{R}}$ both terminate when $[\alpha]$ terminates. Conclusion (2) now follows from Theorem \ref{ThmConseq1}.
\end{proof}
\begin{remark}
Theorem \ref{ThmBarcodeFinal} provides a sufficient condition on the structure of the barcode that guarantees that the simplex terminating the bar matched to the designated bar $[a,b)$ remains constant through $\varepsilon$-perturbations of the filtration function $f$. Going beyond the region of unique critical simplex within the setting of Theorem \ref{ThmBarcodeFinal}, let $\varepsilon_0$ be the maximal $\varepsilon$ for which the terminal simplex terminating the bar matched to the designated bar $[a,b)$ remains unique $\Delta_1$ through $\varepsilon$-perturbations of the filtration function $f$. Proposition \ref{PropStep2} allows us to deduce at least one additional simplex $\Delta_2$ that appears as the terminal simplex terminating the bar matched to the designated bar $[a,b)$ for some $\varepsilon$-perturbations of the filtration function $f$ with $\varepsilon > \varepsilon_0$. Namely, $\Delta_2$ is an $n$-simplex with the function value $f(\Delta_1) \pm 2\varepsilon_0$. If $f$ is generic there is only one such simplex.
To put it differently, we may identify potential simplices generating non-rigidity from the structure of the barcode.
\end{remark}
|
\section{Introduction}
With the rapid development of mobile Internet and big data, a huge amount of data is generated in the network, but the data that users are really interested in a very small portion. To extract the information that users are interested in from the huge amount of data, the information overload problem needs to be solved. Baidu and Google use information retrieval and search engine technology to locate data of interest to users, which to some extent simply solves the information overload problem. However, these search engines do not consider the characteristics of different users, and when users search for the same keywords, the system retrieves the same content, and there is a large amount of redundant information of no value to users, which cannot meet the personalized needs of users. In the era of mobile internet, the user's characteristics and other information should be combined in the massive amount of data to quickly and accurately recommend content to the user, as far as possible to meet the user's personalized needs. Therefore, there is an urgent need to realize high-speed and effective retrieval in tens of thousands of micro-videos. Video data content contains complex meanings, and there are intrinsic connections between video data. In visual modality, association relationships exist between visual features and features, between video clips and clips, and between video semantics and semantics. Through these correlation relationships, the computational cost of semantic detection can be reduced and the quality of search can be improved. Correlations between video semantics have an important role, and problems such as synonymy and polysemy can occur between video semantic contents, which are caused by ignoring the correlations of video semantics and can lead to a series of defects.
In recent years, a great deal of progress has been made in the study of problems related to preference and click-through prediction~\cite{SI2}. The most widely used prediction method in industry is the use of logistic regression (LR) to learn click-through prediction models~\cite{chapelle2014simple,richardson2007predicting,mcmahan2013ad}.LR has the advantage of being simple and very easy to implement for massively real-time parallel processing, but linear models have limited learning ability and cannot capture the information carried by higher-order features (nonlinear information) ~\cite{zhang2016deep}. Joachims et al.~\cite{joachims2002optimizing} proposed a SupportVector Machine (SVM) model to predict ad click-through rates, which can effectively handle multidimensional nonlinear data but cannot predict sparse ads with large data volumes.Lee et al.~\cite{lee2018estimating} modeled the data hierarchically from a tripartite perspective of media, users, and advertisers. Shen et al.~\cite{shen2012personalized} proposed a click-through rate prediction model based on collaborative filtering and tensor decomposition. The model mines users' personalized preferences based on the relationship between users, queries, and documents to improve prediction accuracy.Rendle et al.~\cite{BPR} combined the advantages of support vector machines (SVM) and decomposition models to propose a Factorization Machines (FM) model, which uses decomposition parameters to model all interactions between variables and can be used in very sparse data for parameter estimation and has better prediction quality compared to SVM, moreover, FM is a general-purpose predictor that can be used with any real-valued feature vector.
The current research on micro-videos is mainly focused on the following directions. Micro-video scene classification: e.g., Liu et al.~\cite{liu2017towards} capture the short duration of micro-videos and propose an end-to-end model that jointly models sparsity and multi-sequence structure as a way to capture the temporal structure between video frames and the sparse representation of Micro-videos; Micro-video popularity prediction: Jing et al.~\cite{jing2017low} propose a direct push learning method based on low-rank multi-view embedding to seek a set of specific low-rank constrained projection matrices of views to map multi-view features into a common subspace; in addition, graph regularization terms are constructed to improve generalization ability and further prevent overfitting problems.Micro-video user recommendation: Chen et al.~\cite{chen2020learning} proposed a deep neural network model that fuses multiple interest feature representations of users, and used a fusion of four user interest representations: user information, entry information, history information and neighbor information in the paper. For micro-videos as multimedia contents often have multiple labeled information, the author works on the problem of micro-video multi-label classification. The traditional supervised learning is to learn the mapping relationship from sample space $X$ to label space $Y$, i.e., W: $X\rightarrow Y$~\cite{yu2020personalized}. At this time, $Y$ is mostly a label with unique semantic information, unlike this, in reality, there are co-occurrence properties of each label, for example, "singing" and "playing guitar often appear in the same sample at the same time, and there is a strong semantic correlation between the tags, so the task becomes to learn the mapping relationship from the sample space to the tag space with multiple semantic information, and this relationship is one-to-many. Existing approaches to multi-label learning are as follows: converting multi-label classification problems to binary or multi-classification problems in order to utilize classification algorithms that achieve good results on traditional supervised learning; using low-rank representations to obtain a common low-rank representation, such as Jia et al.~\cite{jia2019facial} proposed a method that uses low-rank structure learning model to stably obtain local correlations between labels; Chen et al.~\cite{d2015review} proposed a graph-based semantic association regularization method for enhancing representation learning, which is mainly used to explore interconnections in terms of personal attributes as a way to perform privacy detection of UGCs. Based on the above analysis, this paper focuses on the research and implementation of micro-video personalized recommendation algorithm with multimodal data sources:
\begin{itemize}
\item For multimodal information, subspace coding learning is introduced to build a coding network from public potential representations to multimodal feature information, taking into account the consistency and complementarity information under each modality to obtain a public representation of the complete eigenvalue.
\item An end-to-end reordering model based on deep learning and attention mechanism, called interest-related product similarity model based on multimodal data, is proposed for providing top-N recommendations.The multimodal feature learning module, interest-related network module and product similarity recommendation module together form the new model.
\item By conducting extensive experiments on publicly accessible datasets, the results demonstrate the state-of-the-art performance of our proposed algorithm and its effectiveness.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{Charts/micro1.pdf}
\caption{An illustration of CF-based recommendation. }
\label{fig:figure1}
\end{figure}
\section{Related Work}
The study of video recommendation systems has affected all aspects of science, society, economy, and life, and they are gradually becoming one of the most powerful and popular tools nowadays as tools that can generate huge business benefits~\cite{KG}. Personalized recommendation system is an intelligent algorithm and decision based on big data. The online recommendation system makes personalized matching by capturing users' historical behavior data (browsing, viewing, purchasing, rating, etc.) and the content information of the sold products (profile, details, applicable scenarios, etc.), and corrects and optimizes its recommendation results based on the feedback results of interaction with users (clicks, ratings, etc.)~\cite{LightGCN,wei2021contrastive}.
\subsection{Collaborative Filtering}
In 1992, Tapestry, the first system using Collaborative Filtering (CF), emerged to predict user preferences by learning from user-product interactions, and soon attracted strong interest in academia and industry because of its data availability and simplicity. The core idea of CF is to generate effective recommendations based on historical user interactions by mining the implied user relationships, product relationships, and user-product relationships. The data collection relies on the back-end system to track users' online footprints, such as browsing, clicking, rating, purchasing, viewing, etc. Specifically, CF~\cite{suganeshwari2016survey} can be divided into memory-based collaborative filtering and model-based collaborative filtering.User-based CF identifies users whose preferences for products are close to those of the users to be predicted, and thus makes recommendations for the users to be predicted based on the preferences of this group of users for new products.Item-based CF is more commonly used, and this algorithm was first proposed by Amazon and put into use~\cite{linden2003amazon}. As in Figure 1, Item-based CF aims to find products that are similar to the new product and use the user base that prefers similar products as the deliverable user base for the new product. In contrast, model-based CF is based on historical rating data and uses machine learning models to mine correlations and make predictions.Simon Funk improved on the traditional SVD by proposing the Funk-SVD algorithm, which represents products and users in the same hidden space to reflect the hidden correlations between users and products. Later, Yehuda Koren trained SVD based on gradient descent method in Netflix algorithm competition and named it as hidden semantic model~\cite{bell2007modeling}. Koren improved and perfected SVD based on SVD to make it applicable to hidden feedback scenario that proposed SVD++ algorithm~\cite{koren2008factorization}. The later proposed TimeSVD++~\cite{koren2009collaborative} considers that the longer the preference is from the current prediction, the less reliable it is, and the closer the preference is to the current prediction, the more influential it is. Factorization Machines (FM) as well as extension algorithms were proposed and continuously improved the recommendation performance~\cite{MF}.
\subsection{Content-based Recommendations}
Unlike collaborative filtering, which requires content information about the item (e.g., title, description) to compare the similarity between items and does not take into account the historical relevance between users and items. The descriptive information of the product is processed by data as a feature vector thus used to create a model reflecting the user's preferences. Content-based recommendation methods utilize information containing user tastes, preferences, and needs as well as a variety of models modeled for specific purposes, where the modeling information can come from explicit or implicit feedback from users~\cite{salton1975vector,jiang2020aspect}, and the available modeling models are various machine learning models, vector space models, etc.Content-based filtering (CBF) models require as input the target user’s behavioral data (but not those of non-target users’) together with item content information, typically represented as real-values vectors. To give some examples in the multimedia domain, item properties can be words or concepts in a text, colors in an image, amount of motion in a movie, or rhythm in a music piece.
\subsection{Deep Learning-based Recommendations}
Due to the excellent performance capabilities of deep learning on many complex tasks, academia and industry have competed to extend it to a wider range of applications and achieve state-of-the-art results so far. Recommender systems are an important part of the industrial landscape. For many online websites and mobile applications, it is an important tool for promoting sales and services. Recently, many researchers have adopted deep learning to further improve the quality of their recommendations~\cite{covington2016deep,KGAT,MMGCN,cheng2016wide,tao2020mgat}.Covington et al.~\cite{covington2016deep} implemented a recommendation algorithm based on a multilayer perceptron on the YouTube platform with 1.9 billion users. Cheng et al.~\cite{cheng2016wide} proposed an app recommendation system for Google Play called Wide\&Deep, which uses a hybrid model of linear and deep models.The news recommendation business in the Yahoo website uses an RNN-based news recommendation system~\cite{okura2017embedding}.Application of MLP in recommendation algorithms:MLP is concise and clear in terms of mathematical theory and has been shown to approximate any network measurable function to any desired accuracy~\cite{hornik1989multilayer}.NCF~\cite{NCF} is a framework that captures both linear and nonlinear relationships between users and products.ACF ~\cite{ouyang2014autoencoder} is the first collaborative recommendation model based on autoencoders. Instead of using the original partial observation vector, it is decomposed into integer ratings. CFN~\cite{strub2016hybrid} takes a vector of users or products with auxiliary information such as user attributes and product descriptions as input then reconstructs it at the output layer, where the loss function is the gap between input and output.
Application of CNNs in recommendation algorithms: CNNs are very effective for processing visual, textual and speech information, and most of them are used to extract features in recommendations. Wang et al.~\cite{wang2017your} used CNNs to extract image features and proposed a visual content-enhanced point-of-interest recommendation system called VPOI, which confirmed the key role of visual features for point-of-interest recommendations. ConTagNet~\cite{rawat2016contagnet} is a tag recommendation system that introduces context-awareness, where one module uses a neural network to learn image features and another module uses a two-layer fully connected feed forward neural network to model the contextual representation, and finally the outputs of the two neural networks are connected into a softmax function thus predicting the probability of each candidate tag. Hidasi et al.~\cite{hidasi2015session} proposed a GRU-based session recommendation system. The input is the product coding status associated with the session say N products, then 1 at the products associated with the session and 0 at the irrelevant products, then all the products are coded as a vector of length N. The output is the maximum likelihood probability that each product may be the next relevant product.RRN~\cite{wu2017recurrent} is a nonparametric recommendation model built on RNNs that can model the seasonal evolution of products and changes in user preferences over time, utilizing two LSTM networks to model dynamic user states and product states, and to account for fixed long-term user interests and static product characteristics, the model also integrates the hidden factor vectors of users and items.
\section{Methodology}
\subsection{Model Framework}
Given a set of micro-video datasets $\mathcal{X}=\{X^{(1)},X^{(2)},\dots,X^{(V)}\}$,where $V$ is the number of views, $X^{(V)}\in {R^{D_v \times N}}$ is the feature matrix under the v-th view, $D_v$is the feature dimension under the v-th view, $N$ is the number of short video samples. Also given the label matrix$Y=Y_{ij}\in \{0,1\}^{N\times C}$, where $C$ is the number of labels. If the i-th micro-video carries the j-th label, then the corresponding $Y_{ij}=1$ , and vice versa, $Y_{ij}=0$.
Assuming the existence of a common subspace that can portray features under different perspectives, with the aim of making full use of information from each perspective.Suppose there is a shared common representation $H$ , which can reconstruct the features of each sample at each viewpoint by a set of mappings, i.e., $X^{(V)}=f_v(H)$, where, for any viewpoint $v$, construct the reconstruction mapping$f_v(\cdot)$ belonging to its corresponding viewpoint. In the following, the reconstructed mappings of sample features are constructed using multilayer perceptual neural networks for each of the three modalities; in addition, for the same sample, different modalities share the same complete common representation $H$ , which encodes the sample features $X^{(V)}$ in multiple modalities.
Given a common representation each perspective is conditionally independent. The key to learning for a multi-view subspace is to effectively encode the information for the different available perspectives as:
\begin{equation}
p(\mathcal{X}|H)=p(X^{(1)}|H)p(X^{(2)}|H)\dots p(X^{(v)}|H),
\end{equation}
For the given $\mathcal{X}$, the likelihood modeling of the common representation$H$ can be represented as:
\begin{equation}
p(\mathcal{X}|H)\propto e^{-\Delta (x,f_v(H))},
\end{equation}
where,$H=\{h_n\}_{n-1}^N\in R^{D\times N}$,$h_n$ corresponding to the embedding representation of n-th sample.It can be found that maximizing the likelihood function is equivalent to minimizing the reconfiguration loss:
\begin{equation}
min{\Delta (x,f_v(H))}=min\sum _{v-1}^v\lvert| U^{(v)}-X^{(v)}\rvert|^2,
\end{equation}
where $X^{(v)}$ is the sample feature matrix at the v-th viewpoint and $U^{(v)}$ is the output of the sample at the v-th viewpoint after reconstruction mapping. The reconstructed loss from input to output of the sample is established and the reconstructed network is trained with its constraints. Its column vector $u^{(v)}_n$ is specified as:
\begin{equation}
u^{(v)}_n=\theta(W^{(v)}_h h_n +b^{(v)}_h),
\end{equation}
where $W^{(v)}_h$ and $b^{(v)}_h$ denote the weight and bias of the network at the v-th viewpoint, respectively. $\theta(\cdot)$ is the activation function.
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{Charts/micro2.pdf}
\caption{An illustration of our model framework.}
\label{fig:figure1}
\end{figure*}
\subsection{Feature Filtering}
Suppose the number of scene categories is $N_{scene}$, for the i-th micro-video, the purpose of scene recognition is to find the the maximum value of the scene prediction probability:
\begin{equation}
p_{v_i}^s=max(p_{v_i}^{s_j}) (1\leqslant j \leqslant N_{scene},1\leqslant i \leqslant n),
\end{equation}
where $v_i$ is the i-th micro-video, and $p_{v_i}^{s_j}$ is the probability value of the i-th micro-video corresponding to the j-th scene. However, it is necessary to keep the probability values of $v_i$ in all scenes to retain as much useful information in the video as possible information:
\begin{equation}
f_{v_i}^s=\{p_{v_i}^{s_j}\} (1\leqslant j \leqslant N_{scene},1\leqslant i \leqslant n),
\end{equation}
Next, behavioral features need to be extracted from the micro-video, using variable convolutional network of dual-flow CNN algorithm for RGB feature and Flow feature extraction. The advantage of the variable convolutional network is that it can change the shape of the convolutional kernel by adding offsets to the convolutional kernel, thus improving the convolutional network's ability to adapt to the image adaptation capability.
Assuming that the number of behavior categories is $N_action$, for the i-th micro-video, the result of behavior recognition can be defined as:
\begin{equation}
p_{v_i}^A=max(p_{v_i}^{a_k}) (1\leqslant k \leqslant N_{action},1\leqslant i \leqslant n),
\end{equation}
where $v_i$ is the i-th micro-video and $p_{v_i}^{a_k}$ is the probability value of the i-th micro-video corresponding to the i-th probability value of the k-th behavior, which can be defined as:
\begin{equation}
p_{v_i}^{a_k}=p_{v_i}^{a_k^{RGB}}p_{v_i}^{a_k^{Flow}},
\end{equation}
Similarly, the behavioral feature extraction part needs to keep the probability values of each micro-video for all behaviors:
\begin{equation}
f_{v_i}^A=\{p_{v_i}^{a_k}\} (1\leqslant k \leqslant N_{action},1\leqslant i \leqslant n),
\end{equation}
According to the above analysis, for the micro-video $V = \{v_1,v_2,\dots,v_n\}$, the joint features are:
\begin{equation}
F_V=\{f_{v_1},f_{v_2},\dots,f_{v_n},\},
\end{equation}
\subsection{Preference Prediction}
After merging the high-impact features extracted by GBDT with discrete features for unique thermal coding, the input feature vector is decomposed into two-by-two factors, and the preference rate of micro-videos is used as the output, and the micro-video preference rate prediction model is defined as:
\begin{equation}
y(x)=w_0 +\sum w_i x_i +\sum_{i=1}^n \sum_{j=i+1}^n (V_i, V_j)x_i x_j,
\end{equation}
where $x_i$ is the value of the i-th feature, $n$ is the dimension of the micro-video feature, and $w_0\in R$is the global bias, $w_i \in R^n$ is the influence factor of the i-th feature, and $V \in R^{n\times h}$ is the interaction parameter between the mutually exclusive feature components. $V_i,V_j$ represent the dot product of two vectors $V_i$ and vector $V_j$ of dimension $h$.
\begin{equation}
(V_i, V_j)=\sum _{f=1}^h v_i \cdot V_{(i,f)},
\end{equation}
where $V_i$ denotes the i-th dimensional vector of the coefficient matrix $V$ and $V_i =(v_{i,1},v_{i,2},\dots,v_{i,h}),h\in N ^0$ is the hyperparameter.
\section{EXPERIMENTS}
\begin{table}
\centering
\renewcommand\arraystretch{1.4}
\caption{ Statistics of the evaluation dataset.}
\label{table_1}
\setlength{\tabcolsep}{2.0mm}
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Dataset}&\textbf{Users}&\textbf{Items}&\textbf{Interactions}&\textbf{Density}\\
\hline
\textbf{MovieLens}&$6040$&$3685$&$998034$&$4.48\%$\\
\hline
\textbf{Amazon}&$39385$&$23033$&$278677$&$0.031\%$\\
\hline
\end{tabular}
\vspace{-2mm}
\end{table}
\subsection{Dataset}
We validated the proposed algorithm on the public dataset MovieLens 1M and a partial subset of the Amazon product dataset Clothing\&Shoes\&Jewelry. Table1 summarizes the statistical properties of the two datasets.
MovieLens: Because standard MovieLens dataset does not contain multimodal information, we used the OMDb API to crawl each movie's corresponding poster and text synopsis as visual information and textual information for validating Multimodal IRIS. We used the presence or absence of historical user ratings as implicit feedback.
Amazon: Amazon includes text and image information for each product. In order to quickly validate our model, we ignore users with less than five historical interactions and remove products with ratings below 5 in preprocessing. We use users' review histories as implicit feedback.
\subsection{Baseline}
VBPR~\cite{he2016vbpr}: VBPR integrates visual information into the prediction of people's preferences. VBPR has a significant improvement over matrix decomposition models that rely only on user hidden vectors and product hidden vectors. We have adapted VBPR to be suitable for learning with logarithmic loss functions.
FISM~\cite{kabbur2013fism}: This model using a mixture of hidden factor models and nearest neighbor based models is consistent with our proposed Multimodal IRIS, and thus the method is a very important benchmarking approach.
NCF~\cite{he2017neural}: NCF is based on a matrix decomposition framework that utilizes MLP to model the nonlinear relationship between users and products. The model achieves the best performance among hidden factor models due to the powerful representation capability of neural networks.
NAIS~\cite{he2018nais}: NAIS is designed to differentiate the importance of different historical products in user modeling. The product similarity recommendation using the attention mechanism is also part of Multimodal IRIS.
\subsection{Evaluation Metrics}
Using all interaction data as positive samples, for each user we randomly select one sample for composing the test set and the rest for the training set. Then we draw K negative samples from the uninteracted products with equal probability for training, K being the hyperparameter used to control the sampling. In addition, the validation phase is not possible for us to sort all the uninteracted products in the validation phase because the set containing all the products is relatively large which leads to a large time consumption. Therefore, for each positive example in the test set, we randomly select 99 negative examples to form a test pair, so that each user in the test set corresponds to a set of 100 candidates, as is common in most research work.
The commonly used measures in typical top-N recommendations based on implicit feedback are Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG). The HR@N is used to measure whether the positive examples in the test set appear in the top-N recommendation list; NDCG@N also considers the position of the positive examples in the test set in the top-N recommendation list, where N is a hyperparameter. We compared the experimental results based on comparing at different N. For both metrics, higher values represent better model performance. The specific calculation is as follows:
\begin{equation}
HR@N=\frac{\sum_{i=1}^N co(i)}{N} ,
\end{equation}
where $co(i)$ indicates whether the positive example retained in the test set appears in the top-N recommendation list of user $i$. If it does, the value is 1, otherwise the value is 0. $N$ indicates the number of users.
\begin{equation}
NDCG@N=\frac{1}{N} \sum_{i=1}^N hits(i),
\end{equation}
\begin{table}
\centering
\renewcommand\arraystretch{1.5}
\caption{ Experimental results on Movielens dataset.}
\label{table_2}
\setlength{\tabcolsep}{2.0mm}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Model}&\textbf{HR@10}&\textbf{NDCG@10}\\
\hline
\textbf{VBPR}&$0.7345$&$0.5187$\\
\hline
\textbf{FISM}&$0.8117$&$0.5753$\\
\hline
\textbf{NCF}&$0.8112$&$0.5637$\\
\hline
\textbf{NAIS}&$0.8129$&$0.5834$\\
\hline
\textbf{Ours}&$0.8287$&$0.5899$\\
\hline
\end{tabular}
\vspace{-2mm}
\end{table}
\begin{table}
\centering
\renewcommand\arraystretch{1.5}
\caption{ Experimental results on Amazon dataset.}
\label{table_2}
\setlength{\tabcolsep}{2.0mm}
\begin{tabular}{|c|c|c|}
\hline
\textbf{Model}&\textbf{HR@10}&\textbf{NDCG@10}\\
\hline
\textbf{VBPR}&$0.3376$&$0.2087$\\
\hline
\textbf{FISM}&$0.4007$&$0.2456$\\
\hline
\textbf{NCF}&$0.3654$&$0.2143$\\
\hline
\textbf{NAIS}&$0.4233$&$0.2431$\\
\hline
\textbf{Ours}&$0.4643$&$0.2879$\\
\hline
\end{tabular}
\vspace{-2mm}
\end{table}
\subsection{Results Analysis}
From these two tables it can be shown that our proposed model outperforms the deep model (VBPR and NCF) and the shallow model (FISM and NAIS). Using only image information as input, the proposed Image IRIS consistently outperforms compared to VBPR on MovieLens and Amazon, and these results indicate that Image IRIS makes fuller use of image information and helps in recommendation. NCF improves the recommendation performance by fusing MF and MLP to learn linear and nonlinear interaction functions between users and products. The improvement of Multimodal IRIS over NCF confirms the effectiveness of multimodal features and neighborhood-based models. Compared to NAIS on the MovieLens dataset, our model improves $0.78\%$ on HR@10, $1.66\%$ on NDCG@10. Also, on the Amazon dataset, our model improves $4.52\%$ on HR@10, $3.87\%$ on NDCG@10. Therefore, we can conclude that the use of multimodal information can improve the recommendation efficiency. It is worth pointing out that the difference in performance between FISM and NAIS on the Amazon dataset is very small (Amazon is two orders of magnitude sparser than MovieLens). Due to the extreme sparsity of the data, it is not possible to efficiently learn the attention model in NAIS by hidden factors alone. Our model produced more significant results on the Amazon dataset compared to MovieLens, which revealed the necessity of adding multimodal information when dealing with sparse datasets. By comparing Image-add-Text IRIS and Our model, we can see that the use of knowledge sharing units in Our model further improves the evaluation metric values compared to the traditional way of using multimodal information, i.e., direct weighted summation of multimodal data, indicating that knowledge transfer helps feature learning. In conclusion, our model provides superior performance compared to existing methods.
\begin{figure}
\centering
\subfigure[NDCG@N on Movielens]{
\includegraphics[width=0.23\textwidth]{Charts/micro3.pdf}
\label{fig_visualize_1_1}
}
\subfigure[HR@N on Movielens]{
\includegraphics[width=0.23\textwidth]{Charts/micro4.pdf}
\label{fig_visualize_2_3}
}
\subfigure[NDCG@N on Amazon]{
\includegraphics[width=0.23\textwidth]{Charts/micro5.pdf}
\label{fig_visualize_1_1}
}
\subfigure[HR@N on Amazon]{
\includegraphics[width=0.23\textwidth]{Charts/micro6.pdf}
\label{fig_visualize_2_3}
}
\vspace{-10pt}
\caption{Comparison of the results of the hidden factor-based model.}
\label{fig_3}
\vspace{-5pt}
\end{figure}
\section{Conclusion}
There are still relatively few relevant studies conducted on micro-videos, and the main work of this thesis is to study and implement a personalized recommendation algorithm based on multimodal micro-video. The multimodal feature fusion multi-label classification model for micro-videos, which integrates subspace coding learning and multi-label relevance learning from multiple perspectives This model integrates subspace coding learning and multi-label relevance learning from multiple perspectives into a unified framework. The framework utilizes subspace coding networks to learn more representations of common representations across multiple views, while using graph convolutional networks to The framework uses a subspace coding network to learn more representations of common representations from multiple perspectives, and a graph convolutional network to mine semantic correlations between labels. The entire network is updated with a stochastic gradient descent-based alternating learning strategy for parameter solving. The interest-related product similarity model based on multimodal data sources is an end-to-end reordering model based on deep learning and attention mechanism for providing top-N recommendations. Finally, the multimodal data feature learning module, IRN and product similarity recommendation module are unified into one integrated model to achieve performance improvement and adapt to the addition or absence of different modal data. The model takes into account the multimodal data that people may pay more attention to when choosing products, thus greatly improving the accuracy and interpretability of the top-N recommendation task. Multimodality on the publicly Experimental results on the open micro-video dataset demonstrate the effectiveness of this method.
\section{DATA PREPARATION}
According to this, we detail the data preparation, including the auxiliary dataset, micro-video dataset construction, and popular topics collection.
\subsection{Auxiliary Dataset}
To train the deep transfer model, the auxiliary dataset should be large-scale, high-quality, and adaptive to our crossmodal multi-label similarity calculation. Overall, Open Images Dataset\footnote{http://storage.googleapis.com/openimages/web/index.html} is adopted as training dataset. It is a large-scale dataset with 9,011,219 images mainly collected from the Flickr\footnote{https://www.flickr.com/}, and each image is annotated with multiple labels over 600 categories. The labels cover the real-life entities and
are manually drawn by professional annotators to ensure the accuracy and consistency.
\subsection{Micro-video Dataset Construction}
To build the set of micro-videos, lack of the ground-truth of each micro-video still poses a big challenge. Previous studies choose the ground-truth thumbnails by some workers, but this strategy is inadaptive to our task. In our method, we suggest that the thumbnail is the frame whose content is largely related to the popular topics of the microvideo shared platform, yet the topics are affected by the realworld users’ interests which are probably different from those annotating workers’. Therefore, we collect the micro-videos with a large number of clicks to build the dataset and treat their posted thumbnails as the ground-truths. Because these posted thumbnails are the first sights of the micro-videos and convey the only information about the content. More clicks on the thumbnail mean it satisfies more users and is closer to the popular topics of the platform.
In practice, we collect some publishers and capture their posted micro-videos associated with the thumbnails. Since the aim is to obtain the micro-videos with a large number of
clicks, whose thumbnails are close to the popular topics, the influence of publishers themselves has to be eliminated. For this, we choose the publishers according to the number of their followers. As can be seen in Figure 4, the number of followers is limited to the range of 300-1000. Ultimately, we achieve 18,279 publishers to form the publishers set and crawl their micro-videos with more than 200 clicks.
To improve the performance of the thumbnail selection, some candidate frames are extracted from each micro-video based on visual quality and representativeness. Specifically, according to the findings [26], three types of low-quality frames would be detected and filtered out, including frames of dark, blurry, and uniform-colored. And we further filter out the near-duplicate frames via a clustering algorithm over the
high-quality frames. As shown in Figure 5, most micro-videos have 7-12 candidate frames. Although the posted thumbnail may not be contained in the set of candidate frames, we believe its nearest neighbor frame in the set is visually closer to the posted thumbnail. Statistically, the average similarity between the posted thumbnail and its nearest neighbor frame is about 93.8\%. Therefore, we take the nearest neighbor frames as the ground-truths for the micro-videos, instead of inserting the posted thumbnails into the sets of candidate frames. To verify the ground-truth, 1, 000 groups of candidate frames are randomly selected and annotated by 10 Instagram active users. From Figure 6, we observe that with the increasing number of clicks, the accuracy of manual labeling is improved. The number of clicks reflects the degree of attractiveness of the thumbnail to the users. Moreover, the accuracy can exceed 60\% in the worst case, proving that the ground-truth is more suitable than other frames as the thumbnail.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{Charts/figure4.png}
\vspace{-1mm}
\caption{Distribution of the number of followers in our dataset.}
\label{fig_4}
\vspace{-5mm}
\end{figure}
\subsection{Popular Topics Collection}
After the micro-video dataset collection and ground-truth construction, we focus on the popular topics collection. Therefore, about 1 million applicable comments are crawled
from Instagram to capture popular topics. Finally, we go through the entity words and obtain 111 high-frequency words to represent the popular topics. Table I lists a part of entity words and their frequencies.
|
\section{Introduction}
\label{sec:Introduciton}
Intelligent reflecting surfaces (IRSs) have been attracted a great research interest due to the lower hardware complexity, physical size and cost of conventional large arrays~\cite{irs_survey1}. An IRS is an electromagnetic two-dimensional surface that is composed of large number of passive reconfigurable meta-material elements, which reflect the incoming signal by introducing a pre-determined phase shift. This phase shift is controlled via external signals by the base station (BS) through a backhaul control link. As a result, the incoming signal from the BS can be manipulated in real-time, thereby, reflecting the received signal toward the users. Hence, the usage of IRS enhances the signal energy received by distant users and expands the coverage of the BS~\cite{lis_channelEst_3}.
The system configuration of the IRS-assisted systems, i.e., designing reflect beamformers, strongly relies on the knowledge of the channel information. In fact, the IRS-assisted systems include multiple communications links, i.e., a direct channel from BS to users and a cascaded channel from BS to users through IRS. This makes the IRS scenario even more challenging than the conventional massive MIMO (multiple-input multiple-output) systems. Furthermore, the wireless channel is dynamic and uncertain because of changing IRS configurations. Consequently, there exists an inherit uncertainty stemming from the IRS configuration and the channel dynamics. These characteristics of IRS make the system design very challenging~\cite{elbir2020IRS_DL_survey,irs_deepDenoisingNN_CE}.
To address the aforementioned uncertainties and non-linearities imposed by channel equalization, hardware impairments, and sub-optimality of high-dimensional problems, model-free techniques have become common in wireless communications~\cite{elbir2020IRS_DL_survey}. In this context, deep learning (DL) is particularly powerful in extracting the features from the raw data by constructing a model-free data mapping. DL-based channel estimation for IRS-assisted wireless systems is considered in~\cite{elbir_LIS,irs_deepDenoisingNN_CE}. Specifically, \cite{elbir_LIS} devise a twin convolutional neural network (CNN), which takes the least squares (LS) estimates of the direct and cascaded channels as input to estimate the channels. A deep denoising neural network (DDNN) approach is proposed in~\cite{irs_deepDenoisingNN_CE}, where the input of the DDNN is a compressive channel estimate that is obtained via orthogonal matching pursuit (OMP) algorithm from the receive antenna array measurements. These DL techniques employ a centralized learning (CL) scheme, wherein the learning model is trained at a parameter server (PS) (possible at the BS) by collecting the training data from the users. The transmission of the training datasets from the users to the PS introduces huge overhead due to the large size of the datasets. This overhead can be reduced by decentralized techniques, such as federated learning (FL), by bringing the learning task to the edge level, e.g., the mobile users~\cite{spm_federatedLearning,fl_By_Google,elbir2021FL4PHY}. In FL, instead of transmitting the whole dataset to the PS, each user computes the local model updates on its local dataset and only transmits the model updates (gradients) to the PS. Due to its recent success on wireless sensor networks~\cite{FL_Gunduz,fl_IoT}, UAV (unmanned aerial vehicle) networks~\cite{FL_Bennis2} and vehicular networks~\cite{elbir2020federated}, FL has been regarded as a promising tool to exhibit a communication-efficient as well as privacy-preserving learning scheme since it does not involve raw data transmission. Recently, FL has been applied for beamformer design problems in massive MIMO~\cite{elbir2020FL_HB} and IRS-assisted systems~\cite{elbir2020_FL_CE}. In~\cite{elbir2020_FL_CE}, FL is applied for channel estimation in IRS-assisted systems, where the proposed approach requires pilot signals at least the number of BS antennas. The present work eliminates such requirement and provides satisfactory channel estimation performance close to CL.
%
In this paper, we propose an FL approach for channel estimation in IRS-assisted wireless systems. We design a single CNN to jointly estimate the direct and cascaded channels in the presence of insufficient pilot signals. During model training, each user computes the model updates by using the local datasets (i.e., input of received pilots and label of channel data) and transmits them to the BS where the model updates are aggregated and sent back to the users. The proposed FL approach is advantageous since it provides decentralized learning, which significantly reduces the transmission overhead compared to the CL-based techniques while maintaining satisfactory channel estimation performance close to CL. Furthermore, the performance of the proposed method can be attributed to the fact that the designed learning model constructs a non-linear mapping from the insufficient number of received pilots to the channel matrix entries. Therefore, the proposed approach is also effective since it does not involve a matrix inversion operation, which may be required in model-based techniques in massive MIMO applications. To the best of our knowledge, this is the first FL-based method for channel estimation with fewer pilot signals.
\section{Signal Model}
\label{sec:FL_IRS}
We consider the downlink channel estimation for IRS-assisted massive MIMO systems, where the BS has $M$ antennas to serve $K$ single-antenna users with the assistance of IRS, which is composed of $L$ reflective elements. The incoming signal from the BS is reflected from the IRS, where each IRS element introduces a phase shift $\varphi_l$, for $l=1,\dots, L$. This phase shift can be adjusted through the PIN (positive-intrinsic-negative) diodes, which are controlled by the IRS-controller connected to the BS over the backhaul link~\cite{lis_channelEst_3,lis_channelEst_4}. As a result, IRS allows the users receive the signal transmitted from the BS when they are distant from the BS or there is a blockage among them. Assume that the BS transmits the $k$-th user the data symbol $s_k\in\mathbb{C}$ by using a baseband precoder $\mathbf{F}= [\mathbf{f}_1,\dots,\mathbf{f}_K]\in\mathbb{C}^{M\times K}$ . Hence, the downlink $M\times 1$ transmitted signal becomes $ \overline{\mathbf{s}} = \sum_{k=1}^{K} \sqrt{\gamma_k} \bar{\mathbf{f}}_ks_k,$ where {\color{black} $\bar{\mathbf{f}}_k = \frac{\mathbf{f}_k}{||\mathbf{f}_k||_2}$ and} $\gamma_k$ denotes the allocated power at the $k$-th user. The transmitted signal is received from the $k$-th user with two components, one of which is through the direct path from the BS and another one is through the IRS. The received signal at the $k$-th user can be given by
\begin{align}
\label{receivedSignal1}
y_k = \big(\mathbf{h}_{\mathrm{BS},k}^\textsf{H} + \mathbf{h}_{\mathrm{IRS},k}^\textsf{H} \boldsymbol{\Psi}^\textsf{H} \mathbf{H}^\textsf{H} \big) \overline{\mathbf{s}} + n_k,
\end{align}
where $n_{k}\in\mathbb{C}$ denotes the spatially and temporarily white Gaussian noise and
$\boldsymbol{\Psi} = \mathrm{diag}\{\boldsymbol{\psi}\}\in\mathbb{C}^{L\times L}$ and $\boldsymbol{\psi} = [\psi_1, \dots, \psi_{L}]^\textsf{T}\in \mathbb{C}^{L}$ is the reflecting beamformer vector, whose $l$-th entry is $\psi_l = a_n e^{j \varphi_l} $, where $a_l\in \{0,1\}$ denotes the on/off stage of the $l$-th element of the IRS and $\varphi_l \in [0, 2\pi]$ is the phase shift introduced by the IRS. In practice, the IRS elements cannot be perfectly turned on/off, hence, they can be modeled as $ a_n=\left\{\begin{array}{cc}
1 - \epsilon_1 & \mathrm{On}\\
0 + \epsilon_0 & \mathrm{Off}
\end{array}\right.,$
for small $\epsilon_1,\epsilon_0 \geq0$, which represents the insertion loss of the reflecting elements~\cite{elbir_LIS,lis_onoff_ICASSP}.
$\mathbf{H}\in \mathbb{C}^{M \times L}$ is the channel between the BS and the IRS and it can be defined as $ \mathbf{H} = \sqrt{\frac{ML }{N_\mathrm{r}}} \sum_{n=1}^{ N_\mathrm{r}} \alpha_{n} \mathbf{a}_\mathrm{BS}( \phi_{n}^{\mathrm{BS}})\mathbf{a}_\mathrm{IRS}( {\phi}_{n}^{\mathrm{IRS}})^\textsf{H},$
where $N_\mathrm{r}$ and $\alpha_{l}^{\mathrm{IRS}}$ are the number of received paths and the complex gain respectively. $\mathbf{a}_\mathrm{BS}( \phi_{n}^{\mathrm{BS}})\in \mathbb{C}^{M}$ and $\mathbf{a}_\mathrm{IRS}( {\phi}_n^\mathrm{IRS})\in \mathbb{C}^{L}$ are the steering vectors corresponding to the BS and IRS with the angle-of-arrival/departure (AoA/AoD) angles $\phi_{n}^{\mathrm{BS}},\phi_{n}^{\mathrm{IRS}}$, respectively. $\mathbf{h}_{\mathrm{BS},k}$ $\in\mathbb{C}^{M}$ and $\mathbf{h}_{\mathrm{IRS},k}\in \mathbb{C}^L$ represent the direct channels for BS-user and IRS-user links, and they can be defined respectively as
\begin{align}
\mathbf{h}_{\mathrm{BS},k} = \sqrt{\frac{M }{N_\mathrm{r}^\mathrm{BS} }} \sum_{n=1}^{ N_\mathrm{r}^\mathrm{BS}} \alpha_{k,n}^{\mathrm{BS}} \mathbf{a}_\mathrm{BS}( \varphi_{k,n}^{\mathrm{BS}}),
\end{align} and
\begin{align}
\mathbf{h}_{\mathrm{IRS},k} = \sqrt{\frac{L }{N_\mathrm{r}^{\mathrm{IRS}} }} \sum_{l=1}^{ N_\mathrm{IRS}} \alpha_{k,n}^{\mathrm{IRS}} \mathbf{a}_\mathrm{IRS}( \varphi_{k,n}^{\mathrm{IRS}}).
\end{align} Here, $N_\mathrm{r}^\mathrm{BS}$, $\alpha_{k,n}^{\mathrm{BS}}$ and $\mathbf{a}_\mathrm{BS}(\varphi_{k,n}^{\mathrm{BS}})$ ($N_\mathrm{r}^\mathrm{IRS}$, $\alpha_{k,n}^{\mathrm{IRS}}$, $\mathbf{a}_\mathrm{IRS}(\varphi_{k,n}^{\mathrm{IRS}})$) are the number of paths, complex gain and the steering vector with the AoA angle $\varphi_{k,n}^{\mathrm{BS}}$ ($\varphi_{k,n}^{\mathrm{IRS}}$) for the BS-user (IRS-user) communication link, respectively.
Let $\boldsymbol{\Lambda}_k = \mathrm{diag}\{ \mathbf{h}_{\mathrm{IRS},k}\}\in \mathbb{C}^{L\times L}$, then the cascaded channel between the BS and the $k$-th user becomes $\mathbf{G}_k = \mathbf{H} \boldsymbol{\Lambda}_k\in \mathbb{C}^{M\times L}$ and (\ref{receivedSignal1}) can be rewritten as
\begin{align}
y_k = \big(\mathbf{h}_{\mathrm{BS},k}^\textsf{H} + \boldsymbol{\psi}^\textsf{H} \mathbf{G}_k^\textsf{H} \big) \overline{\mathbf{s}} + n_k.
\end{align}
Our aim in this work is to estimate the direct and cascaded channels via FL. In the following, we first introduce how input and output data are designed for the proposed learning model, then discuss how FL-based training is performed.
\section{Channel Acquisition With Fewer Pilots In IRS-Assisted Wireless Systems}
The proposed channel estimation scheme has two stages to separately collect the received data for direct and cascaded channels, $\mathbf{h}_{\mathrm{BS},k}$ and $\mathbf{G}_k$, respectively.
In order to estimate the channels $\{\mathbf{h}_{\mathrm{BS},k}, $ $\mathbf{G}_k\}$, $k = 1,\dots,K$, conventional model-based~\cite{lis_channelEst_3,lis_channelEst_4} and model-free~\cite{deepCNN_ChannelEstimation,elbir_LIS,elbir2020_FL_CE} techniques require at least $\bar{M}\geq M$ orthogonal pilot signals, which can be represented as $\mathbf{S} = [\mathbf{s}_1,\dots, \mathbf{s}_M]\in\mathbb{C}^{M\times M}$ whose columns are orthogonal to each other, i.e., $||\mathbf{s}_{m_1}^\textsf{H}\mathbf{s}_{m_2} ||_2^2 =0$ for $m_1 \neq m_2$, $m_1,m_2=1,\dots,M$. In the presence of insufficient pilot signals, i.e., $\bar{M}< M$, using the incomplete pilot matrix $\bar{\mathbf{S}}\in \mathbb{C}^{M\times \bar{M}}$, the received signal at the $k$-th user becomes
\begin{align}
\label{receivedIRS1}
\mathbf{y}_k = (\mathbf{h}_{\mathrm{BS},k}^\textsf{H} + \boldsymbol{\psi}^\textsf{H}\mathbf{G}_k^\textsf{H}) \overline{\mathbf{S}} + \mathbf{n}_k,
\end{align}
where $\mathbf{y}_k = [y_{1,k},\dots, y_{\bar{M},k}]$ and $\mathbf{n}_k = [n_{1,k},\dots, n_{\bar{M},k}]$ are $1\times \bar{M}$ row vectors. We intend to collect the received data with insufficient number of pilots, then construct a learning model which maps the received data to the channel data.
In the first stage, the received data with respect to $\mathbf{h}_{\mathrm{BS},k}$ can be collected when all of the IRS elements are turned off, i.e., $a_l = 0$, for $l=1,\dots,L$. Then, the $1\times \bar{M}$ received signal at the $k$-th user is
\begin{align}
\label{esthdirect}
\mathbf{y}_{\mathrm{D},k} = \mathbf{h}_{\mathrm{BS},k}^\textsf{H}\overline{\mathbf{S}} + \mathbf{n}_{\mathrm{BS},k}.
\end{align}
The conventional model-based methods, such as least squares (LS) and minimum mean-squared-error (MMSE), fail to obtain $\mathbf{h}_{\mathrm{BS},k}$ from $\mathbf{y}_{\mathrm{BS},k}$ due to the lack of sufficient pilot data. Here, we leverage DL to estimate $\mathbf{h}_{\mathrm{BS},k}$ by selecting $\mathbf{y}_{\mathrm{B},k}$ as input to the CNN.
Next, we consider the cascaded channel estimation. We assume that each IRS element is turned on one by one while all the other elements are turned off. This is done by the BS requesting the IRS via a micro-controller device in the backhaul link so that a single IRS element is turned on at a time. Then, the reflecting beamformer vector at the $l$-th frame becomes $\boldsymbol{\psi}^{(l)} = [0,\dots, 0, \psi_l,0,\dots, 0]^\textsf{T}$, where $a_l = \{0: \tilde{l} = 1,\dots L, \tilde{l} \neq l \}$ and the received signal becomes
\begin{align}
\label{esthcascaded}
\mathbf{y}_{\mathrm{C},k}^{(l)} = (\mathbf{h}_{\mathrm{BS},k}^\textsf{H} +{\mathbf{g}_{k}^{(l)}}^\textsf{H} ) \overline{\mathbf{S}} + \mathbf{n}_k,
\end{align}
where $\mathbf{g}_k^{(l)} \in \mathbb{C}^{M}$ is the $l$-th column of $\mathbf{G}_k$, i.e., $ \mathbf{g}_k^{(l)} = \mathbf{G}_k \boldsymbol{\psi}^{(l)},$ where $\psi_l = 1$. Using the estimate of $\mathbf{h}_{\mathrm{BS},k}$ from (\ref{esthdirect}), (\ref{esthcascaded}) can be solved for $\mathbf{g}_k^{(l)}$, $l = 1,\dots, L$, and the cascaded channel $\mathbf{G}_k$ can be estimated. Then, the received data for $l = 1\dots, L$ can be collected as
\begin{align}
\mathbf{Y}_{\mathrm{C},k} =\left[
\mathbf{y}_{\mathrm{C},k}^{(1)^\textsf{T}},\dots,
\mathbf{y}_{\mathrm{C},k}^{(L)^\textsf{T}}\right]^\textsf{T}\in \mathbb{C}^{L\times \bar{M}}.
\end{align}
In order to train the CNN for IRS-assisted massive MIMO scenario, we select the input-output data pair as $\{\mathbf{y}_{\mathrm{D},k},\mathbf{h}_{\mathrm{BS},k} \}$ and $\{\mathbf{Y}_{\mathrm{C},k},\mathbf{G}_{k} \}$ for direct and cascaded channels, respectively. To jointly learn both channels, a single input is constructed to train a single CNN as $ \boldsymbol{\Upsilon}_k = \left[\begin{array}{c}
\mathbf{y}_{\mathrm{D},k}\\
\mathbf{Y}_{\mathrm{C},k}
\end{array}\right] \in \mathbb{C}^{(L + 1)\times \bar{M}}$. We design the input of the CNN as three \textit{channel} of the input data, i.e., $[{\mathcal{X}}_k]_1 = \operatorname{Re}\{ \boldsymbol{\Upsilon}_k\}$ and $[{\mathcal{X}}_k]_2 = \operatorname{Im}\{ \boldsymbol{\Upsilon}_k\}$, $[{\mathcal{X}}_k]_3 = \angle\{ \boldsymbol{\Upsilon}_k\}$, respectively. We can define the output data as $\boldsymbol{\Sigma}_k = \left[\mathbf{h}_{\mathrm{BS},k}, \mathbf{G}_{k}\right] \in \mathbb{C}^{\bar{M}\times (L+1)}$, hence, the output label can be given by a $2{M}(L+1)\times 1$ real-valued vector as
\begin{align}
{\mathcal{Y}}_k = \left[ \mathrm{vec}\{\operatorname{Re} \{\boldsymbol{\Sigma}_k\}\}^\textsf{T}, \mathrm{vec}\{\operatorname{Im} \{\boldsymbol{\Sigma}_k\}\}^\textsf{T} \right]^\textsf{T}.
\end{align} Consequently, the sizes of ${\mathcal{X}}_k$ and ${\mathcal{Y}}_k$ are $(L + 1)\times \bar{M}\times 3$ and $2{M} (L+1)\times 1$, respectively. Once the CNN is trained, the received signals $\mathbf{y}_{\mathrm{D},k}$ and $\mathbf{Y}_{\mathrm{C},k}$ can be collected and fed to the CNN to estimate the channels. We discuss the FL-based model training in the following section.
\section{FL-based Model Training}
Let $\mathcal{D}_k$ be the local dataset of the $k$-user, in which the $i$-th element is given by $\mathcal{D}_i = (\mathcal{X}_k^{(i)},\mathcal{Y}_k^{(i)})$, where $\mathcal{X}_k^{(i)}$ and $\mathcal{Y}_k^{(i)}$ denote the input and output for $i =1,\dots,\textsf{D}_k$ and $\textsf{D}_k = |\mathcal{D}_k|$ ($\textsf{D} = \sum_k \textsf{D}_k $) is the size of the local dataset. We begin by introducing the training concept in conventional CL-based training, then develop FL-based model training.
In CL-based model training for channel estimation~\cite{deepCNN_ChannelEstimation,elbir2019online,elbir_LIS}, the training of the global CNN is performed by collecting the local datasets $\mathcal{D}_k, k = 1,\dots,K$ from the users. Once the BS has collected the whole dataset $\mathcal{D} = \bigcup_{k} \mathcal{D}_k $, the training is performed by solving the following problem
\begin{align}
\label{lossML}
\minimize_{\boldsymbol{\theta}}
\mathcal{F}(\boldsymbol{\theta}) = \frac{1}{\textsf{D}} \sum_{i = 1}^{\textsf{D}}\mathcal{L}(f( \mathcal{X}^{(i)}|\boldsymbol{\theta}),\mathcal{Y}^{(i)} ) ,
\end{align}
where $\boldsymbol{\theta}\in\mathbb{R}^P$ denotes the learnable parameters and $\mathcal{L}(\cdot)$ is the loss function defined by the learning model as
\begin{align}
\mathcal{L}(f(\mathcal{X}^{(i)}|\boldsymbol{\theta}),\mathcal{Y}^{(i)}) = \| f( \mathcal{X}^{(i)}|\boldsymbol{\theta}) - \mathcal{Y}^{(i)} \|_\mathcal{F}^2,
\end{align}
which is the MSE between the label data $\mathcal{Y}^{(i)}$ and the prediction of the CNN, $f( \mathcal{X}^{(i)}|\boldsymbol{\theta})$ for the whole dataset, i.e., $i = 1,\dots, \textsf{D}$. The minimization of the empirical loss $ \mathcal{F}(\boldsymbol{\theta})$ is achieved via gradient descent (GD) by updating the model parameters $\boldsymbol{\theta}_t$ at iteration $t$ as $\boldsymbol{\theta}_{t+1} = \boldsymbol{\theta}_t - \eta_t \mathbf{g}(\boldsymbol{\theta}_t),$
where $\eta_t$ is the learning rate and
\begin{align}
\mathbf{g}(\boldsymbol{\theta}_t)= \nabla_{\boldsymbol{\theta}} \mathcal{F}(\boldsymbol{\theta}_t) =\frac{1}{\textsf{D}} \sum_{i=1}^{\textsf{D}}\nabla_{\boldsymbol{\theta}} \mathcal{L} (f ( \mathcal{X}^{(i)}|\boldsymbol{\theta}_t), \mathcal{Y}^{(i)}),
\end{align}
denotes the \emph{full} or \emph{batch} gradient vector in $\mathbb{R}^P$~\cite{FL_QSGD,fl_convergenceOnNIIDData}. For large datasets, it is computationally inefficient to implement GD, which motivates the use of stochastic GD (SGD), where $\mathcal{D}$ is partitioned into $M_B$ mini-batches as $\mathcal{D} = \bigcup_{m\in \mathcal{M}_B} \mathcal{D}_m$, for $\mathcal{M}_B = \{1,\dots, M_B\}$. Then, $\boldsymbol{\theta}_t$ is updated by $\boldsymbol{\theta}_{t+1} = \boldsymbol{\theta}_t - \eta_t \mathbf{g}_{\mathcal{M}_B}(\boldsymbol{\theta}_t),$
where ${\mathbf{g}}_{\mathcal{M}_B}(\boldsymbol{\theta}_t) =\frac{1}{M_B} \sum_{m=1}^{M_B} \mathbf{g}_m(\boldsymbol{\theta}_t) $ includes the contribution of gradients computed over $\{\mathcal{D}_m\}_{m\in \mathcal{M}_B}$ as $ \mathbf{g}_m(\boldsymbol{\theta}_t) = \frac{1}{\textsf{D}_m} \sum_{i=1}^{\textsf{D}_m}\nabla_{\boldsymbol{\theta}} \mathcal{L} (f ( \mathcal{X}_m^{(i)}|\boldsymbol{\theta}_t), \mathcal{Y}_m^{(i)}) ,$
where ${\textsf{D}_m} = |\mathcal{D}_m|$ is the mini-batch size and ${\mathbf{g}}_{\mathcal{M}_B}(\boldsymbol{\theta}_t)$ satisfies $\mathbb{E}\{ {\mathbf{g}}_{\mathcal{M}_B}(\boldsymbol{\theta}_t) \} = \nabla_{\boldsymbol{\theta}} \mathcal{F}(\boldsymbol{\theta}_t)$. Therefore, SDG provides the minimization of the empirical loss by partitioning the dataset into $M_B$ mini-batches and accelerates the learning process, which is known as mini-batch learning.
In FL, the training dataset $\mathcal{D}$ is partitioned into small portions, i.e., $\mathcal{D}_k$, $k = 1,\dots,K$, however they are available at the users and not transmitted to the BS. Hence, given initial model $\boldsymbol{\theta}_{t-1}$, the $k$-th user aims to solve the following local problem at iteration $t$, i.e.,
\begin{align}
\label{lossFL}
\minimize_{\boldsymbol{\theta}}
\frac{1}{\textsf{D}_k} \sum_{i = 1}^{\textsf{D}_k}\mathcal{L}(f( \mathcal{X}_k^{(i)}|\boldsymbol{\theta}_{t-1}),\mathcal{Y}_k^{(i)} ),
\end{align}
with the use of the local gradient $\mathbf{g}_k(\boldsymbol{\theta}_t)$. The $k$-th user transmits $\mathbf{g}_k(\boldsymbol{\theta}_t)$ to the BS, which receives $\tilde{\mathbf{g}}_k(\boldsymbol{\theta}_t) = \mathbf{g}_k(\boldsymbol{\theta}_t) + \mathbf{e}_{k,t},$
where $\mathbf{e}_{k,t}\in \mathbb{R}^P$ denotes the noise term added onto $\mathbf{g}_k(\boldsymbol{\theta}_t)$ at the $t$-th iteration. Without loss of generality, we assume that $\mathbf{e}_{k,t}$ obeys normal distribution, i.e., $\mathbf{e}_{k,t}\sim \mathcal{N}(0,\sigma_{\boldsymbol{\theta}}^2\mathbf{I}_P)$, for $\mathbf{I}_P$ being $P\times P$ identity matrix with variance $\sigma_{\boldsymbol{\theta}}^2$ and the signal-to-noise-ratio (SNR) in gradient transmission is given by $\mathrm{SNR}_{\boldsymbol{\theta}} = 20\log_{10}\frac{||\mathbf{g}_k(\boldsymbol{\theta}_t)||_2^2}{\sigma_{\boldsymbol{\theta}}^2 } $. Once the gradient data from all users are collected, the BS finally incorporates $\tilde{\mathbf{g}}_k(\boldsymbol{\theta}_t)$ for $k = 1,\dots,K$ to update $\boldsymbol{\theta}_t$ as $\boldsymbol{\theta}_{t+1} = \boldsymbol{\theta}_t - \eta_t \frac{1}{K} \sum_{k=1}^{K} \tilde{\mathbf{g}}_k(\boldsymbol{\theta}_t),$
where $ \tilde{\mathbf{g}}_k(\boldsymbol{\theta}_t) = \frac{1}{\textsf{D}_k} \sum_{i=1}^{\textsf{D}_k}\nabla_{\boldsymbol{\theta}} \mathcal{L} (f ( \mathcal{X}_k^{(i)}|\boldsymbol{\theta}_t), $ $\mathcal{Y}_k^{(i)}) ).$
After model aggregation, the BS returns the updated model parameters $\boldsymbol{\theta}_{t+1}$ to the users, and the received model parameter at the $k$-th user can be written as $\tilde{\boldsymbol{\theta}}_{t+1} = \boldsymbol{\theta}_{t+1} + \mathbf{e}_{k,t+1},$
which will be used for the computation of the gradients in the next iteration.
\subsection{Neural Network Architecture}
\label{sec:NNarch}
We design a single CNN trained on the local datasets of the users via FL. The proposed network architecture is comprised of $10$ layers. The first layer is the input layer, which accepts the input data of size $(L + 1)\times \bar{M}\times 3$. The $\{2,4,6\}$-th layers are the convolutional layers with $N_\mathrm{SF} = 128$ filters, each of which employs a $3\times 3$ kernel for 2-D spatial feature extraction. The $\{3,5,7\}$-th layers are the normalization layers. The eighth layer is a fully connected layer with $N_\mathrm{FCL}=1024$ units, whose main purpose is to provide feature mapping. The ninth layer is a dropout layer with $\kappa=1/2$ probability. The dropout layer applies an $N_\mathrm{FCL}\times 1$ mask on the weights of the fully connected layer, whose elements are uniform randomly selected from $\{0,1\}$. As a result, at each iteration of FL training, randomly selected different set of weights in the fully connected layer is updated. Thus, the use of dropout layer reduces the size of $\boldsymbol{\theta}_t$ and $\mathbf{g}_k(\boldsymbol{\theta}_t)$, thereby, reducing model transmission overhead. Finally, the last layer is output regression layer, yielding the output channel estimate of size $2M (L+1)\times 1$. The CNN architecture is similar to the one in~\cite{elbir2020_FL_CE} with the main difference being that the current model accepts $(L + 1)\times \bar{M}\times 3$ input with fewer pilots instead of $(L + 1)\times {M}\times 3$ as in\cite{elbir2020_FL_CE}.
\subsection{Transmission Overhead}
\label{sec:DataComp}
Transmission overhead can be defined as the size of the transmitted data during model training~\cite{elbir2020_FL_CE,FL_gunduz_fading,FL_Gunduz,fl_By_Google}. Let $\mathcal{T}_\mathrm{FL}$ and $\mathcal{T}_\mathrm{CL}$ denote the transmission overhead of FL and CL, respectively. Then, we can define $\mathcal{T}_\mathrm{CL}$ as $ \mathcal{T}_\mathrm{CL} =
\big(3\bar{M}(L + 1) + 2M (L+ 1)\big) \textsf{D},$
which includes the number of symbols in the uplink transmission of the training dataset from the users to the BS. In contrast, the transmission overhead of FL includes the transmission of $\mathbf{g}_k(\boldsymbol{\theta}_t)$ and $\boldsymbol{\theta}_t$ in uplink and downlink communication for $t = 1,\dots,T$, respectively. Compared to~\cite{elbir2020_FL_CE}, the first term on the right hand side of $\mathcal{T}_\mathrm{CL}$ is computed with $\bar{M}$ instead of $M$. Finally, $\mathcal{T}_\mathrm{FL}$ is given by
\begin{align}
\mathcal{T}_\mathrm{FL} =
2PTK.
\end{align} We can see that the dominant terms are $\textsf{D}$ and $P$, which are the number of training data pairs and the number of CNN parameters, respectively. While $\textsf{D}$ can be adjusted according to the amount of available data at the users, $P$ is usually unchanged during model training. Here, $P$ is computed as
\begin{align}
P = \underbrace{N_\mathrm{CL}(CN_\mathrm{SF} W_x W_y)}_{\mathrm{Convolutional\hspace{1pt} Layers}} + \underbrace{\kappa N_\mathrm{SF} W_x W_yN_\mathrm{FCL} \footnotesize }_{\mathrm{Fully\hspace{1pt} Connected\hspace{1pt} Layers}},
\end{align} where $N_\mathrm{CL}=3$ is the number of convolutional layers and $C=3$ is the number of spatial ``channels''. $W_x=W_y=3$ are the 2-D kernel sizes. As a result, we have $P=600,192$.
\section{Numerical Simulations}
\label{sec:Sim}
The goal of the simulations is to compare the performance of the proposed FL-based channel estimation approach with the state-of-the-art model-based channel estimation technique SF-CNN~\cite{deepCNN_ChannelEstimation} together with the MMSE and LS estimation in terms of normalized MSE (NMSE)~\cite{mimoDeepPrecoderDesign}, defined by $\mathrm{NMSE} = \frac{1}{J_T K} \sum_{i=1}^{J_T} \sum_{k=1}^{K}$ $ \frac{\| \boldsymbol{\Sigma}_{k} - \hat{\boldsymbol{\Sigma}}_{k,i} \|_\mathcal{F}^2}{\| \boldsymbol{\Sigma}_{k} \|_\mathcal{F}^2 },$
for $J_T = 100$ number of Monte Carlo trials. We also present the validation RMSE of the training process, for which the validation dataset includes $20\%$ of the whole dataset $\mathcal{D}$.
The local dataset of each user includes $N=100$ different channel realizations for $K=8$ users. The number of antennas at the BS and the IRS are $M=64$ and $L=64$, respectively. Also, we select $N_\mathrm{r}=5$ and the number of pilots is $\bar{M}=32$, unless stated otherwise. The location of each user is selected as $\phi_{k,n}\in \Phi_k$ and $ {\varphi}_{k,n} \in \bar{\Psi}_k$, where $\Phi_k$ and $\bar{\Psi}_k$ are the equally-divided subregions of the angular domain $\Theta$, i.e., $\Theta =\bigcup_{k} \Phi_k = \bigcup_{k}\bar{\Psi}_k$, respectively. The pilot data are generated as ${\mathbf{S}} = \mathbf{I}_{M}$ and $\overline{\mathbf{S}}$ is composed of the $\bar{M}$ columns of $\mathbf{S}$. During training, $N=200$ channel realizations are conducted and we have added AWGN on the input data for three SNR levels, i.e., SNR$=\{20, 25, 30\}$ dB, for $G=160$ realizations in order to provide robust performance against noisy input~\cite{elbir_LIS}. As a result, the number of input-output pairs in the whole training dataset is $\textsf{D}= 3KNG = 3\cdot8\cdot200\cdot160=768,000$. The proposed CNN model is realized and trained in MATLAB on a PC with a $2304$-core GPU. For CL, we use the SGD algorithm with momentum of $0.9$ and the mini-batch size $M_B = 128$, and update the network parameters with learning rate $0.001$. For FL, we train the CNN for $T=100$ iterations/rounds. Once the training is completed, the labels of the validation data (i.e., $20\%$ of the whole dataset) are used in prediction stage. During the prediction stage, each user estimates its own channel by feeding the CNN with $\boldsymbol{\Upsilon}_k$ and obtains $\hat{\mathbf{h}}_{\mathrm{BS},k}$ and $\hat{\mathbf{G}}_k$ at the output.
Let us first compare the transmission overhead of the proposed FL approach with CL. Using the analysis in Section~\ref{sec:DataComp} and the above system settings, we get $\mathcal{T}_\mathrm{CL} = 1.1\times 10^{10}$ whereas $\mathcal{T}_\mathrm{FL}=9.6\times 10^8$, which is approximately $12$ times lower than $\mathcal{T}_\mathrm{CL}$. This demonstrates the effectiveness of the FL over CL in terms of transmission overhead. The complexity of the proposed CNN approach is also lower than that of SF-CNN, for which the detailed complexity analysis is presented in~\cite{elbir2020_FL_CE}. In the sequel, we present the NMSE performance of the proposed approach.
\begin{figure}[t]
\centering
\subfloat[]{\includegraphics[draft=false,width=\columnwidth]{FL_NMSEAll_IRS.eps}\label{fig_SNR} }
\\
\subfloat[]{\includegraphics[draft=false,width=\columnwidth]{CE_Mbar.eps} \label{fig_Number_of_pilots}}
\caption{ Channel estimation NMSE with respect to (a) $\mathrm{SNR}$ when $\bar{M}=32$ and (b) $\bar{M}$ when $\mathrm{SNR} = 20$ dB. }
\label{fig_SNR_IRS}
\end{figure}
Fig.~\ref{fig_SNR} shows the channel estimation NMSE with respect to SNR. Notice that CL provides better performance than that of FL for the proposed CNN model since it has access to the whole dataset at once. Nevertheless, FL has satisfactory channel estimation performance despite decentralized training and outperforms SF-CNN with CL. Specifically, the proposed CNN with FL and CL have similar NMSE for SNR$\leq 25$ dB and the performance of FL maxes out in high SNR regime. This is because the learning model loses precision due to FL training and cannot perform better. SF-CNN also exhibits similar behavior but performs worse than the proposed method. This is because SF-CNN has convolutional-only layers. In contrast, the proposed CNN includes both convolutional and fully connected layers, exhibiting better feature extraction and data mapping performance.
Fig~\ref{fig_Number_of_pilots} shows the channel estimation NMSE with respect to the number of pilot signals $\bar{M}$ for SNR$=20$ dB. We can see that the proposed approach provides significantly better performance in the presence of insufficient pilot signals, i.e., $\bar{M} < M$, whereas the other algorithms perform poorly since they rely on the complete channel data, demanding $\bar{M}\geq M$.
%
\section{Conclusions}
\label{sec:Conc}
We propose an FL framework for channel estimation in IRS-assisted massive MIMO systems. Via simulations, we demonstrate that FL-based channel estimation approach requires approximately $60\%$ fewer less pilot signals as compared to the model-based techniques. Moreover, FL-based training enjoys approximately $12$ times lower transmission overhead than CL-based training while providing a channel estimation performance close to CL. We further observe that at least $15$ dB SNR on the model parameters is required for reliable channel estimation performance, i.e, $\mathrm{NMSE}\leq 0.01$. As a future work, we plan to study the FL-based models for joint channel estimation and IRS reflect beamformer design.
\clearpage
\balance
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
Advances in astrophysical instrumentation in recent decades have allowed detailed observations of internal structures in Galactic star-forming regions. Among them, the observations of small-spatial scale structures in the Orion Nebula have been pioneering. Since the 1990s, with the first images from the Hubble Space Telescope (HST),
dozens of photoevaporating protoplanetary disks \citep[proplyds,][]{Odell1993}
have been discovered and analyzed in the nebula.
Proplyds generally have a compact semi-circular head
with an extended tail that points away from the ionizing source.
Their greatest concentration is in the central core of the Orion Nebula Cluster,
which contains the four high-mass Trapezium stars,
with \th1C{} (O7~Vp; \citealp{Sota:2011a}) being the most luminous member
and the dominant ionizing source for the nebula.
The Trapezium region is illustrated in Figure~\ref{fig:hst}
with an image that has been high-pass filtered to emphasize sub-arcsecond structures
(\(\SI{1}{arcsec} \approx \SI{0.002}{pc}\) for an assumed distance of \SI{410 \pm 10}{pc},
\citealp{Binder2018}).
High-mass Trapezium stars are labelled in blue
and proplyds are labelled in white following the nomenclature
of \citet{Laques:1979a} for the 6 closest proplyds to \th1C{} (LV knots)
and the coordinate-based names of \citet{Odell:1994a} for the remaining.
In many of the proplyds in the surrounding areas of \th1C,
supersonic gas outflows and jets have been observed.
Several of them have been classified as Herbig-Haro objects (HHs),
some of which are illustrated in Figure~\ref{fig:hst}.
This is the case of HH~514, which originates in the proplyd 170-337 \citep[][]{Bally:1998a},
located \SI{12.61}{arcsec} from \th1C.
A redshifted microjet of size \SI{0.5}{arcsec} protrudes northwards from the proplyd ionization front
and a short chain of redshifted knots
are seen at a distance of \(\approx \SI{4}{arcsec}\)
along the same axis \citep{bally00}
with proper motions consistent with an origin in 170-337 \citep{odellyhenney08}.
These are respectively labelled as ``jet base'' and ``knot'' in Fig.~\ref{fig:hst}.
A blueshifted counterflow is also detected at the base of the jet,
but it is much weaker.
This is unusual for HH flows in Orion,
which tend to be predominantly blueshifted \citep[][]{Henney:2007b}.
A faint fast-moving knot to the S of the proplyd may be associated with this counterjet.
HH~514 is particularly bright in the emission of [Fe\thinspace III] lines,
showing also intense lines of [Fe\thinspace II], [Ni\thinspace II] and [Ni\thinspace III].
In addition, its proximity to $\theta^1$ Ori C and its large propagation velocity
make it an excellent laboratory to explore the phenomenon of destruction of dust grains
by analyzing the abundances of Fe and Ni in the gas.
Many HH objects have relatively large gaseous abundances of Fe and Ni,
whereas in the rest of the Orion Nebula these abundances are much smaller \citep[][]{mendez2021-2}.
In this paper, the third in a series dedicated to photoionized HH objects in the Orion Nebula,
we analyze for the first time the physical conditions, chemical composition and kinematics of HH~514.
We use high-resolution spectroscopy from the Ultraviolet and Visual Echelle Spectrograph (UVES) \citep[][]{Dodorico00}
of the Very Large Telescope
and IFU spectra from MEGARA \citep{gildepaz+18} at the Gran Telescopio Canarias (GTC).
Given the high spectral and spatial resolution of our observations,
we have analyzed two velocity components associated with two areas of HH~514:
the jet base and a northern knot, completely separated from the emission of the Orion Nebula.
Our analysis focuses mainly on estimating the total abundances of Fe and Ni,
which seems to be mainly in gaseous phase due to the strong degree of destruction of dust grains in the shock front.
We also report an overabundance of S and discuss its possible origin.
In Sec.~\ref{sec:data} we describe the observations and the data reduction process.
In Sec.~\ref{sec:physical_cond} we derive the physical conditions of each kinematic component.
In Sec.~\ref{sec:ionic_total_abundances} we derive the ionic and total abundances.
In Sec.~\ref{sec:disc} we discuss our results, focusing on the observed overabundance of S,
the total abundances of Fe and Ni and the kinematic properties of HH~514.
In Sec.~\ref{sec:summary_and_conclusions} we summarize our main conclusions.
Finally in the Appendix~\ref{sec:apendix_A} we add tables of data and figures as supporting material.
\begin{figure*}
\includegraphics[width=\textwidth]{hh514-finding}
\caption{
Location of HH~514 within the inner Orion Nebula,
showing the position and orientation of the UVES spectrograph slit
(dashed green outline).
High mass OB stars are labeled in blue,
prominent proplyds are labeled in white,
and components of Herbig-Haro flows
from \citet{Odell15}
are labeled in yellow.
Background image is a composite of narrow-band observations
with \textit{HST} in the [\ion{O}{3}] \(\lambda 5007\) line
obtained at 3 different epochs over a span of 22 years:
1994 (red channel),
2004 (green channel),
2016 (blue channel).
All images are aligned to the astrometric frame defined in
\citet{Robberto:2013a}
and have been high-pass filtered with a Gaussian kernel of
width 1~arcsec to suppress large-scale brightness variations.
Stationary emission features appear yellow-white in this image,
while moving features show a prism-like separation:
red-green-blue, with blue indicating the leading edge.
}
\label{fig:hst}
\end{figure*}
\section{Observations and data reduction}
\label{sec:data}
\subsection{High spectral resolution observations}
\label{sec:data_uves}
The observations were taken during the nights of 29 and 30 October, 2013, under clear conditions using the UT2 (Kueyen) of the Very Large Telescope (VLT) at Cerro Paranal, Chile, with the Ultraviolet Echelle Spectrograph \citep[UVES;][]{Dodorico00}. The central coordinates of the 10 arcsec slit are: RA(J2000)=05$^h$35$^m$16$^s$.95, DEC(J2000)=$-$05$^{\circ}$23$'$33.72$''$ oriented along the north-south spatial axis as is shown in Fig.~\ref{fig:hst}. The slit width of 1\arcsec provides an effective spectral resolution of $\lambda/\Delta \lambda \approx 6.5 \text{ km s}^{-1}$ in the spectral range between 3100 and 10400 \AA. Analogously to the observations described in the articles on HH~529~II-III and HH~204 \citep[][hereinafter Paper~I and Paper~II, respectively]{mendez2021,mendez2021-2}, three exposures of 150s each of the star GD71 \citep{Moehler14a, Moehler14b} were taken during the same night under similar conditions to achieve the flux calibration of the science data. The configuration of the instrument and the data reduction procedure are described in detail in Paper~I while the main settings of the observations of HH~514 are shown in Table~\ref{tab:obs_set}. We define three spatial cuts along the observed slit as shown in Fig.~\ref{fig:cuts}, covering the main high-velocity components,
which are redshifted by about $150\text{ km s}^{-1}$ from the nebular component.
We name the southernmost high-velocity component (within Cut~1) as the jet and the one located in the central cut as the knot. The nebular emission of Cut~1 also contains the emission of the proplyd 170-337. The complete analysis of the physical conditions and chemical composition of this proplyd will be the subject of a future work (M\'endez-Delgado et al., in prep). In this work we will only mention some results on the abundance of S in 170-337 in Sec.~\ref{subsec:real_overabundance}, which is relevant for our analysis of HH~514. Panel (b) of Fig.~\ref{fig:cuts} shows several components at intermediate redshifted velocities between that of the nebula and HH~514, which are seen mainly in lines of highly ionized ions such as [O\thinspace III] or [Ne\thinspace III]. Unfortunately, their emission is very weak in most spectral lines, so the determination of physical conditions and chemical abundances for those velocity components is not reliable with the present data and they are not studied further here.
They may be associated with the HH~518 flow from the proplyd 159-350
(see Figure~\ref{fig:hst}).
In order to increase the contrast of the HH~514 emission, which is rather weak compared to the main nebular background emission, we subtract Cut~3, which contains only background nebular emission, from Cut~1 and Cut~2, re-scaling its emission by the number of pixels of each cut. In addition to increasing the contrast of the HH~514 emission, this subtraction procedure permits us to remove sky and ghost contamination from the spectra of the HH object. We estimate the flux of the emission lines, their FWHM and their central wavelength by gaussian fittings using the SPLOT task of IRAF\footnote{IRAF is distributed by National Optical Astronomy Observatory, which is operated by Association of Universities for Research in Astronomy, under cooperative agreement with the National Science Foundation} \citep{Tody93} as is described in detail in Paper~I. We achieve the reddening correction by using the same procedure described in Paper~I and Paper~II, using the extinction curve from \citet{Blagrave07}. The resulting values of $\text{c}(\text{H}\beta)$, are presented in Table~\ref{tab:c_extin}. In Table~\ref{tab:sample_spectra} we include the relevant information for some of the emission lines of the spectrum of the jet base, as an example of the tables of the complete spectra that are provided as supplementary material.
\begin{table}
\caption{Main parameters of UVES spectroscopic observations.}
\label{tab:obs_set}
\begin{tabular}{ccccc}
\hline
Date & $\Delta \lambda$& Exp. time &Seeing &Airmass\\
& (\AA) & (s) & (arcsec)&\\
\hline
2013-10-30 & 3100-3885 & 5, 3$\times$180 &0.92&1.11\\
2013-10-30 & 3750-4995 & 5, 3$\times$600 & 0.87 & 1.15\\
2013-10-30 & 4785-6805 & 5, 3$\times$180 &0.92&1.11\\
2013-10-30 & 6700-10420 & 5, 3$\times$600 & 0.87 & 1.15\\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\begin{minipage}{6cm}
\centering\includegraphics[height=2cm,width=\columnwidth]{2D_3729.pdf}
\centerline{(a) [O\thinspace II] $\lambda 3729$.}
\smallskip
\end{minipage}
\begin{minipage}{6cm}
\centering\includegraphics[height=2cm,width=\columnwidth]{2D_4959.pdf}
\centerline{(b) [O\thinspace III] $\lambda 4959$.}
\smallskip
\end{minipage}
\begin{minipage}{6cm}
\centering\includegraphics[height=2cm , width=\columnwidth]{2D_6300.pdf}
\centerline{(c) [O\thinspace I] $\lambda 6300$.}
\smallskip
\end{minipage}
\begin{minipage}{6cm}
\centering\includegraphics[height=2cm , width=\columnwidth]{2D_4649.pdf}
\centerline{(d) O\thinspace II $\lambda 4649$.}
\smallskip
\end{minipage}
\begin{minipage}{12.0cm}
\centering\includegraphics[height=4cm, width=\columnwidth]{2D_4658_cuts_named.pdf}
\centerline{(e) [Fe\thinspace III] $\lambda 4658$.}
\end{minipage}
\caption{\textit{Upper and middle panels:} Sample of representative lines in the bi-dimensional spectrum. The Y axis corresponds to the spatial direction (up north, down south, see Fig.~\ref{fig:hst} for the spatial location of the slit) while the X axis is the spectral direction. All figures are centred at the rest-frame reference wavelength of each line. \textit{Bottom panel:} Emission of the [Fe\thinspace III] $\lambda 4658.17$ line as well as the limits and extension of the different spatial cuts extracted to analyse each velocity component. Cut 1 is at the bottom, which corresponds to the southernmost one. The spatial extension is 2.71 arcsec, 4.18 arcsec and 2.46 arcsec for cuts 1, 2 and 3, respectively. The velocity scale is heliocentric.}
\label{fig:cuts}
\end{figure*}
\subsection{IFU observations}
\label{sec:data_megara}
Optical integral-field spectrograph observations of HH~514 were carried out on the night of 28 February 2022 under dark, photometric sky conditions and a seeing of $\sim$1.2 arcsec using MEGARA \citep{gildepaz+18} at the 10.4 m Gran Telescopio Canarias (GTC). The low-resolution (LR) Volume-Phased Holographic (VPH) grism VPH675-LR was used; this grism covers the 6095--7300 \AA\ wavelength range with a spectral dispersion of 0.287 \AA\ pix$^{-1}$ and an effective resolution of R$\sim$5900. Three 900 s exposures were taken in order to minimize the impact of cosmic-rays in the relatively large field of view of MEGARA (12$''$.5 $\times$ 11$''$.3).
The MEGARA data were reduced using the {\it megaradrp} v0.11 pipeline \citep{pascual+19, pascual+20}. The pipeline uses several python-based recipes to perform bias substraction, fibre-tracing, flat-field correction, wavelength calibration, and spectra extraction. Sky subtraction was avoided owing to the large size of the Orion nebula prevented to place any of the MEGARA sky bundles on a nebular emission-free zone of the sky. Flux calibration was also performed using the spectrophotometric standard star HR\,1544. The final product is a row-stacked spectrum that is converted into a datacube of 0$''$.2 square spaxel on spatial dimensions using the {\it megararss2cube} task of the {\it megara-tools} suite v0.1.1 \citep{gildepaz+20}.
\begin{table}
\caption{Reddening coefficients for each component.}
\label{tab:c_extin}
\begin{tabular}{lccccc}
\hline
& \multicolumn{2}{c}{$\text{c}(\text{H}\beta)$} \\
& Nebular & High velocity\\
\hline
Cut 1 & - & $0.86 \pm 0.05$ \\
Cut 2 & $0.85 \pm 0.03$ &$0.75 \pm 0.07$\\
Cut 3 & $0.85 \pm 0.02$&-\\
\hline
\end{tabular}
\end{table}
\section{Physical Conditions}
\label{sec:physical_cond}
We use the version 1.1.13 of PyNeb \citep{Luridiana15} and the atomic data set shown in Tables 9 and 10 from Paper~II to derive the physical conditions of the different gas components analyzed in this work. The adopted electron density, $n_{\rm e}$, in the nebular components is a weighted average\footnote{The weights were defined as the inverse of the square of the error associated to each density diagnostic.} of the resulting values from the CEL diagnostics
[O\thinspace II] $\lambda3726/\lambda3729$, [S\thinspace II] $\lambda6731/\lambda6716$, [Cl\thinspace III] $\lambda5538/\lambda5518$, [Fe\thinspace II] $\lambda9268/\lambda9052$, [Fe\thinspace III] $\lambda4658/\lambda4702$ and [Ar\thinspace IV] $\lambda4740/\lambda4711$. In addition, we apply the maximum-likelihood procedure described in Sec.~4.2 of Paper~I to derive the physical conditions based on the [Fe\thinspace III] $\lambda \lambda $ 4658, 4702, 4734, 4881, 5011 and 5271 lines. As it was commented in Paper~I, with the adopted atomic data, [Fe\thinspace III] diagnostics are not adequately sensitive to density for values smaller than $10^{3} \text{ cm}^{-3}$, thus if a considerable part of the observed ionized gas has densities of the order or lower than that value (as it is expected in the nebular components), the resulting $n_{\rm e}$ may be biased to that of the densest zones along the line of sight. The value of $n_{\rm e}(\text{O\thinspace II})$ was derived with the available RLs of multiplet 1 and the results are highly consistent with the adopted average density.
In the plasma diagnostics of the jet base, there are two possible intersections of the temperature sensitive curve of [N\thinspace II] $\lambda5755/\lambda6584$ with the density sensitive ones, giving very different values of the electron temperature, $T_{\rm e}$. One gives an extremely high density (larger than 10$^5$ cm$^{-3}$) (shown in Fig.~\ref{fig:plasma}) or an extremely high temperature (higher than 10$^{5}$ K at smaller densities than 10$^{4}$ cm$^{-3}$, beyond the scale of Fig.~\ref{fig:plasma}). The [Fe\thinspace III] $\lambda4658/\lambda4702$, [S\thinspace II] $\lambda$4069/$\lambda$4076 and [O\thinspace II] $\lambda$7319+20/$\lambda$7330+31 density diagnostics confirm the first case. In the two studied components of HH~514 we estimate the convergence of $n_{\rm e}-T_{\rm e}$ by using the [Fe\thinspace III] $\lambda \lambda$ 4658, 4702, 4734, 4881, 5011, 5271 CELs. The resulting density in the jet base of HH~514 is $10^{5.37} \text{ cm}^{-3}$.
In the case of the northern knot of HH~514, the convergence shows a higher dispersion in $n_{\rm e}-T_{\rm e}$ which seems to be due to the mixture of two or more gas components with similar velocities and notably different physical and/or ionization conditions. As it is shown in Fig.~\ref{fig:spatial_dis}, the spatial distribution along the slit of the intensity of the [Fe\thinspace III] $\lambda 4658$ line centred at a heliocentric velocity of $\sim 150 \text{ km s}^{-1}$ is double-peaked in the northern knot of HH~514, which shows that at least two different gas components moving at a similar velocity are integrated in this part of the HH object. However, the separation into two spatial components results in a very low signal-to-noise spectra. Thus, the physical conditions derived in the knot are expected to be more uncertain than those in the jet base.
Based on the adopted density in each component, we estimate $T_{\rm e}$ through various diagnostics shown in Table~\ref{tab:pc}. In the nebular components, we define $T_{\rm e} \text{ (low)}$ as the weighted average of $T_{\rm e} (\text{[N\thinspace II]})$, $T_{\rm e} (\text{[O\thinspace II]})$ and $T_{\rm e} (\text{[S\thinspace II]})$ while $T_{\rm e} \text{ (high)}$ is the weighted average of $T_{\rm e} (\text{[O\thinspace III]})$, $T_{\rm e} (\text{[S\thinspace III]})$ and $T_{\rm e} (\text{[Ar\thinspace III]})$. In both components of HH~514, the estimation of $T_{\rm e} \text{ (low)}$ is quite uncertain, given the low critical density of the diagnoses available for this ionization zone and the high density of these objects. Due to this, and considering the similarity between $T_{\rm e}(\text{[O\thinspace III]})$ in both velocity components, we will adopt the $T_{\rm e} \text{ (low)}$ determined for the jet base --the spectrum with the best signal to noise ratio-- as also representative for the knot.
\begin{table*}
\centering
\caption{ Electron densities and temperatures determined from several diagnostics.}
\label{tab:pc}
\begin{tabular}{ccccc}
\hline
& \multicolumn{1}{c}{Cut 1} & \multicolumn{2}{c}{Cut 2} & \multicolumn{1}{c}{Cut 3} \\
Diagnostic & HH~514~jet base & Nebula & HH~514~knot & Nebula\\
\hline
& \multicolumn{4}{c}{$n_e$(cm$^{\text{-}3}$)}\\
[O\thinspace II] $\lambda$3726/$\lambda$3729 & - &$5980^{+1160} _{-870}$& - & $5610^{+960} _{-800}$\\
[O\thinspace II] $\lambda$7319+20/$\lambda$7330+31 & 263000: & - & - &- \\
[S\thinspace II] $\lambda$6731/$\lambda$6716 & - & $4150^{+1810} _{-1230}$& -& $4240^{+1400} _{-1030}$\\
[S\thinspace II] $\lambda$4069/$\lambda$4076 & 219000: & - &-&-\\
[Cl\thinspace III] $\lambda$5538/$\lambda$5518 & - & $7640^{+970} _{-1020}$& -& $7620^{+1070} _{-960}$\\
[Fe\thinspace II] $\lambda$9268/$\lambda$9052 & >76000 & $4500^{+9740} _{-3590}$ & -& $3540^{+9190} _{-2760}$ \\
[Fe\thinspace III] $\lambda$4658/$\lambda$4702 & 234000: & $8490^{+2490} _{-2350}$& 57000: & $6430^{+2810} _{-2080}$\\
[Ar\thinspace IV] $\lambda$4740/$\lambda$4711 & - &$6730^{+660} _{-690}$& - & $4800^{+570} _{-690}$\\
O\thinspace II$^{*}$ & - &$5640 \pm 510$&-&$4960 \pm 650$ \\
[Fe\thinspace III]$^{*}$ & $232000 \pm 12000$ & $8600 \pm 810$ & $74000 \pm 14000$& $7540\pm 1020$\\
\textbf{Adopted} & \boldmath${232000 \pm 12000}$ & \boldmath${6690 \pm 940}$& \boldmath${74000 \pm 14000}$& \boldmath${5490 \pm 1120 }$\\
& \multicolumn{4}{c}{$T_e$ (K)}\\
T$\left(\mbox{He}\thinspace \mbox{I} \right)$ & $8710:$&$7790 ^{+570} _{-480}$ & $6010:$&$7570 ^{+490} _{-550}$\\
[N\thinspace II] $\lambda$5755/$\lambda$6584 & $12860^{+1240} _{-1170}$& $9750^{+230} _{-220}$&- &$9880^{+230} _{-250}$\\
[O\thinspace II] $\lambda \lambda$ 3726+29/$\lambda \lambda$7319+20+30+31 & $9180^{+1500} _{-630}$& $9500^{+740} _{-610}$& $7830^{+1430} _{-810}$&$10980^{+1040} _{-1320}$\\
[S\thinspace II] $\lambda \lambda$4069+76/$\lambda \lambda$ 6716+31& $20010^{+45030} _{-9150}$& $9410^{+950} _{-1220}$&-&$10090^{+2180} _{-1400}$\\
[O\thinspace III] $\lambda$4363/$\lambda \lambda$4959+5007& $8880^{+250} _{-210}$& $8490^{+90} _{-100}$ & $9330^{+260} _{-230}$& $8400^{+80} _{-70}$\\
[S\thinspace III] $\lambda$6312/$\lambda \lambda$9069+9531 & $8420^{+310} _{-400}$& $8630^{+330} _{-340}$& $8230^{+430} _{-560}$ & $8820^{+330} _{-310}$\\
[Ar\thinspace III] $\lambda$5192/$\lambda$7136 & - & $8070^{+190} _{-160}$&-&$8320^{+230} _{-190}$\\
[Fe\thinspace III]$^{*}$ & $6840 \pm 1640$ & $8960 \pm 850$ & $8180 \pm 1410$& $7850 \pm 830$\\
\textbf{\boldmath${T_e}$ (low) Adopted} & \boldmath${11370 \pm 1820}$ & \boldmath${9530 \pm 100}$& \boldmath${11370 \pm 1820}$ & \boldmath${9920 \pm 200}$\\
\textbf{\boldmath${T_e}$ (high) Adopted} & \boldmath${8880 \pm 250}$ & \boldmath${8430 \pm 150}$ & \boldmath${9330 \pm 260}$ & \boldmath${8410 \pm 100}$\\
\hline
\end{tabular}
\begin{description}
\item $^*$ A maximum likelihood method was used. \\
\end{description}
\end{table*}
\section{Ionic and Total Abundances}
\label{sec:ionic_total_abundances}
We derive ionic abundances of N$^{+}$, O$^{+}$, S$^{+}$, Cl$^{+}$, Ca$^{+}$, Fe$^{+}$ and Ni$^{+}$ based on CELs and using $T_{\rm e} \text{ (low)}$. We derive the abundances of Cl$^{2+}$ and S$^{2+}$ using $T_{\rm e} \text{ ([S\thinspace III])}$ and the ionic abundances of O$^{2+}$, Ne$^{2+}$, Cl$^{3+}$, Ar$^{2+}$, Ar$^{3+}$ and Fe$^{3+}$ using $T_{\rm e} \text{ (high)}$. In the nebular components, we derive the abundance of Fe$^{2+}$ and Ni$^{2+}$ using $T_{\rm e} \text{ (low)}$. However, in both components of HH~514, the derived $T_{\rm e} \text{ ([Fe\thinspace III])}$ indicate a lower temperature than $T_{\rm e} \text{ (low)}$. This can be related to the dust destruction process described in detail in Sec.~\ref{subsec:real_overabundance}, where the filtering and destruction of dust grains increase the relative abundance of metals, cooling the gas. Therefore, in HH~514 we derive the abundances of Fe$^{2+}$ and Ni$^{2+}$ using $T_{\rm e} \text{ (high)}$. In the case of the abundance of Ca$^+$, we must take into account the possible coexistence of Ca$^+$ and H$^0$, due to its low ionization potential. Therefore, the derived Ca$^+$/H$^+$ may be an upper limit to the real abundance.
As discussed in Paper~II, some [Fe\thinspace II] and [Ni\thinspace II] lines can be produced mostly by continuum pumping instead of collisional excitation. We make sure that the estimation of the Fe$^+$ abundance is not affected by fluorescence by using lines of the lower a$^4$F--a$^4$P levels ($\lambda \lambda$ 7155, 8892, 9052, 9267) mostly populated by collisions \citep{Baldwin96}. In the case of Ni$^+$ the fluorescence effects may be larger, considering that the levels involved in the transitions that give rise to the observed optical lines have the same multiplicity. However, if the density is high enough, collisional excitations may dominate the production of [Ni\thinspace II] lines. Following the same procedure described in equation 1 from Paper~II --based on the equation 8 from \citet{Bautista96}--, we estimate a critical density of $n_{\text{cf}}= 55860 \text{ cm}^{-3}$, from which collisions dominate over fluorescence in the production of [Ni\thinspace II] $\lambda 7378$ line at the apparent distance of 12.61 arcsecs from $\theta^1$ Ori C. This implies that in both the jet base and the northern knot --regions with densities larger than $n_{\text{cf}}$--, we can directly calculate the abundance of Ni$^{+}$ based on the intensity of [Ni\thinspace II] $\lambda 7378$ line. However, this is not possible in the nebular components, where --due to the much lower density-- fluorescence should dominate.
In the UVES spectrum of the jet base of HH~514, we are able to estimate an upper limit for the abundance of Fe$^{3+}$ based on the noise level of the continuum at $\sim 6743.30$\AA, where we would expect the emission of [Fe\thinspace IV] $\lambda6740$. We carefully estimated this upper limit, taking into account that at 6742.91\AA~ there may be a relatively weak telluric emission line \citep[][]{Hanuschik03}. However, since Cut~3 was subtracted from Cut~1, --rescaling to the same number of pixels-- to increase the contrast between the nebular emission and the HH~514~jet base spectrum, any contamination by telluric lines has been eliminated. In this case we obtain an upper limit of 12+log(Fe$^{3+}$/H$^+$) $<$ 6.99.
To better constraint the flux of the [Fe\thinspace IV] $\lambda6740$ line in the jet base component of HH~514, we search into the MEGARA deep spectra cube (see Sect.~\ref{sec:data_megara}), analyzing the spatial area that comprises the proplyd 170-337 (which includes the emission of the jet base), subtracting the nebular emission from a surrounding area to optimize the contrast of the high-velocity emission -- as is indicated in Fig.~\ref{fig:megara} -- as well as remove weak sky features. The resulting 1D-spectrum is shown in Fig.~\ref{fig:megara_vs_UVES}, having a higher SNR than the UVES spectra, although with lower spectral resolution. This spectrum allowed us to restrict the upper limit the Fe$^{3+}$ abundance to 12+log(Fe$^{3+}$/H$^+$) $<$ 6.79.
\begin{figure}
\includegraphics[width=\columnwidth]{MEGARA_IFU}
\caption{ Location of the pointing of MEGARA-IFU observations. The field of view of MEGARA is 12$''$.5 $\times$ 11$''$.3 and the observations were centered in proplyd 170-337, where the high-velocity component corresponding to the jet base arise. White rectangles comprise the extracted spaxels to analyse the emission from proplyd 170-337 and those used for the adopted nebular background. The background HST image of the Orion Nebula is a RGB composite \citep[Red: ACS F658N+F775W+F850LP; Green: ACS F555W; Blue: ACS F435W,][]{Robberto:2013a}.}
\label{fig:megara}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{megara_UVES}
\caption{Comparison between MEGARA and UVES one-dimensional spectra of the jet base of HH~514 between 6737-6765\AA. The flux units are normalized. In both cases the nebular background was subtracted. The expected location of the [Fe\thinspace IV] $\lambda 6740$ emission is indicated with a dashed vertical line. The depth of the MEGARA spectrum allows us to observe [Fe\thinspace II] $\lambda 6747$ as well as estimate a tighter upper limit to the [Fe\thinspace IV] $\lambda 6740$ emission. }
\label{fig:megara_vs_UVES}
\end{figure}
In the nebular spectra, we estimate the He$^{+}$, C$^{2+}$, O$^{+}$, O$^{2+}$ and Ne$^{2+}$ abundances making use of RLs following the same procedure described in Paper~I. In the case of the two components of HH~514 analysed in this work, we can only determine the abundances of He$^{+}$ from RLs. In this case, the adopted He$^{+}$ abundance is an average of the resulting values based on the available He\thinspace I singlet lines and triplet ones less affected by self-absorption effects (See Table D14 from Paper~I). The results are shown in Table~\ref{tab:ionic_abundances}.
\begin{table*}
\centering
\caption{Ionic abundances derived in each kinematic component in logarithmic unit with $n(\text{H})=12$.}
\label{tab:ionic_abundances}
\begin{tabular}{ccccccccccccc}
\hline
& \multicolumn{1}{c}{Cut 1} & \multicolumn{2}{c}{Cut 2} & \multicolumn{1}{c}{Cut 3} & \\
Ion & HH~514~jet base & Nebula & HH~514~knot & Nebula \\
\hline
& \multicolumn{4}{c}{CELs}\\
N$^{+}$ & $6.88^{+0.28} _{-0.13}$ &$6.89 \pm 0.02 $ & $6.23^{+0.27} _{-0.14}$& $6.84 \pm 0.03 $\\
O$^{+}$ & $7.93^{+0.51} _{-0.18}$ & $7.81 \pm 0.03 $&$7.40^{+0.49} _{-0.19}$ & $7.69 \pm 0.05 $\\
O$^{2+}$ & $8.30^{+0.05} _{-0.04}$& $8.37 \pm 0.03$& $8.45^{+0.05} _{-0.04}$ &$8.38 \pm 0.02 $\\
Ne$^{2+}$ & $7.56^{+0.06} _{-0.05}$& $7.84^{+0.04} _{-0.03}$ & $7.79^{+0.06} _{-0.05}$&$7.82^{+0.03} _{-0.02}$\\
S$^{+}$ & $6.11^{+0.23} _{-0.12}$ & $5.50 \pm 0.04 $& $5.34^{+0.26} _{-0.14}$&$5.45 \pm 0.05 $\\
S$^{2+}$ & $7.39^{+0.07} _{-0.06}$ &$6.86^{+0.05} _{-0.04}$ & $7.35^{+0.09} _{-0.07}$&$6.84^{+0.05} _{-0.04}$\\
Cl$^{+}$ & - &$3.63 \pm 0.04 $&- &$3.60 \pm 0.04 $\\
Cl$^{2+}$ &- &$4.98^{+0.06} _{-0.05}$&- &$4.99^{+0.06} _{-0.05}$\\
Cl$^{3+}$ & - &$3.79 \pm 0.03 $&- &$3.92 \pm 0.03 $\\
Ar$^{2+}$ & $6.29 \pm 0.04 $& $6.31 \pm 0.03 $& $6.19 \pm 0.04 $&$6.31 \pm 0.02 $\\
Ar$^{3+}$ & - &$5.09^{+0.04} _{-0.03}$& -&$5.15^{+0.03} _{-0.02}$\\
Ca$^{+}$ & $3.71^{+0.20} _{-0.12}$ &- &-&-\\
Fe$^{+}$ & $6.14 \pm 0.05$ & $4.39 \pm 0.05$ & - &$4.32 \pm 0.02$\\
Fe$^{2+}$ & $7.14 \pm 0.02$ & $5.28 \pm 0.01$& $6.83 \pm 0.04$&$5.17 \pm 0.02$\\
Fe$^{3+}$ & < 6.79 &$5.62^{+0.10} _{-0.08}$& -&$5.78 \pm 0.12 $\\
Ni$^{+}$ & $4.94^{+0.15} _{-0.09}$ & - &-&-\\
Ni$^{2+}$ &$5.84^{+0.07} _{-0.06}$ & $4.32 \pm 0.03 $ &$5.62^{+0.08} _{-0.07}$&$4.25 \pm 0.04 $\\
& \multicolumn{4}{c}{RLs}\\
He$^{+}$ &$10.95 \pm 0.04$ & $10.92 \pm 0.02$& $10.96 \pm 0.02$&$10.92 \pm 0.02$\\
C$^{2+}$ & -& $8.37 \pm 0.01 $ &-& $8.37 \pm 0.01$ \\
O$^{+}$ & - & $8.17^{+0.04} _{-0.03}$& -& $8.19 \pm 0.05 $ \\
O$^{2+}$ & - & $8.61 \pm 0.05$ & - & $8.63 \pm 0.05$ \\
Ne$^{2+}$ & -&$8.04 \pm 0.06 $&-& $8.10 \pm 0.08 $ \\
\hline
\end{tabular}
\end{table*}
To estimate the total abundances of C, N, Ne, S and Ar, we use the ionization correction factors (ICFs) from \citet{Amayo2021}. These ICFs, based on photoionization models of giant H~II regions, are consistent with the schemes shown in table~10 of Paper~I. However, they additionally provide the 1-$\sigma$ limits of their predictions, allowing the formal propagation of uncertainties associated to the ICFs. In our nebular spectra, the calculation of the total abundances of O, Cl and Fe did not require ICF and were estimated from the sum of their ionic abundances. In the case of Ar, Ar$^+$ is estimated to contribute only 1 or 2 per cent of the total Ar abundance, having a negligible impact. We estimate the total abundance of Ni as $\text{Fe/Ni}=\text{Fe}^{2+}/\text{Ni}^{2+}$ based on their similar depletion and ionization patterns \citep[][]{mendez2021-2}. In the case of the He abundance in the nebular components, we use the ICF of \citet{kunthsargent83}. Since we have a reliable determination of the abundance of O$^+$ and O$^{2+}$ both with CELs and RLs, we derive the total abundances of He, Ne and C using the degree of ionization estimated from the abundances of the two ionic species of O determined from RLs.
In the case of the jet base of HH~514, we determine the total abundance of Fe as Fe/H = Fe$^{+}$/H$^{+}$+Fe$^{2+}$/H$^{+}$+Fe$^{3+}$/H$^{+}$, summing the estimated upper limit of 12+log(Fe$^{3+}$/H$^{+}$) = 6.79. At first approximation, the total abundance of Fe should be between $7.18 \pm 0.02$ and $7.33 \pm 0.02$, where the value of 12+log(Fe$^{3+}$/H$^{+}$) takes values between 0 and 6.79. \citet{rodriguez05} provides two ICF(Fe) schemes, where the true Fe abundance should be between the limits predicted by both schemes, based on the observed O$^{+}$/O$^{2+}$ and Fe$^{2+}$/O$^{+}$ ratios. In our case, both ICFs match in the value of 7.50. Therefore, considering the observational uncertainties in the abundance of O$^{+}$, this suggest that the upper limit of the value of Fe$^{3+}$/H$^{+}$ should be close to its expected true value and therefore 12+log(Fe/H)$\approx 7.33 \pm 0.02$.
It is remarkable that log(Ni$^{+}$/Fe$^{+}$) = $-1.20 \pm 0.16$, while log(Ni$^{2+}$/Fe$^{2+}$) = $-1.30 \pm 0.07$, which confirms the conclusions of \citet{mendez2021-2} about the similar ionization and depletion patterns of Ni and Fe. Furthermore, this demonstrates that the effects of fluorescence on our Ni$^{+}$ abundance determination are negligible, since we would expect stronger continuum pumping effects for [Ni\thinspace II] than [Fe\thinspace II]. The estimated total abundances are shown in Table~\ref{tab:total_abundances}.
\begin{table*}
\centering
\caption{Total abundances. The units are logarithmic with $n(\text{H})=12$.}
\label{tab:total_abundances}
\begin{tabular}{ccccccccccccc}
\hline
& \multicolumn{1}{c}{Cut 1} & \multicolumn{2}{c}{Cut 2} & \multicolumn{1}{c}{Cut 3} & \\
Ion & HH~514~jet base & Nebula & HH~514~knot & Nebula \\
\hline
& \multicolumn{4}{c}{CELs}\\
N & $7.43 ^{+0.59} _{-0.32}$ & $7.64 \pm 0.16$ &$7.36 ^{+0.58} _{-0.31}$ & $7.69 ^{+0.18} _{-0.16}$ \\
O & $8.45 \pm 0.16$ & $8.48 \pm 0.02$ &$8.49 \pm 0.06$&$8.46 \pm 0.02$\\
Ne & $7.75 ^{+0.19} _{-0.15}$ & $7.97 ^{+0.11} _{-0.10}$& $7.84 ^{+0.11} _{-0.09}$ & $7.92 ^{+0.09} _{-0.08}$ \\
S & $7.43 ^{+0.16} _{-0.12}$ & $6.92 ^{+0.07} _{-0.06}$& $7.49 ^{+0.17} _{-0.14}$ &$6.91 ^{+0.07} _{-0.06}$\\
Cl & - & $5.03 \pm 0.05$&- &$5.04 \pm 0.04$\\
Ar & $6.32 ^{+0.12} _{-0.10}$ & $6.34 ^{+0.06} _{-0.05}$ & $6.29 ^{+0.14} _{-0.12}$& $6.34 \pm 0.04$\\
Fe & $7.33 \pm 0.02$ & $5.80 \pm 0.07$ & - &$5.89 \pm 0.09$\\
Ni & $6.03 \pm 0.08$ & $4.83 \pm 0.08 $&-&$4.96 \pm 0.10$\\
& \multicolumn{4}{c}{RLs}\\
He &$10.95 \pm 0.04$ & $10.95 \pm 0.02$ & $10.96 \pm 0.02$ & $10.95 \pm 0.02$\\
C & - &$8.45 ^{+0.07} _{-0.06}$&-&$8.45 ^{+0.07} _{-0.06}$\\
O & - & $8.74 \pm 0.04$ &-&$8.76 \pm 0.04$\\
Ne & - &$8.17 ^{+0.13} _{-0.11}$&-&$8.26 ^{+0.18} _{-0.14}$\\
\hline
\end{tabular}
\end{table*}
\section{Discussion}
\label{sec:disc}
\subsection{The star/disk/outflow system of 170-337/HH~514}
\label{sec:general-properties}
In this section we combine results from the literature with those of this study
to provide a comprehensive view of the multiple components
that make up the proplyd 170-337 and its outflows.
\subsubsection{The central star and its disk}
\label{sec:central-star-disk}
The central star of 170-337 is of spectral type M2e \citep{Hillenbrand:1997a}
and its mass and age have been determined from fitting to
pre-main-sequence evolutionary tracks,
yielding \SI{0.3}{\msun} and \SI{0.6}{Myr} respectively \citep{Boyden:2020a}.
It was one of the first Orion proplyds to be detected
via the mm-wave dust emission of its protostellar disk \citep{Williams:2005a},
with an estimated dust mass of \SI{23 \pm 5}{\mearth} \citep{Eisner:2018a}.
The disk has also been detected in the CO 3--2 line \citep{Boyden:2020a},
yielding a disk radius of \SI{44 \pm 4}{AU},
which is roughly half of the radius of the proplyd ionization front
(\SI{90}{AU}, \citealp{Henney:1998b}).
Unfortunately, the gas mass cannot be determined from the CO observations
because the line is optically thick.
\subsubsection{Launching regions of the jet and the photoevaporation flow}
\label{sec:launch-regi-jet}
No binary companion to the central star of 170-337 has been detected,
either spectroscopically \citep{Tobin:2009a}
or visually \citep{Duchene:2018a}.
Therefore, it is likely that
both the proplyd and HH flows originate from the same protostellar disk,
but the flows must be launched from very different radii.
In the case of the proplyd, an initially neutral photoevaporation flow
is driven from the disk surface by FUV radiation from the Trapezium stars
\citep{Johnstone:1998a}.
For the high radiation fields found in the Orion Nebula,
the flow is in the supercritical regime \citep{Adams:2004u}
in which the sound speed in the FUV-heated gas (\(\approx \SI{3}{km.s^{-1}}\))
exceeds the local gravitational escape velocity at the disk surface.
For a stellar mass of \SI{0.3}{\msun} (sec.~\ref{sec:general-properties}),
this implies radii greater than about \SI{30}{AU}
and up to the outer radius of the disk (\SI{44}{AU}).
In the case of the jet,
the much higher flow velocity (\(> \SI{100}{km.s^{-1}}\))
implies a launch region that is deeper in the gravitational well
and closer to the star.
Comparison of magneto-centrifugal jet models
with observations of the base of HH~212 \citep{Lee:2017r}
imply that jet launching occurs at disk radii \(< \SI{0.1}{AU}\).
Magnetocentrifugal winds may also be launched from intermediate radii
of \num{1} to \SI{10}{AU} but these will be at lower velocities
(e.g., \citealp{Hasegawa:2022j}).
\subsubsection{Kinematics of the star within the Orion Nebula Cluster}
\label{sec:kinem-star-with}
The line-of-sight radial velocity of the star can be determined from the systemic velocity of the CO emission
\citep{Boyden:2020a},
yielding \SI{23.8 \pm 0.1}{km.s^{-1}} after converting to the heliocentric frame.
This is \SI{-3.7}{km.s^{-1}} with respect to the mean velocity of stars in the core of Orion
(\SI{27.5}{km.s^{-1}}, \citealp{Theissen:2022k}).
The star's tangential velocity in the frame of the cluster can be derived from proper motion measurements \citep{Kim:2019a}
as \(\SI{7.8 \pm 1.1}{km.s^{-1}}\) along \(\text{PA} = \ang{48 \pm 8}\).
The total space velocity of 170-337 is therefore
\SI{8.6}{km.s^{-1}},
which is more than double the 3D velocity dispersion of the cluster
(\SI{3.7}{km.s^{-1}}; \citealp{Theissen:2022k}),
making it a candidate for an escaping star.
It is included in the escape group ESC1 of \citet{Kim:2019a},
but its total proper motion of \SI{4}{mas.yr^{-1}}
is not quite high enough to meet the runaway criterion
used in \citet{McBride:2019a} or \citet{Platais:2020f}.
\subsubsection{Kinematics of the jet}
\label{sec:kinematics-jet}
Proper motion measurements of the jet knot \citep{odellyhenney08},
after compensating for the proper motion of the star and
rescaling to our adopted distance of \SI{410}{pc},
yield a plane-of-sky tangential velocity of \SI{41 \pm 6}{km.s^{-1}},
with a position angle that is consistent with the displacement
of the knot from 177-341 (almost due north).
Line-of-sight radial velocities can be determined from our spectroscopic measurements
(see Table~\ref{tab:sample_spectra})
and are found to be approximately the same for the jet base and the knot,
with no variation between high and low ionization lines,
yielding a heliocentric velocity of \SI{+150 \pm 5}{km.s^{-1}}.
Subtracting the stellar systemic velocity of \SI{24}{km.s^{-1}}
gives a jet propagation velocity of \SI{132 \pm 10}{km.s^{-1}}
at an inclination of \ang{-72 \pm 3} to the plane of the sky.
Note that the brighter northern side of the flow is inclined away from the observer.
Assuming an ionized sound speed of \SI{11}{km.s^{-1}},
then the isothermal Mach number of the jet is \num{12}
and the Mach angle is \ang{4.8},
which should correspond to the half-opening angle of the ionized portion of the jet
after it emerges from the proplyd ionization front.
This is consistent with the observed width of the base of the jet in \textit{HST} images
(e.g., Fig.~\ref{fig:hst}).
\subsubsection{Mass loss rates of photoevaporation flow and jet}
\label{sec:mass-loss-rate}
The mass-loss rate from the proplyd photoevaporation flow is estimated
from model fits to kinematic line profiles \citep{henney99}
to be \(\approx \SI{8e-7}{\msun.yr^{-1}} \),
with an uncertainty of roughly a factor of two.
\citet{Bally:2006a} include HH~514 in a survey of irradiated jets in the Orion Nebula,
calculating a jet mass loss rate of \SI{6e-7}{\msun.yr^{-1}},
which would make it one of the most massive jets in their survey.
However, this estimate is based on analysis of the jet knot and therefore
corresponds to a maximum value of the mass loss,
presumably during an outburst phase.
Assuming that the current conditions at the jet base are more typical,
then a time-averaged jet mass loss can be estimated by scaling the
result from the proplyd photoevaporation flow.
Since the densities at the base of the jet and in the proplyd flow are similar
(section~\ref{sec:physical_cond} and further details in \ref{subsec:proplyd} below),
the mass loss rate scales as the flow velocity times its solid angle.
Assuming a jet half opening angle of \ang{5} (see previous section)
implies a mean jet mass loss rate
of \num{0.05} times that of the proplyd,
yielding \SI{4e-8}{\msun.yr^{-1}} albeit with a large uncertainty.
This is similar to but slightly larger than the jet mass-loss rate
derived for the closer-in proplyd LV~2 (167-317, \citealp{Henney:2002b}).
\subsubsection{Discrepancy between optical jet orientation and CO kinematics}
\label{sec:discr-betw-optic}
Assuming that the jet is perpendicular to the disk,
then the results of sec.~\ref{sec:kinematics-jet} imply that the disk must be close to face-on.
In the disk literature, a different inclination convention is used
in which the \(i\disk\) is the angle between the rotation axis and the line of sight,
meaning that our jet results imply \(i\disk = \ang{18}\) for 170-337.
Studies of the mm continuum and line emission \citep{Eisner:2018a, Boyden:2020a}
find that the disk has a slight elongation in the north-south direction,
which is interpreted as \(i\disk \approx \ang{40}\),
but if the jet results are taken at face value
then this elongation must be an intrinsic asymmetry rather than an inclination effect.
\citet{Boyden:2020a} derive a very low stellar mass (\(< \SI{0.1}{\msun}\))
from kinematic disk modelling.
However, this value is highly dependent on the disk inclination
(\(M \sim (\sin i\disk)^{-2}\)),
so that rescaling to \(i\disk = \ang{18}\) would give \SI{0.3}{\msun},
which is fully compatible with the mass derived from stellar evolution tracks
(see sec.~\ref{sec:central-star-disk} above).
A more serious conflict with the CO kinematics concerns the
sign and orientation of the observed velocity gradient,
which runs north-south with blueshifted CO on the north side
(Figs.~4 and~11 of \citealp{Boyden:2020a}).
This is inconsistent with an origin either in a rotating disk
or in an outflow aligned with the optical jet.
In the first case, the disk rotation axis would have to be perpendicular to the jet axis.
In the second case, one would expect \emph{red}shifted instead of blueshifted CO on the north side.
Two possible explanations for this discrepancy suggest themselves:
(1)~the source, despite lack of evidence, may in fact be binary,
with the optical jet arising from a different star/disk from that seen in CO;
(2)~the CO emission might arise in neither the disk nor the outflow,
but instead in in-falling material from the remnants of the surrounding protostellar envelope.
Both possibilities seem far-fetched,
but in the second case there is a precedent in observations of larger Orion proplyds,
where dense neutral filaments are seen within the shadow of the ionization front.
Examples include 182-413 \citep{Vicente:2013a} and 244-440 \citep{bally00},
which are likely to represent transitional stages between Evaporating Gaseous Globules
\citep{Hester:1996w}
and pure photoevaporating disks.
Although 170-337 is much smaller, it too may contain remnant envelope material
that is still accreting onto the central source.
The same phenomenon may explain asymmetrical dust extinction features
that are seen in the tails of many proplyds,
such as 171-340 (see Fig.~\ref{fig:hst}).
\subsection{On the anomalous [S\thinspace III] emission in HH~514}
\label{subsec:under_TS3}
There are two results for HH~514 included in Tables \ref{tab:pc} and \ref{tab:total_abundances} that we want to highlight: both components show $T_{\rm e} \text{ ([S\thinspace III])}$ $<$ $T_{\rm e} \text{ ([O\thinspace III])}$, and their S/H ratio is around 0.5 dex higher than the value found in the nebular gas. The abnormally large S abundance seems to rely mainly on the S$^{2+}$ one, which represents 91 per cent of the S/H ratio.
The observed trend of $T_{\rm e} \text{ ([S\thinspace III])}$ goes in the opposite direction to the usual temperature stratification observed in H~II regions and simple photoionization models fail to reproduce the observed results \citep[][]{Binette2012}. On the other hand, the overabundance of S is much too large to be produced by errors in the assumed ICF (the abundance of S$^{2+}$ alone is higher than the total S abundance of the Orion Nebula).
In order to be certain that the enhanced S abundance is real,
it is necessary to rule out possible systematic errors in the analysis.
To that end, in the following subsections we explore some scenarios that could explain the observed results.
\subsubsection{Errors in atomic data for analyzing [S\thinspace III] lines?}
\label{subsec:atomic_data}
After analyzing different sets of transition probabilities and collision strengths of different ions, \citet{Juan-de-Dios17} conclude that the derived ionic abundances can have uncertainties of around 0.1-0.2 dex introduced by the atomic data when density is lower than $10^4$ cm$^{-3}$, but they can reach values up to 0.8 dex for higher densities. Since both components of HH~514 have extremely high densities, we expect the uncertainties due to atomic data may be amplified in these objects. Given such expectations, we decided to explore the effect of the atomic data used on the S$^{2+}$ abundance.
Instead of following the methodology of \citet{Juan-de-Dios17} of analyzing all possible combinations of transition probabilities and collision strengths available, we first focus our attention on transition probabilities, discarding those that give clearly wrong results. If we do not do this in the first place, we may include spurious values in the resulting abundance distribution, which would increase the uncertainties associated with the atomic data. Once this first step is done, we check the available collision strengths, analyzing their impact on $T_{\rm e} \text{ ([S\thinspace III])}$ and the S$^{2+}$ abundance.
We use the observed intensity of [S\thinspace III] $\lambda \lambda 9531, 9069, 8829$ lines, arising from the common $^1$D$_2$ upper level and the [S\thinspace III] $\lambda \lambda 3722, 6312$ lines, emitted from the same $^1$S$_0$ level. Their intensity ratios depend exclusively on the transition probabilities. In Table~\ref{tab:atomic_data_test}, we compare the aforementioned [S\thinspace III] line intensity ratios observed in several positions and HH objects of the Orion Nebula and other Galactic H\thinspace II regions for which we have deep high-spectral resolution spectra with the line ratios predicted by the different atomic datasets.
As it is well known, infrared [S\thinspace III] $\lambda \lambda 9531, 9069$ lines can be affected by telluric absorption bands \citep{Noll12}. In \citet{mendez2021} and \citet{mendez2021-2}, we mentioned that [S\thinspace III] $\lambda 9069$ was affected by telluric absorptions, while $\lambda 9531$ remained unaffected in the background emission of the Orion Nebula, while in the HHs both lines avoided the absorption bands. Since in this work we are interested in analyzing the [S\thinspace III] $\lambda$9531/$\lambda$9069 line ratio, we correct the spectra from telluric absorptions around the aforementioned lines. We use the observations of the star GD71 from each data set, normalizing them with the standard tabulated flux. Then, the science spectra were divided by the resulting telluric transmission function around [S\thinspace III] $\lambda \lambda 9531, 9069$. The corrected intensities are shown in Table~\ref{tab:tellabs}.
[S\thinspace III] $\lambda 3722$ is usually blended with the H14 Hydrogen Balmer line. We can not simply subtract the emission of H14 using its theoretical line ratio with respect to other isolated observed H\thinspace I lines, since it is well known that H\thinspace I lines from high quantum numbers $n$ suffer deviations from case ``B'' \citep[][]{mesadelgado09, rodriguez20}. Even using contiguous lines such as H13 or H15 may introduce important deviations \citep[see fig.~A2 of][]{rodriguez20}. Therefore, we consider that it is better to use only the observations where the spectral resolution of the data allows the correct Gaussian deblending of the lines, where the resulting FWHM of [S\thinspace III] $\lambda 3722$ is consistent with that of the isolated [S\thinspace III] $\lambda 6312$ line.
The transition probabilities tested are those available in version 1.1.13 of PyNeb and those from the CHIANTI database V10 \citep[][]{Chianti10}, named by the following acronyms: LL93-HSC95-MZ82b-KS86 \citep[][]{LL93,HSC95,MZ82b, KS86}, MZ82b-HSC95-LL93\citep[][]{MZ82b,HSC95,LL93}, FFTI06 \citep[][]{FFTI06}, TZS19\citep[][]{TZS19} and CHIANTI \citep[][]{Tayal97,FFTI06, Hudson12}. These data encompass all those studied by \citet{Juan-de-Dios17} adding those from TZS19 and CHIANTI. The FFTI06 data file was named as PKW09 in previous versions of PyNeb, alluding to \citet{Podobedova09}, who cited \citet{FFTI06}.
As shown in Table~\ref{tab:atomic_data_test}, the LL93-HSC95-MZ82b-KS86 set of transition probabilities are clearly inconsistent with observations when determining line ratios involving the [S\thinspace III] $\lambda 9531$ line from the $^1$D$_2$-$^3$P$_2$ transition, for which atomic data are taken from \citet{KS86}. This discrepancy forces us to discard the LL93-HSC95-MZ82b-KS86 data set. With respect to the other data sets, the ones most consistent with the observed line ratios are those by FFTI06 and CHIANTI. However, as is shown in Fig.~\ref{fig:A_sii}, the results obtained with the remaining data sets are not very different for the most important optical transitions ([S\thinspace III] $\lambda \lambda 3722, 6312, 8829, 9069, 9531$). In fact, taking the mean and standard deviation of the A-values given by the datasets included in Fig.~\ref{fig:A_sii} excluding those from LL93-HSC95-MZ82b-KS86, we obtain A$_{3722}=(7.06 \pm 0.53) \times 10^{-01}$, A$_{6312}=2.20 \pm 0.08$, A$_{8829}=(6.42 \pm 1.49) \times 10^{-06}$, A$_{9069}=(2.00 \pm 0.13) \times 10^{-02}$ and A$_{9531}=(5.26 \pm 0.35) \times 10^{-02}$. Of these, the one with the largest deviation is A$_{8829}$, about 23 per cent. In any case, this transition probability has very little impact in the determination of $T_{\rm e} \text{ ([S\thinspace III])}$ and the S$^{2+}$ abundance if a line different than [S\thinspace III] $\lambda 8829$ is used, which is always the case. Moreover, if we discard the dataset of TZS19, this deviation is reduced to 4 per cent. Once the LL93-HSC95-MZ82b-KS86 data set is removed, the uncertainties associated with the transition probabilities of the optical lines seem to be around $\sim 5$ percent considering the consistency between the predictions and the observed line intensity ratios included in Table~\ref{tab:atomic_data_test}. The previous results indicate that the available sets of transition probabilities for transitions giving rise to [S\thinspace III] lines show a high degree of consistency. From now on, we adopt the transition probabilities of FFTI06 in our calculations. This was precisely the data set we used in Paper~I and Paper~II.
Analogously to the previous procedure with transition probabilities, we also check the data sets of collision strengths available in PyNeb: GMZ95 \citep[][]{GMZ95}, TG99 \citep[][]{TG99}, HRS12 \citep[][]{HRS12}, GRHK14 \citep[][]{GRHK14} and TZS19 \citep[][]{TZS19}. These includes all the sets analyzed by \citet{Juan-de-Dios17} in addition to TZS19. The CHIANTI data base V10 adopts the same collision strengths as HRS12 by default, so does not need to be considered separately. Comparing the results for the different data sets shown in Fig.~\ref{fig:omega_siii_zoom }, we obtain the following mean collision strengths along with their associated dispersion at 10,000 K for the most important [S\thinspace III] lines: $\Omega_{3722}=(3.65 \pm 0.31) \times 10^{-1}$, $\Omega_{6312}=1.38 \pm 0.04$, $\Omega_{8829}=(7.64 \pm 0.63)\times 10^{-1}$, $\Omega_{9069}=2.28 \pm 0.20$ and $\Omega_{9531}=4.00 \pm 0.26$. The typical dispersion is about $\sim 8$ per cent for most lines, except for [S\thinspace III] $\lambda 6312$, where the dispersion is smaller, $\sim 3$ per cent.
One way in which errors in the collision strengths could explain
the apparent low $T_{\rm e} \text{ ([S\thinspace III])}$ and the high S$^{2+}$ abundance in the HH objects
is by changing the $\frac{\Omega_{\text{Auroral}}}{\Omega_{\text{Nebular}}}=\frac{\Omega(3722)+\Omega(6312)}{\Omega(8829)+\Omega(9069)+\Omega(9531)}$ ratio. When $\frac{\Omega_{\text{Auroral}}}{\Omega_{\text{Nebular}}}$ decreases, $T_{\rm e} \text{ ([S\thinspace III])}$ increases, for a fixed observed [S\thinspace III] $\frac{I(\lambda3722)+I(\lambda6312)}{I(\lambda 8829)+I(\lambda9069)+I(\lambda9531)}$ ratio. Increasing the value of the denominator in the ratio of collision strengths also decreases the derived abundance of S$^{2+}$, which is usually determined from the intensity of [S\thinspace III] $\lambda \lambda 9531, 9069$ lines. However, modifications in the values of $\Omega_{8829}, \Omega_{9069} \text{ and } \Omega_{9531}$ within a dispersion of $\sim8$ per cent are not enough to account for the discrepant values of $T_{\rm e} \text{ ([S\thinspace III])}$ and the S$^{2+}$ abundance.
We will try to solve the aforementioned discrepancies exploring the assumption that the collision strengths of the transitions that give rise to [S\thinspace III] lines are wrong, with $\Omega_{9531}, \Omega_{9069}$ and $\Omega_{8829}$ being underestimated beyond the dispersion of $\sim8$ per cent. We will increase the values of $\Omega_{9531}, \Omega_{9069}$ and $\Omega_{8829}$, until getting $T_{\rm e} \text{ ([O\thinspace III])}=T_{\rm e} \text{ ([S\thinspace III])} $ in the jet base of HH~514. This may be considered a lower limit since it is expected that $T_{\rm e} \text{ ([S\thinspace III])} $ is between $T_{\rm e} \text{ ([O\thinspace III])} $ and $T_{\rm e} \text{ ([N\thinspace II])} $ \citep[][]{Berg20}, the latter being generally higher. To achieve this, the collision strengths should be increased about 40 per cent, as we know, 5 times larger than the dispersion found comparing the available atomic data. In Table~\ref{tab:modified_data_omega} we show the $T_{\rm e} \text{ ([S\thinspace III])} $ and S$^{2+}$ abundances we obtain for each object using the transition probabilities of FFTI06, the different sets of collision strengths considered and the corresponding value of $n_{\rm e}$. In the last column, we include the results obtained increasing by a 40 per cent the $\Omega_{9531}, \Omega_{9069}$ and $\Omega_{8829}$ values of GRHK14.
Now we discuss what other consequences the increase of 40 percent in the collision strengths may have. Using these $\Omega$ modified values, $T_{\rm e} \text{ ([S\thinspace III])}$ matches $T_{\rm e} \text{ ([O\thinspace III])}$ in both components of HH~514\footnote{This is obviously expected for the jet base, since the increase in the collision strengths was chosen to obtain $T_{\rm e} \text{ ([S\thinspace III])}= T_{\rm e} \text{ ([O\thinspace III])}$. The result for the northern knot is independent.}, while the modification implies $T_{\rm e} \text{ ([S\thinspace III])}\approx T_{\rm e} \text{ ([N\thinspace II])}$ for the Orion Nebula. The latter would imply that the S$^{2+}$ and N$^{+}$ ions coexist almost exclusively in the zone of low degree of ionization. However, the large increment we have assumed in the collision strengths are unable to give a \textit{normal}\footnote{A value consistent with that of the Orion Nebula.} S abundance in the high-velocity components. If we adopt $\text{12+log(S/H)}=7.07 \pm 0.03$ from \citet{mendez2021-2} as reference, we still have an overestimation of 0.23 dex in the jet base. A similar difference is obtained in the knot. Moreover, this experiment produces a $\text{log(S/O)} = -1.81 \pm 0.06$ for the nebular component, which is inconsistent with the solar value of $\text{log(S/O)}=-1.58 \pm 0.08$ \citep{lodders19}. Since both S and O are $\alpha$-elements, a constant abundance ratio is expected from their common nucleosynthesis. Therefore, following a \textit{reductio ad absurdum} process, we can say that large errors in the collision strengths $\Omega_{9531}$, $\Omega_{9069}$ and $\Omega_{8829}$ do not solve the observed discrepancies in $T_{\rm e} \text{ ([S\thinspace III])}$ and the total S abundance in HH~514.
\subsubsection{Is the [S\thinspace III] emission affected by the shock?}
\label{subsec:shock_affected}
HH objects are produced by the interaction of high-velocity flows
with the surrounding medium or by internal shocks within the flow.
In both cases, gas is initially heated and compressed in a shock front,
followed by a non-equilibrium cooling zone in which the gas returns to the
photoionization equilibrium temperature while undergoing further compression.
The cooling zone may contribute to observed CEL spectrum,
but the greatest effect is on highly ionized lines that are not typically seen
in \hii{} regions.
In both the jet and knot of HH~514 the low observed line widths of about
\SI{20}{km.s^{-1}}
(see Table~\ref{tab:sample_spectra})
imply that only low velocity internal shocks are present.
For moderate ionization lines such as [\ion{S}{3}], the relative contribution is therefore expected to be small \citep{mendez2021}.
Moreover, the emission from the cooling zone, if any, would affect $T_{\rm e} \text{ ([S\thinspace III])}$ in the opposite direction, increasing the observed [S\thinspace III] $I(\lambda 6312)/I(\lambda 9531)$ line intensity ratio that would produce higher temperatures.
\subsubsection{Telluric absorption bands or sky emission?}
\label{subsec:tell_abs_sky_emmision}
There are no other lines arising from sky emission or the Orion Nebula that could be blended with [S\thinspace III] $\lambda \lambda 6312, 9069, 9531$ ($\lambda \lambda 6315.31,9073.60, 9535.87$) lines at the spectral resolution of our spectra. As commented in Sect.~\ref{subsec:atomic_data}, we correct for telluric absorptions that could affect the intensity of [S\thinspace III] $\lambda \lambda 9531, 9069$ lines in our spectra. Moreover, those [S\thinspace III] lines fall outside the absorption bands in the spectra of HH~514. In any case, in the presence of absorption bands, their effect on [S\thinspace III] $\lambda \lambda 9531, 9069$ lines would fictitiously increase $T_{\rm e} \text{ ([S\thinspace III])}$ and decrease the S$^{2+}$ abundance, which is the opposite of what is found.
\subsubsection{Errors in the reddening correction?}
\label{subsec:errors_CHB}
The [S\thinspace III] $I(\lambda 6312)/I(\lambda 9531)$ line intensity ratio is highly dependent on the reddening correction. However, if we estimate what reddening coefficient is necessary in order to produce $T_{\rm e} \text{ ([S\thinspace III])} \geq T_{\rm e} \text{ ([O\thinspace III])}$, we get $\text{c}(\text{H}\beta)\geq 1.22$. This is at least 44 per cent higher than the estimated value based on the observed Balmer and Paschen H\thinspace I lines, which is simply not possible considering the quality of the flux calibration of our data. There is also no evidence of such an increase in c(H$\beta$) in the MUSE data\citep[][]{Weilbacher15}, which is consistent with our results. In any case, the derivation of a largely incorrect $\text{c}(\text{H}\beta)$ value would produce many different and noticeable inconsistencies in our results that are obviously not observed.
\subsubsection{Is the density overestimated?}
\label{subsec:ov_density}
Let us suppose that the representative density of the [S\thinspace III] zone is overestimated and its true value is more similar to that of the Orion Nebula. As can be seen in panel (a) of Fig.~\ref{fig:plasma}, this would imply that in the jet base, $T_{\rm e} \text{ ([S\thinspace III])} \approx 10200 \text{ K}$ if $n_{\rm e} \leq 10000 \text{ cm}^{-3}$. However, even assuming that value of $T_{\rm e} \text{ ([S\thinspace III])}$, the S$^{2+}$ abundance would be still overestimated. A $T_{\rm e} \text{ ([S\thinspace III])} \approx 13000 \text{ K}$ would be necessary to obtain a {\it normal} S$^{2+}$ abundance, which can not be achieved with any density value.
\subsection{Comparison between HH~514 and its associated proplyd}
\label{subsec:proplyd}
In the previous section, we established that none of the potential systematic biases
can plausibly cause the observed \siii{} emission from the HH~514 jet,
leaving an enhanced sulfur abundance as the most likely explanation.
Given that the ionized base of the jet emerges from the ionization front
of the proplyd 170-337 \citep{bally00},
it is important to compare the abundances in the two components.
In our UVES observations, 170-337 is within Cut~1. We interpolate the nebular emission above and below the zone covered by 170-337 along the slit, so that we can separate the proplyd emission from that of the Orion Nebula.
The line intensities used are shown in Table~\ref{tab:fluxes_proplyd}.
The complete study of 170-337 will be analyzed in a future paper (M\'endez-Delgado et al., in prep).
By analyzing the spectrum, we obtain $n_{\rm e}\approx 10^{5.85 \pm 0.20} \text{ cm}^{-3}$ from the $n_{\rm e}$([S\thinspace II]) $\lambda$4069/$\lambda$4076 diagnostic (no [Fe\thinspace III] emission lines are detected in 170-337). We use those [S\thinspace II] lines -- that originate from the upper levels \chem{^2P} -- because of their very high critical densities, above the $n_{\rm e}$ value obtained for the proplyd.
This density is close to the value of \SI{1.3e6}{cm^{-3}}
found by \citet{henney99} for the same proplyd
from modelling the H\(\alpha\) surface brightness profile.
In Fig.~\ref{fig:density_map} we show the 2-D density map obtained with MEGARA from the [S\thinspace II] $\lambda$6731/$\lambda$6716 ratio. Lines that come from the intermediate \chem{^2D} levels of \chem{S^{+}} and have a considerable lower critical density. In Fig.~\ref{fig:density_map}, we can see that 170-337 shows a rather low contrast with respect to the surroundings, because the emission of [S\thinspace II] $\lambda\lambda$6731, 6716 lines is dominated by the background much lower-density nebular emission. High-density clumps as the proplyd 170-337 can contribute appreciably to the emission of lines arising from the higher levels --as the $T_{\rm e}$-sensitive auroral lines--, which are less affected by collisional deexcitation. The use of incorrect density diagnostics can lead to wrong estimates of physical conditions and chemical abundances \citep[][]{mendez2021-2}.
\begin{figure}
\includegraphics[width=\columnwidth]{density_map.pdf}
\caption{Density map obtained with the intensity ratio [S\thinspace II] $\lambda$6731/$\lambda$6716 using the MEGARA-IFU observations (see Fig.~\ref{fig:megara}). This map reveals in a partial way only, the density structures present in the Field of View (see text).}
\label{fig:density_map}
\end{figure}
By adopting the correct density for the ionized layers of 170-337, we obtain $T_{\rm e} \text{ ([O\thinspace III])} \approx T_{\rm e} \text{ ([S\thinspace III])} \approx T_{\rm e} \text{ ([N\thinspace II])} \approx 8000 \pm 1000 \text{ K}$, in total consistency with the expected values for the gas in photoionization equilibrium within the radiation field of the Orion Trapezium. Although the uncertainties associated with the three temperature diagnostics are high, due to the extremely large density of the proplyd, the convergence of the diagnostics ensures the reliability of the derived physical conditions, yielding the adopted temperature $T_{\rm e}=\SI{8010 \pm 60}{K}$. With the aforementioned physical conditions we obtain $\text{12+log(S/H)} \approx 7.10 \pm 0.04 $ which is consistent with the value derived by \citet{mendez2021-2} for HH~204, representative of the chemical composition of the Orion Nebula. Therefore, the photoionized material of the outer layers of proplyd 170-337 does not show an overabundance of S.
\subsection{Implications of the elevated S abundance in HH~514}
\label{subsec:real_overabundance}
From section~\ref{subsec:under_TS3}, we have ruled out several scenarios that may produce an unrealistic overabundance of S and a low $T_{\rm e} \text{ ([S\thinspace III])}$.
Therefore, it is possible to affirm that the overabundance of S/H in HH~514 is real.
Furthermore, neither the surrounding nebula nor the proplyd photoevaporation flow
show the same overabundance.
The other significant difference between the gas phase abundances in the
two flows is the approximately solar Fe and Ni abundance in HH~514,
compared with severely depleted abundances in the proplyd (similar to the nebula).
This is most simply explained by dust destruction in the HH object,
which liberates Fe and Ni to the gas phase, while it remains locked within grains in the case of the proplyd.
It is therefore reasonable to suppose that the difference in S abundances
might also be related to processes involving dust grains.
It is common to observe relatively high gas-phase Fe and Ni abundances
in HH objects, both in photoionized
\citep[][]{Blagrave06, mesadelgado09, mendez2021, mendez2021-2}
and neutral \citep{Nisini:2002f, Giannini:2015s}
environments.
Depletions with respect to solar abundance of zero to \SI{0.3}{dex} are typically found,
as compared to a depletion of more than \SI{1}{dex} in the general ISM and \hii{} regions
\citep{Rodriguez:2002a, Jenkins09}.
This is usually interpreted as evidence for dust reprocessing in the HH shocks themselves,
although there is also evidence that some reprocessing must have occurred in
the jet launch region \citep{Bohm:2001a, Podio:2009a}.
However, the present study is the first to find a super-solar sulphur abundance in an HH object.
It is well known that planetary formation process is capable of building up large grains in protoplanetary disks \citep[][]{Beckwith2000}.
Recently, \citet{Kama19} have suggested the formation of large sulfide mineral reservoirs during the accretion process from the protoplanetary disk to a central star.
The presence of a giant planet immersed in the disk, can cause a pressure bump that can accumulate the largest dust particles \citep{Pinilla2012, Birnstiel16}, among which would be the sulfides.
If some mechanism allowed these sulfide grains to preferentially enter the jet flow from the inner disk,
and that the grains could subsequently be destroyed,
then this could explain the high sulphur abundances that we observe in HH~514.
The exact mechanism by which this might be accomplished is unclear however.
Simulations of disk winds that include the coupling of different-sized grains
to the flow have recently been developed \citep{Rodenkirch:2022o},
but these do not yet include the effects of external photoevaporation,
nor do they include the wind from the innermost disk that gives rise to the fast jet.
\subsection{The total abundances of Fe and Ni}
\label{subsec:totalFeandNi}
In the jet base component of HH~514 we are able to derive the abundances of Ni$^{+}$, Ni$^{2+}$, Fe$^{+}$ and Fe$^{2+}$. As mentioned in Sec.~\ref{sec:ionic_total_abundances}, despite only having an upper limit to the abundance of Fe$^{3+}$, the ICF schemes of \citet{rodriguez05} indicate that the total abundance of Fe/H should be close to the sum of the abundances of Fe$^{+}$ and Fe$^{2+}$ and the upper limit of Fe$^{3+}$. Both the total abundances of Ni and Fe in the jet base are a factor 34 higher than in the main emission from the Orion Nebula. Moreover, these values are very close to the reference Solar values from \citet{lodders19}. However, this does not necessarily mean that all the dust grains containing Fe and Ni have been destroyed. This is also supported by the evidence of thermal dust IR emission from HH~514 \citep{smith05}, which suggests the survival of some amount of dust in the object. In connection with Sec.~\ref{subsec:real_overabundance}, HH~514 is able to destroy the dust reservoirs of S in a planet-forming disk. However, it can also destroy dust grains of zones of the Orion Nebula outside the S reservoirs, with normal S abundance. To estimate the fraction of Fe released from each kind of dust grains, we can assume that the S excess is produced by the destruction of FeS compounds, as suggested by \citet{Kama19}. In addition, NiS compounds should be present as well. For each atom of S released, one atom of Fe or Ni should be released too. Since the S excess will be equal to the difference between the S abundance in HH~514 and the representative value of the Orion Nebula, then:
\begin{equation*}
\frac{n(\text{S})_{\text{excess}}}{n(\text{H})}\approx\frac{n(\text{Fe+Ni})_{\text{excess}}}{n(\text{H})}\approx\frac{1.05\times n(\text{Fe})_{\text{excess}}}{n(\text{H})}=1.52\times 10^{-5},
\end{equation*}
adopting $\text{12+log(S/H)}=7.07 \pm 0.03$ \citep{mendez2021-2} for the Orion Nebula. Considering the Fe and Ni abundances obtained in the nebular components, this would imply that of the total abundances of Fe/H and Ni/H in the jet base, $\sim$67 per cent would come from dust destruction of S reservoirs, $\sim$30 per cent from the destruction of dust grains of normal S abundance and $\sim$3 per cent from Fe and Ni that were already in the gaseous phase.
\section{Conclusions}
\label{sec:summary_and_conclusions}
We analyze the physical conditions and chemical abundances of the photoionized Herbig-Haro object HH~514 and the Orion Nebula making use of high-spectral resolution spectra taken with UVES at VLT, high-spatial resolution imaging from the HST and deep MEGARA-IFU spectra.
Due to the Doppler shift between the emission of the Orion Nebula and HH~514, we independently study the emission of two high-velocity components: the jet base and a knot. These HHs are extremely dense: $n_{\rm e}(\text{HH~514~jet base})=10^{5.37} \text{ cm}^{-3}$ and $n_{\rm e}(\text{HH~514~knot})=10^{4.87} \text{ cm}^{-3}$. However, they present temperatures consistent with photoionization equilibrium in the Orion Nebula, $T_{\rm e} \text{ ([O\thinspace III])} \sim 9000\text{ K}$.
The total abundances of the high-velocity components are fully consistent with the ones derived in the Orion Nebula with the exception of S, Fe, and Ni. In the case of S, the S$^{2+}$ ion comprises 91 per cent of the total abundance.
Furthermore, we found that $T_{\rm e} \text{ ([S\thinspace III])} < T_{\rm e} \text{ ([O\thinspace III])} $, which is difficult to explain under normal photoionization conditions \citep[][]{Binette2012}.
We rule out the possibility that the S overabundance may be a spurious consequence
of errors in the atomic data or in the observations.
We speculate that it may be due to the preferential incorporation of S-rich dust into
the high-velocity jet from the inner disk, which is then returned to the gas phase
either in the jet launching region or via internal shocks in the jet.
The origin of S-rich dust reservoirs maybe related to the process
that \citet{Kama19} propose to occur in protoplanetary discs,
which consists of the accumulation of large sulfide dust grains
around giant planets during accretion from the disc to the central star.
In the jet base we obtain log(Fe/H)+12$ = 7.33 \pm 0.02$ and log(Ni/H)+12$ = 6.03 \pm 0.08$. These abundances are a factor 34 higher than those derived in the Orion Nebula and are very close to the Solar values \citep[][]{lodders19}.
We estimate that $\sim$67 per cent of the total Fe/H and Ni/H abundances in the jet base have origin in the destruction of sulfide reservoirs, whereas $\sim$30 per cent would come from the destruction of dust grains with a poorer S content, having just $\sim$3 per cent from Fe and Ni previously present in the gaseous phase.
HH~514 destroys dust grains of the Orion Nebula gas more efficiently than other HHs studied in the literature such as HH~202~S, HH~529~II, HH~529~III and HH~204 \citep[][]{mesadelgado09, mendez2021, mendez2021-2}.
Knowing the relationship between the efficiency of dust destruction and the propagation velocity of the HH object is an important issue. In our series of papers about the chemical content of HH objects, we find that gas-phase Fe/H and Ni/H abundances seem to be correlated with the propagation velocity (being HH~529 the slowest and HH~514 the fastest), which is consistent with studies of HH objects propagating in neutral environments \citep[See Fig.~14 from][]{Hartigan2020}. However, it is still not clear if this relation is simply linear or shows a more complex behaviour or even if the dust destruction efficiency reaches a maximum at a given velocity. Some parameters such as the degree of ionization may also play a role. The amount of data is still too limited for addressing this question and analyses of further HH objects are still needed.
Comparison of the optical jet kinematics with the kinematics of a CO molecular emission line
close to the central star yields contradictory results.
If the molecular line traces the disk rotation, then the jet is perpendicular to the rotation axis,
but if the molecular line traces outflow, then the outflow is perpendicular to the jet.
We suggest instead that the molecular line may trace inflowing material
that is the remnant of the cloud core from which the star formed.
\section*{Acknowledgements}
We are grateful to the anonymous referee for his/her helpful comments. This work is based on observations collected at the European Southern Observatory, Chile, proposal number ESO 092.C-0323(A), and on observations made with the Gran Telescopio Canarias (GTC), installed in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofísica de Canarias, in the island of La Palma. The authors appreciate the friendly communication of Dr. Richard Booth on the planet formation processes. We acknowledge support from the Agencia Estatal de Investigaci\'on del Ministerio de Ciencia e Innovaci\'on (AEI-MCINN) under grant {\it Espectroscop\'ia de campo integral de regiones \ion{H}{2} locales. Modelos para el estudio de regiones \ion{H}{2} extragal\'acticas} with reference 10.13039/501100011033. JEM-D thanks the support of the Instituto de Astrof\'isica de Canarias under the Astrophysicist Resident Program and acknowledges support from the Mexican CONACyT (grant CVU 602402). JG-R acknowledges financial support from the Spanish Ministry of Science and Innovation (MICINN) through the Spanish State Research Agency, under Severo Ochoa Centres of Excellence Programme 2020-2023 (CEX2019-000920-S), and from the Canarian Agency for Research, Innovation and Information Society (ACIISI), of the Canary Islands Government, and the European Regional Development Fund (ERDF), under grant with reference ProID2021010074. The authors acknowledge support under grant P/308614 financed by funds transferred from the Spanish Ministry of Science, Innovation and Universities, charged to the General State Budgets and with funds transferred from the General Budgets of the Autonomous Community of the Canary Islands by the MCIU. The paper has been edited using the Overleaf facility.
\section*{DATA AVAILABILITY}
The lines measured in the spectra are entirely available in the supplementary material of this article. Table~\ref{tab:sample_spectra} is an example of its content. The rest of information is found in tables or references of this paper.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{T}{his} document is a template for \LaTeX. If you are
reading a paper or PDF version of this document, please download the
electronic file, trans\_jour.tex, from the IEEE Web site at \underline
{http://www.ieee.org/authortools/trans\_jour.tex} so you can use it to prepare your manuscript. If
you would prefer to use LaTeX, download IEEE's LaTeX style and sample files
from the same Web page. You can also explore using the Overleaf editor at
\underline
{https://www.overleaf.com/blog/278-how-to-use-overleaf-with-}\discretionary{}{}{}\underline
{ieee-collabratec-your-quick-guide-to-getting-started\#.}\discretionary{}{}{}\underline{xsVp6tpPkrKM9}
If your paper is intended for a conference, please contact your conference
editor concerning acceptable word processor formats for your particular
conference.
IEEE will do the final formatting of your paper. If your paper is intended
for a conference, please observe the conference page limits.
\subsection{Abbreviations and Acronyms}
Define abbreviations and acronyms the first time they are used in the text,
even after they have already been defined in the abstract. Abbreviations
such as IEEE, SI, ac, and dc do not have to be defined. Abbreviations that
incorporate periods should not have spaces: write ``C.N.R.S.,'' not ``C. N.
R. S.'' Do not use abbreviations in the title unless they are unavoidable
(for example, ``IEEE'' in the title of this article).
\subsection{Other Recommendations}
Use one space after periods and colons. Hyphenate complex modifiers:
``zero-field-cooled magnetization.'' Avoid dangling participles, such as,
``Using \eqref{eq}, the potential was calculated.'' [It is not clear who or what
used \eqref{eq}.] Write instead, ``The potential was calculated by using \eqref{eq},'' or
``Using \eqref{eq}, we calculated the potential.''
Use a zero before decimal points: ``0.25,'' not ``.25.'' Use
``cm$^{3}$,'' not ``cc.'' Indicate sample dimensions as ``0.1 cm
$\times $ 0.2 cm,'' not ``0.1 $\times $ 0.2 cm$^{2}$.'' The
abbreviation for ``seconds'' is ``s,'' not ``sec.'' Use
``Wb/m$^{2}$'' or ``webers per square meter,'' not
``webers/m$^{2}$.'' When expressing a range of values, write ``7 to
9'' or ``7--9,'' not ``7$\sim $9.''
A parenthetical statement at the end of a sentence is punctuated outside of
the closing parenthesis (like this). (A parenthetical sentence is punctuated
within the parentheses.) In American English, periods and commas are within
quotation marks, like ``this period.'' Other punctuation is ``outside''!
Avoid contractions; for example, write ``do not'' instead of ``don't.'' The
serial comma is preferred: ``A, B, and C'' instead of ``A, B and C.''
If you wish, you may write in the first person singular or plural and use
the active voice (``I observed that $\ldots$'' or ``We observed that $\ldots$''
instead of ``It was observed that $\ldots$''). Remember to check spelling. If
your native language is not English, please get a native English-speaking
colleague to carefully proofread your paper.
Try not to use too many typefaces in the same article. You're writing
scholarly papers, not ransom notes. Also please remember that MathJax
can't handle really weird typefaces.
\subsection{Equations}
Number equations consecutively with equation numbers in parentheses flush
with the right margin, as in \eqref{eq}. To make your equations more
compact, you may use the solidus (~/~), the exp function, or appropriate
exponents. Use parentheses to avoid ambiguities in denominators. Punctuate
equations when they are part of a sentence, as in
\begin{equation}E=mc^2.\label{eq}\end{equation}
Be sure that the symbols in your equation have been defined before the
equation appears or immediately following. Italicize symbols ($T$ might refer
to temperature, but T is the unit tesla). Refer to ``\eqref{eq},'' not ``Eq. \eqref{eq}''
or ``equation \eqref{eq},'' except at the beginning of a sentence: ``Equation \eqref{eq}
is $\ldots$ .''
\subsection{\LaTeX-Specific Advice}
Please use ``soft'' (e.g., \verb|\eqref{Eq}|) cross references instead
of ``hard'' references (e.g., \verb|(1)|). That will make it possible
to combine sections, add equations, or change the order of figures or
citations without having to go through the file line by line.
Please don't use the \verb|{eqnarray}| equation environment. Use
\verb|{align}| or \verb|{IEEEeqnarray}| instead. The \verb|{eqnarray}|
environment leaves unsightly spaces around relation symbols.
Please note that the \verb|{subequations}| environment in {\LaTeX}
will increment the main equation counter even when there are no
equation numbers displayed. If you forget that, you might write an
article in which the equation numbers skip from (17) to (20), causing
the copy editors to wonder if you've discovered a new method of
counting.
{\BibTeX} does not work by magic. It doesn't get the bibliographic
data from thin air but from .bib files. If you use {\BibTeX} to produce a
bibliography you must send the .bib files.
{\LaTeX} can't read your mind. If you assign the same label to a
subsubsection and a table, you might find that Table I has been cross
referenced as Table IV-B3.
{\LaTeX} does not have precognitive abilities. If you put a
\verb|\label| command before the command that updates the counter it's
supposed to be using, the label will pick up the last counter to be
cross referenced instead. In particular, a \verb|\label| command
should not go before the caption of a figure or a table.
Do not use \verb|\nonumber| inside the \verb|{array}| environment. It
will not stop equation numbers inside \verb|{array}| (there won't be
any anyway) and it might stop a wanted equation number in the
surrounding equation.
If you are submitting your paper to a colorized journal, you can use
the following two lines at the start of the article to ensure its
appearance resembles the final copy:
\smallskip\noindent
\begin{small}
\begin{tabular}{l}
\verb+\+\texttt{documentclass[journal,twoside,web]\{ieeecolor\}}\\
\verb+\+\texttt{usepackage\{\textit{Journal\_Name}\}}
\end{tabular}
\end{small}
\section{Units}
Use either SI (MKS) or CGS as primary units. (SI units are strongly
encouraged.) English units may be used as secondary units (in parentheses).
This applies to papers in data storage. For example, write ``15
Gb/cm$^{2}$ (100 Gb/in$^{2})$.'' An exception is when
English units are used as identifiers in trade, such as ``3\textonehalf-in
disk drive.'' Avoid combining SI and CGS units, such as current in amperes
and magnetic field in oersteds. This often leads to confusion because
equations do not balance dimensionally. If you must use mixed units, clearly
state the units for each quantity in an equation.
The SI unit for magnetic field strength $H$ is A/m. However, if you wish to use
units of T, either refer to magnetic flux density $B$ or magnetic field
strength symbolized as $\mu _{0}H$. Use the center dot to separate
compound units, e.g., ``A$\cdot $m$^{2}$.''
\section{Some Common Mistakes}
The word ``data'' is plural, not singular. The subscript for the
permeability of vacuum $\mu _{0}$ is zero, not a lowercase letter
``o.'' The term for residual magnetization is ``remanence''; the adjective
is ``remanent''; do not write ``remnance'' or ``remnant.'' Use the word
``micrometer'' instead of ``micron.'' A graph within a graph is an
``inset,'' not an ``insert.'' The word ``alternatively'' is preferred to the
word ``alternately'' (unless you really mean something that alternates). Use
the word ``whereas'' instead of ``while'' (unless you are referring to
simultaneous events). Do not use the word ``essentially'' to mean
``approximately'' or ``effectively.'' Do not use the word ``issue'' as a
euphemism for ``problem.'' When compositions are not specified, separate
chemical symbols by en-dashes; for example, ``NiMn'' indicates the
intermetallic compound Ni$_{0.5}$Mn$_{0.5}$ whereas
``Ni--Mn'' indicates an alloy of some composition
Ni$_{x}$Mn$_{1-x}$.
\begin{figure}[!t]
\centerline{\includegraphics[width=\columnwidth]{fig1.png}}
\caption{Magnetization as a function of applied field.
It is good practice to explain the significance of the figure in the caption.}
\label{fig1}
\end{figure}
Be aware of the different meanings of the homophones ``affect'' (usually a
verb) and ``effect'' (usually a noun), ``complement'' and ``compliment,''
``discreet'' and ``discrete,'' ``principal'' (e.g., ``principal
investigator'') and ``principle'' (e.g., ``principle of measurement''). Do
not confuse ``imply'' and ``infer.''
Prefixes such as ``non,'' ``sub,'' ``micro,'' ``multi,'' and ``ultra'' are
not independent words; they should be joined to the words they modify,
usually without a hyphen. There is no period after the ``et'' in the Latin
abbreviation ``\emph{et al.}'' (it is also italicized). The abbreviation ``i.e.,'' means
``that is,'' and the abbreviation ``e.g.,'' means ``for example'' (these
abbreviations are not italicized).
A general IEEE styleguide is available at \underline{http://www.ieee.org/authortools}.
\section{Guidelines for Graphics Preparation and Submission}
\label{sec:guidelines}
\subsection{Types of Graphics}
The following list outlines the different types of graphics published in
IEEE journals. They are categorized based on their construction, and use of
color/shades of gray:
\subsubsection{Color/Grayscale figures}
{Figures that are meant to appear in color, or shades of black/gray. Such
figures may include photographs, illustrations, multicolor graphs, and
flowcharts.}
\subsubsection{Line Art figures}
{Figures that are composed of only black lines and shapes. These figures
should have no shades or half-tones of gray, only black and white.}
\subsubsection{Author photos}
{Head and shoulders shots of authors that appear at the end of our papers. }
\subsubsection{Tables}
{Data charts which are typically black and white, but sometimes include
color.}
\begin{table}
\caption{Units for Magnetic Properties}
\label{table}
\setlength{\tabcolsep}{3pt}
\begin{tabular}{|p{25pt}|p{75pt}|p{115pt}|}
\hline
Symbol&
Quantity&
Conversion from Gaussian and \par CGS EMU to SI $^{\mathrm{a}}$ \\
\hline
$\Phi $&
magnetic flux&
1 Mx $\to 10^{-8}$ Wb $= 10^{-8}$ V$\cdot $s \\
$B$&
magnetic flux density, \par magnetic induction&
1 G $\to 10^{-4}$ T $= 10^{-4}$ Wb/m$^{2}$ \\
$H$&
magnetic field strength&
1 Oe $\to 10^{3}/(4\pi )$ A/m \\
$m$&
magnetic moment&
1 erg/G $=$ 1 emu \par $\to 10^{-3}$ A$\cdot $m$^{2} = 10^{-3}$ J/T \\
$M$&
magnetization&
1 erg/(G$\cdot $cm$^{3}) =$ 1 emu/cm$^{3}$ \par $\to 10^{3}$ A/m \\
4$\pi M$&
magnetization&
1 G $\to 10^{3}/(4\pi )$ A/m \\
$\sigma $&
specific magnetization&
1 erg/(G$\cdot $g) $=$ 1 emu/g $\to $ 1 A$\cdot $m$^{2}$/kg \\
$j$&
magnetic dipole \par moment&
1 erg/G $=$ 1 emu \par $\to 4\pi \times 10^{-10}$ Wb$\cdot $m \\
$J$&
magnetic polarization&
1 erg/(G$\cdot $cm$^{3}) =$ 1 emu/cm$^{3}$ \par $\to 4\pi \times 10^{-4}$ T \\
$\chi , \kappa $&
susceptibility&
1 $\to 4\pi $ \\
$\chi_{\rho }$&
mass susceptibility&
1 cm$^{3}$/g $\to 4\pi \times 10^{-3}$ m$^{3}$/kg \\
$\mu $&
permeability&
1 $\to 4\pi \times 10^{-7}$ H/m \par $= 4\pi \times 10^{-7}$ Wb/(A$\cdot $m) \\
$\mu_{r}$&
relative permeability&
$\mu \to \mu_{r}$ \\
$w, W$&
energy density&
1 erg/cm$^{3} \to 10^{-1}$ J/m$^{3}$ \\
$N, D$&
demagnetizing factor&
1 $\to 1/(4\pi )$ \\
\hline
\multicolumn{3}{p{251pt}}{Vertical lines are optional in tables. Statements that serve as captions for
the entire table do not need footnote letters. }\\
\multicolumn{3}{p{251pt}}{$^{\mathrm{a}}$Gaussian units are the same as cg emu for magnetostatics; Mx
$=$ maxwell, G $=$ gauss, Oe $=$ oersted; Wb $=$ weber, V $=$ volt, s $=$
second, T $=$ tesla, m $=$ meter, A $=$ ampere, J $=$ joule, kg $=$
kilogram, H $=$ henry.}
\end{tabular}
\label{tab1}
\end{table}
\subsection{Multipart figures}
Figures compiled of more than one sub-figure presented side-by-side, or
stacked. If a multipart figure is made up of multiple figure
types (one part is lineart, and another is grayscale or color) the figure
should meet the stricter guidelines.
\subsection{File Formats For Graphics}\label{formats}
Format and save your graphics using a suitable graphics processing program
that will allow you to create the images as PostScript (PS), Encapsulated
PostScript (.EPS), Tagged Image File Format (.TIFF), Portable Document
Format (.PDF), Portable Network Graphics (.PNG), or Metapost (.MPS), sizes them, and adjusts
the resolution settings. When
submitting your final paper, your graphics should all be submitted
individually in one of these formats along with the manuscript.
\subsection{Sizing of Graphics}
Most charts, graphs, and tables are one column wide (3.5 inches/88
millimeters/21 picas) or page wide (7.16 inches/181 millimeters/43
picas). The maximum depth a graphic can be is 8.5 inches (216 millimeters/54
picas). When choosing the depth of a graphic, please allow space for a
caption. Figures can be sized between column and page widths if the author
chooses, however it is recommended that figures are not sized less than
column width unless when necessary.
There is currently one publication with column measurements that do not
coincide with those listed above. Proceedings of the IEEE has a column
measurement of 3.25 inches (82.5 millimeters/19.5 picas).
The final printed size of author photographs is exactly
1 inch wide by 1.25 inches tall (25.4 millimeters$\,\times\,$31.75 millimeters/6
picas$\,\times\,$7.5 picas). Author photos printed in editorials measure 1.59 inches
wide by 2 inches tall (40 millimeters$\,\times\,$50 millimeters/9.5 picas$\,\times\,$12
picas).
\subsection{Resolution }
The proper resolution of your figures will depend on the type of figure it
is as defined in the ``Types of Figures'' section. Author photographs,
color, and grayscale figures should be at least 300dpi. Line art, including
tables should be a minimum of 600dpi.
\subsection{Vector Art}
In order to preserve the figures' integrity across multiple computer
platforms, we accept files in the following formats: .EPS/.PDF/.PS. All
fonts must be embedded or text converted to outlines in order to achieve the
best-quality results.
\subsection{Color Space}
The term color space refers to the entire sum of colors that can be
represented within the said medium. For our purposes, the three main color
spaces are Grayscale, RGB (red/green/blue) and CMYK
(cyan/magenta/yellow/black). RGB is generally used with on-screen graphics,
whereas CMYK is used for printing purposes.
All color figures should be generated in RGB or CMYK color space. Grayscale
images should be submitted in Grayscale color space. Line art may be
provided in grayscale OR bitmap colorspace. Note that ``bitmap colorspace''
and ``bitmap file format'' are not the same thing. When bitmap color space
is selected, .TIF/.TIFF/.PNG are the recommended file formats.
\subsection{Accepted Fonts Within Figures}
When preparing your graphics IEEE suggests that you use of one of the
following Open Type fonts: Times New Roman, Helvetica, Arial, Cambria, and
Symbol. If you are supplying EPS, PS, or PDF files all fonts must be
embedded. Some fonts may only be native to your operating system; without
the fonts embedded, parts of the graphic may be distorted or missing.
A safe option when finalizing your figures is to strip out the fonts before
you save the files, creating ``outline'' type. This converts fonts to
artwork what will appear uniformly on any screen.
\subsection{Using Labels Within Figures}
\subsubsection{Figure Axis labels }
Figure axis labels are often a source of confusion. Use words rather than
symbols. As an example, write the quantity ``Magnetization,'' or
``Magnetization M,'' not just ``M.'' Put units in parentheses. Do not label
axes only with units. As in Fig. 1, for example, write ``Magnetization
(A/m)'' or ``Magnetization (A$\cdot$m$^{-1}$),'' not just ``A/m.'' Do not label axes with a ratio of quantities and
units. For example, write ``Temperature (K),'' not ``Temperature/K.''
Multipliers can be especially confusing. Write ``Magnetization (kA/m)'' or
``Magnetization (10$^{3}$ A/m).'' Do not write ``Magnetization
(A/m)$\,\times\,$1000'' because the reader would not know whether the top
axis label in Fig. 1 meant 16000 A/m or 0.016 A/m. Figure labels should be
legible, approximately 8 to 10 point type.
\subsubsection{Subfigure Labels in Multipart Figures and Tables}
Multipart figures should be combined and labeled before final submission.
Labels should appear centered below each subfigure in 8 point Times New
Roman font in the format of (a) (b) (c).
\subsection{File Naming}
Figures (line artwork or photographs) should be named starting with the
first 5 letters of the author's last name. The next characters in the
filename should be the number that represents the sequential
location of this image in your article. For example, in author
``Anderson's'' paper, the first three figures would be named ander1.tif,
ander2.tif, and ander3.ps.
Tables should contain only the body of the table (not the caption) and
should be named similarly to figures, except that `.t' is inserted
in-between the author's name and the table number. For example, author
Anderson's first three tables would be named ander.t1.tif, ander.t2.ps,
ander.t3.eps.
Author photographs should be named using the first five characters of the
pictured author's last name. For example, four author photographs for a
paper may be named: oppen.ps, moshc.tif, chen.eps, and duran.pdf.
If two authors or more have the same last name, their first initial(s) can
be substituted for the fifth, fourth, third$\ldots$ letters of their surname
until the degree where there is differentiation. For example, two authors
Michael and Monica Oppenheimer's photos would be named oppmi.tif, and
oppmo.eps.
\subsection{Referencing a Figure or Table Within Your Paper}
When referencing your figures and tables within your paper, use the
abbreviation ``Fig.'' even at the beginning of a sentence. Do not abbreviate
``Table.'' Tables should be numbered with Roman Numerals.
\subsection{Checking Your Figures: The IEEE Graphics Analyzer}
The IEEE Graphics Analyzer enables authors to pre-screen their graphics for
compliance with IEEE Transactions and Journals standards before submission.
The online tool, located at
\underline{http://graphicsqc.ieee.org/}, allows authors to
upload their graphics in order to check that each file is the correct file
format, resolution, size and colorspace; that no fonts are missing or
corrupt; that figures are not compiled in layers or have transparency, and
that they are named according to the IEEE Transactions and Journals naming
convention. At the end of this automated process, authors are provided with
a detailed report on each graphic within the web applet, as well as by
email.
For more information on using the Graphics Analyzer or any other graphics
related topic, contact the IEEE Graphics Help Desk by e-mail at
<EMAIL>.
\subsection{Submitting Your Graphics}
Because IEEE will do the final formatting of your paper,
you do not need to position figures and tables at the top and bottom of each
column. In fact, all figures, figure captions, and tables can be placed at
the end of your paper. In addition to, or even in lieu of submitting figures
within your final manuscript, figures should be submitted individually,
separate from the manuscript in one of the file formats listed above in
Section \ref{formats}. Place figure captions below the figures; place table titles
above the tables. Please do not include captions as part of the figures, or
put them in ``text boxes'' linked to the figures. Also, do not place borders
around the outside of your figures.
\subsection{Color Processing/Printing in IEEE Journals}
All IEEE Transactions, Journals, and Letters allow an author to publish
color figures on IEEE Xplore\textregistered\ at no charge, and automatically
convert them to grayscale for print versions. In most journals, figures and
tables may alternatively be printed in color if an author chooses to do so.
Please note that this service comes at an extra expense to the author. If
you intend to have print color graphics, include a note with your final
paper indicating which figures or tables you would like to be handled that
way, and stating that you are willing to pay the additional fee.
\section{Conclusion}
A conclusion section is not required. Although a conclusion may review the
main points of the paper, do not replicate the abstract as the conclusion. A
conclusion might elaborate on the importance of the work or suggest
applications and extensions.
\appendices
Appendixes, if needed, appear before the acknowledgment.
\section*{Acknowledgment}
The preferred spelling of the word ``acknowledgment'' in American English is
without an ``e'' after the ``g.'' Use the singular heading even if you have
many acknowledgments. Avoid expressions such as ``One of us (S.B.A.) would
like to thank $\ldots$ .'' Instead, write ``F. A. Author thanks $\ldots$ .'' In most
cases, sponsor and financial support acknowledgments are placed in the
unnumbered footnote on the first page, not here.
\section*{References and Footnotes}
\subsection{References}
References need not be cited in text. When they are, they appear on the
line, in square brackets, inside the punctuation. Multiple references are
each numbered with separate brackets. When citing a section in a book,
please give the relevant page numbers. In text, refer simply to the
reference number. Do not use ``Ref.'' or ``reference'' except at the
beginning of a sentence: ``Reference \cite{b3} shows $\ldots$ .'' Please do not use
automatic endnotes in \emph{Word}, rather, type the reference list at the end of the
paper using the ``References'' style.
Reference numbers are set flush left and form a column of their own, hanging
out beyond the body of the reference. The reference numbers are on the line,
enclosed in square brackets. In all references, the given name of the author
or editor is abbreviated to the initial only and precedes the last name. Use
them all; use \emph{et al.} only if names are not given. Use commas around Jr.,
Sr., and III in names. Abbreviate conference titles. When citing IEEE
transactions, provide the issue number, page range, volume number, year,
and/or month if available. When referencing a patent, provide the day and
the month of issue, or application. References may not include all
information; please obtain and include relevant information. Do not combine
references. There must be only one reference with each number. If there is a
URL included with the print reference, it can be included at the end of the
reference.
Other than books, capitalize only the first word in a paper title, except
for proper nouns and element symbols. For papers published in translation
journals, please give the English citation first, followed by the original
foreign-language citation See the end of this document for formats and
examples of common references. For a complete discussion of references and
their formats, see the IEEE style manual at
\underline{http://www.ieee.org/authortools}.
\subsection{Footnotes}
Number footnotes separately in superscript numbers.\footnote{It is recommended that footnotes be avoided (except for
the unnumbered footnote with the receipt date on the first page). Instead,
try to integrate the footnote information into the text.} Place the actual
footnote at the bottom of the column in which it is cited; do not put
footnotes in the reference list (endnotes). Use letters for table footnotes
(see Table \ref{table}).
\section{Submitting Your Paper for Review}
\subsection{Final Stage}
When you submit your final version (after your paper has been accepted),
print it in two-column format, including figures and tables. You must also
send your final manuscript on a disk, via e-mail, or through a Web
manuscript submission system as directed by the society contact. You may use
\emph{Zip} for large files, or compress files using \emph{Compress, Pkzip, Stuffit,} or \emph{Gzip.}
Also, send a sheet of paper or PDF with complete contact information for all
authors. Include full mailing addresses, telephone numbers, fax numbers, and
e-mail addresses. This information will be used to send each author a
complimentary copy of the journal in which the paper appears. In addition,
designate one author as the ``corresponding author.'' This is the author to
whom proofs of the paper will be sent. Proofs are sent to the corresponding
author only.
\subsection{Review Stage Using ScholarOne\textregistered\ Manuscripts}
Contributions to the Transactions, Journals, and Letters may be submitted
electronically on IEEE's on-line manuscript submission and peer-review
system, ScholarOne\textregistered\ Manuscripts. You can get a listing of the
publications that participate in ScholarOne at
\underline{http://www.ieee.org/publications\_standards/publications/}\discretionary{}{}{}\underline{authors/authors\_submission.html}.
First check if you have an existing account. If there is none, please create
a new account. After logging in, go to your Author Center and click ``Submit
First Draft of a New Manuscript.''
Along with other information, you will be asked to select the subject from a
pull-down list. Depending on the journal, there are various steps to the
submission process; you must complete all steps for a complete submission.
At the end of each step you must click ``Save and Continue''; just uploading
the paper is not sufficient. After the last step, you should see a
confirmation that the submission is complete. You should also receive an
e-mail confirmation. For inquiries regarding the submission of your paper on
ScholarOne Manuscripts, please contact <EMAIL> or call +1 732
465 5861.
ScholarOne Manuscripts will accept files for review in various formats.
Please check the guidelines of the specific journal for which you plan to
submit.
You will be asked to file an electronic copyright form immediately upon
completing the submission process (authors are responsible for obtaining any
security clearances). Failure to submit the electronic copyright could
result in publishing delays later. You will also have the opportunity to
designate your article as ``open access'' if you agree to pay the IEEE open
access fee.
\subsection{Final Stage Using ScholarOne Manuscripts}
Upon acceptance, you will receive an email with specific instructions
regarding the submission of your final files. To avoid any delays in
publication, please be sure to follow these instructions. Most journals
require that final submissions be uploaded through ScholarOne Manuscripts,
although some may still accept final submissions via email. Final
submissions should include source files of your accepted manuscript, high
quality graphic files, and a formatted pdf file. If you have any questions
regarding the final submission process, please contact the administrative
contact for the journal.
In addition to this, upload a file with complete contact information for all
authors. Include full mailing addresses, telephone numbers, fax numbers, and
e-mail addresses. Designate the author who submitted the manuscript on
ScholarOne Manuscripts as the ``corresponding author.'' This is the only
author to whom proofs of the paper will be sent.
\subsection{Copyright Form}
Authors must submit an electronic IEEE Copyright Form (eCF) upon submitting
their final manuscript files. You can access the eCF system through your
manuscript submission system or through the Author Gateway. You are
responsible for obtaining any necessary approvals and/or security
clearances. For additional information on intellectual property rights,
visit the IEEE Intellectual Property Rights department web page at
\underline{http://www.ieee.org/publications\_standards/publications/rights/}\discretionary{}{}{}\underline{index.html}.
\section{IEEE Publishing Policy}
The general IEEE policy requires that authors should only submit original
work that has neither appeared elsewhere for publication, nor is under
review for another refereed publication. The submitting author must disclose
all prior publication(s) and current submissions when submitting a
manuscript. Do not publish ``preliminary'' data or results. The submitting
author is responsible for obtaining agreement of all coauthors and any
consent required from employers or sponsors before submitting an article.
The IEEE Transactions and Journals Department strongly discourages courtesy
authorship; it is the obligation of the authors to cite only relevant prior
work.
The IEEE Transactions and Journals Department does not publish conference
records or proceedings, but can publish articles related to conferences that
have undergone rigorous peer review. Minimally, two reviews are required for
every article submitted for peer review.
\section{Publication Principles}
The two types of contents of that are published are; 1) peer-reviewed and 2)
archival. The Transactions and Journals Department publishes scholarly
articles of archival value as well as tutorial expositions and critical
reviews of classical subjects and topics of current interest.
Authors should consider the following points:
\begin{enumerate}
\item Technical papers submitted for publication must advance the state of knowledge and must cite relevant prior work.
\item The length of a submitted paper should be commensurate with the importance, or appropriate to the complexity, of the work. For example, an obvious extension of previously published work might not be appropriate for publication or might be adequately treated in just a few pages.
\item Authors must convince both peer reviewers and the editors of the scientific and technical merit of a paper; the standards of proof are higher when extraordinary or unexpected results are reported.
\item Because replication is required for scientific progress, papers submitted for publication must provide sufficient information to allow readers to perform similar experiments or calculations and
use the reported results. Although not everything need be disclosed, a paper
must contain new, useable, and fully described information. For example, a
specimen's chemical composition need not be reported if the main purpose of
a paper is to introduce a new measurement technique. Authors should expect
to be challenged by reviewers if the results are not supported by adequate
data and critical details.
\item Papers that describe ongoing work or announce the latest technical achievement, which are suitable for presentation at a professional conference, may not be appropriate for publication.
\end{enumerate}
\section{Reference Examples}
\begin{itemize}
\item \emph{Basic format for books:}\\
J. K. Author, ``Title of chapter in the book,'' in \emph{Title of His Published Book, x}th ed. City of Publisher, (only U.S. State), Country: Abbrev. of Publisher, year, ch. $x$, sec. $x$, pp. \emph{xxx--xxx.}\\
See \cite{b1,b2}.
\item \emph{Basic format for periodicals:}\\
J. K. Author, ``Name of paper,'' \emph{Abbrev. Title of Periodical}, vol. \emph{x, no}. $x, $pp\emph{. xxx--xxx, }Abbrev. Month, year, DOI. 10.1109.\emph{XXX}.123456.\\
See \cite{b3}--\cite{b5}.
\item \emph{Basic format for reports:}\\
J. K. Author, ``Title of report,'' Abbrev. Name of Co., City of Co., Abbrev. State, Country, Rep. \emph{xxx}, year.\\
See \cite{b6,b7}.
\item \emph{Basic format for handbooks:}\\
\emph{Name of Manual/Handbook, x} ed., Abbrev. Name of Co., City of Co., Abbrev. State, Country, year, pp. \emph{xxx--xxx.}\\
See \cite{b8,b9}.
\item \emph{Basic format for books (when available online):}\\
J. K. Author, ``Title of chapter in the book,'' in \emph{Title of
Published Book}, $x$th ed. City of Publisher, State, Country: Abbrev.
of Publisher, year, ch. $x$, sec. $x$, pp. \emph{xxx--xxx}. [Online].
Available: \underline{http://www.web.com}\\
See \cite{b10}--\cite{b13}.
\item \emph{Basic format for journals (when available online):}\\
J. K. Author, ``Name of paper,'' \emph{Abbrev. Title of Periodical}, vol. $x$, no. $x$, pp. \emph{xxx--xxx}, Abbrev. Month, year. Accessed on: Month, Day, year, DOI: 10.1109.\emph{XXX}.123456, [Online].\\
See \cite{b14}--\cite{b16}.
\item \emph{Basic format for papers presented at conferences (when available online): }\\
J.K. Author. (year, month). Title. presented at abbrev. conference title. [Type of Medium]. Available: site/path/file\\
See \cite{b17}.
\item \emph{Basic format for reports and handbooks (when available online):}\\
J. K. Author. ``Title of report,'' Company. City, State, Country. Rep. no., (optional: vol./issue), Date. [Online] Available: site/path/file\\
See \cite{b18,b19}.
\item \emph{Basic format for computer programs and electronic documents (when available online): }\\
Legislative body. Number of Congress, Session. (year, month day). \emph{Number of bill or resolution}, \emph{Title}. [Type of medium]. Available: site/path/file\\
\textbf{\emph{NOTE: }ISO recommends that capitalization follow the accepted practice for the language or script in which the information is given.}\\
See \cite{b20}.
\item \emph{Basic format for patents (when available online):}\\
Name of the invention, by inventor's name. (year, month day). Patent Number [Type of medium]. Available: site/path/file\\
See \cite{b21}.
\item \emph{Basic format}\emph{for conference proceedings (published):}\\
J. K. Author, ``Title of paper,'' in \emph{Abbreviated Name of Conf.}, City of Conf., Abbrev. State (if given), Country, year, pp. \emph{xxxxxx.}\\
See \cite{b22}.
\item \emph{Example for papers presented at conferences (unpublished):}\\
See \cite{b23}.
\item \emph{Basic format for patents}$:$\\
J. K. Author, ``Title of patent,'' U.S. Patent \emph{x xxx xxx}, Abbrev. Month, day, year.\\
See \cite{b24}.
\item \emph{Basic format for theses (M.S.) and dissertations (Ph.D.):}
\begin{enumerate}
\item J. K. Author, ``Title of thesis,'' M.S. thesis, Abbrev. Dept., Abbrev. Univ., City of Univ., Abbrev. State, year.
\item J. K. Author, ``Title of dissertation,'' Ph.D. dissertation, Abbrev. Dept., Abbrev. Univ., City of Univ., Abbrev. State, year.
\end{enumerate}
See \cite{b25,b26}.
\item \emph{Basic format for the most common types of unpublished references:}
\begin{enumerate}
\item J. K. Author, private communication, Abbrev. Month, year.
\item J. K. Author, ``Title of paper,'' unpublished.
\item J. K. Author, ``Title of paper,'' to be published.
\end{enumerate}
See \cite{b27}--\cite{b29}.
\item \emph{Basic formats for standards:}
\begin{enumerate}
\item \emph{Title of Standard}, Standard number, date.
\item \emph{Title of Standard}, Standard number, Corporate author, location, date.
\end{enumerate}
See \cite{b30,b31}.
\item \emph{Article number in~reference examples:}\\
See \cite{b32,b33}.
\item \emph{Example when using et al.:}\\
See \cite{b34}.
\end{itemize}
\section{Introduction}
\label{sec:introduction}
Epilepsy is one of the most common chronic neurodegenerative disorders characterized by recurrent epileptic seizures (ESs) triggered by abnormal discharge of neurons \cite{world2006neurological,yang2020seizure}. Approximately 1\% people worldwide suffer from unprovoked ESs, and the primary treatment for epileptic patients is long-term drug therapy. Moreover, approximately one-third of patients are resistant to anti-epileptic medication \cite{assi2017towards,xu2021trends}. Epileptic patients can benefit from an accurate prediction system that enables early warning minutes before onset. This allows for precautional intervention or therapies, thereby alleviating the psychological burden on patients and the demand for healthcare resources.
Electroencephalography (EEG) is a prevalent neuroimaging approach widely used for epilepsy diagnosis and analysis. Recent research activities in the field of ES prediction are mainly based on either scalp EEG or invasive EEG signals \cite{rasheed2020machine,kuhlmann2018seizure,hu2020epileptic,1198251}. Fig. \ref{fig1} shows an annotated example of a single-channel EEG recording during a seizure occurrence. The recording of epileptic patients can be categorized into several periods: ictal (seizure occurrence), preictal (before seizure onset), postictal (shortly after the seizure), seizure prediction horizon (SPH, a short period before seizure onset in which precautional intervention is required) and interictal (between seizures excluding all other periods).
\begin{figure}[!t]
\centerline{\includegraphics[width=\columnwidth]{figures/periods.pdf}}
\caption{Example of an EEG illustrating different periods during a seizure occurrence.}
\label{fig1}
\end{figure}
Over the past decades, various machine learning (ML) and deep learning (DL) based algorithms have been proposed to improve the performance of ES prediction \cite{7365453,7501827,8239676,zhao2020binary,5415597}. To train a reliable classifier, either ML- or DL-based algorithms requires a large number of annotated data. However, the unavailability of sufficient annotated preictal samples is a major challenge for ES prediction studies. First, the acquisition of sufficient annotated seizure data is costly and time-consuming. Second, the preictal period is only observed if a seizure occurs. Third, only one or a few seizures usually occur per day for most patients, in extreme cases that patients with intractable epilepsy can suffer from multiple seizures per day. Therefore the duration of interictal periods is significantly longer than that of preictal periods. The natural difference between interictal and preictal durations leads to the numbers of the two types of samples for the training set to be significantly imbalanced.
Recently, several researchers have applied generative adversarial network (GAN) technique to ES studies \cite{hartmann2018eeg,rasheed2020generative,truong2019epileptic}. However, they focus on the synthesis of either short-term ($<$5 s) single-channel raw EEG ictal signals or specific features extracted from raw EEG signals. These pioneering efforts cannot be used to synthesize long-term ($>5$ s) multichannel signals. This is primarily because this type of signals easily lead to mode collapse of GAN training. Additionally, the extracted features of preictal and ictal signals are significantly different even when they appear similar. Furthermore, the lack of an evaluation method for synthetic time-series data is always a challenge for GAN studies \cite{esteban2017real}.
In this paper, we propose a data generation method based on GAN to deliver synthetic long-duration multichannel time-series EEG preictal samples to address the problem of scarcity of preictal data and imbalanced training sets. The main contributions of this study are as follows:
\begin{itemize}
\item Generative model architectures based on convolutional neural network (CNN) and recurrent neural network (RNN) have been proposed to generate synthetic single-channel EEG preictal samples. Dedicated statistical evaluation metrics are also proposed to assess the quality of single-channel samples in order to choose the outperformed generative model.
\item A high-quality generative model is chosen to generate synthetic multichannel preictal samples. The effectiveness of the model is confirmed by comparing ES prediction performance without and with data augmentation (DA) techniques. The results reveal that synthetic preictal sample augmentation can improve ES prediction performance.
\end{itemize}
The remaining content of this paper is organized as follows. Section \ref{Related} describes previous studies applying the GAN technique to the field of epilepsy study. Section \ref{Method} elaborates on the methodology of our work including generative model training, performance evaluation and implementation of ES prediction experiments without and with DA. The performance and comparison are presented in Section \ref{Result}. Sections \ref{Dis} and \ref{Conclusion} include a discussion of our results and the main conclusions, respectively.
\section{Related Work}
\label{Related}
\subsection{Literature Review}
Since proposed in 2014 \cite{goodfellow2014generative}, the GAN technique has attracted significant attention from research community owing to its success in the fields of computer vision and natural language processing. Over the past three years, numerous studies that apply the GANs to the time-series biomedical signal research are emerging. In this section, we present a concise literature review that focuses on the application of GANs on EEG signal synthesis.
\begin{figure*}[ht]
\centerline{\includegraphics[width=\textwidth]{figures/flow.pdf}}
\caption{Flow chart of the multichannel preictal signal synthesis architecture.}
\label{fig3}
\end{figure*}
Hartmann et al. \cite{hartmann2018eeg} described an EEG-GAN framework based on deep convolution GAN (DCGAN) to synthesize naturalistic single-channel hand movement EEG signals. They compared several upsampling methods within the CNN-based generator and concluded that the cubic interpolation method outperforms other interpolation methods. Fahimi et al. \cite{fahimi2020generative} introduced a framework based on DCGAN to generat synthetic EEG signals to improve the accuracy of motor task experiment with diverted and focused attention conditions by 7.32\% and 5.45\%, respectively. Aznan et al.\cite{aznan2019simulating} utilized DCGAN, Wasserstein GAN (WGAN), and a variational autoencoder to create synthetic EEG data for improving steady-state visual evoked potential classification tasks. They obtained slightly better performance on single-subject task, and 35\% improved accuracy on the across-subject generalization task. Luo et al. \cite{luo2018eeg} and Jiao et al. \cite{jiao2020driver} proposed a conditional WGAN to augment EEG signals to improve the classification performance in emotion recognition and driver sleepiness detection applications. In addition to CNN-based GAN, researchers \cite{abdelfattah2018augmenting,haradal2018biosignal} utilized RNN-based GAN to augment different bio-signals. Abdelfattah et al. proposed a novel GAN model and demonstrated their model can enhance classification performance on small EEG dataset. Haradal et al. confirmed their RNN-based GAN model is capable of generating synthetic bio-signals using the electrocardiogram and
EEG datasets. Authors in \cite{9206942}, they used bidirectional RNN neurons to generate artificial motor imagery EEG signals, and Short-term Fourier Transform (STFT) and Welch’s power spectral density to evaluate the time-frequency characteristics of artificial signals. The results indicated that GAN can capture important information of beta band. Hwang et al.\cite{JIAO2020100} proposed a scheme for zero-shot EEG signal classification using RNN-based GAN. Their EEG encoder achieves an accuracy of 95.89\% and they reached an accuracy of 39.65\% for classification task with ten untrained EEG classes. In \cite{JIAO2020100}, authors applied RNN-based GAN model to drive sleepiness detection study based on EEG and electrooculography signals. They addressed the change in alpha waves and differentiate the two alpha-related phenomena. In addition, they adopted conditional WGAN to augment dataset and boost classifier performance.
There have also been instances of application of generative models to ES studies. Truong et al. \cite{truong2019semi} proposed a semi-supervised learning approach based on DCGAN to forecast the ESs. They took advantage of features generated from the STFT of EEG signals as real data input to generate synthetic two-dimensional feature samples. Subsequently, a discriminator was used to distinguish real and generated samples. When the training reached equilibrium, the trained discriminator was used as a feature extractor for supervised learning. However, this innovative trial did not outperform the CNN model, which had previously been proposed by the same authors for three different databases \cite{truong2018convolutional}. Rasheed et al. \cite{rasheed2020generative} proposed a generative model based on DCGAN to synthesize the spectrogram of the STFT of EEG signals for ES prediction, which utilized one-class support vector machine algorithm with radial basis function for data validation. Subsequently, the transfer learning technique was used to efficiently train the ES prediction classification model. They found that synthetic spectrogram sample augmentation increased the area under the receiver operating characteristic (ROC) curve (AUC) and sensitivity on two different databases respectively. Pascual et al. \cite{pascual2020epilepsygan} proposed a conditional GAN \cite{mirza2014conditional} approach, namely EpilepsyGAN, to synthesize 4s ictal samples from two selected channels for a seizure detection task. They compared the similarity between real and synthetic ictal samples using a spectral cosine metric, and obtained an average 1.3\% improved sensitivity for 24 patients in the ES detection task. In \cite{GAO20221}, authors took advantage of GAN as a strong candidate for data enhancement intended for tackling imbalanced EEG data distribution. After data augmentation by the GAN, they designed one-dimensional CNN to achieve better performance on seizure detection classification task. You et al. \cite{YOU2020105472} performed automatic seizure detection with the trained GAN as an anomaly detector based on behind-the-ear EEG signals. They eventually achieved AUC of 0.939 and sensitivity of 96.3\%. In \cite{9441413}, authors presented a WGAN-based model to generate multichannel seizure data, and this approach is demonstrated to reduce the false positive rate by 72.72\%, meanwhile slightly increasing detection latency and maintaining sensitivity.
\section{Methodology of Preictal Signal Synthesis}
\label{Method}
\subsection{GAN theory}
The GAN framework consists of two opposing networks that are trained simultaneously to achieve an equilibrium. The first network, which is a generator, attempts to generate artificial realistic samples based on the latent noise input, and the second is a discriminator, which is used to distinguish between artificial and real samples. This adversarial learning aims to train the generator to capture the hidden structure and distribution of real samples to generate artificial samples with the same distribution. The loss function of GAN is defined as:
\begin{equation}
\mathop{min}\limits_{G}\mathop{max}\limits_{D}\mathop{\mathbb{E}}\limits_{\boldsymbol{x}\sim\mathbb{P}_{r}}[logD(\boldsymbol{x})] + \mathop{\mathbb{E}}\limits_{\boldsymbol{\widetilde{x}}\sim\mathbb{P}_{g}}[log(1-D(\boldsymbol{\widetilde{x}}))]
\end{equation}
\noindent where $\mathbb{P}_{r}$ is the real data distribution and $\mathbb{P}_{g}$ represents the model distribution defined by
\begin{equation}
\widetilde{\boldsymbol{x}} = G(\boldsymbol{z})
\end{equation}
\noindent where $\boldsymbol{z}$ is a random vector from the noise distribution.
An improved GAN known as Wasserstein GAN is proposed to overcome the training stability issue of conventional GAN \cite{arjovsky2017wasserstein,gulrajani2017improved}. One drawback of the conventional GAN framework is the training instability issue in which vanishing gradients occur in the generator if the discriminator were trained to optimality during each training iteration. WGAN is intended to minimize the Wasserstein distance (WD) instead of the Jensen-Shannon divergence, which is measured by the original GAN, between real and artificial data distributions. WD, also known as the Earth-Mover distance, is used to calculate the cost of transporting mass from one distribution to another. In the WGAN, discriminator (also called critic according to \cite{arjovsky2017wasserstein}) and generator are trained to optimize the loss function:
\begin{equation}
\mathop{min}\limits_{G}\mathop{max}\limits_{D\in\mathcal{D}}\mathop{\mathbb{E}}\limits_{\boldsymbol{x}\sim\mathbb{P}_{r}}[D(\boldsymbol{x})] + \mathop{\mathbb{E}}\limits_{\boldsymbol{\widetilde{x}}\sim\mathbb{P}_{g}}[D(\boldsymbol{\widetilde{x}})]
\end{equation}
\noindent where $\mathcal{D}$ is the set of 1-Lipschitz functions that aim to introduce Lipschitz continuity to improve the discriminator\cite{gulrajani2017improved}. For practical employment of WGAN, the weights of the discriminator are clipped to a compact interval $[-c,c]$ to enforce the Lipschitz constraint.
\subsection{Data Description}
We employed and evaluated our proposed preictal signal synthesis method based on the CHB-MIT EEG dataset \cite{shoeb2009application}. The CHB-MIT dataset contains scalp EEG signals from 23 cases collected from 22 pediatric patients, and all signals were sampled at 256 Hz. The electrode position system used in this dataset was under the International 10-20 system standard. Fig. \ref{fig2} depicts the 10-20 system electrode placement \cite{klem1999ten}. However, there were electrode configuration differences among patients in terms of practical signal acquisition. We only considered patients with the same electrode configuration and no less than three leading seizures, hence 7 subjects with the same 22 electrode bipolar placement were selected for the experiments. The characteristics of selected patient subjects are shown in Table \ref{tab3} that contains the number of leading and total seizures, preictal hours and detailed electrode placement settings, where FP, F, T, P, C, and O represent frontopolar, frontal, temporal, parietal, central, and occipital respectively, odd number, even number and Z represent left side, right side and midline of the brain.
\subsection{Data Preparation}
Conventionally, preprocessing and preparation of the data is required prior to the training phase. SPH is a very short period prior to the seizure onset providing intervention, it should last a few minutes, 5min is a common choice according to previous studies \cite{daoud2019efficient,truong2018convolutional,truong2019epileptic}. Postictal period usually lasts 5-30min, we expect the postictal signal would not disturb the interictal signal, so that the longest duration 30min is chosen in this work \cite{fisher2000postictal}. The duration of an assumed preictal period varies across seizures and individuals, state-of-the-art studies usually choose the duration of preictal from 15min to 60min, and in terms of the CHB-MIT dataset, the most popular choice is 30min. Therefore, we defined the duration of the SPH, preictal and postictal periods as 5min, 30min and 30min respectively. The preictal and interictal samples are segmented from raw EEG recordings by a time-window without overlapping, the length of the time-window is chosen as 20s because the seizure prediction algorithm used in this work is based on the work published in \cite{9073988}, and the algorithm of the later work required the training and testing samples segmented in 20s. A segment of 20s was used considering much longer continuous EEG signals than 1-5s, meanwhile we considered the issues of computational cost, so that 20s is a good option. In addition, 20s duration provides flexibility for various needs of seizure prediction algorithm design because numerous state-of-the-art algorithms adopted the segments with less than 20s. The length of the single sample should be 5120 (20s $\times$ 256Hz); however, in practical experiments, we down-sampled the real preictal samples from the raw sampling rate to 100Hz intended for training convenience \cite{hussein2019human}. Thus, the length of real preictal sample for GAN training was 2000 (20s $\times$ 100Hz). Furthermore, we aimed to train the channel-wise generators independently, so that the shape of the real preictal samples was 1 $\times$ 2000.
\begin{table}[!t]
\caption{Characteristics of Selected Patient Subjects.}
\begin{center}
\begin{tabular}{ c c c c}
\hline \hline
\makecell{Patient ID} & \makecell{No. of \\ leading seizures \\/ total seizures} & \makecell{Preictal \\ hours} & \makecell{Electrode \\ placement}\\
\hline
\makecell{chb01 \\ chb08 \\ chb09 \\ chb10 \\ chb11 \\ chb19 \\ chb23} & \makecell{3/7 \\ 3/5 \\ 3/4 \\ 6/7 \\ 3/3 \\3/3 \\ 3/7} & \makecell{1.5 \\ 1.5 \\ 1.5 \\ 9 \\ 1.5 \\ 1.5 \\ 1.5} & \makecell{FP1-F7, F7-T7, T7-P7, \\ P7-O1, FP1-F3, F3-C3, \\ C3-P3, P3-O1, FP2-F4, \\ F4-C4, C4-P4, P4-O2, \\ FP2-F8, F8-T8, T8-P8, \\ P8-O2, FZ-CZ, CZ-PZ, \\ P7-T7, T7-FT9, \\FT9-FT10, FT10-T8} \\
\hline \hline
\end{tabular}
\label{tab3}
\end{center}
\end{table}
\begin{figure}[!t]
\centerline{\includegraphics[width=6.5cm]{figures/10-20_system.png}}
\caption{Electrode locations of the International 10-20 system for EEG recording \cite{klem1999ten}.}
\label{fig2}
\end{figure}
\begin{figure}[!t]
\centering
\subfigure[]{
\includegraphics[width=0.95\columnwidth]{figures/G-DC.pdf}} \\
\subfigure[]{
\includegraphics[width=0.93\columnwidth]{figures/D-DC.pdf}}\\
\caption{Convolution neural network-based architecture: \textbf{(a)} and \textbf{(b)} are the generator and discriminator, respectively.}
\label{fig4}
\end{figure}
\begin{figure}[!ht]
\centering
\subfigure[]{\includegraphics[width=0.425\columnwidth]{figures/G-R.pdf}}
\subfigure[]{\includegraphics[width=0.47\columnwidth]{figures/D-R.pdf}}
\caption{Recurrent neural network-based architecture: \textbf{(a)} and \textbf{(b)} are the generator and discriminator, respectively.}
\label{fig5}
\end{figure}
\subsection{Procedure for Preictal EEG Samples Synthesis}
Fig. \ref{fig3} shows the flow of the proposed preictal EEG sample synthesis method. It consists of three stages, i.e., adversarial learning, data evaluation and preictal sample synthesis. In the adversarial learning stage, we consider CNN- and RNN-based generative model architectures for both GAN and WGAN training. Thus, there were four candidates – DCGAN, deep convolution WGAN (DCWGAN), recurrent GAN (RGAN) and recurrent WGAN (RWGAN) for single-channel EEG sample synthesis. In the data evaluation stage, visual inspection and proposed statistical metrics were employed to evaluate the quality of synthetic single-channel samples generated by the four alternatives, and the best performing one was chosen for the latter stage. In the final stage, the high-quality generative model was used to train the channel-wise generators of all channels independently, then identical noise was added to all trained generators to synthesize signal-channel samples for all channels, the synthesized multichannel preictal EEG samples are generated by concatenating these single-channel samples. Subsequently, the effectiveness was evaluated.
\subsection{Generative Model Architecture}
The CNN- and RNN-based generator and discriminator architectures are shown in Fig. \ref{fig4} and Fig. \ref{fig5}, respectively. These figures illustrate the operation and shape of each layer; these architectures are utilized for both GAN and WGAN training. In the CNN-based GAN, the vector from the noise latent space with a shape of 1$\times$100 was fed to the generator which comprised one fully connected (FC) layer and five convolution blocks. These blocks were expected to maintain the network deep enough and the number of FC layer nodes was determined by the shape of the subsequently designed 3D tensor with a depth of 256, which was expected to contain a complex combination of patterns. Except for the last convolution block which used pixel-wise convolution to generate the output with required shape, the remaining blocks implemented the same operations in order, i.e., upsampling with the kernel shape of 1$\times$2, convolution with the kernel shape of 1$\times$5, and a leaky ReLU active function. For the CNN-based discriminator, we utilized four convolution blocks containing the same operations in order: convolution with the kernel shape of 1$\times$5, max pooling with the kernel shape of 1$\times$2, and two FC layers with 64 nodes and 1 node, respectively.
In the RNN-based GAN, noise vectors were input to the generator which consisted of bidirectional long short-term memory (LSTM) cells with 20 time steps which were determined with the aim of processing 20s preictal samples. The output from each cell was fed to a FC layer with shared weights across time steps, and eventually a FC layer was used for generating output with the required shape \cite{hochreiter1997long}. The discriminator used an architecture similar to that of the generator, and there were only two differences that the input should be segmented with several time steps for later bidirectional LSTM cells, and a FC layer with a single neuron was added for sigmoid output generation with distinguishing objectives.
We randomly selected one channel for training different GAN architectures. When the training phase was complete, four different trained generators were used to synthesize single-channel preictal samples, and then we upsampled the synthetic preictal samples from the shape of 1$\times$2000 to the original shape of 1$\times$5120.
\subsection{Evaluation of Generative Models}
\label{evalu}
Evaluation of GAN models usually relies on the evaluation of the performance of artificial outputs generated by generator \cite{borji2019pros}. In this section, we elaborate on visual inspection and proposed statistical metrics for the evaluation of the quality of synthetic single-channel preictal samples generated by the four generators. The outperformed one is chosen to train the generators of all channels according to the evaluation approaches.
For visual inspection, we directly compared synthetic samples with real samples by vision. Although EEG signals are significantly different from images that can be easily recognized visually, we can roughly recognize the quality of EEG signals from the amplitude, frequency and period characteristics by vision. Visual inspection only provides us an intuitive quality of the synthetic EEG samples. Furthermore, reliable statistical evaluation metrics are required.
\begin{algorithm}[!t]
\caption{Algorithm for Multichannel Preictal Sample Synthesis using GAN.}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Require:}}
\renewcommand{\algorithmicensure}{\textbf{Initialization:}}
\REQUIRE $c$, the clipping parameter. $m$, the batch size. $n_{d}$, the number of iterations of the discriminator per generator iteration.
\ENSURE $\theta_{d}$, discriminator's parameters. $\theta_{g}$, generator's parameters. $c=0.01$, $m = 32$
\FOR {$channels = 1,2,\ldots,22$}
\FOR {$epochs=1,2,\ldots,30000$}
\FOR {$n_{d}=1,2,\ldots,5$}
\STATE Sample $\{ \boldsymbol{z}^{(i)} \}^{m}_{i=1} \sim p(\boldsymbol{z})$ a batch from noise prior.
\STATE Sample $\{ \boldsymbol{x}^{(i)} \}^{m}_{i=1} \sim \mathbb{P}_{r}$ a batch from the real single-channel preictal samples.
\STATE Updating the discriminator by ascending its stochastic gradient: \\ $\nabla_{\theta_{d}} \frac{1}{m}\sum\limits_{i=1}^{m} \left[log D(\boldsymbol{x}^{(i)}) + log(1-D(G(\boldsymbol{z}^{(i)}))) \right]$
\STATE Clipping the parameters $\theta_{d} \leftarrow $ clip$(\theta_{d}, -c, c)$
\ENDFOR
\STATE Sample $\{ \boldsymbol{z}^{(i)} \}^{m}_{i=1} \sim p(\boldsymbol{z})$ a batch from noise prior.
\STATE Updating the generator by ascending its stochastic gradient:\\$\nabla_{\theta_{g}} \frac{1}{m}\sum\limits_{i=1}^{m} log(1-D(G(\boldsymbol{z}^{(i)})))$
\ENDFOR
\STATE Generate synthetic single-channel preictal samples based on trained generator per channel: $G_{channels}(\boldsymbol{z})$
\ENDFOR
\STATE Combination of each synthetic single-channel samples to generate multichannel synthetic preictal samples using identical noise vector $\boldsymbol{z}$: \\ $\left[ G_{1}(\boldsymbol{z}),G_{2}(\boldsymbol{z}), \ldots, G_{22}(\boldsymbol{z})\right]^{T}$.
\end{algorithmic}
\end{algorithm}
Numerous statistical evaluation metrics can be used to evaluate the synthetic EEG signal. In this study, we selected three evaluation metrics for our application – frequency domain root mean square error (FDRMSE), Fréchet inception distance (FID) and WD. The root mean square error is widely applied to measure the difference between two signals. However, due to the similarity of real and synthetic samples in the time domain, meaningful information cannot be obtained because significant differences exist among different real samples in the time domain. Thus, the frequency domain information attracts our interest, and we compare the real and synthetic samples by calculating the RMSE of the frequency domain information. The FDRMSE is calculated as:
\begin{equation}
FDRMSE = \sqrt{\frac{1}{N}\sum_{n=1}^{N}(F_{[n]}-\hat{F}_{[n]})^2}
\end{equation}
\noindent where F and $\hat{F}$ represent the coefficients of the real and synthetic samples in the frequency domain.
FID is a common method designed to evaluate the quality of generative models and synthetic samples \cite{hartmann2018eeg,barratt2018note}. Similar to the inception score, FID utilizes a pre-trained classifier model to compare the information of real and synthetic samples representation in the embedding layer of the model. The Fréchet distance is used to calculate the distributions of two values in the embedding layer. Compared to the inception score, FID functions better because it is more robust to noise and sensitive to the quality of generative models.
WD was used as the third statistical metric for evaluation. As mentioned in Section 2.1, WD is a criterion for adversarial learning and can be used to measure the similarity between any two distributions. In this application, WD can also be regarded as sliced WD because the single-channel sample is one-dimensional.
Thus, we considered the frequency coefficients and distributions of the raw data and embedding layer instead of directly comparing the differences from the time domain. These evaluation metrics were used to select the best generative model, that was utilized to train the generators of all channels. Eventually channel-wise trained generators were used to synthesize multichannel preictal samples for following multichannel preictal sample evaluation. The pseudo-code of proposed multichannel preictal sample synthesis method is presented in Algorithm 1.
\begin{figure}[!t]
\centerline{\includegraphics[width=0.5\columnwidth]{figures/preModel.png}}
\caption{Architecture of previous proposed seizure prediction model. In the convolution blocks, the type of operation, the number of kernels and the shape of kernel are clarified. In terms of FC layers, the number of nodes are pointed. Conv: Convolution, FC: Fully connected layer.}
\label{PreModel}
\end{figure}
\subsection{Evaluation of Multichannel Samples}
\begin{figure}[!ht]
\centering
\subfigure[]{
\includegraphics[width=0.95\columnwidth]{figures/signal-real.pdf}}
\subfigure[]{
\includegraphics[width=0.95\columnwidth]{figures/signal-dcgan.pdf}}
\subfigure[]{
\includegraphics[width=0.95\columnwidth]{figures/signal-dcwgan.pdf}}
\subfigure[]{
\includegraphics[width=0.95\columnwidth]{figures/signal-rgan.pdf}}
\subfigure[]{
\includegraphics[width=0.95\columnwidth]{figures/signal-rwgan.pdf}}
\caption{Real and synthetic single-channel EEG samples. Samples with 3 s duration are shown and the amplitudes of all samples are normalized to the same range: \textbf{(a)} Real samples, \textbf{(b)}, \textbf{(c)}, \textbf{(d)} and \textbf{(e)} Synthetic samples generated by DCGAN, DCWGAN, RGAN, RWGAN, respectively. The amplitude values are scaled to [-1,1].}
\label{fig6}
\end{figure}
When the outperformed generative model is chosen, we used this model to train the generators of all channels thereby synthesizing multichannel preictal samples.
After generation of a data pool containing a sufficient number of synthetic multichannel preictal samples, the effectiveness was evaluated by comparing ES prediction performance without and with synthetic samples DA. We only augmented real preictal samples with synthetic samples at different ratios, such as 3, 5 and 10 times, and the real interictal samples were kept unchanged. We utilized the classification algorithm proposed earlier by our group (Xu et al. \cite{9073988}) as a baseline model for subsequent patient-specific ES prediction experiments. Fig. \ref{PreModel} shows the architecture of the model which consists of 3 one-dimensional convolution blocks, 2 two-dimensional convolution blocks and 2 FC layers. This model aims to address raw EEG signal and learn the time-domain representations of EEG signals. We expect that the DA condition can achieve the similar or improved performance compared to the all-real condition, in order to prove the effectiveness of synthetic preictal samples.
Several statistical metrics are considered to evaluate the performance of ES prediction in this study, such as classification accuracy, ROC curve, and AUC.
\section{Results}
\label{Result}
\subsection{Generative Model Selection}
\label{hyperP}
The experiments were carried out on Python and deep learning framework Keras with Tensorflow backend, and the hardware environment is a single NVIDIA 2080Ti GPU machine. The training details are illustrated as follows. The optimizer was Adam with momentum $\beta_1=0.9$, $\beta_2=0.999$, and learning rate $\alpha=1e-5$. The leaky ReLU action function was used a slope of 0.2. A batch size of 32 and 30000 training epochs was set for training the GAN to achieve equilibrium. In addition, we trained the discriminator once and five times per generator training iteration for the conventional GAN and WGAN respectively.
\begin{table}[!t]
\caption{Evaluation scores of synthetic single-channel samples generated by different GAN models. All real and noise samples are used as a control group.}
\begin{center}
\begin{tabular}{p{1.5cm}<{\centering} | p{1.2cm}<{\centering} | p{1.2cm}<{\centering} | p{1.2cm}<{\centering} }
\hline\hline
\textbf{Model} & \textbf{FDRMSE} & \textbf{FID} & \textbf{WD}\\
\hline
DCGAN & 7.10 & 8.8 & 0.66 \\
DCWGAN & \textbf{5.39} & \textbf{7.8} & \textbf{0.45} \\
RGAN & 7.17 & 17.4 & 0.72 \\
RWGAN & 9.32 & 18.5 & 0.92 \\
\hline
Real & 1.89 & 2.2 & 0.14 \\
Noise & 9.52 & 410 & 0.89 \\
\hline\hline
\end{tabular}
\quad
\leftline{\text{\scriptsize{\qquad \qquad FDRMSE: Frequency domain root mean square error}}}
\leftline{\text{\scriptsize{\qquad \qquad FID: Fréchet inception distance \quad WD: Wasserstein distance}}}
\label{tab1}
\end{center}
\end{table}
\begin{table}[!t]
\caption{Performance of seizure prediction experiments via leave-one-seizure-out cross validation approach. Synthetic preictal samples augmentation with three different ratios conditions are compared with all-real condition.}
\begin{center}
\begin{tabular}{ c | c | c | c | c}
\hline \hline
\makecell{Patient ID} & \makecell{All-real\\ ACC \ AUC} & \makecell{3$\times$ DA \\ ACC \ AUC} & \makecell{5$\times$ DA \\ ACC \ AUC} & \makecell{10$\times$ DA \\ ACC \ AUC} \\
\hline
\makecell{chb01 \\ chb08 \\ chb09 \\ chb10 \\ chb11 \\ chb19 \\ chb23} & \makecell{94.3 \ 0.884 \\ 86.6 \ 0.879 \\ 51.5 \ 0.449 \\ 66.7 \ 0.587 \\ 77.0 \ 0.644 \\ 44.4 \ 0.483 \\ 85.3 \ 0.809} & \makecell{95.6 \ 0.902 \\ 88.1 \ 0.881 \\ 51.5 \ 0.507 \\ 66.7 \ 0.613 \\ 77.0 \ 0.654 \\ 44.4 \ 0.501 \\ 90.3 \ 0.812} & \makecell{96.5 \ 0.910 \\ 88.5 \ 0.899 \\ 56.6 \ 0.526 \\ 69.3 \ 0.616 \\ 83.1 \ 0.663 \\ 60.7 \ 0.497 \\ 90.1 \ 0.809} & \makecell{95.9 \ 0.912 \\ 88.8 \ 0.891 \\ 56.5 \ 0.518 \\ 71.0 \ 0.637 \\ 84.7 \ 0.669 \\ 59.3 \ 0.494 \\ 90.1 \ 0.810} \\
\hline
\makecell{Average} & \makecell{73.0 \ 0.676} & \makecell{76.8 \ 0.696} & \makecell{77.8 \ 0.703} & \makecell{78.0 \ 0.704} \\
\hline \hline
\end{tabular}
\quad
\leftline{\text{\scriptsize{\quad DA: Data augmentation \quad ACC: Accuracy (\%) \quad AUC: Area under curve}}}
\label{tab2}
\end{center}
\end{table}
\begin{figure}[!t]
\centerline{\includegraphics[width=\columnwidth]{figures/AUC-figure2.png}}
\caption{Accuracy histogram of seizure prediction task under the all-real and three data augmentation conditions. Average item represents averaged values of total 7 subjects.}
\label{fig7}
\end{figure}
\begin{figure}[!ht]
\centering
\subfigure[]{\includegraphics[width=0.49\columnwidth]{figures/ROC_real_mix.png}}
\subfigure[]{\includegraphics[width=0.49\columnwidth]{figures/ROC_3x_mix.png}}
\subfigure[]{\includegraphics[width=0.49\columnwidth]{figures/ROC_5x_mix.png}}
\subfigure[]{\includegraphics[width=0.49\columnwidth]{figures/ROC_10x_mix.png}}
\caption{ROC of seizure prediction without and with data augmentation by synthetic preictal samples. Each line represents performance of each patient-specific model respectively: \textbf{(a)} The all-real condition without data augmentation, \textbf{(b)-(d)} The data augmentation condition with 3$\times$, 5$\times$, and 10$\times$ synthetic preictal samples, respectively.}
\label{fig8}
\end{figure}
Each trained generator could generate a data pool containing any number of synthetic preictal samples. Fig. \ref{fig6} shows the real and synthetic samples generated by the four different generators, and we display four samples randomly selected from each data pool. According to visual inspection, synthetic samples generated by DCGAN and DCWGAN are better than those generated by RGAN and RWGAN. This is because diversities in frequency, amplitude, and period are expected within the single real sample based on visual inspection, and samples should appear distinct among different real samples. However, the two RNN-based GANs both synthesize samples with similar appearances, and diversities cannot be observed within the single sample. Even though synthetic samples generated by DCGAN and DCWGAN have better appearance, we still cannot distinguish their performance by vision directly. Thus, we need statistical methods to evaluate the quality of samples generated by these two GANs.
Table \ref{tab1} shows the scores of three statistical metrics achieved by the four GANs. We randomly selected 100 samples from the data pool generated by each trained generator for score calculation, and \textit{real} and \textit{noise} items were added as a control group for comparison. We compared two different sample sets where one is selected from real samples and the other one is selected from samples of corresponding items. In terms of \textit{real} item, two different real sample sets, where each includes 100 real preictal samples randomly selected from real data pool, are compared. We accumulated total 100 scores, then averaged them to generate final score shown in Table \ref{tab1}. According to the statistical scores, we can make unanimous conclusion from visual inspection that DCGAN and DCWGAN outperform RGAN and RWGAN. Furthermore, between the two CNN-based GANs, DCWGAN achieves better scores on all three metrics than DCGAN. Therefore, we selected the DCWGAN as a high-quality model to train the generators of all channels. In terms of implementation of experiments, there are 5,778,066 parameters among the proposed DCWGAN model, and training each channel generator would cost $\sim$4h on a single GPU environment.
\begin{figure}[t]
\centerline{\includegraphics[width=\columnwidth]{figures/long-recordings.png}}
\caption{Synthetic and real single-channel EEG recordings in 5 minutes. Blue and red signals stand for synthetic and real recordings, respectively. The amplitude values are scaled to [-1,1].}
\label{fig10}
\end{figure}
\subsection{Synthetic Multichannel Preictal Samples}
\label{multiEva}
Once adversarial training of the chosen model for all channels was complete, we were able to generate synthetic complete preictal samples based on a channel-wise trained generator. With the availability of sufficient synthetic preictal samples, the effectiveness was evaluated by comparing performance of ES prediction experiments without and with augmented synthetic preictal samples. We implemented ES prediction experiments by patient-specific leave-one-seizure-out cross validation approach, preictal samples before one seizure are regarded as testing samples, and the rest preictal samples before all other seizures are used to train the model. We implemented 5 different runs and recorded averaged values of cross validation results for performance comparisons. Only preictal training set is augmented with synthetic samples at different ratios, and the number of interictal samples were always equal to the number of preictal samples for the training and testing phases.
Table \ref{tab2} shows leave-one-seizure-out cross validation performance in the ES prediction experiments under $3\times$, $5\times$, and $10\times$ synthetic preictal sample DA conditions, all of them were used for comparison. From the overall results, all-real condition achieves accuracy of 73.0\% and AUC of 0.676, and the three DA conditions achieve accuracies of 76.8\%, 77.8\%, 78.0\% and AUC of 0.696, 0.703, 0.704, respectively. The results reveal that synthetic preictal sample augmentation can achieve improved performance $\sim$5\% in accuracy and $\sim$3\% in AUC on average compared to the results achieved from all-real data. Fig. \ref{fig7} shows the intuitively manifested histogram of the accuracy under four conditions. We can observe that data augmentation by synthetic preictal samples achieved enhanced performance for all subjects.
Fig. \ref{fig8} shows the ROC curves of patient-specific ES prediction performances without and with DA. As for each subject, we randomly selected one seizure to localize the preictal samples and randomly selected equal number of interictal samples to test the ROC performance. According to the theory of ROC curve , the line closer to the top left area can perform better. We can observe that the curves from figure (a) are distributed around the dash line, while figure (b)-(d) show that the curves are much closer to the top left area. Therefore, Fig. \ref{fig8} can also demonstrate that synthetic preictal sample augmentation can improve the performance of ES prediction.
\begin{table}[]
\caption{COMPARISON OF OUR WORK WITH PRIOR-ART PUBLICATIONS}
\label{com}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
Reference & {\cite{truong2019semi}} & {\cite{pascual2020epilepsygan}} & {\cite{YOU2020105472}} & {\cite{9441413}} & {\cite{rasheed2020generative}} & {\cite{GAO20221}} & \textbf{\begin{tabular}[c]{@{}c@{}}This\\ Work\end{tabular}} \\ \hline
Year & 2019 & 2020 & 2020 & 2021 & 2021 & 2022 & 2022 \\ \hline
\begin{tabular}[c]{@{}c@{}}Seizure\\ Prediction\\ Task?\end{tabular} & \checkmark & × & \checkmark & × & \checkmark & × & \checkmark \\ \hline
\begin{tabular}[c]{@{}c@{}}GAN is\\ used for \\ DA?\end{tabular} & × & \checkmark & × & \checkmark & \checkmark & \checkmark & \checkmark \\ \hline
\begin{tabular}[c]{@{}c@{}}Synthesize\\ Raw EEG?\end{tabular} & × & \checkmark & × & \checkmark & × & \checkmark & \checkmark \\ \hline
\begin{tabular}[c]{@{}c@{}}Synthesize\\ Long-term\\ EEG\\ Signals?\end{tabular} & × & × & × & \checkmark & × & \checkmark & \checkmark \\ \hline
\begin{tabular}[c]{@{}c@{}}LOSC-CV\\ Test?\end{tabular} & × & × & × & × & × & × & \checkmark \\ \hline
\begin{tabular}[c]{@{}c@{}}Prediction\\ Performance\\ is improved?\end{tabular} & × & × & × & × & \checkmark & × & \checkmark \\ \hline
\end{tabular}
\quad
\leftline{\text{\scriptsize{LOSO-CV: Leave-One-Seizure-Out Cross Validation \quad DA: Data Augmentation}}}
\end{table}
\section{Discussion}
\label{Dis}
The results from Section \ref{multiEva} indicate augmentation with synthetic samples can be implemented effectively on either few times (3$\times$) or more times (10$\times$) condition. However, we cannot confirm that which augmentation ratio can bring best performance according to Table \ref{tab2}, because performances are similarly enhanced among three ratios. This is sufficient to demonstrate the goal of enhancing ES prediction by improved prediction performance metrics and addressing the unbalanced dataset issue to achieve robust results. Furthermore, the privacy of patient is an important issue to be considered when medical data are shared, hospitals cannot share private medical data without the permission of patients. Our proposed method that synthesis of effective artificial EEG signals is a solution enabling privacy preservation\cite{van2014systematic,clifton2004privacy}.
Most state-of-the-art seizure prediction machine learning algorithms extracted $<$30s segments from raw recordings as training samples, however, numerous recent studies raised that long-term EEG recordings in several minutes take an important role in overcoming limitations of epilepsy studies\cite{duun2020new,xinghua2020clinical}. Thus, we tested our proposed method to synthesize single-channel EEG recordings in 5 minutes. Because there is a scarce of existing algorithm using recordings in several minutes as training samples, synthetic recordings will be evaluated only by visual inspection and statistical methods. Fig. \ref{fig10} shows 3 randomly selected 5min synthetic and real recordings, we can see that they are quite similar visually. In addition, we evaluated them by 3 statistical methods proposed in Section \ref{evalu}, FDRMSE of 6.58, FID of 8.2 and WD of 0.49 are achieved between synthetic and real recordings. Thus, this can demonstrate similarity between synthetic and real recordings.
In Table \ref{com}, we compared our work with prior-art publications related to the application GAN to epilepsy studies. We can see the merits and advantages of our work according to the table.
Researchers would like to extract various features from raw EEG signals to train ES prediction classification models, few studies have processed raw EEG samples directly. In addition, researchers may need to determine features or raw signals extracted by different time windows, so that the synthesis of specific features and short-duration signals cannot be widely applied to various needs for ES prediction studies. While our synthetic preictal signal is more flexible owing to its raw EEG format.
Mode collapse is one of the most challenging aspects of GAN training. According to the trials of training our proposed GAN, the different noises would converge to the same synthetic single-channel samples if mode collapse occurs. In addition, the hyperparameters, and epochs of adversarial training also have a crucial impact on the successful training of effective generative models. In the practical experiments, we focused on tuning the hyperparameters and the number of epochs to effectively train generative models, eventually parameters are determined as illustrated in Section \ref{hyperP}. In terms of the proposed method, we trained channel-wise generative models to initially synthesize single-channel preictal samples rather than directly synthesize multichannel samples. This is because multichannel EEG signals containing a large number of data dimensions can easily fail to train the GAN. And only a few channels are still informative, it is proved that few channels or even 2 channels are also effective for the seizure prediction task \cite{jana2021deep,salant1998prediction,chang2012channel}. Hence, our single-channel generative model is also flexible for the condition of channel reduction.
\section{Conclusion}
\label{Conclusion}
We proposed a generative model based on DCWGAN to successfully synthesize effective multichannel EEG preictal samples for ES prediction. We initially generated synthetic signal-channel EEG samples and compared the quality for determining the best GAN architecture from four alternatives. Then, complete multichannel samples were synthesized by channel-wise generators. According to the comparisons of ES prediction performance, synthetic preictal sample augmentation can reach or improve the results based on the all-real data condition, which demonstrates the effectiveness of synthetic complete multichannel EEG preictal samples.
To the best of our knowledge, this is the first study to apply a GAN to synthesize artificial multichannel EEG preictal samples for ES prediction. This method provides a solution for the scarcity of a sufficient number of preictal samples as well as privacy concerns. Furthermore, ES prediction is one of the most difficult biomedical applications that relies on EEG signals, and we believe that our method can be applied in further research efforts to synthesize EEG signals for other scenarios.
\section*{Acknowledgment}
Authors would like to acknowledge funding support from Westlake University, Zhejiang Key R$\&$D Program No. 2021C03002 and Bright Dream Joint Institute for Intelligent Robotics.
\par
\bibliographystyle{IEEEtran}
|
\section{(1-1)-dimensional massless scalar fields in Rindler spacetime} \label{massless_scalar_fields_in_Rindler_spacetime}
In this section, we give a brief review of (1-1)-dimensional massless scalar fields in Rindler spacetime. Specifically, we show the explicit form of $\hat{\phi}(t,x)$ -- defined in the main paper as a massless scalar field in Minkowski coordinates -- and its transformed counterparts in Rindler coordinates $\hat{\Phi}_{L,R}(T,X)$. The expression of $\hat{\phi}(t,x)$ ($\hat{\Phi}_{L,R}(T,X)$) will be given in terms of $\hat{a}(k)$ ($ \hat{b}_{L,R}(K)$) defined in the main paper as the annihilation operator for a Minkowski (Rindler) particle with momentum $k$ ($K$). Moreover, we show the explicit expression for the respective conjugate momentum $\hat{\pi}(t,x)$ and $\hat{\Pi}_{L.R}(t,x)$. From the field transformation $\hat{\phi}(t,x) \mapsto \hat{\Phi}_{L,R}(T,X)$ and $\hat{\pi}(t,x) \mapsto \hat{\Pi}_{L,R}(T,X)$ we derive the Bogolyubov transformation $\hat{a}(k) \rightarrow \hat{b}_{L,R}(K)$. Finally we show the canonical commutation relations that define the $\hat{a}(k)$- and $\hat{b}_{L,R}(K)$-algebra.
The massless free scalar field $\hat{\phi}(t,x)$ is solution to the Klein–Gordon equation $\square \hat{\phi}(t,x) = 0$ and writes
\begin{equation} \label{scalar_field_M}
\hat{\phi}(t,x) = \int_{-\infty}^{+\infty} \frac{dk}{\sqrt{2\pi|k|}} \left[ e^{-i|k|ct+ikx} \hat{a}(k) + e^{i|k|ct-ikx}\hat{a}^\dagger (k) \right].
\end{equation}
Analogously, the transformed scalar fields in the Rindler frames $\hat{\Phi}_{L,R}(T,X)$ are still solution to the Klein-Gordon equation and write
\begin{equation} \label{scalar_field_R}
\hat{\Phi}_{L,R}(T,X) = \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} \left[ e^{-i|K|cT+iKX} \hat{b}_{L,R}(K) + e^{i|K|cT-iKX} \hat{b}_{L,R}^\dagger(K) \right].
\end{equation}
The conjugate momentum of $\hat{\phi}(t,x)$ in the Minkowski spacetime is defined as $\hat{\pi}(t,x) = c^{-1} \partial_t \hat{\phi}(t,x)$ and explicitly writes
\begin{equation} \label{classical_momentum_field_M}
\hat{\pi}(t,x) = \int_{-\infty}^{+\infty}dk \sqrt{ \frac{|k|}{2\pi}} i \left[ -e^{-i|k|ct+ikx} \hat{a}(k) + e^{i|k|ct-ikx} \hat{a}^\dagger(k)\right],
\end{equation}
while in the Rindler spacetime $\hat{\Pi}_{L,R}(T,X) = c^{-1} \partial_T \hat{\Phi}_{L,R}(T,X)$ and hence
\begin{equation} \label{classical_momentum_field_R}
\hat{\Pi}_{L,R}(T,X) = \int_{-\infty}^{+\infty}dK \sqrt{ \frac{|K|}{2\pi}} i \left[-e^{-i|K|cT+iKX} \hat{b}_{L,R}(K) + e^{i|K|cT-iKX} \hat{b}^\dagger_{L,R}(K) \right].
\end{equation}
The coefficients $\hat{a}(k)$ can be decomposed as $\hat{a}(k) = \hat{a}_L(k) + \hat{a}_R(k)$ by defining the fields $\hat{\phi}_{L,R}(x) = \theta(s_{L,R} x)\hat{\phi}(0,x)$, $\hat{\pi}_{L,R}(x) = \theta(s_{L,R} x)\hat{\pi}(0,x)$ and the coefficients $\hat{a}_{L,R}(k)$ such that
\begin{subequations}
\begin{equation}
\hat{\phi}_{L,R}(x) = \int_{-\infty}^{+\infty} \frac{dk}{\sqrt{2\pi|k|}} \left[ e^{ikx} \hat{a}_{L,R}(k) + e^{-ikx}\hat{a}_{L,R}^\dagger(k) \right],
\end{equation}
\begin{equation}
\hat{\pi}_{L,R}(x) = \int_{-\infty}^{+\infty}dk \sqrt{ \frac{|k|}{2\pi}} i \left[ -e^{ikx} \hat{a}_{L,R}(k) + e^{-ikx} \hat{a}_{L,R}^\dagger(k)\right].
\end{equation}
\end{subequations}
The Bogoliubov transformations relating $\hat{a}(k)$ and $\hat{b}_{L,R}(K)$ are the following
\begin{equation}\label{Bogolyubov_transformation}
\hat{a}(k) = \int_{-\infty}^{+\infty}dK \left[ \alpha(k,K)\hat{b}_L(K) - \beta^*(k,K) \hat{b}^\dagger_L(K) + \alpha^*(k,K) \hat{b}_R(K) - \beta(k,K) \hat{b}^\dagger_R(K) \right],
\end{equation}
with
\begin{subequations}\label{Bogolyubov_coefficients}
\begin{equation}
\alpha(k,K) = \theta(kK) \sqrt{\frac{K}{k}}F(k,K), \quad \beta(k,K) = \theta(kK) \sqrt{\frac{K}{k}}F(-k,K),
\end{equation}
\begin{equation} \label{F}
F(k,K) = \frac{1}{2 \pi a} \Gamma \left( -\frac{i K}{a} \right) \exp \left( i \frac{K}{a} \ln \frac{|k|}{a} + \text{sign} \left( k \right) \frac{\beta}{4} K \right).
\end{equation}
\end{subequations}
Such transformation can be obtained by requiring that $\hat{\phi}(t,x)$ transforms as a scalar field under any coordinate transformation, while $\hat{\pi}(t,x)$ as a time-derivative of a scalar field. For instance, the coordinate transformation in the right wedge $(t,x) \mapsto (T,X)=(T_R(t,x), X_R(t,x))$ for the fields writes
\begin{subequations}\label{Bogolyubov_transformation_phi_Pi}
\begin{equation}\label{Bogolyubov_transformation_phi}
\hat{\Phi}_R(T,X) = \hat{\phi}(t_R(T,X),x_R(T,X)),
\end{equation}
\begin{align}
\hat{\Pi}_R(T,X) = & \frac{1}{c} \frac{ \partial}{ \partial T} \hat{\Phi}_R(T,X) \nonumber \\
= & \frac{1}{c} \left( \frac{\partial t_R}{\partial T} \frac{ \partial}{ \partial t} + \frac{\partial x_R}{\partial T} \frac{ \partial}{ \partial x} \right) \hat{\phi} (t_R(T, X),x_R(T, X)) \nonumber \\
= & e^{aX} \cosh(a c T) \frac{1}{c} \frac{ \partial}{ \partial t} \hat{\phi} (t_R(T, X),x_R(T, X)) + e^{aX} \sinh(a c T) \frac{ \partial}{ \partial x} \hat{\phi} (t_R(T, X),x_R(T, X)) \nonumber \\
= & e^{aX} \cosh(a c T) \hat{\pi} (t_R(T, X),x_R(T, X)) + e^{aX} \sinh(a c T) \frac{ \partial}{ \partial x} \hat{\phi} (t_R(T, X),x_R(T, X)).
\end{align}
\end{subequations}
Eqs. (\ref{Bogolyubov_transformation_phi_Pi}) result in the following transformation for $\hat{a}_R(k)$
\begin{align} \label{Bogolyubov_transformation_calculation}
\hat{a}_R(k) = & \int_{-\infty}^{+\infty}dx e^{-ikx} \frac{1}{2} \left[ \sqrt{\frac{|k|}{2\pi}} \hat{\phi}_R(x) + \frac{i}{\sqrt{2\pi|k|}} \hat{\pi}_R(x) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dx e^{-ikx} \frac{1}{2} \left[ \sqrt{\frac{|k|}{2\pi}} \theta(x) \hat{\phi}(0,x) + \frac{i}{\sqrt{2\pi|k|}} \theta(x) \hat{\pi}(0,x) \right] \nonumber \\
= & \int_{0}^{+\infty}dx e^{-ikx} \frac{1}{2} \left[ \sqrt{\frac{|k|}{2\pi}} \hat{\phi}(0,x) + \frac{i}{\sqrt{2\pi|k|}} \hat{\pi}(0,x) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dX e^{a X} \exp (- i k x_R(X)) \frac{1}{2} \left[ \sqrt{\frac{|k|}{2\pi}} \hat{\phi}(0,x_R(X)) + \frac{i}{\sqrt{2\pi|k|}} \hat{\pi}(0,x_R(X)) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dX \exp \left( -i\frac{k}{a}e^{aX} \right) \frac{1}{2} \left[ \sqrt{\frac{|k|}{2\pi}} e^{a X} \hat{\phi}(0,x_R(X)) + \frac{i e^{a X} }{\sqrt{2\pi|k|}} \hat{\pi}(0,x_R(X)) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dX \exp \left( -i\frac{k}{a}e^{aX} \right) \frac{1}{2} \left[ \sqrt{\frac{|k|}{2\pi}} e^{aX} \hat{\Phi}_R( 0, X ) + \frac{i}{\sqrt{2\pi|k|}} \hat{\Pi}_R ( 0, X) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dX \frac{1}{2} \left[ \sqrt{\frac{|k|}{2\pi}} \hat{\Phi}_R( 0, X ) ia \frac{d}{dk} + \frac{i}{\sqrt{2\pi|k|}} \hat{\Pi}_R ( 0, X) \right] \exp \left( -i\frac{k}{a}e^{aX} \right) \nonumber \\
= & \int_{-\infty}^{+\infty}dX \frac{1}{2} \left\lbrace \sqrt{\frac{|k|}{2\pi}} \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} \left[ e^{iKX} \hat{b}_R(K) + e^{-iKX} \hat{b}_R^\dagger(K) \right] ia \frac{d}{dk} \right. \nonumber \\
& \left. - \frac{1}{\sqrt{2\pi|k|}} \int_{-\infty}^{+\infty}dK \sqrt{ \frac{|K|}{2\pi}} \left[-e^{iKX} \hat{b}_R(K) + e^{-iKX} \hat{b}^\dagger_R(K) \right] \right\rbrace \exp \left( -i\frac{k}{a}e^{aX} \right) \nonumber \\
= & \int_{-\infty}^{+\infty} \frac{dK}{2} \int_{-\infty}^{+\infty}dX \left\lbrace \frac{1}{2\pi} \sqrt{\left|\frac{k}{K}\right|} \left[ e^{iKX} \hat{b}_R(K) + e^{-iKX} \hat{b}_R^\dagger(K) \right] ia \frac{d}{dk} \right. \nonumber \\
& \left. - \frac{1}{2\pi} \sqrt{\left|\frac{K}{k}\right|} \left[-e^{iKX} \hat{b}_R(K) + e^{-iKX} \hat{b}^\dagger_R(K) \right] \right\rbrace \exp \left( -i\frac{k}{a}e^{aX} \right) \nonumber \\
= & \int_{-\infty}^{+\infty} \frac{dK}{2} \left\lbrace \sqrt{\left|\frac{k}{K}\right|} \left[ \hat{b}_R(K) i a \frac{d}{dk} F(-k,-K) + \hat{b}^\dagger_R(K)i a \frac{d}{dk} F(-k,K) \right] \right. \nonumber \\
& \left. - \sqrt{\left|\frac{K}{k}\right|} \left[ - \hat{b}_R(K) F(-k,-K) + \hat{b}^\dagger_R(K) F(-k,K) \right] \right\rbrace,
\end{align}
where
\begin{equation}
F(k,K) = \int_{-\infty}^{+\infty}\frac{dX}{2\pi}\exp\left(-i K X + i \frac{k}{a} e^{aX} \right)
\end{equation}
is a distribution that can be obtained from the following distributional limit
\begin{equation}\label{F_regularized}
F(k,K) = \lim_{\epsilon \rightarrow 0^+} \int_{-\infty}^{+\infty}\frac{dX}{2\pi} \exp\left((-i K + \epsilon a) X + \left(i \frac{k}{a} - \epsilon\right) e^{aX} \right).
\end{equation}
The distribution $F(k,K)$ of Eq. (\ref{F_regularized}) can be proven to be identical to $F(k,K)$ of Eq. (\ref{F}) -- for such proof see for instance Ref. \cite{Mukhanov2007}. The derivative of $F(k,K)$ with respect to $k$ can be obtained by using integration by parts and taking the distributional limit $\epsilon \rightarrow 0^+$:
\begin{align} \label{dFdk}
\frac{d}{dk} F(k,K) = & \lim_{\epsilon \rightarrow 0^+} \int_{-\infty}^{+\infty}dX \frac{i e^{aX}}{2\pi a} \exp\left((-i K + \epsilon a) X + \left(i \frac{k}{a} - \epsilon\right) e^{aX} \right) \nonumber \\
= & \lim_{\epsilon \rightarrow 0^+} \int_{-\infty}^{+\infty}dX \frac{i}{2\pi a} \exp\left((-i K + \epsilon a) X \right) \frac{1}{a} \left(i \frac{k}{a} - \epsilon\right)^{-1} \frac{d}{dX} \exp \left(\left(i \frac{k}{a} - \epsilon\right) e^{aX} \right) \nonumber \\
= & \lim_{\epsilon \rightarrow 0^+} \left[ \left. \frac{i}{2 \pi a^2} \left(i \frac{k}{a} - \epsilon\right)^{-1} \exp\left((-i K + \epsilon a) X + \left(i \frac{k}{a} - \epsilon\right) e^{aX} \right) \right|_{-\infty}^{+\infty} \right. \nonumber \\
& \left. - \int_{-\infty}^{+\infty}dX\frac{i }{2 \pi a^2} \left(i \frac{k}{a} - \epsilon\right)^{-1} (-iK+\epsilon a) \exp\left((-i K + \epsilon a) X + \left(i \frac{k}{a} - \epsilon\right) e^{aX} \right) \right] \nonumber \\
= & \lim_{\epsilon \rightarrow 0^+} \left[ \frac{iK-\epsilon a}{a (k + i \epsilon a)} \int_{-\infty}^{+\infty} \frac{dX}{2 \pi} \exp\left((-i K + \epsilon a) X + \left(i \frac{k}{a} - \epsilon\right) e^{aX} \right) \right] \nonumber \\
= & \frac{iK}{ak} F(k,K).
\end{align}
In this way the calculation of Eq. (\ref{Bogolyubov_transformation_calculation}) leads to
\begin{align}\label{Bogolyubov_transformation_calculation_2}
\hat{a}_R(k) = & \int_{-\infty}^{+\infty} \frac{dK}{2} \left\lbrace \sqrt{\left|\frac{k}{K}\right|} \left[ \hat{b}_R(K) \frac{K}{k} F(-k,-K) - \hat{b}^\dagger_R(K) \frac{K}{k} F(-k,K) \right] \right. \nonumber \\
& \left. - \sqrt{\left|\frac{K}{k}\right|} \left[ - \hat{b}_R(K) F(-k,-K) + \hat{b}^\dagger_R(K) F(-k,K) \right] \right\rbrace \nonumber \\
= & \int_{-\infty}^{+\infty} \frac{dK}{2} \left\lbrace \sqrt{\left|\frac{k}{K}\right|} \left[ \hat{b}_R(K) \text{sign}(kK) \left| \frac{K}{k} \right| F(-k,-K) - \hat{b}^\dagger_R(K) \text{sign}(kK) \left| \frac{K}{k} \right| F(-k,K) \right] \right. \nonumber \\
& \left. - \sqrt{\left|\frac{K}{k}\right|} \left[ - \hat{b}_R(K) F(-k,-K) + \hat{b}^\dagger_R(K) F(-k,K) \right] \right\rbrace \nonumber \\
= & \int_{-\infty}^{+\infty} \frac{dK}{2} \sqrt{\left|\frac{K}{k}\right|} \left[ \hat{b}_R(K) \text{sign}(kK) F(-k,-K) - \hat{b}^\dagger_R(K) \text{sign}(kK) F(-k,K) \right. \nonumber \\
& \left. + \hat{b}_R(K) F(-k,-K) - \hat{b}^\dagger_R(K) F(-k,K) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dK \sqrt{\left|\frac{K}{k}\right|} \left[ \frac{ 1 + \text{sign}(kK) }{2} F(-k,-K) \hat{b}_R(K) - \frac{ 1 + \text{sign}(kK) }{2} F(-k,K) \hat{b}_R^\dagger(K) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dK \sqrt{\left|\frac{K}{k}\right|} \left[ \theta(kK) F(-k,-K) \hat{b}_R(K) - \theta(kK) F(-k,K) \hat{b}_R^\dagger(K) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dK \left[ \alpha^*(k,K) \hat{b}_R(K) - \beta(k,K) \hat{b}^\dagger_R(K) \right].
\end{align}
Moreover the transformation $a\mapsto-a$ is equivalent to $F(k,K) \mapsto F^*(k,K)$. In this way we can easily prove the following identity from Eq. (\ref{Bogolyubov_transformation_calculation_2})
\begin{equation}\label{Bogolyubov_transformation_calculation_3}
\hat{a}_L(k) = \int_{-\infty}^{+\infty}dK \left[ \alpha(k,K) \hat{b}_L(K) - \beta^*(k,K) \hat{b}^\dagger_L(K) \right].
\end{equation}
Eqs. (\ref{Bogolyubov_transformation_calculation_2}, \ref{Bogolyubov_transformation_calculation_3}) result in Eq. (\ref{Bogolyubov_transformation}).
The inverse equations for Eq. (\ref{Bogolyubov_transformation}) read
\begin{equation}\label{Bogolyubov_transformation_inverse}
b_L(K) = \int_{-\infty}^{+\infty}dk \left[ \alpha^*(k, K)\hat{a}(k) + \beta^*(k,K)\hat{a}^\dagger(k) \right], \quad \hat{b}_R(K) = \int_{-\infty}^{+\infty}dk \left[ \alpha(k, K)\hat{a}(k) + \beta(k,K)\hat{a}^\dagger(k) \right].
\end{equation}
Eq. (\ref{Bogolyubov_transformation_inverse}) will be proven in Sec. \ref{Properties_of_the_Bogolyubov_coefficients}.
Any quantum field in Minkowski space-time must be provided with the following canonical commutation relation
\begin{equation}\label{canonical_commutation_Minkowski}
\left[\hat{a}(k),\hat{a}(k')\right]=0, \quad
\left[\hat{a}(k),\hat{a}^\dagger(k')\right]=\delta(k-k').
\end{equation}
The analogue procedure in the Rindler spacetime consists into imposing the following commutation relations
\begin{equation}\label{canonical_commutation_Rindler}
\left[\hat{b}_L(K),\hat{b}_L(K')\right]=0, \quad \left[\hat{b}_L(K),\hat{b}_L^\dagger(K')\right]=\delta(K-K'), \quad \left[\hat{b}_R(K),\hat{b}_R(K')\right]=0, \quad \left[\hat{b}_R(K),\hat{b}_R^\dagger(K')\right]=\delta(K-K').
\end{equation}
It is possible to prove that the commutation relations (\ref{canonical_commutation_Minkowski}) are compatible with (\ref{canonical_commutation_Rindler}) and give the following additional commutation relations
\begin{equation}\label{canonical_commutation_Rindler_LR}
\left[\hat{b}_L(K),\hat{b}_R(K')\right]=0, \quad \left[\hat{b}_L(K),\hat{b}_R^\dagger(K')\right]=0.
\end{equation}
See Sec. \ref{Properties_of_the_Bogolyubov_coefficients} for such proof.
\section{Properties of the Bogolyubov coefficients}\label{Properties_of_the_Bogolyubov_coefficients}
In Sec. \ref{massless_scalar_fields_in_Rindler_spacetime} we have derived the Bogoliubov transformations relating $\hat{a}(k)$ and $\hat{b}_{L,R}(K)$. The explicit expression for the Bogoliubov coefficients is shown by Eqs. (\ref{Bogolyubov_coefficients}). From such equations we can extract some identities which appear to be useful for some proofs.
For instance, Eq. (\ref{Bogolyubov_transformation_inverse}) can be proven through the following identities
\begin{subequations} \label{identity_k}
\begin{equation} \label{identity_k_3}
\int_{-\infty}^{+\infty}dk [ \alpha(k,K) \alpha(k,K') - \beta(k,K)\beta(k,K') ] = 0,
\end{equation}
\begin{equation} \label{identity_k_4}
\int_{-\infty}^{+\infty}dk [ \alpha(k,K) \beta^*(k,K') - \beta(k,K)\alpha^*(k,K') ] = 0,
\end{equation}
\begin{equation} \label{identity_k_1}
\int_{-\infty}^{+\infty}dk [ \alpha(k,K) \beta(k,K') - \beta(k,K)\alpha(k,K') ] = 0,
\end{equation}
\begin{equation} \label{identity_k_2}
\int_{-\infty}^{+\infty}dk [ \alpha(k,K) \alpha^*(k,K') - \beta(k,K)\beta^*(k,K') ] = \delta(K-K'),
\end{equation}
\end{subequations}
which are valid for any $K, K' \in \mathbb{R}$. Indeed, Eqs. (\ref{identity_k}) and their conjugates result in
\begin{align}
& \int_{-\infty}^{+\infty}dk \left[ \alpha^*(k, K)\hat{a}(k) + \beta^*(k,K)\hat{a}^\dagger(k) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dk \left\lbrace \alpha^*(k, K) \int_{-\infty}^{+\infty}dK' \left[ \alpha(k,K')\hat{b}_L(K') - \beta^*(k,K') \hat{b}^\dagger_L(K') + \alpha^*(k,K') \hat{b}_R(K') - \beta(k,K') \hat{b}^\dagger_R(K') \right] \right. \nonumber \\
& \left. + \beta^*(k,K) \int_{-\infty}^{+\infty}dK' \left[ \alpha^*(k,K')\hat{b}^\dagger_L(K') - \beta(k,K') \hat{b}_L(K') + \alpha(k,K') \hat{b}^\dagger_R(K') - \beta^*(k,K') \hat{b}_R(K') \right] \right\rbrace \nonumber \\
= & \int_{-\infty}^{+\infty}dK' \int_{-\infty}^{+\infty}dk \left[ \alpha^*(k, K) \alpha(k,K')\hat{b}_L(K') - \alpha^*(k, K) \beta^*(k,K') \hat{b}^\dagger_L(K') + \alpha^*(k, K) \alpha^*(k,K') \hat{b}_R(K') \right. \nonumber \\
& - \alpha^*(k, K) \beta(k,K') \hat{b}^\dagger_R(K') + \beta^*(k,K) \alpha^*(k,K')\hat{b}^\dagger_L(K') - \beta^*(k,K) \beta(k,K') \hat{b}_L(K') + \beta^*(k,K) \alpha(k,K') \hat{b}^\dagger_R(K') \nonumber \\
& \left. - \beta^*(k,K) \beta^*(k,K') \hat{b}_R(K') \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dK' \int_{-\infty}^{+\infty}dk \left\lbrace [ \alpha^*(k, K) \alpha(k,K') - \beta^*(k,K) \beta(k,K') ] \hat{b}_L(K') + [ \beta^*(k,K) \alpha^*(k,K') - \alpha^*(k, K) \beta^*(k,K') ] \right. \nonumber \\
& \left. \times \hat{b}^\dagger_L(K') + [\alpha^*(k, K) \alpha^*(k,K')- \beta^*(k,K) \beta^*(k,K')] \hat{b}_R(K') + [ \beta^*(k,K) \alpha(k,K') - \alpha^*(k, K) \beta(k,K')] \hat{b}^\dagger_R(K') \right\rbrace \nonumber \\
= & \int_{-\infty}^{+\infty}dK' \delta(K-K') \hat{b}_L(K') \nonumber \\
= & \hat{b}_L(K)
\end{align}
and
\begin{align}
& \int_{-\infty}^{+\infty}dk \left[ \alpha(k, K)\hat{a}(k) + \beta(k,K)\hat{a}^\dagger(k) \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dk \left\lbrace \alpha(k, K) \int_{-\infty}^{+\infty}dK' \left[ \alpha(k,K')\hat{b}_L(K') - \beta^*(k,K') \hat{b}^\dagger_L(K') + \alpha^*(k,K') \hat{b}_R(K') - \beta(k,K') \hat{b}^\dagger_R(K') \right] \right. \nonumber \\
& \left. + \beta(k,K) \int_{-\infty}^{+\infty}dK' \left[ \alpha^*(k,K')\hat{b}^\dagger_L(K') - \beta(k,K') \hat{b}_L(K') + \alpha(k,K') \hat{b}^\dagger_R(K') - \beta^*(k,K') \hat{b}_R(K') \right] \right\rbrace \nonumber \\
= & \int_{-\infty}^{+\infty}dK' \int_{-\infty}^{+\infty}dk \left[ \alpha(k, K) \alpha(k,K')\hat{b}_L(K') - \alpha(k, K) \beta^*(k,K') \hat{b}^\dagger_L(K') + \alpha(k, K) \alpha^*(k,K') \hat{b}_R(K') \right. \nonumber \\
& - \alpha(k, K) \beta(k,K') \hat{b}^\dagger_R(K') + \beta(k,K) \alpha^*(k,K')\hat{b}^\dagger_L(K') - \beta(k,K) \beta(k,K') \hat{b}_L(K') + \beta(k,K) \alpha(k,K') \hat{b}^\dagger_R(K') \nonumber \\
& \left. - \beta(k,K) \beta^*(k,K') \hat{b}_R(K') \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dK' \int_{-\infty}^{+\infty}dk \left\lbrace [ \alpha(k, K) \alpha(k,K') - \beta(k,K) \beta(k,K') ] \hat{b}_L(K') + [ \beta(k,K) \alpha^*(k,K') - \alpha(k, K) \beta^*(k,K') ] \right. \nonumber \\
& \left. \times \hat{b}^\dagger_L(K') + [\alpha(k, K) \alpha^*(k,K')- \beta(k,K) \beta^*(k,K')] \hat{b}_R(K') + [ \beta(k,K) \alpha(k,K') - \alpha(k, K) \beta(k,K')] \hat{b}^\dagger_R(K') \right\rbrace \nonumber \\
= & \int_{-\infty}^{+\infty}dK' \delta(K-K') \hat{b}_R(K') \nonumber \\
= & \hat{b}_R(K).
\end{align}
Moreover, Eq. (\ref{Bogolyubov_transformation_inverse}) and Eqs. (\ref{identity_k}) can be used in order to prove Eqs. (\ref{canonical_commutation_Rindler}, \ref{canonical_commutation_Rindler_LR}) from Eq. (\ref{canonical_commutation_Minkowski}).
Eqs. (\ref{identity_k}) come from the following chains of identities valid for any $p, q \in \{ 1, -1 \}$
\begin{align} \label{identity_k_5}
&\int_{-\infty}^{+\infty}dk \theta(kK) \theta(kK') \frac{\sqrt{KK'}}{|k|} \left[ F(k,K)F(-pk,-qK') - F(-k,K)F(pk,-qK') \right] \nonumber \\
= & \int_{-\infty}^{+\infty}dk \theta(kK)\theta(KK') \frac{\sqrt{KK'}}{|k|} \frac{1}{(2 \pi a)^2} \Gamma \left( -\frac{i K}{a} \right) \Gamma \left( q \frac{i K'}{a} \right) \left[ \exp \left( i \frac{K-qK'}{a} \ln \frac{|k|}{a} + \text{sign} (k) \frac{\beta}{4} ( K + p q K') \right) \right. \nonumber \\
& \left. - \exp \left( i \frac{K-qK'}{a} \ln \frac{|k|}{a} - \text{sign} (k) \frac{\beta}{4} (K + p q K') \right) \right] \nonumber \\
= & \theta(KK') \sqrt{KK'} \Gamma \left( -\frac{i K}{a} \right) \Gamma \left( q \frac{i K'}{a} \right) \int_{-\infty}^{+\infty}dk \frac{\theta(kK)}{|k|} \frac{1}{(2 \pi a)^2} \left[ \exp \left( i \frac{K-qK'}{a} \ln \frac{|k|}{a} + \text{sign} (K) \frac{\beta}{4} ( K + p q K') \right) \right. \nonumber \\
& \left. - \exp \left( i \frac{K-qK'}{a} \ln \frac{|k|}{a} - \text{sign} (K) \frac{\beta}{4} (K + p q K') \right) \right] \nonumber \\
= & \frac{\theta(KK') \sqrt{KK'}}{2 \pi a} \Gamma \left( -\frac{i K}{a} \right) \Gamma \left( q \frac{i K'}{a} \right) 2 \sinh \left( \text{sign} (K) \frac{\beta}{4} ( K + p q K') \right) \int_{-\infty}^{+\infty}dk \frac{\theta(kK)}{2 \pi a|k|} \exp \left( i \frac{K-qK'}{a} \ln \frac{|k|}{a} \right) \nonumber \\
= & \frac{\theta(KK') \sqrt{KK'}}{ \pi a} \Gamma \left( -\frac{i K}{a} \right) \Gamma \left( q \frac{i K'}{a} \right) \sinh \left( \text{sign} (K) \frac{\beta}{4} ( K + p q K') \right) \int_{0}^{+\infty} \frac{d|k|}{2 \pi a|k|} \exp \left( i \frac{K-qK'}{a} \ln \frac{|k|}{a} \right) \nonumber \\
= & \frac{\theta(KK') \sqrt{KK'}}{\pi a} \Gamma \left( -\frac{i K}{a} \right) \Gamma \left( q \frac{i K'}{a} \right) \sinh \left( \text{sign} (K) \frac{\beta}{4} ( K + p q K') \right) \int_{-\infty}^{+\infty}\frac{d\Xi}{2 \pi} \exp \left( i (K-qK') \Xi \right) \nonumber \\
= & \frac{\theta(KK') \sqrt{KK'}}{\pi a} \Gamma \left( -\frac{i K}{a} \right) \Gamma \left( q \frac{i K'}{a} \right) \sinh \left( \text{sign} (K) \frac{\beta}{4} ( K + p q K') \right) \delta(K-qK') \nonumber \\
= & \frac{\theta(q) |K|}{\pi a} \left| \Gamma \left( \frac{i K}{a} \right) \right|^2 \sinh \left( \text{sign}(K) \frac{\beta}{4} (1+p) K \right) \delta(K-qK') \nonumber \\
= & \theta(q) \left[ \sinh \left( \frac{\beta}{2} |K| \right) \right]^{-1} \sinh \left( \frac{\beta}{4} (1+p) |K| \right) \delta(K-qK'),
\end{align}
where we have replaced the integration variable $|k|$ with
\begin{equation}
\Xi = \frac{1}{a} \ln \left(\frac{|k|}{a}\right)
\end{equation}
and used the following equation for the gamma function
\begin{equation}
|\Gamma(i z)|^2 = \frac{\pi}{|z| \sinh(\pi |z|)},
\end{equation}
which is valid for any real $z$. For $p=-1$ Eq. (\ref{identity_k_5}) gives
\begin{equation}
\int_{-\infty}^{+\infty}dk \theta(kK) \theta(kK') \frac{\sqrt{KK'}}{|k|} \left[ F(k,K)F(k,-qK') - F(-k,K)F(-k,-qK') \right] = 0,
\end{equation}
resulting in Eqs. (\ref{identity_k_3}, \ref{identity_k_4}), while for $p=1$ Eq. (\ref{identity_k_5}) gives
\begin{equation} \label{identity_k_6}
\int_{-\infty}^{+\infty}dk \theta(kK) \theta(kK') \frac{\sqrt{KK'}}{|k|} \left[ F(k,K)F(-k,-qK') - F(-k,K)F(k,-qK') \right] = \theta(q) \delta(K-qK'),
\end{equation}
For $q=-1$ Eq. (\ref{identity_k_6}) gives
\begin{equation}
\int_{-\infty}^{+\infty}dk \theta(kK) \theta(kK') \frac{\sqrt{KK'}}{|k|} \left[ F(k,K)F(-k,K') - F(-k,K)F(k,K') \right] = 0,
\end{equation}
resulting in Eq. (\ref{identity_k_1}), while for $q=1$ Eq. (\ref{identity_k_6}) gives
\begin{equation}
\int_{-\infty}^{+\infty}dk \theta(kK) \theta(kK') \frac{\sqrt{KK'}}{|k|} \left[ F(k,K)F(-k,-K') - F(-k,K)F(k,-K') \right] = \delta(K-K'),
\end{equation}
resulting in Eq. (\ref{identity_k_2}).
Other important identities relating $\alpha(k,K)$ and $\beta(k,K)$ are the following:
\begin{equation} \label{alpha_to_beta}
\alpha(k,K) = \exp \left( \frac{\beta}{2} |K| \right) \beta(k,K) ,
\end{equation}
\begin{equation} \label{identity_K}
\int_{-\infty}^{+\infty}dK 2 \sinh \left( \frac{\beta}{2} |K| \right) \left[ \alpha(k,K)\beta^*(k',K) + \beta^*(k,K)\alpha(k',K) \right] = \delta(k-k').
\end{equation}
Eq. (\ref{alpha_to_beta}) comes from the following identity
\begin{equation} \label{alpha_beta_to_alpha_beta_proof}
F(k,K) = \exp \left( \text{sign}(k) \frac{\beta}{2} K \right) F(-k,K),
\end{equation}
which can be directly extracted from Eq. (\ref{F}). On the other hand, Eq. (\ref{identity_K}) can be proven by the following chain of identities:
\begin{align}
& \int_{-\infty}^{+\infty}dK \theta(kK) \theta(k'K) 2 \sinh \left( \frac{\beta}{2} |K| \right) \frac{|K|}{\sqrt{kk'}} \left[ F(k,K)F(k',-K) + F(k,-K)F(k',K) \right] \nonumber \\
= & \frac{\theta(kk')}{\sqrt{kk'}} \left[ \int_{-\infty}^{+\infty}dK \theta(kK) 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| F(k,K)F(k',-K) \right. \nonumber \\
& \left. + \int_{-\infty}^{+\infty}dK \theta(kK) 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| F(k,-K)F(k',K) \right] \nonumber \\
= & \frac{\theta(kk')}{\sqrt{kk'}} \left[ \int_{-\infty}^{+\infty}dK \theta(kK) 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| F(k,K)F(k',-K) \right. \nonumber \\
& \left. + \int_{-\infty}^{+\infty}dK \theta(-kK) 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| F(k,K)F(k',-K) \right] \nonumber \\
= & \frac{\theta(kk')}{\sqrt{kk'}} \int_{-\infty}^{+\infty}dK \left[ \theta(kK)+\theta(-kK) \right] 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| F(k,K)F(k',-K) \nonumber \\
= & \frac{\theta(kk')}{\sqrt{kk'}} \int_{-\infty}^{+\infty}dK 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| F(k,K)F(k',-K) \nonumber \\
= & \frac{\theta(kk')}{\sqrt{kk'}} \int_{-\infty}^{+\infty}dK 2 \sinh \left( \frac{\beta}{2} |K| \right) \frac{|K|}{(2 \pi a)^2} \left| \Gamma \left( \frac{i K}{a} \right) \right|^2 \exp \left( i \frac{K}{a} \ln \left| \frac{k}{k'} \right| \right) \nonumber \\
= & \frac{\theta(kk')}{\sqrt{kk'}} \int_{-\infty}^{+\infty} \frac{dK}{2 \pi a} \exp \left( i \frac{K}{a} \ln \left| \frac{k}{k'} \right| \right)\nonumber \\
= & \frac{\theta(kk')}{\sqrt{kk'}} \delta \left( \ln \left| \frac{k}{k'} \right| \right) \nonumber \\
= & \frac{\theta(kk')}{|k|} \delta \left( \ln \left| \frac{k}{k'} \right| \right) \nonumber \\
= & \theta(kk') \delta \left( |k| - |k'| \right)\nonumber \\
= & \delta \left( k - k' \right).
\end{align}
\section{Minkowski vacuum states}
In this section, we give a brief review of the Minkowski vacuum states seen by an accelerated observer. The aim is to provide a simple proof for the Rindler effect through some identities shown in the previous sections. Moreover, we provide a proof for Eq. (5) of the main paper and for some other identities which will be used in the next sections.
The Minkowski vacuum states is defined by
\begin{equation} \label{vacuum_state_H_M_in_tilde_H'_M_epsilon}
\hat{a}(k) |0_M\rangle = 0, \quad \forall k \in \mathbb{R}.
\end{equation}
Eq. (\ref{vacuum_state_H_M_in_tilde_H'_M_epsilon}) supplemented with Eqs. (\ref{Bogolyubov_transformation}, \ref{alpha_to_beta}) gives
\begin{equation}
\int_{-\infty}^{+\infty}dK \left\lbrace\beta(k,K) \left[ \exp \left( \frac{\beta}{2} |K| \right) \hat{b}_L(K) - \hat{b}^\dagger_R(K) \right] + \beta^*(k,K) \left[\exp \left( \frac{\beta}{2} |K| \right) \hat{b}_R(K) - \hat{b}^\dagger_L(K) \right] \right\rbrace|0_M\rangle = 0,
\end{equation}
which holds for any real $k$ and resulting, therefore, in Eq. (5).
A non-normalizable state which satisfy Eq. (5) exists and its explicit expression is the following
\begin{equation} \label{vacuum_state_H_M_in_H'_M}
|0_M\rangle \propto \hat{S} | 0_L, 0_R \rangle,
\end{equation}
with
\begin{equation} \label{D}
\hat{S} = \exp \left( \int_{-\infty}^{+\infty} dK \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_L(K) \hat{b}^\dagger_R(K) \right).
\end{equation}
Eq. (\ref{vacuum_state_H_M_in_H'_M}) is identical to Eq. (1) of the main paper and, therefore, it results in the Unruh effect. The fact that the $|0_M\rangle$ in Eq. (\ref{vacuum_state_H_M_in_H'_M}) satisfies Eqs. (5) can be proven in the following way
\begin{align}
\left[ \hat{b}_{L,R}(K) , \hat{S} \right] = & \left[ \hat{b}_{L,R}(K) , \hat{1} +\int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) + \frac{1}{2}\int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) \right. \nonumber \\
& \left. \times \int_{-\infty}^{+\infty} dK_2 \exp \left( -\frac{\beta}{2} |K_2| \right) \hat{b}^\dagger_L(K_2) \hat{b}^\dagger_R(K_2) + \sum_{n=3}^{\infty} \frac{1}{n!} \prod_{i=1}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \right] \nonumber \\
= & \left[ \hat{b}_{L,R}(K) , \hat{1} \right] + \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \left[ \hat{b}_{L,R}(K) , \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) \right] + \frac{1}{2} \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \nonumber \\
& \times \left[ \hat{b}_{L,R}(K) ,\hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1)\right] \int_{-\infty}^{+\infty} dK_2 \exp \left( -\frac{\beta}{2} |K_2| \right) \hat{b}^\dagger_L(K_2) \hat{b}^\dagger_R(K_2) + \frac{1}{2} \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \nonumber \\
& \times \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) \int_{-\infty}^{+\infty} dK_2 \exp \left( -\frac{\beta}{2} |K_2| \right) \left[ \hat{b}_{L,R}(K) ,\hat{b}^\dagger_L(K_2) \hat{b}^\dagger_R(K_2) \right] + \sum_{n=3}^{\infty} \frac{1}{n!} \left\lbrace \int_{-\infty}^{+\infty} dK_1 \right. \nonumber \\
& \times \exp \left( -\frac{\beta}{2} |K_1| \right)\left[ \hat{b}_{L,R}(K) , \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) \right] \prod_{i=2}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \nonumber \\
& + \sum_{j=2}^{n-1} \prod_{i=1}^{j-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \int_{-\infty}^{+\infty} dK_j \exp \left( -\frac{\beta}{2} |K_j| \right) \left[ \hat{b}_{L,R}(K) , \hat{b}^\dagger_L(K_j) \right. \nonumber \\
& \left. \times \hat{b}^\dagger_R(K_j) \right] \prod_{i=j+1}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) + \prod_{i=1}^{n-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \nonumber \\
& \left. \times \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \int_{-\infty}^{+\infty} dK_n \exp \left( -\frac{\beta}{2} |K_n| \right) \left[ \hat{b}_{L,R}(K) , \hat{b}^\dagger_L(K_n) \hat{b}^\dagger_R(K_n) \right] \right\rbrace\nonumber \\
= & \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_{R,L}(K_1) \delta(K-K_1) + \frac{1}{2} \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_{R,L}(K_1) \delta(K-K_1) \nonumber \\
& \times\int_{-\infty}^{+\infty} dK_2 \exp \left( -\frac{\beta}{2} |K_2| \right) \hat{b}^\dagger_L(K_2) \hat{b}^\dagger_R(K_2) + \frac{1}{2} \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) \int_{-\infty}^{+\infty} dK_2 \nonumber \\
& \times \exp \left( -\frac{\beta}{2} |K_2| \right) \hat{b}^\dagger_{R,L}(K_2) \delta(K-K_2) + \sum_{n=3}^{\infty} \frac{1}{n!} \left[ \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_{R,L}(K_1) \delta(K-K_1) \right. \nonumber \\
& \times \prod_{i=2}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) + \sum_{j=2}^{n-1} \prod_{i=1}^{j-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \nonumber \\
& \times \int_{-\infty}^{+\infty} dK_j \exp \left( -\frac{\beta}{2} |K_j| \right) \hat{b}^\dagger_{R,L}(K_j) \delta(K-K_j) \prod_{i=j+1}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \nonumber \\
& \left. + \prod_{i=1}^{n-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \int_{-\infty}^{+\infty} dK_n \exp \left( -\frac{\beta}{2} |K_n| \right) \hat{b}^\dagger_{R,L}(K_n) \delta(K-K_n) \right]\nonumber \\
= & \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) + \frac{1}{2} \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) \int_{-\infty}^{+\infty} dK_2 \exp \left( -\frac{\beta}{2} |K_2| \right) \hat{b}^\dagger_L(K_2) \hat{b}^\dagger_R(K_2) \nonumber \\
& + \frac{1}{2} \int_{-\infty}^{+\infty} dK_1 \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) + \sum_{n=3}^{\infty} \frac{1}{n!} \left[ \exp \left( -\frac{\beta}{2} |K| \right) \right.\nonumber \\
& \times \hat{b}^\dagger_{R,L}(K) \prod_{i=2}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) + \sum_{j=2}^{n-1} \prod_{i=1}^{j-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \nonumber \\
& \times \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) \prod_{i=j+1}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \nonumber \\
& \left. + \prod_{i=1}^{n-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) \right] \nonumber \\
= & \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) \left\lbrace 1 + \frac{1}{2} \int_{-\infty}^{+\infty} dK_2 \exp \left( -\frac{\beta}{2} |K_2| \right) \hat{b}^\dagger_L(K_2) \hat{b}^\dagger_R(K_2) + \frac{1}{2} \int_{-\infty}^{+\infty} dK_1 \right. \nonumber \\
& \times \exp \left( -\frac{\beta}{2} |K_1| \right) \hat{b}^\dagger_L(K_1) \hat{b}^\dagger_R(K_1) + \sum_{n=3}^{\infty} \frac{1}{n!} \left[ \prod_{i=2}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \right.\nonumber \\
& + \sum_{j=2}^{n-1} \prod_{i=1}^{j-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \prod_{i=j+1}^{n} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \nonumber \\
& \left. \left. + \prod_{i=1}^{n-1} \int_{-\infty}^{+\infty} dK_i \exp \left( -\frac{\beta}{2} |K_i| \right) \hat{b}^\dagger_L(K_i) \hat{b}^\dagger_R(K_i) \right] \right\rbrace \nonumber \\
= & \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) \left\lbrace 1 + \int_{-\infty}^{+\infty} dK' \exp \left( -\frac{\beta}{2} |K'| \right) \hat{b}^\dagger_L(K') \hat{b}^\dagger_R(K') \right. \nonumber \\
& \left. + \sum_{n=3}^{\infty} \frac{n}{n!} \left[ \int_{-\infty}^{+\infty} dK' \exp \left( -\frac{\beta}{2} |K'| \right) \hat{b}^\dagger_L(K') \hat{b}^\dagger_R(K') \right]^{n-1} \right\rbrace \nonumber \\
= & \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) \sum_{n=1}^{\infty} \frac{1}{(n-1)!} \left[ \int_{-\infty}^{+\infty} dK' \exp \left( -\frac{\beta}{2} |K'| \right) \hat{b}^\dagger_L(K') \hat{b}^\dagger_R(K') \right]^{n-1} \nonumber \\
= & \exp \left( -\frac{\beta}{2} |K| \right) \hat{b}^\dagger_{R,L}(K) \hat{S}.
\end{align}
Therefore, we obtain
\begin{align} \label{from_left_to_right}
\hat{b}^\dagger_{R,L}(K) \hat{S} | 0_L, 0_R \rangle = & \exp \left( \frac{\beta}{2} |K| \right) \left[ \hat{b}_{L,R}(K) , \hat{S} \right] | 0_L, 0_R \rangle \nonumber \\
= & \exp \left( \frac{\beta}{2} |K| \right) \hat{b}_{L,R}(K) \hat{S} | 0_L, 0_R \rangle,
\end{align}
which means that the $|0_M\rangle$ of Eq. (\ref{vacuum_state_H_M_in_H'_M}) is solution of Eqs. (5).
Eq. (5) can be used to obtain useful identities. For instance, it is possible to move creation $\hat{b}_R^\dagger(K)$ and annihilation $\hat{b}_R(K)$ operators acting from the left of $\hat{\rho}_0$ to its right and the other way round using the following identity and its adjoin
\begin{equation}\label{ordering_Rindler}
\hat{b}^\dagger_R(K) \hat{\rho}_0 = e^{\beta |K|} \hat{\rho}_0 \hat{b}^\dagger_R(K).
\end{equation}
Eq. (\ref{ordering_Rindler}) is a result of Eq. (5):
\begin{align}
\hat{b}^\dagger_R(K) \hat{\rho}_0 = & \text{Tr}_L \left[ \hat{b}^\dagger_R(K) |0_M\rangle \langle 0_M| \right] \nonumber \\
= & \exp \left( \frac{1}{2} \beta |K| \right) \text{Tr}_L \left[ \hat{b}_L(K) |0_M\rangle \langle 0_M| \right] \nonumber \\
= & \exp \left( \frac{1}{2} \beta |K| \right) \text{Tr}_L \left[ |0_M\rangle \langle 0_M| \hat{b}_L(K) \right]\nonumber \\
= & e^{\beta |K|} \text{Tr}_L \left[ |0_M\rangle \langle 0_M| \hat{b}_R^\dagger(K) \right]\nonumber \\
= & e^{\beta |K|} \hat{\rho}_0 \hat{b}^\dagger_R(K).
\end{align}
Eq. (\ref{ordering_Rindler}) gives also the following identity
\begin{align}\label{ordering_Rindler_2}
\hat{b}^\dagger_R(K)\hat{\rho}_0 \hat{b}_R(K') = & e^{\beta |K'|} \hat{b}^\dagger_R(K) \hat{b}_R(K') \hat{\rho}_0 \nonumber \\
= & e^{\beta |K'|} \left[ \hat{b}^\dagger_R(K), \hat{b}_R(K') \right] \hat{\rho}_0 + e^{\beta |K'|} \hat{b}_R(K') \hat{b}^\dagger_R(K) \hat{\rho}_0 \nonumber \\
= & - e^{\beta |K'|} \delta(K-K') \hat{\rho}_0 + e^{\beta (|K|+|K'|)} \hat{b}_R(K') \hat{\rho}_0 \hat{b}^\dagger_R(K),
\end{align}
which will be used in Sec. \ref{Characteristic_function}.
\section{A proof for Eq. (3)}
The main aim of this section is to provide a proof of Eq. (3) through some identities obtained in the previous sections.
Eqs. (5, \ref{Bogolyubov_transformation}, \ref{alpha_to_beta}) can be used in order to derive the following identity
\begin{align} \label{momentum_state}
\hat{a}^\dagger (k) |0_M\rangle = & \int_{-\infty}^{+\infty}dK \left[\alpha^*(k,K)\hat{b}^\dagger_L(K) - \beta(k,K) \hat{b}_L(K) + \alpha(k,K) \hat{b}^\dagger_R(K) - \beta^*(k,K) \hat{b}_R(K) \right] |0_M\rangle \nonumber \\
= & \int_{-\infty}^{+\infty}dK \left[ \exp \left( \frac{\beta}{2} |K| \right) \alpha^*(k,K)\hat{b}_R(K) - \exp \left(-\frac{\beta}{2} |K| \right) \beta(k,K) \hat{b}^\dagger_R(K) \right. \nonumber\\
& \left. + \exp \left( \frac{\beta}{2} |K| \right) \beta(k,K) \hat{b}^\dagger_R(K) - \exp \left(-\frac{\beta}{2} |K| \right) \alpha^*(k,K) \hat{b}_R(K) \right] |0_M\rangle \nonumber\\
= & \int_{-\infty}^{+\infty}dK \left[ 2 \sinh \left( \frac{\beta}{2} |K| \right) \alpha^*(k,K)\hat{b}_R(K) + 2 \sinh \left( \frac{\beta}{2} |K| \right) \beta(k,K) \hat{b}^\dagger_R(K) \right] |0_M\rangle \nonumber\\
= & \int_{-\infty}^{+\infty}dK 2 \sinh \left( \frac{\beta}{2} |K| \right) \left[ \alpha^*(k,K)\hat{b}_R(K) + \beta(k,K) \hat{b}^\dagger_R(K) \right] |0_M\rangle \nonumber\\
= &\int_{-\infty}^{+\infty}dK 2 \sinh \left( \frac{\beta}{2} |K| \right) \theta(kK) \sqrt{\frac{K}{k}} \left[ F(-k,-K)\hat{b}_R(K) + F(-k,K) \hat{b}^\dagger_R(K) \right] |0_M\rangle
\end{align}
Eqs. (2, \ref{momentum_state}) result in
\begin{equation} \label{state}
|\psi\rangle = \int_{-\infty}^{+\infty} dK \left[ \tilde{\psi}_-(K) \hat{b}_R(K) + \tilde{\psi}_+(K) \hat{b}^\dagger_R(K) \right] |0_M\rangle ,
\end{equation}
with
\begin{equation} \label{psi_tilde}
\tilde{\psi}_\pm(K) = 2 \sinh \left( \frac{\beta}{2} |K| \right) \int_{-\infty}^{+\infty} dk \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2\pi}} e^{-ikx} \theta(kK) \sqrt{\frac{K}{k}} F(-k, \pm K).
\end{equation}
It is straightforward to obtain Eq. (3) from Eq. \ref{state} if we prove that the functions $\tilde{\psi}_\pm(K)$ defined in Eq. (\ref{psi_tilde}) are identical to Eq. (4a). This can be seen in the following way
\begin{align}
\tilde{\psi}_\pm(K) = & 2 \sinh \left( \frac{\beta}{2} |K| \right) \int_{-\infty}^{+\infty} dk \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} e^{-ikx} \theta(kK) \sqrt{\frac{K}{k}} \frac{1}{2 \pi a} \Gamma \left(\mp \frac{i K}{a} \right) \exp \left(\pm i \frac{K}{a} \ln \frac{|k|}{a} \mp \text{sign}(k) \frac{\beta}{4} K \right) \nonumber \\
= & 2 \sqrt{|K|} \sinh \left( \frac{\beta}{2} |K| \right) \Gamma \left(\mp \frac{i K}{a} \right) \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} \int_{-\infty}^{+\infty} dk \frac{e^{-i\text{sign}(k)|k|x}}{2 \pi a} \frac{\theta(kK)}{ \sqrt{|k|}} \exp \left(\pm i \frac{K}{a} \ln \frac{|k|}{a} \mp \text{sign}(k) \frac{\beta}{4} K \right) \nonumber \\
= & 2 \sqrt{|K|} \sinh \left( \frac{\beta}{2} |K| \right) \Gamma \left(\mp \frac{i K}{a} \right) \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} \int_{0}^{+\infty} d|k| \frac{e^{-i\text{sign}(K)|k|x}}{2 \pi a \sqrt{|k|}} \exp \left(\pm i \frac{K}{a} \ln \frac{|k|}{a} \mp \text{sign}(K) \frac{\beta}{4} K \right) \nonumber \\
= & 2 \sqrt{|K|} \sinh \left( \frac{\beta}{2} |K| \right) \Gamma \left(\mp \frac{i K}{a} \right) \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} \int_{-\infty}^{+\infty} d\Xi\frac{\sqrt{a}}{2 \pi} \exp \left( \frac{a}{2} \Xi -i \text{sign}(K) a e^{a \Xi} x \pm i K \Xi \mp \frac{\beta}{4} |K| \right) \nonumber \\
= & 2 \sqrt{a|K|} \sinh \left( \frac{\beta}{2} |K| \right) \Gamma \left(\mp \frac{i K}{a} \right) \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} F \left( - \text{sign}(K) a^2 x,\mp K + i \frac{a}{2} \right) \exp \left(\mp \frac{\beta}{4} |K| \right) \nonumber \\
= & 2 \sqrt{a|K|} \sinh \left( \frac{\beta}{2} |K| \right) \Gamma \left(\mp \frac{i K}{a} \right) \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} \frac{1}{2 \pi a} \Gamma \left(\frac{1}{2} \pm \frac{i K}{a} \right) \exp \left( \left(\mp i \frac{K}{a} - \frac{1}{2} \right) \ln |a x| \pm \text{sign} ( x K) \frac{\beta}{4} K \right. \nonumber \\
& \left. - i \text{sign}(x K) \frac{\pi}{4} \mp \frac{\beta}{4} |K|\right) \nonumber \\
= & \frac{\sqrt{|K|}}{\pi \sqrt{a}} \sinh \left( \frac{\beta}{2} |K| \right) \Gamma \left(\mp \frac{i K}{a} \right) \Gamma \left(\frac{1}{2} \pm \frac{i K}{a}\right) \int_{-\infty}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} \exp \left( \left(\mp i \frac{K}{a} - \frac{1}{2} \right) \ln |a x| \pm \text{sign} ( x) \frac{\beta}{4} |K| \right. \nonumber \\
& \left. - i \text{sign}(x K) \frac{\pi}{4} \mp \frac{\beta}{4} |K|\right) \nonumber \\
= & \frac{\sqrt{|K|}}{\pi \sqrt{a}} \sinh \left( \frac{\beta}{2} |K| \right) \Gamma \left(\mp \frac{i K}{a} \right) \Gamma \left(\frac{1}{2} \pm \frac{i K}{a}\right) \left[ \int_{0}^{+\infty} dx \frac{\psi(x)}{\sqrt{2 \pi}} \exp \left( \left(\mp i \frac{K}{a} - \frac{1}{2} \right) \ln |a x| - i \text{sign}(K) \frac{\pi}{4}\right) \right. \nonumber \\
& \left. + \int_{-\infty}^{0} dx \frac{\psi(x)}{\sqrt{2 \pi}} \exp \left( \left(\mp i \frac{K}{a} - \frac{1}{2} \right) \ln |a x| + i \text{sign}(K) \frac{\pi}{4} \mp \frac{\beta}{2} |K|\right) \right] \nonumber \\
= & 2 \sinh \left( \frac{\beta}{2} |K| \right) \exp \left( \mp \frac{\beta}{2} |K|\right) \left\lbrace \left[ \theta(\pm 1) + \tilde{f}_{R \pm}\left( \mp \frac{K}{a} \right) \right] \int_{-\infty}^{+\infty} dX \frac{\psi(x_R(X))}{\sqrt{2 \pi}} \exp \left( \left(\mp i K + \frac{a}{2} \right) X \right) \right. \nonumber \\
& \left. + \tilde{f}_{L \pm}\left( \mp \frac{K}{a} \right) \int_{-\infty}^{+\infty} dX \frac{\psi(x_L(-X))}{\sqrt{2 \pi}} \exp \left( \left(\mp i K + \frac{a}{2} \right) X \right) \right\rbrace \nonumber \\
= & \frac{e^{-\theta(\pm 1)\beta |K|}}{n_0(K)} \left\lbrace \left[ \theta(\pm 1) + \tilde{f}_{R \pm}\left( \mp \frac{K}{a} \right) \right] \int_{-\infty}^{+\infty} dX \psi_R(X) \frac{e^{ \mp i K X }}{\sqrt{2 \pi}} + \tilde{f}_{L \pm}\left( \mp \frac{K}{a} \right) \int_{-\infty}^{+\infty} dX \psi_L(-X) \frac{e^{\mp i K X}}{\sqrt{2 \pi}} \right\rbrace \nonumber \\
= & \frac{e^{-\theta(\pm 1)\beta |K|}}{n_0(K)} \int_{-\infty}^{+\infty} dX \frac{e^{\mp i K X}}{\sqrt{2 \pi}} \left\lbrace \psi_R(X) \left[ \theta(\pm 1) + \tilde{f}_{R \pm}\left( \mp \frac{K}{a} \right) \right] + \psi_L(-X) \tilde{f}_{L \pm}\left( \mp \frac{K}{a} \right) \right\rbrace.
\end{align}
\section{Normalization conditions for single-particle states}
In this section we provide equivalent ways to express the normalization condition $\langle \psi | \psi \rangle = 1$ for different definitions of wave-functions. Moreover, we discuss the case of $\bar{\psi}(x)$ as the wave-function which transforms as a scalar field under Lorentz transformations. We show the explicit normalization condition for $\bar{\psi}(x)$ and a proof for the fact that left-wedge values of $\bar{\psi}(x)$ appear in $\Delta n_R(X)$.
The wave-function $\psi(x)$ is normalized in the following way.
\begin{equation} \label{normalization_x}
\int_{-\infty}^{+\infty} dx \left| \psi(x)\right|^2 = 1.
\end{equation}
Equivalently, the normalization condition $\langle \psi | \psi \rangle = 1$ can be imposed on the wave-function in the momentum space defined as
\begin{equation}
\tilde{\psi}(k) = \int_{-\infty}^{+\infty} dx \frac{e^{- i k x}}{\sqrt{2 \pi}} \psi(x).
\end{equation}
The normalization condition for $\tilde{\psi}(k)$ writes
\begin{equation} \label{normalization_k}
\int_{-\infty}^{+\infty} dk \left|\tilde{\psi}(k)\right|^2 = 1.
\end{equation}
Eq. (\ref{normalization_x}) can be written in terms of $\psi_R (X)$ and $\psi_L(X)$
\begin{equation} \label{normalization_x_2}
P_L + P_R = 1,
\end{equation}
where
\begin{equation}
P_{L,R} = \int_{-\infty}^{+\infty} dX \left| \psi_{L,R}(X)\right|^2
\end{equation}
are associated to the probability of finding the particle in the left and in the right wedge.
Eq. (\ref{normalization_k}) can be written in terms of $\tilde{\psi}_\pm(K)$
\begin{equation} \label{normalization_2}
\int_{-\infty}^{+\infty} dK n_0(K) \left[ \left| \tilde{\psi}_+(K) \exp \left( \frac{\beta}{2}|K| \right) \right|^2 + \left|\tilde{\psi}_-(K)\right|^2 \right] = 1.
\end{equation}
Eq. (\ref{normalization_2}) can be obtained from Eqs. (\ref{identity_K}, \ref{psi_tilde}) and from the following identity
\begin{align} \label{psi_tilde_3}
\tilde{\psi}_\pm(K) \exp \left( \pm \frac{\beta}{2} |K| \right) = & 2 \sinh \left( \frac{\beta}{2} |K| \right) \int_{-\infty}^{+\infty} dk \tilde{\psi}(k) \theta(kK) \sqrt{\frac{K}{k}} F(k, \pm K),
\end{align}
which in turn can be directly obtained from Eqs. (\ref{F}, \ref{psi_tilde}). The proof for Eq. (\ref{normalization_2}) is the following
\begin{align}
& \int_{-\infty}^{+\infty} dK n_0(K) \left[ \left| \tilde{\psi}_+(K) \exp \left( \frac{\beta}{2}|K| \right) \right|^2 + \left|\tilde{\psi}_-(K)\right|^2 \right] \nonumber \\
= & \int_{-\infty}^{+\infty} dK \exp \left( -\frac{\beta}{2}|K| \right) \left[ 2 \sinh\left( \frac{\beta}{2}|K| \right) \right]^{-1} \left[ \left| \tilde{\psi}_+(K) \exp \left( \frac{\beta}{2}|K| \right) \right|^2 + \left|\tilde{\psi}_-(K)\right|^2 \right] \nonumber \\
= & \int_{-\infty}^{+\infty} dK \left[ 2 \sinh\left( \frac{\beta}{2}|K| \right) \right]^{-1} \left[ \tilde{\psi}_+(K) \exp \left( \frac{\beta}{2}|K| \right) \tilde{\psi}^*_+(K) + \tilde{\psi}_-(K) \exp \left( -\frac{\beta}{2}|K| \right) \tilde{\psi}^*_-(K) \right]\nonumber \\
= & \int_{-\infty}^{+\infty} dK 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| \int_{-\infty}^{+\infty} dk \frac{\tilde{\psi}(k)}{\sqrt{|k|}} \theta(kK) \int_{-\infty}^{+\infty} dk' \frac{\tilde{\psi}^*(k')}{\sqrt{|k'|}} \theta(k'K) [F(k, K) F(k', - K) + F(k, - K) F(k', K)]\nonumber \\
= & \int_{-\infty}^{+\infty} dk \tilde{\psi}(k) \int_{-\infty}^{+\infty} dk' \tilde{\psi}^*(k') \int_{-\infty}^{+\infty} dK 2 \sinh \left( \frac{\beta}{2} |K| \right) |K| \frac{\theta(kK) \theta(k'K)}{\sqrt{|k k'|}} [F(k, K) F(k', - K) + F(k, - K) F(k', K)]\nonumber \\
= & \int_{-\infty}^{+\infty} dk \tilde{\psi}(k) \int_{-\infty}^{+\infty} dk' \tilde{\psi}^*(k') \delta(k-k') \nonumber \\
= & \int_{-\infty}^{+\infty} dk \left| \tilde{\psi}(k) \right|^2 \nonumber \\
= & 1.
\end{align}
$\psi(x)$ does not behave as a scalar field under Lorentz transformations. Nevertheless, we can use the definition of a Lorentz-invariant wave-function $\bar{\psi}(x)$ through the following expression
\begin{equation} \label{one_state_invariant}
|\psi\rangle = \int_{-\infty}^{+\infty} dx \bar{\psi}(x) \hat{\phi} (0,x) |0_M\rangle.
\end{equation}
The downside is that the normalization condition for $\bar{\psi}(x)$ is not provided with the interpretation of $\left|\bar{\psi}(x)\right|^2$ as the probability density of finding a particle at position $x$ \cite{Localized}.
A possible way to switch between $\tilde{\psi}(k)$ and $\bar{\psi}(x)$ can be obtained from Eqs. (2, \ref{one_state_invariant}). Indeed, Eq. (2) is equivalent to
\begin{align} \label{one_state_alt}
|\psi\rangle = & \int_{-\infty}^{+\infty} dx \psi(x) \hat{\tilde{a}}^\dagger (x) |0_M\rangle \nonumber\\
= & \int_{-\infty}^{+\infty} dx \psi(x) \int_{-\infty}^{+\infty} dk \frac{e^{-i k x}}{\sqrt{2 \pi}} \hat{a}^\dagger (k) |0_M\rangle \nonumber\\
= & \int_{-\infty}^{+\infty} dk \tilde{\psi}(k) \hat{a}^\dagger (k) |0_M\rangle,
\end{align}
while Eq. (\ref{one_state_invariant})
\begin{align} \label{one_state_invariant_alt}
|\psi\rangle = & \int_{-\infty}^{+\infty} dx \bar{\psi}(x) \hat{\phi} (0,x) |0_M\rangle \nonumber\\
= & \int_{-\infty}^{+\infty} dx \bar{\psi}(x) \int_{-\infty}^{+\infty} \frac{dk}{\sqrt{2\pi|k|}} \left[ e^{ikx} \hat{a}(k) + e^{-ikx}\hat{a}^\dagger (k) \right] |0_M\rangle \nonumber\\
= & \int_{-\infty}^{+\infty} dx \bar{\psi}(x) \int_{-\infty}^{+\infty} \frac{dk}{\sqrt{2\pi|k|}} e^{-ikx}\hat{a}^\dagger (k) |0_M\rangle \nonumber\\
= & \int_{-\infty}^{+\infty} dk \int_{-\infty}^{+\infty} dx \frac{e^{-ikx}}{\sqrt{2\pi|k|}} \bar{\psi}(x) \hat{a}^\dagger (k) |0_M\rangle.
\end{align}
By comparing Eq.(\ref{one_state_alt}) with Eq. (\ref{one_state_invariant_alt}), we obtain the following identity relating $\tilde{\psi}(k)$ and $\bar{\psi}(x)$:
\begin{equation}\label{one_state_invariant_2}
\tilde{\psi}(k) = \int_{-\infty}^{+\infty} dx \frac{e^{-ikx}}{\sqrt{2\pi|k|}} \bar{\psi}(x).
\end{equation}
Eqs. (\ref{normalization_k}, \ref{one_state_invariant_2}) result in the following normalization condition for $\bar{\psi}(x)$
\begin{equation}\label{one_state_invariant_normalization}
\int_{-\infty}^{+\infty} dk \left| \int_{-\infty}^{+\infty} dx \frac{e^{-ikx}}{\sqrt{2\pi|k|}} \bar{\psi}(x) \right|^2 = 1,
\end{equation}
which makes Gaussian functions non-normalizable. Indeed a necessary condition for Eq. (\ref{one_state_invariant_normalization}) is that the absolute value of the Fourier transform of $\bar{\psi}(x)$ is negligible at the origin, which is an impossible condition for Gaussian functions.
The fact that left-wedge values of $\bar{\psi}(x)$ appear in $\Delta n_R(X)$ -- as we have stated in the main paper -- can be proven in the following way:
\begin{align}\label{state_invariant}
| \psi \rangle = & \int_{-\infty}^{+\infty} dx \bar{\psi}(x) \hat{\phi}(x) | 0_M \rangle \nonumber\\
= & \int_{-\infty}^{0} dx \bar{\psi}(x) \hat{\phi}(x) | 0_M \rangle + \int_{0}^{+\infty} dx \bar{\psi}(x) \hat{\phi}(x) | 0_M \rangle \nonumber\\
= & \int_{-\infty}^{0} dx \bar{\psi}(x) \hat{\Phi}_L(0,X_L(x)) | 0_M \rangle + \int_{0}^{+\infty} dx \bar{\psi}(x) \hat{\Phi}_R(0,X_R(x)) | 0_M \rangle \nonumber\\
= & \int_{-\infty}^{0} dx \bar{\psi}(x) \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} \left[ e^{iKX_L(x)} \hat{b}_L(K) + e^{-iKX_L(x)} \hat{b}_L^\dagger(K) \right] | 0_M \rangle \nonumber\\
& + \int_{0}^{+\infty} dx \bar{\psi}(x) \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} \left[ e^{iKX_R(x)} \hat{b}_R(K) + e^{-iKX_R(x)} \hat{b}_R^\dagger(K) \right] | 0_M \rangle \nonumber\\
= & \int_{0}^{+\infty} dx \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} \left\lbrace \bar{\psi}(-x) \left[ e^{-iKX_R(x)} \hat{b}_L(K) + e^{iKX_R(x)} \hat{b}_L^\dagger(K) \right] \right. \nonumber \\
& \left. +\bar{\psi}(x) \left[ e^{iKX_R(x)} \hat{b}_R(K) + e^{-iKX_R(x)} \hat{b}_R^\dagger(K) \right] \right\rbrace | 0_M \rangle \nonumber\\
= & \int_{0}^{+\infty} dx \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} e^{-iKX_R(x)} \left\lbrace \bar{\psi}(-x) \left[ \hat{b}_L(K) + \hat{b}_L^\dagger(-K) \right] +\bar{\psi}(x) \left[ \hat{b}_R(-K) + \hat{b}_R^\dagger(K) \right] \right\rbrace | 0_M \rangle \nonumber\\
= & \int_{0}^{+\infty} dx \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} e^{-iKX_R(x)} \left\lbrace \bar{\psi}(-x) \left[ \exp \left( - \frac{\beta}{2}|K| \right) \hat{b}^\dagger_R(K) + \exp \left( \frac{\beta}{2}|K| \right) \hat{b}_R(-K) \right] \right. \nonumber \\
& \left. +\bar{\psi}(x) \left[ \hat{b}_R(-K) + \hat{b}_R^\dagger(K) \right] \right\rbrace | 0_M \rangle \nonumber\\
= & \int_{-\infty}^{+\infty} \frac{dK}{\sqrt{2\pi|K|}} \int_{0}^{+\infty} dx e^{-iKX_R(x)} \left\lbrace \left[ \exp \left( - \frac{\beta}{2}|K| \right) \bar{\psi}(-x) +\bar{\psi}(x) \right] \hat{b}^\dagger_R(K) \right. \nonumber \\
& \left. +\left[\exp \left( \frac{\beta}{2}|K| \right) \bar{\psi}(-x) + \bar{\psi}(x) \right] \hat{b}_R(-K) \right\rbrace | 0_M \rangle.
\end{align}
By comparing Eq. (\ref{state_invariant}) with Eq. (3), we obtain the expression for $\tilde{\psi}_\pm (K)$ in terms of $\bar{\psi}(x)$:
\begin{equation} \label{Delta_n_invariant}
\tilde{\psi}_\pm (K) = \frac{1}{\sqrt{2\pi|K|}}\int_{0}^{+\infty} dx e^{-iKX_R(x)} \left[ \exp \left( \mp \frac{\beta}{2}|K| \right) \bar{\psi}(-x) +\bar{\psi}(x) \right].
\end{equation}
The functions $\tilde{\psi}_\pm (K)$ appear in the final expression of $\Delta n_R(X)$ through Eqs. (11, 12b). Therefore, Eq. (\ref{Delta_n_invariant}) reveals the fact that left-wedge values of $\bar{\psi}(x)$ appear in $\Delta n_R(X)$.
\section{A proof for Eq. (8)} \label{Characteristic_function}
We want to provide a proof for Eq. (8). Firstly, it is possible to notice that, thanks to Eqs. (\ref{ordering_Rindler}, \ref{ordering_Rindler_2}), Eq. (6) can be put in the following form:
\begin{align}
\hat{\rho} = & \int_{-\infty}^{+\infty} dK \int_{-\infty}^{+\infty} dK' \left[ \tilde{\psi}_-(K) \tilde{\psi}^*_-(K') \hat{b}_R(K)\hat{\rho}_0 \hat{b}^\dagger_R(K') + \tilde{\psi}_-(K) \tilde{\psi}^*_+(K') \hat{b}_R(K)\hat{\rho}_0 \hat{b}_R(K') \right. \nonumber \\
& \left. + \tilde{\psi}_+(K) \tilde{\psi}^*_-(K') \hat{b}^\dagger_R(K)\hat{\rho}_0 \hat{b}^\dagger_R(K') + \tilde{\psi}_+(K) \tilde{\psi}^*_+(K') \hat{b}^\dagger_R(K)\hat{\rho}_0 \hat{b}_R(K') \right] \nonumber \\
= & \int_{-\infty}^{+\infty} dK \int_{-\infty}^{+\infty} dK' \left\lbrace \tilde{\psi}_-(K) \tilde{\psi}^*_-(K') \hat{b}_R(K)\hat{\rho}_0 \hat{b}^\dagger_R(K') + \tilde{\psi}_-(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} \hat{b}_R(K) \hat{b}_R(K')\hat{\rho}_0 \right. \nonumber \\
& \left. + \tilde{\psi}_+(K) \tilde{\psi}^*_-(K') e^{\beta |K|} \hat{\rho}_0 \hat{b}^\dagger_R(K)\hat{b}^\dagger_R(K') + \tilde{\psi}_+(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} \left[ e^{\beta|K|} \hat{b}_R(K') \hat{\rho}_0 \hat{b}^\dagger_R(K) -\delta(K-K') \hat{\rho}_0\right] \right\rbrace .
\end{align}
This last identity, together with Eqs. (9, \ref{normalization_2}) and the definition for $\chi_0[\xi,\xi^*]$ -- i.e.
\begin{equation} \label{chi_0}
\chi_0[\xi,\xi^*] = \text{Tr} \left\lbrace \hat{\rho}_0 \exp \left[ \int_{-\infty}^{+\infty} dK \xi(K) \hat{b}^\dagger_R(K)\right] \exp \left[- \int_{-\infty}^{+\infty} dK \xi^*(K) \hat{b}_R(K) \right] \right\rbrace
\end{equation}
-- gives
\begin{align}
\chi[\xi,\xi^*] = &\int_{-\infty}^{+\infty} dK \int_{-\infty}^{+\infty} dK' \left\lbrace -\tilde{\psi}_-(K) \tilde{\psi}^*_-(K') \frac{\delta}{\delta \xi^*(K)} \frac{\delta}{\delta \xi(K')} + \tilde{\psi}_-(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} \frac{\delta}{\delta \xi^*(K)} \frac{\delta}{\delta \xi^*(K')} \right. \nonumber \\
& \left. + \tilde{\psi}_+(K) \tilde{\psi}^*_-(K') e^{\beta |K|} \frac{\delta}{\delta \xi(K)} \frac{\delta}{\delta \xi(K')} + \tilde{\psi}_+(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} \left[-e^{\beta |K|} \frac{\delta}{\delta \xi(K)} \frac{\delta}{\delta \xi^*(K')} -\delta(K-K') \right] \right\rbrace\chi_0[\xi,\xi^*] \nonumber \\
= &\int_{-\infty}^{+\infty} dK \int_{-\infty}^{+\infty} dK' \left\lbrace \tilde{\psi}_-(K) \tilde{\psi}^*_-(K') [-n_0(K) n_0(K') \xi(K)\xi^*(K') + n_0(K') \delta(K-K') ] \right. \nonumber \\
& + \tilde{\psi}_-(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} n_0(K) n_0(K') \xi(K) \xi(K')+ \tilde{\psi}_+(K) \tilde{\psi}^*_-(K') e^{\beta |K|} n_0(K) n_0(K') \xi^*(K) \xi^*(K') \nonumber \\
& \left. + \tilde{\psi}_+(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} \left[-e^{\beta |K|} n_0(K) n_0(K') \xi^*(K) \xi(K') + e^{\beta |K|} n_0(K')\delta(K-K') -\delta(K-K') \right] \right\rbrace\chi_0[\xi,\xi^*] \nonumber \\
= &\int_{-\infty}^{+\infty} dK \int_{-\infty}^{+\infty} dK' \left\lbrace -\tilde{\psi}_-(K) \tilde{\psi}^*_-(K') n_0(K) n_0(K') \xi(K)\xi^*(K') + \tilde{\psi}_-(K) \tilde{\psi}^*_-(K') n_0(K') \delta(K-K') \right. \nonumber \\
& + \tilde{\psi}_-(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} n_0(K) n_0(K') \xi(K) \xi(K')+ \tilde{\psi}_+(K) \tilde{\psi}^*_-(K') e^{\beta |K|} n_0(K) n_0(K') \xi^*(K) \xi^*(K') \nonumber \\
& \left. + \tilde{\psi}_+(K) \tilde{\psi}^*_+(K') e^{\beta |K'|} \left[-e^{\beta |K|} n_0(K) n_0(K') \xi^*(K) \xi(K') + n_0(K')\delta(K-K') \right] \right\rbrace\chi_0[\xi,\xi^*] \nonumber \\
= &\int_{-\infty}^{+\infty} dK \int_{-\infty}^{+\infty} dK' \left\lbrace n_0(K') \delta(K-K') \left[ \tilde{\psi}_+(K) \tilde{\psi}^*_+(K') e^{\beta|K|} + \tilde{\psi}_-(K) \tilde{\psi}^*_-(K') \right] \right. \nonumber \\
& \left. - n_0(K) n_0(K') \left[\tilde{\psi}_-(K) \xi(K) - \tilde{\psi}_+(K) e^{\beta |K|} \xi^*(K) \right]\left[\tilde{\psi}^*_-(K') \xi^*(K') - \tilde{\psi}^*_+(K') e^{\beta |K'|} \xi(K') \right] \right\rbrace \chi_0[\xi,\xi^*] \nonumber \\
= & \left\lbrace \int_{-\infty}^{+\infty} dK n_0(K) \left[ \left| \tilde{\psi}_+(K) \exp \left( \frac{\beta}{2}|K| \right) \right|^2 + \left|\tilde{\psi}_-(K)\right|^2 \right]\right. \nonumber \\
& \left. - \left| \int_{-\infty}^{+\infty} dK n_0(K) \left[\tilde{\psi}_-(K) \xi(K) - \tilde{\psi}_+(K) e^{\beta |K|} \xi^*(K) \right] \right|^2 \right\rbrace \chi_0[\xi,\xi^*]\nonumber \\
= & \left\lbrace 1 - \left| \int_{-\infty}^{+\infty} dK n_0(K) \left[\tilde{\psi}_-(K) \xi(K) - \tilde{\psi}_+(K) e^{\beta |K|} \xi^*(K) \right] \right|^2 \right\rbrace \chi_0[\xi,\xi^*]
\end{align}
and hence Eq. (8) holds.
\section{Probability density function}
In the present section we show how to derive Eq. (11) from Eq. (10) and Eq. (14) from Eq. (12b).
Eq. (11) can be derived from Eq. (10) through the following chain of identities
\begin{align}
\Delta n_R(X) = & \left\langle \hat{n}_R(X) \right\rangle_{\hat{\rho}} - \left\langle \hat{n}_R(X) \right\rangle_{\hat{\rho}_0} \nonumber \\
= & \left. \int_{-\infty}^{+\infty} dK \frac{e^{-i K X}}{\sqrt{2\pi}} \int_{-\infty}^{+\infty} dK' \frac{e^{i K' X}}{\sqrt{2\pi}} \frac{\delta}{\delta \xi(K)} \left(-\frac{\delta}{\delta \xi^*(K')} \right) \chi_0[\xi,\xi^*]\right|_{\xi=0} - \int_{-\infty}^{+\infty} dK \frac{e^{-i K X}}{\sqrt{2\pi}} \int_{-\infty}^{+\infty} dK' \frac{e^{i K' X}}{\sqrt{2\pi}} \nonumber \\
& \left. \times \frac{\delta}{\delta \xi(K)} \left(-\frac{\delta}{\delta \xi^*(K')} \right) \left| \int_{-\infty}^{+\infty} dK'' n_0(K'') \left[ - \tilde{\psi}_+(K'') e^{\beta |K''|} \xi^*(K'') + \tilde{\psi}_-(K'') \xi(K'') \right] \right|^2 \chi_0[\xi,\xi^*] \right|_{\xi=0} \nonumber \\
& - \left\langle \hat{n}_R(X) \right\rangle_{\hat{\rho}_0} \nonumber \\
= &\int_{-\infty}^{+\infty} dK n_0(K)\int_{-\infty}^{+\infty} dK' n_0(K') \frac{e^{-i (K-K') X}}{2\pi}\left[ \tilde{\psi}^*_+(K) \tilde{\psi}_+(K') e^{\beta (|K|+|K'|)} + \tilde{\psi}_-(K) \tilde{\psi}^*_-(K') \right] \nonumber \\
= &\left| \int_{-\infty}^{+\infty} dK n_0(K)\frac{e^{-i K X}}{\sqrt{2 \pi}} \tilde{\psi}^*_+(K) e^{\beta |K|} \right|^2 + \left| \int_{-\infty}^{+\infty} dK n_0(K)\frac{e^{-i K X}}{\sqrt{2 \pi}} \tilde{\psi}_-(K) \right|^2 \nonumber \\
= &\left| \int_{-\infty}^{+\infty} dK n_0(K)\frac{e^{i K X}}{\sqrt{2 \pi}} \tilde{\psi}_+(K) e^{\beta |K|} \right|^2 + \left| \int_{-\infty}^{+\infty} dK n_0(K)\frac{e^{-i K X}}{\sqrt{2 \pi}} \tilde{\psi}_-(K) \right|^2 \nonumber \\
= & n_+(X) + n_-(X).
\end{align}
It is possible to obtain Eq. (14) from Eq. (12b) and the following chain of identities
\begin{align}
n_\pm (X) = & \left| \int_{-\infty}^{+\infty} dK n_0(K)\frac{e^{\pm i K X}}{\sqrt{2 \pi}} \tilde{\psi}_\pm(K) e^{\theta(\pm 1) \beta |K|} \right|^2 \nonumber \\
= & \left| \int_{-\infty}^{+\infty} dK n_0(K)\frac{e^{- i K X}}{\sqrt{2 \pi}} \tilde{\psi}_\pm(\mp K) e^{\theta(\pm 1) \beta |K|} \right|^2 \nonumber \\
= & \left| \int_{-\infty}^{+\infty} dK \frac{e^{ - i K X}}{\sqrt{2 \pi}} \int_{-\infty}^{+\infty} dX' \frac{e^{ i K X' }}{\sqrt{2 \pi}} \left\lbrace \psi_R(X') \left[ \theta(\pm 1) + \tilde{f}_{R \pm}\left( \frac{K}{a} \right) \right] + \psi_L( - X') \tilde{f}_{L \pm}\left( \frac{K}{a} \right) \right\rbrace \right|^2 \nonumber \\
= & \left| \int_{-\infty}^{+\infty} dX' \left\lbrace \psi_R(X') \left[ \theta(\pm 1) \int_{-\infty}^{+\infty} dK \frac{e^{ i K (X'-X)}}{2 \pi} + \int_{-\infty}^{+\infty} dK \frac{e^{ i K (X'-X)}}{2 \pi} \tilde{f}_{R \pm}\left( \frac{K}{a} \right) \right] \right.\right. \nonumber \\
& \left.\left. + \psi_L( - X') \int_{-\infty}^{+\infty} dK \frac{e^{ i K (X'-X)}}{2 \pi} \tilde{f}_{L \pm}\left( \frac{K}{a} \right) \right\rbrace \right|^2 \nonumber \\
= & \left| \int_{-\infty}^{+\infty} dX' \left\lbrace \psi_R(X') \left[ \theta(\pm 1) \delta(X'-X) + a f_{R \pm}(aX'-aX) \right] + \psi_L( - X') a f_{L \pm} (aX'-aX) \right\rbrace \right|^2 \nonumber \\
= & \left| \theta(\pm 1) \psi_R(X) + \int_{-\infty}^{+\infty} d\xi \psi_R \left( \frac{\xi}{a} \right) f_{R \pm}(\xi-aX) + \int_{-\infty}^{+\infty} d\xi \psi_L \left( - \frac{\xi}{a} \right) f_{L \pm} (\xi-aX) \right|^2 \nonumber \\
= & | \theta(\pm 1) \psi_R(X) + \psi_{R \pm}(X) + \psi_{L \pm}(X) |^2 .
\end{align}
\section{Gaussian wave-packets}
The particular family of Gaussian wave-functions that we have considered are expressed by Eq. (16) with $X$ as variable and $x_0$ and $\sigma$ as parameters. In this section we want to study the behavior of $\psi_{L,R}(X)$, $\psi_{L,R \pm}(X)$ and $\Delta n_R(X)$ with respect to different values of $x_0$ and $\sigma$.
\begin{figure}[h]
\includegraphics[]{psi_R.pdf}
\caption{Profile of $\psi_R(X)$ defined by Eq. (\ref{Gaussian_psi_R}) for different values of $x_0$ and $\sigma$.} \label{psi_R_figure}
\end{figure}
Notice that if we switch the sign of $x_0$ -- i.e. $x_0 \mapsto -x_0$ -- in Eq. (16), we obtain $\psi (-x)$. This results in the possibility to describe $\psi_L(X)$ through $\psi_R(X)$ with opposite $X$ and $x_0$. In other words, the transformation $X \mapsto -X$ and $x_0 \mapsto -x_0$ gives $\psi_L(X) \mapsto \psi_R(X)$ and $\psi_R(X) \mapsto \psi_L(X)$. It is, therefore, sufficient to show $\psi_R(X)$ for different parameters $x_0$ and $\sigma$ in order to describe both $\psi_R(X)$ and $\psi_L(X)$. This has been done in Fig. \ref{psi_R_figure}.
In general, it is possible to expect an almost negligible $\psi_R(X)$ for $x_0<0$ and $\sigma \ll -x_0$, since $\psi(x)$ gives most of its contribution to $\psi_L(X)$ and Eq. (\ref{normalization_x_2}) implies that $P_R \ll 1$ when $P_L \lesssim 1$. Indeed the profile for $a x_0 = -10 $ and $a \sigma = 1$ in Fig. \ref{psi_R_figure} -- marked with a blue solid line -- is not noticeable. On the other hand, a greater $\sigma$ allows $\psi(x)$ to reach the right wedge with more contribution to $\psi_R(X)$. Indeed the wave functions with $x_0 < 0$ and $\sigma \gtrsim -x_0$ are noticeable in Fig. \ref{psi_R_figure} and they are marked with dashed blue and green lines. Moreover, for greater values of $\sigma$, $P_R$ becomes larger together with the area under the curve of $\psi_R(X)$, as it is possible to notice by comparing the dashed blue and green lines with the respective solid lines. Contrary to the case of negative $x_0$, figures with larger area are the ones with $x_0>0$ and $\sigma \ll x_0$, or, in other words, wave-packets well-localized in the right wedge -- see, for instance, the solid red line in Fig. \ref{psi_R_figure}; while smaller areas are given by wave-packets that have a non-negligible probability to be found in $x<0$ -- i.e. $\sigma \gtrsim x_0$, as for the dashed red and yellow lines in Fig. \ref{psi_R_figure}.
In order to conclude the analysis of $\psi_R(X)$ for different values of $\text{sign} (x_0)$, we want to discuss the case of $x_0 = 0$. An interesting property of $\psi_R(X)$ for such case can be obtained from its explicit expression:
\begin{equation} \label{Gaussian_psi_R}
\psi_R(X) = \frac{1}{\sqrt[4]{\pi}\sqrt{\sigma}} \exp \left( \frac{a X}{2} - \frac{\left( e^{a X}-a x_0 \right)^2}{2 (a \sigma)^2}\right).
\end{equation}
Eq. (\ref{Gaussian_psi_R}) can also be put in the following form
\begin{equation} \label{Gaussian_psi_R_2}
\psi_R(X) = \frac{\sqrt{a}}{\sqrt[4]{\pi}} \exp \left( \frac{1}{2} [a X - \ln (a \sigma)] - \frac{1}{2} \left[ e^{a X - \ln (a \sigma)}- \frac{x_0}{\sigma} \right]^2\right).
\end{equation}
In the particular case of $x_0=0$, it is easy to see from Eq. (\ref{Gaussian_psi_R_2}) that for different values of $\sigma$, the profile of $\psi_R(X)$ translates. Specifically for a transformation $\sigma \mapsto \alpha \sigma$ with $\alpha > 0$, $\psi_R(X)$ transforms in the following way: $\psi_R(X) \mapsto \psi_R(X - \ln \alpha)$. Such behavior can be observed from the gray lines in Fig. \ref{psi_R_figure}.
An equivalent expression for $\psi_R(X)$ when $x_0 \neq 0$ is the following
\begin{equation} \label{Gaussian_psi_R_3}
\psi_R(X) = \frac{\sqrt{a}}{\sqrt[4]{\pi}} \sqrt{\frac{|x_0|}{\sigma}} \exp \left( \frac{1}{2} [a X - \ln (a |x_0|)] - \frac{1}{2} \left(\frac{|x_0|}{\sigma} \right)^2 \left[ e^{a X - \ln (a |x_0|)}- \text{sign}(x_0) \right]^2 \right).
\end{equation}
If $x_0 \neq 0$, we can use Eq. (\ref{Gaussian_psi_R_3}) to proof that any transformation $x_0 \mapsto \alpha x_0$ with $\alpha > 0$ acting on $\psi_R(X)$ is equivalent to $\sigma \mapsto \sigma / \alpha$, $a X \mapsto a X - \ln \alpha$. Equivalently, we can say that when $x_0 \neq 0$ and for fixed sign of $x_0$, the only independent variables of $\psi_R(X)$ are $X - X_R(|x_0|)$ and $|x_0|/\sigma$, as we can directly see from Eq. (\ref{Gaussian_psi_R_3}). The invariance under the transformation $x_0 \mapsto x_0/\alpha$, $\sigma \mapsto \sigma / \alpha$, $a X \mapsto a X - \ln \alpha$ for any $\alpha>0$ can be observed in Fig. \ref{psi_R_figure} through the yellow solid line with the dashed red line and the green solid line with the blue dashed line.
Thanks to the possibility to switch from $\psi_R(X)$ to $\psi_L(X)$ by performing the transformation $X \mapsto -X$ and $x_0 \mapsto -x_0$, $\psi_L(X)$ behaves similarly to $\psi_R(X)$ with respect to the property described before. Specifically, when $x_0=0$, the transformation $\sigma \mapsto \alpha \sigma$ with $\alpha > 0$ acting on $\psi_L(X)$ is equivalent to $\psi_L(X) \mapsto \psi_L(X + \ln \alpha)$. Moreover, if $x_0 \neq 0$, invariance under the transformation $x_0 \mapsto x_0/\alpha$, $\sigma \mapsto \sigma / \alpha$, $a X \mapsto a X + \ln \alpha$ for any $\alpha>0$ holds for $\psi_L(X)$. It is now straightforward to derive such properties for $\psi_{L,R\pm}(X)$ and $\Delta n_R (X)$ -- i.e. $\psi_{L,R\pm}(X)$ and $\Delta n_R (X)$ are invariant under $x_0 \mapsto x_0/\alpha$, $\sigma \mapsto \sigma / \alpha$, $X \mapsto X - \ln \alpha$ -- as we have stated in the main paper.
In summary, the behavior of $\psi_R(X)$ with respect to $|x_0|$ and for fixed $\text{sign} (x_0)$ and $\sigma/|x_0|$ corresponds to a translation with respect to $X$. The interesting limits are $a |x_0| \rightarrow 0$ and $a |x_0| \rightarrow \infty$ with finite $\sigma/|x_0|$ since for such limits the profile of $\psi_R(X)$ ``disappears'' at $a X \rightarrow - \infty$ and $a X \rightarrow + \infty$ respectively. Such behavior can be motivated in the following way. The case $a |x_0| \rightarrow 0$ and fixed $ \sigma/|x_0|$ corresponds also to the limit $a |x_0| \rightarrow 0$ and $a \sigma \rightarrow 0$, which, in turns, represents the case in which $\psi(x)$ is well-localized at the horizon. The case $a |x_0| \rightarrow \infty$ and fixed $\sigma/|x_0|$, instead, corresponds to the case in which $\psi(x)$ is localized far from the horizon.
Since we have already explored the behavior of $\psi_R(X)$ with respect to $x_0 \neq 0$ and fixed $\sigma / |x_0|$, we are now interested in the behavior of $\psi_R(X)$ with respect to $\sigma / |x_0|$ for fixed $x_0 \neq 0$, which, in Fig. \ref{psi_R_figure}, corresponds to considering curves with the same color. In limit $\sigma / |x_0| \rightarrow \infty$ for fixed $x_0 \neq 0$, the wave-function $\psi(x)$ ``disappears'' as
\begin{equation}
\psi(x) = \mathcal{O} \left(\sqrt{\frac{|x_0|}{\sigma}}\right).
\end{equation}
In the same manner, $\psi_R(X)$ ``disappears'' since
\begin{equation}
\psi_R(X) = \mathcal{O} \left(\sqrt{\frac{|x_0|}{\sigma}}\right),
\end{equation}
as it is possible to see from Eq. (\ref{Gaussian_psi_R_3}). Moreover, the same behavior can be observed for $\psi_L(X)$, $\psi_{L,R}(X)$ and $\Delta n_R(X)$.
The limit $\sigma / |x_0| \rightarrow 0$ for fixed $x_0$ corresponds to the case of a perfectly localized Minkowski wave-function $\psi(x) = \delta(x-x_0)$. In the case of $\psi_R(X)$, Eq. (\ref{Gaussian_psi_R_3}) gives
\begin{align} \label{limit_small_sigma_x_0_R}
\lim_{\sigma/|x_0| \rightarrow 0} | \psi_R(X) |^2 = & a e^{ a X - \ln (a |x_0|) } \lim_{\sigma/|x_0| \rightarrow 0} \frac{|x_0|}{\sqrt{\pi} \sigma} \exp \left( - \left(\frac{|x_0|}{\sigma} \right)^2 \left[ e^{a X - \ln (a |x_0|)}- \text{sign}(x_0) \right]^2 \right) \nonumber \\
= & \frac{e^{a X}}{|x_0|} \delta \left( e^{a X - \ln (a |x_0|)} - \text{sign}(x_0) \right) \nonumber \\
= &\theta(x_0) \delta \left( X - \frac{ \ln (a x_0) }{a} \right) \nonumber \\
= & \theta(x_0) \delta(X-X_R(x_0)),
\end{align}
which can be visualized through the red and blue lines in Fig. \ref{psi_R_figure}.
Finally, the limit $\sigma / |x_0| \rightarrow 0$ for fixed $x_0$ gives the following identity
\begin{equation} \label{limit_small_sigma_x_0_a_R}
\int_{-\infty}^{+\infty} d\xi \psi_R \left( \frac{\xi}{a} \right) f(\xi) = \theta(x_0) \sqrt{ \frac{2 \sqrt{\pi} a \sigma}{x_0}} f(a X_R(x_0)) + o \left( \sqrt{\frac{\sigma}{|x_0|}} \right),
\end{equation}
which is valid for any function $f(\xi)$. Eq. (\ref{limit_small_sigma_x_0_a_R}) can be proven starting from the following identity
\begin{align}
\lim_{\sigma/|x_0| \rightarrow 0} \sqrt{\frac{|x_0|}{\sigma}}\psi_R \left( \frac{\xi}{a} \right)
= & \lim_{\sigma/|x_0| \rightarrow 0} \frac{\sqrt{|x_0|}}{\sqrt[4]{\pi}\sigma} \exp \left( \frac{\xi}{2} - \frac{x_0^2}{2 \sigma^2} \left(\frac{ e^\xi}{a x_0} - 1\right)^2 \right) \nonumber \\
= & \sqrt{\frac{2\sqrt{\pi}}{|x_0|}} \exp \left( \frac{\xi}{2} \right) \delta \left(\frac{ e^\xi}{a x_0} - 1 \right) \nonumber \\
= & \sqrt{2\sqrt{\pi}|x_0|} a \exp \left(-\frac{\xi}{2} \right) \theta(x_0) \delta (\xi - \ln (a x_0)) \nonumber \\
= & \theta(x_0) \sqrt{ 2 \sqrt{\pi} a} \delta (\xi - a X_R (x_0)) ,
\end{align}
which implies that
\begin{equation}
\lim_{\sigma / x_0 \rightarrow 0} \sqrt{\frac{|x_0|}{\sigma}} \int_{-\infty}^{+\infty} d\xi \psi_R \left( \frac{\xi}{a} \right) f(\xi)= \theta(x_0) \sqrt{ 2 \sqrt{\pi} a } f(a X_R(x_0))
\end{equation}
and hence Eq. (\ref{limit_small_sigma_x_0_a_R}) holds.
Thanks to the possibility of switching between $\psi_R(X)$ and $\psi_L(X)$ trough a change of sign for $X$ and $x_0$, it is possible to write Eq. (\ref{limit_small_sigma_x_0_a_R}) for $\psi_L(X)$
\begin{equation} \label{limit_small_sigma_x_0_a_L}
\int_{-\infty}^{+\infty} d\xi \psi_L \left(-\frac{\xi}{a} \right) f(\xi) = \theta(-x_0) \sqrt{ \frac{2 \sqrt{\pi} a \sigma}{-x_0}} f(a X_L(x_0))+ o \left( \sqrt{\frac{\sigma}{|x_0|}} \right).
\end{equation}
Eqs. (\ref{limit_small_sigma_x_0_a_R}, \ref{limit_small_sigma_x_0_a_L}) give the following result for $\psi_{L,R \pm}(X)$
\begin{subequations} \label{limit_small_sigma_2}
\begin{equation} \label{limit_small_sigma_L_pm}
\psi_{L \pm}(X) = \theta(-x_0) \sqrt{ \frac{2 \sqrt{\pi} a \sigma}{-x_0}} f_{L \pm}(-a X_L(x_0) - a X) + o \left( \sqrt{\frac{\sigma}{|x_0|}} \right),
\end{equation}
\begin{equation} \label{limit_small_sigma_R_pm}
\psi_{R \pm}(X) = \theta(x_0) \sqrt{ \frac{2 \sqrt{\pi} a \sigma}{x_0}} f_{R \pm}(a X_R(x_0) - a X) + o \left( \sqrt{\frac{\sigma}{|x_0|}} \right),
\end{equation}
\end{subequations}
for any $X$.
The order of magnitude for vanishing values of $\psi_{L \pm}(X)$ and $\psi_{R \pm}(X)$ in the case of, respectively, $x_0>0$ and $x_0<0$ can be put in a more strict form of Eqs. (\ref{limit_small_sigma_2}) in the following way. Firstly, let us consider the following chain of inequalities
\begin{align}
\sqrt{\frac{|x_0|}{\sigma}} \exp\left( \frac{x_0^2}{2 \sigma^2} \right) | \psi_{R \pm}(X) | \leq & \sqrt{\frac{|x_0|}{\sigma}} \exp\left( \frac{x_0^2}{2 \sigma^2} \right) \int_{-\infty}^{+\infty} d\xi \left| \psi_{R}\left(\frac{\xi}{a} \right) f_{R \pm}( \xi - aX ) \right| \nonumber \\
\leq & \sqrt{\frac{|x_0|}{\sigma}} \exp\left( \frac{x_0^2}{2 \sigma^2} \right) f^{(max)}_{R \pm} \int_{-\infty}^{+\infty} d\xi \psi_{R}\left(\frac{\xi}{a} \right)\nonumber \\
= & \sqrt{\frac{|x_0|}{\sigma}} \exp\left( \frac{x_0^2}{2 \sigma^2} \right) \frac{f^{(max)}_{R \pm}}{\sqrt[4]{\pi}\sqrt{\sigma}} \int_{-\infty}^{+\infty} d\xi \exp \left( \frac{\xi}{2} - \frac{1}{2}\left(\frac{ e^\xi}{a \sigma} - \frac{x_0}{\sigma}\right)^2 \right)\nonumber \\
= & \frac{f^{(max)}_{R \pm}\sqrt{|x_0|}}{\sqrt[4]{\pi}\sigma} \exp\left( \frac{x_0^2}{2 \sigma^2} \right) \int_{0}^{\infty} d \tilde{\xi} \sigma \sqrt{\frac{a}{|x_0|}} \exp \left( - \frac{1}{2}\left(\frac{ \sigma \tilde{\xi}^2}{|x_0|} - \frac{x_0}{\sigma}\right)^2 \right)\nonumber \\
= & \frac{f^{(max)}_{R \pm} \sqrt{a}}{\sqrt[4]{\pi}} \int_{0}^{\infty} d \tilde{\xi} \exp \left( - \frac{ \sigma^2 \tilde{\xi}^4}{|x_0|^2} - \tilde{\xi}^2 \right)\nonumber \\
\leq & \frac{f^{(max)}_{R \pm} \sqrt{a}}{\sqrt[4]{\pi}} \int_{0}^{\infty} d \tilde{\xi} \exp \left( - \tilde{\xi}^2 \right)\nonumber \\
= & \frac{\sqrt[4]{\pi}}{2}f^{(max)}_{R \pm} \sqrt{a} ,
\end{align}
where $x_0<0$, $f^{(max)}_{R \pm}$ is the maximum value of $|f_{R \pm}( \xi)|$ and
\begin{equation}
\tilde{\xi} = \frac{1}{\sigma} \sqrt{\frac{|x_0|}{a}e^\xi}.
\end{equation}
This leads to
\begin{equation} \label{limit_small_sigma_R_pm_2}
| \psi_{R \pm}(X) | = o \left( \sqrt{\frac{\sigma}{|x_0|}} \exp \left( - \frac{x_0^2}{4 \sigma^2} \right) \right)
\end{equation}
when $x_0<0$. Similarly, when $x_0>0$
\begin{equation} \label{limit_small_sigma_L_pm_2}
| \psi_{L \pm}(X) | = o \left( \sqrt{\frac{\sigma}{|x_0|}} \exp \left( - \frac{x_0^2}{4 \sigma^2} \right) \right).
\end{equation}
Eq. (\ref{limit_small_sigma_x_0_R}) results in Eq. (17a), while Eqs. (\ref{limit_small_sigma_2}, \ref{limit_small_sigma_R_pm_2}, \ref{limit_small_sigma_L_pm_2}) result in Eq. (17b).
|
\section{Introduction}
The time-evolution of both classical and quantum systems may develop memory
effects \cite{vanKampen,breuerbook,vega,wiseman}. Nevertheless, the
characterization and definition of these effects is quite different in both
regimes \cite{BreuerReview,plenioReview}. As is well known, in a classical
regime memory effects can be rigorously defined in a probabilistic approach.
The independence or dependence of conditional probabilities on the previous
system history define respectively the (memoryless) Markovian and
non-Markovian regimes \cite{vanKampen}.
In a quantum regime one is immediately confronted with an extra aspect. In
fact, the state of a quantum system (and consequently its history) can only
be determined by performing a measurement process, which intrinsically
implies a perturbation to its (originally unperturbed) dynamics. Therefore,
the definition of memory effects and quantum non-Markovianity can be tackled
from two intrinsically different approaches. In \textit{non-operational
approaches}, memory effects are defined by taking solely into account the
properties of the unperturbed open system dynamics (its propagator). In
\textit{operational approaches}, memory effects are defined by the
statistical properties of different outcomes associated to system
measurement processes and transformations (like unitary ones).
A wide variety of measures and memory witness has been utilized in the
context of non-operational approaches (see reviews \cit
{BreuerReview,plenioReview}). The first proposals correspond to deviations
of the system propagator from divisibility \cite{divisible}\ and a
non-monotonous behavior of the trace distance\ (TD) between two distinct
system states \cite{BreuerFirst}. In this context, memory effects were
associated to an \textit{environment-to-system backflow of information}:
information stored in the initial system state is transferred to the
environmental degrees of freedom; their influence on the system at later
times implies a backflow of information that leads to memory effects. In
spite of this clear and well-motivated interpretation \cit
{EnergyBackFLow,Energy,HeatBackFLow}, the precise assessment of this concept
is still under debate \cit
{megier,maximal,petruccione,amato,santis,acin,horo,EntroBack}.
The basic idea of operational approaches is to appeal to the standard
definition of memory effects in terms of probabilities \cite{vanKampen}.
Hence, the (quantum) system must be subjected to a set of measurement
processes such that their statistical properties determine the presence or
absence of memory effects \cite{modi,budiniCPF}. The study and understanding
of this approach has been performed in the recent literature \cit
{pollock,pollockInfluence,bonifacio,han,ban,rio}, including alternative
definitions and analysis of information flows \cite{goan,BIF}.
The main goal of this paper is to analyze and to compare how the concept of
environment-to-system backflow of information is interpreted and implemented
in operational and non-operational approaches. As a non-operational memory
witness we take the TD between two different system initial states \cit
{BreuerFirst}, taking also into account the bounds on its revival behavior
that has been characterized recently \cite{EntroBack}. As an operational
memory witness we consider a conditional past-future (CPF) correlation \cit
{budiniCPF}, both in a deterministic and random schemes \cite{BIF}. The
comparison is performed by considering different system-environment models
and analyzing in each case the information flows from the two perspectives.
We consider statistical mixtures of Markovian system evolutions, system
coupled to incoherent \cite{maximal}\ and coherent casual bystander
environments \cite{casual}, which are characterized by a self-dynamics that
is independent of the system degrees of freedom. In addition, we consider
(standard) unitary system-environment models \cite{breuerbook}.
As a specific model, we study a depolarizing map induced by the interaction
of a system with a finite set of incoherent degrees of freedom. In this
regime, as well as in a quantum coherent one, we explain how and why both
approaches lead to different notions of quantum non-Markovianity and
environment-to-system backflows of information.
The paper is outlined as follows. In Sec.~II we review the definition and
main properties of the considered non-operational \cit
{BreuerFirst,EntroBack}\ and operational \cite{budiniCPF,BIF} approaches. In
Sec. III we study both approaches by considering different
system-environment models. In Sec.~IV we study the depolarizing map. In Sec.
V we provide the conclusions.
\section{Quantum non-Markovianity}
Here we briefly review the main characteristics of the different approaches
to quantum non-Markovianity.
\subsection{Non-operational approach}
If the open system is not affected or perturbed during its evolution, the
unique object that allows to defining the presence or absence of memory
effects is its (unperturbed) density matrix propagator. The rigorous theory
of quantum dynamical semigroups \cite{alicki} motivate to associating the
(memoryless) quantum Markovian regime with propagators whose time-evolution
obey a Lindblad equation (or Gorini-Kossakowski-Sudarshan-Lindblad
equation). Consequently, any (scalar) measure or property that quantifies
departures of the system propagator from a Lindblad equation can be taken as
a witness of quantum memory effects.
Lindblad equations lead to completely positive propagators between two
arbitrary times \cite{alicki}. As is well known, completely positive
transformations lead to very specific contractive properties for different
distance measures and entropic quantities \cite{nielsen}. For example, the
TD between two arbitrary density matrixes $\rho $ and $\sigma ,$ defined as
\mathbf{D}(\rho ,\sigma )\equiv (1/2)\mathrm{Tr}|\rho -\sigma |,$ under a
completely positive transformation $\Phi $ fulfills the inequality $\mathbf{
}(\Phi \lbrack \rho ],\Phi \lbrack \sigma ])\leq \mathbf{D}(\rho ,\sigma ).$
Consequently, it is possible to \textit{define} quantum Markovianity by the
condition \cite{BreuerFirst
\begin{equation}
\mathbf{D}(\rho _{t+\tau }^{s},\sigma _{t+\tau }^{s})\leq \mathbf{D}(\rho
_{t}^{s},\sigma _{t}^{s}), \label{Contraction}
\end{equation
where $\rho _{t}^{s}$ and $\sigma _{t}^{s}$ are two arbitrary evolved system
states that differ in their initial conditions, $\rho _{0}^{s}\neq \sigma
_{0}^{s}.$ Alternatively, one can interpret that quantum memory effects are
present whenever this inequality is not fulfilled for a set of two arbitrary
time intervals $t\geq 0$ and $\tau >0.$
In spite of the simplicity and efficacy of the previous theoretical frame,
in general it is not possible to know or infer which physical processes are
involved when the contractive condition~(\ref{Contraction}) is not
fulfilled. A remarkable advance in this direction was recently obtained in
Ref.~\cite{EntroBack} by establishing the inequalit
\begin{eqnarray}
\mathbf{D}(\rho _{t+\tau }^{s},\sigma _{t+\tau }^{s}) &\leq &\mathbf{D}(\rho
_{t}^{s},\sigma _{t}^{s})+\mathbf{D}(\rho _{t}^{e},\sigma _{t}^{e})
\label{General} \\
&&\!\!\!\!\!+\mathbf{D}(\rho _{t}^{se},\rho _{t}^{s}\otimes \rho _{t}^{e})
\mathbf{D}(\sigma _{t}^{se},\sigma _{t}^{s}\otimes \sigma _{t}^{e}). \notag
\end{eqnarray
Here, $\rho _{t}^{se}$ and $\sigma _{t}^{se}$ are the evolved
system-environment states with initial conditions $\rho _{0}^{se}=\rho
_{0}^{s}\otimes \rho _{0}^{e}$ and $\sigma _{0}^{se}=\sigma _{0}^{s}\otimes
\sigma _{0}^{e}.$ As usual, the system and bath states follows from a
partial trace operation, $\rho _{t}^{s}=\mathrm{Tr}_{e}[\rho _{t}^{se}]$ and
$\rho _{t}^{e}=\mathrm{Tr}_{s}[\rho _{t}^{se}].$ The asymmetry between
system and environment $(s\nleftrightarrow e)$ is introduced by taking in
both cases the same initial environmental state, $\rho _{0}^{e}=\sigma
_{0}^{e}.$
The result~(\ref{General}) only relies on the triangle inequality fulfilled
by the TD. Thus, it is valid for arbitrary system-environment models. In
addition, this expression allows to bounding the environment-to-system
backflow of information \textit{defined} by the \textquotedblleft
revivals\textquotedblright
\begin{equation}
\mathbf{D}(\rho _{t+\tau }^{s},\sigma _{t+\tau }^{s})-\mathbf{D}(\rho
_{t}^{s},\sigma _{t}^{s})>0. \label{Revivals}
\end{equation
The remaining (bounding) contributions in the rhs of Eq.~(\ref{General})
have a \textit{clear physical interpretation}. One can relate the
contribution $\mathbf{D}(\rho _{t}^{e},\sigma _{t}^{e})$ to changes in the
environmental state, while the terms $\mathbf{D}(\rho _{t}^{se},\rho
_{t}^{s}\otimes \rho _{t}^{e})+\mathbf{D}(\sigma _{t}^{se},\sigma
_{t}^{s}\otimes \sigma _{t}^{e})$ measure the correlations established
between the system and the environment~\cite{EntroBack}.
Nevertheless, it is important to realize that these physical processes do
not guarantee the developing of revivals. The right conclusion is that
\textit{given} that there exists revivals, their origin can related to
changes in the environmental state or to the establishing of
system-environment correlations.
It was also proven that the inequality~(\ref{General}) remains valid when
the TD is replaced by a telescopic relative entropy and the square root of a
quantum Jensen-Shannon divergence~\cite{EntroBack}. Thus, the interpretation
of the bounds remains the same when using these entropic quantities.
\subsection{Operational approach}
In a probabilistic frame, given a sequence of system states $x\rightarrow
y\rightarrow z$ with joint probability $P(z,y,x),$ Markovianity is defined
by the conditio
\begin{equation}
P(z,y,x)=P(z|y)P(y|x)P(x), \label{MarkovProb}
\end{equation
where $P(b|a)$ denotes in general the conditional probability of $b$ given
a.$ By Bayes rule, the equality~(\ref{MarkovProb}) implies the (memoryless)
condition $P(z|y,x)=P(z|y).$ Similar constraints emerge when considering
higher joint probabilities involving an arbitrary number of events~\cit
{vanKampen}.
For quantum systems, the definition of Markovianity in terms of
probabilities unavoidably implies performing a set of system measurement
processes. In Ref.~\cite{modi}, by means of a process tensor formalism, the
Markovian condition is taken into account for arbitrary (higher order) joint
probabilities. Nevertheless, for \textit{quantum} systems coupled to
standard environment models (standard classical noises and/or unitary
system-environment interaction models), only three measurement events are
enough for detecting departures from a (probabilistic) Markovian regime~\cit
{budiniCPF}. In such a case, the condition~(\ref{MarkovProb}) can be
conveniently rewritten as a CPF independence
\begin{equation}
P(z,x|y)=P(z|y)P(x|y). \label{CPFIndep}
\end{equation
This result follows straightforwardly by using that $P(z,x|y)=P(z,y,x)/P(y),$
where $P(y)=\sum_{z,x}P(z,y,x).$
The CPF independence~(\ref{CPFIndep}) implies that any\ (conditional)
correlation between past and future events witnesses memory effects.
Correspondingly, a CPF correlation is defined as~\cite{budiniCPF
\begin{equation}
C_{pf}(t,\tau )|_{\breve{y}}\overset{d/r}{=}\sum_{z,x}zx[P(z,x|\breve{y
)-P(z|\breve{y})P(x|\breve{y})], \label{CPFCorre}
\end{equation
where $\{x\}$ and $\{z\}$ are the (past and future) measurement outcomes.
The time dependence $(t,\tau )$ emerges because the past, present, and
future measurements are performed at the initial time $t=0,$ at time $t,$
and $t+\tau $ respectively. Evidently $C_{pf}(t,\tau )|_{\breve{y}}$
vanishes in a (probabilistic) Markovian regime [Eq.~(\ref{CPFIndep})].
In Eq.~(\ref{CPFCorre}) it was introduced the change $y\rightarrow \breve{y
, $ which is stretchy related with the definition of memory effects and
information flows in this approach. Two different measurement schemes are
necessary \cite{BIF}. In a deterministic scheme (denoted with the supra $d),$
after the intermediate measurement (whose outcome defines the conditional
property) none change is introduced. Hence, $\breve{y}=y.$ In a random
scheme (denoted with the supra $r),$ after the intermediate measurement the
system state is randomly chosen $(y\rightarrow \breve{y})$ over the set of
possible states associated to the outcomes $\{y\}.$ The CPF correlation is
defined with this renewed conditional state.
In the deterministic scheme, the CPF correlation $[C_{pf}(t,\tau )|_{\breve{
}}\overset{d}{\neq }0]$ detects memory effects [departures with respect to
Eq.~(\ref{MarkovProb}), or equivalently Eq.~(\ref{CPFIndep})] independently
of the specific system-environment model. In the random scheme, a
non-vanishing CPF correlation $[C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{\neq
}0],$ by \textit{definition,} detects the presence of environment-to-system
backflows of information (or bidirectional system-environment information
flows). This relation is motivated by the complementary case $C_{pf}(t,\tau
)|_{\breve{y}}\overset{r}{=}0$ that applies when the environment (which
induces the memory effects $C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{\neq }0)$
is unperturbed by its coupling with the system~\cite{BIF}.
The previous characteristics of the deterministic and random schemes can be
easily understood from the properties of projective measurements performed
over bipartite systems \cite{casual}. Interestingly, the formalism remains
the same and is also valid for purely (classically) incoherent
system-environment arrangements.
\subsection{Bipartite propagator vs. single propagator}
Before comparing both approaches (next section), here we clarify which
dynamical objects determine each one. In the non-operational approach, the
presence of memory effects [TD revivals defined by Eq.~(\ref{Revivals})] can
be determined after knowing solely the system (single) propagator. In
contrast, for determining the bound defined by Eq.~(\ref{General}) it is
necessary to know the bipartite system-environment propagator specified for
a given initial bath state.
In contrast, the operational approach can only be characterized by knowing
(exact or approximate) the bipartite propagator for different
initial bath states (the initial one and the bath state after the
intermediate measurement). As a matter of fact, the CPF correlation~(\re
{CPFCorre}) can be written as a function of the joint probability $P(z
\breve{y},x).$ Assuming that the three measurements are projective ones, in
the deterministic scheme it reads~\cite{BIF
\begin{equation}
\frac{P(z,\breve{y},x)}{P(x)}\overset{d}{=}\mathrm{Tr}_{se}(E_{z}\mathcal{G
_{t+\tau ,t}^{se}[\rho _{\breve{y}}\otimes \mathrm{Tr}_{s}(E_{\breve{y}
\mathcal{G}_{t,0}^{se}[\rho _{x}^{se}])]), \label{Joint_d}
\end{equation
while in the random scheme it is~\cite{BIF
\begin{equation}
\frac{P(z,\breve{y},x)}{P(x)}\overset{r}{=}\mathrm{Tr}_{se}(E_{z}\mathcal{G
_{t+\tau ,t}^{se}[\rho _{\breve{y}}\otimes \mathrm{Tr}_{s}(\mathcal{G
_{t,0}^{se}[\rho _{x}^{se}])])\wp (\breve{y}|x). \label{Joint_r}
\end{equation
In these expressions, $\mathcal{G}_{t+\tau ,t}^{se}$ is the \textit
bipartite propagator} between $t$ and $t+\tau .$ In addition, $E_{m}\equiv
|m\rangle \langle m|$ and $\rho _{m}\equiv |m\rangle \langle m|$ $[m=z
\breve{y},x]$ represent the (positive) effect measurement operators and
post-measurement states respectively. The set $\{|m\rangle \}$ $[m=z,\breve{
},x]$ are the eigenstates of each measured observable. Furthermore, $\rho
_{x}^{se}\equiv \rho _{x}\otimes \rho _{0}^{e}$ and $P(x)=\langle x|\rho
_{0}^{s}|x\rangle .$ The random scheme is parametrized by an arbitrary
conditional probability $\wp (\breve{y}|x)$ that defines the change in the
system state $(y\rightarrow \breve{y})$ after the intermediate measurement.
The different dependence of both approaches on the bipartite propagator
leads to strong different conclusions about memory effects and information
flows, which are analyzed in the next section.
\section{Comparing both approaches}
In order to perform a systematic comparison we consider different
system-environment models and approximations. In general, we assume that the
bipartite system-environment state $\rho _{t}^{se}$ evolves a
\begin{equation}
\frac{d}{dt}\rho _{t}^{se}=(\mathcal{L}_{s}+\mathcal{L}_{e}+\mathcal{L
_{se})[\rho _{t}^{se}], \label{BipartiteEvolution}
\end{equation
where $\mathcal{L}_{s}$ and $\mathcal{L}_{e}$ define the self-dynamics of
the system and the environment respectively, while $\mathcal{L}_{se}$\
defines their mutual interaction. This interaction term may be unitary or
includes dissipative couplings.
\subsection{Born-Markov approximation}
For systems weakly coupled to their environments, the Born-Markov
approximation \cite{breuerbook} allows to write the bipartite state a
\begin{equation}
\rho _{t}^{se}\simeq \rho _{t}^{s}\otimes \rho _{0}^{e}, \label{BMA}
\end{equation
where $\rho _{t}^{s}$ is the system state, while $\rho _{0}^{e}$ is the
(almost) unperturbed environment state.
When this approximation is valid, in the \textit{non-operational approach},
it is simple to check that Eq.~(\ref{General}) reduces to Eq.~(\re
{Contraction}). In fact, $\mathbf{D}(\rho _{t}^{e},\sigma _{t}^{e})=\mathbf{
}(\rho _{t}^{se},\rho _{t}^{s}\otimes \rho _{t}^{e})=\mathbf{D}(\sigma
_{t}^{se},\sigma _{t}^{s}\otimes \sigma _{t}^{e})=0.$ Furthermore, $\rho
_{t}^{s}$ can be well approximated by a Lindblad equation, which guarantees
the absence of any revival in $\mathbf{D}(\rho _{t}^{s},\sigma _{t}^{s}).$
Thus, the dynamics is Markovian.
In the \textit{operational approach}, by introducing the approximation~(\re
{BMA}) into Eqs.~(\ref{Joint_d}) and (\ref{Joint_r}) straightforwardly it
follows that $C_{pf}(t,\tau )|_{\breve{y}}\overset{d/r}{=}0$ [Eq.~(\re
{CPFCorre})]. These results are independent of which observables are
measured. Thus, the dynamics is Markovian.
In this case [Eq.~(\ref{BMA})] both approaches coincides. Strong differences
appear in the cases studied below.
\subsection{Casual bystander environments}
A wide class of \textquotedblleft non-Markovian\textquotedblright\ dynamics
can be derived by assuming that the system interacts with a
\textquotedblleft casual bystander\textquotedblright\ environment. These
baths are defined by the independence of their marginal states $\rho
_{t}^{e}=\mathrm{Tr}_{s}[\rho _{t}^{se}]$ of any degree of freedom of the
system. Alternatively, the time evolution of $\rho _{t}^{e}$ can be written
in the environment Hilbert space without involving any operator or state of
the system. These properties must be valid for arbitrary system and
environment (separable) initial conditions.
For fulfilling the previous properties, the interaction term $\mathcal{L
_{se}$ in the general evolution~(\ref{BipartiteEvolution}) must be
restricted such tha
\begin{equation}
\mathrm{Tr}_{s}(\mathcal{L}_{se}[\rho _{t}^{se}])=\mathcal{A}[\rho _{t}^{e}],
\label{CBE}
\end{equation
where $\mathcal{A}$\ is an arbitrary superoperator acting on $\rho _{t}^{e}$
that does not have any dependence on the system degrees of freedom. In
general, this constraint can only be satisfied by dissipative (non-unitary)
system-environment couplings. On the other hand, the bath dynamics can be
quantum \cite{casual} or a classical (incoherent) one~\cite{maximal}.
In the \textit{non-operational approach}, the independence of the
environment state on the system degrees of freedom cannot be translated to
any restriction on the inequality defined by Eq.~(\ref{General}). In fact,
under the constraint~(\ref{CBE}), the TD may or not present revivals,
\textit{property that can only be cheeked for each specific model}. Thus,
some dynamics are classified as Markovian and other as non-Markovian. The
unique simplification that can be introduced is to assume that the
environment state does not evolve in time, $\rho _{t}^{e}=\rho _{0}^{e},$
that is, the environment begins in its stationary state. In this case, Eq.~
\ref{General}) reduces t
\begin{eqnarray}
\mathbf{D}(\rho _{t+\tau }^{s},\sigma _{t+\tau }^{s})-\mathbf{D}(\rho
_{t}^{s},\sigma _{t}^{s}) &\leq &\mathbf{D}(\rho _{t}^{se},\rho
_{t}^{s}\otimes \rho _{0}^{e}) \notag \\
&&\!\!\!\!\!+\mathbf{D}(\sigma _{t}^{se},\sigma _{t}^{s}\otimes \rho
_{0}^{e}). \label{TdStacion}
\end{eqnarray
Even in this case $(\rho _{t}^{e}=\rho _{0}^{e}),$ the TD may or not present
revivals, that is, depending on the model, the system may be classified as
Markovian or non-Markovian.
In Eq. (\ref{TdStacion}), any environment-to-system backflow of information
can be related to the establishing of the correlations $\mathbf{D}(\rho
_{t}^{se},\rho _{t}^{s}\otimes \rho _{0}^{e})+\mathbf{D}(\sigma
_{t}^{se},\sigma _{t}^{s}\otimes \rho _{0}^{e}).$ Certainly, the
system-environment correlations (always) changes in time.
Nevertheless, even when there is not revivals in the TD system-environment
correlations are established. This feature represent a central problem for
the interpretation of this approach. In addition, here the environment
state is completely independent of the system (and even of time). Thus, the
revivals of the TD must be taken as a (mathematical) model-dependent
property whose origin cannot be related to any physical process that implies
a \textit{physical} transfer of information from the environment to the
system.
A different perspective emerges in the \textit{operational approach}. By
using the independence of the environment state $[\rho _{t}^{e}=\mathrm{Tr
_{s}(\rho _{t}^{se})]$ of any degree of freedom of the system, it is
possible to check that the joint probability~(\ref{Joint_d}) of the
deterministic scheme \textit{does not fulfill} the Markov property~(\re
{MarkovProb}). In contrast, it is simple to check that the joint
probability~(\ref{Joint_r}) of the random scheme \textit{fulfill} the Markov
property~(\ref{MarkovProb}). Consequently, a casual bystander environment
leads to\ the CPF correlations [Eq.~(\ref{CPFCorre})
\begin{equation}
C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{\neq }0,\ \ \ \ \ \ \ \ \ \
C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{=}0. \label{CPFforCBE}
\end{equation
In this approach, the property $C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{\neq
}0,$ valid for any model under the constraint~(\ref{CBE}), implies that the
system dynamics is non-Markovian. Its origin can be related to
the establishing of (arbitrary) system-environment correlations. On the
other hand, the property $C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{=}0,$
which is valid for arbitrary measurement processes and specific models, is
read as the \textit{absence} of bidirectional system-environment information
flows. In fact, given that the environment is characterized by a
self-dynamics that is completely independent of the system, any
environment-to-system backflow of information (as detected in the
non-operational approach) does not rely on any physical process that affect
the environment state neither its dynamics.
The meaning of the previous analysis is clarified by specifying different
bipartite models that fulfill the evolution (\ref{BipartiteEvolution}) and
the constraint~(\ref{CBE}).
\subsubsection{Classical mixture of quantum Markovian dynamics}
Given a set of \textit{different} system Lindblad superoperators $\{\mathcal
L}_{s}^{c}\},$ which may include both unitary and dissipative contributions,
and given a set of normalized positive weights $\{p_{c}\},$
\sum_{c}p_{c}=1, $ a classical statistical mixture of Markovian dynamics is
defined by the bipartite stat
\begin{equation}
\rho _{t}^{se}=\sum_{c}\exp (t\mathcal{L}_{s}^{c})[\rho _{0}^{s}]\otimes
p_{c}|c\rangle \langle c|. \label{Superposition}
\end{equation
Here, $\{|c\rangle \langle c|\}$ is a set of projectors associated to the
environment space. The marginal system and environment states rea
\begin{equation}
\rho _{t}^{s}=\sum_{c}p_{c}\exp (t\mathcal{L}_{s}^{c})[\rho _{0}^{s}],\ \ \
\ \ \ \rho _{t}^{e}=\sum_{c}p_{c}|c\rangle \langle c|. \label{SystemMixture}
\end{equation
Memory effects in this kind of non-Markovian system dynamics has been
explored in the literature \cite{poland}. Notice that the environment does
not have any dynamics. Even more, the system dynamics can be performed by
mixing in a random way (with weight $p_{c})$ each of the evolved Markovian
system states $\exp (t\mathcal{L}_{s}^{c})[\rho _{0}^{s}].$ Thus, the
detection of an environment-to-system backflow of information via Eq.~(\re
{Revivals}) seems to have a formal mathematical interpretation rather than a
physical one. On the other hand, in the operational approach this case is
characterized by Eq.~(\ref{CPFforCBE}), which guaranty the presence of
memory effects $C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{\neq }0,$ but not
any bidirectional information flow, $C_{pf}(t,\tau )|_{\breve{y}}\overset{r}
=}0.$
\subsubsection{Interaction with stochastic classical degrees of freedom}
When the system interacts with stochastic classical degrees of freedom, the
bipartite state can be written a
\begin{equation}
\rho _{t}^{se}=\sum_{c}\rho _{t}^{c}\otimes p_{c}(t)|c\rangle \langle c|.
\label{BipartitoClassical}
\end{equation
In contrast to the previous case [Eq.~(\ref{Superposition})], the weights
\{p_{c}(t)\}$ are time-dependent and the evolution of the states $\{\rho
_{t}^{c}\}$ may involve coupling between all of them. In fact, under the
constraint~(\ref{CBE}), the more general evolution can be written as~\cit
{maximal
\begin{equation}
\frac{d\tilde{\rho}_{t}^{c}}{dt}=\mathcal{L}_{s}^{c}[\tilde{\rho
_{t}^{c}]-\sum_{c^{\prime }}\gamma _{c^{\prime }c}\tilde{\rho
_{t}^{c}+\sum_{c^{\prime }}\gamma _{cc^{\prime }}\mathbb{S}_{cc^{\prime }}
\tilde{\rho}_{t}^{c^{\prime }}]. \label{EvolveClassical}
\end{equation
Here, $\tilde{\rho}_{t}^{c}\equiv p_{c}(t)\rho _{t}^{c}.$ Thus, $p_{c}(t)
\mathrm{Tr}_{s}(\tilde{\rho}_{t}^{c}).$ Furthermore, $\{\mathbb{S
_{cc^{\prime }}\}$ are arbitrary completely positive system transformations,
which are trace preserving, $\mathrm{Tr}_{s}(\mathbb{S}_{cc^{\prime }}[\rho
])=\mathrm{Tr}_{s}(\rho ).$ Consequently, the environment probabilities
\{p_{c}(t)\}$ obey a classical master equatio
\begin{equation}
\frac{dp_{c}(t)}{dt}=-\sum_{c^{\prime }}\gamma _{c^{\prime
}c}p_{c}(t)+\sum_{c^{\prime }}\gamma _{cc^{\prime }}p_{c^{\prime }}(t),
\end{equation
which in turn shows the role played by the coupling rates $\{\gamma
_{c^{\prime }c}\}.$ In contrast, the system dynamics depart from a Markovian
(Lindblad) evolution. From some specific models, it is possible to recover
some phenomenological non-Markovian master equations (see for example \cit
{lidar}).
In the non-operational\ approach it is very difficult to predict if a given
dynamics [Eq.~(\ref{BipartitoClassical})] lead or not to revivals in the TD.
If the incoherent degrees of freedom begin in their stationary state,
p_{c}(0)=\lim_{t\rightarrow \infty }p_{c}(t),$ one is confronted with the
bounds defined by Eq.~(\ref{TdStacion}). Even in this case, one cannot
predict when there exist or not an environment-to-system backflow of
information.
Interestingly, the origin of the contributions $\mathbf{D}(\rho
_{t}^{se},\rho _{t}^{s}\otimes \rho _{0}^{e})+\mathbf{D}(\sigma
_{t}^{se},\sigma _{t}^{s}\otimes \rho _{0}^{e})$ in Eq.~(\ref{TdStacion})
[or in general in Eq.~(\ref{General})] can be easily read from the
evolution~(\ref{EvolveClassical}). In fact, this equation shows that the
system evolution is \textit{totally conditioned} to the environment
dynamics. The contributions $\mathcal{L}_{s}^{c}$ are\ \textquotedblleft
active\textquotedblright\ whenever the environment is in the state
|c\rangle \langle c|.$ Furthermore, the system suffers the transformation
\rho \rightarrow \mathbb{S}_{c^{\prime }c}[\rho ]$ whenever the environment
\textquotedblleft jumps\textquotedblright\ between the states $c\rightarrow
c^{\prime }.$ This is the physical mechanism that leads to the
system-environment correlations, which in turn does not imply any
system-dependent change in the environment state or dynamics. Thus, the
interpretation of revivals in the TD as environment-to-system backflow of
information is again controversial.
Independently of the Lindblad contributions $\{\mathcal{L}_{s}^{c}\},$ the
superoperators $\{\mathbb{S}_{cc^{\prime }}\},$ and rates $\{\gamma
_{c^{\prime }c}\},$ the operational approach is characterized by Eq.~(\re
{CPFforCBE}), that is, the dynamics is non-Markovian $[C_{pf}(t,\tau )|_
\breve{y}}\overset{d}{\neq }0]$ without the developing of any bidirectional
system-environment information flow $[C_{pf}(t,\tau )|_{\breve{y}}\overset{r
{=}0].$
\subsubsection{Environmental quantum degrees of freedom}
The condition Eq.~(\ref{CBE}) can be satisfied even when the environment is
a quantum one, that is, it develops coherent behaviors. In this case, the
bipartite state can be written a
\begin{equation}
\rho _{t}^{se}=\sum \rho _{t}^{c}\otimes p_{c}(t)|c_{t}\rangle \langle
c_{t}|. \label{BipartitoCBE}
\end{equation
In contrast to Eq.~(\ref{BipartitoClassical}), due to the quantum nature of
the environment, the projectors $\{|c_{t}\rangle \langle c_{t}|\}$ are
time-dependent. In fact, they define the base in which the environment
density matrix $\rho _{t}^{e}$ is diagonal. The more general bipartite
evolution~(\ref{BipartiteEvolution}) under the constraint~(\ref{CBE}), in
its diagonal representation, is given by \cite{casual
\begin{eqnarray}
\frac{d}{dt}\rho _{t}^{se} &=&(\mathcal{L}_{s}+\mathcal{L}_{e})[\rho
_{t}^{se}]+\sum_{\alpha }\Gamma _{\alpha }\ B_{\alpha }\mathbb{S}_{\alpha
}[\rho _{t}^{se}]B_{\alpha }^{\dag } \notag \\
&&-\frac{1}{2}\sum_{\alpha }\Gamma _{\alpha }\{B_{\alpha }^{\dag }B_{\alpha
},\rho _{t}^{se}\}_{+}, \label{GeneralSU}
\end{eqnarray
where $\{\cdot ,\cdot \}_{+}$ is an anticommutator operation. Furthermore,
\{B_{\alpha }\}$ are arbitrary environment operators, while $\mathbb{S
_{\alpha }$ are completely positive trace-preserving system superoperators.
The rates $\{\Gamma _{\alpha }\}$ set the environment dynamics. In fact
\begin{equation}
\frac{d}{dt}\rho _{t}^{e}=\mathcal{L}_{e}[\rho _{t}^{e}]+\sum_{\alpha
}\Gamma _{\alpha }\ (B_{\alpha }\rho _{t}^{e}B_{\alpha }^{\dag }-\frac{1}{2
\{B_{\alpha }^{\dag }B_{\alpha },\rho _{t}^{e}\}_{+}), \label{BathLindblad}
\end{equation
which is a Lindblad dynamics completely independent of the system degrees of
freedom. These evolutions recover, as particular cases, some
phenomenological collisional models introduced in the literature (see for
example \cite{collision}).
The physical interpretation of the evolution~(\ref{GeneralSU}) is quite
similar to that of Eq.~(\ref{EvolveClassical}). In fact, here the
application of the system superoperators $\mathbb{S}_{\alpha }$ occur
whenever the environment suffer a transition associated to the operators
B_{\alpha }.$ This (unidirectional) mechanism defines how the
system-environment correlations are build up.
In the non-operational approach, even when the environment begins in its
stationary state $\rho _{0}^{e}=\lim_{t\rightarrow \infty }\rho _{t}^{e}$
[where $\rho _{t}^{e}$ obeys Eq.~(\ref{BathLindblad})], it is not possible
to infer for an arbitrary model the presence or absence of revivals in the
TD [Eq.~(\ref{Revivals})]. In contrast, the operational approach is still
characterized by Eq.~(\ref{CPFforCBE}).
\subsection{Unitary system-environment interactions}
Independently of the specific models, the correlation between the system and
the casual bystander environments introduced previously does
not involve quantum entanglement \cite{entanglement} [see the separable
states Eqs.~(\ref{Superposition}), (\ref{BipartitoClassical}), and (\re
{BipartitoCBE})]. In contrast, quantum entanglement may emerges when
considering Hamiltonian (time-reversible) system-environment interactions.
In fact, solely for special system-environment initial conditions a
bipartite unitary dynamics does not induce quantum entanglement \cite{katarzyna}.
The total Hamiltonian is written a
\begin{equation}
H_{T}=H_{s}+H_{e}+H_{I}. \label{Htotal}
\end{equation
Each contribution correspond respectively to the system, environment, and
interaction Hamiltonians respectively. The bipartite propagator i
\begin{equation}
\mathcal{G}_{t,t_{0}}^{se}[\bullet ]=\exp [-i(t-t_{0})H_{T}]\bullet \exp
[+i(t-t_{0})H_{T}]. \label{Gunitario}
\end{equation}
In the\textit{\ non-operational approach}, each contribution in the rhs of
Eq.~(\ref{General}) make complete sense in this context. In fact, almost all
unitary interactions lead to a change in the environment state and also
induce the developing of (arbitrary) system-environment correlations. When
revivals in the TD develops, Eq.~(\ref{General}) defines a bound with a
clear physical meaning. Nevertheless, in general it is not possible to infer
which kind of dynamics develop or not revivals in the TD. Even for a given
(Hamiltonian) model, depending on the underlying parameters the system
dynamics may be Markovian or not. Consequently, it is not clear which
physical property defines the boundary between Markovian and non-Markovian
dynamics.
In the \textit{operational approach}, given that the state and dynamics of
the environment are in general modified by a unitary interaction, instead of
Eq.~(\ref{CPFforCBE}), here it follow
\begin{equation}
C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{\neq }0,\ \ \ \ \ \ \ \ \ \
C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{\neq }0. \label{CPFUnitary}
\end{equation
Both inequalities can be supported by performing a perturbation theory based
on projectors techniques \cite{bonifacio}. Consistently, it has been shown
that even close to the validity of a Born-Markov approximation the
operational approach can detect memory effects \cite{rio}.
The inequality $C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{\neq }0$ implies
that the system dynamics is non-Markovian (system-environment correlations
are developed during the evolution), while $C_{pf}(t,\tau )|_{\breve{y}
\overset{r}{\neq }0$ detects the presence of bidirectional information
flows. In fact, here the environment state and evolution always depend on
the system degrees of freedom.
There exist a unique exception to Eq.~(\ref{CPFUnitary}), which reduces to
Eq.~(\ref{CPFforCBE}). Hence, even when the environment state is modified,
for any system observables one obtain $C_{pf}(t,\tau )|_{\breve{y}}\overset{
}{=}0.$ While this property is certainly\textit{\ undesirable}, this case
has a clear physical interpretation. It emerges when, in a given
environmental base $\{|e\rangle \},$ the diagonal part of the bipartite
propagator~(\ref{Gunitario}) can be written a
\begin{equation}
\langle e|\mathcal{G}_{t,0}^{se}[\bullet ]|e\rangle =\mathcal{T
_{t,0}^{(e)}\langle e|\bullet |e\rangle , \label{DiagonalCondition}
\end{equation
where $\mathcal{T}_{t,0}^{(e)}$ is a \textit{system} (density matrix)
propagator that parametrically depends on each environmental state
\{|e\rangle \}.$ The condition~(\ref{DiagonalCondition}) is fulfilled,
\textit{for example}, when the environment and interaction Hamiltonians
commutat
\begin{equation}
\lbrack H_{e},H_{I}]=0. \label{Conmutador}
\end{equation}
Introducing the condition~(\ref{DiagonalCondition}) into Eqs.~(\ref{Joint_d
) and~(\ref{Joint_r}) it is possible to check that $C_{pf}(t,\tau )|_{\breve
y}}\overset{d}{\neq }0,$\ and $C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{=}0.$
This last equality \textit{does not imply} that the environment in not
affected. It emerges because the system state assumes the structur
\begin{equation}
\rho _{t}^{s}=\mathrm{Tr}_{e}(\mathcal{G}_{t,0}^{se}[\rho _{0}^{s}\otimes
\rho _{0}^{e}])=\sum_{e}\langle e|\rho _{0}^{e}|e\rangle \mathcal{T
_{t,0}^{(e)}[\rho _{0}^{s}]. \label{SystemRandomUnitary}
\end{equation
Therefore, the system evolution can be written as a statistical
superposition of unitary maps, quite similar to Eq.~(\ref{SystemMixture}).
In consequence, for unitary system-environment models the condition
C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{=}0$ allows to detect when the
system dynamics (even between measurements) can be represented by a \textit
Hamiltonian ensemble}, property that has been of interest in the recent
literature~\cite{nori}.
\section{Example}
In this section we consider an explicit example of the dynamics discussed
previously. The quantum system $(s),$ taken for simplicity as a two-level
system, interacts with an incoherent environment $(e)$ (see Sect. 3.2.2),
which here is defined by four discrete states denoted as $\{|1\rangle
,|2\rangle ,|3\rangle ,|4\rangle \}.$ Correspondingly, the bipartite
system-environment state is written a
\begin{equation}
\rho _{t}^{se}=\sum_{k=1,2,3,4}\tilde{\rho}_{k}(t)\otimes |k\rangle \langle
k|. \label{Bipartita}
\end{equation
The system and environment states then rea
\begin{equation}
\rho _{t}^{s}=\sum_{k=1,2,3,4}\tilde{\rho}_{k}(t),\ \ \ \ \ \ \ \rho
_{t}^{e}=\sum_{k=1,2,3,4}p_{k}(t)|k\rangle \langle k|, \label{RhoSystema}
\end{equation
where $p_{k}(t)=\mathrm{Tr}_{s}[\tilde{\rho}_{k}(t)].$ The evolution of the
unnormalized system states $\{\tilde{\rho}_{k}(t)\}_{k=1}^{k=4}$ is taken as
\begin{subequations}
\label{EvolucionRhoTilde}
\begin{eqnarray}
\frac{d\tilde{\rho}_{4}(t)}{dt} &=&-\gamma \tilde{\rho}_{4}(t)+\phi
\sum_{k=1,2,3}\sigma _{k}\tilde{\rho}_{k}(t)\sigma _{k}, \\
\frac{d\tilde{\rho}_{k}(t)}{dt} &=&-\phi \tilde{\rho}_{k}(t)+\left( \frac
\gamma }{3}\right) \sigma _{k}\tilde{\rho}_{4}(t)\sigma _{k},\ \ \ k=1,2,3.\
\ \ \ \ \ \ \
\end{eqnarray
In this expression, $\gamma $ and $\phi $\ are characteristic coupling
rates. Furthermore, the set of Pauli matrixes is denoted as $(\sigma
_{x},\sigma _{y},\sigma _{z},\mathrm{I})\leftrightarrow (\sigma _{1},\sigma
_{2},\sigma _{3},\sigma _{4}),$ where\ $\mathrm{I}$ is the identity\ matrix
in the two-dimensional system Hilbert space. From Eq.~(\re
{EvolucionRhoTilde}), the evolution of the environment populations is
defined by the following classical master equation
\end{subequations}
\begin{subequations}
\label{Poblaciones}
\begin{eqnarray}
\frac{dp_{4}(t)}{dt} &=&-\gamma p_{4}(t)+\phi \sum_{k=1,2,3}p_{k}(t), \\
\frac{dp_{k}(t)}{dt} &=&-\phi p_{k}(t)+\left( \frac{\gamma }{3}\right)
p_{4}(t),\ \ \ k=1,2,3.\
\end{eqnarray
This equation is completely independent of the system degrees of freedom.
Thus, the evolution (\ref{EvolucionRhoTilde}) has a simple interpretation.
When the environment suffer the transition $|4\rangle \overset{\gamma /3}
\rightarrow }|k\rangle $ or the transition $|k\rangle \overset{\phi }
\rightarrow }|4\rangle $ $(k=1,2,3),$ the transformation $\sigma _{k}\bullet
\sigma _{k}$ is \textit{conditionally} applied over the open quantum system.
Eq.~(\ref{EvolucionRhoTilde}) can be solved after specifying the bipartite
initial conditions. We consider a separable state, $\rho _{0}^{se}=\rho
_{0}^{s}\otimes \rho _{0}^{e},$ which implies $\tilde{\rho}_{k}(0)=\rho
_{0}^{s}p_{k}(0).$ In general, each auxiliary state $\tilde{\rho}_{k}(t)$
can be written as a superposition of the Pauli channels acting on the
initial system state $\rho _{0}^{s},$ that is,
\end{subequations}
\begin{equation}
\tilde{\rho}_{k}(t)=\sum_{j=1,2,3,4}g_{k}^{j}(t)\sigma _{j}\rho
_{0}^{s}\sigma _{j}, \label{RhoTildeK}
\end{equation
where $\{g_{k}^{j}(t)\}$ are (sixteen) scalar functions that depend on time.
Their initial conditions are $g_{k}^{4}(0)=p_{k}(0)$ and $g_{k}^{j}(0)=0,$
with $j=1,2,3,$ and $k=1,2,3,4.$ The evolution of the set $\{g_{k}^{j}(t)\}$
follows after inserting the previous expression for $\tilde{\rho}_{k}(t)$
into Eq.~(\ref{EvolucionRhoTilde}). Consistently with their definition,
p_{k}(t)=\mathrm{Tr}_{s}[\tilde{\rho}_{k}(t)],$ the environment populations
are recovered a
\begin{equation}
p_{k}(t)=\sum_{j=1,2,3,4}g_{k}^{j}(t).
\end{equation}
\subsection{Depolarizing dynamics}
The evolution of the auxiliary states Eq.~(\ref{EvolucionRhoTilde}) is
(structurally) the same for the states $\{\tilde{\rho}_{1}(t),\tilde{\rho
_{2}(t),\tilde{\rho}_{3}(t)\}.$ Thus, if we consider environment initial
conditions where $p_{1}(0)=p_{2}(0)=p_{3}(0),$ from Eqs.~(\ref{RhoSystema})
and (\ref{RhoTildeK}) it follows that the solution map $\rho
_{0}^{s}\rightarrow \rho _{t}^{s}$ must be a depolarizing channel \cit
{nielsen}, that is
\begin{equation}
\rho _{t}^{s}=w(t)\rho _{0}^{s}+\frac{1-w(t)}{3}\sum_{k=1,2,3}\sigma
_{k}\rho _{0}^{s}\sigma _{k}, \label{DepoRho}
\end{equation
where the positive weight $w(t),$\ from Eq.~(\ref{RhoTildeK}), follows a
\begin{equation}
w(t)=\sum_{k=1,2,3,4}g_{k}^{4}(t).
\end{equation
Consistently, $[1-w(t)]/3=\sum_{k=1,2,3,4}g_{k}^{j}(t),$ with $j=1,$ $2,$
3. $
The more natural initial conditions for the environment are their stationary
populations $p_{k}^{\infty }\equiv \lim_{t\rightarrow \infty }p_{k}(t),$
where $p_{k}(t)$ are defined by Eq.~(\ref{Poblaciones}). Straightforwardly,
we get
\begin{equation}
p_{4}^{\infty }=\frac{\phi }{\gamma +\phi },\ \ \ \ \ \ \ \ p_{k}^{\infty }
\frac{1}{3}\frac{\gamma }{\gamma +\phi }\ \ \ \ \ (k=1,2,3).
\label{Estacionarias}
\end{equation
Under the assumption $p_{k}(0)=p_{k}^{\infty },$ after getting the set
\{g_{k}^{j}(t)\}$ in an explicit analytical way, the function $w(t)$ that
characterizes the depolarizing channel Eq.~(\ref{DepoRho}) can be written a
\begin{equation}
w(t)=\frac{(\gamma ^{2}+3\phi ^{2})}{3(\gamma +\phi )^{2}}+\frac{4\gamma
\phi }{3(\gamma +\phi )^{2}}e^{-(\gamma +\phi )t}+\frac{2\gamma }{3(\gamma
+\phi )}e^{-\phi t}, \label{W(t)}
\end{equation
which consistently satisfies $w(0)=1.$ Furthermore, $\lim {}_{t\rightarrow
\infty }w(t)\neq 0.$ On the hand, the environment dynamics is stationary,
that is, $p_{k}(t)=p_{k}(0)=p_{k}^{\infty }$ [Eq.~(\ref{Estacionarias})].
\subsection{Operational vs. non-operational quantum non-Markovianity}
In the\textit{\ non-operational approach}, quantum non-Markovianity is
defined by the revivals in the trace distance between two different initial
states, Eq.~(\ref{Revivals}). By using that $(\mathrm{I}/2)=(\rho
+\sum_{k=1,2,3}\sigma _{k}\rho \sigma _{k})/4$ \cite{nielsen},\ the
depolarizing map (\ref{DepoRho}) can be rewritten as $\rho _{t}^{s}=w(t)\rho
_{0}^{s}+(1/3)[1-w(t)](2\mathrm{I}-\rho _{0}^{s}).$ Thus, the trace distance
straightforwardly can be written a
\begin{equation}
D[\rho _{t}^{s},\sigma _{t}^{s}]=\left\vert \frac{4w(t)-1}{3}\right\vert
D[\rho _{0}^{s},\sigma _{0}^{s}]\equiv d(t)D[\rho _{0}^{s},\sigma _{0}^{s}]
\label{TraceDistance}
\end{equation
where $D[\rho _{0}^{s},\sigma _{0}^{s}]$ is the trace distance between the
two initial states $\rho _{0}^{s}$ and $\sigma _{0}^{s}.$ Notice that the
decay of the trace distance does not depends on the initial states, being
dictated by the function $d(t).$
In Fig.~1(a) we plot the function $d(t)$ for different values of the
characteristic parameter $\phi /\gamma .$ As expected from Eq.~(\ref{W(t)}),
$D[\rho _{t}^{s},\sigma _{t}^{s}]$ decay in a monotonous way without
developing any revival. Thus, under the trace distance criteria, the
dynamics is \textit{Markovian} and there is not any environment-to-system
backflow of information. Nevertheless, notice that for any value of $\phi
/\gamma $ system-environment correlations are built up during the dynamics
[see Eq.~(\ref{Bipartita})]. This feature, which is irrelevant for the TD
decay behavior, is relevant for the CPF correlation.
\begin{figure}[tbp]
\includegraphics[bb=22 350 520 555,angle=0,width=8.5cm]{Figura1_BF.eps}
\caption{(a) Decay of the trace distance $d(t)$ [Eq.~(\protect\re
{TraceDistance})] corresponding to the model (\protect\ref{EvolucionRhoTilde
). (b) Time dependence of the CPF correlation $C_{pf}(t,t)|_{\breve{y}}$ in
the deterministic scheme \protect\cite{CPF} corresponding to the same model.
The value of the quotient $\protect\phi /\protect\gamma $ is indicated in
each plot.}
\end{figure}
In the \textit{operational approach}, the presence of memory effects is
witnessed by the CPF correlation [Eq.~(\ref{CPFCorre})] in the deterministic
scheme. We assume that the three measurements are projective ones, all of
them being performed in the $z$-direction of the Bloch sphere. Furthermore
the initial condition of the system is taken as $\rho _{0}^{s}=|\psi \rangle
\langle \psi |,$ where $|\psi \rangle $ is an eigenstate of the $x$-Pauli
matrix. Explicit general expression for $C_{pf}(t,\tau )|_{\breve{y}}$ in
terms of the coefficients $\{g_{k}^{j}(t)\}$ can be found in Ref.~\cit
{multi} (see corresponding Appendix D). Under the previous assumptions, the
CPF correlation can be obtained in an analytical way, which is written in
\cite{CPF}. Simple expressions are obtained for specific values of the decay
rates. For example, for $\phi =\gamma $ it follow
\begin{eqnarray}
&&C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{=}\frac{4}{81}(1-e^{-\gamma
t})(1-e^{-\gamma \tau }) \\
&&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \times (2+e^{-\gamma t}+e^{-\gamma
\tau }+5e^{-\gamma (t+\tau )}). \notag
\end{eqnarray
Due to the symmetry of the problem, in all cases $C_{pf}(t,\tau )|_{\breve{y
}$ does not depend on the value of the conditional $\breve{y}=\pm 1.$ In
Fig.~1(b) we plot the CPF correlation at equal times, $C_{pf}(t,t)|_{\breve{
}},$ for different values of $\phi /\gamma .$ In contrast to the
non-operational approach, here for all possible values of the characteristic
parameter $\phi /\gamma $ it is fulfilled $C_{pf}(t,\tau )|_{\breve{y}
\overset{d}{\neq }0,$ which indicates a \textit{non-Markovian} regime. In
fact, the system is strongly correlated with the environment [Eq.~(\re
{Bipartita})].
The system-environment correlations emerges due to a unidirectional
dependence of the system dynamics on the environment transitions [Eq.~(\re
{EvolucionRhoTilde})]. In fact, the environment populations do not depend on
the system degrees of freedom [see Eq.~(\ref{Poblaciones})]. These
properties are relevant in the random scheme and imply that $C_{pf}(t,\tau
)|_{\breve{y}}\overset{r}{=}0$ [Eq.~(\ref{CPFforCBE})]. This result is valid
for arbitrary measurement processes, indicating in the operational approach,
the absence of any environment-to-system backflow of information.
\subsection{Environment-to-system backflow of information}
In the previous section we concluded that both approaches differ in the
classification of the dynamics (Markovian vs. non-Markovian), but (due to
different reasons) agree in the absence of any environment-to-system
backflow of information. Here, we show that in general both approaches also
differ in this last aspect. Different mechanisms can be proposed for getting
a revival in the trace distance Eq.~(\ref{TraceDistance}).
\subsubsection{Slow modulation of the stationary environment state}
First, we consider the same model [Eq.~(\ref{EvolucionRhoTilde})], but in
addition it is assumed that the characteristic rates are time dependent,
\gamma \rightarrow \gamma (t),$ $\phi \rightarrow \phi (t),$ wit
\begin{equation}
\gamma (t)=\gamma \lbrack 1+b(t)]>0,\ \ \ \ \ \ \phi (t)=\phi \lbrack
1-b(t)]>0. \label{RateModulated}
\end{equation
Here, $b(t)$ is an arbitrary function of time that fulfill the constraint
-1<b(t)<1.$ The previous structure is chosen for simplifying the argument
and calculus. Nevertheless, we remark that similar dependences can be
implemented in different experimental situations (see for example Ref.~\cit
{Udo}). The more relevant aspect is that the assumption (\ref{RateModulated
) can be implemented by affecting solely the environmental degrees of
freedom [see Eq.~(\ref{Poblaciones})].
In addition, in Eq.~(\ref{RateModulated}) it is assumed tha
\begin{equation}
\left\vert \frac{d}{dt}b(t)\right\vert \ll \gamma ,\ \ \ \ \left\vert \frac{
}{dt}b(t)\right\vert \ll \phi . \label{slow}
\end{equation
Hence, the time dependence of $b(t)$ can be considered slow with respect to
the decay times $(1/\gamma )$ and $(1/\phi ).$ Consequently, the full
dynamics can be described in an adiabatic approximation, where the full
bipartite system in the long time regime $(\gamma t\gg 1,\ \phi t\gg 1)$
rapidly adjust to the instantaneous values of $\gamma (t)$ and $\phi (t).$
In particular, in this regime, the environment populations, from Eq.~(\re
{Estacionarias}), can be written a
\begin{equation}
p_{4}^{\infty }(t)\simeq \frac{\phi }{\gamma +\phi }[1-b(t)],\ \ \ \
p_{k}^{\infty }(t)\simeq \frac{1}{3}\frac{\gamma }{\gamma +\phi }[1+b(t)],
\label{PkSlow}
\end{equation
where $k=1,2,3.$ For simplicity, we assumed that $(\gamma -\phi )\ll (\gamma
+\phi ),$ which allows to approximate $\gamma (t)+\phi (t)=(\gamma +\phi
)+b(t)(\gamma -\phi )\simeq (\gamma +\phi ).$
In the long time regime, the \textit{non-operational approach} is
characterized by the value $\lim_{t\rightarrow \infty }w(t)\neq 0$ [see
Eqs.~(\ref{W(t)}) and~(\ref{TraceDistance})]. For time-independent rates
this quantity can be written in terms of the stationary populations
\{p_{k}^{\infty }\}_{k=1}^{k=4}$ [Eq.~(\ref{Estacionarias})] as
\lim_{t\rightarrow \infty }w(t)=[p_{4}^{\infty
}]^{2}+\sum_{k=1,2,3}[p_{k}^{\infty }]^{2}.$ Given that in the slow
modulation regime [Eq.~(\ref{slow})] these values become time dependent,
p_{k}^{\infty }\longrightarrow p_{k}^{\infty }(t)$ [Eq. (\ref{PkSlow})], it
follows tha
\begin{equation}
w(t)\overset{slow}{\simeq }[p_{4}^{\infty
}(t)]^{2}+\sum_{k=1,2,3}[p_{k}^{\infty }(t)]^{2},\ \ \ \ \gamma t\gg 1,\
\phi t\gg 1.
\end{equation
Therefore, under the previous hypothesis, the stationary values of the TD in
Fig.~1(a) $[d(t)=\left\vert 4w(t)-1\right\vert /3]$ become proportional to
the arbitrary function $b(t).$ This result implies that one can get \textit
arbitrary revivals in the trace distance} [Eq.~(\ref{TraceDistance})] by
choosing different time-dependences of the function $b(t).$ Alternatively,
an \textit{arbitrary environment-to-system backflow of information can be
produced }by changing solely in a slow way the (\textquotedblleft
stationary\textquotedblright ) environment populations. Nevertheless, we
remark that the full dynamics is essentially the same than in the
static-rate case. While one can associate the revivals in the TD to the
system-environment correlations, these correlations have the same origin and
structure than in absence of revivals, that is Fig.~1(a) (static case) and
when $b(t)$ does not lead to revivals.
In the deterministic scheme, the \textit{operational approach} is
characterized by the stationary value \cite{CPF
\begin{equation}
\lim_{\substack{ t\rightarrow \infty \\ \tau \rightarrow \infty }
C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{=}\frac{8\gamma (\gamma -3\phi
)^{2}(\gamma +3\phi )}{81(\gamma +\phi )^{4}},
\end{equation
which can also be written in terms of $\{p_{k}^{\infty }\}_{k=1}^{k=4}$
[Eq.~(\ref{Estacionarias})]. Thus, under the same conditions that guarantee
the slow modulation regime [Eqs.~(\ref{slow}) and (\ref{PkSlow})], the
stationary values of $C_{pf}(t,t)|_{\breve{y}}$ plotted in Fig.~1(b) also
become proportional to the function $b(t).$ Nevertheless, in this approach
this property does not implies the presence of any backflow of information.
In fact, given that the environment state does not depend at all on the
system degrees of freedom, even in the slow modulation regime, it follows
that $C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{=}0$ [Eq.~(\ref{CPFforCBE})].
In this way, it is clear that both the non-operational and operational
approaches also strongly disagree in this aspect.
\subsubsection{Quantum coherent contributions in the environment dynamics}
The system-environment dynamics associated to the depolarizing channel [Eq.~
\ref{EvolucionRhoTilde})] can alternatively be represented through a
Lindblad equation. In fact, the evolution of the bipartite state $\rho
_{t}^{se}$ can be written a
\begin{eqnarray}
\frac{d\rho _{t}^{se}}{dt} &=&+\frac{\gamma }{3}\sum_{k=1,2,3}(B_{k}\sigma
_{k}[\rho _{t}^{se}]\sigma _{k}B_{k}^{\dagger }-\frac{1}{2}\{B_{k}^{\dagger
}B_{k},\rho _{t}^{se}\}_{+}) \notag \\
&&+\phi \sum_{k=1,2,3}(B_{k}^{\dagger }\sigma _{k}[\rho _{t}^{se}]\sigma
_{k}B_{k}-\frac{1}{2}\{B_{k}B_{k}^{\dagger },\rho _{t}^{se}\}_{+}) \notag \\
&&-i[H_{e},\rho _{t}^{se}], \label{BipartitoLindblad}
\end{eqnarray
where the bath operators are $B_{k}\equiv |k\rangle \langle 4|,$ $k=1,2,3.$
As before, $\{|1\rangle ,|2\rangle ,|3\rangle ,|4\rangle \}$ are the
environment base. Defining the states $\tilde{\rho}_{k}\equiv \langle k|\rho
_{t}^{se}|k\rangle ,$ it is simple to check that the first two lines of the
previous Lindblad dynamics recover the time evolution introduced in Eq.~(\re
{EvolucionRhoTilde}).
From Eq.~(\ref{BipartitoLindblad}) it is simple to check that the bath state
$(\rho _{t}^{e}=\mathrm{Tr}_{s}[\rho _{t}^{se}])$ obeys a Lindblad equation
that, even with the extra contribution $-i[H_{e},\rho _{t}^{se}],$ is
independent of the system degrees of freedom. Thus, the environment is still
a casual bystander one [see Eqs.~(\ref{GeneralSU}) and (\ref{BathLindblad
)]. In order to obtain a (system) depolarizing channel [Eq.~(\ref{DepoRho})]
the symmetry between the bath states $\{|1\rangle ,|2\rangle ,|3\rangle \}$
must be granted. For example, the Hamiltonia
\begin{equation}
H_{e}=\frac{\Omega }{2}\sum_{k=1,2,3}(|k\rangle \langle 4|+|4\rangle \langle
k|),
\end{equation
fulfill this property.
\begin{figure}[tbp]
\includegraphics[bb=22 375 550 595,angle=0,width=8.5cm]{Figura2_BF.eps}
\caption{Decay of the trace distance $d(t)$ [Eqs.~(\protect\re
{TraceDistance}) and (\protect\ref{w_Unitario})] corresponding to the model
\protect\ref{BipartitoLindblad}) with $\protect\phi =\protect\gamma $ for
different values of the Hamiltonian frequency $\Omega /\protect\gamma .$}
\end{figure}
In consistence with the solution defined by Eqs.~(\ref{Bipartita}) and (\re
{RhoTildeK}), here the bipartite state is written a
\begin{equation}
\rho _{t}^{se}=\sum_{k=1,2,3,4}(\sigma _{k}\rho _{0}^{s}\sigma _{k})\otimes
\varrho _{t}^{k},
\end{equation
where $\{\varrho _{t}^{k}\}$ are states in the environment Hilbert space. In
order to obtain analytical treatable solutions we assume the bipartite
initial conditio
\begin{equation}
\rho _{0}^{se}=\rho _{0}^{s}\otimes \rho _{0}^{e}=\rho _{0}^{s}\otimes
|4\rangle \langle 4|. \label{CIN4}
\end{equation
Under this assumption $(\rho _{0}^{e}=|4\rangle \langle 4|),$ given that the
underlying system stochastic dynamics associated to Eq.~(\re
{BipartitoLindblad}) is the same than in the incoherent case [Eq.~(\re
{EvolucionRhoTilde})], it follows that the system state goes back to the
initial condition $\rho _{0}^{s}$ whenever the environment goes back to the
state $|4\rangle .$ This property straightforwardly follows from $\sigma
_{k}^{2}=\mathrm{I}.$ Therefore, under the assumption~(\ref{CIN4}), here the
depolarizing map Eq.~(\ref{DepoRho}) is defined with the functio
\begin{equation}
w(t)=\mathrm{Tr}_{e}[\varrho _{t}^{4}]=\langle 4|\rho _{t}^{e}|4\rangle ,
\label{w_Unitario}
\end{equation
where $\rho _{t}^{e}$ is the density matrix of the environment. Consistently
$[1-w(t)]/3=\mathrm{Tr}_{e}[\varrho _{t}^{k}]=\langle k|\rho
_{t}^{e}|k\rangle ,$ with $k=1,$ $2,$ $3.$ In consequence, the decay of the
trace distance is proportional to the bath population $\langle 4|\rho
_{t}^{e}|4\rangle .$ Its explicit analytical expression is rather complex
and non-informative \cite{p4}.
In this alternative situation, it is clear that $H_{e}$ induces intrinsic
quantum coherent oscillations in the environment dynamics, which in turn may
lead to oscillations in the trace distance [Eq.~(\ref{TraceDistance})]. In
Fig.~2 we plot the TD decay $d(t)=\left\vert 4w(t)-1\right\vert /3$ taking
\phi =\gamma $ and for different values of $\Omega /\gamma .$ When $\Omega
/\gamma <1,$ a monotonous decay is observed. Nevertheless, for $\Omega
/\gamma >1$ \textit{revivals} in the TD are observed.
The CPF correlation in the deterministic scheme cannot be calculated in an
analytical way. Nevertheless, given that the system dynamics is still
controlled by the environment (self) transitions, it follows that
C_{pf}(t,\tau )|_{\breve{y}}\overset{d}{\neq }0.$ Thus, the dynamics becomes
non-Markovian in both approaches $(\Omega /\gamma >1).$ Nevertheless, given
that the environment is a casual bystander one, in the random scheme it is
valid that $C_{pf}(t,\tau )|_{\breve{y}}\overset{r}{=}0$ [Eq.~(\re
{CPFforCBE})] for any value of $\Omega /\gamma .$ Consequently, in the same
way as in the previous model [Eq.~(\ref{RateModulated})] the non-operational
and operational approaches gives different results about the presence of
environment-to-system backflows of information.
\section{Summary and conclusions}
The interpretation of quantum memory effects in terms of an
environment-to-system backflow of information is still under a vivid debate.
In this contribution we presented a partial view of this problem by
comparing how this concept is introduced an interpreted in non-operational
and operational approaches to quantum non-Markovianity.
Our main contribution is a comparison between both formalisms for different
environment models. We considered casual bystander environments, which are
characterized by a density matrix that does not depend on the system degrees
of freedom. This class cover classical statistical mixtures of Markovian
dynamics [Eq.~(\ref{Superposition})], interaction with stochastic classical
degrees of freedom [Eq.~(\ref{BipartitoClassical})], and also environmental
quantum degrees of freedom [Eq.~(\ref{BipartitoCBE})]. In addition, we
considered unitary system-environment models [Eq.~(\ref{Htotal})].
As a non-operational approach we used the TD between two system states with
different initial conditions. This formalism is characterized by the bound
Eq.~(\ref{General}). We have argued that, in general, it is not possible to
predict if for a given model the TD presents or not revivals in its time
behavior. This property is valid for all environmental models. In the case
of casual bystander ones, the previous feature represent an obstacle for
giving a consistent physical interpretation of any environment-to-system
backflow of information defined as revivals in the TD [Eq.~(\ref{Revivals
)]. In fact, for these dynamics the system-environment correlations emerges
due to a unidirectional dependence of the system dynamics in the state of
the environment and its transitions. In particular, for stationary
environments it is not possible to know when the system-environment
correlations lead to the presence or absence of backflows of information.
The possibility of getting monotonous decay behaviors of the TD for unitary
interaction models also represents an undesirable property because in
general the environment state is modified by its interaction with the system.
As an operational approach we used a CPF correlation [Eq.~(\ref{CPFCorre})],
which is defined by three consecutive system measurement processes. Both a
deterministic and random schemes were considered [with associated joint
probabilities Eqs.~(\ref{Joint_d}) and~(\ref{Joint_r})]. In the case of
casual bystander environments, the CPF correlation in the deterministic
scheme does not vanish, while in the random scheme it vanishes identically
for any chosen measurement observables [Eq.~(\ref{CPFforCBE})]. Thus, in
this approach, any casual bystander environment leads to a non-Markovian
system dynamics but not any bidirectional information flow is detected. In
the case of Hamiltonian models, in general, in both schemes the CPF
correlation does not vanish indicating a non-Markovian system dynamics and
the presence of bidirectional information flows [Eq.~(\ref{CPFUnitary})]. An
undesirable exception to this last property emerges when the system dynamics
can equivalently be represented by a random unitary map [Eqs.~(\re
{DiagonalCondition}) and (\ref{SystemRandomUnitary})].
As an specific example we considered a system coupled to an
environment able to induce a depolarizing dynamics [Eqs.~(\re
{EvolucionRhoTilde}), (\ref{RateModulated}) and (\ref{BipartitoLindblad})].
We have found that both approaches differ in the Markovian and non-Markovian
regimes as well in the presence or absence of environment-to-system
backflows of information.
In general, both operational and non-operational approaches to quantum
non-Markovianity provide necessary and complementary points of view for
defining and understanding memory effects in open quantum systems. The
present results shed light on some conceptual differences and properties of
these approaches. They may be useful for extending the application of these
formalisms for the understanding of memory effects induced by structured or
spatially extended environments.
\section*{Acknowledgments}
This paper was supported by Consejo Nacional de Investigaciones Cient\'{\i
ficas y T\'{e}cnicas (CONICET), Argentina.
|
\section{Introduction}
Rosetta\footnote{\url{https://rosettaccess.fr/index.php/home-page-english/}} is a French project that aimed to study accessibility solutions for audiovisual content. One of the experiments consisted in designing an automatic translation system from text to Sign Language (SL) displayed through animation of a virtual signer.
The three main contributions concerning SL in this project were the constitution of Rosetta-LSF \citelanguageresource{rosetta-lsf}, an aligned corpus of text and SL captured using a mocap system, a translation system from text to AZee (a representation of SL content), and a system allowing to generate virtual signer animations from AZee input.
This article describes the second contribution: the system of translation from text to AZee.
After an overview of the issues and recent works in the field, we explain our method and design choices and describe the implementation of the translation system. Finally, we give preliminary results and discuss the questions raised for evaluation.
\section{Text-to-Sign translation}
The automatic translation of content from a spoken language into a sign language is a fairly recent and still largely unexplored research topic. Here we are interested in the translation of text as the source format, in our case in French, and video or 3D animation as the target format, in our case French Sign Language (LSF).
In this section, we look at the main challenges encountered with text-to-sign translation.
\subsection{Need for bilingual corpora}
Machine translation (MT) was first developed for spoken languages in their written form using bilingual dictionaries and rule-based systems, that were not easy to develop and maintain. Access to parallel corpora of aligned examples has led to the rise of data-driven approaches, such as Statistical Machine Translation (SMT) that used the frequencies of translation pairs containing source--target pairings of words or phrases. In the current dominant approach, Neural Machine Translation (NMT), which is also data-driven, the source text is encoded into an intermediate representation in the form of numerical vectors to be decoded as a target text. Although the representation is not directly open to interpretation, the practical results largely prevail over former strategies. These methods designed for spoken languages rely on the availability of large volumes of parallel data (of the order of several million sentences), Sign languages being considered as low resource in this respect.
Example-based MT (EBMT) is another data-driven approach based on analogy \cite{nagao:ahi84}. It uses a bilingual corpus that contains texts and their translations. Given a sentence to translate, sentences from this corpus are selected that contain similar sub-sentential components. The similar sentences are then used to translate the sub-sentential components of the original sentence into the target language, and these phrases are put together to form a complete translation. Although the larger the corpus, the better the results will be, this approach can be implemented on smaller corpora and thus may be considered in the case of Sign Language Machine Translation (SLMT).
\subsection{Need for an intermediate representation}
One of the major differences between SLMT and written MT is the difference in channel. Written languages are input to MT systems as streams of discrete tokens---words separated by blanks---whereas SL does not have a written form. SLs are to be considered as face-to-face oral languages. Moreover, they are able to convey simultaneous information by the way of a number of articulators, such as the two hands and arms, but also the torso, shoulders, head, gaze and facial expressions (including a number of facial component movements).
As SL has no written form, many approaches proceed in two steps: a first step transforms the SL content into an intermediate representation, and a second step uses this representation as the input of a synthesis system to control the animation of an avatar in order to display the content in SL.
After a first generation of studies based mainly on the rule-based approach \cite{veale1998challenges,zhao2000machine,marshall-04}), a few ones have investigated Example-Based Machine Translation \cite{morrissey2005example}. They have sometimes be combined with statistical approaches, such as in \newcite{demartino2017signing} who have developed a system that automatically translates Brazilian Portuguese text to Brazilian SL (LIBRAS) by combining SMT with EBMT in case of unseen texts or ambiguous terms dependent on the context and frequency of occurrence in previous translations. To our knowledge, these projects have not led to any follow-up, nor to consumer applications.
The vast majority of projects using an intermediate representation of SL, including the latest ones \cite{gomez2021}, use sequences of glosses, each gloss\footnote{A gloss is a text label, generally a single word, reflecting the meaning of the sign it stands for.} standing for a so-called lexical unit generally restricted to manual activity.
The translation systems then deal with a sequence of tokens and as such, meet the requirements for the approaches designed for sequences.
With this kind of representation though, it is very difficult if not impossible to the handle common SL phenomena like non-manual activity, spatial relations, depicting structures, or the rhythm of the signing production. This results in low quality animations, incomplete if not incomprehensible, and therefore unacceptable by the Deaf community. For this reason, it seems important to consider a richer intermediate representation than mere concatenations of glosses.
Note that in some recent neural-based approaches \cite{stoll2020text2sign}, the use of an intermediate representation is not present. This neural-based approach generates directly photo-realistic continuous sign videos from text inputs. This is still very experimental, and does not offer anonymisation, unlike virtual signer output. Moreover, an avatar's appearance can be suited to different use cases or audiences. For these reasons, avatar-based approaches seem more flexible and appealing.
This work therefore chose to explore EBMT for translation to SL, given that we do not have a large bilingual corpus and that we consider an intermediate representation for SL more appropriate.
\section{Method}
In view of an EBMT approach as explained above, we have created \emph{Rosetta-LSF} \cite{LREC-2022-corpusRosetta} and selected the intermediate representation \emph{AZee} to represent the SL utterances, which we explain in this section.
\subsection{EBMT-type approach}
\label{substitutions}
As explained above, EBMT is a translation mechanism based on analogy from examples.
This means that we can compensate for a missing example by finding one close enough, and working from it to replace what is different.
For example, to translate ``\textit{la présidente parle nerveusement}'' (the president is speaking nervously) when the example is not in the data base of examples, we can hope to work from the translation of ``\textit{le ministre de l'écologie parle nerveusement}'' (the minister of the environment is speaking nervously), with a substitution.
In such candidate segment henceforth, we will call ``anti-matches'' the parts that do not match the query in the segment that otherwise does, and ``corrections'' the respective text parts that would have been a match.
For example, ``\textit{présidente}'' (president) is the anti-match above, and ``\textit{ministre de l'écologie}'' (minister of the environment) its correction.
A hypothesis is that if we find the portions corresponding to each anti-match in the aligned translation, we can attempt to replace them with translations of their corrections.
Our aim is therefore to produce a translation of the source written text into the chosen intermediate representation that reflects the target signed language, AZee.
\subsection{AZee} \label{AZee}
AZee is a formal approach to SL discourse representation \cite{filholNonManualFeaturesRight2014}. It allows to define \emph{production rules} that associate forms to articulate (e.g.\ begin eyebrow raise before $X$) to identified meaning (semantic operations, e.g.\ expression of $Y$ with doubt). By combining them, one can build hierarchically structured \emph{discourse expressions} representing full discourse utterances, determining the forms to produce while exposing the meaning.
For example, consider the four productions rules below:
\begin{itemize}
\item \texttt{info-about}(\emph{topic}, \emph{info}): \emph{info}, which is focused, is given about a \emph{topic};
\item \texttt{nerveusement}(\emph{sig}): \emph{sig} in a nervous, stressed out way;
\item \texttt{président}: president;
\item \texttt{parler}: speak.
\end{itemize}
These can be combined in the expression below, which not only creates a semantic combination interpretable as ``the president is speaking nervously'', but also produces, through recursive application of each rule's forms, the resulting overall signed form with that meaning. A corpus of 120 such expressions has been published by Challant \& Filhol \cite{LREC-2022-corpusAZee}.
\begin{minipage}[t]{1\columnwidth}
\begin{verbatim}
:info-about
'topic
:président
'info
:nerveusement
'sig
:parler
\end{verbatim}
\end{minipage}
\medskip
Since it represents the articulations necessary to convey that meaning, it can be used as the output of our translation system, a lot easier than attempting to generate video frames directly.
Of course this requires to append an animation system to the pipeline, able to render AZee input to SL video. But this is outside the scope of this paper, and assumed for now.
AZee discourse expressions are hierarchical, each nested expression covering a sub-part of the discourse.
So unlike a linear stream like text or video where segments are typically specified with start and length, identifying an AZee ``segment'' can be done through identification of a single node in the expression. This node is the root of a sub-tree (or a leaf) which covers a time segment in the video (the SL capture modelled with the expression, or indeed any avatar animation rendered from the expression).
\subsection{Corpus}
As explained above, we needed a bank of alignments between French text segments and AZee expression nodes.
For this purpose, we worked on a parallel French-LSF corpus, in our case the first task of the Rosetta-LSF corpus \citelanguageresource{rosetta-lsf}.
It consists in 194 French news items, each of 3 to 35 words in length, together with their LSF translations. For instance: ``\textit{L'Everest menacé de réchauffement climatique}'' (Everest threatened with global warming).
The translations were done by a deaf person selected for her experience in producing online LSF content on a regular basis.
All of these news items have been encoded with AZee (section \ref{AZee}), resulting in a corpus of 194 AZee discourse expressions.
These expressions reflect both the forms observed in the news in LSF and the meaning interpreted from it.
From there, we were able to create a bank of AZee--text alignments which are correspondences between an AZee node and a text segment in French.
In an AZee discourse expression, the root node necessarily covers the whole discourse in French, which already serves as an alignment. Besides, each node of the expression represents a portion of the news, which if a text segment is found to be equivalent in meaning, creates a new alignment of finer granularity. Considering each node of the AZee expressions allows to create entries on various levels (from whole news items to single words).
The alignments were created manually, guided by the three following principles.
\begin{description}
\item[Uniqueness] The same text segment cannot be aligned with different AZee sub-expressions.
\item[Maximisation] The text segment has to correspond to the largest AZee sub-expression possible.
\item[Objectivity] Only a text segment present in the news can be aligned: we want to be faithful to the example, not to account for generality by attempting to escape its scope.
\end{description}
After manually proceeding this way on all of the corpus, we found 1812 AZee--text alignments, collected in a file and encoded with the name of the text file in which the news in French is found, the first and last characters of the aligned French segment, the file name of the AZee discourse expression and finally the line number of the aligned AZee expression or sub-expression (node). For example: \[\textrm{RO1}\_\textrm{X0007.Titre1}\ 10\ 4\ \textrm{RO1}\_\textrm{X0007.Titre1.az}\ 7\]
\section{Implementation}
\subsection{General algorithm} \label{general-algorithm}
Let $tr$ be the function that associates to a text query $q$
a set of possible translations for $q$ by analogy based on a corpus of aligned examples.
If the corpus contains alignments in which the text segment is exactly $q$, the set formed
by their aligned AZee expressions specifies an acceptable result for $tr(q)$.
Otherwise, as explained in §\ref{substitutions}, we consider the alignments whose text segments are ``close'' to $q$,
whose differences to $q$ are the ``anti-matches'',
whose translated counterparts in the aligned expression we hope to replace.
By doing this, the global structure of the aligned expression is kept to serve as a template for $tr(q)$,
in which substitutions are made.
Formally, for a given alignment between a text segment $txt$ and an expression $az$
where $txt$ qualifies as close to $q$, let:
\begin{itemize}
\item $\bar{m}_{1},...,\bar{m}_{N}$ be the anti-matches of $txt$, i.e.\ the parts in $txt$ that differ to $q$, where usually $N\le2$;
\item $c_{i}$ be the correction of $\bar{m}_{i}$ ($i\in1..N$), i.e.\ the wanted part of $q$ missing in $txt$;
\item $\bar{az}_{i}$ be the node in $az$ at the root of the sub-expression which translates $\bar{m}_{i}$ ($i\in1..N$).
\end{itemize}
With these notations, our approach is then to find a node $\bar{az}_{i}$ for each $i\in1..N$, and replace it with a translation of $c_{i}$.
For example, to translate query ``\textit{la présidente parle nerveusement}'', we could consider the text segment of the alignment below as close. In the AZee expression, rule \texttt{side-info}, with arguments \emph{focus} and \emph{info}, carries the meaning of \emph{focus} with additional (non-focused) information \emph{info} about it.
\begin{quote}
\begin{description}
\item[Text] ``\textit{le ministre de l'écologie parle nerveusement}''
\item[AZee]
\begin{minipage}[t]{1\columnwidth}
\begin{verbatim}
:info-about
'topic
:side-info (*)
'focus
:ministre
'info
:environnement
'info
:nerveusement
'sig
:parler
\end{verbatim}
\end{minipage}
\end{description}
\end{quote}
The unique anti-match $\bar{m}_{1}$ is ``\textit{le ministre de l'écologie}'', and its correction $\bar{c}_1$ is ``\textit{la présidente}''. We would then want to identify the sub-expression marked \texttt{(*)} as its translation node $\bar{az}_{1}$, for which to substitute a translation of ``\textit{la présidente}''.
If no or several candidates for an $\bar{az}_{i}$ are found, it becomes a lot
less trivial to know what to substitute in $az$ regarding the $i$th anti-match.
For now, we implement translation failure in these cases, forcing each $\bar{az}_{i}$
to be found unique. The translations of $c_i$ can however be numerous, each one becoming an option for the $\bar{az}_{i}$ substitution.
Using our formal notations:
\begin{itemize}
\item finding $\bar{az}_{i}$ means finding a unique node $n$ such that $n$ is an acceptable translation for $\bar{m}_{i}$, in other words such that $n\in tr(\bar{m}_{i})$ ;
\item finding translations for $c_{i}$ implies simply to consider $tr(c_{i})$.
\end{itemize}
Then, any combination of $N$ substitutions $\bar{az}_{i}\rightarrow x,x\in tr(c_{i})$ can be applied to $az$ to create a translation of $q$. The set of all of them is therefore a subset of $tr(q)$ associated with the $txt$--$az$ alignment.
The full set can be specified as the union of such sets, iterating over all
known alignments with a text segment close to $q$.
The approach above yields a recursive definition of $tr(q)$ as it requires
values for $tr(\bar{m}_{i})$ and $tr(c_{i})$ for each anti-match encountered.
The base case for this recursion are the exact matches.
Besides, each anti-match is always a shorter segment than the initial query,
so the only condition for termination of this algorithm is to ensure that
corrections $c_{i}$ are always also shorter than the query, which is clearly
the typical case so adding this constraint will likely result in zero loss.
More than termination, the problem is that of translation failure, which
is all the more likely to happen as the corpus of example alignments is small.
In such cases, we resort to a last fallback where we break the query down into
a partition of smaller text chunks, which we will translate separately and
concatenate in the result with the only reason that it follows the French order.
To do this we apply the AZee production rule \texttt{sign-supported-spoken} which
allows to build utterances based a spoken language literal sequence of items.
For example, one can chunk the query above into ``\textit{la présidente parle}'' + ``\textit{nerveusement}'', find a translation for each chunk separately, say (a) and (b) below, and propose (c) as a final translation.
(a)~\begin{minipage}[t]{.2\textwidth}
\begin{verbatim}
:info-about
'topic
:président
'info
:parler
\end{verbatim}
\end{minipage}
~
(b)~\begin{minipage}[t]{.2\textwidth}
\begin{verbatim}
:nerveux
\end{verbatim}
\end{minipage}
~
(c)~\begin{minipage}[t]{.3\textwidth}
\begin{verbatim}
:sign-supported-spoken
'units
list
:info-about
'topic
:président
'info
:parler
:nerveux
\end{verbatim}
\end{minipage}
\medskip
For a given partition $\left\langle p_{1},p_{2},...,p_{n}\right\rangle $ of $q$,
the combinations $sign-supported-spoken(units=\left\langle x_{1},x_{2},...,x_{n}\right\rangle )$
with $x_{i}\in tr(p_{i})$ constitute a set of possible translations of $q$ with this technique.
By iterating on different partitions and joining all such sets, we generate a last, fallback specification of $tr(q)$.
This is also a recursive definition, whose recursive calls are applied to chunks ($p_{i}$) shorter
then $q$ by construction, so termination is guaranteed as well.
This fallback strategy produces poorer quality Sign Language, and indeed equivalent to literal (word-to-word)
translation if used systematically. But it does allow to juxtapose coarser-grain chunks of Sign when translation
succeeds without resorting to partitioning.
For example, the use of the rule \texttt{nerveux}, generating an additional manual sign meaning ``nervous'', can be judged as poorer LSF than that of \texttt{nerveusement} used further up, which generates a preferred and sufficient facial expression conveying the same meaning.
However, the first chunk was translated as a whole (using \texttt{info-about}), which did avoid the even poorer literal sign sequence below.
\begin{minipage}[t]{.7\columnwidth}
\begin{verbatim}
:sign-supported-spoken
'units
list
:président
:parler
:nerveux
\end{verbatim}
\end{minipage}
\medskip
\subsection{Auxiliary text processing modules}
The practical implementation of the algorithm relies on several text processing modules enhancing analysis to find best correspondences in the existing corpus.
To allow for matching, antimatching and partitions, word-level tokenization is first performed by OpenNMT Tokenizer\footnote{\url{https://github.com/OpenNMT/Tokenizer}}, and flexibility is allowed when finding matching segments for punctuation and articles.
Then the core challenge is to define what kind of ``similarity'' in the source language can produce best candidates for target language generation. As can be seen in the example above, both semantics and syntax come into play to determine similar elements to be replaced or translated separately. In practice, we rely on two types of text analysis at different steps of the algorithm.
\begin{table*}
\centering
\begin{tabular}{llll}
\hline Alignment text & \pbox{1cm}{Common tokens} & Length & Ratio \\
\hline \pbox{20cm}{le superéthanol n'est proposé que dans 1 000 stations-service en \\ france , comme ici \textbf{dans la} \textbf{banlieue de} bordeaux .} & 4 & 22 & 0.18 \\
\hline \pbox{20cm}{comme ici \textbf{dans la banlieue de} bordeaux} & 4 & 7 & 0.57 \\
\hline \textbf{la banlieue de} bordeaux & 3 & 4 & 0.75 \\
\hline \pbox{20cm}{situé \textbf{dans la} province du guizhou , en chine , le mont fanjing attire \textbf{de} \\ nombreux touristes venus découvrir la richesse de ce paysage montagneux .} & 3 & 26 & 0.12 \\
\hline \pbox{20cm}{\textbf{la} villa noailles à hyères \textbf{dans} le var est un château cubiste construit dans \\ les années folles , à la demande d'un couple \textbf{de} mécènes avant-gardiste .} & 3 & 29 & 0.10 \\
\hline
\end{tabular}
\caption{Antimatchable alignments to translate ``dans la banlieue de Gerstheim" (\textit{in the suburbs of Gerstheim})}
\label{tab:antimatches}
\end{table*}
To find the best anti-matches in the current database and replace them by corrections, we use string matching and consider as ``anti-matchable'' all alignments that have tokens in common with the candidate text.
Best matches have been empirically set as the ones with the maximum tokens in common, and either the minimum length in number of tokens or the best ratio of similar tokens over total tokens.
For example, to translate ``\textit{dans la banlieue de Gerstheim}'' (in the suburbs of Gerstheim) by anti-match, the alignments with most tokens in common found in the database are described in Table \ref{tab:antimatches}.
We see that selecting, among the alignments with the highest number of tokens in common (4), the alignment with the lowest length or best ratio between number of similar tokens and length enables to retrieve the most relevant alignment for anti-match: ``\textit{comme ici \textbf{dans la banlieue de} Bordeaux}'' (like here in the suburbs of Bordeaux).
Other sets of metrics could be used successfully, as we found that the selection and ranking of alignments for anti-matching strategy significantly affects the results of the algorithm.
When matching and anti-matching approaches fail, we resort to partitions determined by navigating the syntactic dependency tree obtained using spaCy\footnote{\url{https://spacy.io}}, open-source Python library with off-the-shelf pretrained models and optimized pipelines for Natural Language Processing.
For instance, for the sentence ``\textit{Le couvre-feu cette semaine n'est pas encore arrêté}'' (curfew this week has not yet been stopped), we consider as candidate partitions:
\begin{itemize}
\item ``\textit{le couvre-feu}'' / ``\textit{cette semaine n'est pas encore arrêté}'';
\item ``\textit{le couvre-feu cette semaine n'est}'' / ``\textit{pas encore}'' / ``\textit{arrêté}'';
\item ``\textit{le couvre-feu}'' / ``\textit{cette semaine}'' / ``\textit{n'est pas encore arrêté}''.
\end{itemize}
Our tree exploration makes that some possible partitions such as the following are not explored: ``\textit{Le couvre-feu}'' / ``\textit{cette semaine}'' / ``\textit{n'est pas}'' / ``\textit{encore}'' / ``\textit{arrêté}''.
Dependency parsing does not explore all possible partitions of a sentence but at least constrains the exploration to syntactically valid chunks.
\section{Test}
\subsection{Test set}
Admittedly, the data base is still too small for us to claim a system up to any usable scale. So to test our system, we decided to build a test set by creating sentences mixing segments from different entries of our corpus, and evaluate the produced outputs.
Our test set is composed of 15 sentences and allows to test the algorithm, as presented in the next section.
For instance, the sentence ``\textit{Recul de l'âge légal à la retraite : c'est ce que proposent des retraités pour leurs enfants}'' (Increase of the retirement age: pensioners propose it for their children) was added to the test set and created from the following sentences of the corpus:
\begin{itemize}
\item \textit{\textbf{Recul de l'âge légal à la retraite :} ``Il ne faut pas prendre les Français pour des canards sauvages'', lance Valérie Pécresse.} (Increase of the retirement age: ``We should not take the French for a ride'', shouts Valérie Pécresse.)
\item \textit{Des routes nationales bientôt privatisées ? \textbf{C'est ce que proposent} les sociétés d'autoroutes dans une note interne.} (National roads soon to be privatised? Motorway companies propose it in an internal memo.)
\item \textit{Solidarité : une ancienne abbaye accueille \textbf{des retraités}} (Solidarity: a former abbey hosts pensioners.)
\item \textit{Au Japon, des dizaines de pères français se battent désespérément \textbf{pour} voir \textbf{leurs enfants.}} (In Japan, dozens of French fathers are desperately fighting to see their children.)
\end{itemize}
\subsection{Algorithm on an example}
This section describes the steps taken by the algorithm run on the following example taken from the test set, and the produced AZee description results.
\begin{quote}
``\textit{Alsace : de grands chefs ont vendu leur vaisselle pour les plus modestes comme ici dans la banlieue de Gerstheim.}'' (Alsace: great chefs sold their crockery for the poor like here in the suburbs of Gerstheim.)
\end{quote}
The whole sentence is tested, first for exact matches, then for anti-matching segments but to no avail. So it falls back to partitioning the query, breaking it down into 3 smaller segments as follows: ``\textit{Alsace}'' / ``\textit{de grands chefs}'' / ``\textit{ont vendu leur vaisselle pour les plus modestes comme ici dans la banlieue de Gerstheim}''.
Each segment above is then used as a new (simpler) input query in a recursive call to the algorithm, reported below. See fig.~\ref{fig:partial-matches} for the referenced AZee expression matches.
\begin{description}
\item[``\textit{Alsace}''] An exact-match (d) is found, which is directly returned as an acceptable translation for this segment.
\item[``\textit{de grands chefs}''] Similarly, an exact-match (e) is found.
\item[``\textit{ont vendu leur vaisselle pour} ...''] There is no exact match, and no anti-matching segment is found either to translate this text chunk. So again, the query is broken down into the smaller sub-queries below.
\begin{description}
\item[``\textit{ont vendu leur vaisselle}''] Exact match (f1) found.
\item[``\textit{pour les plus modestes}''] Exact match (f2) found.
\item[``\textit{comme ici dans la banlieue de Gerstheim}''] No exact match, but similar segment found, aligned with (f3'): ``\textit{comme ici dans la banlieue de Bordeaux}'', anti-match ``\textit{Bordeaux}'' to be corrected with ``\textit{Gerstheim}''.
Both the anti-match and the correction trigger a recursive call, the former to decide on a node to change in (f3'); the latter to find what to change it for.
\begin{description}
\item[``\textit{Bordeaux}''] Exact match \texttt{:Bordeaux} found in the alignment base, recognised and unique in (f3')---marked \texttt{(**)} in fig.~\ref{fig:partial-matches}.
\item[``\textit{Gerstheim}''] Exact match (f3'') found in the alignment base.
\end{description}
Let (f3) be the expression (f3') with (f3'') instead of node \texttt{(**)}; (f3) is a resulting translation for this query.
\end{description}
Now that each segment of the inner partition has found a translation, a result can be produced by creating a \texttt{sign-supported-spoken} expression with units (f1), (f2) and (f3) in this order.
\end{description}
Finally and in the same way, a result can be proposed for the outer partition using \texttt{sign-supported-spoken}. The overall expression is therefore the following:
\begin{minipage}[t]{.7\columnwidth}
\begin{verbatim}
:sign-supported-spoken
'units
list
(d)
(e)
:sign-supported-spoken
'units
list
(f1)
(f2)
(f3)
\end{verbatim}
\end{minipage}
\medskip
\begin{figure*}
\begin{center}
(d)~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:category
'cat
:info-about
'topic
:Est
'info
:info-about
'topic
:France
'info
:zone
'elt
:info-about
'topic
:appartenance
'info
:Alsace
\end{verbatim}
\end{minipage}
~
(e)~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:category
'cat
:side-info
'focus
:multiplicity
'elt
:une personne
'info
:zone
'elt
:chef cuisinier
\end{verbatim}
\end{minipage}
~
(f1)~\begin{minipage}[t]{.3\textwidth}
\begin{verbatim}
:info-about
'topic
:là
'info
:info-about
'topic
:all-of
'items
list
:assiette
:assiette
'info
:multiplicity
'elt
:vendre
\end{verbatim}
\end{minipage}
\end{center}
\vspace{1cm}
\begin{center}
(f2)~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:info-about
'topic
:pour
'info
:side-info
'focus
:multiplicity
'elt
:une personne
'info
:info-about
'topic
:comment dire
'info
:difficile
\end{verbatim}
\end{minipage}
~
(f3')~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:info-about
'topic
:exemple
'info
:info-about
'topic
:aussi
'info
:info-about
'topic
:ici
'info
:info-about
'topic
:side-info
'focus
:Bordeaux (**)
'info
:banlieue
'info
:là
\end{verbatim}
\end{minipage}
~
(f3'')~\begin{minipage}[t]{.3\textwidth}
\begin{verbatim}
:category
'cat
:ville
'elt
:fingerspelling
'letters
list
.G
.E
.R
.S
.T
.H
.E
.I
.M
\end{verbatim}
\end{minipage}
\end{center}
\begin{quote}
In English:
\begin{description}
\item[(d)] The Alsace region in the East of France
\item[(e)] Chefs
\item[(f1)] Sell crockery
\item[(f2)] For the poor (people)
\item[(f3')] Like here in the suburbs of Bordeaux
\item[(f3'')] Gerstheim
\end{description}
\end{quote}
\caption{Aligned AZee expressions matched in algorithm run}
\label{fig:partial-matches}
\end{figure*}
\section{Discussion}
First, our anti-matches approach has some advantages, compared to a sequence-based one.
Indeed, structures that are specific to LSF can be found in the final translations, which is not the case when the language is reduced to a sequence of glosses.
In addition, the approach produces results with a certain form of creativity. In LSF, paraphrases or additions are commonly used, and indeed part of our corpus as initially delivered by the translator at the time of video corpus creation. These elements were later aligned as examples, thus frequently appear in the generated translations, although not always strictly necessary. See the example for ``\textit{Alsace}'', which although a single sign exists, is translated to the whole expression (d), typical of LSF when no context yet exists.
Moreover, the output of the algorithm is a set of translations
(built from the various substitution combinations), not necessarily a single expression. This in a way accounts for the reality of the translation task.
For example, to translate ``\textit{Emmanuel Macron}'' into LSF, different possibilities have been used by the translator, hence the different possible AZee output expressions (g), (h) and (i) below.
(g)~\begin{minipage}[t]{.5\columnwidth}
\begin{verbatim}
:Emmanuel Macron
\end{verbatim}
\end{minipage}
(h)~\begin{minipage}[t]{.5\columnwidth}
\begin{verbatim}
:side-info
'topic
:Emmanuel Macron
'info
:président
\end{verbatim}
\end{minipage}
(i)~\begin{minipage}[t]{.5\columnwidth}
\begin{verbatim}
:category
'cat
:side-info
'topic
:une personne
'info
:président
'elt
:Emmanuel Macron
\end{verbatim}
\end{minipage}
\medskip
In our test set, the number of translations proposed for a query ranges from 1 to 12 (average: 4).
At the moment, the order in which the AZee expressions are output is irrelevant. One prospect for this algorithm is to rank them according to some heuristics, for example constraints on preferred AZee rule combinations.
The work presented also has limitations.
We can observe that some anti-matches are incorrect, for instance: ``\textit{des personnes pro-Brexit}'' (pro-Brexit people) vs ``\textit{des personnes manifestent}'' (people demonstrate). The syntactic categories of the anti-match and its correction are not the same (adjective vs. verb), which creates problems during the translation process. If we want to translate ``\textit{des personnes pro-Brexit sont dans la rue}'' (pro-Brexit people are in the street), the algorithm suggests ``\textit{pro-Brexit}'' as an anti-match for ``\textit{manifestent}'', but the result is syntactically unacceptable: ``\textit{*des personnes manifestent sont dans la rue}''.
The syntactic category of each phrase should be taken into account to prevent such errors and to improve the anti-matching results.
Finally, a considerable number of fallbacks are present in the output of the algorithm: 3 per result on average.
As explained in section~\ref{general-algorithm}, this is not ideal, and the size of our corpus is undoubtedly a contributing factor: if we increase the number of examples and alignments, the number of fallbacks will decrease and the quality of the translations should hopefully improve.
\section{Conclusion and prospects}
We have presented a new system of automatic translation from text to AZee, based on an example-based machine translation approach, the hierarchical representation of SL AZee and an aligned corpus of French text and AZee descriptions extracted from the Rosetta-LSF corpus. A prototyping implementation of the system has been made and tested on some examples, thus providing a proof of concept.
The capacities of this system and the size of the corpus still need to be extended before real evaluations can be carried out. But we can already stress that the evaluation of such a system will not be easy, since it proposes a translation from one language to a representation of another language, not readable directly.
Metrics for evaluating the quality of translations, such as the one proposed in the European QT21 project\footnote{\url{https://www.qt21.eu}}, provide a scoring sheet with types of errors produced by the translation system, which allows to highlight the shortcomings of the systems and the aspects to improve.
This project has proposed Multidimensional Quality Metrics (MQM), a framework for describing and defining custom translation quality metrics. It provides a flexible vocabulary of quality issue types, a mechanism to generate quality scores, and mappings to other metrics.
Some of the error categories are defined assuming text as a target, which does not apply in our case. A category called ``fluency'' allows us to evaluate the quality of an utterance, regardless of whether it is the result of a translation. In our case, the target is not even a language utterance, thus this category will need some adjustments.
What remains is the category of errors linked to the translation process itself, categorised as ``accuracy''.
Accuracy errors happen when the target text does not reflect the source text. This could be by addition (of content not present in the source), mistranslation (output content not accurately aligned with the source) or omission (of content that was present in the source).
It would be interesting to study if this kind of evaluation could be adapted to our system.
The issue is to define these categories in the case of SL. It is common or indeed often preferred in SL to introduce contextual information, for example expression (d) figure~\ref{fig:partial-matches} for ``Alsace'', which should not be judged as an unwanted addition.
Furthermore, as AZee can be used to generate virtual signer animations which are directly ``readable'' by language users, fluency error categories could be taken into consideration after this step to complete the evaluation. The establishment of a robust and comprehensive evaluation protocol is clearly a subject of study in its own right, which needs to be pursued in the near future.
\section*{Acknowledgements}
This work has been funded by the Bpifrance investment project ``Grands défis du numérique'', as part of the ROSETTA project (RObot for Subtitling and intElligent adapTed TranslAtion).
We thank Noémie Churlet, Raphaël Bouton and Media'Pi! for their commitment to this project, which would not have had the same validity and impact without them.
\bibliographystyle{plain}
\section{Introduction}
Rosetta\footnote{\url{https://rosettaccess.fr/index.php/home-page-english/}} is a French project that aimed to study accessibility solutions for audiovisual content. One of the experiments consisted in designing an automatic translation system from text to Sign Language (SL) displayed through animation of a virtual signer.
The three main contributions concerning SL in this project were the constitution of Rosetta-LSF \cite{LREC-2022-corpusRosetta}, an aligned corpus of text and SL captured using a mocap system, a transflation system from text to AZee (a representation of SL content), and a system allowing to generate virtual signer animations from AZee input.
This article describes the second contribution: the system of translation from text to AZee.
After an overview of the issues and recent works in the field, we explain our method and design choices and describe the implementation of the translation system. Finally, we give preliminary results and discuss the questions raised for evaluation.
\section{Text-to-Sign translation}
The automatic translation of content from a spoken language into a sign language is a fairly recent and still largely unexplored research topic. Here we are interested in the translation of text as the source format, in our case in French, and video or 3D animation as the target format, in our case French Sign Language (LSF).
In this section, we look at the main challenges encountered with text-to-sign translation.
\subsection{Need for bilingual corpora}
Machine translation (MT) was first developed for spoken languages in their written form using bilingual dictionaries and rule-based systems, that were not easy to develop and maintain. Access to parallel corpora of aligned examples has led to the rise of data-driven approaches, such as Statistical Machine Translation (SMT) that used the frequencies of translation pairs containing source--target pairings of words or phrases. In the current dominant approach, Neural Machine Translation (NMT), which is also data-driven, the source text is encoded into an intermediate representation in the form of numerical vectors to be decoded as a target text. Although the representation is not directly open to interpretation, the practical results largely prevail over former strategies. These methods designed for spoken languages rely on the availability of large volumes of parallel data (of the order of several million sentences), Sign languages being considered as low resource in this respect.
Example-based MT (EBMT) is another data-driven approach based on analogy \cite{nagao:ahi84}. It uses a bilingual corpus that contains texts and their translations. Given a sentence to translate, sentences from this corpus are selected that contain similar sub-sentential components. The similar sentences are then used to translate the sub-sentential components of the original sentence into the target language, and these phrases are put together to form a complete translation. Although the larger the corpus, the better the results will be, this approach can be implemented on smaller corpora and thus may be considered in the case of Sign Language Machine Translation (SLMT).
\subsection{Need for an intermediate representation}
One of the major differences between SLMT and written MT is the difference in channel. Written languages are input to MT systems as streams of discrete tokens---words separated by blanks---whereas SL does not have a written form. SLs are to be considered as face-to-face oral languages. Moreover, they are able to convey simultaneous information by the way of a number of articulators, such as the two hands and arms, but also the torso, shoulders, head, gaze and facial expressions (including a number of facial component movements).
As SL has no written form, many approaches proceed in two steps: a first step transforms the SL content into an intermediate representation, and a second step uses this representation as the input of a synthesis system to control the animation of an avatar in order to display the content in SL.
After a first generation of studies based mainly on the rule-based approach \cite{veale1998challenges,zhao2000machine,marshall-04}), a few ones have investigated Example-Based Machine Translation \cite{morrissey2005example}. They have sometimes be combined with statistical approaches, such as \cite{demartino2017signing} who have developed a system that automatically translates Brazilian Portuguese text to Brazilian SL (LIBRAS) by combining SMT with EBMT in case of unseen texts or ambiguous terms dependent on the context and frequency of occurrence in previous translations. To our knowledge, these projects have not led to any follow-up, nor to consumer applications.
The vast majority of projects using an intermediate representation of SL, including the latest ones \cite{gomez2021}, use sequences of glosses, each gloss\footnote{A gloss is a text label, generally a single word, reflecting the meaning of the sign it stands for.} standing for a so-called lexical unit generally restricted to manual activity.
The translation systems then deal with a sequence of tokens and as such, meet the requirements for the approaches designed for sequences.
With this kind of representation though, it is very difficult if not impossible to the handle common SL phenomena like non-manual activity, spatial relations, depicting structures, or the rhythm of the signing production. This results in low quality animations, incomplete if not incomprehensible, and therefore unacceptable by the Deaf community. For this reason, it seems important to consider a richer intermediate representation than mere concatenations of glosses.
Note that in some recent neural-based approaches \cite{stoll2020text2sign}, the use of an intermediate representation is not present. This neural-based approach generates directly photo-realistic continuous sign videos from text inputs. This is still very experimental, and does not offer anonymisation, unlike virtual signer output. Moreover, an avatar's appearance can be suited to different use cases or audiences. For these reasons, avatar-based approaches seem more flexible and appealing.
This work therefore chose to explore EBMT for translation to SL, given that we do not have a large bilingual corpus and that we consider an intermediate representation for SL more appropriate.
\section{Method}
In view of an EBMT approach as explained above, we have created \emph{Rosetta-LSF} \cite{LREC-2022-corpusRosetta} and selected the intermediate representation \emph{AZee} to represent the SL utterances, which we explain in this section.
\subsection{EBMT-type approach}
\label{substitutions}
As explained above, EBMT is a translation mechanism based on analogy from examples.
This means that we can compensate for a missing example by finding one close enough, and working from it to replace what is different.
For example, to translate ``\textit{la présidente parle nerveusement}'' (the president is speaking nervously) when the example is not in the data base of examples, we can hope to work from the translation of ``\textit{le ministre de l'écologie parle nerveusement}'' (the minister of the environment is speaking nervously), with a substitution.
In such candidate segment henceforth, we will call ``anti-matches'' the parts that do not match the query in the segment that otherwise does, and ``corrections'' the respective text parts that would have been a match.
For example, ``\textit{présidente}'' (president) is the anti-match above, and ``\textit{ministre de l'écologie}'' (minister of the environment) its correction.
A hypothesis is that if we find the portions corresponding to each anti-match in the aligned translation, we can attempt to replace them with translations of their corrections.
Our aim is therefore to produce a translation of the source written text into the chosen intermediate representation that reflects the target signed language, AZee.
\subsection{AZee} \label{AZee}
AZee is a formal approach to SL discourse representation \cite{filholNonManualFeaturesRight2014}. It allows to define \emph{production rules} that associate forms to articulate (e.g.\ begin eyebrow raise before $X$) to identified meaning (semantic operations, e.g.\ expression of $Y$ with doubt). By combining them, one can build hierarchically structured \emph{discourse expressions} representing full discourse utterances, determining the forms to produce while exposing the meaning.
For example, consider the four productions rules below:
\begin{itemize}
\item \texttt{info-about}(\emph{topic}, \emph{info}): \emph{info}, which is focused, is given about a \emph{topic};
\item \texttt{nerveusement}(\emph{sig}): \emph{sig} in a nervous, stressed out way;
\item \texttt{président}: president;
\item \texttt{parler}: speak.
\end{itemize}
These can be combined in the expression below, which not only creates a semantic combination interpretable as ``the president is speaking nervously'', but also produces, through recursive application of each rule's forms, the resulting overall signed form with that meaning. A corpus of 120 such expressions has been published by Challant \& Filhol \cite{LREC-2022-corpusAZee}.
\begin{minipage}[t]{1\columnwidth}
\begin{verbatim}
:info-about
'topic
:président
'info
:nerveusement
'sig
:parler
\end{verbatim}
\end{minipage}
\medskip
Since it represents the articulations necessary to convey that meaning, it can be used as the output of our translation system, a lot easier than attempting to generate video frames directly.
Of course this requires to append an animation system to the pipeline, able to render AZee input to SL video. But this is outside the scope of this paper, and assumed for now.
AZee discourse expressions are hierarchical, each nested expression covering a sub-part of the discourse.
So unlike a linear stream like text or video where segments are typically specified with start and length, identifying an AZee ``segment'' can be done through identification of a single node in the expression. This node is the root of a sub-tree (or a leaf) which covers a time segment in the video (the SL capture modelled with the expression, or indeed any avatar animation rendered from the expression).
\subsection{Corpus}
As explained above, we needed a bank of alignments between French text segments and AZee expression nodes.
For this purpose, we worked on a parallel French-LSF corpus, in our case the first task of the Rosetta-LSF corpus.
It consists in 194 French news items, each of 3 to 35 words in length, together with their LSF translations. For instance: ``\textit{L'Everest menacé de réchauffement climatique}'' (Everest threatened with global warming).
The translations were done by a deaf person selected for her experience in producing online LSF content on a regular basis.
All of these news items have been encoded with AZee (section \ref{AZee}), resulting in a corpus of 194 AZee discourse expressions.
These expressions reflect both the forms observed in the news in LSF and the meaning interpreted from it.
From there, we were able to create a bank of AZee--text alignments which are correspondences between an AZee node and a text segment in French.
In an AZee discourse expression, the root node necessarily covers the whole discourse in French, which already serves as an alignment. Besides, each node of the expression represents a portion of the news, which if a text segment is found to be equivalent in meaning, creates a new alignment of finer granularity. Considering each node of the AZee expressions allows to create entries on various levels (from whole news items to single words).
The alignments were created manually, guided by the three following principles.
\begin{description}
\item[Uniqueness] The same text segment cannot be aligned with different AZee sub-expressions.
\item[Maximisation] The text segment has to correspond to the largest AZee sub-expression possible.
\item[Objectivity] Only a text segment present in the news can be aligned: we want to be faithful to the example, not to account for generality by attempting to escape its scope.
\end{description}
After manually proceeding this way on all of the corpus, we found 1812 AZee--text alignments, collected in a file and encoded with the name of the text file in which the news in French is found, the first and last characters of the aligned French segment, the file name of the AZee discourse expression and finally the line number of the aligned AZee expression or sub-expression (node). For example: \[\textrm{RO1}\_\textrm{X0007.Titre1}\ 10\ 4\ \textrm{RO1}\_\textrm{X0007.Titre1.az}\ 7\]
\section{Implementation}
\subsection{General algorithm} \label{general-algorithm}
Let $tr$ be the function that associates to a text query $q$
a set of possible translations for $q$ by analogy based on a corpus of aligned examples.
If the corpus contains alignments in which the text segment is exactly $q$, the set formed
by their aligned AZee expressions specifies an acceptable result for $tr(q)$.
Otherwise, as explained in §\ref{substitutions}, we consider the alignments whose text segments are ``close'' to $q$,
whose differences to $q$ are the ``anti-matches'',
whose translated counterparts in the aligned expression we hope to replace.
By doing this, the global structure of the aligned expression is kept to serve as a template for $tr(q)$,
in which substitutions are made.
Formally, for a given alignment between a text segment $txt$ and an expression $az$
where $txt$ qualifies as close to $q$, let:
\begin{itemize}
\item $\bar{m}_{1},...,\bar{m}_{N}$ be the anti-matches of $txt$, i.e.\ the parts in $txt$ that differ to $q$, where usually $N\le2$;
\item $c_{i}$ be the correction of $\bar{m}_{i}$ ($i\in1..N$), i.e.\ the wanted part of $q$ missing in $txt$;
\item $\bar{az}_{i}$ be the node in $az$ at the root of the sub-expression which translates $\bar{m}_{i}$ ($i\in1..N$).
\end{itemize}
With these notations, our approach is then to find a node $\bar{az}_{i}$ for each $i\in1..N$, and replace it with a translation of $c_{i}$.
For example, to translate query ``\textit{la présidente parle nerveusement}'', we could consider the text segment of the alignment below as close. In the AZee expression, rule \texttt{side-info}, with arguments \emph{focus} and \emph{info}, carries the meaning of \emph{focus} with additional (non-focused) information \emph{info} about it.
\begin{quote}
\begin{description}
\item[Text] ``\textit{le ministre de l'écologie parle nerveusement}''
\item[AZee]
\begin{minipage}[t]{1\columnwidth}
\begin{verbatim}
:info-about
'topic
:side-info (*)
'focus
:ministre
'info
:environnement
'info
:nerveusement
'sig
:parler
\end{verbatim}
\end{minipage}
\end{description}
\end{quote}
The unique anti-match $\bar{m}_{1}$ is ``\textit{le ministre de l'écologie}'', and its correction $\bar{c}_1$ is ``\textit{la présidente}''. We would then want to identify the sub-expression marked \texttt{(*)} as its translation node $\bar{az}_{1}$, for which to substitute a translation of ``\textit{la présidente}''.
If no or several candidates for an $\bar{az}_{i}$ are found, it becomes a lot
less trivial to know what to substitute in $az$ regarding the $i$th anti-match.
For now, we implement translation failure in these cases, forcing each $\bar{az}_{i}$
to be found unique. The translations of $c_i$ can however be numerous, each one becoming an option for the $\bar{az}_{i}$ substitution.
Using our formal notations:
\begin{itemize}
\item finding $\bar{az}_{i}$ means finding a unique node $n$ such that $n$ is an acceptable translation for $\bar{m}_{i}$, in other words such that $n\in tr(\bar{m}_{i})$ ;
\item finding translations for $c_{i}$ implies simply to consider $tr(c_{i})$.
\end{itemize}
Then, any combination of $N$ substitutions $\bar{az}_{i}\rightarrow x,x\in tr(c_{i})$ can be applied to $az$ to create a translation of $q$. The set of all of them is therefore a subset of $tr(q)$ associated with the $txt$--$az$ alignment.
The full set can be specified as the union of such sets, iterating over all
known alignments with a text segment close to $q$.
The approach above yields a recursive definition of $tr(q)$ as it requires
values for $tr(\bar{m}_{i})$ and $tr(c_{i})$ for each anti-match encountered.
The base case for this recursion are the exact matches.
Besides, each anti-match is always a shorter segment than the initial query,
so the only condition for termination of this algorithm is to ensure that
corrections $c_{i}$ are always also shorter than the query, which is clearly
the typical case so adding this constraint will likely result in zero loss.
More than termination, the problem is that of translation failure, which
is all the more likely to happen as the corpus of example alignments is small.
In such cases, we resort to a last fallback where we break the query down into
a partition of smaller text chunks, which we will translate separately and
concatenate in the result with the only reason that it follows the French order.
To do this we apply the AZee production rule \texttt{sign-supported-spoken} which
allows to build utterances based a spoken language literal sequence of items.
For example, one can chunk the query above into ``\textit{la présidente parle}'' + ``\textit{nerveusement}'', find a translation for each chunk separately, say (a) and (b) below, and propose (c) as a final translation.
(a)~\begin{minipage}[t]{.2\textwidth}
\begin{verbatim}
:info-about
'topic
:président
'info
:parler
\end{verbatim}
\end{minipage}
~
(b)~\begin{minipage}[t]{.2\textwidth}
\begin{verbatim}
:nerveux
\end{verbatim}
\end{minipage}
~
(c)~\begin{minipage}[t]{.3\textwidth}
\begin{verbatim}
:sign-supported-spoken
'units
list
:info-about
'topic
:président
'info
:parler
:nerveux
\end{verbatim}
\end{minipage}
\medskip
For a given partition $\left\langle p_{1},p_{2},...,p_{n}\right\rangle $ of $q$,
the combinations $sign-supported-spoken(units=\left\langle x_{1},x_{2},...,x_{n}\right\rangle )$
with $x_{i}\in tr(p_{i})$ constitute a set of possible translations of $q$ with this technique.
By iterating on different partitions and joining all such sets, we generate a last, fallback specification of $tr(q)$.
This is also a recursive definition, whose recursive calls are applied to chunks ($p_{i}$) shorter
then $q$ by construction, so termination is guaranteed as well.
This fallback strategy produces poorer quality Sign Language, and indeed equivalent to literal (word-to-word)
translation if used systematically. But it does allow to juxtapose coarser-grain chunks of Sign when translation
succeeds without resorting to partitioning.
For example, the use of the rule \texttt{nerveux}, generating an additional manual sign meaning ``nervous'', can be judged as poorer LSF than that of \texttt{nerveusement} used further up, which generates a preferred and sufficient facial expression conveying the same meaning.
However, the first chunk was translated as a whole (using \texttt{info-about}), which did avoid the even poorer literal sign sequence below.
\begin{minipage}[t]{.7\columnwidth}
\begin{verbatim}
:sign-supported-spoken
'units
list
:président
:parler
:nerveux
\end{verbatim}
\end{minipage}
\medskip
\subsection{Auxiliary text processing modules}
The practical implementation of the algorithm relies on several text processing modules enhancing analysis to find best correspondences in the existing corpus.
To allow for matching, antimatching and partitions, word-level tokenization is first performed by OpenNMT Tokenizer\footnote{\url{https://github.com/OpenNMT/Tokenizer}}, and flexibility is allowed when finding matching segments for punctuation and articles.
Then the core challenge is to define what kind of ``similarity'' in the source language can produce best candidates for target language generation. As can be seen in the example above, both semantics and syntax come into play to determine similar elements to be replaced or translated separately. In practice, we rely on two types of text analysis at different steps of the algorithm.
\begin{table*}
\centering
\begin{tabular}{llll}
\hline Alignment text & \pbox{1cm}{Common tokens} & Length & Ratio \\
\hline \pbox{20cm}{le superéthanol n'est proposé que dans 1 000 stations-service en \\ france , comme ici \textbf{dans la} \textbf{banlieue de} bordeaux .} & 4 & 22 & 0.18 \\
\hline \pbox{20cm}{comme ici \textbf{dans la banlieue de} bordeaux} & 4 & 7 & 0.57 \\
\hline \textbf{la banlieue de} bordeaux & 3 & 4 & 0.75 \\
\hline \pbox{20cm}{situé \textbf{dans la} province du guizhou , en chine , le mont fanjing attire \textbf{de} \\ nombreux touristes venus découvrir la richesse de ce paysage montagneux .} & 3 & 26 & 0.12 \\
\hline \pbox{20cm}{\textbf{la} villa noailles à hyères \textbf{dans} le var est un château cubiste construit dans \\ les années folles , à la demande d'un couple \textbf{de} mécènes avant-gardiste .} & 3 & 29 & 0.10 \\
\hline
\end{tabular}
\caption{Antimatchable alignments to translate ``dans la banlieue de Gerstheim" (\textit{in the suburbs of Gerstheim})}
\label{tab:antimatches}
\end{table*}
To find the best anti-matches in the current database and replace them by corrections, we use string matching and consider as ``anti-matchable'' all alignments that have tokens in common with the candidate text.
Best matches have been empirically set as the ones with the maximum tokens in common, and either the minimum length in number of tokens or the best ratio of similar tokens over total tokens.
For example, to translate ``\textit{dans la banlieue de Gerstheim}'' (in the suburbs of Gerstheim) by anti-match, the alignments with most tokens in common found in the database are described in Table \ref{tab:antimatches}.
We see that selecting, among the alignments with the highest number of tokens in common (4), the alignment with the lowest length or best ratio between number of similar tokens and length enables to retrieve the most relevant alignment for anti-match: ``\textit{comme ici \textbf{dans la banlieue de} Bordeaux}'' (like here in the suburbs of Bordeaux).
Other sets of metrics could be used successfully, as we found that the selection and ranking of alignments for anti-matching strategy significantly affects the results of the algorithm.
When matching and anti-matching approaches fail, we resort to partitions determined by navigating the syntactic dependency tree obtained using spaCy\footnote{\url{https://spacy.io}}, open-source Python library with off-the-shelf pretrained models and optimized pipelines for Natural Language Processing.
For instance, for the sentence ``\textit{Le couvre-feu cette semaine n'est pas encore arrêté}'' (curfew this week has not yet been stopped), we consider as candidate partitions:
\begin{itemize}
\item ``\textit{le couvre-feu}'' / ``\textit{cette semaine n'est pas encore arrêté}'';
\item ``\textit{le couvre-feu cette semaine n'est}'' / ``\textit{pas encore}'' / ``\textit{arrêté}'';
\item ``\textit{le couvre-feu}'' / ``\textit{cette semaine}'' / ``\textit{n'est pas encore arrêté}''.
\end{itemize}
Our tree exploration makes that some possible partitions such as the following are not explored: ``\textit{Le couvre-feu}'' / ``\textit{cette semaine}'' / ``\textit{n'est pas}'' / ``\textit{encore}'' / ``\textit{arrêté}''.
Dependency parsing does not explore all possible partitions of a sentence but at least constrains the exploration to syntactically valid chunks.
\section{Test}
\subsection{Test set}
Admittedly, the data base is still too small for us to claim a system up to any usable scale. So to test our system, we decided to build a test set by creating sentences mixing segments from different entries of our corpus, and evaluate the produced outputs.
Our test set is composed of 15 sentences and allows to test the algorithm, as presented in the next section.
For instance, the sentence ``\textit{Recul de l'âge légal à la retraite : c'est ce que proposent des retraités pour leurs enfants}'' (Increase of the retirement age: pensioners propose it for their children) was added to the test set and created from the following sentences of the corpus:
\begin{itemize}
\item \textit{\textbf{Recul de l'âge légal à la retraite :} ``Il ne faut pas prendre les Français pour des canards sauvages'', lance Valérie Pécresse.} (Increase of the retirement age: ``We should not take the French for a ride'', shouts Valérie Pécresse.)
\item \textit{Des routes nationales bientôt privatisées ? \textbf{C'est ce que proposent} les sociétés d'autoroutes dans une note interne.} (National roads soon to be privatised? Motorway companies propose it in an internal memo.)
\item \textit{Solidarité : une ancienne abbaye accueille \textbf{des retraités}} (Solidarity: a former abbey hosts pensioners.)
\item \textit{Au Japon, des dizaines de pères français se battent désespérément \textbf{pour} voir \textbf{leurs enfants.}} (In Japan, dozens of French fathers are desperately fighting to see their children.)
\end{itemize}
\subsection{Algorithm on an example}
This section describes the steps taken by the algorithm run on the following example taken from the test set, and the produced AZee description results.
\begin{quote}
``\textit{Alsace : de grands chefs ont vendu leur vaisselle pour les plus modestes comme ici dans la banlieue de Gerstheim.}'' (Alsace: great chefs sold their crockery for the poor like here in the suburbs of Gerstheim.)
\end{quote}
The whole sentence is tested, first for exact matches, then for anti-matching segments but to no avail. So it falls back to partitioning the query, breaking it down into 3 smaller segments as follows: ``\textit{Alsace}'' / ``\textit{de grands chefs}'' / ``\textit{ont vendu leur vaisselle pour les plus modestes comme ici dans la banlieue de Gerstheim}''.
Each segment above is then used as a new (simpler) input query in a recursive call to the algorithm, reported below. See fig.~\ref{fig:partial-matches} for the referenced AZee expression matches.
\begin{description}
\item[``\textit{Alsace}''] An exact-match (d) is found, which is directly returned as an acceptable translation for this segment.
\item[``\textit{de grands chefs}''] Similarly, an exact-match (e) is found.
\item[``\textit{ont vendu leur vaisselle pour} ...''] There is no exact match, and no anti-matching segment is found either to translate this text chunk. So again, the query is broken down into the smaller sub-queries below.
\begin{description}
\item[``\textit{ont vendu leur vaisselle}''] Exact match (f1) found.
\item[``\textit{pour les plus modestes}''] Exact match (f2) found.
\item[``\textit{comme ici dans la banlieue de Gerstheim}''] No exact match, but similar segment found, aligned with (f3'): ``\textit{comme ici dans la banlieue de Bordeaux}'', anti-match ``\textit{Bordeaux}'' to be corrected with ``\textit{Gerstheim}''.
Both the anti-match and the correction trigger a recursive call, the former to decide on a node to change in (f3'); the latter to find what to change it for.
\begin{description}
\item[``\textit{Bordeaux}''] Exact match \texttt{:Bordeaux} found in the alignment base, recognised and unique in (f3')---marked \texttt{(**)} in fig.~\ref{fig:partial-matches}.
\item[``\textit{Gerstheim}''] Exact match (f3'') found in the alignment base.
\end{description}
Let (f3) be the expression (f3') with (f3'') instead of node \texttt{(**)}; (f3) is a resulting translation for this query.
\end{description}
Now that each segment of the inner partition has found a translation, a result can be produced by creating a \texttt{sign-supported-spoken} expression with units (f1), (f2) and (f3) in this order.
\end{description}
Finally and in the same way, a result can be proposed for the outer partition using \texttt{sign-supported-spoken}. The overall expression is therefore the following:
\begin{minipage}[t]{.7\columnwidth}
\begin{verbatim}
:sign-supported-spoken
'units
list
(d)
(e)
:sign-supported-spoken
'units
list
(f1)
(f2)
(f3)
\end{verbatim}
\end{minipage}
\medskip
\begin{figure*}
\begin{center}
(d)~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:category
'cat
:info-about
'topic
:Est
'info
:info-about
'topic
:France
'info
:zone
'elt
:info-about
'topic
:appartenance
'info
:Alsace
\end{verbatim}
\end{minipage}
~
(e)~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:category
'cat
:side-info
'focus
:multiplicity
'elt
:une personne
'info
:zone
'elt
:chef cuisinier
\end{verbatim}
\end{minipage}
~
(f1)~\begin{minipage}[t]{.3\textwidth}
\begin{verbatim}
:info-about
'topic
:là
'info
:info-about
'topic
:all-of
'items
list
:assiette
:assiette
'info
:multiplicity
'elt
:vendre
\end{verbatim}
\end{minipage}
\end{center}
\vspace{1cm}
\begin{center}
(f2)~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:info-about
'topic
:pour
'info
:side-info
'focus
:multiplicity
'elt
:une personne
'info
:info-about
'topic
:comment dire
'info
:difficile
\end{verbatim}
\end{minipage}
~
(f3')~\begin{minipage}[t]{.25\textwidth}
\begin{verbatim}
:info-about
'topic
:exemple
'info
:info-about
'topic
:aussi
'info
:info-about
'topic
:ici
'info
:info-about
'topic
:side-info
'focus
:Bordeaux (**)
'info
:banlieue
'info
:là
\end{verbatim}
\end{minipage}
~
(f3'')~\begin{minipage}[t]{.3\textwidth}
\begin{verbatim}
:category
'cat
:ville
'elt
:fingerspelling
'letters
list
.G
.E
.R
.S
.T
.H
.E
.I
.M
\end{verbatim}
\end{minipage}
\end{center}
\begin{quote}
In English:
\begin{description}
\item[(d)] The Alsace region in the East of France
\item[(e)] Chefs
\item[(f1)] Sell crockery
\item[(f2)] For the poor (people)
\item[(f3')] Like here in the suburbs of Bordeaux
\item[(f3'')] Gerstheim
\end{description}
\end{quote}
\caption{Aligned AZee expressions matched in algorithm run}
\label{fig:partial-matches}
\end{figure*}
\section{Discussion}
First, our anti-matches approach has some advantages, compared to a sequence-based one.
Indeed, structures that are specific to LSF can be found in the final translations, which is not the case when the language is reduced to a sequence of glosses.
In addition, the approach produces results with a certain form of creativity. In LSF, paraphrases or additions are commonly used, and indeed part of our corpus as initially delivered by the translator at the time of video corpus creation. These elements were later aligned as examples, thus frequently appear in the generated translations, although not always strictly necessary. See the example for ``\textit{Alsace}'', which although a single sign exists, is translated to the whole expression (d), typical of LSF when no context yet exists.
Moreover, the output of the algorithm is a set of translations
(built from the various substitution combinations), not necessarily a single expression. This in a way accounts for the reality of the translation task.
For example, to translate ``\textit{Emmanuel Macron}'' into LSF, different possibilities have been used by the translator, hence the different possible AZee output expressions (g), (h) and (i) below.
(g)~\begin{minipage}[t]{.5\columnwidth}
\begin{verbatim}
:Emmanuel Macron
\end{verbatim}
\end{minipage}
(h)~\begin{minipage}[t]{.5\columnwidth}
\begin{verbatim}
:side-info
'topic
:Emmanuel Macron
'info
:président
\end{verbatim}
\end{minipage}
(i)~\begin{minipage}[t]{.5\columnwidth}
\begin{verbatim}
:category
'cat
:side-info
'topic
:une personne
'info
:président
'elt
:Emmanuel Macron
\end{verbatim}
\end{minipage}
\medskip
In our test set, the number of translations proposed for a query ranges from 1 to 12 (average: 4).
At the moment, the order in which the AZee expressions are output is irrelevant. One prospect for this algorithm is to rank them according to some heuristics, for example constraints on preferred AZee rule combinations.
The work presented also has limitations.
We can observe that some anti-matches are incorrect, for instance: ``\textit{des personnes pro-Brexit}'' (pro-Brexit people) vs ``\textit{des personnes manifestent}'' (people demonstrate). The syntactic categories of the anti-match and its correction are not the same (adjective vs. verb), which creates problems during the translation process. If we want to translate ``\textit{des personnes pro-Brexit sont dans la rue}'' (pro-Brexit people are in the street), the algorithm suggests ``\textit{pro-Brexit}'' as an anti-match for ``\textit{manifestent}'', but the result is syntactically unacceptable: ``\textit{*des personnes manifestent sont dans la rue}''.
The syntactic category of each phrase should be taken into account to prevent such errors and to improve the anti-matching results.
Finally, a considerable number of fallbacks are present in the output of the algorithm: 3 per result on average.
As explained in section~\ref{general-algorithm}, this is not ideal, and the size of our corpus is undoubtedly a contributing factor: if we increase the number of examples and alignments, the number of fallbacks will decrease and the quality of the translations should hopefully improve.
\section{Conclusion and prospects}
We have presented a new system of automatic translation from text to AZee, based on an example-based machine translation approach, the hierarchical representation of SL AZee and an aligned corpus of French text and AZee descriptions extracted from the Rosetta-LSF corpus. A prototyping implementation of the system has been made and tested on some examples, thus providing a proof of concept.
The capacities of this system and the size of the corpus still need to be extended before real evaluations can be carried out. But we can already stress that the evaluation of such a system will not be easy, since it proposes a translation from one language to a representation of another language, not readable directly.
Metrics for evaluating the quality of translations, such as the one proposed in the European QT21 project\footnote{\url{https://www.qt21.eu}}, provide a scoring sheet with types of errors produced by the translation system, which allows to highlight the shortcomings of the systems and the aspects to improve.
This project has proposed Multidimensional Quality Metrics (MQM), a framework for describing and defining custom translation quality metrics. It provides a flexible vocabulary of quality issue types, a mechanism to generate quality scores, and mappings to other metrics.
Some of the error categories are defined assuming text as a target, which does not apply in our case. A category called ``fluency'' allows us to evaluate the quality of an utterance, regardless of whether it is the result of a translation. In our case, the target is not even a language utterance, thus this category will need some adjustments.
What remains is the category of errors linked to the translation process itself, categorised as ``accuracy''.
Accuracy errors happen when the target text does not reflect the source text. This could be by addition (of content not present in the source), mistranslation (output content not accurately aligned with the source) or omission (of content that was present in the source).
It would be interesting to study if this kind of evaluation could be adapted to our system.
The issue is to define these categories in the case of SL. It is common or indeed often preferred in SL to introduce contextual information, for example expression (d) figure~\ref{fig:partial-matches} for ``Alsace'', which should not be judged as an unwanted addition.
Furthermore, as AZee can be used to generate virtual signer animations which are directly ``readable'' by language users, fluency error categories could be taken into consideration after this step to complete the evaluation. The establishment of a robust and comprehensive evaluation protocol is clearly a subject of study in its own right, which needs to be pursued in the near future.
\section*{Acknowledgements}
This work has been funded by the Bpifrance investment project ``Grands défis du numérique'', as part of the ROSETTA project (RObot for Subtitling and intElligent adapTed TranslAtion).
We thank Noémie Churlet, Raphaël Bouton and Media'Pi! for their commitment to this project, which would not have had the same validity and impact without them.
\bibliographystyle{plain}
|
\section{Introduction}
Parody is a figurative device which imitates entities such as politicians and celebrities by copying their particular style or a situation where the entity was involved \citep{rose1993parody}. It is an intrinsic part of social media as a relatively new comedic form \cite{TweetReporting}. A very popular type of parody is political parody, which is used to express political opposition and civic engagement \citep{davis2018seriously}.
One of the hallmarks of parody expression is the deployment of other figurative devices, such as humor and sarcasm, as emphasized on studies of parody in linguistics \citep{haiman1998talk,Parody_Humor}. For example, in Table \ref{tab:tweet} the text expresses sarcasm about Myspace\footnote{\url{https://myspace.com}}
being a \lq winning technology\rq, while mocking the fact that three more popular social media sites were unavailable. This example also highlights the similarities between parody and real tweets, which may pose issues to misinformation classification systems~\cite{mu2020identifying}.
\renewcommand{\arraystretch}{1.0}
\begin{table}[!t]
\footnotesize
\centering
\begin{tabular}{|m{1.5cm}|m{5cm}|}
\hline
\rowcolor[gray]{.7} \textbf{Twitter Handle} &
\texttt{@Queen\_UK} \\
\hline
\textbf{Parody tweet} & Boris Johnson on the phone. Very smug that \colorbox{pink}{\#myspace} hasn't gone down. Says he's always backed \colorbox{pink}{winning technologies} \#whatsappdown \#instagramdown \#FacebookIsDown \\
\hline
\end{tabular}
\caption{Example of a \emph{parody} tweet\protect\footnotemark\ by the Twitter handle @Queen\_UK. Humor and \colorbox{pink}{sarcasm} are expressed simultaneously.}
\label{tab:tweet}
\end{table}
\footnotetext[\thefootnote]{\url{https://twitter.com/Queen\_UK/status/1445103605355323393?t=FGMNsMVFF\_G2tABYxFmkFw\&s=07}}
These figurative devices have so far been studied in isolation to parody.
Previous work on modeling humor in computational linguistics has focused on identifying jokes, i.e., short comedic passages that end with a hilarious line \citep{hetzron1991structure}, based on linguistic features \citep{1400851,purandare-litman-2006-humor,kiddon-brun-2011-thats} and deep learning techniques \citep{chen-soo-2018-humor,weller-seppi-2019-humor,annamoradnejad2020colbert}.
Similarly, computational approaches for modeling sarcasm (i.e., a form of verbal irony used to mock or convey content) in texts have been explored \citep{davidov-etal-2010-semi,gonzalez-ibanez-etal-2011-identifying,liebrecht-etal-2013-perfect,Rajadesingan2015SarcasmDO,ghosh-etal-2020-interpreting,ghosh-etal-2021-laughing}, including multi-modal utterances, i.e. texts, images, and videos \citep{cai-etal-2019-multi,castro-etal-2019-towards,oprea-magdy-2020-isarcasm}.
Recently, parody has been studied with natural language processing (NLP) methods by \citet{maronikolakis-etal-2020-analyzing} who introduced a data set of political parody accounts.
Their method for automatic recognition of posts shared by political parody accounts on Twitter is solely based on vanilla transformer models.
In this paper, we hypothesize that humor and sarcasm information could guide parody specific text encoders towards detecting nuances of figurative language. For this purpose, we propose a \emph{multi-encoder} model (\S \ref{sec:model}) consisting of three parallel encoders that are subsequently fused for parody classification. The first encoder learns parody specific information subsequently enhanced using the representations learned by a humor and sarcasm encoder respectively.
Our contributions are: (1) new state-of-the-art results on political parody detection in Twitter, consistently improving predictive performance over previous work by \citet{maronikolakis-etal-2020-analyzing}; and (2) insights on the limitations of neural models in capturing various linguistic characteristics of parody from extensive qualitative and quantitative analyses.
\section{Multi-Encoder Model for Political Parody Prediction}
\label{sec:model}
\citet{maronikolakis-etal-2020-analyzing} define political parody prediction as a binary classification task where a social media post $T$, consisting of a sequence of tokens $T=\{t_1,...,t_n\}$, is classified as real or parody. Real posts have been authored by actual politicians (e.g., \texttt{\@realDonaldTrump}) while parody posts come from their corresponding parody accounts (e.g., \texttt{\@realDonaldTrFan}).
Parody tends to express complex tangled semantics of both humor and sarcasm simultaneously \citep{haiman1998talk,Parody_Humor}. To better exploit this characteristic of parody, we propose a \emph{multi-encoder} model that consists of three parallel encoders, a feature-fusion layer and a parody classification layer depicted in Fig.\ref{fig:model}.\footnote{Early experiments with multi-task learning did not result in improved performance. The results of these experiments can be found in Appendix \ref{appendix_MTL}.}.
\begin{figure}[t!]
\centering
\includegraphics[width=5.5cm]{Model-Diagram_modified.png}
\caption{The structure of our \emph{multi-encoder} model for combining humor and sarcasm information for political parody prediction.}
\label{fig:model}
\end{figure}
\subsection{Text Encoders}
\paragraph{Parody}
As a task-specific parody encoder, we use the vanilla pretrained BERTweet \citep{nguyen-etal-2020-bertweet}, a BERT \cite{devlin-etal-2019-bert} based model pre-trained on a corpus of English Tweets and fine-tuned on the parody data set (\S \ref{sec:data}).
\paragraph{Humor}
To capture humor specific characteristics in social media text, we use the data set introduced by \citet{annamoradnejad2020colbert} which contains humorous and non-humorous short texts collected from Reddit and Huffington Post. First, we adapt BERTweet using domain-adaptive pre-training \citep{sun2020finetune,gururangan-etal-2020-dont} on 10,000 randomly selected humor-only short texts with masked language modeling.
Subsequently, we use a continual learning strategy \citep{8107520,Sun_Wang_Li_Feng_Tian_Wu_Wang_2020} to gradually learn humor-specific properties by further fine-tuning BERTweet on a humor classification task (i.e., predicting whether a text is humorous or not) by using 40,000 randomly selected humorous and non-humorous short texts from the humor corpus described above (see Figure~\ref{fig:humorEncoder}).
\begin{figure}[t!]
\centering\includegraphics[width=7.5cm]{humor_encoder.png}
\caption{\textit{Humor Encoder}.}
\label{fig:humorEncoder}
\end{figure}
\begin{figure}[t!]
\centering\includegraphics[width=7.5cm]{sarcasm_encoder.png}
\caption{\textit{Sarcasm Encoder}.}
\label{fig:sarcasmEncoder}
\end{figure}
\paragraph{Sarcasm}
Similar to humor, we extract sarcasm-related semantic information from a post $T$ by using sarcasm annotated data sets from \citet{oprea-magdy-2020-isarcasm} and \citet{Rajadesingan2015SarcasmDO}. The first data set consists of 777 and 3,707 sarcasm and non-sarcasm posts from Twitter and the second data set consists of 9,104 sarcasm and more than 90,000 non-sarcasm posts from Twitter. We first perform domain-adaptive pre-training of BERTweet on all sarcastic posts with masked language modeling. Then, we fine-tune the model on a sarcasm classification task, similar to the humor encoder (see Figure~\ref{fig:sarcasmEncoder}). For the fine-tuning step, we use the 9,881 sarcastic tweets and 10,000 randomly sampled non-sarcasm tweets from the two data sets (i.e., 3,707 from the first and 6,293 from the second).
We compute parody $f^t$, humor $f^h$, and sarcasm $f^s$ representations by extracting the `classification' [CLS] token from each encoder respectively, where $f \in \mathbf{R}^{768}$.
\subsection{Combining Encoders}
We explore three approaches to combine $f^t$, $f^h$, and $f^s$ representations.
\paragraph{Concatenation} First, the three text representations are simply concatenated to form a combined representation $f \in \mathbf{R}^{768\times3}$.
\paragraph{Self-Attention} We also use a 4-head self-attention\footnote{Early experimentation with larger attention heads did not improve results in the dev set.} mechanism \cite{NIPS2017_3f5ee243} on $f^t, f^h, f^s$. The goal is to find correlations between representations and learn the contribution of each encoder in the final representation.
\paragraph{Max-Pooling} Finally, we perform a max-pooling operation on each dimension of $f^t$, $f^h$, $f^s$ to obtain a representation $f \in \mathbf{R}^{768}$. The aim is to use the most dominant features learned by each encoder.
\subsection{Classification}
Finally, we pass the combined representation $f$ to a classification layer with a sigmoid activation function for predicting whether a post is a parody or not. Three encoders are fine-tuned simultaneously on the parody data set (\S \ref{sec:data}).\footnote{Early experimentation with humor and sarcasm encoders frozen during the fine-tuning process did not show any performance improvement.}
\section{Experimental Setup}
\subsection{Data}
\label{sec:data}
We use the data set introduced by \citet{maronikolakis-etal-2020-analyzing} which contains 131,666 tweets written in English, with 65,956 tweets from political parody accounts and 65,710 tweets posted by real politician accounts. The data set is publicly available\footnote{\url{https://archive.org/details/parody_data_acl20}} and allows us to compare our results to state-of-the-art parody detection methods.
We use the three data splits provided: (i) \emph{Person Split}, each split (train, dev, test) contains tweets from different real -- \emph{parody} account pairs; (ii) \emph{Gender Split}, two different splits based on the gender of the politicians (i.e., female accounts in train/dev and male in test, and male accounts in train/dev and female in test); \emph{Location Split}, data is split according to the location of the politicians in three groups (US, UK, Rest of the World or RoW). Each group is assigned to the test set and the other two groups to the train and dev sets.
\subsection{Baselines}
We compare our \emph{multi-encoder} models with
transformers for parody detection \citep{maronikolakis-etal-2020-analyzing}: {\bf BERT} \citep{devlin-etal-2019-bert} and {\bf RoBERTa} \citep{liu2019roberta}. Also, we compare our models to {\bf BERTweet} \citep{nguyen-etal-2020-bertweet}.
\subsection{Implementation details}
\label{imp_details}
\paragraph{Humor Encoder}
For adaptive pre-training, the batch-size is set to $16$ and the number of training epochs is set to $3$ with a learning rate of $2e^{-5}$. For humor classification, we use batch size of $128$ and the number of epochs is set to $2$ with a learning rate of $3e^{-5}$.
\paragraph{Sarcasm Encoder}
We pretrain using a batch-size of $16$ over $5$ epochs with a learning rate of $2e^{-5}$. For fine-tuning on a sarcasm classification task, we use the $9,881$ sarcasm tweets and $10,000$ randomly sampled non-sarcasm tweets from the two data sets (i.e., $3,707$ from the first and $6,293$ from the second) using the same hyperparameters to the humor-specific encoder.
\paragraph{Multi-encoder}
For the complete \emph{multi-encoder} model, we use a batch size of $128$ and the learning rate is set to $2e^{-5}$. The entire model is fine-tuned for $2$ epochs
\subsection{Evaluation}
We evaluate the performance of all models using F1 score as \citet{maronikolakis-etal-2020-analyzing}. Results are obtained over $3$ runs using different random seeds reporting average and standard deviation.
\section{Results}
\label{sec:results}
\subsection{Predictive Performance}
Table \ref{tab:person} shows the results for parody detection on the \emph{Person Split}.
We observe that \textit{BERTweet} has the best performance (F1: $90.72$) among transformer-based models (\textit{BERT}, \textit{RoBERTa}, \textit{BERTweet}), outperforming previous state-of-the-art by \citet{maronikolakis-etal-2020-analyzing}. This is due to the fact that \textit{BERTweet} has been specifically pre-trained on Twitter text. Similar behavior is observed on the \emph{Gender} and \emph{Location} splits (see Table~\ref{tab:gender} and \ref{tab:location} respectively).
\renewcommand{\arraystretch}{1.1}
\begin{table}[!t]
\small
\centering
\begin{tabular}{|l| c|}
\rowcolor[gray]{.7}\multicolumn{2}{|c|}{\textbf{Person}} \\\hline
\rowcolor[gray]{.7}\multicolumn{1}{|c|}{\textbf{Model}}& \multicolumn{1}{|c|}{\textbf{F1}}
\\ \hline
\rowcolor[gray]{0.9}\multicolumn{2}{|l|}{\textbf{Single-Encoder}}\\
\cellcolor[gray]{1}BERT$^{\ast\ast}$ & $87.65\pm0.18$
\\
\cellcolor[gray]{1}RoBERTa$^{\ast\ast}$ & $89.66\pm0.33$
\\
\cellcolor[gray]{1}BERTweet & $90.72\pm0.31$
\\\hline
\rowcolor[gray]{0.9} \multicolumn{2}{|l|}{\textbf{Multi-encoder (Ours)}}\\
\cellcolor[gray]{1}Concatenation & $88.99\pm0.17$
\\
\cellcolor[gray]{1}Self-Attention & $\textbf{91.19} \pm\textbf{0.31}$
\\
\cellcolor[gray]{1}Max-Pooling & $91.05\pm0.30
\\ \hline
\end{tabular}
\caption{F1-scores for parody detection on the \emph{Person Split}. $^{\ast\ast}$ Results from \citet{maronikolakis-etal-2020-analyzing}.
Best results are in bold.}
\label{tab:person}
\end{table}
\begin{table}[!t]
\renewcommand{\arraystretch}{1.1}
\centering
\small
\begin{tabular}{ |l| c|c|}
\hline
\rowcolor[gray]{.7}\multicolumn{3}{|c|}{\textbf{Gender}} \\\hline
\rowcolor[gray]{.7}\multicolumn{1}{|c|}{\textbf{Model}} & \textbf{M$\to$F} & \textbf{F$\to$M} \\ \hline
\rowcolor[gray]{0.9} \multicolumn{3}{|l|}{\textbf{Single-Encoder}}\\
\cellcolor[gray]{1}BERT$^{\ast\ast}$ & $85.85\pm0.28$ & $84.40\pm0.35$\\
\cellcolor[gray]{1}RoBERTa$^{\ast\ast}$ & $87.11\pm0.31$ & $84.87\pm0.38$\\
\cellcolor[gray]{1}BERTweet & $88.01\pm0.29$ & $85.57\pm0.27$\\ \hline
\rowcolor[gray]{0.9}\multicolumn{3}{|l|}{\textbf{Multi-encoder (Ours)}} \\
\cellcolor[gray]{1}Concatenation & $86.84\pm0.15$ & $84.21\pm0.22$\\
\cellcolor[gray]{1}Self-Attention & $\textbf{89.97}\pm\textbf{0.34}$ & $\textbf{88.56}\pm\textbf{0.39}$\\
\cellcolor[gray]{1}Max-Pooling & $88.39\pm0.27$ & $86.89\pm0.56$\\ \hline
\end{tabular}
\caption{F1-scores on the \emph{Gender Split}.
$^{\ast\ast}$ Results from \citet{maronikolakis-etal-2020-analyzing}. Best results are in bold.}
\label{tab:gender}
\end{table}
Our proposed \emph{multi-encoder} achieves the best performance when using \textit{Self-Attention} to combine the three parallel encoders (F1: $91.19$; $89.97$, $88.56$; $88.37$, $87.91$, $87.16$; for \emph{Person}, \emph{Gender}, and \emph{Location} splits respectively). Moreover, it outperforms the best single-encoder model \textit{BERTweet} in the majority of cases which corroborates that parody detection benefits from combining general contextual representations with humor and sarcasm specific information, as humor and sarcasm are important characteristics of parody \citep{haiman1998talk,Parody_Humor}. On the other hand, simply concatenating the three parallel encoders degrades the performance across different splits (\emph{Person}: $88.99$; \emph{Gender}: $86.84$, $84.21$ \emph{Location}: $85.41$, $84.74$, $83.62$). This happens because the concatenation operation treats the three encoders as equally important. While humor and sarcasm are related to parody, they may not necessarily have the same relevance as indicators of parody.
Our best performing model (\textit{Self-Attention}) outperforms the vanilla \textit{BERTweet} by $3$ F1 points when trained on female accounts and by almost $2$ F1 points when trained on male accounts. We speculate that the additional linguistic information from the two encoders (i.e., sarcasm and humor) is more beneficial in low data settings. The number of female politicians is considerably smaller than males in the data set (see \citet{maronikolakis-etal-2020-analyzing} for more details).
\renewcommand{\arraystretch}{1.1}
\begin{table}[!t]
\centering
\small
\resizebox{\linewidth}{!}{
\begin{tabular}{ |l|c|c|c|}
\hline
\rowcolor[gray]{.7}\multicolumn{4}{|c|}{\textbf{Location}}\\\hline
\rowcolor[gray]{.7}\multicolumn{1}{|c|}{\textbf{Model}} & \begin{tabular}{c}\textbf{UK+US}\\ \textbf{$\to$ RoW}\end{tabular} & \begin{tabular}{c}\textbf{RoW+US} \\ \textbf{$\to$ UK} \end{tabular}
& \begin{tabular}{c}\textbf{RoW+UK} \\ \textbf{$\to$ US}\end{tabular}\\ \hline
\rowcolor[gray]{0.9} \multicolumn{4}{|l|}{\textbf{Single-Encoder}}\\
\cellcolor[gray]{1}BERT$^{\ast\ast}$ & $86.69\pm0.45$ & $83.78\pm0.19$ & $83.12\pm0.60$\\
\cellcolor[gray]{1}RoBERTa$^{\ast\ast}$ & $87.70\pm0.45$ & $85.10\pm0.27$ & $85.99\pm0.61$\\
\cellcolor[gray]{1}BERTweet & $88.21\pm0.26$ & $87.85\pm0.24$ & $\textbf{87.18}\pm\textbf{0.41}$\\ \hline
\rowcolor[gray]{0.9} \multicolumn{4}{|l|}{\textbf{Multi-encoder (Ours)}}\\
\cellcolor[gray]{1}Concatenation & $85.41\pm0.26$ & $84.74\pm0.20$ & $83.62\pm0.35$\\
\cellcolor[gray]{1}Self-Attention & $\textbf{88.37}\pm\textbf{0.28}$ & $\textbf{87.91}\pm\textbf{0.19}$ & $87.16\pm0.37$\\
\cellcolor[gray]{1}Max-Pooling & $88.25\pm0.39$ & $86.49\pm0.33$ & $86.54\pm0.41$\\ \hline
\end{tabular}}
\caption{F1-scores on the \emph{Location Split}.
$^{\ast\ast}$ Results from \citet{maronikolakis-etal-2020-analyzing}. Best results are in bold.
}
\label{tab:location}
\end{table}
\subsection{Ablation Study}
We also examine the effect of combining parody-specific representations with humor and sarcasm information by running an ablation study. We compare performance of four models: using parody representations only (P), and combining parody representations with humor (P+H), or sarcasm (P+S) information, as well as with both (P+S+H). The results of this analysis are depicted in Tables~\ref{tab:person_complete}, \ref{tab:gender_complete} and \ref{tab:location_complete}. We observe that both sarcasm and humor contribute to the performance gain, but using both is more beneficial. Modelling sarcasm leads to more gains than humor and this could be attributed to the characteristics of the parody corpus, namely that it focuses primarily on the political domain, which have a high sarcastic component \citep{anderson2017social}.
\renewcommand{\arraystretch}{1.2}
\begin{table}[!h]
\small
\centering
\begin{tabular}{|l| c|}
\rowcolor[gray]{.7}\multicolumn{2}{|c|}{\textbf{Person}} \\\hline
\rowcolor[gray]{.7}\multicolumn{1}{|c|}{\textbf{Model}}& \multicolumn{1}{|c|}{\textbf{F1}}
\\ \hline
\rowcolor[gray]{0.9}\multicolumn{2}{|l|}{\textbf{Single-Encoder}}\\
\cellcolor[gray]{1}BERTweet (P) & $90.72\pm0.31$
\\\hline
\rowcolor[gray]{0.9} \multicolumn{2}{|l|}{\textbf{Multi-encoder (Ours)}}\\
\cellcolor[gray]{1}Concatenation (P+S+H) & $88.99\pm0.17$
\\
\cellcolor[gray]{1}Concatenation (P+S) & $90.51\pm0.26$
\\
\cellcolor[gray]{1}Concatenation (P+H) & $89.98\pm0.23$
\\
\cellcolor[gray]{1}Self-Attention (P+S+H) & $\textbf{91.19} \pm\textbf{0.31}$
\\
\cellcolor[gray]{1}Self-Attention (P+S) & $91.14\pm0.40$
\\
\cellcolor[gray]{1}Self-Attention (P+H)& $90.98\pm0.36$
\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+S+H) & $91.05\pm0.30
\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+S) & $91.06\pm0.39
\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+H) & $90.78\pm0.42
\\ \hline
\end{tabular}
\caption{F1-scores for parody detection on the \emph{Person Split} with various settings: parody (P) representations only, and combining parody representations with humor (P+H), or sarcasm (P+S) information, as well as with both (P+S+H). Best results are in bold.
}
\label{tab:person_complete}
\end{table}
\renewcommand{\arraystretch}{1.2}
\begin{table}[!t]
\centering
\small
\begin{tabular}{ |l| c|c|}
\hline
\rowcolor[gray]{.7}\multicolumn{3}{|c|}{\textbf{Gender}} \\\hline
\rowcolor[gray]{.7}\multicolumn{1}{|c|}{\textbf{Model}} & \textbf{M$\to$F} & \textbf{F$\to$M} \\ \hline
\rowcolor[gray]{0.9} \multicolumn{3}{|l|}{\textbf{Single-Encoder}}\\
\cellcolor[gray]{1}BERTweet (P) & $88.01\pm0.29$ & $85.57\pm0.27$\\ \hline
\rowcolor[gray]{0.9}\multicolumn{3}{|l|}{\textbf{Multi-encoder (Ours)}} \\
\cellcolor[gray]{1}Concatenation (P+S+H)& $86.84\pm0.15$ & $84.21\pm0.22$\\
\cellcolor[gray]{1}Concatenation (P+S)& $86.93\pm0.40$ & $83.70\pm0.41$\\
\cellcolor[gray]{1}Concatenation (P+H)& $86.58\pm0.31$ & $83.34\pm0.38$\\
\cellcolor[gray]{1}Self-Attention (P+S+H)& $\textbf{89.97}\pm\textbf{0.34}$ & $\textbf{88.56}\pm\textbf{0.39}$\\
\cellcolor[gray]{1}Self-Attention (P+S)& $89.49\pm0.37$ & $88.23\pm0.44$\\
\cellcolor[gray]{1}Self-Attention (P+H)& $88.71\pm0.42$ & $87.62\pm0.50$\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+S+H)& $88.39\pm0.27$ & $86.89\pm0.56$\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+S)& $88.36\pm0.46$ & $86.55\pm0.49$\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+H)& $88.14\pm0.52$ & $86.53\pm0.53$\\\hline
\end{tabular}
\caption{
F1-scores for parody detection on the \emph{Gender Split} with various settings: parody (P) representations only, and combining parody representations with humor (P+H), or sarcasm (P+S) information, as well as with both (P+S+H). Best results are in bold.}
\label{tab:gender_complete}
\end{table}
\renewcommand{\arraystretch}{1.2}
\begin{table}[!t]
\centering
\small
\resizebox{\linewidth}{!}{
\begin{tabular}{ |l|c|c|c|}
\hline
\rowcolor[gray]{.7}\multicolumn{4}{|c|}{\textbf{Location}}\\\hline
\rowcolor[gray]{.7}\multicolumn{1}{|c|}{\textbf{Model}} & \begin{tabular}{c}\textbf{UK+US}\\ \textbf{$\to$ RoW}\end{tabular} & \begin{tabular}{c}\textbf{RoW+US} \\ \textbf{$\to$ UK} \end{tabular}
& \begin{tabular}{c}\textbf{RoW+UK} \\ \textbf{$\to$ US}\end{tabular}\\ \hline
\rowcolor[gray]{0.9} \multicolumn{4}{|l|}{\textbf{Single-Encoder}}\\
\cellcolor[gray]{1}BERTweet (P) & $88.21\pm0.26$ & $87.85\pm0.24$ & $\textbf{87.18}\pm\textbf{0.41}$\\ \hline
\rowcolor[gray]{0.9} \multicolumn{4}{|l|}{\textbf{Multi-encoder (Ours)}}\\
\cellcolor[gray]{1}Concatenation (P+S+H) & $85.41\pm0.26$ & $84.74\pm0.20$ & $83.62\pm0.35$\\
\cellcolor[gray]{1}Concatenation (P+S) & $85.92\pm0.24$ & $85.67\pm0.18$ & $84.09\pm0.39$\\
\cellcolor[gray]{1}Concatenation (P+H) & $85.39\pm0.29$ & $85.33\pm0.26$ & $83.75\pm0.44$\\
\cellcolor[gray]{1}Self-Attention (P+S+H) & $\textbf{88.37}\pm\textbf{0.28}$ & $\textbf{87.91}\pm\textbf{0.19}$ & $87.16\pm0.37$\\
\cellcolor[gray]{1}Self-Attention (P+S) & $88.24\pm0.33$ & $87.88\pm0.23$ & $86.47\pm0.32$\\
\cellcolor[gray]{1}Self-Attention (P+H) & $88.13\pm0.35$ & $87.05\pm0.28$ & $85.36\pm0.40$\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+S+H) & $88.25\pm0.39$ & $86.49\pm0.33$ & $86.54\pm0.41$\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+S) & $88.28\pm0.42$ & $87.83\pm0.39$ & $86.56\pm0.36$\\
\cellcolor[gray]{1}Max-Pooling\hspace{0.44em} (P+H) & $88.22\pm0.52$ & $86.44\pm0.42$ & $85.96\pm0.45$\\
\hline
\end{tabular}
}
\caption{
F1-scores for parody detection on the \emph{Location Split} with various settings: parody (P) representations only, and combining parody representations with humor (P+H), or sarcasm (P+S) information, as well as with both (P+S+H). Best results are in bold.}
\label{tab:location_complete}
\end{table}
\section{Error Analysis}
Finally, we perform an error analysis to examine the behavior and limitations of our best-performing model (\emph{multi-encoder} with Self-Attention).
The next two examples correspond to real tweets that were misclassified as parody:
\begin{itemize}
\item [(1)] \textit{Congratulations, <mention>! <url>.}
\item [(2)] \textit{It's a shame that Boris isn't here answering questions from the public this evening.}
\end{itemize}
We speculate that the model misclassified these tweets as parody because they contain terms that are related to sarcastic short texts such as user mentions, punctuation marks (\textit{!}), and negation (\textit{isn't}) \cite{gonzalez-ibanez-etal-2011-identifying,Parody_Humor}.
The following two examples correspond to parody tweets that were misclassified as real:
\begin{itemize}
\item [(3)] \textit{Hey America, it’s time to use your safe word.}
\item [(4)]\textit{I fully support the Digital Singles Market.}
\end{itemize}
Example (3) is a call-to-action message, while Example (4) is a statement expressing support for a particular subject. These statements are written in a style that is similar to political slogans or campaign speeches \citep{fowler2021political} that the model fails to recognise. As a result, in addition to humor and sarcasm semantics, the model might be improved by integrating knowledge from the political domain such as from political speeches.
\section{Conclusion}
In this paper, we studied the impact of jointly modelling figurative devices to improve predictive performance of political parody detection in tweets. Our motivation was based on studies in linguistics which emphasize the humorous and sarcastic components of parody \citep{haiman1998talk,Parody_Humor}. We presented a method that combines parallel encoders to capture parody, humor, and sarcasm specific representations from input sequences, which outperforms previous state-of-the-art proposed by \citet{maronikolakis-etal-2020-analyzing}.
In the future, we plan to combine information from other modalities (e.g., images) for improving parody detection~\citep{sanchez-villegas-aletras-2021-point,sanchez-villegas-etal-2021-analyzing}.
\section*{Acknowledgements}
We would like to thank all reviewers for their valuable feedback. DSV is supported by the Centre for Doctoral Training in Speech and Language Technologies (SLT) and their Applications funded by the
UK Research and Innovation grant EP/S023062/1.
|
\section{Introduction}
\label{sec:intro}
Ideally, the observations in a sample would be independent of each
other and would each contribute roughly the same amount of information
about the parameter(s) of interest. From the earliest days of
econometrics, it has been recognized that this ideal situation often
does not apply to time\kern 0.04167em-series data, because there may be serial
correlation. But it has taken much longer for econometricians to
realize that it generally does not apply to cross\kern 0.04167em-section data
either. The first important step, following \citet{White_1980}, was to
allow for heteroskedasticity of unknown form, and for quite some time
this was the default in empirical work that used cross\kern 0.04167em-section
data. More recently, however, it has become common for investigators
to drop the assumption of independence as well as the assumption of
homoskedasticity.
There are many ways in which cross\kern 0.04167em-section observations might be
dependent, and sometimes it is possible to model this dependence
explicitly. For example, there is a large literature on spatial
econometrics and statistics, in which each observation is associated
with a point in space, and the correlation between any two observations
is assumed to depend (usually in a rather simple parametric way) on the
distance between them. See, among many others, \citet{Anselin_1988},
\citet*{Gelfand_2010}, and \citet{Corrado_2012}. However, there are a
great many cases in which either the ``distance'' between any pair of
observations cannot be measured, or the correlation between them is
not related to distance in any way that can readily be modeled.
A more widely applicable approach, on which we focus in this paper, is
to employ cluster-robust inference in the context of least-squares
estimation. This approach has become increasingly popular over the past
quarter century and is now used routinely in a great deal of empirical
microeconomic work. The idea is to divide the sample into $G$ disjoint
clusters. Depending on the nature of the data, the clusters might
correspond to classrooms, schools, families, villages, hospitals, firms,
industries, years, cities, counties, states, or countries. This list is
by no means exhaustive. Any pattern of heteroskedasticity and/or
dependence is allowed within each cluster, but it is assumed that there
is independence across clusters and that the assignment of observations
to clusters is known.
Under these assumptions, it is easy to compute cluster-robust standard
errors that can be used to produce asymptotically valid inferences;
see \Cref{sec:CRVE}. However, these inferences may not be at all
reliable in finite samples. Hypothesis tests may reject far more often
than they should. Less commonly, they may reject far less often. In
consequence, the actual coverage of confidence intervals may differ
greatly from their nominal coverage. Therefore, in practice, using
cluster-robust inference often requires a good deal of care.
There are several recent survey papers on cluster-robust inference,
including \citet{CM_2015}, \citet{JGM-CJE}, \citet{Esarey_2019}, and
\citet{MW-survey}. \citet*{CGH_2018} surveys a broader class of
methods for various types of dependent data. Although there will
inevitably be some overlap with these papers, our aim is to provide a
guide to empirical practice rather than a survey of the extant
literature. We therefore apologize for any missing references and
refer the reader to the survey papers just mentioned for more complete
bibliographies. Our guide is closely based on the econometric theory
and simulation evidence that is currently available. When the theory
is clear and the evidence is strong, we make definitive
recommendations for empirical practice. However, when the theory is
less clear or the evidence is weak, our recommendations are more
guarded.
This guide does not discuss models with clustered data estimated by
instrumental variables (IV). For such models, neither the current state
of econometric theory nor the available simulation evidence allows us
to make recommendations with any confidence. The number of
over-identifying restrictions and the strength of the instruments can
greatly affect the reliability of finite\kern 0.04167em-sample IV inference, and
dealing with these issues may often be even more important than dealing
with the issues associated with clustering. There is an enormous
literature on the topic of weak instruments; see \citet*{Andrews_2019}
for a recent survey. That paper suggests that, when the disturbances
of a regression model are independent and homoskedastic, it is generally
possible to obtain reliable (although perhaps imprecise) inferences even
when the instruments are quite weak. However, it also states that this
is not the case, in general, when there is heteroskedasticity and/or
clustering.
In \Cref{sec:CRVE}, we obtain the (true) variance matrix for the
coefficient estimators in a linear regression model with clustered
data. The form of this matrix depends on critical assumptions about the
score vectors for each cluster. In practice, inference must be based
on a cluster-robust variance estimator, or CRVE, which estimates the
unknown variance matrix. We discuss the three CRVEs that are commonly
encountered.
\Cref{sec:why} deals with the important and sometimes controversial
issue of when to use cluster-robust inference. It also illustrates how
complicated patterns of intra-cluster correlation can arise in the
context of a simple factor model, introduces the concept of leverage
at the cluster level, discusses the role of cluster fixed effects, and
describes several procedures for deciding the level at which to
cluster.
\Cref{sec:asytheory} concerns the key issue of asymptotic inference.
It explains how to obtain asymptotically valid inferences and
discusses what determines how reliable, or unreliable, these
inferences are likely to be in practice. In many cases, bootstrap
inference tends to be more reliable than asymptotic inference.
\Cref{sec:bootstrap} describes two methods for bootstrap inference,
namely, the pairs cluster bootstrap and the restricted version of the
wild cluster bootstrap, which is called the WCR bootstrap. The former
has the advantage of being applicable to a wide variety of econometric
models, while the latter is only applicable to regression models with
clustered data, for which it typically performs better. For clustered
linear regression models, both of these methods can be remarkably
inexpensive to implement, even for very large samples. We recommend
that, in most cases, the WCR bootstrap be among the methods employed
for inference.
\Cref{sec:other} goes on to discuss some related inferential
procedures. The first of these uses an alternative critical value
estimated from the data, and the second is randomization inference,
which can work well in certain cases where even the WCR bootstrap
fails. \Cref{sec:report} discusses what an empirical investigator
should report in order to convince the reader that results are
reliable. \Cref{sec:empirical} presents an empirical example that uses
individual data to study the effects of the minimum wage on the labor
supply of teenagers. \Cref{sec:conc} provides a summary of the main
points of the paper. This is presented in the form of a short
checklist or guide for empirical researchers on what to do in
practice, with references to relevant sections.
\section{Cluster-Robust Variance Estimators}
\label{sec:CRVE}
\subsection{The Clustered Regression Model}
\label{subsec:Model}
Throughout the paper, we deal with the linear regression model $y_i =
{\bm{x}}_i^\top\kern -0.08333em{\bm\beta} + u_i$, which, if the data have been divided into
$G$ disjoint clusters, can be rewritten as
\begin{equation}
\label{eq:lrmodel}
{\bm{y}}_g ={\bm{X}}_g{\bm\beta} + {\bm{u}}_g, \quad g=1,\ldots,G.
\end{equation}
Here ${\bm{X}}_g$ is an $N_g\times k$ matrix of exogenous regressors,
${\bm\beta}$ is a $k$-vector of coefficients, ${\bm{y}}_g$ is an $N_g$-vector
of observations on the regressand, and ${\bm{u}}_g$ is an $N_g$-vector of
disturbances (or error terms). Thus ${\bm{X}}_g$, ${\bm{y}}_g$, and ${\bm{u}}_g$
stack the ${\bm{x}}_i^\top$, $y_i$, and $u_i$, respectively. In many
cases, the regressors will include cluster fixed effects; see
\Cref{subsec:FE}. Since the \th{g} cluster has $N_g$ observations,
the sample size is $N = \sum_{g=1}^G N_g$. The ${\bm{X}}_g$ may of course
be stacked into an $N\times k$ matrix ${\bm{X}}$\kern -0.08333em, and likewise the
${\bm{y}}_g$ and ${\bm{u}}_g$ may be stacked into $N$-vectors ${\bm{y}}$ and
${\bm{u}}$, so that \eqref{eq:lrmodel} can be rewritten in the usual way
as ${\bm{y}} = {\bm{X}}\!{\bm\beta} + {\bm{u}}$.
It is assumed that the data are generated by \eqref{eq:lrmodel} with
${\bm\beta}={\bm\beta}_0$. Under this assumption, the OLS estimator of
${\bm\beta}$ is
\begin{equation*}
\hat{\bm\beta} = ({\bm{X}}^\top{\bm{X}})^{-1}{\bm{X}}^\top{\bm{y}}
= {\bm\beta}_0 + ({\bm{X}}^\top{\bm{X}})^{-1}{\bm{X}}^\top{\bm{u}}.
\end{equation*}
It follows that
\begin{equation}
\hat{\bm\beta} - {\bm\beta}_0 = ({\bm{X}}^\top{\bm{X}})^{-1}\sum_{g=1}^G {\bm{X}}_g^\top{\bm{u}}_g
= \Big(\kern -0.08333em\sum_{g=1}^G{\bm{X}}_g^\top{\bm{X}}_g\Big)^{\!-1} \sum_{g=1}^G {\bm{s}}_g,
\label{eq:betahat}
\end{equation}
where ${\bm{s}}_g = {\bm{X}}_g^\top {\bm{u}}_g$ denotes the $k\times1$ score
vector corresponding to the \th{g} cluster. For a correctly specified
model, ${\rm E}({\bm{s}}_g)={\bm{0}}$ for all $g$. From the rightmost expression
in \eqref{eq:betahat}, the distribution of the OLS estimator
$\hat{\bm\beta}$ depends on ${\bm{u}}$ only through the distribution of the
score vectors ${\bm{s}}_g$. Ideally, the sum of the ${\bm{s}}_g$, suitably
normalized, would be well approximated by a multivariate normal
distribution with mean zero.
Because we can always divide the sample into $G$ clusters in any way
we like, \eqref{eq:betahat} is true for any distribution of the
disturbance vector ${\bm{u}}$. Dividing the sample into clusters only
becomes meaningful if we further assume that
\begin{equation}
\label{eq:Sigma_g}
{\rm E}({\bm{s}}_g{\bm{s}}_g^\top) = {\bm{\Sigma}}_g \quad\mbox{and}\quad
{\rm E}({\bm{s}}_g{\bm{s}}_{g'}^\top) = {\bm{0}}, \quad g,g'=1,\ldots,G,\quad g'\ne g,
\end{equation}
where the variance matrix of the scores for the \th{g} cluster,
${\bm{\Sigma}}_g$, is a $k\times k$ symmetric, positive semidefinite matrix.
The second assumption in \eqref{eq:Sigma_g} is the key one. It states
that the scores for every cluster are uncorrelated with the scores for
every other cluster. In contrast, the first assumption imposes no real
limitations, so that the ${\bm{\Sigma}}_g$ matrices may display any patterns
of heteroskedasticity and/or within-cluster dependence. Indeed, one
motivation for using cluster-robust inference is that it is robust
against both heteroskedasticity and intra-cluster dependence without
imposing any restrictions on the (unknown) form of either of them.
For now, we will simply assume that \eqref{eq:Sigma_g} holds for some
specified division of the observations into clusters. Although the
choice of clustering structure is often controversial, or at least
somewhat debatable, the structure is almost always assumed known in
both theoretical and applied work. The important issue of how to choose
the clustering structure will be discussed below in \Cref{subsec:level}.
It follows immediately from \eqref{eq:betahat} that an estimator of the
variance of $\hat{\bm\beta}$ should be based on the usual sandwich formula,
\begin{equation}
\label{eq:trueV}
({\bm{X}}^\top{\bm{X}})^{-1} \Big(\kern -0.08333em\sum_{g=1}^G {\bm{\Sigma}}_g\kern -0.08333em\Big)
({\bm{X}}^\top{\bm{X}})^{-1}.
\end{equation}
Of course, this matrix cannot be computed, because we need to estimate
the ${\bm{\Sigma}}_g$. This can be done in several ways, as we discuss in
\Cref{subsec:CVmats}.
As \eqref{eq:betahat} makes clear, it is the properties of the score
vectors that matter for inference. Of course, those properties are
inherited from the properties of the disturbances and the regressors.
If ${\bm{\Omega}}_g = {\rm E} ({\bm{u}}_g{\bm{u}}_g^\top |{\bm{X}})$ denotes the conditional
variance matrix of ${\bm{u}}_g$, then
\begin{equation}
{\bm{\Sigma}}_g = {\rm E}({\bm{X}}_g^\top{\bm{\Omega}}_g{\bm{X}}_g), \quad g=1,\ldots,G.
\label{sigomega}
\end{equation}
Thus, instead of making assumptions directly about the ${\bm{\Sigma}}_g$, as
we did in \eqref{eq:Sigma_g}, it may be more illuminating to make
assumptions about the ${\bm{\Omega}}_g$ and the ${\bm{X}}_g$. If
${\rm E}({\bm{u}}_g{\bm{u}}_{g'}^\top |{\bm{X}}) = {\bm{0}}$ for all $g' \neq g$, then
the second assumption in \eqref{eq:Sigma_g} will hold. It will also
hold if the regressors are exogenous and uncorrelated across clusters
even when the disturbances are not.
Since the score vector ${\bm{s}}_g$ can be written as $\sum_{i=1}^{N_g}
{\bm{s}}_{gi} = \sum_{i=1}^{N_g} {\bm{X}}^\top_{gi} u_{gi}$, where ${\bm{X}}_{gi}$
is the \th{i} row of ${\bm{X}}_g$ and $u_{gi}$ is the \th{i} element of
${\bm{u}}_g$, the outer product of the score vector with itself is seen to
be
\begin{equation}
{\bm{s}}_g{\bm{s}}_g^\top =
\Big(\kern -0.08333em\sum_{i=1}^{N_g} {\bm{X}}^\top_{gi} u_{gi}\Big)
\Big(\kern -0.08333em\sum_{i=1}^{N_g} {\bm{X}}^\top_{gi} u_{gi}\Big)^{\!\!\top} =
\sum_{i=1}^{N_g} \sum_{j=1}^{N_g} {\bm{X}}^\top_{gi} {\bm{X}}_{gj} u_{gi} u_{gj}
= \sum_{i=1}^{N_g} \sum_{j=1}^{N_g} {\bm{s}}_{gi} {\bm{s}}_{gj}^\top.
\label{eq:opscores}
\end{equation}
When ${\rm E}(u_{gi}^2 | {\bm{X}}) = \sigma^2$ and ${\rm E}(u_{gi} u_{gj} | {\bm{X}}) =
0$ for $i \neq j$, then ${\rm E} ( {\bm{s}}_g{\bm{s}}_g^\top | {\bm{X}} ) = \sigma^2
{\bm{X}}_g^\top\kern -0.08333em {\bm{X}}_g$. In that case, we would replace ${\bm{\Sigma}}_g$
with $\sigma^2 ({\bm{X}}_g^\top{\bm{X}}_g )$ in \eqref{eq:trueV} and obtain
the classic result that $\var (\hat{\bm\beta} | {\bm{X}} ) =
\sigma^2({\bm{X}}^\top {\bm{X}})^{-1}$.
Taking expectations in \eqref{eq:opscores} and defining the covariance
matrix ${\bm{\Sigma}}_{g,ij} = {\rm E} ( {\bm{s}}_{gi}{\bm{s}}^\top_{gj} )$, we find
that, in general, ${\bm{\Sigma}}_g = \sum_{i=1}^{N_g}\sum_{j=1}^{N_g}{\rm E}
({\bm{s}}_{gi} {\bm{s}}_{gj}^\top)
=\sum_{i=1}^{N_g}\sum_{j=1}^{N_g}{\bm{\Sigma}}_{g,ij}$. In the special case
where the score vectors ${\bm{s}}_{gi}$ are uncorrelated within each
cluster, i.e.\ where ${\bm{\Sigma}}_{g,ij}={\bm{0}}$ for $i \neq j$, we find
that ${\bm{\Sigma}}_g = \sum_{i=1}^{N_g}{\rm E} ({\bm{s}}_{gi}{\bm{s}}_{gi}^\top) =
\sum_{i=1}^{N_g} {\bm{\Sigma}}_{g,ii}$. The difference between these two
expressions for ${\bm{\Sigma}}_g$ is
\begin{equation}
\sum_{i=1}^{N_g} \sum_{j=1}^{N_g} {\rm E}\big({\bm{s}}_{gi}{\bm{s}}_{gj}^\top\big)
- \sum_{i=1}^{N_g} {\rm E}\big({\bm{s}}_{gi} {\bm{s}}_{gi}^\top\big)
= \sum_{i=1}^{N_g} \sum_{j\ne i} {\bm{\Sigma}}_{g,ij}.
\label{eq:diffop}
\end{equation}
The rightmost expression in \eqref{eq:diffop} is just the summation of
the $N_g^2 - N_g$ matrices that correspond to the off-diagonal
elements of ${\bm{\Sigma}}_g$. It equals zero whenever there is no
intra-cluster correlation, but in general it is $O(N_g^2)$. Therefore,
incorrectly assuming that the scores are not correlated within
clusters potentially leads to much larger errors of inference when
clusters are large than when they are small. For sufficiently large
values of $N_g$, these errors may be large even when all of the
${\bm{\Sigma}}_{g,ij}$ for $i\ne j$ are very small \citep{JGM_2016}.
The famous ``Moulton factor'' \citep{Moulton_1986} gives the ratio of
the true variance of an OLS coefficient, from \eqref{eq:trueV}, to the
variance based on the classic formula $\sigma^2({\bm{X}}^\top{\bm{X}})^{-1}$
under the assumption that both the disturbances and the regressor of
interest (after other regressors have been partialed out) are
equi-correlated within clusters; see \Cref{subsec:sources}. If the
scores were scalars with intra-cluster correlation $\rho_s$, and the
cluster sizes were constant, say $N_g = M$\kern -0.08333em, then the Moulton factor
would be $1+(M-1)\rho_s$. The second term is proportional to the
number of observations per cluster, so the mistakes made by not
clustering can be enormous when clusters are large.
Since the disturbances in \eqref{eq:lrmodel} are neither independent
nor homoskedastic, it seems relevant to consider GLS estimation, even
though OLS estimation is almost always used in practice. If we were
willing to specify a simple parametric form for the ${\bm{\Omega}}_g$
matrices, then we could use feasible GLS. For example, if we assumed
that the disturbances were equi-correlated within each cluster, that
would be equivalent to specifying a random-effects model; see
\Cref{subsec:sources}. In practice, however, the regressors in
\eqref{eq:lrmodel} very often include cluster fixed effects
(\Cref{subsec:FE}), and the latter remove whatever intra-cluster
correlation a random-effects specification induces. So we would need
to specify a more complicated model if we wanted to use feasible GLS.
In any case, specifying a parametric form for the intra-cluster
correlations would imply making assumptions much stronger than those
in \eqref{eq:Sigma_g}, and this would violate the principal objective
of cluster-robust inference, namely, to be robust to arbitrary and
unknown dependence and heteroskedasticity within clusters.
\subsection{Three Feasible CRVEs}
\label{subsec:CVmats}
The natural way to estimate \eqref{eq:trueV} is to replace the
${\bm{\Sigma}}_g$ matrices by their empirical counterparts, which are the
outer products of the empirical score vectors $\hat{\bm{s}}_g =
{\bm{X}}_g^\top \hat{\bm{u}}_g$ with themselves. If, in addition, we multiply
by a correction for degrees of freedom, we obtain
\begin{equation}
\mbox{CV$_{\tn1}$:}\qquad
\frac{G(N-1)}{(G-1)(N-k)}
({\bm{X}}^\top{\bm{X}})^{-1}
\Big(\kern -0.08333em\sum_{g=1}^G \hat{\bm{s}}_g\hat{\bm{s}}_g^\top\Big) ({\bm{X}}^\top{\bm{X}})^{-1}.
\label{eq:CV1}
\end{equation}
At present, this is by far the most widely used CRVE in practice.
Observe that, when $G=N$\kern -0.08333em, CV$_{\tn1}$ reduces to the familiar
HC$_1$ estimator \citep{MW_1985} that is robust only to
heteroskedasticity of unknown form.
The empirical score vectors $\hat{\bm{s}}_g$ are not always good
estimators of the ${\bm{s}}_g$. CV$_{\tn1}$ attempts to compensate for
this by including a degrees-of-freedom factor. Two alternative CRVEs,
proposed in \citet{BM_2002}, instead replace the empirical score
vectors $\hat{\bm{s}}_g$ by modified score vectors that use transformed
residuals. The first of these is
\begin{equation}
\mbox{CV$_{\tn2}$:}\qquad
({\bm{X}}^\top{\bm{X}})^{-1}\Big(\kern 0.08333em\sum_{g=1}^G \grave{\bm{s}}_g\grave{\bm{s}}_g^\top
\Big)({\bm{X}}^\top{\bm{X}})^{-1},
\label{eq:CV2}
\end{equation}
where $\grave{\bm{s}}_g = {\bm{X}}_g^\top{\bm{M}}_{gg}^{-1/2}\kern 0.08333em\hat{\bm{u}}_g$, with
${\bm{M}}_{gg} = {\bf I}_{N_g} - {\bm{X}}_g({\bm{X}}^\top{\bm{X}})^{-1}\kern -0.08333em{\bm{X}}_g^\top$.
Thus ${\bm{M}}_{gg}$ is the \th{g} diagonal block of the projection matrix
${\bm{M}}_{\bm{X}}$, which satisfies $\hat{\bm{u}} = {\bm{M}}_{\bm{X}}{\bm{u}}$, and
${\bm{M}}_{gg}^{-1/2}$ is its inverse symmetric square root. The CV$_2$
estimator reduces to the familiar HC$_2$ estimator when $G=N$. If the
variance matrix of every ${\bm{u}}_g$ were proportional to an identity
matrix, then CV$_{\tn2}$ would actually be unbiased \citep{PT_2018}.
The second alternative CRVE is
\begin{equation}
\mbox{CV$_{\tn3}$:}\qquad
\frac{G-1}{G}
({\bm{X}}^\top{\bm{X}})^{-1}\Big(\kern 0.08333em\sum_{g=1}^G \acute{\bm{s}}_g\acute{\bm{s}}_g^\top
\Big)({\bm{X}}^\top{\bm{X}})^{-1},
\label{eq:CV3}
\end{equation}
where $\acute{\bm{s}}_g = {\bm{X}}_g^\top{\bm{M}}_{gg}^{-1}\kern 0.08333em\hat{\bm{u}}_g$. As we
discuss in \Cref{subsec:leverage}, CV$_{\tn3}$ is actually a jackknife
estimator which generalizes the familiar HC$_3$ estimator of
\citet{MW_1985}.
As written in \eqref{eq:CV2} and \eqref{eq:CV3}, both CV$_{\tn2}$ and
CV$_{\tn3}$ are computationally infeasible for large samples, because
they involve the $N_g\times N_g$ matrices ${\bm{M}}_{gg}$. However,
\citet*{NAAMW_2020} proposes a more efficient algorithm for both of
them, and \citet*{MNW-influence} provides an even more efficient one
for CV$_{\tn3}$ by exploiting the fact that it is a jackknife
estimator; see \Cref{subsec:leverage}. Moreover, when one of the
regressors is a fixed-effect dummy for cluster $g$, the ${\bm{M}}_{gg}$
matrices are singular. This problem can be avoided, and some computer
time saved, by partialing out the fixed-effect dummies as discussed in
\Cref{subsec:leverage} and \citet{PT_2018}.
It seems plausible that both CV$_{\tn2}$ and CV$_{\tn3}$ should
perform better in finite samples than CV$_{\tn1}$, because the
modified score vectors $\grave{\bm{s}}_g$ and $\acute{\bm{s}}_g$ ought to
provide better approximations to the ${\bm{s}}_g$ than do
the~$\hat{\bm{s}}_g$. We would expect tests based on CV$_{\tn3}$ to be
more conservative than ones based on CV$_{\tn2}$, just as ones based
on HC$_3$ are more conservative than ones based on HC$_2$, because the
$\acute{\bm{s}}_g$ are ``shrunk'' more than the $\grave{\bm{s}}_g$. Simulation
evidence dating back to \citet{BM_2002} suggests that CV$_{\tn3}$
typically yields the most reliable tests, but that they can sometimes
under-reject; see also \citet*{MNW-bootknife}.
\section{Why and How to Cluster}
\label{sec:why}
We cannot hope to obtain reliable inferences when using clustered data
unless we know the actual clustering structure, at least to a good
approximation. Thus, before specifying any clustering structure, we
need to think about how intra-cluster correlations may arise and why
independence across clusters may, or may not, be plausible for that
structure.
\citet*{AAIW_2017} distinguishes between two alternative approaches to
inference, referred to as ``model-based'' and ``design-based.'' The
model-based approach is the traditional one, according to which every
sample is treated as a random outcome, or realization, from some
data-generating process (DGP), which in our context is the model
\eqref{eq:lrmodel}, and the objective is to draw inferences about
parameters in the DGP, which are interpreted as features of the
population. In particular, the DGP is the source of randomness and
hence an important determinant of the clustering structure.
The ``design-based'' approach to inference, which is analyzed in
detail in \citet{AAIW_2017}, is conceptually different from the
model-based framework. It involves thinking about the population,
estimand, sampling scheme, parameters of interest, and even the notion
of statistical uncertainty in a different manner. For example, according
to the design-based approach, statistical uncertainty does not derive
from a DGP, but rather is induced solely by the sampling uncertainty
coming from sampling from a fixed population \citep{AAIW_2020}.
Therefore, under this approach, the sampling process is an important
determinant of the clustering structure.
In some cases, the two approaches make use of similar inferential
procedures, but in others they employ quite different ones. Both
approaches may be informative about the choice of whether to cluster
and at which level, although their motivations may differ. We follow
most of the existing literature and focus exclusively on the model-based
approach in the remainder of this paper.
\subsection{Modeling Intra-Cluster Dependence}
\label{subsec:sources}
Intra-cluster correlations of the disturbances and regressors, and
hence of the scores, can arise for many reasons. By making the
assumptions in \eqref{eq:Sigma_g} and using cluster-robust inference,
we avoid the need to model these correlations. Nevertheless, it can be
illuminating to consider such models in order to learn about
intra-cluster dependence and their consequences. The simplest and most
popular model is the random-effects, or error-components, model
\begin{equation}
\label{eq:REmodel}
u_{gi} = \lambda\kern 0.04167em \varepsilon_g + \varepsilon_{gi},
\end{equation}
where $u_{gi}$ is the disturbance for observation $i$ within cluster
$g$, $\varepsilon_{gi} \sim {\rm iid}(0, \omega^2)$ is an
idiosyncratic shock for observation $i$, $\varepsilon_g \sim {\rm
iid}(0, 1)$ is a cluster-wide shock for cluster $g$, and the two
shocks are independent. The model \eqref{eq:REmodel} implies that the
variance matrix ${\bm{\Omega}}_g$ of the $u_{gi}$ for cluster $g$ has a very
simple form with diagonal elements equal to $\lambda^2 + \omega^2$
and off-diagonal elements equal to $\lambda^2$. Thus the disturbances
within every cluster are equi-correlated, with correlation coefficient
$\lambda^2/(\lambda^2 + \omega^2)$.
Although the random-effects model \eqref{eq:REmodel} has a long and
distinguished history, it is almost certainly too simple. As we
discuss in \Cref{subsec:FE}, it is very common in modern empirical
practice to include a set of cluster fixed effects among the regressors
in \eqref{eq:lrmodel}. In the case of \eqref{eq:REmodel}, these fixed
effects are simply estimates of the $\lambda\kern 0.04167em \varepsilon_g$, and by
including them we therefore remove all of the intra-cluster correlation.
Thus, if the random-effects model \eqref{eq:REmodel} were correct,
there would be no need to worry about cluster-robust inference whenever
the regressors included cluster fixed effects. Note also that inclusion
of cluster fixed effects usually comes at the price of larger standard
errors on the coefficients of interest.
In practice, however, it usually seems to be the case that we need
both cluster fixed effects and a CRVE; see
\Cref{subsec:FE,subsec:placebo}. This implies that whatever process is
generating the intra-cluster correlations must be more complicated
than \eqref{eq:REmodel}. A simple example is the (very standard)
factor model
\begin{equation}
\label{eq:factor}
u_{gi} = \lambda_{gi}\kern 0.08333em \varepsilon_g + \varepsilon_{gi},
\end{equation}
which differs from \eqref{eq:REmodel} in one important respect. The
effect of the cluster-wide shock $\varepsilon_g$ on $u_{gi}$ is given
not by a coefficient $\lambda$ but by a weight, or factor loading,
$\lambda_{gi}$. These factor loadings could be either fixed parameters
or random variables. They determine the extent to which observation
$i$ within cluster $g$ is affected by the cluster-wide shock
$\varepsilon_g$.
As an example, if the observations were for individual students, the
clusters denoted classrooms, and the outcome were student achievement,
then $\varepsilon_{gi}$ would measure unobserved student-specific
characteristics, $\varepsilon_g$ would measure unobserved teacher
quality (and perhaps other features of the class), and $\lambda_{gi}$
would measure the extent to which the disturbance term for student $i$
is affected by teacher quality. Clearly, the $\lambda_{gi}$ do not
need to be the same for all~$i$. Similar motivating examples based on
\eqref{eq:factor} can easily be given in many fields, including labor
economics, health economics, development economics, and financial
economics.
To verify that the factor model in \eqref{eq:factor} generates
dependence within clusters, it suffices to derive the second-order
moments of the $u_{gi}$. We find that ${\rm E} (u_{gi}) = 0$ and $\var
(u_{gi}) = \lambda_{gi}^2 + \omega^2$. The cluster dependence is
characterized by $\cov (u_{gi},u_{gj}) = \lambda_{gi} \lambda_{gj}$,
which differs across $(i,j)$ pairs and is zero only when the factor
loadings are zero. In the context of the classroom example, the
intra-cluster covariances would be zero only if the teacher had no
effect on student achievement. Moreover, the correlations would be
fully captured by classroom fixed effects if and only if
$\lambda_{gi}$ were the same for all~$i$.
The factor model in \eqref{eq:factor} is discussed in terms of the
disturbances $u_{gi}$ rather than the scores. There are at least two
simple cases in which the same model structure, and in particular the
same within-cluster correlation structure, applies to the scores. The
first is when a regressor is generated by a model similar to
\eqref{eq:factor}, but possibly with different parameters. The second
is when a regressor only varies at the cluster level, as is often the
case for dummy variables, especially treatment dummies.
The model \eqref{eq:factor} has only one clustering dimension, but the
idea does not apply exclusively to cross\kern 0.04167em-section data. For
example, if the observations also had a time dimension, we could
replace each of the $\varepsilon_g$ by a time\kern 0.04167em-series process at
the cluster level. This would yield a pattern where, within a cluster,
observations that were closer together in time would be more
correlated than observations that were further apart. In
\Cref{subsec:emp-placebo}, we generate placebo regressors in this way.
For panel data, it is possible that, in addition to correlation within
cross\kern 0.04167em-sectional units across time periods, there may be
correlation within time periods across cross\kern 0.04167em-sectional units. This
leads to two\kern 0.04167em-way clustering, which is discussed in
\Cref{subsec:twoway}.
In principle, we might be able to estimate a factor model like
\eqref{eq:factor} and use it to obtain feasible GLS estimates, as
mentioned at the end of \Cref{subsec:Model}. However, this would be
relatively complicated and rather arbitrary, since there are many
plausible ways in which the details of \eqref{eq:factor} could be
specified. Moreover, any factor model would necessarily impose far
stronger restrictions than the weak assumptions given in
\eqref{eq:Sigma_g}. Thus estimating any sort of factor model would
inevitably require far more effort, and surely result in much more
fragile inferences, than simply employing OLS estimation together with
a~CRVE.
\subsection{Do Cluster Fixed Effects Remove Intra-Cluster Dependence?}
\label{subsec:FE}
Investigators very often include fixed effects at the cluster level
among the regressors. There are generally good reasons for doing so.
The cluster fixed effects implicitly model a large number of possibly
omitted explanatory variables without assuming, implausibly, that the
omitted variables are uncorrelated with the included ones.
It is sometimes believed that fixed effects remove any within-cluster
dependence and hence eliminate the need to use a CRVE. However, as was
pointed out by \citet{Arellano_1987}, that is in fact only true under
very special circumstances. Including cluster fixed effects in any
regression model forces the intra-cluster sample average to be zero
for each cluster. In particular, including cluster fixed effects
transforms the factor model \eqref{eq:factor} into
\begin{equation}
\label{eq:FEscore}
u_{gi} - \bar{u}_g = ( \lambda_{gi} -\bar\lambda_g )
\varepsilon_g + ( \varepsilon_{gi} - \bar\varepsilon_g ),
\end{equation}
where the averages are taken across observations within each cluster,
so that, for example, $\bar{u}_g = N_g^{-1}\sum_{i=1}^{N_g}u_{gi}$. The
intra-cluster covariance for \eqref{eq:FEscore} is
\begin{equation}
\label{eq:FEcov}
\cov (u_{gi}- \bar{u}_g, u_{gj}-\bar{u}_g) =
( \lambda_{gi} -\bar\lambda_g )( \lambda_{gj}-\bar\lambda_g ),
\end{equation}
which is zero if and only if $\lambda_{gi}$ is the same for all~$i$.
In other words, the random-effects model \eqref{eq:REmodel} is the
\emph{only} model within the class of factor models \eqref{eq:factor}
for which including cluster fixed effects can remove all intra-cluster
dependence. Some dependence necessarily remains whenever there is any
variation in factor loadings across observations within clusters.
Furthermore, \eqref{eq:FEcov} strongly suggests that, whether or not a
regression model includes cluster fixed effects, the scores will tend
to be clustered whenever within-cluster dependence can be approximated
by a factor model like \eqref{eq:factor}. Including fixed effects will
almost always reduce the intra-cluster correlations, but rarely will it
entirely eliminate them. Because even very small intra-cluster
correlations can have a large effect on standard errors when the
clusters are large (see \eqref{eq:diffop} and the discussion that
follows) it generally seems unwise to assume that cluster fixed effects
make it unnecessary to use a~CRVE.
In view of these arguments, it has become quite standard in modern
empirical practice both to include cluster fixed effects (and perhaps
other fixed effects as well) and also to employ cluster-robust
inference. The empirical example in \Cref{sec:empirical} is typical in
these respects. Of course, cluster fixed effects cannot be included
when the regressor of interest is a treatment dummy and treatment is
at the cluster level, since the treatment dummy and the fixed effects
would be perfectly collinear. This problem does not arise for
difference\kern 0.04167em-in-differences (DiD) regressions,
because only some observations in the treated clusters are treated. In
recent empirical work with non-staggered adoption of treatment, the
regressions almost always include at least two sets of fixed effects,
one for time periods and one for cross-sectional units, with
clustering typically by the latter; see \Cref{subsec:placebo}.
\subsection{At What Level Should We Cluster?}
\label{subsec:level}
In many cases, there is more than one level at which we could cluster.
For example, with data on educational outcomes, we may be able to
cluster by classroom, by school, or perhaps by school district. With
data that are coded geographically, we may be able to cluster by
county, by state, or even by region. Choosing the right level at which
to cluster is not always easy, and choosing the wrong level can have
serious consequences.
Suppose, for concreteness, that there are two possible levels of
clustering, coarse and fine, with one or more fine clusters nested
within each of the coarse clusters. When there are $G$ coarse
clusters, the middle matrix in \eqref{eq:trueV} is $\sum_{g=1}^G
{\bm{\Sigma}}_g$. If each coarse cluster contains $M_g$ fine clusters
indexed by $h$, then ${\bm{\Sigma}}_g$ can be written as
\begin{equation}
\label{eq:midcoarse}
{\bm{\Sigma}}_g = \sum_{h_1=1}^{M_g} \sum_{h_2=1}^{M_g} {\bm{\Sigma}}_{g,h_1h_2},
\end{equation}
where ${\bm{\Sigma}}_{g,h_1h_2}$ denotes the covariance matrix of the scores
for fine clusters $h_1$ and $h_2$ within coarse cluster~$g$. Under the
assumption of fine clustering, ${\bm{\Sigma}}_{g,h_1h_2} = {\bm{\Sigma}}_{gh}$
when $h_1=h_2=h$ and ${\bm{\Sigma}}_{g,h_1h_2} = {\bm{0}}$ when $h_1\ne h_2$,
so that the middle matrix in \eqref{eq:trueV} reduces to
$\sum_{g=1}^G \sum_{h=1}^{M_g} {\bm{\Sigma}}_{gh}$.
From \eqref{eq:midcoarse}, the difference between the middle matrices
for coarse and fine clustering is
\begin{equation}
\label{eq:middiff}
\sum_{g=1}^G {\bm{\Sigma}}_g - \sum_{g=1}^G \sum_{h=1}^{M_g} {\bm{\Sigma}}_{gh} =
2 \sum_{g=1}^G \, \sum_{h_1=1}^{M_g}\,
\sum_{h_2=h_1+1}^{M_g}\!\!\!{\bm{\Sigma}}_{g,h_1h_2}.
\end{equation}
The finest possible level of clustering is no clustering at all. In
that case, the right-hand side of \eqref{eq:middiff} reduces to the
right-hand side of \eqref{eq:diffop}, because the fine clusters within
each coarse cluster are just the individual observations.
Under the assumption of fine clustering, the terms on the right-hand
side of \eqref{eq:middiff} are all equal to zero. Under the assumption
of coarse clustering, however, at least some of them are non-zero, and
\eqref{eq:middiff} must therefore be estimated. If we cluster at the
fine level when coarse clustering is appropriate, the CRVE is
inconsistent. On the other hand, if we cluster at the coarse level
when fine clustering is appropriate, the CRVE has to estimate
\eqref{eq:middiff} even though it is actually zero. This makes the
CRVE less efficient than it should be, leading to loss of power, or,
equivalently, to confidence intervals that are unnecessarily long,
especially when the number of coarse clusters is small.
Using simulation methods, \citet{MW-survey} investigates the
consequences on hypothesis tests of clustering at an incorrect level.
Clustering at too fine a level generally leads to serious
over-rejection, which becomes worse as the sample size increases with
the numbers of clusters at all levels held constant. This is exactly
what we would expect; see the discussion following \eqref{eq:diffop}.
Clustering at too coarse a level also leads to both some over-rejection
and some loss of power, especially when the number of clusters is small.
Two rules of thumb are commonly suggested for choosing the right
level of clustering. The simplest is just to cluster at the coarsest
feasible level \citep[Section~IV]{CM_2015}. This may be attractive
when the number of coarse clusters $G$ is reasonably large, but it can
be dangerous when $G$ is small, or when the clusters are heterogeneous
in size or other features; see \Cref{subsec:failure}.
A more conservative rule of thumb is to cluster at whatever level
yields the largest standard error(s) for the coefficient(s) of
interest \citep[Section~8.2]{MHE_2008}. This rule will often lead to
the same outcome as the first one, but not always. When $G$ is small,
cluster-robust standard errors tend to be too small, sometimes much
too small (\Cref{subsec:failure}). Hence, the second rule of thumb is
considerably less likely to lead to severe over-rejection than the
first one. However, because it is conservative, it can lead to loss of
power (or, equivalently, confidence intervals that are unnecessarily
long).
When the regressor of interest is a treatment dummy, and the level at
which treatment is assigned is known, then it generally makes sense
to cluster at that level \citep*{BDM_2004}. If treatment is assigned by
cluster, whether for all observations in each cluster or just for some
of them, as in the case of DiD models, then the scores will be
correlated within the treated clusters whenever there is any
intra-cluster correlation of the disturbances. Thus it never makes
sense to cluster at a level finer than the one at which treatment is
assigned. If we are certain that clusters are treated at random, then
it also does not make sense to cluster at a coarser level. However,
when there are two or more possible levels of clustering, it may not
be realistic to assume that treatments are independent across
finer-level clusters within the coarser-level ones. Unless we are
certain that this is actually the case, it may be safer to cluster at
a coarser level than the one at which treatment was supposedly
assigned. For example, it may make sense to cluster by school instead
of by classroom even when treatment was supposed to be assigned by
classroom.
Instead of using a rule of thumb, we can test for the correct level of
clustering. The best-known such test is an ingenious but indirect one
proposed in \citet{IM_2016}. It requires the model to be estimated
separately for every coarse cluster, something that is not possible
when the regressor of interest is invariant within some clusters, as
is typically the case for treatment models and DiD models. It is also
invalid if the parameter of interest has different meanings for
different clusters; see \Cref{subsec:small}. When it is valid, the
test statistic compares the observed variation of the estimates across
clusters with an estimate of what that variation would be if
clustering were actually at a finer level.
\citet*{MNW-testing} proposes direct tests called score\kern 0.04167em-variance
tests, which compare the variance of the scores for two nested levels
of clustering. For example, when there is just one coefficient, the
empirical analog of \eqref{eq:middiff} is a scalar that, when divided
by the square root of an estimate of its variance, is asymptotically
distributed as~${\rm N}(0,1)$. The null hypothesis is that the (true)
standard errors are the same for fine and coarse clustering, and the
(one\kern 0.04167em-sided) alternative is that they are larger for the latter
than for the former. \citet{MNW-testing} also proposes wild (cluster)
bootstrap implementations of these tests to improve their
finite\kern 0.04167em-sample properties.
\citet{Cai_2021} takes a different approach, proposing a test for the
level of clustering that is based on randomization inference
(\Cref{subsec:RI}). This test is designed for settings with a small
number of coarse clusters and a small number of fine clusters within
each of them.
It seems natural to cluster at the coarse level when a test rejects
the null hypothesis, and to cluster at the fine level when it does
not. However, choosing the level of clustering in this way is a form
of pre\kern 0.04167em-testing, which can lead to estimators with distributions
that are poorly approximated by asymptotic theory, even in large
samples \citep{LP_2005}. Using a pre\kern 0.04167em-test in this way will
inevitably lead to over-rejection when there is actually coarse
clustering, the standard errors for coarse clustering are larger than
the ones for fine clustering, and the test incorrectly fails to reject
the null hypothesis. Thus using such a test is less conservative than
relying on the second rule of thumb discussed above. On the other hand,
we may feel more comfortable with the second rule of thumb when it
agrees with the outcomes of one or more tests for the level of
clustering.
\subsection{Leverage and Influence}
\label{subsec:leverage}
As will be explained in \Cref{sec:asytheory}, asymptotic inference
depends on being able to apply laws of large numbers and central limit
theorems to functions of the (empirical) score vectors. How well those
theorems work depends on how homogeneous the score vectors are across
clusters. When they are quite heterogeneous, asymptotic inference may
be problematic; see \Cref{subsec:failure}. It is therefore desirable
to measure the extent of cluster-level heterogeneity.
Classic measures of observation-level heterogeneity are leverage and
influence \citep*{BKW_1980,CH_1986}. These are generalized to
cluster-level measures in \citet{MNW-influence}. One possible
consequence of heterogeneity is that the estimates may change a lot
when certain clusters are deleted. When this is the case, a cluster is
said to be influential. In order to identify individually influential
clusters, first construct the matrices ${\bm{X}}_g^\top{\bm{X}}_g$ and the
vectors ${\bm{X}}_g^\top{\bm{y}}_g$, for $g=1,\ldots,G$. Then
\begin{equation}
\label{eq:delone}
\hat{\bm\beta}^{(g)} = ({\bm{X}}^\top{\bm{X}} - {\bm{X}}_g^\top{\bm{X}}_g)^{-1}
({\bm{X}}^\top{\bm{y}} - {\bm{X}}_g^\top{\bm{y}}_g)
\end{equation}
is the vector of least squares estimators when cluster $g$ is deleted.
It should not be expensive to compute $\hat{\bm\beta}^{(g)}$ for every
cluster using~\eqref{eq:delone}. Note, however, that we cannot partial
out regressors other than cluster fixed effects (see below) prior to
computing the $\hat{\bm\beta}^{(g)}$, because the latter would then depend
indirectly on the observations for the \th{g} cluster.
When there is a parameter of particular interest, say~$\beta_j$, then
it will often be a good idea to report the $\hat\beta_j^{(g)}$ for
$g=1,\ldots,G$ in either a histogram or a table. If
$\hat\beta_j^{(h)}$ differs a lot from $\hat\beta_j$ for some cluster
$h$\kern -0.08333em, then cluster $h$ is evidently influential. In a few extreme
cases, there may be a cluster $h$ for which it is impossible to
compute $\hat\beta_j^{(h)}$\kern -0.08333em. If so, then the original estimates
should probably not be believed. This will happen, for example, when
cluster $h$ is the only treated one, and we will see in
\Cref{subsubsec:few} that inference is extremely unreliable in that
case.
The $\hat{\bm\beta}^{(g)}$ are of interest even when there is no reason to
expect any clusters to be influential. As \citet{MNW-bootknife} shows,
an alternative way to write CV$_{\tn3}$ is
\begin{equation}
\mbox{CV$_{\tn3}$:}\qquad
\frac{G-1}{G} \sum_{g=1}^G (\hat{\bm\beta}^{(g)} - \hat{\bm\beta})
(\hat{\bm\beta}^{(g)} - \hat{\bm\beta})^\top.
\label{eq:jackvar}
\end{equation}
This is the matrix version of the classic jackknife variance estimator
given in \citet{Efron_81} and others. Unless all clusters are very
small, \eqref{eq:jackvar} is enormously faster to compute than
\eqref{eq:CV3}. The \texttt{Stata} command \texttt{summclust}, which
is described in detail in \cite{MNW-influence}, calculates CV$_{\tn3}$
standard errors based on \eqref{eq:jackvar}.
As pointed out in \citet{BKW_1980} and \citet{CH_1986}, it is often
valuable to identify high-leverage observations as well as influential
ones. It is perhaps even more valuable to identify high-leverage
clusters \citep{MNW-influence}. Loosely speaking, a high-leverage
cluster is one whose regressors contain a lot of information. At the
observation level, high-leverage observations are associated with a
high value of $h_i$, the \th{i} diagonal element of ${\bm{H}} = {\bm{P}}_{\bm{X}}
= {\bm{X}} ({\bm{X}}^\top{\bm{X}})^{-1}{\bm{X}}^\top$. The analog of $h_i$ in the
cluster case is the $N_g\times N_g$ matrix ${\bm{H}}_g =
{\bm{X}}_g({\bm{X}}^\top{\bm{X}})^{-1} {\bm{X}}_g^\top$\kern -0.08333em. Since it is not feasible
to report the ${\bm{H}}_g$, we suggest that investigators instead report
their traces, which are
\begin{equation}
\label{eq:traceXX}
L_g = \Tr({\bm{H}}_g) =
\Tr\!\big({\bm{X}}_g^\top\!{\bm{X}}_g({\bm{X}}^\top\!{\bm{X}})^{-1}\big), \quad
g=1,\ldots,G.
\end{equation}
These are easy to compute because we have already calculated
$({\bm{X}}^\top{\bm{X}})^{-1}$ and the ${\bm{X}}_g^\top{\bm{X}}_g$. For any cluster
that contains just one observation, $L_g$ reduces to the usual measure
of leverage at the observation level. High-leverage clusters can be
identified by comparing the $L_g$ to their own average, which is
$k/G$. If, for some $h$, $L_h$ is substantially larger than $k/G$,
then cluster $h$ has high leverage. This can happen either because
$N_h$ is much larger than $G/N$ or because the matrix ${\bm{X}}_h$ is
somehow extreme relative to the other ${\bm{X}}_g$ matrices, or both. For
example, $L_h$ is likely to be much larger than $k/G$ if cluster $h$
is one of just a few treated clusters.
Regression models often include cluster fixed effects. It is
computationally attractive to partial them out before estimation
begins, using for example the \texttt{areg} procedure in
\texttt{Stata}. When one of the regressors is a fixed-effect dummy for
cluster $g$, the matrices ${\bm{X}}^\top{\bm{X}} - {\bm{X}}_g^\top{\bm{X}}_g$ are
singular. However, the problem solves itself if we partial out the
fixed-effect dummies and replace ${\bm{X}}$ by $\tilde{\bm{X}}$ and ${\bm{y}}$ by
$\tilde{\bm{y}}$, the matrix and vector of deviations from cluster means.
For example, the \th{gj} element of $\tilde{\bm{y}}$ is $y_{gj} -
N_g^{-1}\sum_{i=1}^{N_g} y_{gi}$. Since this depends only on
observations for cluster~$g$, the jackknife CV$_{\tn3}$ estimator
\eqref{eq:jackvar} remains valid.
In \Cref{sec:report}, we discuss what quantities investigators should
report in any empirical analysis that involves cluster-robust
inference. In addition to measures of influence and leverage, these
may include measures of partial leverage (the analog of leverage for a
single coefficient) and summary statistics based on either leverage,
partial leverage, or the effective number of clusters. An example is
provided in \Cref{subsec:emplever}.
\subsection{Placebo Regressions}
\label{subsec:placebo}
An interesting way to assess the validity of alternative standard
errors is to run ``placebo regressions.'' The idea, first suggested in
\citet{BDM_2004}, is to start with a model and dataset, then generate
a completely artificial regressor at random, add it to the model, and
perform a $t$-test of significance. This is repeated a large number of
times, and the rejection frequency is observed. The artificial
regressor is often a dummy variable that is referred to as a ``placebo
law'' or ``placebo treatment.'' Using such a dummy variable is natural
because, for any level of intra-cluster correlation of the
disturbances, the intra-cluster correlation of the scores is greatest
for regressors that do not vary within clusters. However, any
artificial regressor that is not completely uncorrelated within
clusters can potentially be used.
Because a placebo regressor is artificial, we would expect valid
significance tests at level~$\alpha$ to reject the null close to
$\alpha\%$ of the time when the experiment is repeated many times.
Following the lead of \citet{BDM_2004} by using models for
log-earnings based on age, education, and other personal
characteristics, together with data taken from the Current Population
Survey, several papers \citep*{JGM_2016,MW-JAE,Brewer_2018} find that
not clustering, or clustering at below the state level, leads to
rejection rates far greater than~$\alpha$. In
\Cref{subsec:emp-placebo}, we find similar results for the datasets
used in our empirical example. Our findings, and those of the papers
cited above, all suggest that using a state\kern 0.04167em-level CRVE is
important for survey data that samples individuals from multiple
states. If we fail to do so, we will find, with probability much
higher than $\alpha$, that nonsense regressors apparently belong in
the model.
Since the empirical score vectors are $\hat{\bm{s}}_g =
{\bm{X}}_g^\top\hat{\bm{u}}_g$, a placebo\kern 0.04167em-regressor experiment should lead
to over-rejection whenever both the regressor and the residuals
display intra-cluster correlation at a coarser level than the one at
which the standard errors are clustered. As in \Cref{subsec:level},
suppose there are two potential levels of clustering, fine and coarse,
with the fine clusters nested within the coarse clusters. If the
placebo regressor is clustered at the coarse level, we would expect
significance tests based on heteroskedasticity-robust standard errors
to over-reject whenever the residuals are clustered at either level.
Similarly, we would expect significance tests based on
finely-clustered standard errors to over-reject whenever the residuals
are clustered at the coarse level. \Cref{tab:placebo} in
\Cref{subsec:emp-placebo} displays both of these phenomena.
Placebo regressions can provide useful guidance as to the correct
level of clustering. However, using the rejection rates for placebo
regressions with different levels of clustering as informal tests is
really a form of pre\kern 0.04167em-testing. Thus, like using the formal tests
discussed in \Cref{subsec:level}, doing this seems very likely to
yield less conservative inferences than simply relying on the second
rule of thumb.
\subsection{Two\kern 0.08333em-\kern -0.08333em Way Clustering}
\label{subsec:twoway}
Up to this point, we have assumed that there is clustering in only one
dimension. However, there could well be clustering in two or more
dimensions. With data that have both a spatial and a
temporal dimension, there may be clustering by jurisdiction and also
by time period. In finance, there is often clustering by firm and
by year. Thus, instead of \eqref{eq:lrmodel}, we might have
\begin{equation}
\label{eq:modelgh}
{\bm{y}}_{gh} = {\bm{X}}_{gh} {\bm\beta} + {\bm{u}}_{gh},\quad g=1,\ldots,G,
\;\; h=1,\ldots,H,
\end{equation}
where the vectors ${\bm{y}}_{gh}$ and ${\bm{u}}_{gh}$ and the matrix
${\bm{X}}_{gh}$ contain, respectively, the rows of ${\bm{y}}$, ${\bm{u}}$, and
${\bm{X}}$ that correspond to both the \th{g} cluster in the first
clustering dimension and the \th{h} cluster in the second one. The
$GH$ clusters into which the data are divided in \eqref{eq:modelgh}
represent the intersection of the two clustering dimensions.
If there are $N_g$ observations in the \th{g} cluster for the first
dimension, $N_h$ observations in the \th{h} cluster for the second
dimension, and $N_{gh}$ observations in the \th{gh} cluster for the
intersection, the number of observations in the entire sample is
$N=\sum_{g=1}^G N_g=\sum_{h=1}^H N_h=\sum_{g=1}^G \sum_{h=1}^H
N_{gh}$, where $N_{gh}$ might equal 0 for some values of $g$ and $h$.
The scores for the clusters in the first dimension are
${\bm{s}}_g={\bm{X}}_g^\top{\bm{u}}_g$, for the clusters in the second dimension
${\bm{s}}_h={\bm{X}}_h^\top{\bm{u}}_h$, and for the intersections
${\bm{s}}_{gh}={\bm{X}}_{gh}^\top{\bm{u}}_{gh}$. If, by analogy with
\eqref{eq:Sigma_g}, we assume that
\begin{equation}
\label{eq:varmats}
{\bm{\Sigma}}_g = {\rm E}({\bm{s}}_g {\bm{s}}_g^\top), \;
{\bm{\Sigma}}_h = {\rm E}({\bm{s}}_h {\bm{s}}_h^\top), \;
{\bm{\Sigma}}_{gh} = {\rm E}({\bm{s}}_{gh} {\bm{s}}_{gh}^\top),\;
{\rm E}({\bm{s}}_{gh} {\bm{s}}_{g'h'}^\top)={\bm{0}} \; \text{for } g\neq g', h \neq h' ,
\end{equation}
then the variance matrix of the scores is seen to be
\begin{equation}
\label{eq:scorevar}
{\bm{\Sigma}} = \sum_{g=1}^G {\bm{\Sigma}}_g
+ \sum_{h=1}^H {\bm{\Sigma}}_h
- \sum_{g=1}^G\sum_{h=1}^H {\bm{\Sigma}}_{gh}.
\end{equation}
The last condition in \eqref{eq:varmats} means that the scores are
assumed to be independent whenever they do not share a cluster along
either dimension. The third term in \eqref{eq:scorevar} must be
subtracted in order to avoid double counting. It is important to
distinguish between two\kern 0.04167em-way clustering and clustering by the
intersection of the two dimensions. If we assumed the latter instead
of the former, then all three terms on the right-hand side of
\eqref{eq:scorevar} would be equal, and consequently ${\bm{\Sigma}} =
\sum_{g=1}^G \sum_{h=1}^H {\bm{\Sigma}}_{gh}$. Thus these assumptions are
radically different.
An estimator of the variance matrix of $\hat{\bm\beta}$ is
\begin{equation}
\label{eq:betavar}
\widehat\var(\hat{\bm\beta}) =
({\bm{X}}^\top{\bm{X}})^{-1} \hat{\bm{\Sigma}} ({\bm{X}}^\top{\bm{X}})^{-1}\kern -0.08333em, \quad
\hat{\bm{\Sigma}} = \sum_{g=1}^G \hat{\bm{s}}_g\hat{\bm{s}}_g^\top +
\sum_{h=1}^H \hat{\bm{s}}_h\hat{\bm{s}}_h^\top -
\sum_{g=1}^G\sum_{h=1}^H \hat{\bm{s}}_{gh}\hat{\bm{s}}_{gh}^\top.
\end{equation}
Here $\hat{\bm{\Sigma}}$ is an estimator of \eqref{eq:scorevar}, with the
empirical scores defined in the usual way; for example, $\hat{\bm{s}}_g =
{\bm{X}}_g^\top\hat{\bm{u}}_g$. In practice, each of the matrices on the
right-hand side of the second equation in \eqref{eq:betavar} is
usually multiplied by a scalar factor, like the one in \eqref{eq:CV1},
designed to correct for degrees of freedom. Because the third term is
subtracted, the matrix $\hat{\bm{\Sigma}}$ may not always be positive
definite. This problem can be avoided by omitting the third term,
which is asymptotically valid under some assumptions
\citep*{Davezies_2021,MNW_2021}. Another possibility is to use an
eigenvalue decomposition \citep*{CGM_2011}, although this merely
forces the variance matrix to be positive semidefinite.
The idea of two\kern 0.04167em-way clustering can, of course, be generalized to
three\kern 0.04167em-way clustering, four-way clustering, and so on. However, the
algebra rapidly becomes daunting. If there were three clustering
dimensions, for example, the analog of \eqref{eq:scorevar} would have
seven terms.
Two\kern 0.04167em-way clustering seems to have been suggested first in
\citet{MH_2006} and rediscovered independently by \citet*{CGM_2011}
and \citet{Thompson_2011}. Although two\kern 0.04167em-way clustering has been
widely used in empirical work, the asymptotic theory to justify it is
much more challenging than the theory for the one\kern 0.04167em-way case, and
this theory is still under active development
\citep*{Chiang_2020,Chiang_2021JBES,Davezies_2021,MNW_2021,Menzel_2021}.
In view of this, and because of the technical difficulties involved,
we will focus mainly on one\kern 0.04167em-way clustering in the remainder of the
paper.
\section{Asymptotic Inference}
\label{sec:asytheory}
For the regression model \eqref{eq:lrmodel}, inference is commonly based
on the $t$-statistic,
\begin{equation}
t_a = \frac{{\bm{a}}^\top(\hat{\bm\beta} -
{\bm\beta}_0)}{({\bm{a}}^\top\hat{\bm{V}}{\bm{a}})^{1/2}}\kern 0.08333em,
\label{eq:tstat}
\end{equation}
where the hypothesis to be tested is
${\bm{a}}^\top{\bm\beta}={\bm{a}}^\top{\bm\beta}_0$, with ${\bm{a}}$ a known $k$-vector.
Here $\hat{\bm{V}}$ denotes one of CV$_{\tn1}$, CV$_{\tn2}$, or
CV$_{\tn3}$, given in \eqref{eq:CV1}, \eqref{eq:CV2}, and either
\eqref{eq:CV3} or \eqref{eq:jackvar}, respectively. In many cases,
just one element of ${\bm{a}}$, say the \th{j}, equals~1, and the
remaining elements equal~0, so that \eqref{eq:tstat} is simply
$\hat\beta_j - \beta_{j0}$ divided by its standard error. When there
are $r>1$ linear restrictions, which can be written as ${\bm{R}}{\bm\beta} =
{\bm{r}}$ with ${\bm{R}}$ an $r\times k$ matrix, inference can be based on the
Wald statistic,
\begin{equation}
W = ({\bm{R}}\hat{\bm\beta} - {\bm{r}})^\top
({\bm{R}}\kern 0.04167em\hat{\bm{V}}\kern -0.08333em{\bm{R}}^\top)^{-1} ({\bm{R}}\hat{\bm\beta} - {\bm{r}}).
\label{Waldstat}
\end{equation}
Of course, when $r=1$, the $t$-statistic \eqref{eq:tstat} is just the
signed square root of a particular Wald statistic with
${\bm{R}}={\bm{a}}^\top$ and ${\bm{r}} = {\bm{a}}^\top{\bm\beta}_0$.
By letting the sample size become arbitrarily large, one can
frequently obtain a tractable asymptotic distribution for any test
statistic of interest, including \eqref{eq:tstat} and
\eqref{Waldstat}. Ideally, this would provide a good approximation to
the actual distribution. With clustered data, there is more than one
natural way to let the sample size become large, because we can make
various assumptions about what happens to $G$ and the $N_g$ as we let
$N$ tend to infinity. Which assumptions it is appropriate to use, and
how well the resulting approximations work, will depend on the
characteristics of the sample and the (unknown) DGP.
In order for inferences based on the statistics \eqref{eq:tstat} and
\eqref{Waldstat} to be asymptotically valid, two key asymptotic
results must hold. First, a central limit theorem (CLT) must apply to
the sum of the score vectors ${\bm{s}}_g$ in \eqref{eq:betahat}. In the
limit, after appropriate normalization, the vector $\sum_{g=1}^G
{\bm{s}}_g$ needs to follow a multivariate normal distribution with
variance matrix $\sum_{g=1}^G {\bm{\Sigma}}_g$. Second, again after
appropriate normalization, a law of large numbers (LLN) must apply to
the matrices $\sum_{g=1}^G \hat{\bm{s}}_g\hat{\bm{s}}_g^\top$, $\sum_{g=1}^G
\grave{\bm{s}}_g\grave{\bm{s}}_g^\top$, or $\sum_{g=1}^G
\acute{\bm{s}}_g\acute{\bm{s}}_g^\top$ in the middle of the variance matrix
estimators \eqref{eq:CV1}, \eqref{eq:CV2}, or \eqref{eq:CV3}, so that
they converge to $\sum_{g=1}^G {\bm{\Sigma}}_g$. We refer to ``appropriate
normalization'' here rather than specifying the normalization factors
explicitly because, with clustered data, the issue of normalization is
a very tricky one; see \Cref{subsec:large}. For asymptotic inference
to be reliable, we need both the CLT and the LLN to provide good
approximations.
There are currently two quite different types of assumptions on which
the asymptotic theory of cluster-robust inference can be based. The
most common approach, and we believe usually the most appropriate one,
is to let the number of clusters tend to infinity. We refer to this as
the ``large number of clusters'' approach and discuss it in
\Cref{subsec:large}. An alternative approach is to hold the number of
clusters fixed and let the number of observations within each cluster
tend to infinity. We refer to this as the fixed-$G$ or ``small number
of large clusters'' approach and discuss it in \Cref{subsec:small}.
Some of the material in \Cref{subsec:large,subsec:small} is quite
technical, but it helps to explain when and why asymptotic inference
can fail.
Inference based on asymptotic theory often performs well, but it can
perform poorly in some commonly-encountered situations that are
discussed in \Cref{subsec:failure}. We therefore do not recommend
relying only on CV$_{\tn1}$ and asymptotic theory. Because the
bootstrap methods to be discussed in \Cref{sec:bootstrap} can work
much better than asymptotic methods when the latter do not work well,
we recommend that they be used almost all the time, at least to verify
that both approaches yield similar results. In particular, we
recommend using one or more variants of the wild cluster restricted,
or WCR, bootstrap (\Cref{subsec:wild}) as a matter of routine.
\subsection{Asymptotic Theory: Large Number of Clusters}
\label{subsec:large}
The simplest assumption about how the sample size goes to infinity is
that every cluster has a fixed number of observations, say $M$\kern -0.08333em.
Then $N=MG$, and both $N$ and $G$ go to infinity at the same rate.
Thus the appropriate normalizing factor for the parameter estimator is
either $\sqrt{G}$ or $\sqrt{N}$. In this case, it is not difficult to
show that $\sqrt{G}(\hat{\bm\beta} - {\bm\beta}_0)$ is asymptotically
multivariate normal with variance matrix equal to the probability
limit of $G$ times the right-hand side of \eqref{eq:trueV}. Moreover,
the latter can be estimated consistently by $G$ times the CV$_{\tn1}$,
CV$_{\tn2}$, or CV$_{\tn3}$ matrices. The first proof for this case of
which we are aware is in \citet[Chapter~6]{White_1984}; see also
\citet{Hansen_2007}.
In actual samples, clusters often vary greatly in size, so it is
usually untenable to assume that every cluster has the same number of
observations. The assumption that $G$ is proportional to $N$ may be
relaxed by allowing $G$ to be only approximately proportional to
$N$\kern -0.08333em, so that $G/N$ is roughly constant as $N\to\infty$. This
implies that all the clusters must be small. In this case, the quality
of the asymptotic approximations is not likely to be harmed much by
moderate variation in cluster sizes. If a sample has, say, 500
clusters that vary in size from 10 to 50 observations, we would expect
asymptotic inference to perform well unless there is some other reason
(unrelated to cluster sizes) for it to fail.
\citet*{DMN_2019} and \citet{HansenLee_2019} take a more flexible
approach, with primitive conditions that restrict the variation in the
$N_g$ relative to the sample size. These conditions allow some
clusters to be ``small'' and others to be ``large'' in the sense that
some but not all $N_g \to \infty$ as $N \to \infty$. Although a key
assumption is that $G\to\infty$ (i.e., is ``large''), the appropriate
normalization factor for $\hat{\bm\beta} - {\bm\beta}_0$ is usually not
$\sqrt{G}$. Instead, this factor depends in a complicated way on the
regressors, the relative cluster sizes, the intra-cluster correlation
structure, and interactions among these; some examples of different
normalizing factors are given in the papers cited above. For this
reason, the key result that the $t$-statistic defined in
\eqref{eq:tstat} is asymptotically distributed as standard normal is
derived assuming that the rate at which $\hat{\bm\beta} - {\bm\beta}_0$ tends
to zero is unknown. Of course, this result also justifies using the
$t(G-1)$ distribution, which is more conservative, is widely used, and
is derived from the theory discussed in \Cref{subsec:small}.
The application of a CLT to $\sum_{g=1}^G {\bm{s}}_g$, appropriately
normalized, requires a restriction on the amount of heterogeneity that
is allowed. Otherwise, just a few clusters might dominate the entire
sample in the limit, thus violating the Lindeberg or Lyapunov
conditions for the CLT. The necessary restrictions on the
heterogeneity of clusters may be expressed in terms of two key
parameters. The first of these parameters is the number of moments
that is assumed to exist for the distributions of ${\bm{s}}_{gi}$
(uniformly in $g$ and~$i$). We denote this parameter by $\gamma > 2$.
When more moments exist, the distributions of ${\bm{s}}_{gi}$ are closer
to the normal distribution, and hence the sample will feature fewer
outliers or other highly leveraged observations or clusters.
In the clustered regression model, the variance of the scores, which
is often referred to as the Fisher information matrix, is given by
$\mathcal{J}_N = \sum_{g=1}^G \var ({\bm{s}}_g )$. When appropriately
normalized, $\mathcal{J}_N$ converges to a nonzero and finite matrix
$\mathcal{J}$\kern -0.08333em. The rate of convergence $\eta_N$ is defined
implicitly by $\eta_N^{-1} \mathcal{J}_N \to \mathcal{J}$\kern -0.08333em. This
rate is the second key parameter. One interpretation of $\eta_N$ can
be found in \eqref{eq:trueV}, from which the stochastic order of
magnitude of $\hat{\bm\beta} - {\bm\beta}_0$ is seen to be
$O_P(\eta_N^{1/2}\!/N)$. In general, $\eta_N \geq N$, with the
equality holding whenever there is no intra-cluster correlation. The
larger the value of $\eta_N$, the more slowly does $\hat{\bm\beta}$
converge to ${\bm\beta}_0$.
The conditions required on the heterogeneity of clusters to apply a
CLT can be stated in terms of the parameters $\gamma$ and $\eta_N$.
Specifically, when expressed in our notation, Assumption~3 of
\citet{DMN_2019} states the following condition:
\begin{equation}
\label{RateCondition}
\Bigl(\frac{\eta_N^{1/2}}{N}\Bigr)^{\!\frac{-2\gamma}{2\gamma-2}}
\,\frac{\sup_g N_g}{N} \longrightarrow 0.
\end{equation}
Because $\eta_N = o ( N^2)$ for consistency of $\hat{\bm\beta}$, the
condition in \eqref{RateCondition} makes it clear that we cannot allow
a single cluster to dominate the sample, in the sense that its size is
proportional to~$N$\kern -0.08333em. More generally, \eqref{RateCondition} shows
that there is a tradeoff between information accumulation and
variation in cluster sizes, as measured by the largest cluster size.
To interpret this tradeoff, we will consider three different
implications of \eqref{RateCondition}.
First, when $\gamma$ increases, so that more moments exist, condition
\eqref{RateCondition} becomes less strong. In particular, when the
scores are nearly normally distributed, in the sense that all their
moments exist, then $\gamma=\infty$, which implies that
$-2\gamma/(2\gamma-2) = -1$. In this case, \eqref{RateCondition}
reduces to $\eta_N^{-1/2} \sup_g N_g \to 0$, so that the size of the
largest cluster must increase more slowly than the square root of the
rate at which the Fisher information matrix converges. When
$\gamma<\infty$, so that there are fewer moments, then the rate at
which $\sup_g N_g$ is allowed to increase becomes smaller.
Second, suppose that the scores are uncorrelated, or more generally
that a CLT applies to $N_g^{-1/2}{\bm{s}}_g$, as assumed in
\citet*{BCH_2011} and \citet{IM_2010,IM_2016}; see
\Cref{subsec:small}. In this case, $\var ({\bm{s}}_g)= O(N_g)$, so that
$\eta_N = N$. Condition \eqref{RateCondition} is then
$N^{-(\gamma-2)/(2\gamma-2)} \sup_g N_g \to 0$. If the scores are
nearly normal, then it reduces further to $N^{-1/2} \sup_g N_g \to 0$,
so that the size of the largest cluster must increase no faster than
the square root of the sample size. Once again, the fewer moments
there are, the more slowly $\sup_g N_g$ is allowed to increase.
Third, suppose that the scores are generated by the factor model in
\eqref{eq:factor}, or by the simpler random-effects model
\eqref{eq:REmodel}. Then $\var ({\bm{s}}_g)=O(N_g^2)$. If, in addition,
$\inf_g N_g$ and $\sup_g N_g$ are of the same order of magnitude, then
$\eta_N = N \sup_g N_g$ and the condition \eqref{RateCondition}
collapses to $N^{-1} \sup_g N_g \to~0$, regardless of the number of
finite moments.
One possibly surprising implication of the above discussion is that,
when there is more intra-cluster correlation, so that $\eta_N$ is
relatively large, then greater heterogeneity of cluster sizes is
allowed. That is, a higher degree of intra-cluster correlation implies
a faster rate of convergence, $\eta_N$, of the Fisher information
matrix, which in turn allows a larger $\sup_g N_g$ in
\eqref{RateCondition}. The intuition is that greater intra-cluster
correlation reduces the effective cluster size, as measured by the
amount of independent information a cluster contains. In the extreme
case in which all observations in the \th{g} cluster are perfectly
correlated, the size of the cluster is effectively~1 and not $N_g$.
Note, however, that large clusters are implicitly weighted more
heavily than small clusters even in this extreme case.
Although it is impossible to verify condition \eqref{RateCondition} in
any finite sample, investigators can always observe the $N_g$. The
discussion of \eqref{RateCondition} above suggests that asymptotic
inference tends to be unreliable when the $N_g$ are highly variable,
especially when a very few clusters are unusually large. This is
exacerbated when the distribution of the data is heavy-tailed (has
fewer moments), but is mitigated when the clusters have an
approximate factor structure. Circumstances in which asymptotic
inference can be unreliable are discussed in more detail in
\Cref{subsec:failure}.
\subsection{Asymptotic Theory: Small Number of Large Clusters}
\label{subsec:small}
A few authors have assumed that $G$ remains fixed (i.e., is ``small'')
as $N\to\infty$, while the cluster sizes diverge (i.e., are
``large''). Notably, \citet{BCH_2011} proves that, for CV$_{\tn1}$,
the $t$-statistic \eqref{eq:tstat} follows the $t(G-1)$ distribution
asymptotically. An analogous result for the Wald statistic
\eqref{Waldstat} is discussed in \Cref{subsubsec:several}. These
results provide useful approximations, but they are proven under some
very strong assumptions. In particular, all the clusters are assumed
to be the same size $M$\kern -0.08333em. In addition, the pattern of dependence
within each cluster is assumed to be such that a CLT applies to the
normalized score vectors $M^{-1/2}{\bm{s}}_g$ for all $g=1,\ldots,G$, as
$M\to\infty$.
This second assumption is crucial, as it limits the amount of
dependence within each cluster and requires it to diminish quite
rapidly as $M\to\infty$. Although \citet{BCH_2011} discusses a
particular model for which this requirement holds, it rules out simple
DGPs such as the factor model \eqref{eq:factor}, with or without
cluster fixed effects. It even rules out the random-effects model
\eqref{eq:REmodel}, which is the most common model of intra-cluster
correlation. For all these models, no CLT can possibly apply to the
vector~$M^{-1/2} {\bm{s}}_g$. Consider, for example, the factor model
\eqref{eq:factor}, and suppose that ${\bm{x}}_g=1$. In this case,
\begin{equation}
\label{clt:variance}
\var ( M^{-1/2} {\bm{s}}_g ) = \frac1M\!\sum_{i,j=1}^{M}\!\cov (u_{gi},u_{gj})
= \frac1M\sum_{i=1}^{M} \lambda_{gi}^2
+ \frac2M\sum_{i=1}^{M} \sum_{j=i+1}^{M}\!\lambda_{gi}\lambda_{gj}.
\end{equation}
Because of the double summation, the second term on the right-hand side
of \eqref{clt:variance} clearly does not converge as $M \to \infty$
unless additional, and very strong, assumptions are made.
Another, quite different, approach to inference when $G$ is fixed is
developed in \citet{IM_2010}. The parameter of interest is a scalar,
say $\beta$, which can be thought of as one element of ${\bm\beta}$. The
key idea is to estimate $\beta$ separately for each of the $G$
clusters. This yields estimators $\hat\beta_g$ for $g=1,\ldots,G$.
Inference is then based on the average, say $\bar\beta$, and standard
error, say $s_{\kern -0.08333em\hat\beta}$, of the $\hat\beta_g$. \citet{IM_2010}
shows that the test statistic $\sqrt{G}(\bar\beta-\beta_0)
/s_{\kern -0.08333em\hat\beta}$ is approximately distributed as $t(G-1)$ when all
clusters are large and a CLT applies to $N_g^{-1/2} {\bm{s}}_g$ for
each~$g$. However, as we saw above, this assumption cannot hold even
for the simple random-effects or factor models.
A practical problem with this procedure is that $\beta$ may not be
estimable for at least some clusters. For models of treatment effects
at the cluster level, this will actually be the case for every
cluster. For difference\kern 0.04167em-in-differences (DiD) models with
clustering at the jurisdiction level, it will be the case for every
jurisdiction that is never treated. \citet{IM_2016} suggests a way to
surmount this problem by combining clusters into larger ones that
allow $\beta$ to be estimated for each of them. Even when $\beta$
itself can be estimated for each cluster, however, the full model may
not be estimable. This can happen, for example, when there are fixed
effects for categorical variables, and not all categories occur in
each cluster. In such cases, the interpretation of $\beta$ may differ
across clusters.
Although the estimators and test statistics proposed in
\citet{IM_2010,IM_2016} differ from the more conventional ones studied
in \citet{BCH_2011}, both approaches lead to $t$-statistics that
follow the $t(G-1)$ distribution asymptotically. This distribution has
in fact been used in \texttt{Stata} as the default for
CV$_{\tn1}$-based inference for many years. For small values of $G$,
using it can lead to noticeably more accurate, and more conservative,
inferences than using the $t(N-k)$ or normal distributions. However,
as we discuss in the next subsection and in \Cref{sec:bootstrap},
inferences based on the $t(G-1)$ distribution are often not nearly
conservative enough, especially when $G$ is small.
\subsection{When Asymptotic Inference Can Fail}
\label{subsec:failure}
Whenever we rely on asymptotic theory, we need to be careful. What is
true for infinitely large samples may or may not provide a good
approximation for any actual sample. Unless the very strong
assumptions discussed in \Cref{subsec:small} are satisfied, we cannot
expect to obtain reliable inferences when $G$ is small.
Unfortunately, there is no magic number for $G$ above which asymptotic
inference can be relied upon. It is sometimes claimed that asymptotic
inference based on CV$_{\tn1}$ is reliable when $G \geq 50$, or even
(partly in jest) when $G \geq 42$ \citep{MHE_2008}, but this is not
true. In very favorable cases, inference based on CV$_{\tn1}$ and the
$t(G-1)$ distribution can be fairly reliable when $G=20$, but in
unfavorable ones it can be unreliable even when $G=200$ or more; see
\Cref{subsubsec:het,subsubsec:few}. Moreover, there is evidence
\citep{BM_2002,MNW-bootknife} that inference based on CV$_{\tn3}$
tends to be more reliable, sometimes much more reliable, than
inference based on other CRVEs.
\subsubsection{Cluster Heterogeneity}
\label{subsubsec:het}
What determines whether a case is favorable or unfavorable for a given
$G$ is mostly the heterogeneity of the cluster score vectors.
Unfortunately, the latter cannot be observed directly. We can observe
the empirical score vectors, but they can sometimes differ greatly
from the true ones (\Cref{subsubsec:few}). We can also observe cluster
sizes, which the discussion in \Cref{subsec:large} focused on. These
are often particularly important, but any form of heterogeneity can
have serious consequences. This includes both heteroskedasticity of
the disturbances at the cluster level and systematic variation across
clusters in the distribution of the regressors. In general, the number
of clusters $G$ and the extent to which the distribution of the scores
varies across clusters will determine the quality of the asymptotic
approximation.
In principle, a poor asymptotic approximation could lead $t$-tests
based on the $t(G-1)$ distribution either to under-reject or
over-reject. We have never observed $t$-tests based on CV$_{\tn1}$ or
CV$_{\tn2}$ to under-reject in any simulation experiments, but we have
observed ones based on CV$_{\tn3}$ to do so. Unless $G$ is fairly
large, it is not difficult to find cases in which \mbox{$t$-tests}
based on any of these CRVEs reject much more than their nominal level.
Wald tests of several restrictions typically perform even worse; see
\Cref{subsubsec:several}.
As we discussed in \Cref{subsec:large}, the condition
\eqref{RateCondition} imposes a restriction on the size of the largest
cluster relative to the sample size. Thus, the quality of the
asymptotic approximation will surely diminish as the size of the
largest cluster increases relative to the average cluster size, and
over-rejection will consequently increase. This conjecture is
supported by simulation evidence in \citet{MW-JAE} and
\citet{DMN_2019}, as well as by analytic results based on Edgeworth
expansions in the latter paper.
There are at least two situations in which cluster-robust $t$-tests
and Wald tests are at risk of over-rejecting to an extreme extent,
namely, when one or a few clusters are unusually large, or when only a
few clusters are treated. In both of these cases, one cluster, or just
a few of them, have high leverage, in the sense that omitting one of
these clusters has the potential to change the OLS estimates
substantially; see \Cref{subsec:leverage}. Since both of these
situations can occur even when $G$ is not small, all users of
cluster-robust inference need to be on guard for them.
The first case in which conventional inference fails is when one or a
very few clusters are much larger than the others. This implies that
the distributions of the score vectors for those clusters are much
more spread out than the ones for the rest of the clusters. An extreme
example is studied in \citet[Figure~3]{DMN_2019}. When half the sample
is in one large cluster, rejection rates for $t$-tests based on
CV$_{\tn1}$ actually increase as $G$ increases, approaching 50\% at
the 5\% level for $G=201$. Unfortunately, this extreme case is
empirically relevant. Because roughly half of all incorporations in
the United States are in Delaware, empirical studies of state laws and
corporate governance encounter precisely this situation whenever they
cluster at the state level \citep{HS_2020}. \citet{MNW-bootknife}
studies a case with cluster sizes proportional to incorporations,
where almost 53\% of the observations are in the largest cluster.
Although $t$-tests based on all three CRVEs over-reject, those based
on CV$_{\tn3}$ do so much less severely than ones based on CV$_{\tn1}$
and CV$_{\tn2}$.
Not all forms of heterogeneity are harmful. In particular, having some
extremely small clusters in a sample generally does not cause any
problems, so long as there is not too much heterogeneity in the
remainder of the sample. For example, suppose that a sample consists
of, say, 25 large clusters, each with roughly 200 observations, and 15
tiny clusters, each with just one or a handful of observations. Except
in very unusual cases, the coefficient estimates and their
$t$-statistics would hardly change if we were to drop the tiny
clusters, so this sample is better thought of as having 25 equal-sized
clusters. The asymptotic approximations would perform just about the
same whether or not the tiny clusters were included. Of course, if we
changed this example so that there were 5 large clusters and 15 tiny
ones, then asymptotic inference would surely be very problematic,
because there would effectively be just 5 clusters.
\subsubsection{Treatment and Few Treated Clusters}
\label{subsubsec:few}
The second case in which conventional inference fails is when the
regressor of interest is a treatment dummy, and treatment occurs only
for observations in a small number of clusters. In such cases, the
empirical score vectors for the treated clusters, even when they have
been modified, can provide very poor estimates of the actual score
vectors.
Suppose that $d_{gi}$ is the value of the treatment dummy for
observation $i$ in cluster $g$, and let $s^d_g$ denote the element of
${\bm{s}}_g$ corresponding to the dummy. Consider first the extreme case
in which only some or all of the observations in the first cluster are
treated. Then $s^d_g = \sum_{i=1}^{N_g} d_{gi} u_{gi}$ is equal to
$\sum_{i=1}^{N_1} d_{1i} u_{1i}$ for $g=1$ and to~0 for all $g \neq
1$. Thus the scores corresponding to the treatment dummy equal zero
for the control clusters. Moreover, because the treatment regressor
must be orthogonal to the residuals, the empirical score $\hat s^d_1 =
0$. Since the actual score $s^d_1 \neq 0$, this implies that
\eqref{eq:CV1} provides a dreadful estimate of \eqref{eq:trueV}, at
least for the elements corresponding to the coefficient on the
treatment dummy. In consequence, the CV$_{\tn1}$ standard error of
this coefficient can easily be too small by a factor of five or more.
When more than one cluster is treated, the problem is not as severe,
because the $\hat s^d_g$ now sum to zero over the observations in all
the treated clusters. This causes them to be too small, but not to the
same extent as when just one cluster is treated; see
\citet{MW-JAE,MW-EJ}.
How well the empirical scores mimic the actual scores depends on the
sizes of the treated and control clusters, the values of other
regressors, and the number of treated observations within the treated
clusters. Thus all these things affect the accuracy of cluster-robust
standard errors and the extent to which $t$-statistics based on them
over-reject. As the number of treated clusters, say $G_1$, increases,
the problem often goes away fairly rapidly. But increasing $G$ when
$G_1$ is small and fixed does not help and may cause over-rejection to
increase. For models where all observations in each cluster are either
treated or not, having very few control clusters is just as bad as
having very few treated clusters. The situation is more complicated
for DiD models, however; see \citet{MW-TPM}.
It seems plausible that the modified empirical score vectors used by
CV$_{\tn2}$ and CV$_{\tn3}$ may mimic the actual score vectors more
accurately than the unmodified ones used by CV$_{\tn1}$. In fact,
there is some evidence \citep{MNW-bootknife} that $t$-tests based on
CV$_{\tn2}$ over-reject less severely than ones based on CV$_{\tn1}$
when there are few treated clusters, and that $t$-tests based on
CV$_{\tn3}$ over-reject considerably less severely. However, even
though CV$_{\tn3}$ can perform much better than CV$_{\tn1}$, it
still tends to over-reject when $G_1$ is very small.
\subsubsection{Testing Several Restrictions}
\label{subsubsec:several}
Most of the literature on cluster-robust inference has focused on
$t$-tests, but the cluster-robust Wald tests defined in
\eqref{Waldstat} also generally over-reject in finite samples. In
fact, they tend to do so more severely as $r$, the number of
restrictions, increases; see \citet{PT_2018} and \citet{JGM-fast}. As
is well known, this phenomenon occurs for Wald tests of all kinds. The
problem may well be unusually severe in this case, however, because
all CRVEs have rank at most $G$ and, in many cases, only $G-1$. On the
other hand, although the true variance matrix $\var(\hat{\bm\beta})$ is
unknown in finite samples, it will normally have full rank~$k$. As $r$
increases, the inverse of ${\bm{R}}^\top\hat{\bm{V}}{\bm{R}}$ thus seems likely to
provide an increasingly poor approximation to the inverse of
${\bm{R}}^\top\!\var(\hat{\bm\beta}){\bm{R}}$\kern 0.04167em.
\citet{BCH_2011} proves a result that helps to mitigate this problem.
Under the very strong assumptions discussed in \Cref{subsec:small},
where $G$ is fixed, every cluster is the same size $M$, and the amount
of within-cluster dependence is limited, the paper shows that
$r(G-1)/(G-r)$ times the appropriate quantile of the $F(r,G-r)$
distribution provides an asymptotic critical value for the Wald
statistic \eqref{Waldstat} based on CV$_{\tn1}$. Equivalently,
$(G-r)/(r(G-1))$ times $W$ is shown to follow the $F(r,G-r)$
distribution asymptotically in this special case. However, the WCR
bootstrap (\Cref{subsec:wild}) can provide a much better
approximation.
\subsection{Cluster-Robust Inference in Nonlinear Models}
\label{subsec:nonlinear}
Although cluster-robust inference is most commonly used with the
linear regression model \eqref{eq:lrmodel}, it can actually be
employed for a wide variety of models estimated by maximum likelihood
or the generalized method of moments (GMM); see
\citet{HansenLee_2019}.
Consider a model characterized by the log-likelihood function
\begin{equation}
\ell({\bm{\theta}}) = \sum_{g=1}^G \sum_{i=1}^{N_g} \ell_{gi}({\bm{\theta}}),
\label{logl}
\end{equation}
where ${\bm{\theta}}$ is the $k\times 1$ parameter vector to be estimated,
and $\ell_{gi}({\bm{\theta}})$ denotes the contribution to the log-likelihood
made by the \th{i} observation within the \th{g} cluster. Let
$\hat{\bm{\theta}}$ denote the vector that maximizes \eqref{logl},
${\bm{s}}_{gi}({\bm{\theta}})$ the $k\times 1$ vector of the first derivatives
of $\ell_{gi}({\bm{\theta}})$ (that is, the score vector), and
${\bm{H}}_{gi}({\bm{\theta}})$ the $k\times k$ Hessian matrix of the second
derivatives. Further, let $\hat{\bm{s}}_g = \sum_{i=1}^{N_g}
{\bm{s}}_{gi}(\hat{\bm{\theta}})$ and $\hat{\bm{H}} = \sum_{g=1}^G \sum_{i=1}^{N_g}
{\bm{H}}_{gi}(\hat{\bm{\theta}})$. Then \citet[Theorem~10]{HansenLee_2019} shows
(using somewhat different notation) that the cluster-robust variance
estimator for the maximum likelihood estimator $\hat{\bm{\theta}}$ is
\begin{equation}
\widehat\var(\hat{\bm{\theta}} - {\bm{\theta}}_0) =
\hat{\bm{H}}^{-1} \bigg(\sum_{g=1}^G \hat{\bm{s}}_g \hat{\bm{s}}_g^\top\!\bigg)
\kern 0.08333em\hat{\bm{H}}^{-1}.
\label{mlvar}
\end{equation}
The resemblance between \eqref{mlvar} and the CV$_{\tn1}$ variance
matrix in \eqref{eq:CV1} is striking. Indeed, since the Hessian is
proportional to ${\bm{X}}^\top{\bm{X}}$ for the linear regression model,
CV$_{\tn1}$ without the leading scalar factor is really just a special
case of \eqref{mlvar}.
The variance matrix estimator \eqref{mlvar} can be used for a wide
variety of models estimated by maximum likelihood. In fact,
\texttt{Stata} has been using it for various models, including probit
and logit, for some years. \citet[Theorem~12]{HansenLee_2019} provides
a similar result for GMM estimation, which is also very widely
applicable. More recently, the fixed-$G$ approach discussed in
\Cref{subsec:small} has been applied to GMM estimation by
\citet{Hwang_2021}. It leads to a novel inferential procedure that
involves modifying the usual asymptotic $t$ and $F$ statistics, but it
requires that cluster sizes be approximately equal.
Unfortunately, very little currently seems to be known about the
finite\kern 0.04167em-sample properties of tests based on \eqref{mlvar} or its
GMM analog. They are probably worse than those of tests based on
\eqref{Waldstat}. It seems quite plausible that bootstrapping, to
which we turn in the next section, would help, and \citet{Hwang_2021}
discusses one bootstrap method for GMM estimation. However,
bootstrapping nonlinear models tends to be computationally expensive,
and the properties of applicable bootstrap procedures are largely
unknown at the present time.
\section{Bootstrap Inference}
\label{sec:bootstrap}
Instead of basing inference on an asymptotic approximation to the
distribution of a statistic of interest, it is often more reliable to
base it on a bootstrap approximation. In \Cref{subsec:bootstrap}, we
briefly review some key concepts of bootstrap testing and bootstrap
confidence intervals. Then, in \Cref{subsec:pairs,subsec:wild}, we
discuss bootstrap methods for regression models with clustered data.
These methods, in particular the wild cluster restricted (WCR)
bootstrap to be discussed in \Cref{subsec:wild}, can be surprisingly
inexpensive to compute and often lead to much more reliable inferences
than asymptotic procedures. We therefore recommend that at least one
variant of the WCR bootstrap be used almost all the time.
\subsection{General Principles of the Bootstrap}
\label{subsec:bootstrap}
Suppose we are interested in a test statistic $\tau$, which might be a
$t$-statistic or a Wald statistic. Instead of using $P$~values or
critical values taken from an asymptotic distribution, we can use ones
from the empirical distribution function (EDF) of a large number of
bootstrap test statistics. This EDF often provides a good
approximation to the unknown distribution of~$\tau$. In order to
obtain the EDF, we need to generate $B$ bootstrap samples and use each
of them to compute a bootstrap test statistic, say $\tau^*_b$, for
$b=1,\ldots,B$.
Precisely how the bootstrap samples are generated is critical, and we
will discuss some methods for doing so in the next two subsections.
The choice of $B$ also matters. Ideally, it should be reasonably large
\citep{DM_2000} and satisfy the condition that $\alpha(B+1)$ is an
integer for any $\alpha$ (the level of the test) that may be of
interest \citep{RM_2007}. In general, the computational cost of
generating the bootstrap test statistics is proportional to $B$ times
$N$\kern -0.08333em, so that bootstrapping can be expensive. However, as we discuss
in \Cref{subsec:wild,subsec:pairs}, surprisingly inexpensive methods
are available for linear regression models with clustered
disturbances. Unless computational cost is an issue, $B=9,\tn999$ and
even $B=99,\tn999$ are generally good choices.
The EDF of the $\tau^*_b$ often provides a better approximation to
$F(\tau)$, the distribution of~$\tau$, than does its asymptotic
distribution. This can sometimes be shown formally, but generally only
under strong assumptions and at the cost of a great deal of algebra
\citep[Section~5]{DMN_2019}. For the model \eqref{eq:lrmodel},
however, the intuition is quite simple. In many cases, the poor
finite\kern 0.04167em-sample properties of test statistics based on CV$_{\tn1}$
arise because $\sum_{g=1}^G \hat{\bm{s}}_g\hat{\bm{s}}_g^\top$, provides a
poor approximation to $\sum_{g=1}^G{\bm{\Sigma}}_g$. Often the bootstrap
analog of the former provides a similarly poor approximation to the
bootstrap analog of the latter. If so, then it is plausible that the
empirical distribution of the $\tau^*_b$ will differ from the
asymptotic distribution of the $\tau^*_b$ in roughly the same way as
the distribution of~$\tau$ differs from its asymptotic distribution.
In that case, the EDF of the bootstrap test statistics should provide
a reasonably good approximation to $F(\tau)$.
The EDF of the $\tau^*_b$ may be obtained by sorting the $\tau^*_b$
from smallest to largest. Number $(1-\alpha)(B+1)$ then provides an
estimate of the $1-\alpha$ quantile, which may be used as the critical
value for an upper-tail test at level~$\alpha$. Identical inferences
will be obtained by calculating the upper-tail bootstrap $P$~value,
\begin{equation}
\hat{P}^*(\tau) = \frac{1}{B} \sum_{b=1}^B {\mathbb I}(\tau^*_b > \tau),
\label{bootp}
\end{equation}
and rejecting the null hypothesis whenever $\hat{P}^*(\tau) < \alpha$.
Here $\tau$ could be either the Wald statistic \eqref{Waldstat} or the
absolute value of the $t$-statistic \eqref{eq:tstat}.
Setting $\tau = |t_a|$ in \eqref{bootp} imposes symmetry on the
bootstrap distribution of $t_a$. In many cases, it makes sense to do
this, because cluster-robust $t$-statistics for linear regression
models with exogenous regressors are often symmetrically distributed
around the origin, at least to a good approximation. When $\tau =
|t_a|$, the quantity $\hat{P}^* (\tau)$ defined in \eqref{bootp} is a
symmetric bootstrap $P$~value for a two\kern 0.04167em-sided test of
${\bm{a}}^\top({\bm\beta}-{\bm\beta}_0)=0$; see~\eqref{eq:tstat}.
In dynamic models, nonlinear models, and models estimated by
instrumental variables, however, it is common for the coefficients of
interest to be biased. This causes the associated $t$-statistics to
have non-zero means in finite samples. In such cases, it makes sense
to use the equal-tail bootstrap $P$~value,
\begin{equation}
\hat{P}^*_{\rm et}(\tau) = \frac{2}{B}\min\Big(\sum_{b=1}^B{\mathbb I}(\tau^*_b >
\tau),\; \sum_{b=1}^B{\mathbb I}(\tau^*_b \le \tau)\kern -0.08333em\Big).
\label{equaltail}
\end{equation}
Here we compute upper-tail and lower-tail $P$~values, take the minimum
of them, and then multiply by 2 to ensure that the nominal level of the
test is correct.
There are many ways to construct a bootstrap confidence interval for a
regression coefficient $\beta$ of which we have an estimate
$\hat\beta$. A method that is conceptually (but not always
computationally) simple is to invert a bootstrap test. This means
finding two values of the coefficient, say $\beta_l$ and $\beta_u$,
with $\beta_u > \beta_l$ and normally on opposite sides of
$\hat\beta$, such that
\begin{equation}
\hat{P}^*_{\rm et}\big(t(\beta=\beta_l)\big) = \alpha
\quad\mbox{and}\quad
\hat{P}^*_{\rm et}\big(t(\beta=\beta_u)\big) = \alpha.
\label{eq:confint}
\end{equation}
Here $t(\beta=\beta_c)$, for $c=l$ and $c=u$, is a cluster-robust
$t$-statistic for the hypothesis that $\beta=\beta_c$. The desired
$1-\alpha$ confidence interval is then $[\beta_l,\, \beta_u]$. When
the bootstrap DGP imposes the null hypothesis, the distribution of the
bootstrap samples depends on the value of $\beta_c$. Solving the two
equations in \eqref{eq:confint} therefore requires iteration; see
\citet{Hansen_1999} and \citet{JGM_2015,JGM-fast}. In general, these
methods tend to be expensive, but this is not the case for the WCR
bootstrap to be discussed in \Cref{subsec:wild}.
For bootstrap confidence intervals, it is common to use bootstrap DGPs
that do not impose the null hypothesis, because no iteration is then
required. The simplest method is just to calculate the standard
deviation of the $\hat\beta^*_b$ and use this number, say
$s^*(\hat\beta)$, as an estimator of the standard error of $\hat\beta$.
The confidence interval is then
\begin{equation}
\label{eq:sstarint}
\big[\hat\beta - c_{1-\alpha/2}\, s^*(\hat\beta), \;
\hat\beta + c_{1-\alpha/2}\, s^*(\hat\beta)\big],
\end{equation}
where $c_{1-\alpha/2}$ is the $1 - \alpha/2$ quantile of (in this
case) the $t(G-1)$ distribution. A better approach, at least in
theory, is to use the studentized bootstrap, or percentile\kern 0.04167em-$t$,
confidence interval advocated in \citet{Hall_1992}, which is
\begin{equation}
\label{eq:studboot}
\big[\hat\beta - s_\beta\kern 0.08333em c^*_{\smash{1-\alpha/2}}, \;
\hat\beta - s_\beta\kern 0.08333em c^*_{\smash{\alpha/2}}\big],
\end{equation}
where $s_\beta$ is the standard error of $\hat\beta$ from the CRVE,
and $c^*_z$ denotes the $z$ quantile of the bootstrap
$t$-statistics~$\tau_b^*$. Although the higher-order theory in
\citet{DMN_2019} does not explicitly deal with confidence intervals, it
strongly suggests that the intervals \eqref{eq:sstarint} and
\eqref{eq:studboot} should not perform as well as inverting a bootstrap
test based on a bootstrap DGP that imposes the null hypothesis.
Simulation results in \citet{JGM_2015} are consistent with these
predictions. However, the intervals \eqref{eq:sstarint} and
\eqref{eq:studboot} have the advantage that they are easy to compute.
No iteration is required, and a single set of bootstrap samples can be
used to compute confidence intervals for all the parameters of interest.
\subsection{Pairs Cluster Bootstrap}
\label{subsec:pairs}
The most important aspect of any bootstrap procedure is how the
bootstrap samples are generated. The only procedure applicable to
every model that uses clustered data is the pairs cluster bootstrap,
which is also sometimes referred to as the cluster bootstrap, the
block bootstrap, or resampling by cluster. The pairs cluster bootstrap
works by grouping the data for every cluster into a $[{\bm{y}}_g,{\bm{X}}_g]$
pair and then resampling from the $G$ pairs. Every bootstrap sample
is constructed by choosing $G$ pairs at random with equal
probability~$1/G$.
Although this procedure ensures that every bootstrap sample contains
$G$ clusters, the number of observations inevitably varies across the
bootstrap samples, unless all cluster sizes are the same. The size of
the bootstrap samples can vary greatly, because the largest clusters
may be over-represented in some bootstrap samples and
under-represented in others. This limits the ability of the bootstrap
samples to mimic the actual sample. So does the fact that the
${\bm{X}}^\top{\bm{X}}$ matrix is different for every bootstrap sample.
Because the pairs cluster bootstrap does not impose the null
hypothesis, care must be taken when calculating the bootstrap test
statistics. If the null hypothesis is that $\beta=\beta_0$, the actual
$t$-statistic will have numerator $\hat\beta-\beta_0$, but the
bootstrap $t$-statistic must have numerator $\hat\beta^*_b-\hat\beta$,
where $\hat\beta^*_b$ is the estimator of $\beta$ for the \th{b}
bootstrap sample. In this case, $\hat\beta$ is the parameter value
associated with the bootstrap DGP. Because the bootstrap DGP does not
impose the null hypothesis, the pairs cluster bootstrap cannot be used
to construct the confidence interval \eqref{eq:confint}, but it can be
used to construct the intervals \eqref{eq:sstarint} and
\eqref{eq:studboot}.
\citet[Section VI]{CM_2015} discusses several problems that can arise
with the pairs cluster bootstrap and sensibly suggests that
investigators should examine the empirical distributions of the
bootstrap coefficient estimates and test statistics. For example, if
the bootstrap distribution has more than one mode, then it probably
does not provide a good approximation to the actual distribution. This
can happen when one or two clusters are very different from all the
others; see \Cref{subsec:leverage}.
\citet{FP_2019} proposes several bootstrap procedures that can be
thought of as variants of the pairs (not pairs cluster) bootstrap. The
first step is to run regressions at either the individual level or the
group $\times$ time\kern 0.04167em-period level, then aggregate the residuals so
that there is just one residual per cluster, and finally run bootstrap
regressions on the resampled residuals. These procedures include
parametric methods to correct for the heteroskedasticity generated by
variation in the number of observations per group. Remarkably, they
can work well even with just one treated cluster. However, this is
possible only because, unlike methods based on a CRVE, they do not
allow for unrestricted heteroskedasticity.
In general, the pairs cluster bootstrap is expensive to compute.
However, a computational shortcut for linear regression models can
make it feasible even when $N$ and $B$ are both large; see
\citet{JGM-fast}. Nevertheless, we do not recommend this method for
the linear regression model \eqref{eq:lrmodel}, because, as we discuss
in the next subsection, a much better method is available. With
nonlinear models such as the probit model, the pairs cluster bootstrap
may be attractive even though it can be expensive. However, we cannot
recommend it without reservation, because it appears that very little
is known about its finite\kern 0.04167em-sample properties. Simulation evidence
suggests that, at least in some cases, it can either over-reject or
under-reject severely \citep{MW-TPM,JGM-fast}.
\subsection{Wild Cluster Bootstrap}
\label{subsec:wild}
The restricted wild cluster, or WCR, bootstrap was first suggested in
\citet{CGM_2008}. Until recently, the only variant of it ever used in
practice is based on CV$_{\tn1}$, and, at time of writing, this
classic variant is the only one for which software is readily
available. Other variants will be discussed briefly at the end of
this subsection.
Suppose that $\tilde{\bm\beta}$ denotes the OLS estimator of ${\bm\beta}$
subject to the restriction ${\bm{a}}^\top{\bm\beta}={\bm{a}}^\top{\bm\beta}_0$, which
is to be tested, and $\tilde{\bm{u}}_g = {\bm{y}}_g - {\bm{X}}_g\tilde{\bm\beta}$
denotes the vector of restricted residuals for the \th{g} cluster.
Then the traditional way to write the WCR bootstrap DGP is
\begin{equation}
{\bm{y}}_g^{*b} = {\bm{X}}_g\tilde{\bm\beta} + {\bm{u}}_g^{*b}, \quad
{\bm{u}}_g^{*b} = v_g^{*b} \tilde{\bm{u}}_g, \quad g=1,\ldots,G,
\label{eq:wcr}
\end{equation}
where the $v_g^{\ast b}$ are independent realizations of an auxiliary
random variable $v^*$ with zero mean and unit variance. In practice,
the best choice for $v^*$ is usually the Rademacher distribution, in
which case $v^*$ equals 1 or $-1$ with equal probabilities
\citep{DF_2008,DMN_2019}. This imposes symmetry on the bootstrap
disturbances.
Instead of using \eqref{eq:wcr}, we can generate the bootstrap scores
directly as
\begin{equation}
\label{eq:scoreboot}
{\bm{s}}_g^{*b} = v_g^{*b}\tilde{\bm{s}}_g, \quad g=1,\ldots,G,
\end{equation}
where $\tilde{\bm{s}}_g = {\bm{X}}_g^\top \tilde{\bm{u}}_g$ is the score vector
for the \th{g} cluster evaluated at the restricted estimators. Plugging
the ${\bm{s}}_g^{*b}$ into \eqref{eq:betahat} then yields the bootstrap
estimators $\hat{\bm\beta}^*_b$. This method is computationally inexpensive
\citep{JGM-fast}. It also provides an intuitive justification for the
WCR bootstrap because, as we stressed in \Cref{sec:asytheory}, the
finite\kern 0.04167em-sample properties of cluster-robust tests depend mainly on
the properties of the scores.
\citet{DMN_2019} establishes the asymptotic validity of the classic
WCR bootstrap with CV$_{\tn1}$ standard errors and also studies the
unrestricted wild cluster, or WCU, bootstrap. The difference between
the WCR and WCU bootstraps is that, for the latter, the unrestricted
scores $\hat{\bm{s}}_g$ are used instead of the restricted ones
$\tilde{\bm{s}}_g$ in the bootstrap DGP~\eqref{eq:scoreboot}. In general,
the WCU bootstrap does not perform as well in finite samples as the
WCR one. The paper contains both theoretical results (based on
Edgeworth expansions) and simulation results to support this
assertion. However, the WCU bootstrap has the advantage that the
bootstrap DGP does not depend on the restrictions to be tested. The
same set of bootstrap samples can therefore be used to perform tests
on any restriction or set of restrictions and/or to construct
confidence intervals based on either \eqref{eq:sstarint} or
\eqref{eq:studboot} for any coefficient of interest.
\citet{DMN_2019} also proves the asymptotic validity of two variants
of the ordinary wild bootstrap, restricted (WR) and unrestricted (WU),
for the model \eqref{eq:lrmodel}. The ordinary wild bootstrap uses $N$
realizations of $v^*$\kern -0.08333em, one for each observation, instead of just
$G$. This means that the disturbances for the bootstrap samples are
uncorrelated within clusters. Although this implies that the
distribution of the $\hat{\bm\beta}^*_b$ cannot possibly match that of
$\hat{\bm\beta}$, it often does not prevent the distribution of the
$\tau^*_b$ from providing a good approximation to the distribution
of~$\tau$. With one important exception (see below), the WR bootstrap
seems to work less well than the WCR bootstrap, as asymptotic theory
predicts. It can also be much more expensive to compute when $N$ is
large.
The classic versions of the WCR and WCU bootstraps that use
CV$_{\tn1}$ are surprisingly inexpensive to compute. The computations
require only the matrices ${\bm{X}}_g^\top{\bm{X}}_g$ and the vectors
${\bm{X}}_g^\top{\bm{y}}_g$; see \citet{JGM-fast}. The calculations can be
made even faster by rewriting the bootstrap test statistic so that it
depends on all the sample data in the same way for every bootstrap
sample; see \citet*{RMNW}. The only thing that varies across the
bootstrap samples is the $G$-vector ${\bm{v}}^{*b}$ of realizations of the
auxiliary random variable. There are some initial computations that
may be expensive when $N$ is large, but they only have to be done
once. After that, the ${\bm{v}}^{*b}$ and the results of the initial
computations are used to compute all the bootstrap test statistics.
This fast procedure is implemented in the package \texttt{boottest}
for both \texttt{Stata} and \texttt{Julia}; for details, see
\citet{RMNW}. In \Cref{sec:empirical}, there is an illustration of how
fast it can be; see the notes to \Cref{tab:teenreg}. Importantly,
\texttt{boottest} not only computes WCR bootstrap $P$~values for both
$t$-tests and Wald tests; it also computes WCR bootstrap confidence
intervals based on \eqref{eq:confint}. The package has many other
capabilities as well.
In many cases (exceptions will be discussed below), the classic
CV$_{\tn1}$ variant of the WCR bootstrap yields very accurate
inferences. These are generally more accurate than those for the pairs
cluster or WCU bootstraps. In addition to \citet{DMN_2019}, see
\citet{CGM_2008}, \citet{MW-JAE}, and \citet{JGM-fast}. Because the
${\bm{X}}_g$ are always the same, the wild cluster bootstrap is able to
replicate what is often the main source of heterogeneity, namely,
variation in cluster sizes, in every bootstrap sample. This is not the
case for the pairs cluster bootstrap, and it surely contributes to the
superior accuracy of inferences based on the WCR bootstrap.
Of course, no bootstrap method can work perfectly. Not surprisingly,
the performance of the WCR bootstrap using CV$_{\tn1}$ tends to
deteriorate as $G$ becomes smaller, as the number of regressors
increases, and as the clusters become more heterogeneous. In
particular, it can sometimes perform very badly when the number of
treated clusters $G_1$ is very small \citep{MW-JAE, MW-TPM, MW-EJ}.
This is true both for pure treatment models, where all the
observations in each cluster are either treated or not, and for DiD
models, where only some observations in the treated clusters are
treated.
Unlike other methods, which generally over-reject severely when there
are few treated clusters (\Cref{subsubsec:few}), the WCR bootstrap
usually under-rejects in this case. This happens because the
distribution of the bootstrap statistics $\tau^*_b$ depends on the
value of the actual test statistic~$\tau$. The larger is $\tau$, the
more dispersed are the $\tau^*_b$. This makes $\hat P^*(\tau)$ in
\eqref{bootp} larger than it should be. In the most extreme case of
just one treated cluster, rejection frequencies may be essentially
zero. In this case, the bootstrap distribution is often bimodal
\citep[Figure~4]{MW-JAE}, so that plotting it can provide a useful
diagnostic. When there are few treated clusters and the WCR bootstrap
$P$~value seems suspiciously large, it may be worth trying the
ordinary wild restricted (WR) bootstrap, which can sometimes work
surprisingly well in this context \citep{MW-EJ}. This is a case where
randomization inference can be attractive; see \Cref{subsec:RI}.
We recommend using at least one variant of the WCR bootstrap
(preferably with at least $B=9,\tn999$) almost all the time. All
variants are often remarkably inexpensive to compute, and they often
seem to work well. When $G$ is not too small and the clusters are not
too heterogeneous, WCR bootstrap $P$~values and confidence intervals
may be quite similar to ones based on CV$_{\tn1}$ or CV$_{\tn3}$ with
$t(G-1)$ critical values. In that case, it is likely that
finite\kern 0.04167em-sample issues are not severe, and there is probably no need
to do anything else. When there is a large discrepancy between the
results of different methods, however, it would make sense to try
several bootstrap methods and maybe some of the methods discussed in
\Cref{sec:other}.
The classic version of the WCR bootstrap can sometimes work remarkably
well even when $G$ is very small. In fact, \citet*{CSS_2021} shows
that it can yield exact inferences in certain cases where $N$ is large
and $G$ is small. These results are obtained by exploiting the
relationship between the WCR bootstrap with Rademacher auxiliary
random variables and randomization inference (\Cref{subsec:RI}).
However, they require rather strong homogeneity conditions on the
distribution of the covariates across clusters, as well as limits on
the amount of dependence within each cluster similar to those in
\citet{BCH_2011}.
At this point, a word of warning is in order. Almost all the
simulation results that we have referred to are based on models with
one or just a few regressors, and these regressors are typically
generated in a fairly simple way. Moreover, almost all existing
simulations focus on $t$-statistics rather than Wald statistics. There
is evidence that rejection frequencies for all methods increase as the
number of regressors increases, and that Wald tests are less reliable
than $t$-tests; see \citet[Section~C.2]{DMN_2019} and
\citet{MNW-bootknife} on the former point, \citet{PT_2018} on the
latter, and \citet{JGM-fast} on both of them. Thus the classic WCR
bootstrap may perform less well in empirical applications with large
numbers of regressors than it has typically done in simulations,
especially when there is more than one restriction.
When $G$ is small, the WCR bootstrap encounters an important practical
problem. For the Rademacher distribution, or any other two\kern 0.04167em-point
distribution, the number of possible bootstrap samples is just
$2^G$\kern -0.08333em. \citet{Webb_2014} proposes a six-point distribution which
largely solves this problem, because $6^G$ is much larger than
$2^G$\kern -0.08333em. This distribution seems to work almost as well as Rademacher
for most values of $G$, and sometimes much better when $G$ is very
small. Whenever either $2^G$ (for Rademacher) or $6^G$ (for six-point)
is smaller than the chosen value of $B$, we can enumerate all possible
bootstrap samples instead of drawing them at random. For example,
when $G=16$, there are just $65,\tn536$ Rademacher bootstrap samples
to enumerate (one of which is identical to the actual sample). This
eliminates simulation randomness from the bootstrap procedure. In
fact, \texttt{boottest} uses enumeration by default whenever $B$ is
greater than the number of possible bootstrap samples.
Most of the discussion above pertains to the classic variant of the
WCR bootstrap, which uses \eqref{eq:scoreboot} to generate bootstrap
samples and CV$_{\tn1}$ to calculate standard errors.
\citet{MNW-bootknife} introduces three new variants. The simplest of
these still uses \eqref{eq:scoreboot}, but the actual and bootstrap
test statistics employ CV$_{\tn3}$ standard errors, computed using
\eqref{eq:jackvar}. Two somewhat more complicated variants generate
the bootstrap scores using an alternative to \eqref{eq:scoreboot} that
implicitly involves a transformation of the restricted scores similar
to the one that yields the $\acute{\bm{s}}_g$ in \eqref{eq:CV3}. One of
these new variants uses CV$_{\tn1}$ standard errors, and the other
uses CV$_{\tn3}$ standard errors. Simulation results in
\citet{MNW-bootknife} suggest that all three new variants of the WCR
bootstrap outperform the classic variant in a number of circumstances.
In particular, the tendency to over-reject seems to increase much more
slowly for the new variants than for the classic one as the number of
regressors increases and as the variation in cluster sizes increases.
However, the new variants still tend to under-reject severely when the
number of treated clusters is small. We hope that \texttt{Stata} and/or
\texttt{R} packages able to compute these new variants will become
available in the near future.
\section{Other Inferential Procedures}
\label{sec:other}
Bootstrap methods are not the only way to obtain inferences more
accurate than those given by cluster-robust $t$-tests and confidence
intervals using the $t(G-1)$ distribution. Numerous other procedures
have been proposed, which broadly fall into two categories that we
discuss in the following two subsections. Except in the case of
treatment models with very few clusters or very few treated clusters
(\Cref{subsec:RI}), we do not recommend any of these procedures in
preference to CV$_{\tn3}$ combined with the $t(G-1)$ distribution or
the WCR bootstrap, especially the newest variants of the latter.
However, when the recommended methods yield conflicting results, it is
surely a good idea to try other methods as well.
\subsection{Alternative Critical Values}
\label{subsec:altcrit}
Critical values for cluster-robust test statistics can be based on
various approximations. The first paper to take this approach seems to be
\citet{BM_2002}. It suggests methods for calculating approximations to
CV$_{\tn2}$ or CV$_{\tn3}$ $t$-statistics based on the Student's $t$
distribution with an estimated degrees-of-freedom (d-o\kern 0.04167em-f)
parameter. These employ what is called a ``Satter\-thwaite
approximation'' to calculate the d-o\kern 0.04167em-f. This is done under the
assumption that the variance matrix of ${\bm{u}}$ is proportional to an
identity matrix. The d-o\kern 0.04167em-f parameter is different for every
hypothesis to be tested, and it can be much less than~$G-1$.
\citet{Imbens_2016} proposes a similar procedure for $t$-tests based
on CV$_{\tn2}$ under the assumption that the variance matrix of ${\bm{u}}$
corresponds to a random-effects model. As we saw in
\Cref{subsec:sources}, such a model implies that the disturbances
within each cluster are equi-correlated, and the intra-cluster
correlation $\rho$ must be estimated from the residuals. When cluster
fixed effects are partialed out, doing so absorbs any random effects,
making this approach inapplicable.
\citet{AY-exact} proposes a related method that uses CV$_{\tn1}$
instead of CV$_{\tn2}$. It involves two steps. In the first step, the
CV$_{\tn1}$ standard error for the coefficient of interest is
multiplied by a factor greater than one. In the second step, a
d-o\kern 0.04167em-f parameter is calculated. The standard error and the
d-o\kern 0.04167em-f parameter can then be used to compute either a $t$-statistic
and its $P$~value or a confidence interval. A \texttt{Stata} package
called \texttt{edfreg} is available.
The three procedures just discussed are described in some detail in
\citet[Appendix~B]{MW-EJ}. However, the the CV$_2$-based and
CV$_3$-based procedures can be very expensive as described there, and
\citet{NAAMW_2020} provides a better way to compute them. Limited
simulation evidence suggests that the performance of Young's method is
similar to those of the two methods based on CV$_{\tn2}$. However,
this evidence is by no means definitive, because the simulations focus
on only a narrow set of treatment models.
Using Hotelling's $T^{\tk2}$ distribution with estimated degrees of
freedom, \citet{PT_2018} generalizes the CV$_{\tn2}$ procedure of
\citet{BM_2002} to the case of Wald tests based on \eqref{Waldstat}.
Simulations suggest that the resulting tests always reject less often
than standard Wald tests. They rarely over-reject but often
under-reject, and they sometimes do so quite severely. The
\texttt{clubSandwich} package for \texttt{R} and the
\texttt{reg$\_$sandwich} package for \texttt{Stata} implement the
procedures discussed in \citet{PT_2018}.
Although the procedures discussed in this subsection have some
theoretical appeal and seem to work well in many cases, we are not
aware of any evidence that they outperform either tests based directly
on CV$_{\tn3}$ or the wild cluster bootstrap for a range of models and
DGPs. One limitation is that the Student's $t$ distribution is not very
flexible. Even when the d-o\kern 0.04167em-f parameter is estimated very well, the
best we can hope for is that a test based on this distribution will be
accurate for some level of interest. It may well over-reject at some
levels and under-reject at others.
Our recommendation is to use the methods discussed in this subsection
primarily to confirm (or perhaps cast doubt on) the results of the
methods that we recommend when there is concern about the reliability
of the latter. Cases of particular concern are ones with few but
balanced clusters (say, $G \leq 12$), ones with balanced but few
treated (or few control) clusters (say $G_1 \leq 6$ or $G-G_1 \leq
6$), ones with seriously unbalanced cluster sizes (even when $G$ is
quite large), ones with treated clusters that are unusually large or
small, and ones with any sort of heterogeneity that causes a few
clusters to have high leverage; see \Cref{subsec:leverage}. It is
always reassuring when several methods yield essentially the same
inferences.
\subsection{Randomization Inference}
\label{subsec:RI}
Randomization inference (RI) was proposed by \citet{Fisher_1935} as a
distribution-free way to perform hypothesis tests in the context of
experiments. RI tests are also called permutation tests.
\citet[Chapter~15]{Lehmann_2005} gives a formal introduction, and
\citet[Chapter~15]{Imbens_2015} provides a more accessible discussion.
The key idea of RI is to compare an outcome $\tau$ that is actually
observed with a set of outcomes that might have been observed if
treatment had been assigned differently. The outcome could be a sample
average, a coefficient estimate, or a test statistic.
Specifically, consider a clustered regression model with treatment at
the cluster level, which could be a DiD model where only some
observations within the treated clusters receive treatment. Then
$\tau$ might be the average treatment effect for some outcome measure.
RI procedures may be attractive when it is plausible that treatment
was assigned at random and the researcher is interested in the sharp
null hypothesis that the treatment has no effect. They can sometimes
yield reliable results even when the number of clusters is small
and/or the number of treated clusters is small.
Suppose there are $G$ clusters, $G_1$ of which received treatment. The
number of ways in which treatment could have been assigned to $G_1$
out of the $G$ clusters is
\begin{equation}
{}_{G}C_{G_1} = \frac{G\kern 0.04167em!}{G_1!(G-G_1)!}\kern 0.08333em.
\label{Nassign}
\end{equation}
One of these corresponds to the actual assignment, and the remaining
$S = {}_{G}C_{G_1} - 1$ are called re\kern 0.04167em-randomizations. Each
re\kern 0.04167em-randomization involves pretending that a particular set of
$G_1$ clusters was treated, with the remaining $G-G_1$ serving as
controls. The values of the dependent variable do not change across
re\kern 0.04167em-randomizations, only the values of the treatment dummy. For
every re\kern 0.04167em-randomization, indexed by $j$, we could calculate a test
statistic $\tau^*_j$. If the observable characteristics of the
clusters were all the same, it would make sense to compare $\tau$ with
the empirical distribution of the $\tau^*_j$. To do so, it is natural
to calculate the $P$~value for an upper-tail test as either
\begin{equation}
P_1^*(\tau) = \frac1S \sum_{j=1}^S {\mathbb I}(\tau^*_j \ge \tau) \quad\mbox{or}
\quad P_2^*(\tau) = \frac1{S+1}\Big(\!1 + \sum_{j=1}^S
{\mathbb I}(\tau^*_j \ge \tau)\kern -0.08333em\Big).
\label{Pvalues}
\end{equation}
Here $P_2^*$ implicitly includes the actual assignment to treatment,
and $P_1^*$ omits it.
When $\alpha(S+1)$ is an integer for a test at level $\alpha$, both $P$
values in \eqref{Pvalues} yield the same result, and the test is exact
if the distributions of the $\tau^*_j$ are the same as that of~$\tau$.
However, $P_1^*$ and $P_2^*$ can differ noticeably when $S$ is small.
The latter is more conservative and seems to be more popular. For
moderate values of $S$, it is often easy to enumerate all of the
possible $\tau^*_j$, but this is infeasible even when $G$ is not
particularly large. In such cases, we must choose a number of
re\kern 0.04167em-randomizations, say $B=99,\tn999$, at random. In principle,
these should be drawn without replacement, but that is not important
if $B$ is small relative to~$S$.
It seems to be widely believed that tests based on RI are always
exact. This is not true. When treatment is not assigned at random, or
when the observed characteristics of the treated clusters differ
systematically from those of the controls, we cannot expect the
distributions of $\tau$ and the $\tau^*_j$ to coincide.
For DiD models, \citet{CT_2011} proposes a test that is very similar
to an RI test based on the OLS estimate of the coefficient on a
treatment dummy. That paper also shows how to obtain a confidence
interval by inverting the test. \citet{MW-RI} studies two procedures
for DiD tests, one based on coefficient estimates (called RI-$\beta$)
and one based on cluster-robust $t$-statistics (called RI-$t$). Both
procedures work very well when the clusters are homogeneous and $G$ is
reasonably large, even when~$G_1=1$. However, when the treated
clusters are systematically larger or smaller than average, neither
RI-$\beta$ nor \mbox{RI-$t$} tests perform well, although the latter
typically perform better. In such cases, it appears that $G$ may have
to be quite large (much larger than for the WCR bootstrap) before
either procedure works really well, even when $G_1$ is not
particularly small.
\citet{AY_fisher} contains an interesting recent application of RI,
where the RI-$\beta$ and \mbox{RI-$t$} procedures are applied to
regressions for the results of 53 randomized experiments in published
papers. \texttt{Stata} packages called \texttt{randcmd} and
\texttt{randcmdci} are available. In many cases, estimates that were
originally reported to be significant are not significant according to
the RI procedures. Not surprisingly, this is particularly true for
regressions with a few high-leverage clusters or observations.
There is evidently a close relationship between RI and the wild cluster
bootstrap. Evaluating all possible bootstrap samples by enumeration is
quite similar to evaluating all possible re\kern 0.04167em-randomizations. The
results of \citet{CSS_2021} and \citet{MW-RI} strongly suggest that
homogeneity across clusters is more important for RI than for the WCR
bootstrap. For the former, the regressors change as we compute each of
the $\tau^*_j$, but the regressand stays the same. For the latter, the
regressand changes as we compute each of the $\tau^*_b$, but the
regressors stay the same. Thus, the empirical distribution to which
$\tau$ is being compared is conditional on the actual regressors
(including the clusters actually treated) for the WCR bootstrap, but
not for the RI procedures discussed above.
The RI methods discussed so far are not the only ones.
\citet*{CRS_2017} proposes ``approximate randomization tests'' based
on the cluster-level estimators of \citet{IM_2010}; see
\citet*{CCKS_2021} for a guide to this approach. In models for
treatment effects, every cluster must include both treated and
untreated observations. This can be accomplished by merging clusters,
but at the cost of making $G$ smaller with a resulting loss in power.
\citet{Hag_2019a,Hag_2019b} develops RI tests that can be used even
when $G$ is quite small and there is substantial heterogeneity across
clusters. These tests do not require cluster-level estimation, but
$G_1$ and $G-G_1$ should both be no less than~4. \citet{HS_2020}
develops an RI procedure based on RI-$t$ for the case in which one
cluster is much larger than any of the others.
An alternative way to deal with the problem of one or very few treated
clusters, which can involve an RI-like procedure, is the method of
synthetic controls surveyed in \citet{Abadie_2021}.
\section{What Should Investigators Report?}
\label{sec:report}
Many studies fail to report enough information to convince readers
that their empirical results should be believed. Investigators often
simply report test statistics or confidence intervals based on
CV$_{\tn1}$ together with the $t(G-1)$ distribution. Although results
of this type may be reliable, they often will not be. Unless the
number of clusters is fairly large and evidence demonstrates a
convincing degree of homogeneity across clusters, results from at
least one or two other methods should be reported. These might include
tests and/or confidence intervals based on CV$_{\tn3}$, $P$ values
and/or confidence intervals based on one or more variants of the WCR
bootstrap discussed in \Cref{subsec:wild}, or perhaps results from
some of the alternative methods discussed in \Cref{sec:other}.
It is important to report some key information about the sample as a
matter of routine. The fundamental unit of inference when the
observations are clustered is not the observation but the cluster;
this is evident from \eqref{eq:betahat} and \eqref{eq:trueV}. The
asymptotic theory discussed in \Cref{subsec:large} therefore depends
on $G$, not~$N$\kern -0.08333em. With the exception of certain very special cases
discussed in \Cref{subsec:small}, asymptotic approximations tend to
work poorly when there are few clusters. It is therefore absolutely
essential to report the number of clusters, $G$, whenever inference is
based on a~CRVE. This is even more important than reporting~$N$\kern -0.08333em.
Moreover, because the distributional approximations perform best when
the scores are homogeneous across clusters, and the most important
source of heterogeneity is often variation in cluster sizes, it is
extremely important to report measures of this variation. At a
minimum, we believe that investigators should always report the median
cluster size and the maximum cluster size, in addition to $N$ and $G$.
When $G$ is small, or when the distribution of the $N_g$ is unusual,
it would be good to report the entire distribution of cluster sizes in
the form of either a histogram or a table.
Of course, clusters can be heterogeneous in many ways beyond their
sizes. In \Cref{subsec:leverage}, we discussed a measure of leverage
at the cluster level. Another measure is the partial leverage of each
cluster for the coefficient(s) of interest, which generalizes the
observation-level notion of partial leverage of \citet{CW_1980}. If
$\acute{\bm{x}}_j$ denotes the vector of residuals from regressing
${\bm{x}}_j$ on all the other regressors, and $\acute{\bm{x}}_{gj}$ is the
subvector of $\acute{\bm{x}}_j$ corresponding to the \th{g} cluster, then
the partial leverage of cluster $g$ for the \th{j} coefficient is
\begin{equation}
\label{partlev}
L_{gj} =
\frac{\acute{\bm{x}}_{gj}^\top\acute{\bm{x}}_{gj}}{\acute{\bm{x}}_j^\top\acute{\bm{x}}_j}\kern 0.08333em.
\end{equation}
When a cluster has high partial leverage for the \th{j} coefficient,
removing that cluster has the potential to change the estimate of the
\th{j} coefficient substantially.
A popular way to quantify the heterogeneity of clusters is to
calculate $G_j^*$\kern -0.08333em, the ``effective number of clusters'' for the
\th{j} coefficient, as proposed in \citet*{CSS_2017}. This number is
always less than $G$, and it can provide a useful warning when $G_j^*$
is much smaller than~$G$. The value of $G_j^*$ depends on an unknown
parameter $\rho$, the intra-cluster correlation of the disturbances
when they are assumed to be equi-correlated. \citet{CSS_2017} suggests
setting $\rho=1$, as a sort of worst case. However, since cluster
fixed effects absorb all intra-cluster correlation for the random
effects model \eqref{eq:REmodel}, only $\rho=0$ makes sense for models
with such fixed effects, and trying to use $\rho\ne0$ can lead to
numerical instabilities. \citet{MNW-influence} discusses these issues
and shows how to calculate $G_j^*(\rho)$ efficiently for any value of
$\rho$ in models without cluster fixed effects.
Using \eqref{partlev} and the definition of $G_j^*(0)$, it can be
shown that the latter is a monotonically decreasing function of the
squared coefficient of variation of the partial leverages $L_{gj}$,
which is defined as
\begin{equation*}
V_s(L_{j\bullet}) = \frac{1}{G\kern 0.08333em \bar L_j^2}
\sum_{g=1}^G (L_{gj} - \bar L_j)^2,
\end{equation*}
where $\bar L_j$ is the sample mean of the $L_{gj}$. Thus $G_j^*(0)$
and $V_s(L_{j\bullet})$ provide exactly the same information. When the
partial leverage for the \th{j} coefficient is the same for every
cluster, $V_s(L_{j\bullet}) = 0$ and $G_j^*(0) = G$. When the $L_{gj}$
differ greatly, $V_s(L_{j\bullet})$ is large, and $G_j^*(0) <\kern -0.08333em< G$.
We suggest that investigators should routinely report the leverages
$L_g$ defined in \eqref{eq:traceXX}, the partial leverages $L_{gj}$
for the coefficient(s) of interest, and the $\hat\beta_j^{(g)}$ for
the same coefficient(s), at least when they provide information beyond
that in the distribution of the cluster sizes. When $G$ is small, it
may be feasible to report all these numbers. Otherwise, it may be
feasible to graph them, as in \Cref{fig:EDF}, or to report summary
statistics such as $G_j^*(0)$ or $V_s(L_{j\bullet})$. Reporting
measures of influence and leverage should help to identify cases in
which inference may be unreliable, as well as sometimes turning up
interesting, or possibly disturbing, features of the data.
When there is clustering in two or more dimensions
(\Cref{subsec:twoway}), we recommend that investigators compute
everything just suggested for each of the clustering dimensions and
also for their intersection(s). If the results are interesting, they
should be reported. For two\kern 0.04167em-way clustering, this might mean
reporting, or at least summarizing, up to three sets of leverages,
partial leverages, and $\hat\beta_j^{(g)}$\kern -0.08333em.
\section{Empirical Example}
\label{sec:empirical}
We illustrate many of our recommendations by revisiting a
long-standing empirical question in labor economics, namely, the
impact of the minimum wage on young people. In the past two decades,
many U.S.\ states have significantly increased their minimum wages. In
fact, from 2000 to 2019, every state increased the nominal minimum
wage by at least 27\%, and six states doubled it. Moreover, recent
proposals to increase the national minimum wage to \$15 per hour have
reinvigorated the debate about the effects of minimum wages.
Some classic references on the impact of the minimum wage are
\citet{Mincer_1976} and \citet{CK_1994}. The latter paper was among
the very first and most influential applications of the DiD
methodology, which continues to be used in this literature. For
example, \citet{Seattle_2017} uses a DiD analysis to study the effects
of a large increase in the minimum wage in Seattle.
\cite{Wolfson_2019} and \cite{Neumark_2021} survey many recent studies
on the impacts of minimum wages. Both conclude that the majority of
studies, but not all, find dis-employment effects that are
concentrated among teenagers and those with low levels of education.
\citet{Manning_2021} explores why the evidence on employment effects
is mixed.
Instead of using a DiD approach, we exploit state\kern 0.04167em-level
differences in the minimum wage and analyze their impacts on
labor-market and education outcomes at the individual level. Although
we treat the minimum wage as exogenous in our analysis, we hesitate to
call our estimates causal. There is reason to believe that
state\kern 0.04167em-level minimum wages may be endogenous, because states may be
more likely to increase them during good economic times. Moreover, it
is possible that the effects of the minimum wage differ depending on
the state of the economy. However, we ignore these issues, because our
principal objective is to illustrate the importance of clustering for
statistical inference.
The model we estimate is
\begin{equation}
\label{reg:wage}
y_{ist} = \alpha + \beta\kern 0.08333em \textrm{mw}_{\kern -0.08333em st} + {\bm{Z}}_{ist}{\bm{\gamma}}
+ \textrm{year}_t\kern 0.08333em{\bm{\delta}}_t + \textrm{state}_s\kern 0.08333em{\bm{\delta}}_s
+ u_{ist}.
\end{equation}
Here $y_{ist}$ is the outcome of interest for person $i$ in state $s$
in year~$t$. There are three outcome variables. ``Hours'' records the
usual hours worked per week, which is defined only for employed
individuals. ``Employed'' is a binary variable equal to~1 if person
$i$ is employed and to~0 if they are either unemployed or not in the
labor force. ``Student'' is a binary variable equal to~1 if person $i$
is currently enrolled in school and to~0 otherwise. The parameter of
interest is $\beta$, which is the coefficient on $\textrm{mw}_{\kern -0.08333em
st}$, the minimum wage in state $s$ at time~$t$. The row vector
${\bm{Z}}_{ist}$ collects a large set of individual-level controls,
including race, gender, age, and education. There are also year and
state fixed effects. \cite{NW_2007} estimates models similar to
\eqref{reg:wage} with individual-level data, clustering at the state
level.
Data at the individual level from the American Community Survey (ACS)
were obtained from IPUMS \citep{IPUMS_2020} and cover the years
2005--2019. The minimum wage data were provided by
\cite{Neumark_2019}, and we have collapsed them to state\kern 0.04167em-year
averages to match the ACS frequency. Following previous literature, we
restrict attention to teenagers aged 16--19. We keep only individuals
who are ``children'' of the respondent to the survey and who have
never been married. We drop individuals who had completed one year of
college by age~16 and those reporting in excess of 60~hours usually
worked per week. We also restrict attention to individuals who
identify as either black or white.
We consider six different clustering structures that lead to nine
different standard errors for $\hat\beta$. These are no clustering
(with HC$_1$ standard errors), one\kern 0.04167em-way clustering at either the
state\kern 0.04167em-year, state, or region\footnote{Regions are defined as the
U.S.\ Census Divisions, with the following partitioning of states. New
England: CT, MA, MN, NH, RI, VT; Middle Atlantic: NJ, NY, PA; South
Atlantic: DC, DE, FL, GA, MD, NC, SC, VA, WV; East South Central: AL,
KY, MS, TN; East North Central: IL, IN, MI, OH, WI; West North
Central: IA, KS, MN, MO, ND, NE, SD; West South Central: AR, LA, OK,
TX; Mountain: AZ, CO, ID, MT, NM, NV, UT, WY; Pacific: AK, CA, HI, OR,
WA.} level (with both CV$_{\tn1}$ and CV$_{\tn3}$ standard errors),
and two\kern 0.04167em-way clustering by state and year or by region and year
(with CV$_{\tn1}$ standard errors, the only ones currently available
for multi-way clustering). Early empirical research on the impacts of
the minimum wage would have used either conventional or HC$_1$
standard errors, but modern studies would almost always cluster at
some level.
Since the minimum wage is invariant within state\kern 0.04167em-year clusters, it
seems highly likely that the scores are correlated within them, and we
therefore consider (at least) state\kern 0.04167em-year clustering. However,
after a state has increased its minimum wage, it almost always remains
at the new level until it is increased again. This implies that
minimum wages must be correlated across years within each state.
Unless the disturbances happen to be uncorrelated across years within
states, the scores will therefore be correlated, which suggests that
state\kern 0.04167em-level clustering may be appropriate. We consider
region-level clustering based on the nine census divisions because
there may be correlations among nearby states. Largely for
completeness, we also consider two\kern 0.04167em-way clustering by either state
or region and year.
\begin{table}[tp]
\caption{Estimated Impact of the Minimum Wage}
\label{tab:teenreg}
\vspace*{-0.5em}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}
lld{2.4}d{2.4}d{2.4}}
\toprule
\multicolumn{1}{l}{Clustering level} & & \multicolumn{1}{c}{Hours} &
\multicolumn{1}{c}{Employed} & \multicolumn{1}{c}{Student} \\
\midrule
& $\hat \beta$ & -0.1539 & -0.0037 & 0.0022 \\
\vspace{-11pt}\\
None: HC$_1$ & $t$-statistic & -5.4469 & -5.2801 & 4.9719 \\
& $P$~value, $N(0,1)$ & 0.0000 & 0.0000 & 0.0000 \\
\vspace{-11pt}\\
State\kern 0.04167em-year: CV$_{\tn1}$ & $t$-statistic & -3.3823 & -2.6492 & 4.0776 \\
& $P$~value, $t(764)$ & 0.0008 & 0.0082 & 0.0001 \\
& $P$~value, WCR & 0.0027 & 0.0145 & 0.0005 \\
\vspace{-11pt}\\
State\kern 0.04167em-year: CV$_{\tn3}$ & $t$-statistic & -2.9466 & -2.3199 & 3.6095 \\
& $P$~value, $t(764)$ & 0.0033 & 0.0206 & 0.0003 \\
\vspace{-11pt}\\
State: CV$_{\tn1}$ & $t$-statistic & -2.4696 & -1.3679 & 2.9780 \\
& $P$~value, $t(50)$ & 0.0170 & 0.1775 & 0.0045 \\
& $P$~value, WCR & 0.0362 & 0.2141 & 0.0238 \\
\vspace{-11pt}\\
State: CV$_{\tn3}$ & $t$-statistic &-2.2925 &-1.2203 &2.7884 \\
& $P$~value, $t(50)$ &0.0261 &0.2281 &0.0075 \\
\vspace{-11pt}\\
Region: CV$_{\tn1}$ & $t$-statistic & -2.2478 & -1.0230 & 3.1743 \\
& $P$~value, $t(8)$ & 0.0548 & 0.3362 & 0.0131 \\
& $P$~value, WCR & 0.0527 & 0.3826 & 0.0430 \\
\vspace{-11pt}\\
Region: CV$_{\tn3}$ & $t$-statistic &-2.0407 &-0.8827 &2.6715 \\
& $P$~value, $t(8)$ &0.0756 &0.4031 &0.0283 \\
\vspace{-11pt}\\
State \& year: two\kern 0.04167em-way CV$_{\tn1}$ & $t$-statistic & -2.5197 & -1.4776 & 3.4443 \\
& $P$~value, $t(14)$ & 0.0245 & 0.1617 & 0.0039 \\
& $P$~value, WCR (year) & 0.1281 & 0.2148 & 0.0034 \\
\vspace{-11pt}\\
Region \& year: two\kern 0.04167em-way CV$_{\tn1}$ & $t$-statistic & -2.2842 & -1.0999 & 3.5766 \\
& $P$~value, $t(8)$ & 0.0517 & 0.3034 & 0.0072 \\
& $P$~value, WCR (region) & 0.0736 & 0.3711 & 0.0367 \\
\bottomrule
\end{tabular*}
\vskip 6pt {\footnotesize \textbf{Notes:} There are 765 state\kern 0.04167em-year
clusters, 51 state clusters, and 9 region clusters, with
$492{,\kern 0.04167em}827$ observations in the hours dataset and
$1{,\kern 0.04167em}531{,\kern 0.04167em}360$ observations in the employment/student
dataset. The bootstrap dimension for two\kern 0.04167em-way clustering is given
in parentheses. Classic WCR bootstrap $P$~values employ $B=99,999$,
mostly using the Rademacher distribution. When bootstrapping by
region, they are calculated using Webb's six-point distribution. When
bootstrapping by year, they are calculated by enumerating the
Rademacher distribution, so that $B=32{,}768$. Obtaining the
CV$_{\tn1}$ and WCR bootstrap results in this table took 6 minutes and
12 seconds using \texttt{Stata}~16 and \texttt{boottest}~2.5.3 on one
core of an Intel~i9 processor running at 3.6~GHz.}
\end{table}
\Cref{tab:teenreg} presents the estimates of $\beta$ from
regression~\eqref{reg:wage} for all three regressands, along with
$t$-statistics and $P$~values for each of the clustering levels
considered. The coefficients are negative for the hours and employment
regressands and positive for the student one. Under the surely false
assumption of no clustering, all of these coefficients are extremely
significant, with $P$ values below $10^{-6}$. We do not compute
bootstrap $P$ values, because it would be prohibitively costly, and
they must all be very close to zero.
When we instead cluster at the state\kern 0.04167em-year level, the
$t$-statistics become smaller, especially for the employment model.
Nevertheless, clustering at this level still leads us to conclude that
all three coefficients are significant, with WCR bootstrap $P$ values
ranging from 0.0005 to 0.0145. However, some of our conclusions change
radically when we cluster at the state or region levels. For all three
outcome variables, the $t$-statistics become smaller, and the $P$
values (especially the CV$_{\tn3}$ and bootstrap ones) become larger.
The bootstrap $P$ values often differ substantially from the ones
based on the $t(G-1)$ distribution, which is expected given that the
clusters are either very unbalanced or small in number
(\Cref{subsec:emplever}). Clustering by region always yields larger
bootstrap $P$ values than clustering by state.
If instead we adopt the conservative approach of clustering at the
level with the largest standard error, we cluster at the state level
for the student model, but at the region level for the hours and
employment models. Because the number of regions is so small, however,
the latter results may not be reliable, even when we use CV$_{\tn3}$
or the WCR bootstrap. In our view, it is probably most reasonable to
cluster at the state level in all three cases.
Luckily, the choice between state and region clustering does not
matter much. In either case, we find from \Cref{tab:teenreg} that an
increase in the minimum wage is associated with a decrease in
employment, but the coefficient is not significant at any conventional
level using either state or region clustering. It is also associated
with an increased probability of being a student, which is significant
at the 5\% level regardless of clustering level. For hours worked, the
coefficient is negative. It is significant at the 5\% level for state
clustering, but not for region clustering.
The table also reports asymptotic and bootstrap $t$-statistics and $P$
values for two\kern 0.04167em-way clustering, either by state and year or by
region and year. The bootstrap method we use combines the one\kern 0.04167em-way
WCR bootstrap with the two\kern 0.04167em-way variance matrix \eqref{eq:betavar}.
This can be done in two different ways, corresponding to each of the
two clustering dimensions. Based on simulation evidence in
\citet{MNW_2021}, we bootstrap by the dimension with the smallest
number of clusters. Since two\kern 0.04167em-way clustering does not change most
of the results very much, however, there appears to be little reason
to use it in this case.
We tentatively conclude that an increase in the minimum wage is
associated with a significant decrease in hours worked and
a significant increase in the likelihood of being a student.
\citet{Seattle_2017} find a similar reduction in hours following a
minimum wage increase in Seattle, and the effects of the minimum wage
on school enrollment are discussed in \citet{NW-1995}. For employment,
we obtain a small negative effect, but it is not close to being
statistically significant when we cluster at either the state or
region levels. Thus our results are consistent with, and add support
for, the conclusions of \citet{Manning_2021} about the ``elusive''
employment effect of the minimum wage.
\begin{table}[tp]
\caption{Score\kern 0.08333em-\kern -0.08333em Variance Tests for Level of Clustering}
\label{tab:teenscore}
\vspace*{-0.5em}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}
ld{2.4}d{2.4}d{2.4}d{2.4}d{2.4}d{2.4}}
\toprule
& \multicolumn{1}{c}{Hours} & \multicolumn{1}{c}{Employed} &
\multicolumn{1}{c}{Student} & \multicolumn{1}{c}{Hours} &
\multicolumn{1}{c}{Employed} & \multicolumn{1}{c}{Student} \\
\midrule
& \multicolumn{3}{c}{H$_{\rm none}$ vs.\ H$_{\text{state\kern 0.04167em-year}}$} &
\multicolumn{3}{c}{H$_{\text{state\kern 0.04167em-year}}$ vs.\ H$_{\rm state}$} \\
\cmidrule{2-4}\cmidrule{5-7}
$\hat \tau$ statistic & 10.1108 & 18.2723 & 2.8558 & 2.2800 & 7.9453 & 3.1105 \\
$P$~value, asymptotic & 0.0000 & 0.0000 & 0.0021 & 0.0113 & 0.0000 & 0.0009 \\
$P$~value, bootstrap & 0.0000 & 0.0000 & 0.0026 & 0.0227 & 0.0000 & 0.0085 \\
\midrule
& \multicolumn{3}{c}{H$_{\rm none}$ vs.\ H$_{\rm state}$} &
\multicolumn{3}{c}{H$_{\text{state\kern 0.04167em-year}}$ vs.\ H$_{\rm region}$} \\
\cmidrule{2-4}\cmidrule{5-7}
$\hat \tau$ statistic & 10.9183 & 38.9394 & 4.6608 & 2.5226 & 11.2952 & 1.2701 \\
$P$~value, asymptotic & 0.0000 & 0.0000 & 0.0000 & 0.0058 & 0.0000 & 0.1020 \\
$P$~value, bootstrap & 0.0000 & 0.0000 & 0.0007 & 0.0279 & 0.0000 & 0.1388 \\
\midrule
& \multicolumn{3}{c}{H$_{\rm none}$ vs.\ H$_{\rm region}$} &
\multicolumn{3}{c}{H$_{\rm state}$ vs.\ H$_{\rm region}$} \\
\cmidrule{2-4}\cmidrule{5-7}
$\hat \tau$ statistic & 9.4096 & 49.6559 & 2.7201 & 0.6757 & 2.2566 & -0.3206 \\
$P$~value, asymptotic & 0.0000 & 0.0000 & 0.0033 & 0.2496 & 0.0120 & 0.6257 \\
$P$~value, bootstrap & 0.0000 & 0.0120 & 0.0000 & 0.2662 & 0.0410 & 0.5497 \\
\bottomrule
\end{tabular*}
\vskip 6pt
{\footnotesize \textbf{Notes:} There are 765 state\kern 0.04167em-year clusters, 51
state clusters, and 9 region clusters. The test statistic $\hat\tau$
is asymptotically distributed as ${\rm N}(0,1)$. All $P$ values are for
one\kern 0.04167em-sided tests. The bootstrap $P$~values were
calculated with $B=99,999$.}
\end{table}
To confirm the choice of clustering level, we use the
score\kern 0.04167em-variance tests of \citet{MNW-testing} to test for the
appropriate level of clustering.\footnote{We do not calculate the
tests proposed in \citet{IM_2016}, because \eqref{reg:wage} cannot be
estimated on a cluster-by-cluster basis for state\kern 0.04167em-level or
state\kern 0.04167em-year clusters without dropping some regressors.}
\Cref{tab:teenscore} presents results from six tests for each of the
three models. Following a systematic, sequential testing approach, we
would test independence vs.\ state\kern 0.04167em-year, then state\kern 0.04167em-year vs.\
state, and finally state vs.\ region. Apart from the possibility of
two\kern 0.04167em-way clustering, we would conclude that the appropriate level
of clustering is that of the first non-rejected null hypothesis. For
the hours model, the score\kern 0.04167em-variance tests marginally favor region
clustering over state clustering, but, for the employment and student
models, they clearly favor state clustering. Thus, these tests lend
support to our decision to cluster at the state level; see the
discussion of \Cref{tab:teenreg}.
\subsection{Cluster Heterogeneity}
\label{subsec:emplever}
\begin{table}[tp]
\caption{Summary Statistics for Cluster Heterogeneity}
\label{tab:clustsum}
\vspace*{-0.5em}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}
ld{4.0}d{4.2}d{7.0}d{6.0}d{6.0}d{6.0}d{6.0}d{6.0}}
\toprule
Clustering & \multicolumn{1}{c}{$G$} & \multicolumn{1}{c}{$G_\beta^*(0)$}
& \multicolumn{1}{c}{$\bar{N}_g$} & \multicolumn{1}{c}{min.} &
\multicolumn{1}{c}{$1^{\rm st}$ quart.} & \multicolumn{1}{c}{median} &
\multicolumn{1}{c}{$3^{\rm rd}$ quart.} & \multicolumn{1}{c}{max.} \\
\midrule
\multicolumn{8}{l}{Hours data: $N = 492{,}827$} \\
\midrule
State\kern 0.04167em-year &765 &79.4 &644 &6 &176 &480 &860 &3{,}052 \\
State &51 &16.2 &9{,}663 &258 &2{,}495 &7{,}082 &13{,}481 &35{,}995 \\
Year &15 &6.6 &32{,}855 &28{,}262 &28{,}839 &30{,}733 &40{,}224 &40{,}394 \\
Region &9 &7.5 &54{,}759 &27{,}849 &37{,}396 &50{,}489 &65{,}389 &96{,}337 \\
\midrule
\multicolumn{8}{l}{Employment and student data: $N = 1{,}531{,}360$} \\
\midrule
State\kern 0.04167em-year &765 &66.0 &2{,}002 &42 &524 &1{,}413 &2{,}426 &10{,}794 \\
State &51 &13.1 &30{,}027 &927 &7{,}363 &22{,}845 &37{,}020 &144{,}914 \\
Year &15 &6.5 &102{,}091
&92{,}701 &95{,}589 &102{,}319 &108{,}858 &110{,}528 \\
Region &9 &7.0 &170{,}151
&74{,}172 &104{,}703 &181{,}767 &208{,}099 &291{,}955 \\
\bottomrule
\end{tabular*}
\vskip 6pt
{\footnotesize \textbf{Notes:} The values of $G_\beta^*(0)$
are calculated using 28 regressors after the state dummies have been
partialed out. The $\beta$ subscript emphasizes the fact that they
correspond to the coefficient $\beta$ in \eqref{reg:wage}. Because
there are state fixed effects, values of $G_\beta^*(1)$ are not
reported; see \citet{MNW-influence}.}
\end{table}
As we discussed in \Cref{sec:report}, investigators should be
suspicious of any results that are overly dependent on very few
clusters. \Cref{tab:clustsum} reports several summary statistics for
cluster heterogeneity. Specifically, it reports $G$, the number of
clusters and $G_\beta^*(0)$, the effective number of clusters
(\Cref{sec:report}), as well as the average, minimum, first and third
quartiles, median, and maximum of the $N_g$. These statistics suggest
that the state\kern 0.04167em-year clusters are extremely unbalanced. Although
there are $G=765$ clusters, the effective number $G_\beta^*(0)$ is
smaller than $G$ by a factor of roughly 10 to 12. The maximum cluster
size is over six times the median, and the third quartile is nearly
twice the median.
From \Cref{tab:clustsum}, we also see that the state clusters are
extremely unbalanced, based both on their sample sizes and on the
values of $G_\beta^*(0)$. The region clusters are fairly balanced in
terms of their sample sizes, with the maximum $N_g$ only about four
times the minimum. The values of $G_\beta^*(0)$ are also not
worrisome. The year clusters (which we only use in two\kern 0.04167em-way
clustering) are very balanced in terms of their sample sizes but much
less so in terms of $G^*(0)$. We would expect asymptotic tests based
on CV$_{\tn1}$ to perform rather poorly in all cases, because there is
a lot of cluster heterogeneity for clustering by state\kern 0.04167em-year and
state, a small number of clusters for clustering by region, and a
small number of effective clusters for clustering by year.
For each of the three levels of one\kern 0.08333em-way clustering, we calculate
measures of leverage, partial leverage, and influence using the
\texttt{summclust} package in \texttt{Stata} \citep{MNW-influence}.
However, we only report results for state\kern 0.04167em-level clustering. Since
state\kern 0.04167em-year clustering is strongly rejected against state\kern 0.04167em-level
clustering (\Cref{tab:teenscore}), it does not seem interesting to
discuss it. We also do not discuss region-level clustering, because
there are just nine regions, and there is not much evidence in favor
of clustering at that level.
\begin{figure}[tb]
\begin{center}
\caption{\label{fig:EDF}{Empirical Distribution Functions for 51 State
Clusters}}
{\includegraphics[width=1.00\textwidth]{gfigAv3.pdf}}
\end{center}
\textbf{Notes:} In both panels, each of the three staircase curves is the EDF
for the $N_g$, the $L_g$, or the $L_{g\beta}$. They were plotted in
that order. When they overlap, only the last one to be plotted is
visible. The $N_g$ and the $L_g$ have been rescaled so that they sum
to unity, like the $L_{g\beta}$.
\end{figure}
The empirical distributions of the $N_g$, the $L_g$, and the
$L_{g\beta}$ (the partial leverages) are shown in \Cref{fig:EDF},
where the $N_g$ and the $L_g$, which actually sum to $N$ and $k$,
respectively, have been rescaled so that they sum to unity. In a
perfectly balanced case, where all of the $N_g$ (or $L_g$ or
$L_{g\beta}$) are identical, these EDFs would simply be vertical lines
at~$1/G$. Both datasets are evidently far from being perfectly
balanced. The $L_g$ are essentially proportional to the $N_g$ and
extremely highly correlated with them; the correlations are 0.9986 for
the employment/student dataset and 0.9982 for the hours dataset. This
is evident in both panels, where the EDFs for the $N_g$ and the $L_g$
are extremely similar. Thus, for leverage, it appears that there is no
heterogeneity in the clusters other than what is implied by the
cluster sizes.
In contrast, the partial leverages vary much more across states than
do the other two measures. The largest values of the $L_{g\beta}$ are
considerably larger than the largest values of the $N_g$ and the
$L_g$, especially for the employment/student dataset; note that all of
these are for California. Not coincidentally, more of the $L_{g\beta}$
than of the other measures are smaller than $1/G$, the average. The
correlations between the $L_{g\beta}$ and the $L_g$ are 0.9150 for the
student/employment dataset and 0.9165 for the hours dataset. The
correlations between the $L_{g\beta}$ and the $N_g$ are also very
similar for the two datasets, at 0.9209 and 0.9163.
The substantial amount of cluster heterogeneity that is evident in
both \Cref{tab:clustsum} and \Cref{fig:EDF} suggests that inference
based on CV$_{\tn1}$ and the $t(50)$ distribution may not be reliable.
Further evidence on this point will be provided in
\Cref{subsec:emp-placebo}.
For the student model, we find no evidence of influential clusters.
The $\hat \beta^{(g)}$ range from 0.00185 to 0.00246, with $\hat \beta
= 0.00221$. On the other hand, for the employment model, there are two
noticeable values of~$\hat\beta^{(g)}$: Texas has $\hat\beta^{(g)} =
-0.00529$ and New York has $\hat\beta^{(g)} = -0.00203$, with $\hat\beta
= -0.00367$ and all the remaining $\hat\beta^{(g)}$ in the interval
$[-0.00455 ; -0.00302]$. For comparison, the standard error of
$\hat\beta$ is $0.00268$.
For the hours model, four values of the $\hat\beta^{(g)}$ stand out.
They are Minnesota ($-0.1776$) and Texas ($-0.1770$) at one end
and Arizona ($-0.1274$) and New York ($-0.1227$) at the other. For
comparison, $\hat\beta = -0.1539$, and all the other $\hat\beta^{(g)}$
lie in the interval $[-0.1713; -0.1406]$. In this case, the standard
error of $\hat\beta$ is $0.0623$.
\subsection{Placebo Regressions for the Empirical Example}
\label{subsec:emp-placebo}
As we discussed in \Cref{subsec:placebo}, placebo regressions provide
a simple way to check the level at which the residuals are clustered,
even when the pattern of intra-cluster correlation is unknown and
perhaps very complicated. If a placebo regressor is clustered at, say,
the state level, then the empirical scores will also be clustered at
that level unless the residuals are clustered only at a finer level.
When the residuals display intra-cluster correlation at the state
level, we would therefore expect placebo\kern 0.04167em-regression tests with
standard errors clustered at that level to reject roughly as often as
they should. But we would expect placebo regressions with standard
errors clustered at finer levels to over-reject.
We perform two sets of placebo\kern 0.04167em-regression experiments for each of
the three equations estimated in \Cref{tab:teenreg}. In the first set,
the placebo regressor is a DiD-style treatment dummy similar to the
ones used in \citet{BDM_2004} and the other papers cited in
\Cref{subsec:placebo}. Treatment is randomly applied to 15, 20, 25,
30, or 35 states. For each state, it begins randomly in any year
excluding 2005 (to avoid collinearity with the state fixed effects)
and continues through 2019.\footnote{Of course, this sort of DiD model
with two\kern 0.04167em-way fixed effects is somewhat obsolete; see
\cite{CS_2021} and other papers cited therein. But we would still
expect to find no effects for placebo treatments beyond those
attributable to chance.} Rejection percentages are shown in the top
panel of \Cref{tab:placebo}. The first number in each pair is the
smallest rejection percentage observed over the five experiments for
each equation, and the second number is the largest one. The numbers
of treated and control states are deliberately chosen not to be
extreme, so as to avoid the issues discussed in \Cref{subsubsec:few}.
\begin{table}[tp]
\caption{Rejection Percentages for Placebo Regressions}
\label{tab:placebo}
\vspace*{-0.5em}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}
lcd{2.1}d{2.1}cd{2.1}d{2.1}cd{2.1}d{2.1}cd{2.1}d{2.1}cd{2.1}d{2.1}cd{2.1}d{2.1}}
\toprule
&& \multicolumn{8}{c}{DiD-type treatment} &&
\multicolumn{8}{c}{State\kern 0.04167em-level AR(1) component}\\
\cmidrule{3-10}\cmidrule{12-19}
Method && \multicolumn{2}{c}{Hours} &&
\multicolumn{2}{c}{Employed} && \multicolumn{2}{c}{Student} &&
\multicolumn{2}{c}{Hours} && \multicolumn{2}{c}{Employed} &&
\multicolumn{2}{c}{Student}\\
\midrule
HC$_1$, ${\rm N}(0,1)$ && 45.9 &48.8 &&61.2 &62.3 &&42.4 &43.7
&&37.8 &26.0 &&54.3 &44.4 &&28.6 &21.7 \\
\midrule
State\kern 0.04167em-year\\
CV$_{\tn1}$, $t(764)$ &&24.8 &27.5 &&30.6 &32.2 &&28.6 &29.6
&&16.4 &16.3 &&21.1 &25.2 &&14.6 &16.2 \\
CV$_{\tn3}$, $t(764)$ &&19.3 &21.3 &&24.2 &25.9 &&23.2 &24.5
&&11.4 &13.3 &&15.4 &20.5 &&10.0 &13.7 \\
CV$_{\tn1}$, WCR &&20.6 &22.7 &&25.7 &27.1 &&25.1 &26.3
&&13.2 &14.6 &&17.3 &22.5 &&11.8 &14.6 \\
\midrule
State\\
CV$_{\tn1}$, $t(50)$ &&7.1 &9.3 &&7.7 &10.0 &&7.0 &8.4
&&7.5 &6.6 &&9.0 &8.3 &&6.5 &6.9 \\
CV$_{\tn3}$, $t(50)$ &&5.2 &6.6 &&5.3 &6.6 &&4.9 &6.0
&&5.5 &5.1 &&5.8 &6.0 &&4.8 &5.4 \\
CV$_{\tn1}$, WCR &&4.8 &5.7 &&4.6 &5.3 &&5.1 &5.8
&&5.2 &4.8 &&5.3 &5.5 &&5.1 &5.4 \\
\midrule
Region\\
CV$_{\tn1}$, $t(8)$ &&7.6 &8.5 &&7.3 &8.1 &&7.8 &9.1
&&7.6 &7.2 &&8.2 &8.0 &&7.8 &7.2 \\
CV$_{\tn3}$, $t(8)$ &&4.6 &5.3 &&4.5 &5.5 &&4.9 &5.4
&&5.1 &5.0 &&5.3 &5.2 &&4.9 &4.9 \\
CV$_{\tn1}$, WCR &&5.4 &6.2 &&4.9 &5.7 &&6.0 &6.8
&&6.2 &6.0 &&6.0 &6.1 &&6.1 &5.6 \\
\bottomrule
\end{tabular*}
\vskip 6pt
{\footnotesize \textbf{Notes:} The numbers are rejection percentages
at the nominal $5\%$ level based on 10,\tkk000 simulations when a
single placebo regressor is added to \eqref{reg:wage}. For the
DiD-type treatment, the first number in each pair is the smallest
rejection percentage over all parameter values used to simulate the
placebo regressor, and the second is the largest; see text. For the
state\kern 0.04167em-level AR(1) component, the first number is for $\rho=0.5,
\delta=0.9$, and the second is for $\rho=0.8, \delta=0.5$. There are
765 state\kern 0.04167em-year clusters, 51 state clusters, and 9 region clusters.
The classic WCR bootstrap uses $B=999$.}
\end{table}
For the second set of simulation experiments, the placebo regressor is
generated by
\begin{equation}
x_{ist} = \delta\kern 0.04167em v_{st} + (1-\delta) \epsilon_{ist},
\quad v_{st} = \rho\kern 0.04167em v_{s,t-1} + e_{st}, \;\; 0 \le \rho < 1, \;\;
0 \le \delta \le 1,
\label{xprocess}
\end{equation}
where the $\epsilon_{ist}$ and the $e_{st}$ are independent standard
normals. Thus the $v_{st}$ are 51 separate stationary AR(1) processes,
and $x_{ist}$ is a weighted average of $v_{st}$ and $\epsilon_{ist}$.
When either $\rho=0$ or $\delta=0$, the $x_{ist}$ are independent.
When both $\rho$ and $\delta$ are positive, they are correlated within
both state\kern 0.04167em-years and states. They are never correlated across
states within regions. The extent to which the $x_{ist}$ are
correlated within state\kern 0.04167em-years and states depends on both of the
parameters in \eqref{xprocess}. For simplicity, we report rejection
percentages for just two cases. In the first case, $\rho=0.5$ and
$\delta=0.9$, so there is a lot of correlation within each
state\kern 0.04167em-year. In the second case, $\rho=0.8$ and $\delta=0.5$, so
there is less correlation within state\kern 0.04167em-years but more correlation
across years within each state.
The first row in \Cref{tab:placebo} shows that failing to cluster
always leads to severe over-rejection, and the next three rows show
that clustering at the state\kern 0.04167em-year level also leads to substantial
over-rejection. This is true for all methods of inference and for both
DGPs, but more so for the DiD-type one. Rows 5 through 7 show that
there can be noticeable over-rejection for clustering at the state
level when using CV$_{\tn1}$ with $t(G-1)$ critical values, especially
for the DiD-type DGP. This is to be expected given the unbalanced
cluster sizes at the state level. Both CV$_{\tn3}$ with $t(G-1)$
critical values and the classic WCR bootstrap perform much better. For
the AR(1) placebo regressors, both CV$_{\tn3}$ and the WCR bootstrap
perform very well indeed when there is clustering at either the state
or region level. For both designs, the performance of CV$_{\tn3}$ at
the region level is remarkably good. However, the WCR bootstrap tends
to outperform it slightly at the state level.
The placebo\kern 0.04167em-regression results in \Cref{tab:placebo} are largely
consistent with those of the score\kern 0.04167em-variance tests. They suggest
that the CV$_{\tn3}$ and classic WCR bootstrap results with
state\kern 0.04167em-level clustering in \Cref{tab:teenreg} can probably be
relied upon, but that the results without clustering or with
state\kern 0.04167em-year clustering should not be believed.
\section{Conclusion: A Summary Guide}
\label{sec:conc}
We conclude by presenting a brief summary guide. This is essentially a
checklist for cluster-robust inference in regression models. The first
three items should be dealt with as soon as possible in the process of
specifying and estimating a model. The remaining ones should be kept
in mind throughout the process of estimation and inference.
\begin{enumerate}
\item List all plausible clustering dimensions and levels for the data
at hand and make an informed decision regarding the clustering
structure. The decision may depend on what is to be estimated and why.
Formal tests (\Cref{subsec:level}) can be helpful in making this
decision. In some cases, placebo regressions
(\Cref{subsec:placebo,subsec:emp-placebo}) may also be informative. A
conservative approach is simply to choose the structure with the
largest standard error(s) for the coefficient(s) of interest, subject
to the number of clusters not being so small that inference risks
being unreliable.
\item For each of the plausible levels of clustering, report the number
of clusters, $G$, and a summary of the distribution of the cluster
sizes (the $N_g$). We suggest reporting at least the minimum, maximum,
mean, and median of the $N_g$. These can be reported in tabular form,
as in \Cref{tab:clustsum}. It may also be helpful to present the
distribution(s) of the $N_g$ graphically using box plots, histograms,
or EDFs like the ones in \Cref{fig:EDF}.
\item For the key regression specification(s) considered, report
information about leverage, partial leverage, and influence
(\Cref{subsec:leverage,sec:report,subsec:emplever}), including the
effective number(s) of clusters for the coefficient(s) of interest.
This may be particularly informative for difference\kern 0.04167em-in-differences
and other treatment models. Inferences may not be reliable when a few
clusters are highly influential or have high (partial) leverage.
\item In addition to, or instead of, the usual CV$_{\tn1}$ CRVE,
employ the CV$_{\tn3}$ CRVE and at least one variant of the restricted
wild cluster (WCR) bootstrap (\Cref{subsec:wild}) as a matter of
course for both tests and confidence intervals. In many cases,
especially when $G$ is reasonably large and the clusters are fairly
homogeneous, these methods will yield very similar inferences that can
likely be relied upon. However, in the event that they differ, it
would be wise to try other methods as well, including additional
variants of the WCR bootstrap and some of the alternative methods
discussed in \Cref{sec:other}.
\item For models with treatment at the cluster level, where either the
treated clusters or the controls are few in number and/or atypical,
cluster-robust inference can be quite unreliable
(\Cref{subsec:failure}), even when it is based on CV$_{\tn3}$ or the
WCR bootstrap. In such cases, it is important to verify that the
results are (or perhaps are not) robust. This can often be done by
using methods based on randomization inference (\Cref{subsec:RI}).
\end{enumerate}
\setlength{\bibsep}{1.5pt}
|
\section{Introduction}
\label{sec: introduction}
Environment perception in Autonomous Vehicles (AV) is a challenging problem. With the current approach of using only on-board sensors to solve the perception problem, it is impossible to sense occluded areas and mitigate the effects of sensor outage. Complex traffic intersections with buildings close to the curb may minimize the field of view of an AV's sensors. Integration of smart infrastructure nodes (sensing and compute) on roads where AVs operate can help overcome these challenges. The elevated and static view-point of the smart sensors enables them to observe the environment, detect more objects in the scene, and communicate that information to AVs. AVs can fuse that information with their own sensor measurements and augment their situational awareness. Fisheye cameras are well known for their low cost and wide Field of View (FoV), making them suitable for such smart infrastructure based sensing applications. However, the fisheye camera needs to provide this information in the same coordinate frame as the vehicle. For this reason, they need to be localized or registered within the same map that is used by the AV for navigation. In this work, we propose a method to localize a downward looking static smart infrastructure fisheye camera in a prior map consisting of a metric satellite image, and a co-registered LiDAR map of ground points with their LiDAR reflectivity values. An overview of the approach is shown in Figure \ref{fig: cameraregistrationinpriormap}.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.4\textwidth]{figures/camera_registration_inpriormap.png}
\caption{Overview of our two step approach to camera localization. The coordinates shown are in meter, wrt the origin of the map. We start with a noisy GPS initialization (cyan), then perform feature matching between satellite image and rectified fisheye image to obtain an inital camera pose - PnP Localization (Green), and finally maximize the Mutual Information between LiDAR map and the fisheye image to obtain a refined camera localization estimate.}
\label{fig: cameraregistrationinpriormap}
\vspace{-20pt}
\end{figure}
\section{Related Work}
\label{sec: relatedwork}
There are several contributions on localization of camera images in prior maps (satellite imagery or LiDAR generated 3D maps). Satellite maps can be procured easily from third party sources \cite{vora2020aerial, maxar}, and with the widespread use of LiDARs in autonomous driving, we now have several high definition map providers which provide dense 3D map of the environment \cite{vora2019high}. This has made localization of cheaper sensors like cameras on prior maps an active area of research with an ultimate aim of using cheaper sensors on-board an AV. \cite{Yu2020MonocularCL} presents monocular perspective camera localization in pre-built 3D LiDAR maps using 2D-3D line correspondences. This method shows promise only in structured environment where lines can be easily detected in both camera and the LiDAR map. \cite{Wolcott2014VisualLW} presents LiDAR map based monocular camera localization in urban environment. Unlike \cite{Yu2020MonocularCL} which depends on detection of geometric primitives like lines in both the sensing modalities, \cite{Wolcott2014VisualLW} uses dense appearance based approach which work in unstructured environments. In \cite{Wolcott2014VisualLW}, given an initial belief of the camera pose, they generate several synthetic views of the environment by projecting the LiDAR map points using a perspective camera model, and compare these synthetic views against the live camera feed. The synthetic view which maximizes the Normalized Mutual Information (NMI) between the real image gray scale values and the projected points' LiDAR reflectivity values, is the solution to the localization problem. \cite{Wolcott2014VisualLW} draws inspiration from work done on 3D-LiDAR Camera extrinsic calibration described in \cite{MIGP}, which uses maximization of Mutual Information (MI) for calibrating a 3D-LiDAR Camera pair. \cite{Viswanathan} presents a camera localization technique which matches ground imagery obtained by cameras onboard an AV to the available satellite imagery. The camera images are warped to obtain a bird's eye view (BEV) of the ground. Next, the BEV image is matched with the given satellite imagery using using SIFT \cite{Lowe:2004:DIF:993451.996342} features.
While the above mentioned approaches provide solutions for perspective cameras, our focus is to localize (estimate $[\mathbf{^{C}R_W}, \mathbf{^{C}t_W}]$ in Equation \ref{eqn: fisheyeprojectionmodel}) a downward looking static fisheye camera (Figure \ref{fig: fisheyeimage}) in a prior map (Figure \ref{fig: priormap}) which consists of 2D satellite imagery with metric information (Figure \ref{fig: satimage}) and 3D LiDAR map of ground points (Figures \ref{fig: intensityimage}, \ref{fig: zheightimage}). We initialize the camera localization using feature matching as done in \cite{Viswanathan}, and refine the camera localization using maximization of a MI based cost function as done in \cite{Wolcott2014VisualLW} and \cite{MIGP}.
\section{Overview}
\label{sec: overview}
In this section we provide an overview of the various components of our implementation.
\subsection{Fisheye Camera}
\label{sec: fisheyecamera}
\vspace{-30pt}
\begin{figure}[H]
\centering
\subfloat[Fisheye image captured at a traffic intersection in our simulator]{\includegraphics[width=0.225\textwidth]{figures/fisheyesampleimage.png}\label{fig: fisheyeimage}}
\quad
\subfloat[Rectification of fisheye image to perspective image]{\includegraphics[width=0.225\textwidth]{figures/perspectifiedfisheyeimage.png}\label{fig: perspectifiedfisheyeimage}}
\caption{Fisheye Image (Figure \ref{fig: fisheyeimage}) and its perspective rectification (Figure \ref{fig: perspectifiedfisheyeimage})}
\label{fig: fisheyevsperspectiveprojection}
\vspace{-10pt}
\end{figure}
We use the fisheye camera projection model proposed in \cite{MeiRives} to project 3D point $\mathbf{P_{W}} = [X_W, Y_W, Z_W]$ defined in the prior map coordinate frame $\mathbf{W}$ to a 2D point $\mathbf{p}$ on the fisheye camera image plane using Equation \ref{eqn: fisheyeprojectionmodel}.
\begin{align}
\textbf{p} &= \mathbf{\Pi}(K, D, \xi, [\mathbf{^{C}R_W}, \mathbf{^{C}t_W}], \mathbf{P_{W}})
\label{eqn: fisheyeprojectionmodel}
\end{align}
Here $\mathbf{\Pi}()$ is the projection function, $K=\begin{bmatrix}
f_x & s & c_x \\
0 & f_y & c_y \\
\end{bmatrix}$, $D = [k_1, k_2, p_1, p_2]$ and $\xi$ are the camera intrinsics, and $[\mathbf{^{C}R_W}$, $\mathbf{^{C}t_W}]$ is the camera extrinsic. $\mathbf{^{C}R_W}$ $\in SO(3)$ is an orthonormal rotation matrix and $\mathbf{^{C}t_W}$ $\in R^{3}$ is a 3D vector. The goal of this work is to estimate the unknown camera extrinsic $[\mathbf{^{C}R_W}$, $\mathbf{^{C}t_W}]$ in the prior map frame $\mathbf{W}$.
\subsubsection{Intrinsic Calibration} \label{sec: intrinsiccalibration} We estimate the intrinsic parameters $K$, $D$ and $\xi$ by collecting several images of a large checkerboard at different poses, and feeding those images to the omnidirectional camera calibrator in OpenCV\cite{opencv_library}, which provides an implementation\footnote{\url{https://docs.opencv.org/4.5.2/dd/d12/tutorial_omnidir_calib_main.html}} of the intrinsic calibration technique presented in \cite{boli}.
\subsubsection{Rectification} \label{sec: perspectiverectification} The intrinsic camera calibration parameters are used to rectify the fisheye image (Figure \ref{fig: fisheyeimage}) into its corresponding perspective image (Figure \ref{fig: perspectifiedfisheyeimage}) utilizing OpenCV's rectification routines. Although perspective rectification results in loss of field of view, it makes the application of computer vision algorithms developed for perspective images possible for fisheye images.
\subsection{Prior Map}
\label{sec: priormap}
\begin{figure*}[!ht]
\centering
\subfloat[Satellite Map]{\includegraphics[width=0.3\textwidth]{figures/satellite_image_new.png}\label{fig: satimage}}
\quad
\subfloat[LiDAR Map: Reflectivity of Ground Points]{\includegraphics[width=0.3\textwidth]{figures/intensity_image_new.png}\label{fig: intensityimage}}
\quad
\subfloat[LiDAR Map: Height of Ground Points]{\includegraphics[width=0.3\textwidth]{figures/height_image_new.png}\label{fig: zheightimage}}
\caption{\textbf{Prior Map:} Figures \ref{fig: satimage}, \ref{fig: intensityimage} $\&$ \ref{fig: zheightimage} show the components of our prior map. The full map is not shown in the interest of space.}
\label{fig: priormap}
\end{figure*}
The prior map (Figure \ref{fig: priormap}) consists of two important components which are registered and expressed in the frame of reference $\mathbf{W}$. They are:
\subsubsection{Satellite Map} \label{sec: satmap}
The satellite map is a metric Bird's Eye View (BEV) satellite image (Figure \ref{fig: satimage}). In our case, a pixel on the image corresponds to 0.1 m on the ground.
\subsubsection{LiDAR Map} \label{sec: LiDARmap}
The prior LiDAR map is built using an offline mapping process described in \cite{Wolcott2014VisualLW}. Broadly, a survey vehicle equipped with several 3D LiDAR scanners and a high end inertial navigation system is manually driven and sensor data is collected in the environment we want to map. Next, an offline pose-graph optimization SLAM (Simultaneous Localization and Mapping) problem is solved to obtain the accurate global pose of the vehicle. Finally, a dense ground point mesh is constructed from the optimized pose graph using region growing techniques which gives a dense 3D point cloud map. The ground points from this dense 3D cloud are used to generate the LiDAR ground reflectivity image (Figure \ref{fig: intensityimage}) and ground height image (Figure \ref{fig: zheightimage}). The LiDAR Map (Figures \ref{fig: intensityimage} and \ref{fig: zheightimage}) is aligned with the satellite imagery (Figure \ref{fig: satimage}) using the GPS measurements from the inertial navigation system.
\section{Problem Formulation}
\label{sec: problemformulation}
\begin{figure*}[!ht]
\centering
\includegraphics[width=\textwidth]{figures/FisheyeRegistrationBlockDiagram.png}
\caption{\textbf{Overview of the method:} The block diagram shows the two steps involved in our approach. In Step 1, we match features between the perspective projection of fisheye image and a cropped satellite map to initialize camera pose, and in Step 2 we refine this initialization by maximization of Mutual Information between fisheye image and prior 3D-LiDAR map.}
\label{fig: fisheyeregistrationblockdiagram}
\vspace{-20pt}
\end{figure*}
The goal of this work is to estimate the unknown camera pose $[\mathbf{^{C}R_W}$, $\mathbf{^{C}t_W}]$ in the prior map frame $\mathbf{W}$. We assume that we have a noisy estimate of the fisheye camera's (GPS) position (no orientation) in the map which helps us reduce the search space in the prior map. We follow a two step approach to register the camera in the prior map, the details of which are presented in Sections \ref{sec: initsparsefeaturematching} and \ref{sec: refinement_of_localization}, and a broad overview is provided in Figure \ref{fig: fisheyeregistrationblockdiagram}.
\subsection{Initialization using sparse feature matching}
\label{sec: initsparsefeaturematching}
Traditionally available feature detection, description and matching techniques are usually suitable for perspective images only. Therefore, we rectify the fisheye image into the corresponding perspective image as explained in Section \ref{sec: perspectiverectification}, and use SuperGlue \cite{sarlin20superglue}, a pre-trained deep learning based feature matching algorithm, for matching features (Figure \ref{fig: featurematches}) between the rectified fisheye image and the cropped satellite image (cropped using GPS initialization, refer Figure \ref{fig: fisheyeregistrationblockdiagram}). The matched features are used to solve a Perspective-n-Point (PnP) problem \cite{P3P, RANSAC} to estimate the initial pose of camera (also called the PnP estimate) in the prior map reference frame $\mathbf{W}$. As we know the metric scale of the satellite image (1 pixel = 0.1 m), we obtain the camera pose in metric units.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth]{figures/superglue_matches.png}
\caption{SuperGlue \cite{sarlin20superglue} Matching between Satellite Image (Left) $\&$ perspective projection of fisheye image (Right)}
\label{fig: featurematches}
\vspace{-15pt}
\end{figure}
\subsection{Refinement of camera localization using Maximization of Mutual Information}
\label{sec: refinement_of_localization}
Mutual Information (MI) has been used in several fields for registering data from multi-modal sensors \cite{paulviola, MaesMI}. We refine the initial camera pose estimate from Section \ref{sec: initsparsefeaturematching} by maximizing the Mutual Information (MI) between the LiDAR reflectivity of ground points and the fisheye grayscale values at the pixel locations onto which the LiDAR points are projected using the camera pose $[\mathbf{^{C}R_W}, \mathbf{^{C}t_W}]$.
\subsubsection{Theory} MI (Equation \ref{eqn: mutualinformationdefinition}) provides a way to statistically measure mutual dependence between two random variables $X$ and $Y$.
\begin{equation}
MI(X, Y) = H(X) + H(Y) - H(X, Y)
\label{eqn: mutualinformationdefinition}
\end{equation}
Where $H(X)$ and $H(Y)$ are the Shannon entropy over random variables $X$ and $Y$ respectively, and $H(X, Y)$ is the joint Shannon entropy over the two random variables:
\begin{align}
H(X) &= - \sum_{x \in X} p_{X}(x) \log p_{X}(x) \label{eqn: marginalX}\\
H(Y) &= - \sum_{y \in X} p_{Y}(y) \log p_{Y}(y) \label{eqn: marginalY}\\
H(X, Y) &= - \sum_{x \in X} \sum_{y \in Y} p_{X, Y}(x, y) \log p_{XY}(x, y) \label{eqn: jointXY}
\end{align}
The entropy $H(X)$ of a random variable $X$ denotes the amount of uncertainty in $X$, whereas $H(X, Y)$ is the amount of uncertainty when the random variables $X$ and $Y$ are co-observed. The formulation of MI in Equation \ref{eqn: mutualinformationdefinition} shows that maximization of $MI(X, Y)$ is achieved by minimization of the joint entropy $H(X, Y)$, which coincides with minimization of dispersion of two random variable's joint histogram.
\subsubsection{Mathematical Formulation}
\label{sec: mathematicalformulation}
Let $\{\mathbf{P_{W_{i}}}; i = 1,2, \dots, n\}$ be the set of 3D points whose coordinates are known in the prior map reference frame $\mathbf{W}$ and let $\{X_{i}; i = 1,2, \dots, n\}$ be the corresponding reflectivity values for these points ($X_i \in [0, 255]$). Equation \ref{eqn: fisheyeprojectionmodel} presents the relationship between $\mathbf{P_{W_{i}}}$ and its image projection $\mathbf{p_{i}}$ as a function of $[\mathbf{^{C}R_W}, \mathbf{^{C}t_W}]$.
Let $\{Y_i; i = 1, 2, \dots, n\}$ be the grayscale intensity of the pixels $\mathbf{p_{i}}$ where ${\mathbf{P_{W_{i}}}}$ project onto, such that:
\begin{align}
Y_i = I(\mathbf{p_{i}})
\end{align}
where $Y_i \in [0, 255]$ and $I$ is the grayscale fisheye image. Therefore, $X_i$ is an observation of the random variable $X$, and for a given $[\mathbf{^{C}R_W}, \mathbf{^{C}t_W}]$, $Y_i$ is an observation of random variable $Y$. The marginal ($p_{X}(x)$, $p_{Y}(y)$) and joint ($p_{X, Y}(x, y)$) probabilities of the random variables $X$ and $Y$, required for calculating MI (Equation \ref{eqn: mutualinformationdefinition}), can be estimated using a normalized histogram (Equation \ref{eqn: pdfdefinition}):
\begin{align}
\hat{p}(X = k) = \frac{x_k}{n}, k \in [0, 255]
\label{eqn: pdfdefinition}
\end{align}
where $x_k$ is the observed counts of the intensity value $k$.
\subsubsection{Global Optimization}
$\mathbf{^CR_W} \in SO(3)$ is an orthonormal rotation matrix which can be parameterized as Euler angles $[\phi, \theta, \psi]^{\top}$ and $\mathbf{^{C}t_{W}} = [x, y, z]^{\top}$ is an Euclidean 3-vector. $\psi$ is the rotation of the camera along its principal axis. In our context, the fisheye camera is facing vertically downward so we do not refine the $\phi$ $\&$ $\theta$ and leave it at what the feature matching based technique (Section \ref{sec: initsparsefeaturematching}) determines it to be, which is very close to 0. Therefore, as far as rotation variables are concerned, we refine only $\psi$. We represent all the variables to be optimized together as $\Theta = [x, y, z, \psi]^{\top}$. The optimization is posed as a maximization problem:
\begin{align}
\hat{\Theta} = \argmax_{\Theta} MI(X, Y; \Theta)
\label{eqn: globaloptimization}
\end{align}
\section{Experiments and Results}
\label{sec: experimentsandresults}
This section describes the experiments performed to evaluate the proposed technique using data obtained from both our simulator and real world sensor.
\subsection{Simulation Studies}
\label{sec: simulationstudies}
We first validate our approach on a simulator which is built using data from real sensors. The Mathworks' tool RoadRunner \cite{roadrunner} is used to generate the 2D features like lane geometry and lane markings with the satellite map used as a reference. The 3D structures are created using the Unreal Engine Editor \cite{unrealeditor} with the help of real satellite and 3D LiDAR maps. Since the simulated environment is created using the prior map components, it can safely be assumed that the simulator aligns with the real world to a high degree of accuracy. In order to generate the fisheye images, we model a fisheye camera in Unreal Engine using the equidistant model with a field of view of 180 degrees. We demonstrate our approach in simulation for the fisheye image shown in Figure \ref{fig: fisheyeimage}. The fisheye image is first rectified (Section \ref{sec: perspectiverectification}) to generate Figure \ref{fig: perspectifiedfisheyeimage}, which is used for estimating the initial camera pose using the approach in Section \ref{sec: initsparsefeaturematching}. Next, the initialization is refined using maximization of MI (Section \ref{sec: refinement_of_localization}).
\vspace{-20pt}
\begin{figure}[!ht]
\centering
\subfloat[]{\includegraphics[width=0.225\textwidth]{figures/loc1_ht1_pnp_marked_cyan.pdf}\label{fig: loc1_ht1_pnp_marked}}
\quad
\subfloat[]{\includegraphics[width=0.225\textwidth]{figures/loc1_ht1_mi_cyan.pdf}\label{fig: loc1_ht1_mi}}
\caption{Two step approach for camera localization in prior map : Figure \ref{fig: loc1_ht1_pnp_marked} shows the projection of 3D-LiDAR ground points (cyan) using the initial estimate obtained using feature matching (PnP Estimate from Section \ref{sec: initsparsefeaturematching}), and Figure \ref{fig: loc1_ht1_mi} shows the projection of 3D-LiDAR ground points using the refined camera pose estimate from maximization of MI (from Section \ref{sec: refinement_of_localization}). The misalignment visible in Figure \ref{fig: loc1_ht1_pnp_marked}, is absent in Figure \ref{fig: loc1_ht1_mi} (best viewed digitally).}
\label{fig: improvementwithMIbasedlocalization}
\vspace{-15pt}
\end{figure}
We qualitatively validate the camera localization (Figure \ref{fig: improvementwithMIbasedlocalization}) estimate ($[\mathbf{^{C}R_W}, \mathbf{^{C}t_W}]$) by projecting points from the 3D-LiDAR map (Figures \ref{fig: intensityimage} and \ref{fig: zheightimage}) onto the fisheye image (Figure \ref{fig: fisheyeimage}). As shown in Figure \ref{fig: loc1_ht1_pnp_marked}, the projection of LiDAR map points on the fisheye image obtained using the initial camera pose are not well aligned.
When we plot (Figure \ref{fig: xperturbatioMI}) the MI around the initial camera pose, we observe that it is not at its maximum at the initial estimate (also called PnP Estimate), thus holding the promise for further improvement. Similarly, Figure \ref{fig: costfunction2d} presents the surface plot of MI, which shows the presence of a global maximum in each sub-plot. Therefore, on solving the optimization problem posed in Equation \ref{eqn: globaloptimization} we obtain camera pose estimate which maximizes the MI between the two modalities and results in negligible misalignment of the projected LiDAR map points in Figure \ref{fig: loc1_ht1_mi}.
\begin{figure}[!ht]
\centering
\includegraphics[width = 0.4\textwidth]{figures/x_MI.png}
\caption{
Plot of MI around the PnP estimate (from Section \ref{sec: initsparsefeaturematching}). Plot shows that MI is not at maximum at the PnP estimate, therefore the maximization of MI may reduce the misalignment in projection of 3D-LiDAR points visible in Figure \ref{fig: loc1_ht1_pnp_marked}}\label{fig: xperturbatioMI}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.4\textwidth]{figures/costsurface.png}
\caption{\textbf{Plot of MI by perturbing two degrees of freedom around the PnP Estimate} (i.e. the initial estimate from Section \ref{sec: initsparsefeaturematching}). The cost function from single Image LiDAR Scan pair is not differentiable at several points. Hence, we use an exhaustive grid search around the initialization point to arrive at solution where MI is maximized.}
\label{fig: costfunction2d}
\vspace{-10pt}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.4\textwidth]{figures/perturbation_collage.png}
\caption{Performance of MI based fisheye camera localization refinement for different initial conditions. Here we perform 100 independent trials. \textcolor{red}{+} is the initialization, \textcolor{green}{*} is the final result.}
\label{fig: mi_randominitializaton}
\vspace{-10pt}
\end{figure}
We run 100 independent trials to evaluate the robustness of the MI based refinement method (Section \ref{sec: refinement_of_localization}) to change in initialization (Figure \ref{fig: mi_randominitializaton}). The translation parameters show greater variance when compared to yaw $\psi$. The higher variance of the translation variables can be attributed to the fact that the MI based cost function is less sensitive to changes in translation variables, especially in the outdoor scenario where most of the points lie in the far field.
In the limiting case, far away points are considered to be points at infinity, represented as $[x, y, z, 0]^{T}$, which, under camera projection (Equation \ref{eqn: fisheyeprojectionmodel}), render the translation variable ($\mathbf{^{C}t_{W}}$) in the optimization problem (Equation \ref{eqn: globaloptimization}) un-observable.
This result is also presented in \cite{MIGP}, specifically when discussing sensor registration in an outdoor environment using only a single image - LiDAR scan pair, which is similar to our situation.
\subsection{Real World Experiments}
In order to demonstrate the validity of our algorithm in realistic situations, we conducted experiments with data collected from a real fisheye camera (Figure \ref{fig: realfisheyedatacollection}).
\subsubsection{System Description}
We use a fisheye lens Fujinon FE185C057HA 2/3 inch sensor, which provides $185^{\circ}$ of vertical and horizontal FoV. Our camera is a 5MP Sony IMX264. We mount our sensor from a tripod (Figure \ref{fig: realfisheyedatacollection}), looking vertically down, and capture images of the environment. Our ultimate goal is to mount these cameras at challenging intersections for navigation of autonomous vehicles, and use the proposed method to register them in a prior map. We use an iPhone to provide us an approximate GPS location (without the orientation) of the fisheye camera, which is used to limit the search space in the prior map. We use a high accuracy RTK-GPS (uBlox ZED F9P GNSS + uBlox antenna ANN-MB-00) unit to measure GPS co-ordinates of distinctive corners on road markings that can be used for quantifying the accuracy of camera localization (Figure \ref{fig: reprojectionErrors}).
\vspace{-10pt}
\begin{figure}[H]
\centering
\includegraphics[width=0.4\textwidth]{figures/real_fisheye_data_collection.pdf}
\caption{Collecting data for real experiments using a tripod mounted downward looking fisheye camera. Our ultimate goal is to mount these cameras, along with our smart infrastructure nodes, at challenging intersections for navigation of autonomous vehicles.}
\label{fig: realfisheyedatacollection}
\vspace{-15pt}
\end{figure}
\subsubsection{Results} We present results with real world data from two different locations in Figure \ref{fig: misalignmentvsalignmentsimulateddata_rdloc} which qualitatively demonstrate the incremental improvement in camera localization using our two step approach. While the projection of LiDAR points onto the fisheye image using the initial camera localization appears misaligned in Figures \ref{fig: misalignmentinprojection_rdloc1} and \ref{fig: misalignmentinprojection_rdloc2}, the misalignment is reduced when the LiDAR points are projected using refined camera localization in Figures \ref{fig: properalignment_rdloc1} and \ref{fig: properalignment_rdloc2}. We quantify the veracity of camera localization by measuring the average reprojection error for points on the fisheye image whose GPS locations we have measured using high accuracy RTK-GPS. We manually mark these points on the fisheye image, and measure the difference between them and the reprojection of the corresponding 3D point in the prior map onto the fisheye image, using the estimated camera localization. The results presented in Figure \ref{fig: reprojectionErrors} show that the reprojection error on the fisheye image plane reduces when we refine the initial camera localization by maximizing the mutual information.
\begin{figure*}[!ht]
\centering
\subfloat[Location 1 - Projection of LiDAR points (cyan) using PnP Estimate]{\includegraphics[width=0.45\textwidth]{figures/pnp_loc1_cyan.pdf}\label{fig: misalignmentinprojection_rdloc1}}
\quad
\subfloat[Location 1 - Projection of LiDAR points (cyan) using MI Estimate]{\includegraphics[width=0.45\textwidth]{figures/max_mi_loc1_cyan.pdf}\label{fig: properalignment_rdloc1}}
\subfloat[Location 2 - Projection of LiDAR points (cyan) using PnP Estimate]{\includegraphics[width=0.45\textwidth]{figures/pnp_loc2_cyan.pdf}\label{fig: misalignmentinprojection_rdloc2}}
\quad
\subfloat[Location 2 - Projection of LiDAR points (cyan) using MI Estimate]{\includegraphics[width=0.45\textwidth]{figures/max_mi_loc2_cyan.pdf}\label{fig: properalignment_rdloc2}}
\caption{\textbf{Real World Experiments: }Projection of 3D-LiDAR ground points (cyan) on to fisheye camera using the initial camera pose (PnP Estimate) (Figure \ref{fig: misalignmentinprojection_rdloc1}, \ref{fig: misalignmentinprojection_rdloc2}) and MI based refinement of initial camera pose (Figure \ref{fig: properalignment_rdloc1}, \ref{fig: properalignment_rdloc2}). The misalignment of LiDAR points visible in highlighted areas in Figures \ref{fig: misalignmentinprojection_rdloc1}, \ref{fig: misalignmentinprojection_rdloc2}, are minimized in Figures \ref{fig: properalignment_rdloc1}, \ref{fig: properalignment_rdloc2}}
\label{fig: misalignmentvsalignmentsimulateddata_rdloc}
\vspace{-15pt}
\end{figure*}
\begin{figure*}[!ht]
\centering
\subfloat[Location1 - Average Reprojection Error with PnP = \textcolor{red}{20.52} pixel, and with Maximization of MI = \textcolor{blue}{9.12} pixel]{\includegraphics[width=0.45\textwidth]{figures/repErr_loc1.pdf}\label{fig: repErrLoc1}}
\quad
\subfloat[Location2 - Average Reprojection Error with PnP = \textcolor{red}{26.40} pixel, and with Maximization of MI = \textcolor{blue}{13.11} pixel]{\includegraphics[width=0.45\textwidth]{figures/repErr_loc2.pdf}\label{fig: repErrLoc2}}
\caption{\textcolor{green}{Green Circle} - Hand annotated corner point whose GPS location was measured using a high accuracy RTK-GPS unit, \textcolor{red}{Red Circle} - Projection of corner point's position onto Fisheye Image using PnP estimate (Section \ref{sec: initsparsefeaturematching}), \textcolor{blue}{Blue Circle} - Projection of corner point's position onto Fisheye Image using MI estimate (Section \ref{sec: refinement_of_localization}).}
\label{fig: reprojectionErrors}
\vspace{-20pt}
\end{figure*}
\section{Discussion $\&$ Conclusion}
\label{sec: conclusion}
We present an approach to localize a smart infrastructure node equipped with a fisheye camera. The downward facing fisheye image is registered to a prior map, comprising of a co-registered satellite image and a ground reflectivity/height map from LiDAR-SLAM. Our two-step approach uses feature matching between the rectified fisheye image and the satellite imagery to get an initial camera pose, followed by maximization of MI between the fisheye image and 3D LiDAR map to refine the initial camera localization. Since we have only a single camera image to register against (the smart infrastructure node is static), the cost surface may not always be smooth \cite{MIGP} and therefore not differentiable - leading to the failure of gradient descent methods. Hence, we use an exhaustive grid search method to find the optimal camera pose. Such a search may be time-consuming (depending on the number of 3D LiDAR map points used for calculating MI (Equation \ref{eqn: mutualinformationdefinition}), the interval of exhaustive grid search and the available compute power), and not suitable for real-time operation. This is acceptable for our application, because we need to localize the smart infrastructure node once at install and this can be an offline process. Moreover, this method can be accelerated by use of GPUs.
|
\section{Introduction and Problem Outline}
Semantic segmentation is the process of identifying specific regions of an image by labeling each pixel as being part of a class. This process is has a wide range of applications in image analysis, from the development of self-driving vehicles -- that use computer vision to understand where roads, pedestrians, and other vehicles are -- to organizational image search tools -- that allow users to automatically sort their pictures by content. The enormous amounts of image data generated each day by people around the world provide the field of semantic-segmentation with an ever-growing domain of both data and possible application.
One application of semantic segmentation in medical imaging is in the diagnosis of Chiari Malformation type I (CMI), a condition affecting the skull and brain that is estimated to affect slightly fewer than 1 in 1,000 people \cite{AANS}. Although most of these cases are asymptomatic, some can lead to extreme pain and require immediate surgery. In such cases CMI can be accompanied by severe head and neck pain, headaches, dizziness, and impaired vision \cite{BologneseEtAl2019}.
Although CMI is typically diagnosed when an individual's cerebellar tonsils (TP) extend more than 5mm below the foramen magnum, many experts recommend rejecting this criterion due to a weak correlation between CMI and TP \cite{BologneseEtAl2019}. A recent study also found that there is a significant increase in neural tissue movement in CMI patients compared to controls, especially in the brain stem and cerebellum \cite{NwotchouangEtAl2020}. This was found using a dynamic MRI technique called Displacement Encoding with Stimulated Echoes (DENSE): a type of MR imaging that records information about movement in the brain and that is sensitive enough to capture micrometer scale deformations that occur with each heartbeat. Measurements related to the displacement in the brain stem and cerebellum could perhaps become part of a new, more reliable basis for the diagnosis of CMI.
The purpose of this study is to create image segmentation algorithms that automatically identify these regions and conduct further analysis with DENSE MRI data. One of the most limiting steps in diagnosis based on this new criterion is the accurate segmentation of the brain stem and cerebellum in the DENSE-MRI data. Currently, the brain regions are labeled manually, which is a time consuming and, thus, expensive process \cite{NwotchouangEtAl2020}. This study compares two distinct methods to automate the segmentation of the brain stem and cerebellum: an atlas-based image registration approach, and a convolutional-neural-network-based approach (paper overview shown in \textbf{Figure \ref{fig:paper-overview}}).
\begin{figure}[t]
\centering
\includegraphics[width=.8\textwidth]{paper-overview.png}
\caption{An overview of the setup of this study. The current method of diagnosing Chiari from brain displacement data (center gold path) involves a radiologist manually drawing a mask that identifies the brain stem and cerebellum, and computing the spatial average and temporal peak displacement on those regions as a biomarker for CMI diagnosis. The brain stem peak-displacement biomarker is shown for each method in the white box on the left, while the cerebellum's is shown in the grey on the right. An atlas-based algorithm (left blue path) and a neural-network-based approach (right red path) were developed in this study to segment the cerebellum and brain stem given DENSE MR imaging and produce biomarkers automatically. Results were compared between the two methods and to the results produced from the manual process, that are treated as truth.}
\label{fig:paper-overview}
\end{figure}
The data set that is used to automate this process includes DENSE MR images and masks identifying the cerebellum and brain stem for patients from Emory University Hospital's Department of Radiology and Imaging Science.
The data set has been split into training data, which is used to create of the neural-network-based and atlas-based segmentation methods, and testing data, which is used to gauge the generalization of the methods to unseen data.
In investigating atlas-based image registration, our study builds on methods from a MATLAB toolbox called FAIR \cite{Modersitzki2009}. Image registration from this toolbox relies on the idea of a template image and a reference image. In the case of segmenting MR images, the reference is the magnitude image (described further in \textbf{Section \ref{data subsection}}) that needs segmentation and the template is some magnitude image that has already been segmented. The main idea of this approach is to find a transformation that approximately aligns the template image to the reference.
Another approach is found in convolutional neural networks (CNNs), a category of machine learning. Using MR images from the data set as input, a model is trained in a supervised fashion to output a segmentation with three classes: brain stem, cerebellum, and background. In this study, a CNN made for biomedical imaging called U-Net \cite{RonnebergerEtAl2015} is implemented the python-based machine learning library PyTorch \cite{paszke2017automatic}. This network, combined with an optimizer and loss function, loops over a section of the data set, called the training set, to produce an output mask by classifying each pixel as background, brain stem, or cerebellum. This output is then compared to the corresponding known mask. With each iteration, the CNN learns from this comparison and updates its parameters to produce more accurate results. The model can then be tested on images in the data set, not used in training, as a means to see how well it segments new MR images.
Overall, as a single method of semantic segmentation, the neural-network based approach outperformed the atlas-based approach. However, both methods worked better for different regions; while the neural-network based method was able to produce a more accurate biomarker for the cerebellum in a majority of cases, the atlas-based method was able to produce a more accurate biomarker for the brain stem in a majority of cases. Together, for our testing set, the methods were able to produce segmentations of the brain stem and cerebellum that were, on average, 86\% similar to radiologist drawn segmentations (measured with Dice similarity index, described in \textbf{Section \ref{dice subsection}}) and that were able to predict the manually produced biomarkers for the brain stem and cerebellum with an average 4\% relative error each.
\section{Problem Description}
This section begins with a detailed description of the dataset used, then an introduction to the measures used to evaluate both segmentation and diagnosis. This is followed, in the next sections, first by an explanation of the atlas-based segmentation method used, then by an explanation of the convolutional neural-network-based approach.
\subsection{Data and Pre-Processing}
\label{data subsection}
This study uses 256x256 pixel DENSE MR image data and associated masks for 63 patients.
We randomly split the data into a training set, containing 51 examples used in the creation of the neural-network-based and atlas-based methods, and and a test set, consisting of 12 examples which we use to gauge the generalization of the methods. As is typical in MR imaging, the DENSE MRI data is stored as a complex-valued array.
The images typically visualized in MRI are based on magnitude, where the grey value of each pixel is created from the magnitude of the complex value in that pixel’s position (shown in \textbf{Figure \ref{fig:data-example}}(1)). The color of each of the pixels in these magnitude images is controlled by the chemical makeup of the material that exists in that position in the brain, and therefore represent the brain's anatomy.
Since the brightness and contrast levels varied across the magnitude images in a way that made segmentation more difficult, we pre-processed the data using histogram normalization. To this end, we used the MATLAB function histeq, which approximately linerarizes the cumulative distribution function and helps create a more uniform set of images. \textbf{Figure \ref{fig:normal}} compares a sample of original and normalized images. For more explanation of this process, see \cite{Coste}.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{50-full-full-dataset.png}
\caption{From left to right, an example of one patient's (1) magnitude MRI image; (2) mask; (3) DENSE images representing one cardiac gate each; (4) temporal mean DENSE image; and (5) temporal peak DENSE image.}
\label{fig:data-example}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.6\textwidth]{normalization-example.png}
\caption{A sample of sixteen magnitude images, original (top) and normalized using histogram normalization (bottom).}
\label{fig:normal}
\end{figure}
DENSE imaging encodes displacement information in the phase of each complex number corresponding to the intensity of movement associated with each pixel position. For each subject many DENSE MRI scans were created corresponding to gates that occur over the patient's cardiac cycle (shown in \textbf{Figure \ref{fig:data-example}}(3)). The number of DENSE images varies per patient based on their heartbeat. The visual interpretation of these phase images is perhaps less intuitive than the magnitude images, mapping movement rather than anatomy.
Generally, areas of large movement in brain tissue appear lighter than those of small movement, but in areas of especially high movement -- especially outside the brain where cerebrosprinal fluid (CSF) moves much more quickly than the micrometer deformations in the brain tissue -- phase wrapping occurs. Because there is no distinction between the gray scale color of a pixel stored with a phase $\phi$ and $\phi + 2\pi$, for example, they appear the same. These regions with high movement appear as random noise because the movement is beyond the smaller scale for which the phase encoding was calibrated. For more on phase wrapping, see \cite{phasewrap}.
In this study, we follow \cite{NwotchouangEtAl2020} and remove the temporal dimension of the DENSE MRI data by computing a mean DENSE image, showing the mean value per pixel over time (\textbf{Figure \ref{fig:data-example}}(4)), and a peak DENSE image, showing the maximum value per pixel over time (\textbf{Figure \ref{fig:data-example}}(5)), for each subject.
This representation of the DENSE MRI data over time also helps reduce the regions of random noise and summarize over the patient's whole cardiac cycle. For more discussion of this choice and of other ways to represent this data, see \cite{NwotchouangEtAl2020}.
The masks associated with the magnitude images classify each pixel as being part of the background, the cerebellum, or the brain stem, and were drawn by researchers in Emory's Department of Radiology and Imaging Science (shown in \textbf{Figure \ref{fig:data-example}}(2)) as part of the study~\cite{NwotchouangEtAl2020}.
\subsection{Notation}
\label{Notation}
The neural-network-based and atlas-based approaches use image data in different ways. The first keeps the images as MRI discrete data while the latter uses these data to build a continuous model through interpolation.
In the neural-network-based method, the input, $I \in \ensuremath{\mathds{R}}^{256 \times 256 \times 3}$, is made of 3 channels: the normalized magnitude, peak DENSE, and mean DENSE images. These are each 256x256 matrices with entries corresponding to the gray values.
The atlas-based approach treats an image as an interpolated function, ${\cal I} : \Omega \to \ensuremath{\mathds{R}}$ where the domain $\Omega \subseteq \ensuremath{\mathds{R}}^2$ corresponds to coordinates of pixels in the image grid. The multilevel processes, described in \textbf{Section \ref{Registration}}, rely on this continuous model of our data. The range of this function, called on a specific 256x256 grid, creates the discrete 256x256 pixel images. To distinguish these understandings, calligraphic letters have been used to denote continuous functions while regular script denotes discrete. This notation, and the notation of other equations related to the atlas-based method, are formatted in the style of FAIR \cite{Modersitzki2009}. This MATLAB toolbox and its associated textbook~\cite{Modersitzki2009} can be referred to for a more complete discussion of image registration.
These two understandings of images lend also themselves to two conceptions of the outputted masks. In the neural-network-based method, the output mask is a discrete 256x256 matrix which predicts a class $k \in \{0,1,2\}$ for each pixel where 0 corresponds to background, 1 to cerebellum, and 2 to brain stem. $M$ symbolizes the known target masks while $M_{p}$ corresponds to the model predicted mask.
Analogously, in the continuous function sense, let $C$, $B$, and $T$ be the sets representing the cerebellum, brain stem, and the total union between them respectively. Then let $\chi_C, \chi_B : \Omega \to \{0, 1\}$ be the characteristic functions for the cerebellum and brain stem respectively. Masks that display these regions will hence be defined as ${\mathcal{M}} : \Omega \to \{0, 1, 2\}$ where
\begin{equation}
{\mathcal{M}}(x) = \chi_C(x) + 2\chi_B(x).
\end{equation}
In addition to this, we define the total (T) mask characteristic function to be $\chi_T : \Omega \to \{0, 1\}$ where $\chi_T = \chi_{C \cup B}$.
Using this notation, the template mask will be referred to as ${\mathcal{M}}_{\mathcal{T}}$ and the reference mask as ${\mathcal{M}}_{\mathcal{R}}$.
\subsection{Evaluation Measures}
Our ultimate goal is to automatically segment the brain stem and cerebellum so that we could produce peak-displacement biomarkers for those regions that match radiologist produced ones. The cerebellum's average peak-displacement, for example, is calculated by averaging the peak-displacement of each pixel in the DENSE MRI that is classified as being part of the cerebellum (and likewise for the brain stem). In addition to producing accurate biomarkers we also quantify the accuracy of the segmentations to ensure the plausibility of the biomarker.
Radiologist drawn masks that identify what displacement data to average to obtain the biomarkers are treated as the gold standard. It may be that these algorithmic methods can produce better, more consistent masks, but that lies outside the scope of this study. To balance the dual concerns of producing masks and biomarkers that match manually produced ones, Dice similarity indices and biomarkers were considered at each step.
\subsubsection{Dice Index}
\label{dice subsection}
To evaluate the success of both the neural-network-based and atlas-based segmentation, we use the Dice similarity index. Given two subsets of $A,B \subset \Omega$, the Dice index is computed as
\begin{equation}\label{eq:Dice}
D_{\rm{Dice}}[A, B]
= \rm{Dice}[\chi_{A}, \chi_{B}]
= 2 \int_{\Omega} \frac{\chi_{A}(x) \chi_{B}(x)}{\chi_{A}(x)+\chi_{B}(x)} dx,
\end{equation}
where $\chi_{A}(x), \chi_{B}(x): \Omega \to \{0,1\}$ represent the characteristic functions for the sets $A$ and $B$, respectively. This results in a measure that ranges from 0, in the case of two disjoint regions, to 1, in the case of identical regions \cite{Monteux2019}.
We use this measure primarily to evaluate mask similarity and determine if algorithmically predicted masks reasonably match radiologist drawn masks. Our masks define three sets of special interest: the set of pixels identified as the brain stem, the set identified as the cerebellum, and the set identified as background. Computing the Dice similarity for the cerebellum and brain stem, especially, between a true mask and predicted mask for the same brain quantifies how well the segmentation method is working; in both the neural-network-based and atlas-based approaches, algorithms and parameters were chosen to maximize this number.
\subsubsection{Biomarkers} Recently, the spatial mean and temporal maximum of the magnitude displacement measured in DENSE-MRI over the cerebellum and brain stem have shown promise as a biomarker for diagnosing CMI~\cite{NwotchouangEtAl2020}. For example, once a segmentation is identified for the brain stem, the value of each pixel classified as a part of this region from the peak-displacement DENSE image data is averaged (or analogously, for the cerebellum). Though the world ``biomarker" applies broadly, and attributes including age, weight, and sex could each be considered biomarkers that are relevant to CMI, when we refer to biomarkers throughout this paper, we are referring to the displacement averages across the brain stem and cerebellum described above. Building on the results of \cite{NwotchouangEtAl2020}, the biomarkers produced from manual segmentation will be treated as the truth, and our aim is to match these with our automated approaches.
\section{Atlas-Based Segmentation}
This section includes discussion of the way a template image is chosen, a way of averaging the segmentation results, the parameters and functions used in image registration (associated with the toolbox FAIR \cite{Modersitzki2009}), and examples of the method in full.
\subsection{Overview of the Registration Pipeline}
\begin{figure}[t]
\centering
\includegraphics[width=.7\textwidth]{Atlas/atlas-based-example.png}
\caption{A general overview of the registration process, that involves finding a transformation between a template and reference and then applying that transformation to the template mask to predict a reference mask.}
\label{fig:atlas-example}
\end{figure}
Given a new image without a mask, our method compares the new image to a set of 51 labeled images in order to predict a mask for the new image. To allow for a pixel-to-pixel comparison, we solve a set of registration problems. Following the language used in FAIR, the new image is called the reference, while the known labeled images are templates.
Here, we discuss how templates are chosen to be compared to new references along with an averaging method to reduce outlying results, and in the next section we will describe the registration performed in each of these comparisons.
\subsubsection{Choosing a Template Image}
\label{Template}
Perhaps the most critical factor in the success of the atlas-based segmentation is picking an appropriate template image for a new reference; when the brains in a template and reference image are already similar -- in their shape and in how they are positioned -- the registration is best set up for success. However, choosing the most similar template image is not a trivial task. While the Dice index between the template and reference masks could provide a useful measure, as it captures the alignment we are interested in, the reference mask is what we aim to produce so this index cannot be calculated before registration. For this reason, we had to rely on calculating some type of similarity metric between the magnitude images. There is a wide range of options for this that value different attributes. Three of these options that we explored are the Normalized Gradient Field (NGF), Sum of Squares Distance (SSD), and weighted Sum of Squares Distance (wSSD).
A NGF method works by aligning edges identified in the magnitude images by the gradient -- relying more on the shape and location of the brains (for more on NGFs, see \cite[Sec. 7.4]{Modersitzki2009}). This showed a slight correspondence, but was much less successful than using SSD (this metric is described in more detail in \textbf{Section \ref{Registration}}). Using a wSSD metric allowed the program to focus on the section of the brain with the brain stem and cerebellum. While this can be effective, the main flaw was that the skull shape was an important factor that was being ignored. The final similarity metric that we decided on was regular SSD because it seemed to be the overall best and most adaptable.
In \textbf{Figure \ref{fig:ssd_similarity}}, 50 image registrations were run with one example patient used as reference and each of our other 50 training images used to see how well the SSD works as a similarity measure for the magnitude images, with the SSD measure plotted against the ideal, mask-similarity Dice measure. Note that whereas a higher Dice index represents higher similarity, the opposite is true for the SSD measure; if two exactly similar images were compared, the SSD would measure zero, while the Dice index would measure one. Both before and after registration, the SSD and Dice index are, to an extent, negatively correlated, which helps justify our use of it as a way to chose initially similar templates for our reference. Though it is by no means a perfect measure, for our case, it provides an intuitive and computationally efficient measure that helps improve registration results.
\begin{figure}[t]
\centering
\begin{tabular}{c}
\textbf{Dice Mask Similarity versus SSD} \\
\includegraphics[width=0.9\textwidth]{Atlas/ssd-dice-scatter.png}
\end{tabular}
\caption{Plot of the Dice similarity index and SSD between an example reference image and all template images in our training set. The red circles represent the measures before the registration is performed (with a corresponding solid red least squares line), while the blue squares represent them after it is performed (with a corresponding solid red least squares line). Black dotted lines connect points representing the same templates before and after registration.}
\label{fig:ssd_similarity}
\end{figure}
\subsubsection{Averaging the Registrations}
\label{Averaging}
The significance of edge cases in the registration creates a need for less dependence on any individual template image. To accomplish this, the registration was run on multiple template images and the resulting masks were averaged together. Then the resulting average mask could be converted to a binary image and used as a more definitive prediction of the segmentation.
To be more specific, the variable $n$ was used to correspond to the number of patients that the registration program is run on. Using the same reasoning behind choosing a template as \textbf{Section \ref{Template}}, the SSD is calculated for each template and the lowest $n$ are selected. These images are then summed together and divided by $n$ to get the average mask. Finally, a threshold is used to convert this average transformed mask into a binary mask again. For discussion of how the values of $n$ and threshold were chosen, see the beginning of \textbf{Section \ref{AtlasResults}}.
\subsection{Registration Algorithm}
\label{Registration}
The goal of the image registration is to find some transformation $y : \Omega \to \ensuremath{\mathds{R}}^2$ such that
\begin{equation}
{\mathcal{T}}[y](x) \approx {\mathcal{R}}(x), \quad \text{ for all } \quad x \in \Omega.
\end{equation}
Note that ${\mathcal{T}}[y](x) = {\mathcal{T}}(y(x))$ is written this way to emphasize that the function $y$ is an input as well. \textbf{Section \ref{Notation}} provides insight into other notational choices.
We find that function $y$ by minimizing the nonconvex objective functional
\begin{equation}
{\cal{J}}_{\rm{atlas}}[y] = D_{\rm{SSD}}[{\mathcal{T}}[y], {\mathcal{R}}]+ \alpha {\cal S}[y],
\label{joint_objective}
\end{equation}
where $D_{\rm{SSD}}$ is the sum of squares distance measure, $\alpha$ is the regularization parameter, and $\cal S$ is the regularizing functional. These parameters along with some fundamental design decisions are described in more detail below: \\
\textbf{Distance}: When performing the image registration, the algorithm tries to minimize the distance measure. There are many options for how to pick this distance measure, but the perhaps simplest and most common is the Sum of Squares Distance (SSD) given by
\begin{equation}
D_{\rm{SSD}}[{\mathcal{T}}[y], {\mathcal{R}}] = \frac{1}{2} \int_{\Omega} ({\mathcal{T}}[y](x) - {\mathcal{R}}(x))^2 dx.
\end{equation}
The SSD measure penalizes different brightness levels. As a result, altering the contrast of the images through normalization can make the SSD measure perform much better. \\
\textbf{Regularizer}: When only focusing on the distance measure, the image registration program can come up with some unrealistic scenarios (for further discussion, see \cite{Fischer_2008}). Sometimes it compresses the template image to fit within some small region of the reference. Other times it may create visible folds in the image. One way to counteract this is by introducing a regularizer.
Two commonly used regularizers implemented in FAIR are elastic and hyperelastic \cite{BurgerEtAl2013}. Both of these support the same idea but vary in severity. An elastic regularizer will try to keep the image registration close to the original orientation by introducing linear strain to any change. A hyperelastic regularizer does this with a nonlinear strain equation.
Because of this, the elastic regularizer is only accurate for smaller displacements. For anything larger, the elastic regularizer risks becoming inaccurate and causing folding in the transformation grid. The more realistic, yet computationally more expensive hyperelastic regularizer also prevents folding. Either of these regularizers would work for the Chiari data, but hyperelastic was used since seems to remove more of the edge cases and its computational costs for two-dimensional problems is feasible. \\
\textbf{Regularization Parameter}: When constructing the objective functional from \textbf{Equation \ref{joint_objective}}, there is a trade-off between whether to minimize the distance or follow the regularizer. The parameter $\alpha$ controls the balance between allowing for displacement between the original and transformed image, to better the distance measure, or to penalizing displacement, to preserve the general shape of the image throughout the registration \cite{Modersitzki2009}.
Note that in the objective functional equation, $\alpha$ corresponds to the regularization parameter. This coefficient was chosen to be 500 in our final algorithm based on trial and error, but the parameter selection could be further refined for better results. \\
\textbf{Multilevel Optimization}: As common in FAIR, Gauss Newton optimization was applied to a discretization of the objective functional in Equation~\ref{joint_objective} to generate a transformation from template to reference. Because discretizing objective functional on a 256x256 grid can lead to many local minima, and the convergence of this Gauss Newton to an appropriate minimum is dependent on an initial guess, a multilevel approach was used. This involves creating coarser discretizations of the problem and leverages the continuous image model that we can find through linear interpolation of the image data as detailed below.
We model the template and reference images as continuous functions, ${\mathcal{T}} : \Omega \to \ensuremath{\mathds{R}}$ and ${\mathcal{R}} : \Omega \to \ensuremath{\mathds{R}}$ where $\Omega \subseteq \ensuremath{\mathds{R}}^2$ corresponds to coordinates of pixels in the image grid. Coarser representations are created by changing the grid for which our interpolated image function is called, that contain fewer pixels: in our case, three additional representations, with 32x32, 64x64, and 128x128 pixels each. The optimization is applied in a way that moves up through these representations, applying optimization on the coarsest first and using the found transformation as the initial guess for the image at the next level.
The coarsest representation, described above, is also used in an initial pre-registration step where a parametric transformation is estimated to better align the template and reference images. This step, much like the step to choose appropriate template images, helps simplify the registration. One natural instinct for this rough transformation may be to simply translate and rotate the template image to align the reference. This would be a rigid transformation. There are advantages to keeping the transformations rigid, especially in cases where we expect little difference in the shapes of two images, but more flexibility can be added by using an affine transformation. This allows for additional linear transformations, such as shearing and scaling. We use affine transformations because of this flexibility to further align our images of often widely varied brain shapes.
\subsection{Examples}
This subsection compares two examples of how the registration performed --- one showing a successful registration and another showing how it can fail.
Using the same parameters described in \textbf{Section \ref{Registration}}, a single registration was run on both patients. \textbf{Figure \ref{patient_mask}} compares how well the masks were aligned before and after the registration. The transformations found for each patient were reasonable and proved to be somewhat successful. However, there were still some regions that required more attention. Particularly, the lower right of the cerebellum overextends in both patients.
\begin{figure}[t]
\centering
\begin{tabular}{ccc}
& \textbf{${\mathcal{R}}$ with ${\mathcal{M}}_{\mathcal{T}}$ and ${\mathcal{M}}_{\mathcal{R}}$} & \textbf{${\mathcal{R}}$ with ${\mathcal{M}}_{\mathcal{T}}[y]$ and ${\mathcal{M}}_{\mathcal{R}}$} \\
\vspace{3mm}
\rotatebox[origin=c]{90}{Patient 1} &
\raisebox{-0.5\height}{\includegraphics[width=0.35\textwidth]{Atlas/42/42_mask_1.jpg}} & \raisebox{-0.5\height}{\includegraphics[width=0.35\textwidth]{Atlas/42/42_mask_2.jpg}} \\
\rotatebox[origin=c]{90}{Patient 2} &
\raisebox{-0.5\height}{\includegraphics[width=0.35\textwidth]{Atlas/47/47_mask_1.jpg}} & \raisebox{-0.5\height}{\includegraphics[width=0.35\textwidth]{Atlas/47/47_mask_2.jpg}}
\end{tabular}
\caption{The mask transformations from patients 1 and 2. The first column of images are the reference and reference mask (yellow) with the template mask on top (blue); the second column of images are the same except with the blue outline representing the transformed template mask. The Dice similarities are in \textbf{Table \ref{patient_dice}}.}
\label{patient_mask}
\end{figure}
In an attempt to smooth some of the outlying regions, the average program was run on both patients as shown in \textbf{Figure \ref{patient_avg}} and \textbf{Table \ref{patient_dice}}. Tuning the parameters of the average program (as seen in \textbf{Section \ref{AtlasResults}}) resulted in it performing better on most patients. Averaging this with other patients lessened the severity of this by eliminating the outliers that result from running the registration once.
\begin{figure}[t]
\centering
\begin{tabular}{ccccc}
&\textbf{Single ${\mathcal{T}}[y]$} & \textbf{Average ${\mathcal{T}}[y]$} & \textbf{Average ${\mathcal{T}}[y] > 0.5$} & \\
\vspace{3mm}
\rotatebox[origin=c]{90}{Patient 1} &
\raisebox{-0.5\height}{\includegraphics[width=0.25\textwidth]{Atlas/42/42_avg_1.jpg}} & \raisebox{-0.5\height}{\includegraphics[width=0.25\textwidth]{Atlas/42/42_avg_2.jpg}} & \raisebox{-0.5\height}{\includegraphics[width=0.25\textwidth]{Atlas/42/42_avg_3.jpg}} & \multirow{2}{*}{\includegraphics[width=0.045\textwidth]{Atlas/42/42_avg_4.jpg}} \\
\rotatebox[origin=c]{90}{Patient 2} &
\raisebox{-0.5\height}{\includegraphics[width=0.25\textwidth]{Atlas/47/47_avg_1.jpg}} & \raisebox{-0.5\height}{\includegraphics[width=0.25\textwidth]{Atlas/47/47_avg_2.jpg}} & \raisebox{-0.5\height}{\includegraphics[width=0.25\textwidth]{Atlas/47/47_avg_3.jpg}}
\end{tabular}
\caption{Single registration compared to the average registration program on both patients. The left column of images shows the single registration from before, the middle column is the average of all the pixels with $n = 10$ patients, and the right column is the same figure after using a threshold of $0.5$. This is the same mask as \textbf{Figure \ref{patient_mask}} and the similarity is shown under the average column of \textbf{Table \ref{patient_dice}}.}
\label{patient_avg}
\end{figure}
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& \multicolumn{3}{c|}{Patient 1} & \multicolumn{3}{c|}{Patient 2} \\
\cline{2-7}
& Original & Transformed & Average & Original & Transformed & Average \\
\hline
Brain stem & 0.734 & 0.896 & 0.928 & 0.675 & 0.845 & 0.805 \\
\hline
Cerebellum & 0.659 & 0.910 & 0.936 & 0.805 & 0.855 & 0.850 \\
\hline
Full Mask & 0.701 & 0.902 & 0.932 & 0.728 & 0.849 & 0.824 \\
\hline
\end{tabular}
\caption{Dice similarities of the patient masks from \textbf{Figure \ref{patient_mask}} before and after the transformation along with the average version from \textbf{Figure \ref{patient_avg}}.}
\label{patient_dice}
\end{table}
Looking at the particular patients from \textbf{Table \ref{patient_dice}}, the averaging improved the similarity of patient 1 by around 2-3\% but decreased the similarity of patient 2 by around the same amount (particularly in the brain stem). This could represent the need for a more definitive metric or a larger dataset; either of these solutions would aim to increase the quality of the chosen template.
\section{Neural-Network-Based Semantic Segmentation}
The overarching goal of this approach is to find a relationship between the input DENSE imaging data and the known target segmentation masks through a function which directly labels the image. We can define our input $I \in \ensuremath{\mathds{R}}^{256\times256\times3}$ and masks $M \in \{0,1,2\}^{256 \times 256}$. As stated previously in \textbf{Section \ref{Notation}}, $M$ symbolizes the known target masks while $M_{p}$ corresponds to the model predicted mask.
\subsection{U-Net: A Convolutional Neural Network}
Given an input image $I$, our goal is to estimate a three-dimensional tensor $P$ in the set
\begin{equation}
\Delta = \left\{Q \in \ensuremath{\mathds{R}}^{256 \times 256 \times 3} : Q_{i,j,k} \geq 0 \text{ }\forall \text{ } i,j,k \text{ and } \sum_{k=0}^{2} Q_{i,j,k}=1\right\},
\end{equation}
where $P_{i,j,k}$ is the probability the pixel at $i,j$ belongs to class $k$. The relationship between $I$ and $P$ can then be represented by a parameterized function
\begin{equation}
f_{\theta}: \ensuremath{\mathds{R}}^{256\times 256\times 3} \to \Delta,
\end{equation}
where $\theta$ represents the weights of the model.
To then create a mask $M_p$, each pixel is assigned to the class with the highest probability; i.e., we maximize over the third dimension in $P$.
Our model is a convolutional neural network (CNN).
CNNs are a type of neural network which process data that have a ``grid like topology," such as image data (a 2D grid of pixels), and utilizes a convolution operator. Since the data in this study consist of multidimensional arrays of pixels and the goal is to correctly predict which class each pixel belongs to, a CNN is a natural choice. Many deep networks require thousands of images, however, our data set, as is common in biomedical applications, is much smaller.
A CNN architecture that has been successful in other biomedical segmentation tasks with limited amounts of training data is the U-Net~\cite{RonnebergerEtAl2015}. A U-Net consists of two main sections:
the contracting path (encoder) and the expansive path (decoder) shown on the left and right side of \textbf{Figure \ref{fig:Unet-Arch}}, respectively.
\begin{figure}[t]
\centering
\includegraphics[width=.9\textwidth]{MachineLearning/UNet_Model_Arch-Reduced.pdf}
\caption{U-Net architecture used in our study based on the figure provided in U-Net: Convolutional Networks for Biomedical Image Segmentation \cite{RonnebergerEtAl2015}. Here numbers placed on top of the boxes represent number of channels while those on the side represent image dimensions.}
\label{fig:Unet-Arch}
\end{figure}
In the first layer of the contracting path, a 3x3 convolution is applied to each channel increasing the number of total feature channels. This is followed by a max pooling operation to break down the image resolution by decreasing image size to better identify features. A similar convolution structure is applied at each layer doubling the number feature channels. This leads to the bottom most layer where 1x1 convolutions are used, maintaining number of channels. Increasing the size of the image data, the expansive path (decoder) replaces the max pooling with upsampling operators. In the same formation as the contracting path, 2x2 up convolutions are applied except these halve the number of feature channels. During the process, high resolution features found during encoding are `copied and cropped' and combined with the decoded output increasing dimension size and providing information that will be used in the convolution layer to produce more accurate outputs.
As the input images were run through the U-Net, the network used information from each channel to assign each pixel a probability for each class. These values are used to calculate the loss and to find the final predicted segmentation, $M_p$.
\subsection{Training the Model}
In order to get a successful model, we want the predicted mask $M_p$ and the target mask $M$ to be as similar as possible. The first step in maximizing similarities is to create a satisfactory model. This was done through supervised training of the network using a training set, $\{\left(I_{t},M_{t}\right)\}_{t=1,2,\ldots41}$, a subset of the 51 DENSE MR images and corresponding masks $\{(I,M)\}$. To test the accuracy of the model on `new' data, a validation set, $\{\left(I_{v},M_{v}\right)\}_{v=1,2,\ldots,10}$, was defined using the remaining 10 DENSE images and masks not included in training.
The U-Net was setup based on code provided by Aman Arora \cite{UNetCode}. The network was then trained by looping a sequence of inputting $I_{t}$ to the U-Net, computing the loss between $f_{\theta}(I_{t})$ and $M_{t}$, and updating the model with the optimizer for a set number of iterations. In each iteration, the updated model was also run on the validation data to gauge its generalization; the validation loss was used to tune hyperparameters but was neither used to calculate gradients nor to change model parameters.
Comparing the probability matrix $P = f_{\theta}(I)$ to the corresponding $M$, for both training and validation data, shows the success of the model and can be numerically calculated using a loss function.
By maximizing the probability at the class corresponding to that of the known mask in the same position, the model error is minimized. In other words, our goal is to maximize
\begin{equation}\label{eq:lossMax}
[f_{\theta}(I)]_{i,j,[M]_{i,j}}, \quad \text{for all} \quad i = 1, \ldots, 256, j=1,\ldots,256,
\end{equation}
which is equivalent to minimizing
\begin{equation}\label{eq:lossMin}
-\log \left( [f_{\theta}(I)]_{i,j,[M]_{i,j}} \right), \quad \text{for all} \quad i,j \in [1,256],
\end{equation}
where $i,j$ corresponds to the pixel position.
Averaging \textbf{Equation \ref{eq:lossMin}} over all training image sand all pixels results in the objective function
\begin{equation}\label{eq:objFunct}
{\cal{J}}_{CNN}(\theta) = \frac{1}{41\cdot256^{2}} \sum_{t=1}^{41}\sum_{i,j = 1}^{256} -\log\left([f_{\theta}(I_{e})]_{i,j,[M_{e}]_{i,j}}\right).
\end{equation}
It can be seen that ${\cal{J}}_{CNN}$ equals the cross-entropy loss between $[f_{\theta}(I)]_{i,j,:}$ and the probability distribution defined by the standard basis vector, $[M]_{i,j}$, which calculates how far the model prediction is from the expected output.
Inputting $f_{\theta}(I)$ and $M$ in the loss function is a forward pass; this is done for both training and validation data. By executing a backwards pass on the loss, gradients can be calculated and used by the optimizer to update model parameters. As emphasized before, a backward pass was only executed with respect to the training data.
A limited-memory Broyden–Fletcher–Goldfarb–Shanno (lBFGS) algorithm from the PyTorch library \cite{paszke2017automatic} was used to optimize the overall loss function, shown previously in \textbf{Equation \ref{eq:objFunct}}. lBFGS is a Quasi-Newton method which uses a line search to automatically find the optimal learning rate $\epsilon$ \cite{Goodfellow:2016wc} which satisfies the strong Wolfe conditions set in the algorithm parameters. The linesearch ensures decrease of the objective function while keeping $\epsilon$ from becoming too small.
The lBFGS algorithm uses the gradient calculated from the loss function and the learning rate found in the line search to update the model weights. The images were then, once again, run through the U-Net in hopes of finding a better segmentation. The model with the lowest validation loss was saved.
\section{Results}
This section begins with results from the processes of creating the atlas-based and neural-network-based methods; overall results on the final testing set follow.
\subsection{Atlas-Based Image Registration Results}
\label{AtlasResults}
When creating the averaging program, two main parameters were chosen arbitrarily: $n$ and threshold. The $n$ refers to the number of template images to average and the threshold refers to which values should be counted when converting the average image to binary. In order to decide on a more definitive value for these parameters, a program was made to iterate through different possible values and find which performed the best.
This program was run over two metrics for how well it performed: Dice similarity and biomarker difference. Both of these metrics compare the atlas-based generated mask (average program) to the manually segmented mask and then average together all of the patients. These produced approximately the same results since the biomarker difference depends on how similar the masks are. A plot showing averaged Dice indices and absolute biomarker error over changes of both $n$ and the threshold can be seen in \textbf{Figure \ref{choosing_n}}.
\begin{figure}[t]
\centering
\begin{tabular}{cc}
\multirow{2}{*}{\textbf{Average Dice vs $n$}} & \textbf{Average Absolute Difference} \\
& \textbf{in Biomarker vs $n$} \\
\hspace{-5mm}
\includegraphics[width=0.48\textwidth]{Atlas/Choosing-Parameters/n-dice-grid.png} & \includegraphics[width=0.48\textwidth]{Atlas/Choosing-Parameters/n-biom-grid.png}
\end{tabular}
\caption{The left image shows the average Dice index plotted against the number of templates used in registration, and the right shows the average absolute difference between predicted and actual biomarker plotted against the number of templates used in registration. For both these figures, the average was computed over all 41 brains in the training set. The different curves correspond to different thresholds chosen for registration.}
\label{choosing_n}
\end{figure}
To choose these $n$ and threshold parameters, we performed a grid search, first using a wide range of $n$ and threshold values (ranging from 5 to 40 and 10\% to 90\%, respectively). We chose to complete all further registrations with $n=10$ and a threshold of 50\%, as this maximizes the average Dice index. The maximizer of the Dice index was chosen over the minimizer of the biomarker error because of how sensitive this biomarker measure is to slightly different segmentations. We expect that the behavior of the Dice index average will generalize better to other data.
\subsection{Neural-Network-Based Semantic Segmentation Results}
While training the model, cross-entropy loss and Dice similarities were used as a measure of model success. Our final model was saved at the lowest validation loss found, which occurred at iteration 309, and was used to produce the final results shown below. Cross-entropy loss and Dice similarities throughout model training are plotted in \textbf{Figure \ref{fig:MLPlots}} alongside a comparison of true to predicted masks for validation data using the final model. Final results of the model used can be found in \textbf{Table \ref{tab:mlComp}}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
& Cross-Entropy & \multicolumn{3}{c|}{Dice} \\
\cline{3-5}
& Loss & Background & Cerebellum & Brain Stem\\
\hline
Training Data & 0.0180 & 0.9962 & 0.9216 & 0.9116 \\
\hline
Validation Data & 0.0229 & 0.9953 & 0.8965 & 0.8919\\
\hline
\end{tabular}
\caption{Cross-Entropy loss and Dice similarity of each class for training and validation data sets for the neural-network based segmentation. The results shown here correspond to the loss and the respective Dice value in \textbf{Figure \ref{fig:MLPlots}} at iteration 309 where minimum validation loss was found.}
\label{tab:mlComp}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{MachineLearning/FinalTotalPlots.png}
\caption{Dice similarities of each class for training (top left) and validation (top right) data over training iterations, cross-entropy loss for each set over training iterations (bottom left), and a comparison of true (top) to predicted (bottom) masks for validation data (bottom right).}
\label{fig:MLPlots}
\end{figure}
The model was then used on the test data previously set aside. Biomarkers were calculated using test data segmentations which were compared to the corresponding biomarkers produced in the atlas-based method and those from the known masks $M$.
\subsection{Comparison Results}
When analyzing atlas-based (AB) and neural-network-based (NN) in the end, we relied on the Dice and biomarker (BM) metrics. \textbf{Table \ref{tab:beast}} shows the samples provided in the testing set and compares these methods to the manually segmented mask (true). The masks that were compared are shown in \textbf{Figure \ref{fig:result-masks}}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
Test Set &
\multicolumn{4}{c|}{Dice Similarity Index}&
\multicolumn{6}{c|}{Displacement Biomarker}
\\
\cline{2-11}
Subject &
\multicolumn{2}{c|}{Brain Stem}&
\multicolumn{2}{c|}{Cerebellum} &
\multicolumn{3}{c|}{Brain Stem} &
\multicolumn{3}{c|}{Cerebellum} \\
\cline{2-11}
\# & AB & NN & AB & NN & True & AB & NN & True & AB & NN \\
\hline
1 & 0.70 & \textbf{0.83} & 0.81 & \textbf{0.84} & 2.32 & \textbf{2.37} & 2.25 & 1.97 & 1.31 & \textbf{1.62} \\
\rowcolor{Gray}
2 & 0.87 & \textbf{0.90} & \textbf{0.89} & 0.86 & 0.69 & 0.68 & \textbf{0.68} & 0.47 & 0.45 & \textbf{0.45} \\
3 & 0.81 & \textbf{0.84} & 0.87 & \textbf{0.88} & 1.25 & 1.13 & \textbf{1.15} & 1.34 & 1.03 & \textbf{1.22} \\
\rowcolor{Gray}
4 & \textbf{0.84} & 0.81 & \textbf{0.88} & 0.71 & 1.06 & 0.99 & \textbf{1.00} & 0.58 & 0.53 & \textbf{0.54} \\
5 & 0.82 & \textbf{0.88} & 0.88 & \textbf{0.90} & 1.02 & \textbf{0.99} & 0.98 & 0.64 & 0.58 & \textbf{0.59} \\
\rowcolor{Gray}
6 & 0.53 & \textbf{0.74} & 0.60 & \textbf{0.67} & 0.85 & \textbf{0.77} & 0.76 & 0.54 & 0.90 & \textbf{0.60} \\
7 & 0.90 & \textbf{0.94} & 0.92 & \textbf{0.93} & 0.86 & 0.84 & \textbf{0.85} & 0.48 & 0.45 & \textbf{0.47} \\
\rowcolor{Gray}
8 & 0.87 & \textbf{0.90} & 0.90 & \textbf{0.96} & 0.89 & \textbf{0.89} & 0.88 & 0.48 & 0.51 & \textbf{0.49} \\
9 & \textbf{0.91} & 0.90 & \textbf{0.90} & 0.87 & 0.73 & \textbf{0.73} & 0.74 & 0.52 & \textbf{0.51} & 0.51 \\
\rowcolor{Gray}
10 & 0.83 & \textbf{0.85} & \textbf{0.92} & 0.87 & 1.05 & \textbf{0.97} & 1.20 & 1.24 & \textbf{1.07} & 0.97 \\
11 & 0.84 & \textbf{0.88} & 0.85 & \textbf{0.91} & 1.48 & \textbf{1.48} & 1.47 & 0.99 & 0.85 & \textbf{0.90} \\
\rowcolor{Gray}
12 & \textbf{0.87} & 0.82 & 0.92 & \textbf{0.93} & 0.58 & \textbf{0.58} & 0.59 & 0.37 & 0.39 & \textbf{0.37} \\
\hline
\end{tabular}
\caption{Results of atlas-based (AB) and neural-network-based (NN) semantic segmentation approaches on the testing set. The bold-face numbers represent the best results for each subject (closest to $1$ for Dice similarity, closest to true value for the biomarkers).}
\label{tab:beast}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth]{Result-Masks/results-horiz.png}
\caption{Comparison of the true masks (top), atlas-based predicted (middle), and neural-network-based predicted (bottom) for all subjects in the test data.}
\label{fig:result-masks}
\end{figure}
The bars chart in \textbf{Figure \ref{fig:test-result-plot}} can offer more visual insight into which method did better on each patient. Some other representations and related results of \textbf{Table \ref{tab:beast}} are shown in \textbf{Table \ref{tab:analytics}}, \textbf{Figure \ref{fig:full-scatter}} and \textbf{Figure \ref{fig:box}}.
\begin{figure}[t]
\centering
\begin{tabular}{cc}
\multicolumn{2}{c}{\textbf{Dice similarity to true masks}}\\
{}&{}\\
Brain Stem & Cerebellum\\
\includegraphics[width=.5\textwidth]{bs-dice.png} &
\includegraphics[width=.5\textwidth]{cb-dice.png}\\
{}&{}\\
\multicolumn{2}{c}{\textbf{Relative biomarker difference from true values}}\\
{}&{}\\
Brain Stem & Cerebellum\\
\includegraphics[width=.5\textwidth]{bs-biom.png} &
\includegraphics[width=.5\textwidth]{cb-biom.png}\\
\end{tabular}
\caption{Bar graph displaying atlas-based (blue) and neural-network-based (orange) results from testing set in terms of Dice similarity and relative biomarker error when compared to target (manually generated) results. Note that higher values are better for the Dice similarity while a lower difference is desirable for biomarkers. Negative relative differences indicate that manually generated results are larger than those predicted (and vice versa).}
\label{fig:test-result-plot}
\end{figure}
\begin{table}[t]
\centering
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{Analytics} &
\multicolumn{2}{c|}{BS DICE} &
\multicolumn{2}{c|}{CB DICE} &
\multicolumn{3}{c|}{BS BIOM} &
\multicolumn{3}{c|}{CB BIOM} \\
\cline{2-11}
& AB & NN & AB & NN & True & AB & NN & True & AB & NN \\
\hline
Mean: & 0.82 & \textbf{0.86} & \textbf{0.86} & 0.86 & 1.06 & 1.04 & \textbf{1.05} & 0.80 & 0.71 & \textbf{0.72} \\
Avg. Error: & 0.18 & \textbf{0.14} & \textbf{0.14} & 0.14 & --- & \textbf{0.04} & 0.04 & --- & 0.16 & \textbf{0.04} \\
\hline
\end{tabular}
\caption{Summary of results on testing set. Dice error computed as $1-D_{Dice}$. Biomarker error (relative) computed as $|\text{prediction} - \text{true}| / \text{true}$.}
\label{tab:analytics}
\end{table}
\begin{figure}[t]
\centering
\begin{tabular}{c}
\textbf{Relative Biomarker Difference versus Dice Similarity}\\
\includegraphics[width=.8\textwidth]{full-results-scatter.png}\\
\end{tabular}
\caption{Scatter plot showing Dice similarity versus absolute relative biomarker difference split between both regions (x's for brain stem, squares for cerebellum) and both methods (blue for atlas-based, red for neural-network-based).}
\label{fig:full-scatter}
\end{figure}
\begin{figure}[t]
\centering
\begin{tabular}{c}
\textbf{Error Spread in Dice and Biomarker Similarities}\\
\includegraphics[width=.8\textwidth]{Whisker.png}\\
\end{tabular}
\caption{Box plot showing errors of predicted masks and biomarkers. The Dice error (left) is calculated as $1 - D_{Dice}$ between the manually and automatically produced masks. The biomarker error (right) is a relative error between the manually and automatically produced biomarkers ($|\text{prediction} - \text{true}| / \text{true}$). These errors are split between the brain stem and cerebellum, where blue indicates the errors of the atlas-based method and red indicates the error of the neural-network-based method.}
\label{fig:box}
\end{figure}
\section{Discussion}
In this paper, we developed and compared two methods -- one atlas-based and one neural-network-based -- that identify the cerebellum and brain stem in a given MR image.
The goal of the segmentation is to produce a displacement biomarkers that can aid in the diagnosis of CMI. After developing and evaluating these methods, we found interesting differences between the accuracy of each method in terms of the mask similarity and biomarker similarity.
In 9 of the 12 test cases for the brain stem and 8 out of 12 cases for the cerebellum, the neural-network-based method produced higher similarity masks (\textbf{Table \ref{tab:beast}}). For the brain stem, this corresponded to an average mask similarity that was higher for the neural-network-based method. However, the atlas-based method produced slightly more similar masks \textit{on average} for the cerebellum than the neural-network-based method, even though it performed relatively worse in a majority of cases.
In terms of biomarker similarity, although the neural-network-based approach produced better segmentations (in 9 of 12 cases and on average), the atlas-based method produced more accurate biomarkers in 8 of 12 cases and on average in terms of lowest relative error. For the cerebellum, the neural-network-based method produced better biomarkers in 10 of 12 cases, and on average.
These differences between the Dice similarity and biomarker similarity for the brain stem show that although the neural-network produced segmentations that were more similar to those manually drawn in a pixel-wise sense, the atlas-based method's segmentations led to an average of displacement values that more closely matched the average displacement biomarker from manual segmentations. The differences between which method performed best in a majority of cases versus on average are due to the spreads of error across these test subjects. \textbf{Figures \ref{fig:test-result-plot}} and \textbf{\ref{fig:box}} and \textbf{Tables \ref{tab:beast}} and \textbf{\ref{tab:analytics}} present these results in further detail.
The methods produce comparable results, and if a just single method were to be chosen to complete all analyses, we would recommend the neural-network-based approach, as it produces the best segmentations and biomarkers on average, notably outperforming the atlas-based method overall. However, a combination of the two -- with the atlas-based method used to produce the brain stem biomarker and the neural-network-based method used to produce the cerebellum biomarker -- may be more likely produce the best displacement biomarkers for a new patient based on the above results.
We anticipate that further work and further data may improve the accuracy of the neural-network-based segmentation more quickly than it would atlas-based. This further work on the neural-network-based method could include adding a regularizer, similar to that used in the atlas-method (described in \textbf{Section \ref{Registration}}), that could eliminate non-contiguous pixels from predicted masks, resulting in even better segmentations. A multi-batch approach, that iteratively creates training and validation sets across non-testing data, could also further help improve results, to maximally take advantage of the currently relatively small data set.
As more DENSE images are collected and labeled, both methods may produce better results; the neural-network can be trained with a larger data set, further increasing the accuracy of the segmentations, and the bank of images the atlas-based method compares a new reference to could be more likely to find a close match in a larger bank of images that could also improve the registration and resulting masks.
An interesting possibility for further work may also lie in examining the accuracy of radiologist predicted masks. Though they were treated as truth in this study, manually drawn segmentations can vary in perhaps significant ways. In the context of Chiari diagnosis with a displacement biomarker, to draw accurate masks of the brain stem and cerebellum on a new MRI is an especially challenging problem; both the cerebellum and brain stem are bordered by regions of high movement cerebrospinal fluid (CSF). If borders are drawn too wide, those high-movement pixels may be included in the displacement averages and may inflate them inaccurately. We experimented with calculating biomarkers while algorithmically ignoring values above a certain threshold as a way to ignore high movement cerebrospinal fluid. This method did not help us match manually predicted biomarkers, but it may be a low-cost and helpful method to implement even when an expert radiologist is drawing segmentations to make them more accurately include only areas with brain-tissue scale movements.
\section*{Acknowledgements}
We want to sincerely thank Dr. Lars Ruthotto and Justin Smith for their mentorship and help throughout the project, as well as Dr. John Oshinski for providing us with this data set and collaborating with us. We would further like to thank all the mentors at Emory's 2021 REU/RET Program. This work was supported by the US National Science Foundation award DMS 2051019.
\bibliographystyle{abbrv}
|
\chapter{Introduction}\label{chap.intro}
Classically, a set relation $R$ on finitely many sets $\{A_{i}\}_{I}$ is a subset $R\subseteq\prod A_{i}$ of the product of those sets. A salient feature of set relations is that they may be compared: given $R, R'$ set relations on $\{A_{i}\}_{I}$, it is reasonable to ask whether $R\subseteq R'$. More still, we may note that this partial ordering underlies in particular a meet-semilattice structure on set relations on $\{A_{i}\}_{I}$; there is a maximal set relation $\prod A_{i}$ and given two set relations $R,R'$ we may compute their meet as $R\cap R'$.
All of the structure we have highlighted so far is ``local'', it is particular to a fixed family of sets $\{A_{i}\}_{I}$ and set relations thereupon. Should we allow ourselves to involve multiple such families, we see that set relations support further structure still. For example, given set relations $R \subseteq X \times Y$ and $S \subseteq Y \times Z$, their composite set relation $R\mathbin{\fatsemi}_{Y} S\subseteq X\times Z$ may be described by the formula \[ R \mathbin{\fatsemi}_{Y} S = \big\{(x,z) \mid \exists y [R(x,y)\wedge S(y,z)]\big\}\ . \] This form of composition interacts with all of the ``local'' structure we listed above, and is in fact well defined for arbitrarily finitely many set relations $R_{k}$ over arbitrarily finitely many shared indices $F\subseteq I$---although we would quickly exhaust the utility of our notation $R\mathbin{\fatsemi}_{Y} S$ above in an attempt to write this precisely. Further examination of such operations reveals that we wish to deal precisely with those operations and structures permitted by and expressible in \emph{regular logic}: that fragment of first order logic generated by equality ($=$), true ($\const{true}$), conjunction ($\wedge$), and existential quantification ($\exists$).
Having thus noted key aspects of the calculus of set relations, and their connection to regular logic, our motivating question is as follows.
\begin{question}\hypertarget{question}
What is an \emph{ergonomic} syntax for \emph{relations}?
\end{question}\newcommand{\hyperlink{question}{question}}{\hyperlink{question}{question}}
Let us now give meaning to the emphasised terms in the above question, beginning with the notion of \emph{relation}. Although we began our exploration by teasing apart the structure of set relations, we wish to find a suitably general syntax for \emph{any} class of objects supporting the structure of regular logic. Certainly at first we may straightforwardly generalise the notion of a set relation $R\subseteq\prod A_{i}$ to any sufficiently structured category by asking for a monomorphism $R\rightarrowtail\prod A_{i}$. However, we mean something broader even than this. Our interest is in allowing the type of things upon which relations are defined and the type of relations to be entirely different. For instance, we hope to also encompass conjunctive queries from database theory; essentially, those database queries that can be expressed using regular logic \cite{chandra1977optimal}. Another, abundant and indeed motivating source of examples where the type of relations differs from the type of objects is the class \emph{regular theories}.
A regular theory comprises a collection $\Sigma\text{-sort}$ of sorts, a collection $\Sigma\text{-rel}$ of relation symbols, and a collection $\mathbb T$ of axioms. From the first collection we build the notion of \emph{context}, a list of sorts $\sigma_{i}\in\Sigma\text{-sort}$ from which we are permitted to draw abstract variables. We then allow ourselves to form \emph{formulae} in contexts by inductively applying the structures of regular logic ($=$, $\const{true}$, $\wedge$, $\exists$) to abstract variables and relation symbols on those variables. Finally we impose an ordering of \emph{provability} according to the laws of deduction of regular logic, as well as our stated axioms. As formulae host all of the same structures and properties we extracted from set relations above, we see that we are naturally led to consider formulae in a regular theory as relations. Note however that contexts are the type of things upon which formulae are defined and so we must allow for a difference between our relations and the things upon which they are defined.
Having attended to the word \emph{relation}, let us now give concrete meaning to our question by elaborating \emph{ergonomic}. By this term mean firstly the following rigorous notions. A relation $R$ defined on some finite collection of objects $\{A_{i}\}_{I}$ does not naturally admit notions of domain and codomain. While it may be possible in practice to choose somehow a division of the $A_{i}$, nevertheless on our view an ergonomic syntax for relations should not impose arbitrary measures such as a privileging of certain objects as domain or codomain. In a similar vein, while we only displayed explicitly a binary composition $R\mathbin{\fatsemi}_{Y}S$ for relations above, we noted that arbitrarily finitely many relations may be composed, over arbitrarily finitely shared indices, simultaneously. An ergonomic syntax then should directly support such unbiased, multi-ary operadic compositions.
However, there are also some non-rigorous criteria we wish to ascribe to our notion of ergonomics. The syntax, whatever its form, should intuitively encode the various deductions of regular logic. For instance, it should be evident in the syntax that we can eliminate any conjunct $\exists_{x'}[x=x']$ whenever $x$ is already present in the context. Furthermore, given the operadic nature of composition, our notion of ergonomic includes also the requirement of a graphical syntax.
Now that our \hyperlink{question}{question}{} has been understood, we can turn our attention to the state of the art. The classical categorical syntax for relations is that of regular categories. These are categories which abstract enough of the structure of the category of sets to house regular logic, and thus to deal with relations. Unfortunately, by design regular categories do not capture the notion of relation we outlined above: a relation $R$ on objects $\{A_{i}\}_{I}$ in a regular category $\cat R$ is in particular an object $R\in\ob\cat R$ and so is necessarily of the same sort as the objects upon which it is defined. In order to remedy this, work was done on ``functionally complete `bicategories of relations'{}''\cite{carboni1987cartesian}\footnote{for another, ultimately equivalent approach in this general theme see \cite{freyd1990categories}}, or more recently and equivalently on \emph{relational po-categories} \cite{fong2019regular}---structures which form the progenitor of this paper. In common to both of these approaches is a privileging of relations to live between objects, instead of among them, thereby freeing the type of relations to differ from the type of objects. Moreover, these approaches benefit from sharing the same ``category theory'' as regular categories: \cite[Theorem 7.3]{fong2019regular} proves that $2$-category $\CCat{RlPoCat}$ of relational po-categories is equivalent to the $2$-category $\CCat{RgCat}$ of regular categories. While this viewpoint does admit a pleasing graphical syntax for relations, nevertheless it suffers from the technical requirement of privileging a sense of ``binarity'' for relations: as morphisms, relations require a choice of domain and codomain, and composition is presented as a binary operation.
To organise more complicated multi-way composites of relations, many fields have developed some notion of wiring diagram. A good amount of recent work, including but not limited to control theory \cite{bonchi2014categorical,baez2015categories,fong2016categorical}, database theory and knowledge representation \cite{bonchi2018graphical,patterson2017knowledge}, electrical engineering \cite{baez2018compositional}, and chemistry \cite{baez2017compositional}, all serve to demonstrate the link between these languages and an ergonomic syntax for relations.
In a similar vein, we desire to use an ergonomic syntax to describe those relational structures arising from regular theories. Consider in particular the regular theory of a pre-order. This regular theory has a single sort $S$, and a single relation symbol $P$. Formally its axioms, reflexivity and transitivity, are implications between certain formulae in the theory. However, an ergonomic syntax would instead allow us to describe these axioms as follows.
\begin{diagram*}
\node(A)[]{
\begin{tikzpicture}[inner WD]
\node(1)[minimum size=0, inner sep=0]{};
\draw(1) -- ++(-2,0) -- ++(2,0);
\node(2)[pack] at ($(1)+(6,0)$) {$P$};
\draw(2.west) -- ++(-0.5, 0);
\draw(2.east) -- ++(+0.5, 0);
\path(1) -- node(a)[midway]{$\leq$} ($(2)-(1,0)$);
\end{tikzpicture}
};
\node(B)[right=3cm of A.east,anchor=west]{
\begin{tikzpicture}[inner WD]
\node(1)[pack]{$P$};
\node(2)[below= of 1,pack]{$P$};
\coordinate(l) at ($(1)!0.5!(2)$){};
\draw(1) -- (2);
\draw(1.north) -- ++(0, +0.5);
\draw(2.south) -- ++(0, -0.5);
\node(3)[right=4 of l,pack]{$P$};
\draw(3.north) -- ++(0,+0.5);
\draw(3.south) -- ++(0,-0.5);
\path(l-|2.east) -- node[midway]{$\leq$} (3);
\end{tikzpicture}
};
\end{diagram*}
The above-left diagram is our desired syntactic representation of reflexivity: whenever we have a wire---which represents an abstract variable of the unique sort $S$---we may deduce the presence of a $P$ relation between that wire and itself. Similarly, in the above-right diagram we see our desired syntactic representation of $\exists_{y}[xPy\wedge yPz]\vdash xPz$, that is, transitivity: given two abstract variables and an existentially quantified variable shared between two copies of $P$, we may deduce a $P$ relation directly between the two abstract variables.
How then do we go about answering our \hyperlink{question}{question}{} and rigorously develop an ergonomic syntax for relations which affords us the above pictures? This paper and its companion \cite{grl2} are intended to serve as just such a means. Our first step, and the main thrust of this paper, is to introduce the notion of \emph{regular calculi}.
A regular calculus $\rc P$ comprises first the data of a symmetric monoidal \emph{po-category} $\cont P$---a category whose homs are posets and whose compositional and monoidal structures are monotonic; see \cref{def.pocats_funs_laxnts}. This po-category additionally has a structured notion of \emph{wiring diagram}, a purely combinatorial way to describe how objects are connected to one-another in $\cont P$ governed by cospans of sets and for which we develop a graphical syntax in \cref{chap.wires}. A regular calculus then additionally comprises the data of a \emph{right ajoint lax} po-functor $P$ from $\cont P$ to posets (\cref{def.ajax}). These pieces we assemble as follows. We think of the objects of $\Gamma\in\ob\cont P$ as contexts for relations in some relational theory, each poset $P(\Gamma)$ as the poset of relations in the context $\Gamma$ ordered by implication, the right adjoint lax structure on $P$ gives meets of relations $R\wedge R'\in P(\Gamma)$ and $\const{true}_{\Gamma}\in P(\Gamma)$, and each wiring diagram $\Gamma\to\Gamma'$ gives a method for converting relations in the context $\Gamma$ to relations in the context $\Gamma'$ by using equality ($=$), true ($\const{true}$), conjunction ($\wedge$), and existential quantification ($\exists$).
The structured notion of wiring diagram in the symmetric monoidal po-category of contexts $\cont P$ of a regular calculus $\rc P$ automatically extends to a graphical language for describing relations which arise out of regular logic operations on others. For instance,
in a fixed context $\Gamma$ from which we draw variables, and from relations $R_{1}$, $R_{2}$, and $R_{3}$ of arity 3, 3, and 4 respectively, we might wish to construct a composite relation $S$ graphically by specifying how the various $R_{i}$ are connected and share variables. By using the graphical notation for regular calculi we develop here, we will be able to rigorously draw and interpret the following \emph{graphical term} of the regular calculus $\rc P$
\begin{diagram*}[][][penetration=0, inner WD, pack size=9pt, link size=2pt, scale=2, baseline=(out)]
\node[packs] at (-1.5,-1) (f) {$R_{3}$};
\node[packs] at (0,1.9) (g) {$R_{1}$};
\node[packs] at (1.5,-1) (h) {$R_{2}$};
\node[outer pack, inner sep=34pt] at (0,.2) (out) {};
\node[above left=0.25 and 0.25 of out]{$S$};
\node[link] at ($(f)!.5!(h)$) (link1) {};
\node[link] at (-2.4,-.25) (link2) {};
\node[link] at ($(f.75)!.5!(g.-135)$) (link3) {};
\draw[fr_out] (out.270) to (link1);
\draw[fr_out] (out.190) to (link2);
\draw[fr_out] (out.155) to (link3);
\draw[fr_out] (out.-35) to (h.-30);
\draw[to_out,fr_out] (out.15) to[out=-165,in=-110] (out.70);
\draw (f.30) to[out=0,in=130] (link1);
\draw (f.-30) to[out=0,in=-130] (link1);
\draw (h.180) to (link1);
\draw (g.-60) to (h.120);
\draw (f.45) to (g.-105);
\draw (f.75) to (link3);
\draw (g.-135) to (link3);
\end{diagram*}
and recognise that it represents the following relation
\[
S(y,z,z',x,x',z'') = \exists\, \tilde{x},\tilde{y}, \left[R_1(\tilde{x},\tilde{y},y) \wedge R_2(x',\tilde{x},x) \wedge R_3(y,\tilde{y},x',x') \wedge (z=z')\right]\ .
\] The structure governing these graphical terms, and the notion of wiring diagram, is here formalised as a \emph{supply for the po-prop} $\mathbb{W}$. This object $\mathbb{W}$ is morally the po-category $\poCat{Cospan}^{\tn{co}}$ of cospans of finite sets, and affords us intuitive encodings of the various theorems of regular logic. Indeed, that our wiring diagrams admit definition purely combinatorially in terms of four generators and some relations is a central aspect of our development; see \cref{sec.depict_ww}. For example, the relation $\exists_{x'}[x=x']$ is rendered as a graphical term as below left, and one of the equations in $\mathbb{W}$ dictates precisely the equality below, therefore enabling us to eliminate any occurrences as indicated below right.
\begin{diagram*}
\node(A){
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(-1,0);
\draw (mu) to[out=60, in=180] +(.5,.5) node[link] {};
\draw (mu) to[out=-60, in=180] +(1,-.5) -- +(.5,0);
\end{tikzpicture}
};
\node(B)[right=of A.east,anchor=west]{
\begin{tikzpicture}[WD]
\draw (0,0) -- (2,0);
\end{tikzpicture}
};
\path (A) -- node[midway]{$=$} (B);
\end{diagram*}
Let us now return to our motivating \hyperlink{question}{question}{}. Note that regular calculi do not impose any constraints on the type of relations, and moreover do not enforce a notion of ``binarity'': relations have no chosen domain or codomain, and composition---as seen in the diagram above---is fully multi-ary and operadic. Thus, in making the notion of graphical term rigorous through the course of this work, we will see that regular calculi do present an ergonomic syntax for relations.
To claim that regular calculi are truly a syntax for relations, however, we must additionally establish that they may stand in wherever we would have otherwise used a different syntax to capture regular logic and relations.
The first such case we wish to draw attention to is the that of models for regular theories. Given a regular theory $(\Sigma,\mathbb T)$, one can describe what it means to model the theory in a regular category $\cat R$ and this gives a functor $\mathbb T\model(-)\colon \CCat{RgCat}\to\CCat{Cat}$. A major triumph of the classical work on categorical regular logic is to provide a representation\footnote{More correctly, a bi-representation for the $2$-functorial extension} for this functor: there is a \emph{syntactic regular category} $\cat C^{\mathrm{reg}}_{\mathbb T}$ associated to $(\Sigma,\mathbb T)$ and an appropriately natural family of equivalences $\CCat{RgCat}(\cat C^{\mathrm{reg}}_{\mathbb T},\cat R)\eqv\mathbb T\operatorname{-Mod}(\cat R)$---see, for instance, \cite{butz1998regular} for exposition in this vein.
We are able to give a complete analogy for regular calculi. Recall that we wish to understand regular calculi as housing a regular theory of relations. To this end, we expect that we should be able to extract from each regular calculus $\rc P$ a \emph{syntactic relational po-category} $\poCat{Syn}\rc P$ in analogy with the classical process $(\Sigma,\mathbb T)\mapsto\cat C^{\mathrm{reg}}_{\mathbb T}$ for regular theories to regular categories. Indeed, we provide such a construction which proceeds, in particular, by defining the objects of $\poCat{Syn}\rc P$ to be pairs $(\Gamma,\theta)$ of contexts $\Gamma\in\ob\cont P$ and relations $\theta$ in context $\Gamma$. In this way we see that $\poCat{Syn}$ models the syntax of the regular calculus closely. An immediate benefit of working categorically is that we are able to extend this construction to a $2$-functor $\poCat{Syn}\colon\CCat{RgCalc}\to\CCat{RlPoCat}$ from regular calculi to relational po-categories.
To complete the analogy then we need two additional ingredients. Thinking once more of regular calculi as housing regular theories of relations, there is an appropriate notion of \emph{model for a regular calculi} and in \cref{def.models} we describe a $2$-functor $\rc P\model(\cdot)\colon\CCat{RlPoCat}\to\CCat{Cat}$ sending a relational po-category $\ccat{R}$ to the $2$-category of models of $\rc P$ in $\ccat{R}$. Then, given a regular theory $(\Sigma,\mathbb T)$, in \cref{con.rgcalc_rgtheory} we prove the existence of an associated regular calculus $(\cont{\mathbb T},F_{\mathbb T})$.
With the ingredients prepared we are able to give the analogous theory of models. First, using tools developed in the companion, we prove that our notion of syntactic po-category suitably generalises the classical notion. This result appears as \cref{thm.syn_equiv} below.
\begin{altheorem}
Given a regular theory $(\Sigma,\mathbb T)$, the syntactic po-category $\poCat{Syn}(\ccat{C}_{\mathbb T},F_{\mathbb T})$ of the regular calculus associated to the theory is equivalent to the po-category of relations $\poCat{Rel}(\cat C^{\mathrm{reg}}_{\mathbb T})$ of the regular category associated to the theory. That is, under the equivalence between relational po-categories and regular categories, the two constructions agree.
\end{altheorem}
Using this we prove additionally that our notion of model of a regular calculus suitably generalises the classical notion. This result appears as \cref{thm.model_equiv} below.
\begin{altheorem}
There is an equivalence of categories $(\ccat{C}_{\mathbb T},F_{\mathbb T})\model(\poCat{Rel}\cat R)\eqv \mathbb T\model(\cat R)$ appropriately natural in regular categories $\cat R$, where $\poCat{Rel}\ccat{R}$ is the po-category of relations of $\cat R$.
\end{altheorem}
In fact, the above theorem is proven as a corollary of another result which shows that we may suitably substitute regular calculi as a syntax for relations in another fashion. Let us now turn our attention to the interaction between regular calculi and relational po-categories.
We have already demonstrated that regular calculi provide a home for regular theories, and so we would be remiss to omit addressing the question of regular categories---equivalently the relational po-categories. We show that from a relational po-category $\ccat{R}$ one may construct a regular calculus $\poCat{Prd}\ccat{R}$ through a process we call \emph{taking predicates}. Given a regular category $\cat{R}$, viewed as its po-category of relations $\poCat{Rel}\cat{R}$, our construction yields the regular calculus $\poCat{Prd}\poCat{Rel}\cat{R}$ whose contexts are the objects of $\cat{R}$ and whose relations $\theta$ in context $r$ are precisely the sub-objects $\theta\rightarrowtail r$, exactly as we might have hoped. This assignment of relational po-categories $\ccat{R}\mapsto\poCat{Prd}\ccat{R}$ to regular calculi we extend to a $2$-functor $\poCat{Prd}\colon\CCat{RlPoCat}\to\CCat{RgCalc}$.
Recall however that we have the opposed $2$-functor $\poCat{Syn}\colon\CCat{RgCalc}\to\CCat{RlPoCat}$ of the syntactic po-category construction, and so with $\poCat{Prd}$ above we prove our first comparison theorem---appearing as part of \cref{thm.main} in this paper and whose proof is elaborated in the companion.
\begin{altheorem}
The $2$-functors $\poCat{Syn}\colon\CCat{RgCalc}\to\CCat{RlPoCat}$ and $\poCat{Prd}\colon\CCat{RlPoCat}\to\CCat{RgCalc}$ form a bi-adjunction $\poCat{Syn}\dashv_{\mathrm{bi}}\poCat{Prd}$.
\end{altheorem}
By a \emph{bi-adjunction} here we mean the appropriate notion of $2$-dimensional adjunction where the equations on the unit and co-unit now hold only up to invertible $3$-dimensional morphisms, each of which satisfy some appropriate equation. As adjunctions transfer a wealth of category-theoretic aspects, so too do bi-adjunctions---this bi-adjunction affords us a rich comparison of the $2$-category theory of regular calculi and relational po-categories. However, from the point of view of providing an ergonomic syntax for relational po-categories, or equivalently regular categories, it is as yet unsatisfactory.
Without a stronger theorem we cannot be sure that by working syntactically in the regular theory carried by the regular calculus $\poCat{Prd}\ccat{R}$ we are in fact working in the relational po-category $\ccat{R}$ itself. That is, we wish to know: is there an equivalence between the syntax $\poCat{Syn}\poCat{Prd}\ccat{R}$ given by our regular calculus approach and the relational po-category $\ccat{R}$? To this end we prove as part of our main theorem the following general answer to this question.
\begin{altheorem}
The co-unit of the bi-adjunction $\poCat{Syn}\dashv_{\mathrm{bi}}\poCat{Prd}$ is an adjoint equivalence, so relational po-categories are pseudo-reflective in regular calculi.
\end{altheorem}
By \emph{pseudo-reflective} here we mean the appropriate $2$-dimensional version of the analogous ordinary category theoretic notion of a fully-faithful inclusion of sub-categories which admits a left adjoint. Among other things, this result tells us that we may freely embed relational po-categories and their morphisms into regular calculi by taking predicates, and that all graphical manipulations and syntactical operations hold in the original object: $\poCat{Syn}\poCat{Prd}\ccat{R}\eqv\ccat{R}$.
As \cite{fong2019regular} proves that regular categories and relational po-categories have equivalent $2$-categories, we have in fact also obtained the following theorem.
\begin{altheorem}
The $2$-category of regular categories is pseudo-reflective in regular calculi.
\end{altheorem}
Finally, the presence of this syntax-semantics pseudo-reflection allows us to deduce further interesting consequences: in \cref{cor.ladj_birep} we prove that taking the regular category of left adjoints in a relational po-category is suitably represented as a $2$-functor, and in \cref{thm.reg_lex} we use it to give an alternate proof of an adjunction between finite limit categories and regular categories.
\section{Outline}
We have striven, where reasonable, to render this paper as self-contained as possible. Where we make use of results from the body of work of \cite{fong2019hypergraph,fong2019regular,fong2019supplying} we are careful to cite them or reprove them in our context. With that in mind, this paper is organised as follows.
\cref{chap.background} presents the setting of symmetric monoidal po-categories and morphisms thereof in which we will be working. \cref{chap.wires} introduces the po-prop for wiring $\mathbb{W}$, the notion of supply for a po-prop, and develops our graphical notation for $\mathbb{W}$ as well as for symmetric monoidal po-categories which supply it. Then in \cref{chap.regular_calculi} we define the central structures of this paper, the regular calculi and their morphisms, by way of the notions of right adjoint monoid and right ajax po-functor. Additionally, among our examples in this section, we show that regular theories give rise to regular calculi. In \cref{chap.relational_pocats} we recall the axiomatisation of relational po-categories, and construct and study the $2$-functor $\poCat{Prd}$ which takes a relational po-category to its regular calculus of predicates. In \cref{chap.graphical_reglog} we develop our graphical formalism for regular calculi by defining graphical terms and establishing key lemmas which afford us intuitive means of graphical reasoning and manipulations, with examples drawn from regular theories and relational po-categories. Additionally in that section we sketch the construction of the syntactic po-category of a regular calculus. Finally in \cref{chap.preview} we preview some of the results and details that appear in the companion, including the statement of our main theorem \cref{thm.main} and its various corollaries.
\section{Related work}
Graphical formalisms for fragments of first order logic are an old subject, dating back to the existential graphs of Pierce in the late 19th Century \cite[Volume 4, Book II]{peirce1974collected}. More recently, Brady and Trimble \cite{brady2000categorical,brady2000string} and Haydon and Sobocinski \cite{haydon2020compositional} have sought to reinterpret Pierce's work within a categorical setting, providing a foundation for Pierce's diagrams through the string diagrams for monoidal categories defined by Joyal and Street \cite{joyal1991geometry}.
In the context of regular logic, Bonchi, Seeber, and Sobocinski's work on graphical conjunctive queries \cite{bonchi2018graphical}, as well as the last two authors' precursor version of this work \cite{fong2020string}, give alternate categorical, diagrammatic approaches to regular logic. Like regular calculi, these categorical approaches to regular logic and relational structures feature Frobenius monoids as a central way of capturing equality of variables or, graphically, the lines that can be drawn between relation symbols. A difference in our emphasis is that our diagrams are not simply string diagrams in a monoidal category; they are distinct, combinatorial objects inspired by an operadic notion of composition.
Brady and Trimble investigate Pierce's diagrams via Lawvere's notion of hyperdoctrine \cite{lawvere1969adjointness}, a foundational tool in categorical logic. A hyperdoctrine studies a logical system as a fibration, presenting it as a collection of posets indexed by a suitable base category. The required structure of this base category, such as having finite products, allows interpretation of certain syntactic aspects, while interpretations of connectives and quantifiers are studied through adjunctions. In particular, regular logic is modelled through the notion of an elementary existential doctrine \cite{lawvere1970equality}.
An elementary existential doctrine is a functor from the opposite of a category with finite products to the category of meet-semilattices, such that certain assignments of the functor have adjoints, and such that these adjoints obey Beck-Chavelley and Frobenius reciprocity conditions. Similar to elementary existential doctrines, regular calculi index posets. The difference lies in the choice of base category: while the base category of a hyperdoctrine is modelled on aspects of the category of finite sets and functions---that is, the free finite product category---the base category of a regular calculus is modelled on the po-category of finite sets and co-relations (certain equivalence classes of cospans)---or, more precisely, the free relational po-category.
Our regular calculus approach then expresses the logical structure through adjunctions in the base category and the right ajax condition. This leads to a more compact definition than that of an elementary existential doctrine, without the Beck-Chavelley and Frobenius reciprocity axioms. Indeed, we believe the regular calculus definition provides some motivation for how these axioms arise.
Many of the aspects of our work with regular calculi here have precursors in work done with elementary existential doctrines. The syntactic category, also known as a Lindenbaum--Tarski category of a logical theory or hyperdoctrine is a well-known construction (see for example \cite{butz1998regular}); our syntactic po-category is a relational analogue. Our completion theorem lies within the tradition of the completion theorems of Maietti and Rosolini \cite{maietti2013quotient}, with Trotta's recent paper on existential completions of primary doctrines perhaps the closest analogue \cite{trotta2020existential}.
Finally, the adjunction between hyperdoctrine-style logical presentations and categorical structure within the context of regular logic is also explored by Bonchi, Santamaria, Seeber, and Sobocinski \cite{bonchi2021doctrines}, who detail an adjunction between elementary existential doctrines and cartesian bicategories.
\section{Acknowledgements}
The second- and third-named authors would like thank Paolo Perrone for comments that have improved this article and Christina Vasilakopoulou for finding an error in a previous version, which led us to this fully 2-categorical formulation. The first-named author would like to thank Emily Riehl for conversations which informed the present structure of this paper and the companion. We acknowledge support from AFOSR grants FA9550-17-1-0058 and FA9550-19-1-0113.
\chapter{Background on symmetric monoidal po-categories}\label{chap.background}
To develop the theory of regular calculi and to state and prove our main results we will make extensive use of the language of symmetric monoidal po-categories and various higher morphisms thereof. In \cref{sec.two_cats} below we recall briefly the needed notions of \emph{oplax-natural transformation}, \emph{modification}, and \emph{adjunction in a $2$-category}. Readers familiar with these notions are invited to omit this section. Following this, in \cref{sec.po_categories} we will observe the several significant specialisations of these notions to the \emph{po-categorical} setting and cement terminology therein. Finally, in \cref{sec.symm_mom_po_cat} we will recall the notion of \emph{symmetric monoidal po-category} and various morphisms thereof.
Before we proceed with this background, let us pause a moment to record the salient features of our notation in this paper. While we endeavour to be standard in most aspects, it may nevertheless be useful to note the following.
\begin{itemize}
\item By a poset we mean a partially ordered set, that is, $x\leq y\wedge y\leq x\Rightarrow x=y$.
\item We typically denote composition in diagrammatic order, so the composite of $f\colon A\to B$ and $g\colon B\to C$ is $f\mathbin{\fatsemi} g\colon A\to C$. We often denote the identity morphism $\id_c\colon c\to c$ on an object $c\in\cat{C}$ simply by the name of the object, $c$. Thus if $f\colon c\to d$, we have $(c\mathbin{\fatsemi} f)=f=(f\mathbin{\fatsemi} d)$.
\item We may denote the unique map from an object $c$ to a terminal $1$ as $!\colon c \to 1$, and we denote the top element of any poset $P$ by $\const{true}\in P$.
\item We denote the universal map into a product by $\pair{f,g}$ and the universal map out of a coproduct by $\copair{f,g}$.
\item Given a natural number $n\in\mathbb{N}$, we write $\ord n$ for the set $\{1,2,\ldots,n\}\in\Cat{FinSet}$; in particular $\ord{0}=\varnothing$.
\item We will write $c\tpow n$ in a monoidal category to denote the left-associated $n$-fold iterated binary tensor product $(\cdots((c\otimes c)\otimes c)\cdots)\otimes c$.
\item Given a lax monoidal functor $F\colon\cat{C}\to\cat{D}$, we denote the \emph{laxators} by $\varphi\colon I\to F(I)$ and $\varphi_{c,c'}\colon F(c)\mathbin{\otimes'} F(c')\to F(c\otimes c')$ for objects $c,c'\in\ob\cat{C}$. If $F$ is strong, then we will make use of the same notation, but refer to these maps as \emph{strongators} instead.
\item Where our arguments make use of more than one dimension, we will write the morphisms with Latin letters, the $2$-morphisms with Greek letters, and the $3$-morphisms with Hebrew letters. For instance, $2$-functors will be denoted by $F$, $G$, \ldots, oplax-natural transformations will be denoted by $\alpha$, $\beta$, \ldots, and modifications will be denoted by $\aleph$, $\beth$, \ldots
\end{itemize}
\section{Background on \texorpdfstring{$2$}{2}-categories}\label{sec.two_cats} We will take for granted the notion of $2$-category and $2$-functor, but briefly recall here the definitions of higher morphisms between these. The reader already comfortable with such notions is nevertheless encouraged to review the various specialisations obtained in the \emph{po-categorical} setting in \cref{sec.po_categories}, and the later background on \emph{symmetric monoidal po-categories} in \cref{sec.symm_mom_po_cat}.
\begin{definition}[Oplax-natural transformations \& modifications]\label{def.strong_twont}
Given a pair of parallel $2$-functors $F,G\colon\cat{K}\to\cat{L}$, an \define{oplax-natural transformation} $\alpha\colon F\Rightarrow G$ comprises the data of object components $\alpha_c\colon Fc\to Gc$ for each object $c\in\ob\cat{K}$, and morphism components $\alpha_h\colon (\alpha_c\mathbin{\fatsemi} Gh)\Rightarrow(Fh\mathbin{\fatsemi} \alpha_{c'})$ for each morphism $h\colon c\to c'$ of $\cat{K}$. These morphism components are required to be natural with respect to $2$-morphisms of $\cat{K}$, and are required to be compatible with identity morphisms and composition in $\cat{K}$. For details see, for example, \cite[Definition 4.2.1]{JohYau}.
An oplax-natural transformation $\alpha$ is \define{pseudo-natural} when each morphism component $\alpha_{h}$ is a $2$-isomorphism, and is \define{$2$-natural} when each morphism component $\alpha_{h}$ is an identity.
A \define{modification} $\aleph\colon \alpha\threecellarr\beta$ between oplax-natural transformations $\alpha,\beta\colon F\Rightarrow G$ comprises the data of object components $\aleph_c\colon\alpha_c\Rightarrow\beta_c$ in $\cat{L}(Fc,Gc)$ for each object $c\in\ob\cat{K}$, which are required to be compatible with the morphism components of $\alpha$ and $\beta$. For details see, for example, \cite[Definition 4.4.1]{JohYau}.
\end{definition}
Recall that, given a $2$-category $\cat{K}$, an \define{adjunction in $\cat{K}$} consists of a pair of objects $c,d\in\ob\cat{K}$, a pair of morphisms $l\colon c\to d$ and $r\colon d\to c$, and a pair of 2-morphisms $\eta\colon d\Rightarrow (l\mathbin{\fatsemi} r)$ and $\epsilon\colon (l\mathbin{\fatsemi} r)\Rightarrow c$ such that the following pair of diagrams, the \define{triangle equalities}, are rendered commutative.
\begin{diagram}[][][baseline=(b.base)]\label{eqn.adjunction}
\node(1)[]{$l$};
\node(2)[right= of 1]{$lrl$};
\node(3)[below= of 2]{$l$};
\draw[n](1)to node[la,above]{$\eta\mathbin{\fatsemi} l$}(2);
\draw[n](2)to node[la,right]{$l\mathbin{\fatsemi}\epsilon$}(3);
\draw[d](1)to(3);
\node(4)[right= 3cm of 2]{$r$};
\node(5)[below= of 4]{$rlr$};
\node(6)[right= of 5]{$r$};
\draw[n](4)to node(b)[la,left]{$r\mathbin{\fatsemi}\eta$}(5);
\draw[n](5)to node[la,below]{$\epsilon\mathbin{\fatsemi} r$}(6);
\draw[d](4)to(6);
\end{diagram}
One may verify that adjunctions compose, and so by $\Funr{LAdj}\cat{K}$ we denote the $1$-category with the same objects as $\cat{K}$ and whose morphisms are the data of left adjoints $(l,r,\eta,\epsilon)$ in $\cat{K}$.
For given data $(l,r,\eta,\epsilon)$ as above, the property of being an adjunction is expressed equationally in the compositions of the ambient $2$-category. As such, we obtain the following lemma.
\begin{lemma}
Let $F\colon\cat{K}\to\cat{L}$ be a $2$-functor. The assignment $(l,r,\eta,\epsilon)\mapsto (Fl,Fr,F\eta,F\epsilon)$ sends adjunctions in $\cat{K}$ to adjunctions in $\cat{L}$, and so gives rise to a functor between the categories of left adjoints, $\Funr{LAdj} F\colon\Funr{LAdj}\cat{K}\to\Funr{LAdj}\cat{L}$. Moreover, this assignment $F\mapsto\Funr{LAdj} F$ of $2$-functors itself functorial in $2$-functors and so extends to a functor $\Funr{LAdj}\colon2\!\CCat{Cat}\to\CCat{Cat}$.\hfill$\qed$
\end{lemma}
In fact more is true, $\Funr{LAdj}$ is a $2$-functor when the $2$-morphisms in $2\!\CCat{Cat}$ are themselves required to be left adjoints, but we will not need this fact in this generality.
\section{Po-categories}\label{sec.po_categories}
The theory of $2$-categories specialises significantly to the context of \emph{po-categories}, and so we recall briefly the appropriate definitions now.
\begin{definition}[Po-category]\label{def.pocats_funs_laxnts}
A \define{po-category $\ccat{C}$} is a locally-posetal $2$-category, that is, it is an ordinary $2$-category $\ccat{C}$ for which the category $\ccat{C}(c,c')$ of morphisms between any two objects is a partially ordered set.
A \define{po-functor} $F\colon\ccat{C}\to\ccat{D}$ between po-categories is an ordinary $2$-functor, but we may summarise this structure by requiring that $F$ is an ordinary functor of the underlying $1$-categories and that the functions $F_{c,c'}\colon\ccat{C}(c,c')\to\ccat{D}(Fc,Fc')$ are monotonic for all objects $c,c'\in\ccat{C}$.
An \define{oplax-natural transformation} $\alpha\colon F \Rightarrow G$ between po-functors $F,G\colon\ccat{C}\to\ccat{D}$ is an ordinary oplax-natural transformation between the $2$-functors $F$ and $G$. However, all of the compatibility conditions are degenerate and so the data is merely a collection of morphisms $\alpha_c\colon Fc \to Gc$ which satisfy $F(f) \mathbin{\fatsemi} \alpha_{c'}\leq \alpha_c \mathbin{\fatsemi} G(f)$ for all morphisms $f\colon c \to c'$ of $\ccat{C}$. In particular, a \define{$2$-natural transformation} of po-functors is merely a natural transformation of the underlying functors.
Modifications are especially degenerate. Given parallel oplax-natural transformations $\alpha,\beta\colon F\Rightarrow G$, we write $\alpha\leq\beta$ if for each $c\in\ccat{C}$ there is an inequality $\alpha_c\leq\beta_c$ in $\ccat{D}(Fc,Gc)$ between $c$-components. Thus we are motivated in writing $[\ccat{C},\ccat{D}]$ for the po-category of po-functors, oplax-natural transformations, and modifications; we call it the \define{po-category of po-functors from $\ccat{C}$ to $\ccat{D}$}.
\end{definition}
\begin{notation}[Po-categories]
To distinguish po-categories from $1$-categories we will write po-categories with double-struck letters, as in $\ccat{C},\ccat{D},\ldots$, and reserve script for $1$-categories, as in $\cat{C},\cat{D},\ldots$
\end{notation}
Note that in any $2$-category, any two right adjoints to a given morphism are isomorphic, so in a po-category, a given morphism has \emph{at most one} right adjoint.
\begin{definition}[Left adjoint oplax-natural transformation]\label{def.ladj_lax_nt}
Let $\ccat{C}$ and $\ccat{D}$ be po-categories. A \define{left adjoint oplax-natural transformation} is a left adjoint in the po-category $[\ccat{C},\ccat{D}]$ of \cref{def.pocats_funs_laxnts}.
\end{definition}
As a consequence of the posetal nature of the po-category of po-functors $[\ccat{C},\ccat{D}]$ we may freely pass the structure of left adjointness between oplax-natural transformations and their components, in the following sense.
\begin{lemma}\label{lemma.adj_in_pocat}
Let $F,G\colon\ccat{C}\to\ccat{D}$ be po-functors, and let $\lambda\colon F\Rightarrow G$ and $\rho\colon F\Rightarrow G$ be opposed oplax-natural transformations. Then $\lambda$ is left adjoint to $\rho$ if and only if for each $c\in\ccat{C}$ the component $\lambda_c\colon Fc\to Gc$ are left adjoint to $\rho_c\colon Gc\to Fc$ in $\ccat{D}$.
\end{lemma}
\begin{proof}
The forward direction is true even for $2$-categories that aren't locally posetal; the backwards direction holds since the uniqueness of $2$-morphisms in a po-category implies that the triangle equalities \cref{eqn.adjunction} hold trivially.
\end{proof}
Observe that an invertible modification whose eventual codomain is a po-category is necessarily an equality. As such, between po-categories the notions of $2$-dimensional and $1$-dimensional equivalence coincide.
\begin{definition}\label{def.po_equiv}
An \define{equivalence of po-categories} is an equivalence in the $1$-category of po-categories and po-functors. We say that a po-functor $F\colon\ccat{C}\to\ccat{D}$ is \define{fully-faithful} when the morphism $F_{c,c'}\colon\ccat{C}(c,c')\to\ccat{D}(Fc,Fc')$ is an isomorphism of posets for each $c,c'\in\ob\ccat{C}$. Furthermore, we say that \define{$F$ is essentially surjective} if one can construct a function sending $d\in\ob\ccat{D}$ to a pair $(c\in\ob\ccat{C}, Fc\iso d)$, and we will tend to leave the function implicit.
\end{definition}
We will take for granted the following extension of the classical result relating fully-faithful essentially surjective functors and equivalences.
\begin{lemma}
Given a po-functor $F\colon\ccat{C}\to\ccat{D}$, the data of an equivalence on $F$ is equivalently the data of essential surjectivity of $F$ and the property of fully-faithfulness for $F$.\hfill$\qed$
\end{lemma}
\section{Symmetric monoidal po-categories}\label{sec.symm_mom_po_cat}
We will have a great deal of use for symmetric monoidal po-categories. These objects may be viewed as 3-categories which are ``petite'' in two dimensions: they are locally po-categorical, and they have only one object. While it is notoriously difficult to give a full definition of monoidal bi-category, all of the coherence issues degenerate for po-categories. As such it is conceptually simpler to think of symmetric monoidal po-categories instead as symmetric monoidal $1$-categories with extra structure: hom-sets are equipped with an order, and the monoidal operation is monotonic on morphisms.
\begin{definition}[Symmetric monoidal po-category]\label{def.symm_mon_po}
A \define{symmetric monoidal structure} on a po-category $\ccat{C}$ consists of a symmetric monoidal structure $(\otimes, I, \lambda, \rho)$ on its underlying $1$-category, such that $\otimes$ is a additionally a po-functor. That is, $(f_1\otimes g_1)\leq(f_2\otimes g_2)$ whenever $f_1\leq f_2$ and $g_1\leq g_2$. Recall that this means $\lambda$ and $\rho$ are automatically $2$-natural (\cref{def.pocats_funs_laxnts}).
A \define{strong symmetric monoidal po-functor} $(F,\varphi)\colon(\ccat{C},\otimes, I)\to(\ccat{D},\mathbin{\otimes'},I')$ is a po-functor $F\colon\ccat{C}\to\ccat{D}$ whose underlying functor is strong symmetric monoidal. Recall that this means that the \define{strongators} $\varphi_{c,c'}\colon Fc\mathbin{\otimes'} Fc'\to F(c\otimes c')$ are automatically $2$-natural.
We will frequently wish to apply the qualifier ``monoidal'' to various forms of natural transformations; by a \define{monoidal `\emph{adjective}' natural transformation} we will always mean an `\emph{adjective}' natural transformation whose components additionally obey the monoidal natural transformation conditions strictly. For example, a \define{monoidal left adjoint oplax-natural transformation} $\alpha\colon(F,\varphi)\Rightarrow(G,\psi)$ is a left adjoint oplax-natural transformation $\alpha\colon F\Rightarrow G$ whose components $\alpha_{c}\colon Fc\to Gc$ additionally obey the monoidal natural transformation conditions strictly---for instance $\varphi_{I}\mathbin{\fatsemi}\alpha_{I}=\psi_{I}$.
\end{definition}
\begin{notation}[Symmetry isomorphisms]
If $(\ccat{C},\otimes,I)$ is a symmetric monoidal po-category, $m,n\in\mathbb{N}$ are natural numbers, and $c\colon \ord{m}\times \ord{n}\to\ccat{C}$ is a family of objects in $\ccat{C}$, then there is a canonical natural isomorphism
\begin{equation}\label{eqn.symmetry}
\sigma\colon
\bigotimes_{i\in\ord{m}}\bigotimes_{j\in\ord{n}}c(i,j)\Too{\cong}
\bigotimes_{j\in\ord{n}}\bigotimes_{i\in\ord{m}}c(i,j)\ .
\end{equation}
We will refer to these $\sigma$ as the \define{symmetry} isomorphisms, though note that generally such isomorphisms involve associators and unitors too.
\end{notation}
\chapter{Supplying wires}\label{chap.wires}
Our work toward understanding graphical regular logic begins with the establishment of the supporting machinery which was developed in \cite{fong2019regular,fong2019supplying} and whose salient details we recall here. In order to render our graphical terms, we will need already the more primitive notion of \emph{wiring diagram}. The somehow prototypical case of these is the generic structure supporting a basic graphical calculus, viz., the \emph{po-prop} for wiring $\mathbb{W}$. Once we have gained some proficiency in this context, we will see how the notion of \emph{supply} for a po-prop allows us to understand wiring diagrams in any po-category which supplies $\mathbb{W}$.
\section{The po-prop \texorpdfstring{$\mathbb{W}$}{W} for wiring}\label{sec.prop_ww}
\begin{definition}[Po-prop]\label{def.props}
A \define{po-prop} is a symmetric strict monoidal po-category $\mathbb{P}$ whose monoid of objects is isomorphic to $(\mathbb{N},0,+)$. A \define{po-prop functor} $F\colon\mathbb{P}\to\mathbb{Q}$ is a bijective-on-objects symmetric strict monoidal po-functor.
\end{definition}
The first, and indeed most important example we intend to consider is the po-prop for wiring. Consider the symmetric monoidal $2$-category $(\poCat{Cospan}^{\tn{co}},\varnothing,+)$, i.e.\ the $2$-dual of cospans between finite sets.
\begin{definition}[$\mathbb{W}$]\label{def.wwiring}
The \define{po-prop for wiring}, $\mathbb{W}$, is the local poset reflection of the full and locally full sub-$2$-category of $(\poCat{Cospan}^{\tn{co}},\varnothing,+)$ spanned by the finite ordinals $\ord{n}$.
\end{definition}
Even though we have presented $\mathbb{W}$ in a way that emphasises ``binarity''---its morphisms have domains and codomains and its composition is that of an ordinary category---nevertheless as we shall see in the next section, and in particular in \cref{ex.wiring_diagram}, the combinatorial nature of $\mathbb{W}$ affords us a completely unbiased way to depict morphisms.
In \cref{prop.ww_explicit} we shall give a more explicit description of the hom posets of $\mathbb{W}$, after which we will exhibit the generating morphisms and relations for $\mathbb{W}$ graphically. Note that $\mathbb{W}$ is almost the prop of equivalence relations, also known as corelations \cite{coya2017corelations}, but without the ``extra'' law, which would equate the following cospans.
\begin{equation}
\label{eqn.extra_law}
\begin{gathered}
\ord0\to \ord0\leftarrow \ord0\\
\text{``truth''}
\end{gathered}\qquad\text{and}\qquad
\begin{gathered}
\ord0\to \ord1\leftarrow \ord0\\
\text{``inhabitedness''}
\end{gathered}
\end{equation}
We'll see later that equating these two morphisms bears the consequence of forcing every sort in a regular theory to be inhabited.
\begin{proposition}\label{prop.ww_explicit}
The hom-posets of $\mathbb{W}$ admit the following explicit description:
\[
\mathbb{W}(\ord m,\ord n)\cong
\begin{cases}
\{0\leq 1\}^{\tn{op}}&\tn{ if }m=n=0\\
\er^{\tn{op}}(m+n)&\tn{ if }m+n\geq 1
\end{cases}
\]
where $\{0\leq 1\}$ is the poset of booleans, and $\er(p)$ is the poset of equivalence relations on the set $\ord{p}$, ordered by inclusion.
\end{proposition}
\begin{proof}
For any $m,n$, may identify $\mathbb{W}(\ord m,\ord n)^{\tn{op}}$ with the poset reflection of $\poCat{Cospan}(\ord{m},\ord{n})$. But $\poCat{Cospan}(\ord{m},\ord{n})$ is the coslice category $\ord{m+n}/\Cat{FinSet}$, consisting of finite sets $S$ equipped with functions $\ord{m+n}\to S$. If $m+n=0$, then the poset reflection is that of $\Cat{FinSet}$, namely $\{0\leq1\}$; otherwise it may be identified with the poset of equivalence relations on $\ord{m+n}$. Indeed, every function $\ord{m+n}\to S$ factors as an epi followed by a mono, and every mono out of a nonempty set has a retraction.
\end{proof}
The proof just given suggests two further results on the structure of $\mathbb{W}$ which we record here.
\begin{corollary}\label{cor.jointly_surj_w}
Every morphism $\omega\colon\ord n\to\ord m$ in $\mathbb{W}$ with $m+n>0$ admits a unique representation as a jointly surjective cospan of finite sets $\ord n\to\ord n_{w}\leftarrow \ord m$.\hfill$\qed$
\end{corollary}
\begin{remark}\label{rem.canonical_decomp}
At this point we have succeeded in showing that there are in fact \emph{canonical choices} of representative cospans for every morphism of $\mathbb{W}$. When the domain or codomain are inhabited, then the above corollary uniquely determines a jointly epimorphic cospan, and in the remaining case we choose $\ord 0\to \ord 0\leftarrow\ord 0$ and $\ord 0\to \ord 1\leftarrow\ord 0$ as our representatives for the two distinct elements of $\mathbb{W}(\ord 0,\ord0)$.
\end{remark}
\begin{corollary}\label{cor.mor_tensors_ww}
In $\mathbb{W}$, for an arbitrary morphism $\omega\colon\ord n\to\ord m$, we have:
\begin{enumerate}
\item $\omega+\id_{\ord 0}=\omega$
\item $\omega+\eta\mathbin{\fatsemi}\epsilon=\omega$ if and only if $\omega\ne\id_{\ord0}$.\hfill$\qed$
\end{enumerate}
\end{corollary}
\begin{remark}
Later on we shall see that tensor of morphisms in $\mathbb{W}$ bears interpretation as a form of logical conjunction (\cref{con.rgcalc_rgtheory}), and so we may understand the above corollary, in the language of \eqref{eqn.extra_law}: logical conjunction by ``truth'' is the identity, while logical conjunction by ``inhabitedness'' is the identity when we have any context involving the implicit sort. In addition, the logical conjunction of ``inhabitedness'' with ``inhabitedness'' is once more ``inhabitedness''.
\end{remark}
Now that we have explicated the structure of the hom posets of $\mathbb{W}$, let us turn our attention to its morphisms, and for this purpose introduce a graphical notation prototypical of those to come. $\mathbb{W}$ may be generated by four morphisms, and we list these generating morphisms, their canonical cospan representatives in $\poCat{Cospan}^{\tn{co}}$ (see \cref{rem.canonical_decomp}), and their graphical icons in the table below.
\setlength{\belowrulesep}{0pt} \setlength{\tabcolsep}{8pt} \renewcommand{\arraystretch}{1.5}
\begin{equation}\label{eqn.generating_wires}
\begin{tabular}{|c|c|c|}\toprule[1pt]
Morphism in $\mathbb{W}$ & Corresponding cospan & Icon \\\toprule[1pt]
$\epsilon\colon \ord 1\to \ord0$&
$\ord1\to \ord1\leftarrow \ord0$&
\begin{tikzpicture}[WD]
\node[link] (epsilon) {};
\draw (epsilon) to +(-.8,0);
\end{tikzpicture}\\\hline
$\delta\colon \ord1\to \ord2$&
$\ord1\to \ord1\leftarrow \ord2$&
\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-1,0);
\draw (delta) to[out=60, in=180] +(1,.5);
\draw (delta) to[out=-60, in=180] +(1,-.5);
\end{tikzpicture}\\\hline
$\eta\colon \ord0\to \ord1$&
$\ord0\to \ord1\leftarrow \ord1$&
\begin{tikzpicture}[WD]
\node[link] (eta) {};
\draw (eta) to +(.8,0);
\end{tikzpicture}\\\hline
$\mu\colon \ord2\to \ord1$&
$\ord2\to \ord1\leftarrow \ord1$&
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(1,0);
\draw (mu) to[out=120, in=0] +(-1,.5);
\draw (mu) to[out=-120, in=0] +(-1,-.5);
\end{tikzpicture}\\\toprule[1pt]
\end{tabular}
\end{equation}
\setlength{\tabcolsep}{3pt}
\noindent As the underlying 1-category of $\mathbb{W}$ is a symmetric monoidal category, we may use these icons as primitive generators in the usual string diagram language in the sense of Joyal and Street \cite{joyal1991geometry}. These generators satisfy the following equations and inequalities involving additionally the symmetry $\ord 2\to\ord 2$ of $\mathbb{W}$, and we render these constraints graphically with composition indicated via horizontal juxtaposition and tensor indicated via vertical juxtaposition.
\begin{equation}\label{eqn.equations_wires}
\renewcommand{\arraystretch}{2}
\begin{array}{rl<{\qquad}rl<{\qquad}rl}
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(-1,0);
\draw (mu) to[out=60, in=180] +(.5,.5) coordinate (out1);
\draw (mu) to[out=-60, in=180] +(.5,-.5) coordinate (out2);
\coordinate (end) at ($(out1)+(1,0)$);
\draw (out1) to[out=0, in=180] (out2-|end);
\draw (out2) to[out=0, in=180] (out1-|end);
\end{tikzpicture}
&\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(-1,0);
\draw (mu) to[out=60, in=180] +(1,.5);
\draw (mu) to[out=-60, in=180] +(1,-.5);
\end{tikzpicture}
&
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(-1,0);
\draw (mu) to[out=60, in=180] +(.5,.5) node[link] {};
\draw (mu) to[out=-60, in=180] +(1,-.5) -- +(.5,0);
\end{tikzpicture}
&\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD,baseline=-5pt]
\draw (0,0) -- (2,0);
\end{tikzpicture}
&
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(-1,0);
\draw (mu) to[out=60, in=180] +(.5,.5) node[link] (mu2) {};
\draw (mu2) to[out=60, in=180] +(1,.5) coordinate (end);
\draw (mu2) to[out=-60, in=180] +(1,-.5);
\draw (mu) to[out=-60, in=180] +(1,-.5) coordinate (h);
\draw (h) -- (h-|end);
\end{tikzpicture}
&\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(-1,0);
\draw (mu) to[out=-60, in=180] +(.5,-.5) node[link] (mu2) {};
\draw (mu2) to[out=60, in=180] +(1,.5) coordinate (end);
\draw (mu2) to[out=-60, in=180] +(1,-.5);
\draw (mu) to[out=60, in=180] +(1,.5) coordinate (h);
\draw (h) -- (h-|end);
\end{tikzpicture}
\\
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(1,0);
\draw (mu) to[out=120, in=0] +(-.5,.5) coordinate (out1);
\draw (mu) to[out=-120, in=0] +(-.5,-.5) coordinate (out2);
\coordinate (end) at ($(out1)+(-1,0)$);
\draw (out1) to[out=180, in=0] (out2-|end);
\draw (out2) to[out=180, in=0] (out1-|end);
\end{tikzpicture}
&\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(1,0);
\draw (mu) to[out=120, in=0] +(-1,.5);
\draw (mu) to[out=-120, in=0] +(-1,-.5);
\end{tikzpicture}
&
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(1,0);
\draw (mu) to[out=120, in=0] +(-.5,.5) node[link, anchor=center] {};
\draw (mu) to[out=-120, in=0] +(-1,-.5) -- +(-.5,0);
\end{tikzpicture}
&\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD,baseline=-5pt]
\draw (0,0) -- (2,0);
\end{tikzpicture}
&
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(1,0);
\draw (mu) to[out=120, in=0] +(-.5,.5) node[link, anchor=center] (mu2) {};
\draw (mu2) to[out=120, in=0] +(-1,.5) coordinate (end);
\draw (mu2) to[out=-120, in=0] +(-1,-.5);
\draw (mu) to[out=-120, in=0] +(-1,-.5) coordinate (h);
\draw (h) -- (h-|end);
\end{tikzpicture}
&\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(1,0);
\draw (mu) to[out=-120, in=0] +(-.5,-.5) node[link, anchor=center] (mu2) {};
\draw (mu2) to[out=120, in=0] +(-1,.5) coordinate (end);
\draw (mu2) to[out=-120, in=0] +(-1,-.5);
\draw (mu) to[out=120, in=0] +(-1,.5) coordinate (h);
\draw (h) -- (h-|end);
\end{tikzpicture}
\\
\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-.75,0);
\draw (delta) to[out=60, in=180] +(.5,.5) coordinate (out1);
\draw (delta) to[out=-60, in=180] +(.5,-.5) coordinate (out2);
\node[link, right=1 of delta] (mu) {};
\draw (mu) -- +(.75,0);
\draw (mu) to[out=120, in=0] +(-.5,.5) -- (out1);
\draw (mu) to[out=-120, in=0] +(-.5,-.5) -- (out2);
\end{tikzpicture}
&\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD,baseline=-5pt]
\draw (0,0) -- (1.5,0);
\end{tikzpicture}
&
\multicolumn{4}{c}{$
\begin{tikzpicture}[WD]
\coordinate (htop);
\coordinate[below=.7 of htop] (hmid);
\coordinate[below=.7 of hmid] (hbot);
\node[link, left=.5] (dotL) at ($(htop)!.5!(hmid)$) {};
\node[link, right=.5] (dotR) at ($(hbot)!.5!(hmid)$) {};
\draw (dotL) -- +(-1,0) coordinate (l);
\draw (dotR) -- +(1,0) coordinate (r);
\draw (dotL) to[out=60, in=180] (htop);
\draw (dotL) to[out=-60, in=180] (hmid);
\draw (hmid) to[out=0, in=120] (dotR);
\draw (hbot) to[out=0, in=-120] (dotR);
\draw (htop) -- (htop-|r);
\draw (hbot) -- (hbot-|l);
\end{tikzpicture}
\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-.5,0) coordinate (end);
\draw (delta) to[out=60, in=180] +(1,.5);
\draw (delta) to[out=-60, in=180] +(1,-.5);
\node[link, left=1 of delta] (mu) {};
\draw (mu) -- (end);
\draw (mu) to[out=120, in=0] +(-1,.5);
\draw (mu) to[out=-120, in=0] +(-1,-.5);
\end{tikzpicture}
\;\raisebox{2pt}{=}\quad
\begin{tikzpicture}[WD]
\coordinate (htop);
\coordinate[below=.7 of htop] (hmid);
\coordinate[below=.7 of hmid] (hbot);
\node[link, left=.5] (dotL) at ($(hbot)!.5!(hmid)$) {};
\node[link, right=.5] (dotR) at ($(htop)!.5!(hmid)$) {};
\draw (dotL) -- +(-1,0) coordinate (l);
\draw (dotR) -- +(1,0) coordinate (r);
\draw (dotL) to[out=-60, in=180] (hbot);
\draw (dotL) to[out=60, in=180] (hmid);
\draw (hmid) to[out=0, in=-120] (dotR);
\draw (htop) to[out=0, in=120] (dotR);
\draw (htop) -- (htop-|l);
\draw (hbot) -- (hbot-|r);
\end{tikzpicture}
$}
\\
\begin{tikzpicture}[WD,baseline=-5pt]
\draw (0,0) -- (1.5,0);
\end{tikzpicture}
&\;\raisebox{2pt}{$\leq$}\quad
\begin{tikzpicture}[WD,baseline=-5pt]
\node[link] (epsilon) {};
\draw (epsilon) to +(-.8,0);
\node[link, right=.5 of epsilon] (eta) {};
\draw (eta) to +(.8,0);
\end{tikzpicture}
&
\begin{tikzpicture}[WD,baseline=-5pt]
\node[link] (epsilon) {};
\draw (epsilon) to +(-.8,0);
\node[link, left=1 of epsilon] (eta) {};
\draw (eta) to +(.8,0);
\end{tikzpicture}
&\;\raisebox{2pt}{$\leq$}\quad
\id_{\ord0}
&
\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-.5,0) coordinate (end);
\draw (delta) to[out=60, in=180] +(1,.5);
\draw (delta) to[out=-60, in=180] +(1,-.5);
\node[link, left=1 of delta] (mu) {};
\draw (mu) -- (end);
\draw (mu) to[out=120, in=0] +(-1,.5);
\draw (mu) to[out=-120, in=0] +(-1,-.5);
\end{tikzpicture}
&\;\raisebox{2pt}{$\leq$}\quad
\begin{tikzpicture}[WD]
\draw (0, 0) -- (1.5, 0);
\draw (0, .75) -- (1.5, .75);
\end{tikzpicture}
\end{array}
\end{equation}
We refer to the composites $\eta\mathbin{\fatsemi}\delta$ and $\mu\mathbin{\fatsemi}\epsilon$ as the \define{cup} and the \define{cap}; they are denoted $\const{cup}$ and $\const{cap}$ and are depicted as follows:
\begin{equation}\label{eqn.cap_cup}
\begin{array}{cccc}
\begin{tikzpicture}[inner WD]
\draw (.5, .5) -- (0,.5) to[out=180, in=180] (0, -.5) -- (.5, -.5);
\end{tikzpicture}
&\;\raisebox{2pt}{$\coloneqq$}\quad
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\node[link, left=.3 of mu] (eta) {};
\draw (mu) -- (eta);
\draw (mu) to[out=60, in=180] +(.5,.5) -- +(.25,0);
\draw (mu) to[out=-60, in=180] +(.5,-.5) -- +(.25,0);
\end{tikzpicture}
&\qquad\text{and}\qquad
\begin{tikzpicture}[inner WD]
\draw (-.5, .5) -- (0,.5) to[out=0, in=0] (0, -.5) -- (-.5, -.5);
\end{tikzpicture}
&\;\raisebox{2pt}{$\coloneqq$}\quad
\begin{tikzpicture}[WD]
\node[link] (mu) {};
\node[link, right=.3 of mu] (eta) {};
\draw (mu) -- (eta);
\draw (mu) to[out=120, in=0] +(-.5,.5) -- +(-.25,0);
\draw (mu) to[out=-120, in=0] +(-.5,-.5) -- +(-.25,0);
\end{tikzpicture}
\end{array}
\end{equation}
It follows from \eqref{eqn.equations_wires} that $\const{cap}$ and $\const{cup}$ satisfy the ``yanking'' or adjunction identities
\begin{equation}\label{eqn.yanking}
\begin{tikzpicture}[inner WD]
\draw (1, 1) --
(-.25, 1) to[out=180, in=180]
(-.25, 0) --
(.25, 0) to[out=0, in=0]
(.25, -1) --
(-1, -1);
\node[font=\normalsize] at (3,0) {=};
\draw (5,0) -- (7,0);
\node[font=\normalsize] at (9,0) {=};
\draw (13, -1) --
(11.75, -1) to[out=180, in=180]
(11.75, 0) --
(12.25, 0) to[out=0, in=0]
(12.25, 1) --
(11, 1);
\end{tikzpicture}
\end{equation}
\renewcommand{\arraystretch}{1}
The equations in the first and second lines of \eqref{eqn.equations_wires} are known as the \define{(co)commutativity, (co)unitality, and (co)associativity} equations for comonoids and monoids, respectively. The equations in the next line are known as the \define{special law} and the \define{frobenius law}. We refer to the inequalities in the last line as the \define{adjunction inequalities}, because they show up as the unit and co-unit of adjunctions, as we see next in the following proposition.
\begin{proposition}\label{prop.left_adj_in_ww}
With notation as in \eqref{eqn.generating_wires}, there are adjunctions
\begin{equation}\label{eqn.ww_adjunctions}
\adj{\ord1}{\epsilon}{\eta}{\ord0}
\qquad\text{and}\qquad
\adj{\ord1}{\delta}{\mu}{\ord2}.
\end{equation}
\end{proposition}
\begin{proof}
The inequalities $\id_{\ord1}\leq(\epsilon\mathbin{\fatsemi}\eta)$,\quad $(\eta\mathbin{\fatsemi}\epsilon)\leq\id_{\ord1}$,\quad $\id_{\ord2}\leq(\mu\mathbin{\fatsemi}\delta)$, and the equation $\id_{\ord1}=(\delta\mathbin{\fatsemi}\mu)$ are all shown in \eqref{eqn.equations_wires}, which itself is proved via computations in $\poCat{Cospan}^{\tn{co}}$. The required equalities \eqref{eqn.adjunction} are automatic in a po-category.
\end{proof}
\begin{remark}\label{rem.surprising_inequality}
The perhaps surprising half of the ``special law'', i.e.\ the inequality $(\delta\mathbin{\fatsemi}\mu)\leq\id_{\ord1}$ not arising from adjointness, is in fact derivable from the rest of the structure:
\[
\begin{tikzpicture}
\node (P1) {
\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-.75,0);
\draw (delta) to[out=60, in=180] +(.5,.5) coordinate (out1);
\draw (delta) to[out=-60, in=180] +(.5,-.5) coordinate (out2);
\node[link, right=1 of delta] (mu) {};
\draw (mu) -- +(.75,0);
\draw (mu) to[out=120, in=0] +(-.5,.5) -- (out1);
\draw (mu) to[out=-120, in=0] +(-.5,-.5) -- (out2);
\end{tikzpicture}
};
\node (P2) [right=1 of P1] {
\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-.75,0);
\draw (delta) to[out=60, in=180] +(.5,.5) coordinate (out1);
\draw (delta) to[out=-60, in=180] +(.5,-.5) coordinate (out2);
\node[link, right=1 of delta] (mu) {};
\draw (mu) to[out=120, in=0] +(-.5,.5) -- (out1);
\draw (mu) to[out=-120, in=0] +(-.5,-.5) -- (out2);
\draw (mu) -- +(.75,0) node[link] (delta2) {};
\draw (delta2) to[out=60, in=180] +(.5,.5) -- +(.5,0);
\draw (delta2) to[out=-60, in=180] +(.5,-.5) node[link] {};
\end{tikzpicture}
};
\node (P3) [right=1 of P2] {
\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-.75,0);
\draw (delta) to[out=60, in=180] +(.5,.5) -- +(.5,0);
\draw (delta) to[out=-60, in=180] +(.5,-.5) node[link] {};
\end{tikzpicture}
};
\node (P4) [right=1 of P3] {
\begin{tikzpicture}[WD]
\draw (0,0) -- (1.5,0);
\end{tikzpicture}
};
\node at ($(P1.east)!.5!(P2.west)$) {$=$};
\node at ($(P2.east)!.5!(P3.west)$) {$\leq$};
\node at ($(P3.east)!.5!(P4.west)$) {$=$};
\end{tikzpicture}
\]
More generally, if $f\colon \ord{m}\twoheadrightarrow \ord{n}$ is any surjective function then $f^{\dagger}\mathbin{\fatsemi} f=\id$, where $f^{\dagger}$ is the transpose (left adjoint) of $f$.
\end{remark}
\begin{definition}[Sub-props of $\mathbb{W}$]\label{def.comon_mon_selfduals}
The prop $\mathbb{W}$ contains several other important full sub-props:
\begin{itemize}
\item That generated by $\epsilon$ and $\delta$ is called the \define{prop for cocommutative comonoids}.
\item That generated by $\eta$ and $\mu$ is called the \define{prop for commutative monoids}.
\item That generated by $\const{cup}$ and $\const{cap}$ \eqref{eqn.cap_cup} is called the \define{prop for self-duals}.
\end{itemize}
In fact these three props are equivalent to the monoidal category of finite sets, its opposite, and the category of unoriented cobordisms, respectively. See also \cite[Section 3]{fong2019supplying}.
\end{definition}
\begin{notation}[General arity $\delta$ and $\mu$]
We will adopt the convention that for $n\in\mathbb{N}$ we will write $\delta^{n}\colon\ord 1\to\ord n$ and $\mu^{n}\colon \ord n\to\ord 1$ for the maps associated to the cospans $\ord 1\to\ord1\leftarrow\ord n$ and $\ord n\to\ord1\leftarrow\ord 1$ respectively. In particular we have $\delta^{0}=\epsilon$, $\mu^{0}=\eta$, and $\delta^{1}=\mu^{1}=\id_{\ord 1}$ among other identities. Graphically we might render these morphisms as
\begin{diagram*}
\node(D){\begin{tikzpicture}[WD]
\node[link] (delta) {};
\draw (delta) -- +(-1,0);
\draw (delta) to[out=60, in=180] +(1,1.3);
\draw (delta) to[out=-60, in=180] +(1,-1.3);
\draw (delta) to[out=60, in=180] +(1,0.8) node(t){};
\draw (delta) to[out=-60, in=180] +(1,-0.8) node(b){};
\path (t) -- node[right]{$m$} (b);
\path (t) -- node[sloped]{$\ldots$} (b);
\end{tikzpicture}};
\node(M)[right=2cm of D]{\begin{tikzpicture}[WD]
\node[link] (mu) {};
\draw (mu) -- +(+1,0);
\draw (mu) to[out=120, in=0] +(-1,1.3);
\draw (mu) to[out=-120, in=0] +(-1,-1.3);
\draw (mu) to[out=120, in=0] +(-1,0.8) node(t2)[anchor=center]{};
\draw (mu) to[out=-120, in=0] +(-1,-0.8) node(b2)[anchor=center]{};
\path (t2) -- node[left]{$n$} (b2);
\path (t2) -- node[sloped,anchor=center]{$\ldots$} (b2);
\end{tikzpicture}};
\path (D) -- node[midway]{and} (M);
\end{diagram*}
for $\delta^{m}$ and $\mu^{n}$ respectively.
\end{notation}
\section{Picturing morphisms in \texorpdfstring{$\mathbb{W}$}{W}}\label{sec.depict_ww}
The po-category $\mathbb{W}$ forms the foundation of our diagrammatic language for regular logic. We have already seen that morphisms in $\mathbb{W}$ can be given a graphical description by depicting generating morphisms using special icons, and working in the usual Joyal--Street string diagram language for morphisms in symmetric monoidal categories.
We now present an alternate way to depict morphisms in $\mathbb{W}$. While the description we provide here is presently informal, later we shall describe a rigorous manner in which to draw, interpret, and manipulate these pictures.
\begin{notation}[Objects as IO shells]
By definition, an object $\ord n\in\mathbb{W}$ is a finite set and we represent it graphically by a circle with $n$ ports around the exterior.
\begin{equation}\label{eqn.shell_pic}
\ord{n}=
\begin{tikzpicture}[inner WD, baseline=(rho)]
\node[pack,minimum size = 3ex] (rho) {};
\draw (rho.180) to +(180:2pt);
\draw (rho.110) to +(110:2pt);
\node at ($(rho.0)+(60:4pt)$) {$\vdots$};
\node at ($(rho.0)+(0:7pt)$) {\big\}};
\node at ($(rho.0)+(0:29pt)$) {$n-3$ ports};
\draw (rho.-110) to +(-110:2pt);
\end{tikzpicture}
\end{equation}
Our convention will be for the ports to be numbered clockwise from the left of the circle, unless otherwise indicated. We refer to such an annotated circle as an \define{input-output shell}, or \define{IO shell}.
\end{notation}
\begin{notation}[Picturing morphisms of $\mathbb{W}$] \label{notation.picturing_ww}
Given a morphism $\omega\colon \ord n_1+ \dots + \ord n_s \to \ord n_\mathrm{out}$ in $\mathbb{W}$, we associate to it the canonical representation as a cospan of finite sets assured by \cref{rem.canonical_decomp}.
\begin{equation}\label{eqn.cospan}
\ord{n}_1+\cdots+\ord{n}_s
\To{\copair{\omega_1,\ldots,\omega_s}}
\ord{n}_\omega
\From{\omega_\mathrm{out}}
\ord{n}_\mathrm{out}
\end{equation}
From this we draw our depiction of $\omega$ as follows.
\begin{enumerate}[nolistsep, noitemsep]
\item Draw the IO shell for $\ord n_\mathrm{out}$.
\item Draw each object $\ord n_i$, for $i=1,\dots,s$, as non-overlapping IO shells inside the $\ord n_\mathrm{out}$ IO shell.
\item For each $i \in \ord{n}_\omega$, draw a black dot anywhere in the region interior to the $\ord n_\mathrm{out}$ IO shell but exterior to all the $\ord n_i$ IO shells.
\item For each element $(i,j) \in \sum_{i=1,\dots,s,\mathrm{out}} \ord{n}_i$, draw a wire connecting the $j$th port on the object $\ord n_i$ to the black dot $\omega_i(j)$.
\end{enumerate}
For a more compact notation, we may also neglect to explicitly draw the object $\ord n_\mathrm{out}$, leaving it implicit as comprising the wires left dangling on the boundary of the diagram.
\end{notation}
\begin{remark}
While we take the convention of \cref{notation.picturing_ww} for drawing morphisms of $\mathbb{W}$, note that once we determine the domain and codomain of the morphism, the rest of the data is combinatorial. This implies that, up to a choice of domain and codomain, the morphism represented by any picture is invariant under topological deformation.
\end{remark}
\begin{example} \label{ex.wiring_diagram}
Here is the combinatorial data of a morphism $\omega\colon \ord n_1+ \ord n_2 + \ord n_3 \to \ord n_\mathrm{out}$ in $\mathbb{W}$, together with its depiction:
\begin{align*}
\parbox{3.5in}{\raggedright
$n_1 = 3$,\quad $n_2 = 3$,\quad $n_3 = 4$,\quad $n_\mathrm{out} = 6$, \quad $n_\omega = 7$\\
$\omega_1(1)=4$, $\omega_1(2)=2$, $\omega_1(3)=1$,\\
$\omega_2(1)=6$, $\omega_2(2)=4$, $\omega_2(3)=5$,\\
$\omega_3(1)=1$, $\omega_3(2)=2$, $\omega_3(3)=\omega_3(4)=6$,\\
$\omega_\mathrm{out}(1)=1$, $\omega_\mathrm{out}(2)=\omega_\mathrm{out}(3)=3$\\
$\omega_\mathrm{out}(4)=5$, $\omega_\mathrm{out}(5)=6$, $\omega_\mathrm{out}(6)=7$.}
&\qquad
\begin{tikzpicture}[penetration=0, inner WD, pack size=9pt, link size=2pt, font=\tiny, scale=2, baseline=(out)]
\node[packs] at (-1.5,-1) (f) {$3$};
\node[packs] at (0,1.9) (g) {$1$};
\node[packs] at (1.5,-1) (h) {$2$};
\node[outer pack, inner sep=34pt] at (0,.2) (out) {};
\node[link,label=90:$6$] at ($(f)!.5!(h)$) (link1) {};
\node[link,label=0:$7$] at (-2.4,-.25) (link2) {};
\node[link,label={[label distance=0pt,xshift=-2pt]90:$1$}] at ($(f.75)!.5!(g.-135)$) (link3) {};
\begin{scope}[label distance=-6pt]
\draw[fr_out] (out.270) to (link1);
\draw[fr_out] (out.190) to (link2);
\draw[fr_out] (out.155) to (link3);
\draw[fr_out] (out.-35) to node[pos=.5,link,label=10:$5$]{} (h.-30);
\draw[to_out,fr_out] (out.15) to[out=-165,in=-110] node[pos=.5,link,label
distance=3pt,label=200:$3$]{} (out.70);
\draw (f.30) to[out=0,in=145] (link1);
\draw (f.-30) to[out=0,in=-145] (link1);
\draw (h.180) to (link1);
\draw (g.-60) to node[pos=.5,link,label={[xshift=2pt]10:$4$}]{} (h.120);
\draw (f.45) to node[pos=.5,link,label=-10:$2$]{} (g.-105);
\draw (f.75) to (link3);
\draw (g.-135) to (link3);
\end{scope}
\end{tikzpicture}
\end{align*}
\end{example}
\begin{example}\label{ex.no_inner}
Note that we may have $s=0$, in which case there are no inner IO shells.
For example, the following has $n_\omega=2$, and $n_{\mathrm{out}}=4$.
\[
\begin{tikzpicture}[inner WD]
\node[link] (dot270) {};
\node[outer pack, fit=(dot270), inner sep=12pt] (outer) {};
\node[link] at ($(outer.0) - (0:5pt)$) (dot0) {};
\draw[fr_out] (outer.0) -- (dot0);
\draw[fr_out] (outer.70) -- (dot270);
\draw[fr_out] (outer.180) -- (dot270);
\draw[fr_out] (outer.300) -- (dot270);
\pgfresetboundingbox
\useasboundingbox (outer.270) rectangle (outer.north);
\end{tikzpicture}
\]
In addition we may have $\ord n_{\mathrm{out}}=\ord 0$ in which case we render the two possible morphisms $\id_{\ord 0}$ and $\eta\mathbin{\fatsemi}\epsilon$ in $\mathbb{W}(\ord0,\ord0)$ as the empty picture and a free floating dot, respectively.
\end{example}
\begin{remark}\label{rem.multiple}
When multiple wires meet at a point, our convention will be to draw a dot iff the number of wires is different from two.
\begin{center}
\begin{tikzpicture}[unoriented WD, font=\small]
\node["1 wire"] (P1) {
\begin{tikzpicture}[inner WD, surround sep=4pt]
\node[link] (dot) {};
\node[outer pack, fit=(dot)] (outer) {};
\draw[to_out] (dot) -- (outer.west);
\end{tikzpicture}
};
\node[right=4 of P1, "2 wires"] (P2) {
\begin{tikzpicture}[inner WD, surround sep=4pt]
\node[link, white] (dot) {};
\node[outer pack, fit=(dot)] (outer) {};
\draw[fr_out,to_out] (outer.west) -- (outer.east);
\end{tikzpicture}
};
\node[right=4 of P2, "3 wires"] (P3) {
\begin{tikzpicture}[inner WD, surround sep=4pt]
\node[link] (dot) {};
\node[outer pack, fit=(dot)] (outer) {};
\draw[to_out] (dot) -- (outer.west);
\draw[to_out] (dot) -- (outer.east);
\draw[to_out] (dot) -- (outer.south);
\end{tikzpicture}
};
\node[right=4 of P3, "4 wires"] (P4) {
\begin{tikzpicture}[inner WD, surround sep=4pt, font=\tiny]
\node[link] (dot) {};
\node[outer pack, fit=(dot)] (outer) {};
\draw[to_out] (dot) -- (outer.west);
\draw[to_out] (dot) -- (outer.east);
\draw[to_out] (dot) -- (outer.south);
\draw[to_out] (dot) -- (outer.north);
\end{tikzpicture}
};
\node[right=5 of P4, "$\cdots$\quad etc."] (etc){};
\pgfresetboundingbox
\useasboundingbox ($(P1.north west)+(0,10pt)$) rectangle ($(etc)+(0,-5pt)$);
\end{tikzpicture}
\end{center}
When wires intersect and we do not draw a black dot, the intended interpretation is that the wires are \emph{not connected}:
\begin{tikzpicture}[baseline=(P1.-20)]
\node (P1) {
\begin{tikzpicture}[inner WD, surround sep=2pt, font=\tiny]
\node[link] (dot) {};
\node[outer pack, fit=(dot)] (outer) {};
\draw[to_out] (dot) -- (outer.west);
\draw[to_out] (dot) -- (outer.east);
\draw[to_out] (dot) -- (outer.south);
\draw[to_out] (dot) -- (outer.north);
\end{tikzpicture}
};
\node[right=0.25 of P1] (P2) {
\begin{tikzpicture}[inner WD, surround sep=2pt, font=\tiny]
\node[link, white] (dot) {};
\node[outer pack, fit=(dot)] (outer) {};
\draw[fr_out,to_out] (outer.east) -- (outer.west);
\draw[fr_out,to_out] (outer.south) -- (outer.north);
\end{tikzpicture}
};
\path (P1) -- node[midway](N){$\neq$} (P2);
\end{tikzpicture}.
\end{remark}
The following examples give a flavor of how composition, monoidal product, and
$2$-morphisms are represented using this graphical notation.
\begin{example}[Composition as substitution]\label{ex.comp_as_subst}
Composition of morphisms is described by \define{nesting} of their depictions. Let $\omega'\colon \ord n'\to \ord n_1$ and $\omega\colon \ord n_1 \to \ord n_\mathrm{out}$ be morphisms in $\mathbb{W}$. Then the composite morphism $\omega'\mathbin{\fatsemi} \omega\colon \ord n' \to \ord n_\mathrm{out}$ is depicted by
\begin{enumerate}[nolistsep, noitemsep]
\item drawing $\omega'$ inside the inner circle of the picture for $\omega$,
\item erasing the IO shell representing $\ord n_1$,
\item amalgamating any connected black dots into a single black dot,
\item removing either
\begin{enumerate}[nolistsep,noitemsep, label=(\roman*)]
\item all but one of the black dots not connected to a IO shell (if $n' = n_\mathrm{out} = 0$) or
\item all black dots not connected to a IO shell (if $n' \ne 0$ or $n_\mathrm{out}\ne 0$).
\end{enumerate}
\end{enumerate}
Note that step 3 corresponds to taking pushouts in $\Cat{FinSet}$, while step 4 corresponds to taking the poset reflection.
As a shorthand for composition, we simply draw one morphism directly substituted into another, as per step 1. For example, we have
\begin{center}
\begin{tikzpicture}[unoriented WD, font=\small]
\node["$\omega'$"] (P1a) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs] (a) {};
\node[outer pack, inner sep=10pt, fit=(a)] (outer) {};
\node[link] (link1) at ($(a.west)!.6!(outer.west)$) {};
\node[link] (link3) at ($(a.-20)!.5!(outer.-20)$) {};
\draw[fr_out] (outer.west) -- (link1);
\draw[to_out] (a.40) -- (outer.45);
\draw (a.-20) -- (link3);
\draw[to_out] (link3) -- (outer.0);
\draw[to_out] (link3) -- (outer.-45);
\end{tikzpicture}
};
\node[right=1 of P1a, "$\omega$"] (P1b) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs] (c) {};
\node[outer pack, inner sep=10pt, fit=(c)] (outer2) {};
\node[link] (link4) at ($(c.west)!.4!(outer2.west)$) {};
\node[link] (link5) at ($(c.20)!.5!(outer2.20)$) {};
\draw (c.west) -- (link4);
\draw (c.45) -- (link5);
\draw (c.0) -- (link5);
\draw[to_out] (link5) -- (outer2.20);
\draw[to_out] (c.-45) -- (outer2.-45);
\end{tikzpicture}
};
\node[right=3 of P1b] (P2) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs] (a) {};
\node[outer pack, inner sep=7pt, fit=(a)] (c) {};
\node[outer pack, inner sep=5pt, fit=(c)] (outer2) {};
\node[link] (link1) at ($(a.west)!.6!(c.west)$) {};
\node[link] (link3) at ($(a.-20)!.5!(c.-20)$) {};
\draw (c.west) -- (link1);
\draw (a.40) -- (c.45);
\draw (a.-20) -- (link3);
\draw (link3) -- (c.0);
\draw (link3) -- (c.-45);
\node[link] (link4) at ($(c.west)!.4!(outer2.west)$) {};
\node[link] (link5) at ($(c.20)!.5!(outer2.20)$) {};
\draw (c.west) -- (link4);
\draw (c.45) -- (link5);
\draw (c.0) -- (link5);
\draw[to_out] (link5) -- (outer2.20);
\draw[to_out] (c.-45) -- (outer2.-45);
\end{tikzpicture}
};
\node[right=3 of P2] (P3) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs] (c) {};
\node[outer pack, inner sep=10pt, fit=(c)] (outer2) {};
\node[link] at ($(c.180)!.5!(outer2.180)$) {};
\node[link] (link) at ($(c.0)!.5!(outer2.0)$) {};
\draw (c.25) -- (link);
\draw (c.-25) -- (link);
\draw[to_out] (link) -- (outer2.15);
\draw[to_out] (link) -- (outer2.-15);
\end{tikzpicture}
};
\node[right=3 of P3, "$\omega'\mathbin{\fatsemi}\omega$"] (P4) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs] (c) {};
\node[outer pack, inner sep=10pt, fit=(c)] (outer2) {};
\node[link] (link) at ($(c.0)!.5!(outer2.0)$) {};
\draw (c.25) -- (link);
\draw (c.-25) -- (link);
\draw[to_out] (link) -- (outer2.15);
\draw[to_out] (link) -- (outer2.-15);
\end{tikzpicture}
};
\node (P1) at ($(P1a.east)!.5!(P1b.west)$) {$\mathbin{\fatsemi}$};
\node (E1) at ($(P1b.east)!.5!(P2.west)$) {$=$};
\node[above=0 of E1] {\tiny 1};
\node (E2) at ($(P2.east)!.5!(P3.west)$) {$=$};
\node[above=0 of E2] {\tiny 2, 3};
\node (E3) at ($(P3.east)!.5!(P4.west)$) {$=$};
\node[above=0 of E3] {\tiny 4};
\end{tikzpicture}
\end{center}
For the more general $n$-ary or operadic case, we may obtain the composite
\[
(\ord n_1 + \dots + \ord n_{i-1} + \omega' +
\ord n_{i+1} + \dots + \ord n_j) \mathbin{\fatsemi} \omega
\]
of any two morphisms $\omega'\colon \ord n'_1+ \dots + \ord n'_{s'} \to \ord n_i$ and $\omega\colon \ord n_1+ \dots + \ord n_s \to \ord n_\mathrm{out}$, with $1\leq i\leq s$, by substituting the picture for $\omega'$ into the $i$\textsuperscript{th} inner circle of the picture for $\omega$, and following a procedure similar to that in \cref{ex.comp_as_subst}.
\end{example}
\begin{example}[Monoidal product as juxtaposition]
Recall \cref{cor.mor_tensors_ww}: $\id_{\ord 0}$ acts as an identity under tensoring of morphisms. In our graphical notation the monoidal product of two morphisms in $\mathbb{W}$ is simply their juxtaposition if neither morphism is $\eta\mathbin{\fatsemi}\epsilon$. For example, we might have:
\begin{center}
\begin{tikzpicture}[unoriented WD, font=\small]
\node (P1a) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs] (a) {};
\node[packs, below right=.1 and 2 of a] (b) {};
\node[outer pack, fit=(a) (b)] (outer) {};
\draw (a.180) -- (a.180-|outer.west);
\draw (a.20) to[out=0, in=120] (b.140);
\draw (a.-30) to[out=-40, in=180] (b.180);
\draw (b.east) -- (b.east-|outer.east);
\end{tikzpicture}
};
\node[below=1 of P1a] (P1b) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs, below=2 of $(a)!.5!(b)$] (c) {};
\node[link, right=.8 of c] (link1) {};
\node[outer pack, fit=(link1) (c)] (outer) {};
\draw (c.20) -- (link1);
\draw (c.-20) -- (link1);
\draw[to_out] (link1) -- (link1-|outer.east);
\end{tikzpicture}
};
\node (P1) at ($(P1a.south)!.5!(P1b.north)$) {$+$};
\node[right=5 of $(P1a.north)!.5!(P1b.south)$] (e) {$=$};
\node[right=1 of e] (P2) {
\begin{tikzpicture}[inner WD, pack size=8pt]
\node[packs] (a) {};
\node[packs, below right=.1 and 2 of a] (b) {};
\node[packs, below=2 of $(a)!.5!(b)$] (c) {};
\node[link, right=.8 of c] (link1) {};
\node[outer pack, fit=(a) (b) (c)] (outer) {};
\draw (a.180) -- (a.180-|outer.west);
\draw (a.20) to[out=0, in=120] (b.140);
\draw (a.-30) to[out=-40, in=180] (b.180);
\draw[to_out] (b.east) -- (b.east-|outer.east);
\draw (c.20) -- (link1);
\draw (c.-20) -- (link1);
\draw (link1) -- (link1-|outer.east);
\end{tikzpicture}
};
\pgfresetboundingbox
\useasboundingbox (P1b.190-|P1a.west) rectangle (P2.east|-P1a.150);
\end{tikzpicture}
\end{center}
By \cref{cor.mor_tensors_ww}, $\omega+\eta\mathbin{\fatsemi}\epsilon=\omega$ if and only if $\omega\ne\id_{\ord 0}$, and so we render the monoidal product $\omega+\eta\mathbin{\fatsemi}\epsilon$ as $\omega$, unless $\omega=\id_{\ord 0}$, in which case it is once more $\eta\mathbin{\fatsemi}\epsilon$.
\end{example}
\begin{example}[$2$-morphisms as breaking wires and removing disconnected black dots]\label{lemma.breaking}
Let $\omega,\omega'\colon \ord n \to \ord n_\mathrm{out}$ be morphisms in $\mathbb{W}$. Each is canonically represented by a cospan of finite sets, \[ \ord n\to \ord n_{\omega}\leftarrow \ord n_\mathrm{out} \qquad\text{and}\qquad \ord n\to \ord n_{\omega'}\leftarrow \ord n_\mathrm{out}\ .\] By definition, there exists a $2$-morphism $\omega\leq\omega'$ iff there is a function $x\colon \ord n_{\omega'}\to \ord n_\omega$ making the requisite diagrams commute. For any element $i\in\ord n_\omega$, the pre-image $x^{*}(i)$ is either empty, has one element, or has multiple elements. In the first case, the pair of pictures depicting each side of the inequality $\omega\leq\omega'$ would show dot $i$ being removed; in the second case, it would show dot $i$ remaining as it was; and in the third case, it would show a connection being broken at dot $i$. For example, we have $2$-morphisms
\begin{tikzpicture}[inner WD,baseline=(P1.-20)]
\node (P1) {
\begin{tikzpicture}[inner WD,baseline=(current bounding box.south)]
\node[link] (dot) {};
\node[outer pack, surround sep=3pt, fit=(dot)] (outer) {};
\end{tikzpicture}
};
\node[right=2 of P1] (P2) {
\begin{tikzpicture}[inner WD,baseline=(current bounding box.south)]
\node[link, white] (dot) {};
\node[outer pack, surround sep=3pt, fit=(dot)] (outer) {};
\end{tikzpicture}
};
\node at ($(P1.east)!.5!(P2.west)$) {$\leq$};
\end{tikzpicture}
and
\begin{tikzpicture}[inner WD,baseline=(P3.-20)]
\node (P3) {
\begin{tikzpicture}[inner WD,baseline=(current bounding box.south)]
\node[link] (dot) {};
\node[outer pack, surround sep=3pt, fit=(dot)] (outer) {};
\draw[to_out] (dot) -- (outer.0);
\draw[to_out] (dot) -- (outer.120);
\draw[to_out] (dot) -- (outer.240);
\end{tikzpicture}
};
\node[right=2 of P3] (P4) {
\begin{tikzpicture}[inner WD,baseline=(current bounding box.south)]
\node[link, draw=none,fill=none] (fake dot) {};
\node[outer pack, surround sep=3pt, fit=(fake dot)] (outer) {};
\node[link] (dot) at ($(outer.0)+(0:-5pt)$) {};
\draw[to_out] (dot) -- (outer.0);
\draw[to_out,fr_out] (outer.120) to[out=300, in=60] (outer.240);
\end{tikzpicture}
};
\node at ($(P3.east)!.5!(P4.west)$) {$\leq$};
\end{tikzpicture}.
\end{example}
\section{Supply}\label{sec.supply}
It often happens that every object in a symmetric monoidal category $\cat{C}$ is equipped with the same sort of algebraic structure -- say coming from a prop $\mathbb{P}$ -- with the property that these algebraic structures are compatible with the monoidal structure. In \cite{fong2019supplying}, we refer to this situation by saying that $\cat{C}$ \emph{supplies} $\mathbb{P}$. For our purposes we need to slightly generalise this theory, from props to po-props and from symmetric monoidal categories $\cat{C}$ to symmetric monoidal po-categories $\ccat{C}$.
\begin{definition}[Supply]\label{def.supply}
Let $\mathbb{P}$ be a po-prop and $\ccat{C}$ a symmetric monoidal po-category. A \define{supply of $\mathbb{P}$ in $\ccat{C}$} consists of a strong monoidal po-functor $s_c\colon\mathbb{P}\to\ccat{C}$ for each object $c\in\ccat{C}$, such that
\begin{enumerate}[label=(\roman*)]
\item $s_c(m)=c\tpow{m}$ for each $m\in\mathbb{N}$,
\item the strongator $c\tpow{m}\otimes c\tpow{n}\to c\tpow{(m+n)}$ is equal to the associator for each $m,n\in\mathbb{N}$,
\item the following diagrams commute for every $c,d\in\ccat{C}$ and $\mu\colon m\to n$ in $\mathbb{P}$, where the $\sigma$'s are the symmetry isomorphisms from \eqref{eqn.symmetry}.
\begin{diagram}[][][baseline=(s.base)]\label{eqn.supply_commute_tensors}
\node(1)[]{$c\tpow m\otimes d\tpow m$};
\node(2)[right=2cm of 1.east,anchor=west]{$c\tpow n\otimes d\tpow n$};
\node(3)[below= of 1]{$(c\otimes d)\tpow m$};
\node(4)[below= of 2]{$(c\otimes d)\tpow n$};
\draw[a](1)to node[la,above]{$s_{c}(\mu)\otimes s_{d}(\mu)$}(2);
\draw[a](3)to node[la,below]{$s_{c\otimes d}(\mu)$}(4);
\draw[a](1)to node(s)[la,left]{$\sigma$}(3);
\draw[a](2)to node[la,right]{$\sigma$}(4);
\node(5)[right=2cm of 2.east]{$I$};
\node(6)[right= of 5]{$I$};
\node(7)[below= of 5]{$I\tpow m$};
\node(8)[below= of 6]{$I\tpow n$};
\draw[a](5)to node[la,left]{$\sigma$}(7);
\draw[a](6)to node[la,right]{$\sigma$}(8);
\draw[a](7)to node[la,below]{$s_{I}(\mu)$}(8);
\draw[d](5)to(6);
\end{diagram}
\end{enumerate}
We often denote the morphism $s_c(\mu)$ in $\ccat{C}$ simply by $\mu_c\colon c\tpow{m}\to c\tpow{n}$ for typographical reasons; i.e.\ we elide explicit mention of $s$.
We further say that $f\colon c\to d$ in $\ccat{C}$ is a \define{lax $s$-homomorphism} (resp.\ \define{oplax $s$-homomorphism}) if, for each $\mu\colon m\to n$ in the prop $\mathbb{P}$, there is a $2$-morphism as shown in the left-hand (resp.\ right-hand) diagram:
\begin{diagram*}[.][4][node distance=1.5cm]
\node(1)[] {$c\tpow{m}$};
\node(2)[right of= 1]{$d\tpow{m}$};
\node(3)[below of= 1]{$c\tpow{n}$};
\node(4)[below of= 2]{$d\tpow{n}$};
\draw[a](1)to node[la,above]{$f\tpow m$}(2);
\draw[a](1)to node[la,left]{$\mu_{c}$}(3);
\draw[a](2)to node[la,right]{$\mu_{d}$}(4);
\draw[a](3)to node[la,below]{$f\tpow n$}(4);
\path (2) to node[la,sloped]{$\geq$}(3);
\node(1)[right=3cm of 2]{$c\tpow{m}$};
\node(2)[right of= 1] {$d\tpow{m}$};
\node(3)[below of= 1] {$c\tpow{n}$};
\node(4)[below of= 2] {$d\tpow{n}$};
\draw[a](1)to node[la,above]{$f\tpow{m}$}(2);
\draw[a](1)to node[la,left]{$\mu_{c}$}(3);
\draw[a](2)to node[la,right]{$\mu_{d}$}(4);
\draw[a](3)to node[la,below]{$f\tpow{n}$}(4);
\path (3) to node[la,sloped]{$\leq$}(2);
\end{diagram*}
Since $\ccat{C}$ is locally posetal, if $f$ is both a lax and an oplax $s$-homomorphism, then these diagrams commute and we simply say $f$ is an \define{$s$-homomorphism}.
We say that $\ccat{C}$ \define{(lax-/oplax-) homomorphically supplies $\mathbb{P}$} if every morphism $f$ in $\ccat{C}$ is a (lax/oplax) $s$-homomorphism.
\end{definition}
\begin{example}
An important class of examples of homomorphic supply are those categories with finite products. By the main theorem of \cite{fox1976coalgebras}, replicated below, such categories are precisely the discretely ordered po-categories that homomorphically supply cocommutative comonoids (\cref{def.comon_mon_selfduals}).
\end{example}
\begin{proposition}\label{prop.fox}
A category $\cat{C}$ has finite products iff it can be equipped with a homomorphic supply of commutative comonoids. If $\cat{C}$ and $\cat{D}$ have finite products, a functor $\cat{C}\to\cat{D}$ preserves them iff it preserves the supply of comonoids.
\end{proposition}
\begin{example}
We shall meet another large class of examples of lax homomorphic supply in \cref{chap.relational_pocats}, wherein we shall find that regular categories are equivalently po-categories with a lax homomorphic supply of $\mathbb{W}$ and some additional structure.
\end{example}
Now that we have established the definition of supply we collect some results which we will variously leverage in our later sections.
\begin{notation}[Coproduct of symmetric monoidal po-categories]
We will write $\bigsqcup$ to denote the coproduct of symmetric monoidal po-categories in the $2$-category \textsf{SMC} of symmetric monoidal po-categories, symmetric monoidal po-functors, and monoidal natural transformations.
\end{notation}
\begin{warning}\label{warn.coprod}
The coproduct of symmetric monoidal categories in \textsf{SMC} \emph{does not} coincide with the po-categorical coproduct in $\CCat{PoCat}$. Instead, for a set $J$ and symmetric monoidal po-categories $\{\ccat{C}_{j}\}_{j\in J}$,
\[
\ob\Big(\bigsqcup_{j\in J}\ccat{C}_{j}\Big)\coloneqq\Big\{ (c_{j})\in\prod_{j\in J}\ob\ccat{C}_{j}\ \Big|\ c_{j}=I_{j} \text{ for all but finitely many } j\in J\Big\}\ .
\]
See \cite[Theorem 2.2 \& Appendix A]{fong2019supplying} for details.
\end{warning}
\begin{lemma}\label{lemma.irritating_supply_coprod}
Let $\ccat{C}$ be a symmetric monoidal po-category supplying $\mathbb{P}$ a po-prop. Then for any set $J$, the supply $s$ of $\mathbb{P}$ in $\ccat{R}$ extends to a supply $\tilde s$ of $\ccat{C}$ in $\bigsqcup_{J}\ccat{C}$ such that \[\tilde s_{(c_{j})_{J}}\mathbin{\fatsemi}\pi_{i} =
\begin{cases}
I,& c_{i}=I\\
s_{c_{i}},& \mathrm{otherwise}
\end{cases}\quad ,\]
as functors $\mathbb{P}\to\ccat{C}$. If in particular $\ccat{C}$ is symmetric strict monoidal then $\tilde s$ satisfies $\tilde s_{(c_{j})_{j}}\mathbin{\fatsemi}\pi_{i}=s_{c_{i}}$.
\end{lemma}
\begin{proof}
The supply conditions (\cref{def.supply}) hold ``point-wise'' for each $c_{i}$, and as the symmetries of $\bigsqcup_{J}\ccat{C}$ are point-wise those of $\ccat{C}$, the functors $\tilde s_{(c_{i})_{I}}$ constitute a supply of $\mathbb{P}$ in $\bigsqcup\ccat{C}$.
\end{proof}
In what follows, we do not produce proofs for \cref{prop.supply_coprod,prop.p_supplies_itself,cor.change_of_supply} as these were essentially proven in \cite[Propositions 3.13, 3.14, and 3.21]{fong2019supplying}; the change from props to po-props makes no difference in this context.
\begin{proposition}\label{prop.supply_coprod}
A supply $s$ of $\mathbb{P}$ in $\ccat{C}$ induces a strong monoidal po-functor \[s^{\sqcup}\colon\bigsqcup\limits_{\ob\ccat{C}}\mathbb{P}\to\ccat{C}\] uniquely determined by $\iota_{c}\mathbin{\fatsemi} s^{\sqcup}=s_{c}$ for each $c\in\ccat{C}$ and inclusion $\iota_{c}\colon\mathbb{P}\rightarrowtail\bigsqcup_{\ob\ccat{C}}\mathbb{P}$.\hfill$\qed$
\end{proposition}
\begin{proposition}\label{prop.p_supplies_itself}
Let $\mathbb{P}$ be a po-prop. Then there is a supply of $\mathbb{P}$ in $\mathbb{P}$ where the functors $s_{c}$ for $c\in\ob\mathbb{P}$ and $\mu\colon m\to n$ of $\mathbb{P}$ are given by
\[ s_{c}(\mu)\colon s_{c}(m)=\underbrace{c+\ldots +c}_{m}=\underbrace{m+\ldots+m}_{c}\xrightarrow{\mu+\ldots+\mu} \underbrace{n+\ldots+n}_{c}=\underbrace{c+\ldots+c}_{n}=s_{c}(n)\ .\hfill\qed\]
\end{proposition}
\begin{proposition}[Change of supply]\label{cor.change_of_supply}
Let $G\colon\mathbb{P}\to\mathbb{Q}$ be a po-prop functor. For any supply $s$ of $\mathbb{Q}$ in $\ccat{C}$, we have a supply $(G\mathbin{\fatsemi} s)$ of $\mathbb{P}$ in $\ccat{C}$.\hfill$\qed$
\end{proposition}
\begin{example}\label{ex.wiring_comonoids}
Any supply of $\mathbb{W}$ in $\ccat{C}$ induces a supply of cocommutative comonoids in $\ccat{C}$ by change of supply and \cref{def.comon_mon_selfduals}, and hence implies that $\ccat{C}$ has finite products (among left adjoints) by \cite{fox1976coalgebras}.
\end{example}
\begin{example} \label{ex.W_self_dual}
We say that a symmetric monoidal po-category $(\ccat{C},I,\otimes)$ is self-dual compact closed iff it supplies self duals (see \cref{def.comon_mon_selfduals}). Thus $\mathbb{W}$ is self-dual compact closed by \cref{prop.p_supplies_itself}, as is any po-category $\ccat{C}$ supplying $\mathbb{W}$ by \cref{cor.change_of_supply}.
\end{example}
\begin{example}\label{ex.corel}
Consider the po-prop of finite sets and corelations, where a morphism $m\to n$ is an equivalence relation on $m+n$, and the order is given by coarsening. Since this po-prop receives a po-prop functor from $\mathbb{W}$ (see \cref{prop.ww_explicit}), it supplies $\mathbb{W}$ by \cref{prop.p_supplies_itself,cor.change_of_supply}.
\end{example}
\begin{definition}[Preservation of supply]\label{def.preserve_supply}
Let $\mathbb{P}$ be a po-prop, $\ccat{C}$ and $\ccat{D}$ symmetric monoidal po-categories, and suppose $s$ is a supply of $\mathbb{P}$ in $\ccat{C}$ and $t$ is a supply of $\mathbb{P}$ in $\ccat{D}$. We say that a strong symmetric monoidal po-functor $(F,\varphi)\colon\ccat{C}\to\ccat{D}$ \define{preserves the supply} if the strongators $\varphi$ provide an isomorphism $t_{Fc}\cong (s_c\mathbin{\fatsemi} F)$ of po-functors $\mathbb{P}\to\ccat{D}$ for each $c\in\ccat{C}$.
\end{definition}
Unpacking, a strong monoidal po-functor $(F,\varphi)$ preserves the supply iff the following diagram commutes for each morphism $\mu\colon m\to n$ in $\mathbb{P}$ and object $c\in\ccat{C}$:
\begin{diagram}[][][baseline=(p.base)]\label{eqn.unpack_preserve_supply}
\node(1)[]{$F(c)\tpow m$};
\node(2)[right= of 1]{$F(c)\tpow n$};
\node(3)[below= of 1]{$F(c\tpow m)$};
\node(4)[below= of 2]{$F(c\tpow n)$};
\draw[a](1)to node[la,above]{$\mu_{F(c)}$}(2);
\draw[a](3)to node[la,below]{$F(\mu_{c})$}(4);
\draw[a](1)to node(p)[la,left]{$\varphi$}(3);
\draw[a](2)to node[la,right]{$\varphi$}(4);
\end{diagram}
One common example of preservation of supply arises from the supply of a po-prop $P$ in itself.
\begin{lemma}\label{lemma.supply_preserves_supply}
Let $\ccat{C}$ be a po-category supplying $\mathbb{P}$ a po-prop, and let $c\in\ob\ccat{C}$ be an object. Then the strong symmetric monoidal po-functor $s_{c}\colon\mathbb{P}\to\ccat{C}$ determined by the supply of $\mathbb{P}$ in $\ccat{C}$ preserves the supply of $\mathbb{P}$ in itself of \cref{prop.p_supplies_itself}.
\end{lemma}
\begin{proof}
By unwinding this claim we may see that it reduces to iterated applications of the supply conditions \eqref{eqn.supply_commute_tensors} for $s_{c}$.
\end{proof}
\section{Supplying \texorpdfstring{$\mathbb{W}$}{W}}\label{sec.ww_suppliers}
Our definition of regular calculi will be built upon a po-category supplying $\mathbb{W}$. The supply of $\mathbb{W}$ provides a notion of equality: roughly speaking, the morphisms in $\mathbb{W}$, represented by cospans, allow the marking of any number of objects as `the same', and composition in $\mathbb{W}$ takes the transitive closure of two markings, re-completing the result to an equivalence relation. Graphically, we reflect this in our depicting of the morphisms as $\mathbb{W}$ as wirings: wires connect objects that are to be considered `the same'.
In this section we explore a few properties of po-categories supplying $\mathbb{W}$, emphasising the sense in which their underlying 1-categories can be understood as undirected thus fulfilling a promise from the introduction. We have already seen a hint of this in \cref{ex.W_self_dual}, where we noted that any po-category supplying $\mathbb{W}$ has an induced supply of self-dualities (i.e.\ is self-dual compact closed). This implies that we may bijectively move tensor factors of the domain to the codomain and visa-versa by making use of the supplied images of the morphisms $\const{cup}$ and $\const{cap}$. Roughly speaking, this allows us to treat the distinction between domain and codomain as a matter of bookkeeping. We give two important examples of this, depicting them in string diagrams.
The first is the transpose of a morphism, which exchanges domain and codomain.
\begin{definition}[Transpose]
Let $f\colon c \to d$ be a morphism in a self dual compact closed po-category $\ccat{C}$. We define its \define{transpose} $f^{\dagger}\colon d \to c$ to be the morphism
\[
f^{\dagger}
\coloneqq\quad
\begin{tikzpicture}[inner WD, baseline=(f.-20)]
\node[oshellr, syntax] (f) {$f$};
\draw (f.west) --
++(-.5, 0) to[out=180, in=180]
++(0, 1.25) --
++(4.5, 0)
;
\draw (f.east) --
++(.5,0) to[out=0, in=0]
++(0, -1.25) --
++(-4.5, 0)
;
\end{tikzpicture}
\]
\end{definition}
\begin{notation}[Transpose]\label{notation.transpose}
Note that we have used the asymmetric icon
$
\begin{tikzpicture}[inner WD, baseline=(f.-20)]
\node[oshellr, syntax] (f) {$f$};
\draw (f.west) -- +(-.5,0);
\draw (f.east) -- +(.5,0);
\end{tikzpicture}
$
to denote $f$. This asymmetry allows us to distinguish the domain against the codomain without appealing to the orientation of the picture: the domain $c$ is attached at largest angle of the kite, while the domain $d$ is attached at the least angle.
In line with this convention, and the adjunction equations \eqref{eqn.yanking}, we depict the transpose simply by the 180-degree rotation of $f$:
\[
f^{\dagger} =
\begin{tikzpicture}[inner WD, baseline=(f.-20)]
\node[oshellr, syntax] (f) {$f^{\dagger}$};
\draw (f.west) -- +(-.5,0);
\draw (f.east) -- +(.5,0);
\end{tikzpicture}
=
\begin{tikzpicture}[inner WD, baseline=(f.-25)]
\node[oshelll, syntax] (f) {$f\phantom{^{\dagger}}$};
\draw (f.west) -- +(-.5,0);
\draw (f.east) -- +(.5,0);
\end{tikzpicture}
\]
We will make liberal use of asymmetric icons when it will help to have rotation denote transpose.
\end{notation}
Transposition gives an isomorphism $\ccat{C}(c,d) \cong \ccat{C}(d,c)$ of posets. The second example of moving factors between domain and codomain is given by naming and unfolding. These give an isomorphism $\ccat{C}(c,d) \cong \ccat{C}(I,c\otimes d)$.
\begin{definition}[Name and unfolding]\label{def.name_unfolding}
Let $f\colon c\to d$ and $g\colon I \to c\otimes d$ be morphisms in a self dual compact closed po-category. We define the \define{name} $\name{f}\colon I\to c\otimes d$ of $f$ and the \define{unfolding} $\unname{g}\colon c\to d$, to be respectively:
\begin{center}
\begin{tikzpicture}[inner WD, syntax]
\node[oshellr] (f) {$f$};
\draw (f.east) --
++(.5,0) -- +(.5, 0) coordinate (end)
node[right] {$c$}
;
\draw (f.west) --
++(-.5, 0) to[out=180, in=180]
++(0, 1.5) coordinate (top) --
(top-|end)
node[right] {$d$};
;
\node[oshellr,right=4cm of end] (g) {$g\vphantom{f}$};
\draw (g.30) --
++(1, 0) to[out=0, in=0]
++(0, 1) --
++(-4, 0)
node[left] {$c$};
;
\draw (g.-30) --
++(2, 0)
node[right] {$d$};
;
\begin{scope}[font=\normalsize]
\node[left=1.5 of f] {$\name{f}\coloneqq$};
\node[below=.5 of f] (formula) {$\const{cup}_c\mathbin{\fatsemi}(\id_c\otimes f)$};
\node[left=1.5 of g] {$\unname{g}\coloneqq$};
\node at (g|-formula) {${\rho_{c}}^{\text{-}1}\mathbin{\fatsemi}(\id_c\otimes g)\mathbin{\fatsemi}(\const{cap}_c\otimes\id_d)\mathbin{\fatsemi}\lambda_{d}$};
\end{scope}
\end{tikzpicture}
\end{center}
\end{definition}
We invite the reader to construct the pleasing graphical proof of the following.
\begin{lemma}\label{lemma.name_unfolding_iso}
The maps $\name{\cdot}$ and $\unname{\cdot}$ are mutually inverse.\hfill$\qed$
\end{lemma}
More generally, this `undirected' nature of 1-morphisms is a manifestation of the fact that the underlying 1-category of a po-category supplying $\mathbb{W}$ is a \emph{hypergraph category}. While we will not go into this in detail, to give the rough idea, recall that we refer to the generating morphisms of $\mathbb{W}$ as $(\eta,\mu,\epsilon,\delta)$; see \eqref{eqn.generating_wires}. These morphisms allow, visually, the termination, combining, initialisation, and splitting of wires. Moreover, the equations relating these generators in \eqref{eqn.equations_wires} define what is known as a \emph{special commutative frobenius monoid}, and capture precisely the fact that only the connectivity of their composites matter---for example, the special law says that splitting a wire and then combining the resulting pair is the same as doing nothing to that wire. For more on hypergraph categories, see \cite{fong2019hypergraph}. In our work here, rather than explore the intricacies of hypergraph categories, we will use wiring diagrams to capture this invariance.
\section{Wiring diagrams for po-categories supplying \texorpdfstring{$\mathbb{W}$}{W}}\label{sec.wds_for_ww_suppliers}
We have used two methods for depicting morphisms in $\mathbb{W}$. The first is the string diagrams of Joyal--Street, which we used in \cref{sec.prop_ww}, and which works for any monoidal category. In these diagrams, the domain of the morphism is represented on the left of the diagram, and the codomain is on the right. The second method is the pictures of \cref{sec.depict_ww}. For these diagrams, the domain of the morphism is represented by the interior blue circles, while the codomain is represented by the outer circle.
We may extend this latter notation to a notation for morphisms any po-category $\ccat{C}$ supplying $\mathbb{W}$, which we term \emph{wiring diagrams}. The change in our notation is essentially that each wire is labelled by an object in $\ccat{C}$ and that in addition to IO shells, we may also have kites labelled by morphisms of $\ccat{C}$. We first define these IO shells and kites, and describe how to draw them.
\begin{definition}[IO shell]
Let $\ccat{C}$ be a po-category supplying $\mathbb{W}$. An \define{input-output shell}, or \define{IO shell}, $\Gamma = (n,\tau)$ in $\ccat{C}$ is a function $\tau\colon \ord{n} \to \ob\ccat{C}$.
\end{definition}
Whereas in picturing morphisms of $\mathbb{W}$ (\cref{sec.depict_ww}) IO shells stand for objects, in a po-category $\ccat{C}$ supplying $\mathbb{W}$ IO shells are intended to capture ``contexts'': specified, ordered collections of ``types''. These IO shells will help us organise and capture the calculus we need in order to support a regular logic of relations.
\begin{notation}[Depicting IO shells]
We depict an IO shell as a circle with $n$ ports, labelled clockwise starting from its left with the objects $\tau(i) \in \ob\ccat{C}$:
\[
\Gamma =
\begin{aligned}
\begin{tikzpicture}[inner WD]
\node[pack,minimum size=16pt] (rho) {};
\draw (rho.180) to[pos=1] node[left] (w) {$\tau(1)$} +(180:2pt);
\draw (rho.110) to[pos=1] node[above] (n) {$\tau(2)$} +(110:2pt);
\node at ($(rho.0)+(60:5pt)$) {$\vdots$};
\draw (rho.-110) to[pos=1] node[below] (e) {$\tau(n)$} +(-110:2pt);
\end{tikzpicture}
\end{aligned}
\]
We shall abuse notation by writing $\Gamma$ for both the pair $(n,\tau)$ as well as the tensor product $\Gamma \coloneqq \tau(1) \otimes \dots \otimes \tau(n)\in\ob\ccat{C}$.
\end{notation}
\begin{definition}[Kite]
Let $\ccat{C}$ be a po-category supplying $\mathbb{W}$. A \define{kite} $(\kappa,f) \coloneqq (a,b,\tau_l,\tau_r,f)$ is a pair of natural numbers $a,b$, functions $\tau_l\colon \ord{a} \to \ob\ccat{C}$, $\tau_r\colon \ord{b} \to \ob\ccat{C}$, and a morphism $f\colon \bigotimes_{i \in \ord{a}} \tau_l(i) \to \bigotimes_{i \in \ord{b}} \tau_r(i)$ in $\ccat{C}$.
\end{definition}
\begin{notation}[Depicting kites]
We depict a kite as a kite with $a$ ports on the side with greater angle, $b$ ports on the side with lesser angle, with each port labelled by the objects $\tau_l(i), \tau_r(i) \in \ob\ccat{C}$ in sequence, and the kite itself labelled by $f$:
\[
\kappa =
\begin{tikzpicture}[WD, baseline=(rho.base)]
\node[oshellr,syntax,minimum size=21pt] (rho) {$f$};
\draw (rho.120) to[pos=1] node[left,la] (n) {$\tau_l(1)$} +(150:10pt);
\draw (rho.-120) to[pos=] node[left,la] (e) {$\tau_l(a)$} +(210:10pt);
\path (n) to node[sloped]{$\ldots$} (e);
\draw (rho.60) to[pos=1] node[right,la] (m) {$\tau_r(1)$} +(30:10pt);
\draw (rho.-60) to[pos=1] node[right,la] (d) {$\tau_r(b)$} +(330:10pt);
\path (m) to node[sloped]{$\ldots$} (d);
\end{tikzpicture}
\]
As we do for IO shells, we shall abuse notation by writing $\kappa$ for both the tuple $(a,b,\tau_l,\tau_l)$ as well as the tensor product $\kappa \coloneqq \tau_l(1) \otimes \dots \otimes \tau_l(a) \otimes \tau_r(1) \otimes \dots \otimes \tau_r(b)$.
\end{notation}
A wiring diagram consists of IO shells and kites wired together by morphisms in the supply of $\mathbb{W}$. Recall from \cref{prop.supply_coprod} that a supply of $\mathbb{W}$ in $\ccat{C}$ induces a strong monoidal functor $s^\sqcup \colon \bigsqcup_{\ob\ccat{C}} \mathbb{W} \to \ccat{C}$.
\begin{definition}[Wiring diagram]\label{def.extended_wiring_diagram}
Let $\ccat{C}$ be a po-category supplying $\mathbb{W}$. A \define{wiring diagram for $\ccat{C}$} is:
\begin{enumerate}[label=(\roman*)]
\item natural numbers $s$ and $k$,
\item for each $i\in\{1,\dots, s,\text{out}\}$, an IO shell $\Gamma_i$,
\item for each $i\in \{1, \dots, k\}$, a kite $(\kappa_i, f_i)$,
\item a morphism $\omega^\ast\colon \Gamma_1 \otimes \dots \otimes \Gamma_s \otimes
\kappa_1 \otimes \dots \otimes \kappa_k
\to \Gamma_\mathrm{out}$ in the image of $s^\sqcup$ in $\ccat{C}$.
\end{enumerate}
\end{definition}
\begin{notation}[Wiring diagrams for a supply]\label{notation.wiring_diagrams_supplying_ww}
We depict wiring diagrams for $\ccat{C}$ in much the same way as we depict wiring diagrams for $\mathbb{W}$ (see \cref{notation.picturing_ww}), except that wires are now labelled by objects of $\ccat{C}$, and we draw kites as well as IO shells.
To give a bit more detail, since the morphism $\omega^\ast$ is in the image of $\bigsqcup_{\ob\ccat{C}} \mathbb{W}$ under $s^\sqcup$, it may be canonically represented by a cospan of finite sets
\[
\begin{tikzcd}
\ord{n_1}+\dots+\ord{n_s}+\ord{a_1}+\ord{b_1}+\dots+\ord{a_k}+\ord{b_k} \ar[r] \ar[dr,"{[\tau_1,\dots,\tau_s,\tau_{l_1},\dots,\tau_{r_k}]}"']
& \ord{n_\omega} \ar[d,"\tau_\omega"]
& \ord{n_{\text{out}}} \ar[l] \ar[dl,"\tau_{\text{out}}"] \\
& \ob\ccat{C}
\end{tikzcd}
\]
such that the above diagram commutes. We depict this cospan as we do in \cref{notation.picturing_ww}, labelling each wire with the object of $\ccat{C}$ indicted by the $\tau_i$.
\end{notation}
\begin{example} \label{ex.a_wiring_diagram}
Here's a depiction of a wiring diagram in which $s = k= 2$, $\Gamma_1\coloneqq c_1\otimes c_2\otimes c_3$, $\Gamma_2\coloneqq c_2 \otimes c_5\otimes c_6\otimes c_3$, $\kappa_1 = c_1\otimes c_5 \otimes c_5$, $\kappa_2 = c_4 \otimes c_5$, and $\Gamma_\mathrm{out}\coloneqq c_5\otimes c_4\otimes c_5\otimes c_3$. Note in particular that $c_{5}$ appears twice in $\Gamma_{\mathrm{out}}$.
\begin{center}
\examplewiringdiagram{}
\end{center}
\end{example}
Each wiring diagram itself represents a morphism $\Gamma_1 \otimes \dots \otimes \Gamma_s \to \Gamma_{\text{out}}$ in $\ccat{C}$.
\begin{definition}[The morphism represented by a wiring diagram] \label{def.morphism_represented_by_wd}
Let
\[
(s,k,\{\Gamma_{i}\}_{i\in\ord s},\, \{(\kappa_i,f_i)\}_{i \in \ord k},\, \omega^\ast)
\]
be a wiring diagram. We say that this wiring diagram \define{represents} the following morphism of $\ccat{C}$.
\[
\omega \coloneqq \big(\Gamma_1 \otimes \dots \otimes \Gamma_s \otimes \name{f_1} \otimes \dots \otimes \name{f_k}\big) \mathbin{\fatsemi} \omega^\ast \colon \Gamma_1 \otimes \dots \otimes \Gamma_s \longrightarrow \Gamma_{\text{out}}
\]
\end{definition}
\begin{example}
The wiring diagram in \cref{ex.a_wiring_diagram} represents the following morphism $\Gamma_1\otimes \Gamma_2\to \Gamma_\mathrm{out}$ of $\ccat{C}$. Note in particular that $c_{5}$ appears twice in $\Gamma_{\mathrm{out}}$, that is, $\Gamma_{\mathrm{out}}$ includes the $c_{5}\otimes c_{5}$ up to symmetry.
\begin{center}
\begin{tikzpicture}
\begin{scope}[font=\scriptsize,x=1em, decoration={brace, amplitude=4pt},y=3ex]
\node (a1) {$1$};
\node[below=0 of a1] (a2) {$2$};
\node[below=0 of a2] (a3) {$3$};
\node[below=0 of a3] (b1) {$2$};
\node[below=0 of b1] (b2) {$5$};
\node[below=0 of b2] (b3) {$6$};
\node[below=0 of b3] (b4) {$3$};
\node[right=15] at (b1) (out1) {$5$};
\node[below=0 of out1] (out2) {$4$};
\node[below=0 of out2] (out3) {$5$};
\node[below=0 of out3] (out4) {$3$};
\draw[decorate] ($(a3.west)+(-5pt, -3pt)$) to node[left=6pt] {$\Gamma_1$} ($(a1.west)+(-5pt, 3pt)$);
\draw[decorate] ($(b4.west)+(-5pt, -3pt)$) to node[left=6pt] {$\Gamma_2$} ($(b1.west)+(-5pt, 3pt)$);
\draw[decorate] ($(out1.east)+(5pt, 3pt)$) to node[right=6pt] {$\Gamma_\mathrm{out}$} ($(out4.east)+(5pt, -3pt)$);
\end{scope}
\begin{scope}[inner WD]
\node[oshellr, syntax, below=of b4] (f1) {$\name{f_1}$};
\node[oshellr, syntax, below=of f1] (f2) {$\name{f_2}$};
\draw(a1) -- (a1-|f1.east) coordinate(a1e);
\draw(f1.45) -- (f1.45-|f1.east) coordinate(f1t);
\draw(f1t) to[out=0,in=0] (a1e);
\draw(b2) -- (b2-|f1.east) coordinate(b2e);
\draw(f1.-45) -- (f1.-45-|f1.east) coordinate(f1b);
\draw(f1b) to[out=0,in=0] (b2e);
\draw(f1) to ($(f1)+(1.5,0)$) to[out=0,in=180] ($(out1)-(1.5,0)$) to (out1);
\draw(f2.45) to ($(f2.45)+(1.5,0)$) to[out=0,in=180] ($(out2)-(1,0)$) to (out2);
\draw(f2.-45) to ($(f2.-45)+(2,0)$) to[out=0,in=180] ($(out3)-(0.5,0)$) to (out3);
\draw(b3) -- (b3-|f1.east) node[link]{};
\draw(a2) to (a2-|f1.east) to[out=0,in=0] (b1-|f1.east) to (b1);
\node(l)[right=2 of b2,link]{};
\draw(a3) to ($(a3)+(1,0)$) to[out=0,in=90] (l);
\draw(b4) to ($(b4)+(1,0)$) to[out=0,in=270] (l);
\draw(l) to ($(l)+(1,0)$) to[out=0,in=180] ($(out4)-(1,0)$) to (out4);
\end{scope}
\end{tikzpicture}
\end{center}
Note that we may use the manipulations afforded to us by the string diagram calculus to draw this as the following string diagram instead
\begin{diagram*}[,][out4]
\begin{scope}[font=\scriptsize,x=1em, decoration={brace, amplitude=4pt},y=3ex]
\node (a1) {$1$};
\node[below=0 of a1] (a2) {$2$};
\node[below=0 of a2] (a3) {$3$};
\node[below=0 of a3] (b1) {$2$};
\node[below=0 of b1] (b2) {$5$};
\node[below=0 of b2] (b3) {$6$};
\node[below=0 of b3] (b4) {$3$};
\coordinate (a2a3) at ($(a2)!.5!(a3)$);
\node[right=6] at (a2a3) (out1) {$5$};
\node[below=0 of out1] (out2) {$4$};
\node[below=0 of out2] (out3) {$5$};
\node[below=0 of out3] (out4) {$3$};
\draw[decorate] ($(a3.west)+(-5pt, -3pt)$) to node[left=6pt] {$\Gamma_1$} ($(a1.west)+(-5pt, 3pt)$);
\draw[decorate] ($(b4.west)+(-5pt, -3pt)$) to node[left=6pt] {$\Gamma_2$} ($(b1.west)+(-5pt, 3pt)$);
\draw[decorate] ($(out1.east)+(5pt, 3pt)$) to node(go)[right=6pt] {$\Gamma_\mathrm{out}$} ($(out4.east)+(5pt, -3pt)$);
\end{scope}
\begin{scope}[inner WD]
\coordinate (helper) at ($(a2a3)!.5!(out1)$);
\coordinate (a1b1) at ($(a1)!.5!(b1)$);
\coordinate (a2b4) at ($(a2)!.5!(b4)$);
\coordinate (a3b3) at ($(a3)!.4!(b3)$);
\coordinate (out4out5) at ($(out2)!.5!(out3)$);
\node[oshellr, syntax] at (helper|-a1b1) (f1) {$f_1$};
\node[oshellr, syntax] at (helper|-out4out5) (f2) {$f_2$};
\node[link, right=of b3] (dot6) {};
\node[link] at (helper|-out4) (dot3) {};
\draw (a1) to[out=0,in=150] (f1);
\draw (b2) to[out=0,in=210] (f1);
\draw (f1) -- (out1);
\draw (b3) -- (dot6);
\draw (a2) to[out=0, in=0] (b1);
\draw (a3) to[out=0, in=135] (dot3);
\draw (b4) to[out=0, in=225] (dot3);
\draw (dot3) -- (out4);
\draw (out2) to[bend right=15pt] (f2);
\draw (out3) to[bend left=15pt] (f2);
\end{scope}
\end{diagram*}
which is a bit easier to read.
\end{example}
\begin{remark}
Wiring diagrams are robust to topological deformation, so long as we keep the connections to the outer IO shell---our ``context'', so to speak---consistent. For example, the following pictures all represent the same wiring diagram and indeed same morphism of $\ccat{C}$:
\begin{center}
\begin{tikzpicture}[inner WD]
\coordinate (g1);
\coordinate[below=2 of g1] (g2);
\node[link] at ($(g1)!.5!(g2)+(1,0.3)$) (dot) {};
\node[oshelll, syntax, right=.5 of dot] (f0) {$f$};
\node[oshelll, syntax, left=.5 of g1,inner sep=1pt] (g0) {$g$};
\draw (f0) -- (dot);
\draw (g0) -- (g1);
\draw (g1) to[out=0, in=120] (dot);
\draw (g2) to[out=0, in=-120] (dot);
\draw (g2) -- +(-3,0);
\draw (dot) to[out=-45, in=180] +(1,-1.3) -- +(3.5,0);
\draw (f0.east) to[out=0, in=0] +(0,2) -- ($(dot-|g1)+(0,2)$);
\draw ($(dot-|g1)+(0,2)$) -- +(-3,0);
\end{tikzpicture}
\hspace{1.5cm}
\begin{tikzpicture}[inner WD]
\node[link] (dot) {};
\node[oshelld, syntax, above=.5 of dot] (f0) {$f$};
\node[oshelld, syntax, below=.5 of dot,inner sep=1pt] (g0) {$g$};
\draw (f0) -- (dot);
\draw (g0) -- (dot);
\draw (dot) -- +(-3,0);
\draw (dot) -- +(3,0);
\draw (f0) to[out=90, in=0] +(-1,1.5) -- +(-2,0);
\end{tikzpicture}
\hspace{1.5cm}
\begin{tikzpicture}[inner WD]
\coordinate (g1);
\coordinate[below=2 of g1] (g2);
\node[link] at ($(g1)!.5!(g2)+(1,0)$) (dot) {};
\node[oshellr, syntax, left=.5 of g1] (f0) {$f$};
\node[oshellu, syntax, above=1 of dot,inner sep=1pt] (g0) {$g$};
\draw (f0.west) -- +(-1,0);
\draw (f0.east) -- (g1);
\draw (g1) to[out=0, in=120] (dot);
\draw (g2) to[out=0, in=-120] (dot);
\draw (dot) -- +(2,0);
\draw (g0) -- (dot);
\draw (g2) -- (g2-|f0.west) -- +(-1,0);
\end{tikzpicture}
\end{center}
\end{remark}
\begin{remark}\label{rem.suppress}
In the coming sections, where it causes no ambiguity, we will gradually suppress infer-able details in our wiring diagrams, as well as choose topological layouts for wiring diagrams that emphasise easy reading and understanding, rather than strictly following our procedure for laying them out as per \cref{notation.picturing_ww}.
\end{remark}
\begin{definition}[Bare wiring diagram]
A wiring diagram is a \define{bare wiring diagram} if $k=0$; that is, if it contains no kites.
\end{definition}
\begin{remark} \label{rem.bare_wd_represents_itself}
In the case of a bare wiring diagram, $\omega= \omega^\ast \colon \Gamma_1 \otimes \dots \otimes \Gamma_s \to \Gamma_{\text{out}}$.
\end{remark}
\begin{proposition}\label{prop.barebare}
Let $ \big(s,k,\{\Gamma_{i}\}_{i\in\ord s},\, \{(\kappa_i,f_i)\}_{i \in \ord k},\, \omega^\ast\big) $ be a wiring diagram in a po-category $\bigsqcup_{J} \mathbb{W}$ that is a coproduct of copies of $\mathbb{W}$. Then
\[
\big(s,0,\{\Gamma_{i}\}_{i\in\ord s},\, \varnothing,\, (\Gamma_1 \otimes \dots \otimes \Gamma_s \otimes \name{f_1} \otimes \dots \otimes \name{f_k}\big) \mathbin{\fatsemi} \omega^\ast)\big)
\]
is a bare wiring diagram representing the same morphism, where $\name{f}\colon I\to\dom(f)\otimes\cod(f)$ is the name of $f$, see \cref{def.name_unfolding}.
\end{proposition}
\begin{proof}
This follows from \cref{rem.bare_wd_represents_itself} and the fact that all morphisms in $\bigsqcup_{J} \mathbb{W}$ are in the image of the supply functor $s^\sqcup$.
\end{proof}
Although wiring diagrams in their general form will play a critical role in our upcoming explorations of regular calculi and regular logic, a corollary of our main theorem (\cref{cor.vague_eqv}) is that bare wiring diagrams suffice in general for our purposes.
\chapter{Regular calculi}\label{chap.regular_calculi}
Graphical regular logic is the graphical representation of the logical calculus carried by a \emph{regular calculus}: a structure understanding contexts, predicates, and the regular fragment of logic thereupon. In this section we will introduce the framework of \emph{right ajax po-functors} which will allow us to compactly capture all of the desired properties of regular calculi. We will then define regular calculi in terms of these right ajax po-functors and assemble regular calculi into a suitable $2$-category.
\section{Right ajax po-functors and right adjoint monoids}
\begin{definition}[Right ajax po-functor]\label{def.ajax}
Let $\ccat{C}$ and $\ccat{D}$ be symmetric monoidal po-categories. A \define{right adjoint-lax} or \define{right ajax} po-functor $F\colon \ccat{C} \to \ccat{D}$ is a lax symmetric monoidal po-functor for which the laxators are right adjoints.
We denote the laxators by $\rho$ and their left adjoints by $\lambda$:
\begin{equation}\label{eqn.ajax}
\adjr{I}{\rho}{\lambda}{F(I)}
\qquad\text{and}\qquad
\adjr{F(c)\otimes F(c')}{\rho_{c,c'}}{\lambda_{c,c'}}{F(c\otimes c')}.
\end{equation}
\end{definition}
\begin{example}[Cartesian monoidal po-categories]\label{ex.terminal_ajax}
If $\ccat{D}$ is a cartesian monoidal po-category then every symmetric monoidal po-category $\ccat{C}$ has a canonical right ajax po-functor $\ccat{C}\to\ccat{D}$, viz., the constant po-functor at the terminal object of $\ccat{D}$. This po-functor is easily seen to be terminal in the po-category of right ajax po-functors $\ccat{C}\to\ccat{D}$ and monoidal $2$-natural transformations. Once we have proven \cref{prop.adjoint_monoids} below we will see that this is a special case of the composition of po-functors $\ccat{C}\to1\to\ccat{D}$ and the fact that $1$ is an \emph{right adjoint monoid} in $\ccat D$. Although we do not prove it here, for a class of symmetric monoidal po-categories $\ccat{C}$ we shall see in the companion that the initial right ajax po-functor $\ccat{C}\to\poCat{Poset}$ is given by $\ccat{C}(I,-)$.
\end{example}
\begin{example}[Represented po-functors]\label{ex.ajax_defer}
As we shall see later in \cref{sec.regular_calc,sec.prd} there are many important classes of examples of right ajax po-functors. One such class, established in \cref{prop.prerel_ajax}, is the represented po-functor $\ccat{R}(I,-)\colon\ccat{R}\to\poCat{Poset}$ for any \emph{prerelational} po-category $\ccat{R}$. Examples of prerelational po-categories include the po-category $\Funr{LAdj}\cat{R}$ of left adjoints in a regular category $\ccat{R}$---see \cref{thm.carboni_equivalence} for details.
\end{example}
\begin{warning}
If $F$ is a right ajax po-functor, the left adjoints $\lambda_{c,c'}$ of the laxators \emph{do not} in general equip $F$ with the structure of an oplax monoidal functor. \emph{A priori}, the naturality squares for the laxators $\rho$ are only lax-naturality squares for their left adjoints $\lambda$.
\end{warning}
Let us record two immediate but nevertheless useful consequences of the definition of right ajax po-functors.
\begin{lemma}\label{lemma.ajax}\
\begin{enumerate}
\item Every strong monoidal functor between monoidal po-categories is right ajax.
\item The composite of right ajax po-functors is again right ajax.
\end{enumerate}
\end{lemma}
To understand right ajax po-functors and their importance in the story of regular logic, it is useful to introduce the notion of \emph{right adjoint monoid}. Let us write $1$ for the terminal monoidal po-category.
\begin{proposition}\label{prop.adjoint_monoids}
Let $(\ccat{C},I,\otimes)$ be a symmetric monoidal po-category. There is a bijection between:
\begin{enumerate}
\item The set of right ajax po-functors $1\to\ccat{C}$,
\item The set of commutative monoid objects $(c,\mu,\eta)$ such that $\mu$ and $\eta$ are right adjoints,
\item The set of cocommutative comonoid objects $(c,\delta,\epsilon)$ such that $\delta$ and $\epsilon$ are left adjoints.
\end{enumerate}
\end{proposition}
\begin{proof}
$(1)\Leftrightarrow(2)$: The set $\Cat{Lax}(1,\ccat{C})$ of symmetric lax monoidal po-functors $1\to\ccat{C}$ may be seen to be in bijection with the set of commutative monoid objects $(c,\mu,\eta)$ in $\ccat{C}$. Indeed, $\eta$ and unit $\mu$ come from the 0-ary and $2$-ary laxators respectively: $\eta=\rho$ and $\mu=\rho_{1,1}$. Hence the added condition that $\eta$ and $\mu$ have left adjoints is precisely the right ajax condition.
\noindent
\hphantom{\textit{Proof.} }
$(2)\Leftrightarrow(3)$: This bijection is implemented by taking adjoints. That is, given an object $c\in\ccat{C}$ and two adjunctions
\begin{equation}\label{eqn.adjmon}
\adjr{I}{\eta}{\epsilon}{c}
\qquad\text{and}\qquad
\adjr{c\otimes c}{\mu}{\delta}{c},
\end{equation}
it may be verified that $\mu$ and $\eta$ satisfy the commutative monoid laws if and only if $\delta$ and $\epsilon$ satisfy the cocommutative comonoid laws. \qedhere
\end{proof}
To summarise, if $(c,\rho,\lambda)\colon 1\to\ccat{C}$ is a right ajax po-functor then the corresponding monoid and comonoid structures on $c$ are given by
\begin{equation}\label{eqn.monoid_comonoid_ajax}
\eta=\rho\qquad
\mu=\rho_{1,1}
\qquad\text{and}\qquad
\epsilon=\lambda\qquad
\delta=\lambda_{1,1}
\end{equation}
\Cref{prop.adjoint_monoids} motivates the following definition.
\begin{definition}[Right adjoint monoid]
Let $(\ccat{C},I,\otimes)$ be a monoidal po-category. A \define{right adjoint monoid} in $\ccat{C}$ is a commutative monoid object $(c,\mu,\eta)$ in $\ccat{C}$ such that $\mu$ and $\eta$ are right adjoints.
\end{definition}
Right adjoint monoids are like \emph{internal $\wedge$-semilattices}; see \cite[Chapter 5]{schalk1994algebras} and references therein. To get a feel for why this might be so, it helps to first recall a po-categorical analogue of a classical lemma of \cite{fox1976coalgebras}.
\begin{lemma}\label{lemma.comonoids_unique}
Let $\ccat{C}$ be a monoidal po-category. If the monoidal structure is cartesian (given by finite products in the underlying 1-category) then every object has a unique comonoid structure, and it is cocommutative.
\end{lemma}
\begin{proof}
Since the unit object is terminal, the maps $c\times\epsilon$ and $\epsilon\times c$ are forced to be the projections $c\times c\to c$, so $\delta$ is forced to be the diagonal. Commutativity follows by universal property arguments.
\end{proof}
\begin{example}[Right adjoint monoids in $\poCat{Poset}$ are $\wedge$-semilattices]\label{prop.adjmon_msl}
A poset $P\in\poCat{Poset}$ is an adjoint monoid iff it is a meet-semilattice, in which case $\eta=\const{true}$ and $\mu=\wedge$. To see this we can use \cref{lemma.comonoids_unique}. This states that any poset $P$ has a unique comonoid structure given by the terminal and diagonal maps $\epsilon\colon P\to 1$ and $\delta\colon P\to P\times P$. Thus $P$ is an adjoint monoid iff these maps have right adjoints as in \eqref{eqn.adjmon}, which holds iff $\eta$ is a top element and $\mu$ is a meet.
Note however that a map of right adjoint monoids in $\poCat{Poset}$, defined to be a monoidal $2$-natural transformation (\cref{def.symm_mon_po}) between the corresponding right ajax po-functors $1\to\poCat{Poset}$ (\cref{prop.adjoint_monoids}), does \emph{not} preserve meets---instead it is merely a monotone map between posets which happen to be $\wedge$-semilattices. This is a deliberate choice, and will feature in our coming definition of morphism of regular calculus.
\end{example}
\begin{example}[Right adjoint monoids in a relational po-category are objects]\label{prop.adjmon_reg}
In later sections we will have cause to consider \emph{relational} and \emph{prerelational} po-categories. By \cref{thm.carboni_equivalence}, every relational po-category is equivalent to the po-category of left adjoints $\Funr{LAdj}\cat R$ in a regular category $\cat R$. In particular, $\Funr{LAdj}\cat R$ is cartesian monoidal and so \cref{lemma.comonoids_unique} above applies to relational po-categories: each object has a unique cocommutative comonoid structure. By \cref{prop.adjoint_monoids} $(2)\Leftrightarrow(3)$ we see that each object in fact must have a unique commutative monoid structure.
\end{example}
It will be useful to know that, just as lax monoidal functors map monoids to monoids, right ajax po-functors map right adjoint monoids to right adjoint monoids.
\begin{proposition}\label{prop.ajax_pres_adjmon}
Right ajax po-functors send right adjoint monoids to right adjoint monoids.
\end{proposition}
\begin{proof}
The composite of right ajax po-functors is again right ajax, so the result follows from \cref{prop.adjoint_monoids}.
\end{proof}
Although this result may seem anodyne at present, we will consistently leverage this in our development of graphical regular calculus in \cref{chap.graphical_reglog} as it lends our graphical calculus its ``regular'' aspect.
\section{Regular calculi as indexed right adjoint monoids}\label{sec.regular_calc}
Now that we have understood the notion of right ajax po-functor, we move to define one of the central notions in this paper.
\begin{definition}[Regular calculus]\label{def.reg_calc}
A \define{regular calculus} P consists of a pair $(\cont{P},P)$ where $(\cont{P},I,\otimes)$ is a symmetric monoidal po-category supplying $\mathbb{W}$ and $P\colon\cont{P}\to\poCat{Poset}$ is a right ajax po-functor, whose laxators we denote by $\const{true}$ and $\boxplus$:
\begin{equation}\label{eqn.ajax_reg_calc}
\adjr{1}{\const{true}}{\lambda_{I}}{P(I)}
\qquad\text{and}\qquad
\adjr{P(\Gamma_1)\times P(\Gamma_2)} {\boxplus_{\Gamma_{1},\Gamma_{2}}}{\lambda_{\Gamma_{1},\Gamma_{2}}} {P(\Gamma_1\otimes \Gamma_2)}.
\end{equation}
We call $\cont{P}$ the po-category of \define{contexts}, and $P$ the \define{predicates} po-functor. If $\Gamma\in\ob\cont P$ is a context, and $\theta,\theta'\in P(\Gamma)$ are predicates in context $\Gamma$ such that $\theta\leq \theta'$, then we say that \define{$\theta$ entails $\theta'$}, and write $\theta\vdash \theta'$. Finally, a regular calculus is termed \define{bare} if $\cont{P}=\bigsqcup_{J}\mathbb{W}$ is a coproduct\footnote{see \cref{warn.coprod}} of copies of $\mathbb{W}$.
\end{definition}
As a first example let us cast $\wedge$-semilattices as regular calculi.
\begin{example}[$\wedge$-semilattices are regular calculi]\label{ex.reg_calc_meet_sl}
Observe that the terminal po-category $1$ supplies $\mathbb{W}$. By \cref{prop.adjoint_monoids,prop.adjmon_msl} a right ajax po-functor $1\to\poCat{Poset}$ is the same as a $\wedge$-semilattice. Hence a regular calculus whose po-category of contexts is terminal is a $\wedge$-semillatice.
Should we fix a $\wedge$-semilattice $L$ viewed as a right ajax po-functor $L\colon1\to\poCat{Poset}$, we may extend this idea to endow any po-category $\ccat{C}$ which supplies $\mathbb{W}$ with the structure of a regular calculus by considering the composite po-functor $\ccat{C}\To{!}1\To{L}\poCat{Poset}$.
\end{example}
This phenomenon, that regular calculi induce $\wedge$-semilattice structures on their posets of predicates, is not special to factoring through the terminal po-category $1$. In the following sense it must occurs point-wise for any regular calculus.
\begin{proposition}\label{prop.meet_sl}
If $\rc{P}$ is a regular calculus then the poset $P(\Gamma)$ has the structure of a $\wedge$-semilattice for each $\Gamma\in\cont P$.
\end{proposition}
\begin{proof} Recall that by supplying $\mathbb{W}$, $\cont P$ has a supply of chosen right adjoint monoid structures $(\eta_{\Gamma},\, \mu_{\Gamma})$ for each object $\Gamma\in\cont P$. The result then follows from fact that ajax po-functors send right adjoint monoids to right adjoint monoids by \cref{prop.ajax_pres_adjmon}, as well as \cref{prop.adjmon_msl}.
Explicitly, \eqref{eqn.ajax} and \eqref{eqn.ww_adjunctions} give rise to the following adjunctions.
\[
\begin{tikzpicture}[baseline=(1.base)]
\node(1)[]{\ensuremath{1}};
\node(2)[right= of 1]{\ensuremath{P(I)}};
\node(3)[right= of 2]{\ensuremath{P(\Gamma)}};
\draw[a]($(1.east)+(0,5pt)$)to node(left)[la,above]{\ensuremath{\const{true}}}($(2.west)+(0,5pt)$);
\draw[a]($(2.west)-(0,5pt)$)to node(right)[la,below]{\ensuremath{\lambda_{I}}}($(1.east)-(0,5pt)$);
\draw[a]($(2.east)+(0,5pt)$)to node(left)[la,above]{\ensuremath{P(\eta_{\Gamma})}}($(3.west)+(0,5pt)$);
\draw[a]($(3.west)-(0,5pt)$)to node(right)[la,below]{\ensuremath{P(\epsilon_{\Gamma})}}($(2.east)-(0,5pt)$);
\path (1) -- node[midway,rotate=90]{$\dashv$} (2);
\path (2) -- node[midway,rotate=90]{$\dashv$} (3);
\end{tikzpicture}\quad\text{and}\quad
\begin{tikzpicture}[baseline=(1.base)]
\node(1)[]{\ensuremath{P(\Gamma)\times P(\Gamma)}};
\node(2)[right= of 1]{\ensuremath{P(\Gamma\otimes\Gamma)}};
\node(3)[right= of 2]{\ensuremath{P(\Gamma)}};
\draw[a]($(1.east)+(0,5pt)$)to node(left)[la,above]{\ensuremath{\boxplus_{\Gamma,\Gamma}}}($(2.west)+(0,5pt)$);
\draw[a]($(2.west)-(0,5pt)$)to node(right)[la,below]{\ensuremath{\lambda_{\Gamma,\Gamma}}}($(1.east)-(0,5pt)$);
\draw[a]($(2.east)+(0,5pt)$)to node(left)[la,above]{\ensuremath{P(\mu_{\Gamma})}}($(3.west)+(0,5pt)$);
\draw[a]($(3.west)-(0,5pt)$)to node(right)[la,below]{\ensuremath{P(\delta_{\Gamma})}}($(2.east)-(0,5pt)$);
\path (1) -- node[midway,rotate=90]{$\dashv$} (2);
\path (2) -- node[midway,rotate=90]{$\dashv$} (3);
\end{tikzpicture}
\qedhere
\]
\end{proof}
This result motivates the following notation.
\begin{notation}[The $\wedge$-semilattice on each object in a regular calculus]
In a regular calculus $\rc P$, for each context $\Gamma\in\ob\cont P$ we denote the $\wedge$-semilattice structure on $P(\Gamma)$ given by \cref{prop.meet_sl} by:
\begin{equation}\label{eqn.meet_sl}
\adjr{1}{\const{true}_{\Gamma}}{!}{P(\Gamma)}
\qquad\text{and}\qquad
\adjr{P(\Gamma)\times P(\Gamma)}{\wedge_{\Gamma}}{\Delta_{\Gamma}}{P(\Gamma)}.
\end{equation}
In keeping with the conventions thus far, for $n\in\mathbb{N}$ we will write $\wedge^{n}_{\Gamma}$ for the composite map $\boxplus^{n}_{\Gamma}\mathbin{\fatsemi} P(\mu_{\Gamma}^{n})\colon P(\Gamma)^{\times n}\to P(\Gamma)$. In the case of $n=0$ we see that $\wedge_{\Gamma}^{0}=\const{true}_{\Gamma}$, and we will freely confuse this map $\const{true}_{\Gamma}$ with the top element $\const{true}_{\Gamma}(*)$. Where it will cause no ambiguity we will omit the label $\Gamma$ on the maps $\const{true}$, $\wedge$, $\wedge^{n}$, $\boxplus$, and so forth.
\end{notation}
Further examples of regular calculi abound. In the next section we will meet our first large class of regular calculi, those arising syntactically from ``regular theories''. After that, \cref{chap.relational_pocats} is devoted to constructing another large class, those arising semantically from regular categories.
\section{The \texorpdfstring{$2$}{2}-category \texorpdfstring{$\mathcal{R}\Cat{gCalc}$}{RgCalc} of regular calculi}
The central result of the companion paper compares the $2$-category theory of regular calculi with that of regular categories. In order to effect such a comparison we must first define suitable notions of $1$- and $2$-morphisms of regular calculi.
\begin{definition}[Morphism of regular calculi]\label{def.reg_calc_morphisms}
Let $(\cont P, P)$ and $(\cont{P'}, P')$ be regular calculi. A \define{morphism of regular calculi} from $(\cont P, P)$ to $(\cont{P'}, P')$ is a pair $(F,F^\sharp)$ where $(F,\varphi)\colon\cont P\to\cont{P'}$ is a strong monoidal po-functor preserving the supply and $F^\sharp$ is a monoidal $2$-natural transformation (\cref{def.symm_mon_po}) as follows.
\begin{diagram*}
\node(1)[]{$\cont {P\phantom{'}}$};
\node(2)[below=0.50cm of 1]{$\cont{P'}$};
\node(3)[right=1.00cm of 2]{$\poCat{Poset}$};
\draw[a](1)to node(F)[la,left]{$F$}(2);
\draw[a,rc](1) to (3|-1) to (3);
\draw[a](2)to node(P')[la,below]{$P'$}(3);
\node[above,la] at (P'|-1) {$P$};
\cell[right,la]{(P'|-1)}{(P'.north)}{$F\smash{^{\sharp}}$};
\end{diagram*}
A \define{$2$-morphism of regular calculi} from $(F,F^{\sharp})$ to $(G,G^{\sharp})$ is the data of a monoidal left adjoint oplax-natural transformation (\cref{def.symm_mon_po,def.ladj_lax_nt}) $\alpha\colon F\Rightarrow G$ which satisfies the property $F^{\sharp}\mathbin{\fatsemi} P'\alpha\leq G^{\sharp}$ as explicated below.
\begin{equation}\label{eqn.two_morphism_modification}
\begin{tikzpicture}[baseline=(F.base)]
\node(1)[]{$\cont P$};
\node(2)[below=1cm of 1]{$\cont{P'}$};
\node(4)[right=3cm of 2]{$\poCat{Poset}$};
\draw[a](2)to node(Pp)[la,below]{$P'$}(4);
\node(P)[la,yshift=-4pt] at (Pp|-1.north east){$P$};
\draw[a,rc](1) -- (1-|4) -- (4);
\draw[a,bend right](1)to node(F)[la,left]{$F$}(2);
\draw[a,bend left](1)to node(G)[la,right]{$G$}(2);
\cell[la,above,inner sep=5pt]{(F)}{(G)}{$\alpha$};
\draw[n,bend right,shorten <=6pt,shorten >=6pt](P)to node(Fs)[]{}(Pp);
\draw[n,bend left,shorten <=4pt, shorten >=1pt](P)to node(Gs){}(Pp);
\node(Fs)[la,left] at (Fs|-F) {$F\smash{^{\sharp}}$};
\node(Gs)[la,right] at (Gs|-G) {$G\smash{^{\sharp}}$};
\path (Fs.east) -- (Gs.west) node[midway,la]{$\leq$};
\end{tikzpicture}
\quad\parbox{3.5cm}{i.e., for each $\Gamma\in\cont P$,\\an inequality in $\poCat{Poset}\big(P(\Gamma),P'G(\Gamma)\big)$:}\quad
\begin{tikzpicture}[baseline=(Fs.base)]
\node(1)[]{$P(\Gamma)$};
\node(2)[below=0.75cm of 1]{$P'F(\Gamma)$};
\node(3)[right=0.75cm of 2]{$P'G(\Gamma)$};
\draw[a](1)to node(Fs)[la,left]{$F^{\sharp}_{\Gamma}$}(2);
\draw[a,rc] (1) -- (1-|3) -- (3);
\draw[a](2)to node(P')[la,below]{$P'\alpha_{\Gamma}$}(3);
\node[la,above](Gs) at (P'|-1) {$G^{\sharp}_{\Gamma}$};
\path (P') to node[la,sloped,pos=0.45]{$\leq$}(Gs);
\end{tikzpicture}
\end{equation}
\end{definition}
\begin{remark}
The reader would be justified in wondering whether these particular choices of $1$- and $2$-morphisms preserves all of the relevant structure at hand. We provide some discussion on this topic in \cref{sec.graphical_morphisms}, and for greater detail still direct the reader to the companion paper \cite{grl2} in which we show that these morphisms support the construction of a pseudo-reflection of regular categories into regular calculi.
\end{remark}
It is a straightforward if lengthy task to verify that morphisms of regular calculi obey strict composition laws, and moreover that $2$-morphisms may be composed along shared morphism boundaries as well as whiskered with morphisms in a functorial manner. With these facts, we are justified in making the following definition.
\begin{definition}[$\mathcal{R}\Cat{gCalc}$]\label{def.rgcalc_2cat}
The \define{$2$-category of regular calculi}, denoted $\mathcal{R}\Cat{gCalc}$, has as objects regular calculi, and as $1$- and $2$-morphisms the $1$- and $2$-morphisms of regular calculi of \cref{def.reg_calc_morphisms}.
\end{definition}
\begin{example}\label{ex.rgcalc_subcat}
Recall from \cref{ex.reg_calc_meet_sl} that a regular calculus $\rc P$ with a terminal context po-category $\cont P=1$ is a $\wedge$-semilattice. It is furthermore the case that the full subcategory of $\CCat{RgCalc}$ spanned by the regular calculi with terminal context po-category is terminal is isomorphic to the full subcategory of $\poCat{Poset}$ spanned by the $\wedge$-semilattices. Note in particular that the morphisms here are arbitrary monotone maps, and not just the $\wedge$-preserving ones.
\end{example}
\section{The bare regular calculus of a regular theory}\label{sec.regular_theories}
Now that we have defined regular calculi, we turn our attention to the first interesting class of examples, viz., the \emph{regular theories}. Indeed the main purpose of regular calculi is to capture syntactic regular logic of relations: as a data structure to support the understanding of regular logic in graphical terms. To illustrate this role, we now explore how regular calculi can be constructed from presentations of regular theories, and hence how they are a tool for describing regular theories in categorical terms.
To begin let us briefly recall the definition of a regular theory. For further detail we direct the curious reader to \cite[Definition D.1.1.1]{Johnstone:2002a}, for example.
\begin{definition}[Regular theory]
A \define{regular theory} $((\Sigma\text{-sort},\Sigma\text{-rel}),\mathbb T)$ is the data of
\begin{enumerate}
\item a set $\Sigma\text{-sort}$ of sorts
\item a set $\Sigma\text{-rel}$ of relation symbols, each of which is paired with a set of sorts
\item a set $\mathbb T$ of axioms
\end{enumerate}
From $\Sigma\text{-sort}$ we define \emph{contexts}: formal lists $\vec X$ of sorts $X_{i}\in\Sigma\text{-sort}$ where repetition is allowed. Using this, and the first two sets above, we inductively define the collection of formulae over formal variables associated to \emph{suitable} contexts, by using the relation symbols and the regular fragment of logic $(\const{true},\exists,=,\wedge)$. For instance, if $\varphi(\vec x)$ is a formula in variables $\langle x_{1} : X_1,\ x_2 : X_{2}\rangle$ in the suitable context $\vec X=\langle X_{1},\ X_{2}\rangle$, and $\psi(y)$ is a formula in variable $y:Y$, then $\exists_{y}[\varphi(\vec x)\wedge\psi(y)]$ is a formula in the context $\vec X$ once more. This construction depends on a notion of {suitability}: a context $\vec X$ is suitable for a formula $\varphi(\vec x)$ if the variables $x_{i}\in\vec x$ may be drawn appropriately from the sorts $X_{i}\in\vec X_{i}$. Finally, the set of axioms $\mathbb T$ is a collection of sequents of the form $\varphi\vdash\psi$ which we take in addition to those sequents provided by regular fragment of logic.
\end{definition}
\begin{remark}
Note that we have defined a regular theory to appear without ``function symbols''. In the companion \cite{grl2} we sketch a proof that this is an anodyne assumption inasmuch as the categories of models are concerned. The essence of this replacement is that function symbols may instead appear as relation symbols along with axioms asserting their functionality, and that models are sufficiently structured so as to prove that the internal relations thus formed are in fact the graphs of morphisms.
\end{remark}
Now that we have recalled the notion of regular theory we are prepared to give a construction assigning to each regular theory $(\Sigma,\mathbb T)$ a bare regular calculus $(\ccat{C}_{\mathbb T},F_{\mathbb T})$.
\begin{theorem}\label{con.rgcalc_rgtheory}
Given a regular theory $(\Sigma,\mathbb T)$, we may define a regular calculus
\[(\ccat C_{\mathbb T},F_{\mathbb T}) \quad\text{where}\quad \ccat C_{\mathbb T}\coloneqq\displaystyle{\bigsqcup_{\Sigma\text{-sort}}}\mathbb{W}\ ,\] and $F_{\mathbb T}$ sends each object, viewed as a context, to the poset of equivalence classes of regular formulas in that context quotiented by $\alpha$-equivalence.
\end{theorem}
\begin{proof}
Let us begin by more precisely stating the assignment of $F_{\mathbb T}$ on objects.
An object of $\ccat C_{\mathbb T}$ is a list $\langle(\sigma_{i},n_{i})\rangle$ of pairs of sorts $\sigma_{i}\in\Sigma\text{-sort}$ and natural numbers $n_{i}\in\mathbb N$. From this list, let us form the context given by concatenating each $\sigma_{i}$ repeated $n_{i}$ many times, and write this as $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}}$. If $n_{i}=0$ then $\sigma_{i}$ does not appear in the context. We thus set $F_{\mathbb T}(\left< (\sigma_{i},n_{i})\right>)$ to be the set of all formulae $\varphi$ over $\Sigma$ for which the context formed by $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}}$ is suitable. We quotient this set by $\alpha$-equivalence, and impose the order associated to provability in the sense of \cite{Johnstone:2002a}. Finally, we perform the poset quotient of what is \emph{a priori} a pre-order.
Next we must give the data of the functorial action $F_{\mathbb T}\colon\ccat C_{\mathbb T}\to \poCat{Poset}$. An arbitrary morphism in $\ccat C_{\mathbb T}$ is decomposable as a list of morphisms $\left< f_{i}\right>$, and may be further decomposed as a composite of morphisms which are the identity in all but one index. Thus, to give the functorial action it suffices to specify the action of $F_{\mathbb T}$ on a morphism of a single index, and ensure that functoriality holds. Moreover, observe that the formulae which are suitable for a given context are precisely the same as the formulae which are suitable for any permutation of that context, so without loss of generality we may decompose an arbitrary morphism in $\ccat C_{\mathbb T}$ as a finite sequence of composites which are the identity in all but their last index. We may yet further simplify this task by recalling \eqref{eqn.generating_wires}, that in each index, $\mathbb{W}$ is generated by four morphisms---$\eta$, $\epsilon$, $\mu$, $\delta$---and some relations governing their composites. Let us therefore address each one of these in turn.
For a fixed sort $\sigma$, let us define $F_{\mathbb T}(\id+\ldots+\id+\eta_{\sigma})$ to be the function which weakens a formula $\varphi(\vec x)$ in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}}$ to a formula in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}},\sigma$ by adding a free variable $y:\sigma$. Then let us set $F_{\mathbb T}(\id+\ldots+\id+\epsilon_{\sigma})$ to be the assignment of formulae $\varphi(\vec x,y)$ in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}},\sigma$ to formulae $\exists_{y}\varphi(\vec x,y)$ in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}}$. Next, we set $F_{\mathbb T}(\id+\ldots+\id+\epsilon_{\mu})$ as the mapping taking a formula $\varphi(\vec x, y,y')$ in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}},\sigma^{2}$ where $y,y':\sigma$ to the formula $\mu(\varphi)(\vec x,y)\coloneqq\varphi(\vec x,y,y)$ in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}},\sigma$. Finally, define $F_{\mathbb T}(\id+\ldots+\id+\delta_{\mu})$ to take a formula $\varphi(\vec x,y)$ in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}},\sigma$ to the formula $\delta(\varphi)(\vec x,y,y')\coloneqq [\varphi(\vec x,y)\wedge y=y']$ in context $\sigma_{1}^{n_{1}},\ldots,\sigma_{k}^{n_{k}},\sigma^{2}$. It is straightforward that all these assignments respect $\alpha$-equivalence, and respect the order of provability and the quotient thereby.
Checking that $F_{\mathbb T}$ is po-functorial amounts to checking that the relations exhibited in \eqref{eqn.equations_wires} hold. For instance, in $\mathbb{W}$ we have $\delta\mathbin{\fatsemi}\mu=\id$ and $\mu\mathbin{\fatsemi}\delta\leq\id_{\ord 2}$. Under $F_{\mathbb T}$ these become $\mu(\delta(\varphi))(\vec x,y)=\delta(\varphi)(\vec x,y,y)=[\varphi(\vec x,y)\wedge y=y]=\varphi(\vec x,y)$ and $\delta(\mu(\varphi))(\vec x,y,y')=[\mu(\varphi)(\vec x,y)\wedge y=y']=[\varphi(\vec x,y,y)\wedge y=y']\vdash \varphi(\vec x,y,y')$.
Finally it remains to give the ajax structure of $F_{\mathbb T}$. First let us define laxator $1\to F_{\mathbb T}(\left<\right>)$ to pick out the formula $\top$ in the empty context, and this is evidently a right adjoint to $!\colon F_{\mathbb T}(\left<\right>)\to 1$ for $\varphi\vdash \top$. Next let us define the laxator \[\boxplus\colon F_{\mathbb T}(\left<(\sigma_{i},n_{i})\right>)\times F_{\mathbb T}(\left<(\sigma'_{i},n'_{i})\right>)\to F_{\mathbb T}(\left<(\sigma_{i},n_{i})\right>+\left<(\sigma'_{i},n'_{i})\right>)\] to be the function sending pairs of formulae $\varphi(\vec x),\psi(\vec y)$ to the formula $\varphi(\vec x)\wedge \psi(\vec y)$ where by construction the contexts for $\vec x$ and $\vec y$ are disjoint in the codomain. In this sense the laxators behave as a form of ``exterior conjunction''. We claim that this map is a right adjoint in $\poCat{Poset}$. To see this, consider the opposed map $\lambda$ which sends a formula $\gamma(\vec x,\vec y)$ to the pair $\left<\exists_{\vec y}\gamma(\vec x,\vec y),\exists_{\vec x}\gamma(\vec x,\vec y) \right>$. First observe that \[\gamma(\vec x,\vec y)\vdash \exists_{\vec y}\gamma(\vec x,\vec y) \wedge \exists_{\vec x}\gamma(\vec x,\vec y) = \boxplus(\lambda (\gamma(\vec x,\vec y)))\ .\] In the other direction $\lambda(\boxplus(\left<\varphi(\vec x),\varphi(\vec y)\right>))$ has first component $\exists_{y}(\varphi(\vec x)\wedge\psi(\vec y))$ but this entails in particular $\varphi(\vec x)$, and similarly so for the second component. Thus $\boxplus$ is a right adjoint.
To conclude the construction it remains to check that $\top$ and $\boxplus$ obey the symmetric lax-monoidal coherence conditions and are $2$-natural. The former conditions reduce to the associativity, commutativity and unitality of conjunction, while the latter may be checked on each generator and relation.
\end{proof}
Let us turn to some examples of this construction, beginning with the simplest. In \cref{sec.graph_terms_regul} later we shall find pleasing graphical renditions of some of the formulae we discuss below.
\begin{example}[The theory of an inhabited sort]\label{ex.rgth_inhabited}
The theory of an inhabited sort is a regular theory. In this regular theory, $\Sigma\text{-sort}=\{S\}$ the unique sort, and $\Sigma\text{-rel}=\emptyset$ is empty. There is precisely one axiom in this theory, viz., $\mathbb T=\{\vdash\exists_{s}\top\}$---the sort $S$ is inhabited.
The regular calculus $(\ccat{C}_{\mathbb T},F_{\mathbb T})$ has $\ccat{C}_{\mathbb T}=\mathbb{W}$, and $F_{\mathbb T}(\ord n)$ is the set of regular formulae in $n$-many variables $s_{i}$. We wish to draw attention to the formula $[\exists_{s}\top]\in F_{\mathbb T}(\ord 0)$ which is asserted by our axiom. Note that this formula is obtained from $\top\in F_{\mathbb T}(\ord 0)$ by the functorial action $F_{\mathbb T}(\eta\mathbin{\fatsemi}\epsilon:\ord 0\to\ord0)(\top)$ using the unique morphism $\eta\mathbin{\fatsemi}\epsilon\ne\id_{\ord 0}\in\mathbb{W}(\ord0,\ord 0)$. That is, without this extra morphism $\ord 0\to\ord1\leftarrow\ord0$ of \eqref{eqn.extra_law} we would not be able to speak of inhabitedness of a sort in a bare regular calculus.
\end{example}
It is also of interest to consider a regular theory whose corresponding regular calculus has a genuine non-trivial inequality.
\begin{example}[The theory of a pre-order]\label{ex.rgth_preorder}
The theory of a pre-order is a regular theory. In this regular theory, $\Sigma\text{-sort}=\{X\}$ the unique sort, and $\Sigma\text{-rel}=\{P\}$ the pre-order relation symbol. The axiom set of this theory is small enough to be stated completely: $\mathbb T=\{x\vdash xPx,\, \exists_{y}[xPy\wedge yPz]\vdash xPz\}$, reflexivity and transitivity.
The regular calculus $(\ccat{C}_{\mathbb T},F_{\mathbb T})$ has $\ccat{C}_{\mathbb{T}}=\mathbb{W}$ and $F_{\mathbb T}(\ord n)$ is the set of regular formulae in $n$-many variables $x_{i}$ and the relation symbol $P$, ordered by provability on $\mathbb T$. In particular, observe that $F_{\mathbb T}(\ord 2)$ contains the sub-poset $\{[x=x'] \leq [xPx']\leq\top\}$ which displays two genuine inequalities imposed by the axioms of the theory.
\end{example}
Some more algebraic examples include the theories of monoids and categories.
\begin{example}[The theory of a monoid]\label{ex.rgth_monoid}
The theory of a monoid is a regular theory. Here $\Sigma\text{-sort}=\{M\}$, $\Sigma\text{-rel}=\{\star,e\}$ where the $\star$ is a ternary relation symbol and $e$ is a unary relation symbol, both on $M$. Our axioms include the usual axioms of a monoid, say $m\star e=m'\vdash m=m'$, but phrased relationally instead, $\star(m,e(m'),m'')\vdash m=m'$. We also include axioms to ensure that what are \emph{a priori} the relations of $\star$ and $e$ are in fact functions, for example, $\vdash\exists_{m}e(m)$ and $e(m)\wedge e(m')\vdash m=m'$.
The regular calculus $(\ccat{C}_{\mathbb T},F_{\mathbb T})$ associated to the theory $(\Sigma,\mathbb T)$ of monoids therefore has as its category of contexts $C_{\mathbb T}=\mathbb{W}$, and $F_{\mathbb T}(\ord n)$ is the collection of formulae in $n$-many variables of this theory. For example, $F(\ord 2)$ has the formula for expressing commutativity: $\exists_{m,m'}[\star(x,y,m)\wedge\star(y,x,m')\wedge m=m']\in F(\ord 2)$.
\end{example}
\begin{example}[The theory of a category]\label{ex.rgth_cat}
The theory of a category is also a regular theory. Here $\Sigma\text{-sort}=\{O,M\}$, $\Sigma\text{-rel}=\{\dom,\cod,\id,\comp\}$ where $\id,\dom,\cod$ are all binary relations on $O,M$, and $\comp$ is a binary relation on $M,M$.
The axioms of the regular theory of categories include once more all the usual axioms, say $\vdash \dom\id(x)=x$, but phrased relationally instead, $\id(x,f)\wedge\dom(f,y)\vdash x=y$. We also include axioms to ensure that $\dom,\cod,\id$ are functions, as well as axioms to ensure that $\comp$ is a function whenever it is defined.
The regular calculus $(\ccat{C}_{\mathbb T},F_{\mathbb T})$ associated to the theory $(\Sigma,\mathbb T)$ of categories therefore has as its category of contexts $C_{\mathbb T}=\mathbb{W}\sqcup\mathbb{W}$, and $F_{\mathbb T}([\ord n, \ord m])$ is the collection of formulae in $n$ object variables and $m$ morphism variables. For example, $F([\ord 0, \ord 2])$ contains the formula for isomorphisms (where $f,g:M$ are the isomorphism pair):
\begin{align*}
\exists_{h,k:M}\exists_{x,x',y,y':O}\big[&
[\dom(f,x)\wedge\cod(f,y)\wedge\dom(g,y')\wedge\dom(g,x')\wedge(x=x')\wedge(y=y')]\\
&\wedge [\comp(f,g,h)\wedge\comp(g,f,k)]\wedge[\id(x,h)\wedge\id(y,k)] \big]
\end{align*}
\end{example}
Once we pass to the graphical regular logic of \cref{chap.graphical_reglog} we will see that all three example formulas above---the inhabitedness of a sort, commutativity in monoids, and isomorphisms in categories---admit simple and intuitive graphical depictions.
To conclude this section, let us pause to consider the question of models. Classically, one defines what it means for a regular category to be a model of a given regular theory $(\Sigma,\mathbb T)$. One then also defines a ``syntactic regular category'' $\cat{C}^{\mathrm{reg}}_{\mathbb T}$ and the theoretical edifice is considered satisfactory when regular functors $\cat{C}^{\mathrm{reg}}_{\mathbb T}\to\cat{R}$ are equivalent to models in $\cat{R}$.
In order to replicate this process for regular calculi we will need to develop two main tools in the coming section. First we must find a structure that plays the role of regular categories in the story of models; for us these are the ``relational po-categories'' of \cref{chap.relational_pocats} next. We must also provide the construction analogous to the syntactic regular category; for us this is the ``syntactic po-category'' construction of \cref{sec.syn} which takes a regular calculus to a relational po-category. With these tools we will, in the companion \cite{grl2}, prove a satisfactory model theorem---\cref{thm.model_equiv} below---for regular calculi which subsumes that sketched for regular categories above.
\chapter{Relational po-categories}\label{chap.relational_pocats}
The goal of this section is at first to axiomatise those po-categories whose objects, morphisms, and $2$-morphisms arise as the objects, relations, and inclusions of relations in a regular category. We call these po-categories \emph{relational} and introduce this notion and the $2$-category $\CCat{RlPoCat}$ of such in \cref{sec.relational}. As we shall see there, relational po-categories admit tidy definition in terms of the notions of supply of \cref{sec.supply}, the po-category for wiring $\mathbb{W}$, and an extra piece of structure due to Freyd and Scedrov \cite{freyd1990categories}. Moreover, as we note in \cref{thm.carboni_equivalence}, the $2$-category of relational po-categories is appropriately equivalent to the $2$-category of regular categories. The fundamental impetus for so characterising regular categories is the construction of the $2$-functor $\poCat{Prd}$ of \cref{sec.prd}, from relational po-categories to regular calculi. In our main theorem, \cref{thm.main}, we show that this $2$-functor $\poCat{Prd}$ is $2$-dimensionally fully-faithful and is a $2$-dimensional right adjoint, and thereby affirm that regular calculi and graphical logic may be leveraged as tools for studying relational po-categories and so regular categories.
\section{The \texorpdfstring{$2$}{2}-category \texorpdfstring{$\mathcal{P}\Cat{rlPoCat}$}{PrlPoCat} of prerelational po-categories}\label{sec.prerel}
The first step towards characterising relations in regular categories is the notion of a prerelational category. Whereas relational po-categories play host to a full regular structure, prerelational po-categories support, roughly speaking, only the underlying cartesian structure.
\begin{definition}[Prerelational po-category] \label{def.prerel_pocat}
A symmetric monoidal po-category $(\ccat{R},I,\otimes)$ is \define{prerelational} if it supplies wiring $\mathbb{W}$, such that the induced supply of cocommutative comonoids (\cref{ex.wiring_comonoids}) is lax homomorphic.
\end{definition}
\begin{remark}\label{rem.lax_hom_prerel}
The lax homomorphicity condition requires that for every $f\colon c\to d$ in $\ccat{R}$, both $f\mathbin{\fatsemi}\epsilon_d\leq\epsilon_c$ and $f\mathbin{\fatsemi}\delta_d\leq\delta_c\mathbin{\fatsemi} (f\otimes f)$. In graphical notation:
\begin{equation}\label{eqn.lax_hom_prerel}
\begin{tikzpicture}[inner WD,baseline=(1.-30)]
\node(1){};
\node(f)[syntax, oshellr,right=1.25 of 1] (f) {$f$};
\node(2)[right=1.25 of f,link]{};
\draw(1) to node[above]{$c$} (f);
\draw(f) to node[above]{$d$} (2);
\end{tikzpicture}\
\leq\
\begin{tikzpicture}[inner WD,baseline=(1.-30)]
\node(1)[]{};
\node(2)[right=1.25 of 1,link]{};
\draw(1) to node[above]{$d$} (2);
\end{tikzpicture}\hspace{0.75cm}\text{and}\hspace{0.75cm}
\begin{tikzpicture}[inner WD,baseline=(1.-30)]
\node(1){};
\node(f)[syntax, oshellr,right=1.25 of 1] (f) {$f$};
\node(2)[right=1 of f,link]{};
\coordinate (d1) at ($(2)+(1,1.3)$);
\coordinate (d2) at ($(2)+(1,-1.3)$);
\draw(1) to node[above]{$c$} (f);
\draw(f) to node[above]{$d$} (2);
\draw(2) to [out=60, in=180] (d1) -- ++(1,0);
\draw(2) to [out=-60, in=180] (d2) -- ++(1,0);
\end{tikzpicture}\ \leq \
\begin{tikzpicture}[inner WD,baseline=(1.-30)]
\node(1){};
\node(2)[right=1 of 1,link]{};
\coordinate (d1) at ($(2)+(1,1.3)$);
\coordinate (d2) at ($(2)+(1,-1.3)$);
\draw(1) to node[above]{$c$} (2);
\draw(2) to [out=60, in=180] (d1) to node[syntax,oshellr,pos=0.3]{$f$} ++(3,0) node[above]{$d$};
\draw(2) to [out=-60, in=180] (d2) to node[syntax,oshellr,pos=0.3]{$f$} ++(3,0) node[below]{$d$};
\end{tikzpicture}
\end{equation}
\end{remark}
As follows from the mate calculus, or indeed through pleasing graphical manipulations, we may re-characterise prerelational po-categories in terms of the induced supply of commutative monoids instead.
\begin{lemma} \label{lemma.lax_comons_oplax_mons}
Suppose $\ccat{C}$ supplies wiring $\mathbb{W}$. Then the induced supply of comonoids is lax homomorphic iff the supply of monoids is oplax homomorphic.
\end{lemma}
\begin{proof}
Note that given $l$ left adjoint to $r$, we always have that $a \le (l \mathbin{\fatsemi} b)$ iff $(r \mathbin{\fatsemi} a) \le b$ and that $(a \mathbin{\fatsemi} l)\le b$ iff $a \le (b \mathbin{\fatsemi} r)$. The result now follows from \cref{prop.left_adj_in_ww}.
\end{proof}
\begin{example}\label{ex.ww_prerelational}
The po-category $\mathbb{W}$ is prerelational. It supplies $\mathbb{W}$ by \cref{prop.p_supplies_itself}. To see that the supply of comonoids is lax homomorphic, take a cospan $m\To{f}n\From{g}p$. In the case of the diagonal, the necessary composites are computed as pushouts (e.g.\ $n +_m n$), and the inequality from \eqref{eqn.lax_hom_prerel} is indicated by the dotted arrow, in the diagram
\[
\begin{tikzcd}[row sep=3pt]
&&
p\ar[dl, "g"']\ar[dr, equal]\\&
n\ar[r, equal]&
n&
p\ar[l, "g"]\\
m\ar[ur, "f"]\ar[dr, equal]&&&&
p+p\ar[ul, "\copair{p,p}"']\ar[dl, "g+g"]\\&
m\ar[r]&
n+_m n\ar[uu, dashed]&
n+n\ar[l]\\&&
m+m\ar[ul, "\copair{m,m}"]\ar[ur, "f+f"']
\end{tikzcd}
\]
The case of co-units is similar.
\end{example}
\begin{example}
The po-category of finite sets and corelations is also prerelational; see \cref{ex.corel}. More generally, given any category $\cat{C}$ with finite limits, the poset reflection of $\poCat{Span}(\cat{C})$ is prerelational. As a proof sketch, by Fox's theorem \cite{fox1976coalgebras}, any category $\cat C$ with finite limits homomorphically supplies $\Cat{FinSet}^{\tn{op}}$ and with some work this may transferred to a suitable supply of $\mathbb{W}$ in $\poCat{Span}(\cat C)$.
\end{example}
\begin{example}
If $\cat{C}$ is regular, then $\poCat{Rel}(\cat{C})$ is prerelational, see \cite{fong2019regular}.
\end{example}
A class of examples for which we will have use later is given by the following lemma.
\begin{lemma}\label{lemma.irritating_prerelational_coprod}
Let $J$ be a set and $\ccat{R}$ a prerelational po-category supplying $\mathbb{W}$. Then with induced supply of $\mathbb{W}$ in $\bigsqcup_{J}\ccat{R}$ of \cref{lemma.irritating_supply_coprod}, the symmetric monoidal po-category $\bigsqcup_{J}\ccat{R}$ is prerelational.
\end{lemma}
\begin{proof}
Direct computation.
\end{proof}
By {\cite[Corollary 6.2]{fong2019regular}}, if $f\colon r\to s$ is a left adjoint in a prerelational po-category $(\ccat{R},I,\otimes)$ then its right adjoint is its transpose $f^{\dagger}\colon r\to s$. With this fact we are justified in introducing the following notation.
\begin{notation}[Left adjoints]
We shall denote \funcrinl{$f$} by \ladjinl{$f$} when $f\colon r\to s$ is known to be a left adjoint. In keeping with \cref{notation.transpose}, we shall denote its right adjoint and transpose $f^{\dagger}\colon s\to r$ as \radjinl{$f$}.
\end{notation}
\begin{remark}
Left adjoints in a prerelational po-category are profitably understood as the true morphisms of the cartesian $1$-category whose structure has been elaborated. As such, we should expect all analogues of cartesian results for left adjoints, such as: two left adjoints $f,g\colon r\to s\otimes t$ are equal iff their ``projections'' under $\epsilon_{s}$ and $\epsilon_{t}$ agree; left adjoints are monoid homomorphisms strictly, for to them the structure is cartesian; a ``natural transformation'' between ``cartesian'' po-functors of prerelational po-categories is automatically ``monoidal'' if its components are left adjoints. Indeed we make the first and last results precise and prove them as \cref{lemma.pres_all_struct_pre,lemma.ladjlaxnat_m} later, while the middle appears as \cite[Corollary 6.2]{fong2019regular}. This serves as additional motivation to distinctly signify left adjoints graphically.
\end{remark}
\begin{definition}
The \define{$2$-category $\CCat{PrlPoCat}$ of prerelational po-categories} has as objects the prerelational po-categories, as morphisms the strong symmetric monoidal po-functors, and as $2$-morphisms the left adjoint oplax-natural transformations $\alpha\colon F\Rightarrow G$; see \cref{def.ladj_lax_nt}.
\end{definition}
Although it is not immediate, the $1$ and $2$-morphisms we have chosen above are indeed appropriate for respecting \emph{all} the structure present in prerelational po-categories. To show this we appeal to a few results of \cite{fong2019regular}.
\begin{lemma}[{\cite[Proposition 6.22]{fong2019regular}}]\label{lemma.pres_all_struct_pre}
If $\ccat{R}$ and $\ccat{R}'$ are prerelational po-categories and $F\colon\ccat{R}\to\ccat{R}'$ is any strong symmetric monoidal po-functor, then $F$ automatically preserves the supply of $\mathbb{W}$.
\end{lemma}
\begin{lemma}\label{lemma.prerelational_epsilon_joint_mono}
If $\ccat{R}$ is prerelational, and $f,g\colon r\to s\otimes s'$ are two left adjoints, then $f=g$ iff
\begin{center}
\begin{tikzpicture}[inner WD,baseline=(f.base)]
\node(f)[syntax,funcr]{$f$};
\draw (f.west) -- +(-1,0);
\draw ($(f.east)+(0,0.3)$) -- +(1,0) node[link]{};
\draw ($(f.east)-(0,0.3)$) -- +(1,0);
\node(g)[right=2cm of f,syntax,funcr]{$g$};
\draw (g.west) -- +(-1,0);
\draw ($(g.east)+(0,0.3)$) -- +(1,0) node[link]{};
\draw ($(g.east)-(0,0.3)$) -- +(1,0);
\node at ($(f)!0.5!(g)$) {$=$};
\end{tikzpicture}\hspace{1cm} and \hspace{1cm}
\begin{tikzpicture}[inner WD,baseline=(f.base)]
\node(f)[syntax,funcr]{$f$};
\draw (f.west) -- +(-1,0);
\draw ($(f.east)+(0,0.3)$) -- +(1,0);
\draw ($(f.east)-(0,0.3)$) -- +(1,0) node[link]{};
\node(g)[right=2cm of f,syntax,funcr]{$g$};
\draw (g.west) -- +(-1,0);
\draw ($(g.east)+(0,0.3)$) -- +(1,0);
\draw ($(g.east)-(0,0.3)$) -- +(1,0) node[link]{};
\node at ($(f)!0.5!(g)$) {$=$};
\end{tikzpicture}\quad .
\end{center}
\end{lemma}
\begin{proof}
Certainly $f=g$ implies the given condition. For the converse recall that by \cite[Corollary 6.2]{fong2019regular} left adjoints are comonoid homomorphisms -- the inequalities of \eqref{eqn.lax_hom_prerel} are equalities --, so that we may argue as follows.
\[
\begin{tikzpicture}[inner WD]
\node(f)[syntax,funcr]{$f$};
\draw (f.west) -- +(-1,0);
\draw ($(f.east)+(0,0.3)$) -- +(1,0);
\draw ($(f.east)-(0,0.3)$) -- +(1,0);
\node(f2)[syntax,funcr,right=1.5cm of f]{$f$};
\draw (f2.west) -- +(-1,0);
\draw($(f2.east)+(0,0.3)$) to[out=45,in=180] +(1,1) -- +(1,0) node[link](upper){};
\draw($(f2.east)-(0,0.3)$) to[out=-45,in=180] +(1,-1) -- +(1,0) node[link](lower){};
\draw(upper) to[out=45,in=180] +(1,1) -- +(3,0);
\draw(upper) to[out=-45,in=180] ++(1,-1) -- ++(0.5,0) to[out=0,in=180] ++(2,-1) -- ++(0.5,0) node[link]{};
\draw(lower) to[out=45,in=180] ++(1,1) -- ++(0.5,0) to[out=0,in=180] ++(2,1) -- ++(0.5,0) node[link]{};
\draw(lower) to[out=-45,in=180] +(1,-1) -- +(3,0);
\node[font=\normalsize] at ($(f)!0.5!(f2)$) {$=$};
\node(dot)[link,right=2.75cm of f2]{};
\draw (dot.west) -- +(-1,0);
\draw(dot) to[out=45,in=180] +(1,1) -- +(1,0) node(upf)[syntax,funcr]{$f$};
\draw(dot) to[out=-45,in=180] +(1,-1) -- +(1,0) node(dnf)[syntax,funcr]{$f$};
\draw($(upf.east)+(0,0.3)$) -- +(1,0);
\draw($(upf.east)-(0,0.3)$) -- +(1,0) node[link]{};
\draw($(dnf.east)+(0,0.3)$) -- +(1,0) node[link]{};
\draw($(dnf.east)-(0,0.3)$) -- +(1,0);
\node[font=\normalsize] at ($(f2)!0.75!(dot)$) {$=$};
\node(dot2)[link,right=2cm of dot]{};
\draw (dot2.west) -- +(-1,0);
\draw(dot2) to[out=45,in=180] +(1,1) -- +(1,0) node(upg)[syntax,funcr]{$g$};
\draw(dot2) to[out=-45,in=180] +(1,-1) -- +(1,0) node(dng)[syntax,funcr]{$g$};
\draw($(upg.east)+(0,0.3)$) -- +(1,0);
\draw($(upg.east)-(0,0.3)$) -- +(1,0) node[link]{};
\draw($(dng.east)+(0,0.3)$) -- +(1,0) node[link]{};
\draw($(dng.east)-(0,0.3)$) -- +(1,0);
\node[font=\normalsize] at ($(dot)!0.7!(dot2)$) {$=$};
\node(g)[funcr,syntax,right=2.25cm of dot2]{$g$};
\draw (g.west) -- +(-1,0);
\draw ($(g.east)+(0,0.3)$) -- +(1,0);
\draw ($(g.east)-(0,0.3)$) -- +(1,0);
\node[font=\normalsize] at ($(dot2)!0.6!(g)$) {$=$};
\end{tikzpicture}
\qedhere
\]
\end{proof}
\begin{lemma}\label{lemma.ladjlaxnat_m}
Let $(F,\varphi),(G,\psi)\colon\ccat{R}\to\ccat{R}'$ be strong symmetric monoidal po-functors between prerelational po-categories, and let $\alpha\colon F\Rightarrow G$ be a left adjoint oplax-natural transformation. Then $\alpha$ is a monoidal oplax-natural transformation (\cref{def.symm_mon_po}).
\end{lemma}
\begin{proof}
We must show that the diagrams
\begin{center}
\begin{tikzpicture}[node distance=1.25cm,baseline=(phi.base)]
\node(1)[]{$I'$};
\node(2)[below of= 1]{$FI$};
\node(3)[right of= 2]{$GI$};
\draw[a](1)to node(phi)[la,left]{$\varphi_{I}$}(2);
\draw[a](1)to node[la,auto,squeeze]{$\psi_{I}$}(3);
\draw[a](2)to node[la,below]{$\alpha_{I}$}(3);
\end{tikzpicture}
\hspace{1cm} and \hspace{1cm}
\begin{tikzpicture}[node distance=1.25cm,baseline=(phi.base)]
\node(1)[]{$Fr\otimes'Fs$};
\node(2)[right=1.5cm of 1]{$Gr\otimes'Gs$};
\node(3)[below of=1]{$F(r\otimes s)$};
\node(4)[below of=2]{$G(r\otimes s)$};
\draw[a](1)to node[la,above]{$\alpha_{r}\otimes\alpha_{s}$}(2);
\draw[a](3)to node[la,below]{$\alpha_{r\otimes s}$}(4);
\draw[a](1)to node(phi)[la,left]{$\varphi_{r,s}$}(3);
\draw[a](2)to node[la,right]{$\psi_{r,s}$}(4);
\end{tikzpicture}
\end{center}
are strictly commutative for all objects $r,s\in\ob\ccat{R}$. Observe that all the strongators are isomorphisms and so are, in particular, left adjoints. Thus these are diagrams of left adjoints and by \cite[Proposition 6.5]{fong2019regular} it suffices to show in the first case that $\varphi_{I}\mathbin{\fatsemi} \alpha_{I}\leq \psi_{I}$, equivalently $\alpha_{I}\mathbin{\fatsemi}\psi_{I}^{\text{-}1}\leq\varphi_{I}^{\text{-}1}$. In the second case we will use \cref{lemma.prerelational_epsilon_joint_mono} to show $\varphi_{r,s}\mathbin{\fatsemi}\alpha_{r\otimes s}\mathbin{\fatsemi} \psi^{\text{-}1}=\alpha_{r}\otimes \alpha_{s}$.
In the first case, by \cref{lemma.pres_all_struct_pre} above, we have the equalities $\varphi_{I}^{\text{-}1}=\epsilon'_{FI}\colon FI\to I$ and $\psi_{I}^{\text{-}1}=\epsilon'_{GI}\colon GI\to I$ as both $(F,\varphi)$ and $(G,\psi)$ preserve the supply of $\mathbb{W}$. As $\ccat{R}'$ is prerelational, $\alpha_{I}\colon FI\to GI$ must be lax homomorphic and so we have \[\alpha_{I}\mathbin{\fatsemi}\psi_{I}^{\text{-}1}=\alpha_{I}\mathbin{\fatsemi}\epsilon'_{GI}\leq\epsilon'_{FI}=\varphi_{I}^{\text{-}1}\ ,\] as desired.
To show the equality of the left adjoints $\varphi_{r,s}\mathbin{\fatsemi}\alpha_{r\otimes s}\mathbin{\fatsemi} \psi^{\text{-}1}$ and $\alpha_{r}\otimes \alpha_{s}$ let us begin by writing $\pi'_{Gr}\coloneqq(\epsilon'_{Gr}\otimes\id_{Gs})\mathbin{\fatsemi} \lambda'_{Gs}\colon Gr\otimes'Gs\to Gs$ and likewise for $\pi'_{Gs}$,$\pi'_{Fr}$, and $\pi'_{Fs}$. \Cref{lemma.prerelational_epsilon_joint_mono} shows that it is enough to prove that
\[
\varphi_{r,s}\mathbin{\fatsemi}\alpha_{r\otimes s}\mathbin{\fatsemi} \psi^{\text{-}1}\mathbin{\fatsemi}\pi'_{Gr} =\alpha_{r}\otimes \alpha_{s}\mathbin{\fatsemi} \pi'_{Gr}\quad\text{and}\quad
\varphi_{r,s}\mathbin{\fatsemi}\alpha_{r\otimes s}\mathbin{\fatsemi} \psi^{\text{-}1}\mathbin{\fatsemi}\pi'_{Gs} =\alpha_{r}\otimes \alpha_{s}\mathbin{\fatsemi} \pi'_{Gs}\ .
\]
The proofs that these equalities hold involve large diagrams which make use of the commutativity of the triangle established above, \cite[Proposition 6.5]{fong2019regular} to see that the oplax-naturality squares for $\alpha$ on left adjoints are actually strict, and \Cref{lemma.pres_all_struct_pre} for equalities like $G(\epsilon_{r})\mathbin{\fatsemi}\psi_{I}^{\text{-}1}=\epsilon'_{Gr}$. These diagrams are not especially illuminating and so we have not included them here. Nevertheless, with the aforementioned techniques, it is possible to show that both terms of the above-left claimed equality reduce to $\pi'_{Fs}\mathbin{\fatsemi}\alpha_{s}$, while for the above-right claimed equality the reduced form is $\pi'_{Fr}\mathbin{\fatsemi}\alpha_{r}$.
\end{proof}
\section{The \texorpdfstring{$2$}{2}-category \texorpdfstring{$\mathcal{R}\Cat{lPoCat}$}{RlPoCat} of relational po-categories}\label{sec.relational}
To make a prerelational po-category relational, we need tabulations. The following definition is due to Freyd and Scedrov \cite{freyd1990categories}.
\begin{definition}[Tabulation] \label{def.tabulation}
Suppose $\ccat{C}$ supplies $\mathbb{W}$ and let $f\colon r \to s$ be a morphism in $\ccat{C}$. A \define{tabulation $(f_R,f_L)$ of $f$} is a factorisation $r\xrightarrow{f_R}\tab f\xrightarrow{f_L}s$ of $f$ where
\begin{enumerate}[label=(\roman*)]
\item $f_R\colon r\to \tab f$ is a right adjoint in $\ccat{C}$;
\item $f_L\colon \tab f\to s$ is a left adjoint in $\ccat{C}$; and
\item $\hat{f}\mathbin{\fatsemi}\hat{f}^{\dagger}=\id_{\tab f}$, where $\hat{f}\coloneqq\delta_{\tab f}\mathbin{\fatsemi}(f_L\otimes f_R^{\dagger})$; in pictures
\begin{equation}\label{eqn.tabulation}
\begin{tikzpicture}[baseline=(P1)]
\node (P1) {
\begin{tikzpicture}[inner WD, baseline=(dot1)]
\node[funcr, syntax, minimum size=3ex] (fl) {$f_L$};
\node[funcl, syntax, right=of fl, minimum size=3ex] (fl*) {$f_L$};
\node[funcr, syntax, below=of fl, minimum size=3ex] (fr!) {$f_R^{\dagger}$};
\node[funcl, syntax, minimum size=3ex] at (fl*|-fr!) (fr) {$f_R^{\dagger}$};
\draw (fl.east) -- (fl*.west);
\draw (fr!.east) -- (fr.west);
\node[link] at ($(fl.west)!.5!(fr!.west)+(-1,0)$) (dot1) {};
\node[link] at ($(fl*.east)!.5!(fr.east)+(1,0)$) (dot2) {};
\draw (fl.west) to[out=180, in=60] (dot1);
\draw (fr!.west) to[out=180, in=-60] (dot1);
\draw (dot1) to node[above, font=\scriptsize] {$\tab f$} +(-2,0);
\draw (fl*.east) to[out=0, in=120] (dot2);
\draw (fr.east) to[out=0, in=-120] (dot2);
\draw (dot2) to node[above, font=\scriptsize] {$\tab f$} +(2,0);
\end{tikzpicture}
};
\node at ($(P1.east)+(.5,0)$) {$=$};
\draw ($(P1.east)+(1,0)$) to node[above, font=\scriptsize] {$\tab f$} +(1,0);
\end{tikzpicture}
\end{equation}
\end{enumerate}
\end{definition}
\begin{definition}[Relational po-category]\label{def.rl2cat}
A \define{relational po-category} is a prerelational po-category $\ccat{R}$ in which additionally every morphism has a a chosen tabulation.
The \define{$2$-category $\CCat{RlPoCat}$ of relational po-categories} is the $2$-full sub-$2$-category $\CCat{PrlPoCat}$ whose objects are relational po-categories. That is, it has as objects relational po-categories, as morphisms strong symmetric monoidal po-functors, and as $2$-morphisms the left adjoint oplax-natural transformations.
\end{definition}
By \cref{lemma.pres_all_struct_pre} we thus know that strong symmetric monoidal po-functors between relational po-categories preserve the supply, but in fact more is true.
\begin{lemma}[{\cite[Proposition 6.22]{fong2019regular}}]\label{lemma.pres_all_struct}
If $\ccat{R}$ and $\ccat{R}'$ are relational po-categories and $F\colon\ccat{R}\to\ccat{R}'$ is any strong symmetric monoidal po-functor, then $F$ automatically preserves the supply of $\mathbb{W}$ and tabulations.
\end{lemma}
\begin{remark}
A prerelational po-category is exactly what Carboni and Walters called a \emph{`bicategory of relations'} (quotation marks are an explicit part of the their terminology), and a relational po-category is exactly what they called a \emph{functionally complete `bicategory of relations'}. There are a few, ultimately immaterial differences in the definition; see \cite[Section 8.1]{fong2019regular} for details.
\end{remark}
\begin{example}\label{ex.ww_not_relational}
As seen in \cref{ex.ww_prerelational}, the po-category $\mathbb{W}$ is prerelational. However, it is not relational: the cospan $\ord 0\to \ord 1\leftarrow \ord 0$ in $\mathbb{W}$ does not have a tabulation. On the other hand, the po-category of finite sets and corelations is relational.
\end{example}
Every regular category has a po-category of relations, and it is relational; this mapping extends to a $2$-functor $\poCat{Rel}\colon\CCat{RgCat}\to\CCat{RlPoCat}$ sending a regular category $\ccat{R}$ to the po-category with the same objects and with hom-posets given by
\begin{equation}\label{eqn.def_rel}
\poCat{Rel}(\ccat{R})(r_1,r_2)\coloneqq\Cat{Sub}_{\ccat{R}}(r_1\times r_2).
\end{equation}
In the other direction, the category of left adjoints in any relational po-category is regular; this also extends to a $2$-functor $\Funr{LAdj}\colon\CCat{RlPoCat}\to\CCat{RgCat}$. It turns out these functors form an equivalence of $2$-categories. Indeed, this is the Carboni-Walters idea, although they did not explicitly give the full $2$-categorical account. However, this equivalence was proven as the main theorem in \cite[Theorem 7.3]{fong2019regular}.
\begin{theorem}\label{thm.carboni_equivalence}
The $2$-functors $\poCat{Rel}\colon\CCat{RgCat}\leftrightarrows\CCat{RlPoCat}\ \,\llap{\normalfont:}\Funr{LAdj}$ form an equivalence of $2$-categories. Their underlying 1-functors moreover form an equivalence between the underlying 1-categories.
\end{theorem}
\begin{example}
Under the equivalence from \cref{thm.carboni_equivalence}, the relational po-category of finite sets and equivalence relations corresponds to the regular category $\Cat{FinSet}^{\tn{op}}$. This is \emph{almost} the free regular category on one object, but not quite: it is the free regular category in which every object $x$ is inhabited (the unique map $x\to 1$ is a regular epi).
\end{example}
\begin{remark}
In fact, more is true of $\Funr{LAdj}$. As a consequence of the main result \cref{thm.main} and the work of companion paper \cite{grl2}, in \cref{cor.ladj_birep} we record a result of the companion: $\Funr{LAdj}$ is ``bi-represented'' by the relational po-category, $\poCat{Syn}\poCat{Prd}\mathbb{W}$, the syntactic category of the regular calculus of predicates in $\mathbb{W}$. That is, understanding the regular category of left adjoints of a relational po-category is equivalent to mapping into it from $\poCat{Syn}\poCat{Prd}\mathbb{W}$.
\end{remark}
\section{The predicates \texorpdfstring{$2$}{2}-functor \texorpdfstring{$\mathbb{P}\Cat{rd}\colon\mathcal{P}\Cat{rlPoCat}\to\mathcal{R}\Cat{gCalc}$}{Prd}}\label{sec.prd}
We are now ready to address the second large source of regular calculi: in this section we will construct a $2$-functorial assignment $\CCat{RgCat}\to\CCat{RgCalc}$ of regular categories to regular calculi.
To begin, let us recall that \cref{thm.carboni_equivalence} gives a $2$-equivalence between the $2$-category of regular categories $\mathcal{R}\Cat{gCat}$ and that of relational po-categories $\mathcal{R}\Cat{lPoCat}$ (\cref{def.rl2cat}). With this in mind, our goal may be achieved by instead constructing a $2$-functor $\poCat{Prd}\colon\CCat{RlPoCat}\to\mathcal{R}\Cat{gCalc}$. However, it turns out that our construction of this to-be $2$-functor does not depend on the presence of tabulators, and so we further factor our promised assignment as follows:
\[
\mathcal{R}\Cat{gCat}\overset{\ref{thm.carboni_equivalence}}\eqv\CCat{RlPoCat}\rightarrowtail\CCat{PrlPoCat}\To{\poCat{Prd}}\CCat{RgCalc}\ .
\]
With that let us turn our attention to the construction of the to-be $2$-functor $\poCat{Prd}$.
Our first step will be to associate a right ajax po-functor to every prerelational po-category $\ccat{R}$. In the coming proposition we will show that $\ccat{R}(I,-)$ is right ajax. While we will have no use for the following, it is more generally true that $\ccat{R}(r,-)$ is right ajax for all objects $r\in\ob\ccat{R}$.
\begin{proposition}\label{prop.prerel_ajax}
Let $(\ccat{R},\otimes,I)$ be a prerelational po-category. The representable po-functor $\ccat{R}(I,-)\colon\ccat{R}\to\poCat{Poset}$ has a lax monoidal structure whose laxators are right adjoints, and thus supports the structure of a regular calculus on $\ccat{R}$.
\end{proposition}
\begin{proof}
For every $r\in\ccat{R}$, we have a poset $\ccat{R}(I,r)$, and this is clearly functorial in $r$. Consider the proposed adjunctions below:
\begin{equation}\label{eqn.prd_right_ajax}
\adjr{1}{\id_{I}}{!}{\ccat{R}(I,I)}
\qquad\text{and}\qquad
\adjr{\ccat{R}(I,r)\times \ccat{R}(I,r')}{\otimes\mathbin{\fatsemi} (\delta_{I})^{*}}{\pi}{\ccat{R}(I,r\otimes r')}.
\end{equation}
The poset map labelled $\id_{I}$ sends 1 to the identity; that labelled $!$ is uniquely determined; that labelled $\otimes\mathbin{\fatsemi}(\delta_{I})^{*}$ sends $\pair{f,f'}$ to $f\otimes f'$ pre-composed with $\delta_{I}\colon I\to I\otimes I$; and that labelled $\pi$ sends $h\colon I\to r\otimes r'$ to the pair $\pi(h)\coloneqq\pair{h\mathbin{\fatsemi}(r\otimes\epsilon_{r'})\mathbin{\fatsemi}\rho_{r}, h\mathbin{\fatsemi}(\epsilon_r\otimes r')\mathbin{\fatsemi}\lambda_{r'}}$. The coming contents of the proof rely on the lax comonoid homomorphism condition in \cref{def.prerel_pocat}.
To see that the first pair of maps form an adjunction, take any $s\in\ccat{R}(I,I)$. By definition of supply \eqref{eqn.supply_commute_tensors} we have $\epsilon_I=\id_I$, so lax comonoidality implies $s=s\mathbin{\fatsemi}\epsilon_I\leq\epsilon_I=\id_I$, as desired. Let us pause here to note that the same supply conditions imply the equality of morphisms $\delta_{I}={\lambda_{I}}^{\text{-}1}={\rho_{I}}^{\text{-}1}\colon I\to I\otimes I$.
To see that the second pair of maps form an adjunction, take $\pair{f,f'}\in\ccat{R}(I,r)\times\ccat{R}(I,r')$ and $h\in\ccat{R}(I,r\otimes r')$ and consider the below arguments, where we have used the lax comonoidality properties of \eqref{eqn.lax_hom_prerel} and the facts $\epsilon_{I}=\id_{I}$ and $\delta_{I}={\lambda_{I}}^{\text{-}1}$ from above.
\[(\otimes\mathbin{\fatsemi}(\delta_{I})^{*}\mathbin{\fatsemi}\pi)\left(\left<
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$f\vphantom{'}$};
\draw(1.east) -- +(1,0);
\end{tikzpicture},
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$f'$};
\draw(1.east) -- +(1,0);
\end{tikzpicture}\right>\right)=
\left<
\begin{tikzpicture}[inner WD,baseline=(2.north)]
\node(1)[syntax,oshellr]{$f\vphantom{'}$};
\node(2)[syntax,oshellr,below= 1 of 1]{$f'$};
\node(dot)[link,right= 1 of 2]{};
\draw(1.east) -- +(1,0);
\draw(2.east) -- (dot);
\end{tikzpicture}\ \ ,\ \
\begin{tikzpicture}[inner WD,baseline=(2.north)]
\node(1)[syntax,oshellr]{$f\vphantom{'}$};
\node(2)[syntax,oshellr,below= 1 of 1]{$f'$};
\node(dot)[link,right= 1 of 1]{};
\draw(2.east) -- +(1,0);
\draw(1.east) -- (dot);
\end{tikzpicture}\right>\leq
\left<
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$f\vphantom{'}$};
\draw(1.east) -- +(1,0);
\end{tikzpicture},
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$f'$};
\draw(1.east) -- +(1,0);
\end{tikzpicture}\right>
\]
\[
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$h$};
\draw(1) to[out=45,in=180] +(1,1) -- +(1,0);
\draw(1) to[out=-45,in=180] +(1,-1) -- +(1,0);
\end{tikzpicture}\quad=\quad
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$h$};
\draw(1) to[out=45,in=180] +(2,2) -- +(1,0) node[link](upper){};
\draw(1) to[out=-45,in=180] +(2,-2) -- +(1,0) node[link](lower){};
\draw(upper) to[out=45,in=180] +(1,1) -- +(1,0);
\draw(upper) to[out=-45,in=180] +(1,-1) -- +(1,0) node[link]{};
\draw(lower) to[out=45,in=180] +(1,1) -- +(1,0) node[link]{};
\draw(lower) to[out=-45,in=180] +(1,-1) -- +(1,0);
\end{tikzpicture}\quad=\quad
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$h$};
\draw(1) to[out=45,in=180] +(2,2) -- +(1,0) node[link](upper){};
\draw(1) to[out=-45,in=180] +(2,-2) -- +(1,0) node[link](lower){};
\draw(upper) to[out=45,in=180] +(1,1) -- +(3,0);
\draw(upper) to[out=-45,in=180] ++(1,-1) -- ++(0.5,0) to[out=0,in=180] ++(2,-2) -- ++(0.5,0) node[link]{};
\draw(lower) to[out=45,in=180] ++(1,1) -- ++(0.5,0) to[out=0,in=180] ++(2,2) -- ++(0.5,0) node[link]{};
\draw(lower) to[out=-45,in=180] +(1,-1) -- +(3,0);
\end{tikzpicture}\quad\leq\quad
\begin{tikzpicture}[inner WD,baseline=(c)]
\node(1)[oshellr,syntax]{$h$};
\draw(1) to[out=45,in=180] +(1,1) -- +(1,0);
\draw(1) to[out=-45,in=180] +(1,-1) -- +(1,0) node[link]{};
\node(2)[oshellr,syntax,below=1.5 of 1]{$h$};
\draw(2) to[out=45,in=180] +(1,1) -- +(1,0) node[link]{};
\draw(2) to[out=-45,in=180] +(1,-1) -- +(1,0);
\coordinate (c) at ($(1)!0.5!(2)-(0,0.25)$);
\end{tikzpicture}\quad=(\pi\mathbin{\fatsemi}\otimes\mathbin{\fatsemi}(\delta_{I})^{*})\
\begin{tikzpicture}[inner WD,baseline=(1.-20)]
\node(1)[oshellr,syntax]{$h$};
\draw(1) to[out=45,in=180] +(1,1) -- +(1,0);
\draw(1) to[out=-45,in=180] +(1,-1) -- +(1,0);
\end{tikzpicture}
\]
It remains to establish that $(\ccat{R}(I,-),\id_{I},\otimes\mathbin{\fatsemi}(\delta_{I})^{*})$ assembles into a lax monoidal functor. The conditions on $!$ are all automatic by the terminality of $1\in\poCat{Poset}$, and so we have reduced our claim to the assertion that the following diagram is commutative.
\[\begin{tikzpicture}
\node(1)[]{$\ccat{R}(I,(r\otimes r')\otimes r'')$};
\node(2)[right=3 of 1]{$\ccat{R}(I,r\otimes(r'\otimes r''))$};
\node(3)[below of= 1]{$\ccat{R}(I,r\otimes r')\times\ccat{R}(I,r'')$};
\node(4)[below of= 2]{$\ccat{R}(I,r)\times\ccat{R}(I,r'\otimes r'')$};
\node(5)[below of= 3]{$(\ccat{R}(I,r)\times\ccat{R}(I,r'))\times\ccat{R}(I,r'')$};
\node(6)[below of= 4]{$\ccat{R}(I,r)\times(\ccat{R}(I,r')\times\ccat{R}(I,r''))$};
\draw[a](1)to node[la,above]{$\alpha^{\ccat{R}}_{*}$}(2);
\draw[a](1)to node[la,left]{$\pi$}(3);
\draw[a](2)to node[la,right]{$\pi$}(4);
\draw[a](3)to node[la,left]{$\pi\times\id$}(5);
\draw[a](4)to node[la,right]{$\id\times\pi$}(6);
\draw[a](5)to node[la,below]{$\alpha^{\poCat{Poset}}$}(6);
\end{tikzpicture}\]
The commutativity of this diagram on $f\in\ccat{R}(I,(r\otimes r')\otimes r'')$ is equivalently three equalities between the corresponding components in $\ccat{R}(I,r)\times(\ccat{R}(I,r')\times\ccat{R}(I,r''))$. To show all three of these equalities it is advantageous to cast $\delta_{I}={\lambda_{I}}^{\text{-}1}$ which follows from \eqref{eqn.supply_commute_tensors} and the classical fact $\lambda_{I}=\rho_{I}$ in monoidal categories. In this way, the second of these equalities follows only from formal arguments on the axioms of a monoidal category, while the first and third follow from additional equations on $\epsilon$ assured by definition of supply \eqref{eqn.supply_commute_tensors}.
\end{proof}
Thus, if $(\ccat{R},I,\otimes)$ is a prerelational po-category, we define
\begin{equation}\label{eqn.prd_obs}
\poCat{Prd}(\ccat{R},I,\otimes)\coloneqq\left(\ccat{R}\To{\ccat{R}(I,-)}\poCat{Poset}\right)
\text{ with}\hspace{0.5cm}
\begin{gathered}
\adjr{1}{\id}{!}{\ccat{R}(I,I)}\\
\adjr{\ccat{R}(I,r)\times \ccat{R}(I,r')}{\otimes\mathbin{\fatsemi} (\delta_{I})^{*}}{\pi}{\ccat{R}(I,r\otimes r')}
\end{gathered}
\end{equation}
where the right ajax structure indicated on $\ccat{R}(I,-)$ was constructed in \eqref{eqn.prd_right_ajax} above.
Next let $(F,\varphi)\colon\ccat{R}\to\ccat{R}'$ be a morphism of prerelational po-categories, that is, a strong symmetric po-functor. Recall that by \cref{lemma.pres_all_struct_pre}, $F$ preserves the supply of $\mathbb{W}$. We define $\poCat{Prd}(F,\varphi)\coloneqq (F,F^\sharp)$, where the monoidal natural transformation $F^{\sharp}$ below-left is given in components $F^{\sharp}_{r}\colon\ccat{R}(I,r)\to\ccat{R}'(I',Fr)$ as below-right.
\begin{equation}\label{eqn.prd_mors}
\poCat{Prd}(F,\varphi)\coloneqq (F,F^\sharp)\text{ where}\hspace{1cm}
\begin{gathered}
\begin{tikzpicture}
\node(1)[]{$\ccat{R}\phantom{'}$};
\node(2)[below=0.50cm of 1]{$\ccat{R}'$};
\node(3)[right=1.00cm of 2]{$\poCat{Poset}$};
\draw[a](1)to node(F)[la,left]{$F$}(2);
\draw[a,rc](1) to (3|-1) to (3);
\draw[a](2)to node(P')[la,below]{$\ccat{R}'(I,-)$}(3);
\node[above,la] at (P'|-1) {$\ccat{R}(I,-)$};
\cell[right,la]{(P'|-1)}{(P'.north)}{$F^{\sharp}$};
\end{tikzpicture}\\
F^{\sharp}_{r}\coloneqq\ccat{R}(I,r)\xrightarrow{F} \ccat{R}'(FI,Fr)\xrightarrow{(\varphi_{I})^{*}}\ccat{R}'(I',Fr)
\end{gathered}
\end{equation}
It is straightforward to check that this definition is 1-functorial in $(F,\varphi)$.
Finally, let $\alpha\colon (F,\varphi)\Rightarrow (G,\psi)$ be a $2$-morphism of prerelational po-categories, that is, a left adjoint oplax-natural transformation. We wish to provide a $2$-morphism of regular calculi of the form $\poCat{Prd}(\alpha)\coloneqq(\alpha,F^{\sharp}\leq G^{\sharp})\colon\poCat{Prd}(F,\varphi)\Rightarrow\poCat{Prd}(G,\psi)$.
To achieve this, let us note that \cref{lemma.ladjlaxnat_m} implies that $\alpha$ is, in fact, a monoidal oplax-natural transformation. As such, it remains to provide the data of a modification $F^{\sharp}\leq G^{\sharp}$ as in \eqref{eqn.two_morphism_modification}.
The required inequality $F^{\sharp}_{r}\mathbin{\fatsemi} P'\alpha_{r}\leq G^{\sharp}_{r}$ for each $c\in\ob\ccat{R}$ follows from the monoidal nature of $\alpha$ and the oplax-naturality of $\alpha$.
\begin{equation}\label{eqn.prd_twomors}
\poCat{Prd}(\alpha)\coloneqq(\alpha,F^{\sharp}\mathbin{\fatsemi}\ccat{R}'(I',\alpha)\leq G^{\sharp})\text{ where}\hspace{1cm}
\begin{tikzpicture}[node distance=1.25cm,baseline=(3.base)]
\node(1)[]{$I'$};
\node(2)[right of= 1]{$I'$};
\node(3)[below of= 1]{$FI$};
\node(4)[below of= 2]{$GI$};
\node(5)[below of= 3]{$Fr$};
\node(6)[below of= 4]{$Gr$};
\draw[d](1)to(2);
\draw[a](1)to node[la,left]{$\varphi_{I}$}(3);
\draw[a](2)to node[la,right]{$\psi_{I}$}(4);
\draw[a](3)to node[la,above]{$\alpha_{I}$}(4);
\draw[a](3)to node[la,left]{$Ff$}(5);
\draw[a](4)to node[la,right]{$Gf$}(6);
\draw[a](5)to node[la,below]{$\alpha_{r}$}(6);
\path(5) to node[la,sloped]{$\leq$} (4);
\path(3) to node[la,sloped]{$=$} (2);
\end{tikzpicture}
\end{equation}
Again, it is straightforward to see that this is $2$-functorial by transitivity.
In summary, we have demonstrated the following.
\begin{proposition}
The assignments of objects, morphisms and $2$-morphisms given by \eqref{eqn.prd_obs}, \eqref{eqn.prd_mors}, and \eqref{eqn.prd_twomors} assemble into a $2$-functor $\poCat{Prd}\colon\CCat{PrlPoCat}\to\CCat{RgCalc}$.
\end{proposition}
We will henceforth freely confuse the $2$-functor $\poCat{Prd}$ with its restriction to relational po-categories \[\CCat{RlPoCat}\rightarrowtail\CCat{PrlPoCat}\To{\poCat{Prd}}\CCat{RgCalc}\ .\]
\chapter{Graphical regular logic} \label{chap.graphical_reglog}
In this section we finally develop our graphical formalism for regular logic by defining the notion of graphical term, showing how these graphical terms represent predicates in contexts, and explaining how to reason with them. We sketch how the collection of graphical terms, together with our graphical reasoning, allows us to form the ``syntactic po-category'' of a regular calculus in a $2$-functorial fashion. In the companion paper \cite{grl2}, we make extensive use of this graphical regular logic to prove the that regular categories are ``pseudo-reflective'' in regular calculi by means of our syntactic po-category construction.
\section{Graphical terms}
Given a regular calculus $(\cont{P},P)$, graphical terms provide representations of its predicates, i.e.\ the elements in $P(\Gamma)$ for various contexts $\Gamma\in\cont{P}$.
We invite the reader to recall \cref{def.extended_wiring_diagram}, our definition of wiring diagrams in a po-category supplying $\mathbb{W}$, as well as \cref{notation.wiring_diagrams_supplying_ww}, our graphical notation therefor, before considering this next definition and its accompanying notation.
\begin{definition}(Graphical term)\label{def.graphical_term}
Let $(\cont{P},P)$ be a regular calculus. A \define{graphical term} is the data of
\begin{enumerate}[label=(\roman*)]
\item a wiring diagram $(s,k,\{\Gamma_{i}\}_{i\in\ord s},\, \{(\kappa_i,f_i)\}_{i \in \ord k},\, \omega^\ast\colon \Gamma_1\otimes \dots \otimes \Gamma_s \otimes \kappa_1 \otimes \dots \otimes \kappa_k \to \Gamma_\mathrm{out})$ in $\cont P$,
\item for each $i\in\ord s$, a predicate $\theta_i \in P(\Gamma_i)$ in context $\Gamma_{i}=\bigotimes_{j\in\ord{n_{i}}} \tau_{i}(j)$.
\end{enumerate}
We will choose to suppress the details of the wiring diagram and notate such a graphical term by $(\theta_1,\dots,\theta_s;\omega)$, where $\omega$ is the morphism of $\ccat{C}$ represented by the wiring diagram, as given by \cref{def.morphism_represented_by_wd}. If $s=0$ then a graphical term $(;\omega)$ is simply a morphism $\omega\colon I\to\Gamma_\mathrm{out}$.
We say that the graphical term $t = (\theta_1,\dots,\theta_s; \omega)$ \define{represents} the predicate
\begin{equation}\label{eqn.graph_term_reps}
\church{ t } \coloneqq P(\omega)(\theta_1\boxplus^{s}\cdots\boxplus^{s}\theta_s) \in P(\Gamma_\mathrm{out})
\end{equation}
where $\boxplus^{s}$ is the $s$-ary laxator $\boxplus^{s}\colon \prod_{i\in\ord s} P(\Gamma_{i})\to P(\bigotimes_{\ord s} \Gamma_{i})$. In particular, if $s=0$ then a graphical term $(;\omega)$ represents the predicate $P(\omega)(\const{true})$. We extend the equality and implication of the poset $P(\Gamma_{\mathrm{out}})$ to graphical terms $t$, $t'$ via $\church{-}$ in the following sense: we say that \define{$t$ implies $t'$} and that \define{$t$ equals $t'$} when $\church{t}\vdash\church{t'}$ and $\church{t}=\church{t'}$ respectively. In a slight abuse of notation we will write $t\vdash t'$ and $t=t'$ for this implication and equality.\footnote{In this sense graphical terms inherit a pre-order as well as an equivalence relation, relative to which anti-symmetry of the pre-order holds.}
\end{definition}
\begin{example}
When $s=1$ and $\omega=\id$ is the identity, then $\church{(-;\Gamma)}\colon P(\Gamma)\to P(\Gamma)$ is also the identity. More generally for any $s$, when $\omega=\id$, the map \[\church{(-,\,\ldots,\,-;\textstyle\bigotimes_{i\in\ord s} \Gamma_{i})}\colon\prod_{i\in\ord s} P(\Gamma_i)\to P(\textstyle\bigotimes_{i\in\ord s} \Gamma_i)\] is $\boxplus^{s}$, the $s$-ary laxator. We shall see other special cases in \cref{prop.diagrams_meet}.
\end{example}
\begin{notation}[Graphical terms]
We draw a graphical term $(\theta_1,\dots,\theta_s; \omega)$ by drawing the morphism $\omega$ as in \cref{notation.wiring_diagrams_supplying_ww} and annotating the $i$\textsuperscript{th} inner IO shell with its corresponding predicate $\theta_i$. In the case that $\omega$ is the identity morphism, we may simply draw the contexts annotated by their predicates. For instance,
\begin{diagram*}[.][A]
\node(A) {
\begin{tikzpicture}[inner WD, baseline=(rho)]
\node[pack,minimum size = 3ex] (rho) {$\theta$};
\draw (rho.180) to[pos=1] node[left] {$\tau(1)$} +(180:2pt);
\draw (rho.110) to[pos=1] node[above] {$\tau(2)$} +(110:2pt);
\node at ($(rho.10)+(10:6pt)$) {$\vdots$};
\draw (rho.-110) to[pos=1] node[below] {$\tau(s)$} +(-110:2pt);
\end{tikzpicture}
};
\node(B)[left=3.5cm of A.west, anchor=east]{$\church{(\theta;\id_{\bigotimes\tau(i)})}$};
\path(B) -- node[midway]{is represented by} (A);
\end{diagram*}
\end{notation}
\begin{example}
Let $\rc P$ be a regular calculus, and let $\theta_{i}\in P(\Gamma_{i})$ for $i\in\{1,2,3\}$ be predicates, let $f_{1}\colon\Gamma_{1}\to \Gamma'_{1}$ and $f_{2}\colon \Gamma_{1}\otimes \Gamma_{2}\to\Gamma'_{2}$ be morphisms of $\cont P$. Let us write $\sigma\colon \Gamma_{1}\otimes\Gamma_{1}\tpow2\otimes\Gamma_{2}\to\Gamma_{1}\tpow2\otimes(\Gamma_{1}\otimes\Gamma_{2})$ for the appropriate symmetry of $\cont P$, then the predicate \[\left\llbracket\big(\theta_{1},\theta_{2},\theta_{3};(\Gamma_{1}\otimes\delta_{\Gamma_{1}}\otimes \Gamma_{2})\mathbin{\fatsemi}\sigma\mathbin{\fatsemi}((\mu_{\Gamma_{1}}\mathbin{\fatsemi} f_{1})\otimes f_{2})\big)\right\rrbracket\]
is represented by the graphical term
\begin{diagram*}[.][][inner WD]
\node(t1)[pack]{$\theta_{1}$};
\draw(t1.east) -- +(1, 0) node(l1)[link]{};
\draw(l1) to[out=-90, in=180] +(1, -1.5) -- +(2, 0) node(f2)[oshellr,syntax,anchor=140]{$f_{2}$};
\draw(f2.east) -- +(1, 0);
\node(t2)[pack] at (t1|-f2.220) {$\theta_{2}$};
\draw(t2.east) -- (t2-|f2.220);
\draw(l1) to[out=+90, in=180] +(1, +1.5) -- +(1,0) node(l2)[link]{};
\node(f1)[oshellr,syntax,anchor=west] at (l2-|f2.west) {$f_{1}$};
\draw(f1.east) -- +(1, 0);
\draw(l2) -- (f1.west);
\node(t3)[pack, above of=l2, anchor=south]{$\theta_{3}$};
\draw(t3.south) -- (l2);
\end{diagram*}
\end{example}
\begin{example}
We saw in \cref{ex.reg_calc_meet_sl} that right ajax po-functors $1\to\poCat{Poset}$ are $\wedge$-semilattices. The corresponding diagrammatic language has no wires, since $1$ comprises only the monoidal unit. The semantics of an arbitrary graphical term $(\theta_1,\,\ldots,\,\theta_s;\id)$ is simply the meet $\theta_1\wedge\cdots\wedge\theta_s$.
\end{example}
\begin{remark}\label{rem.graphical_regular}
Graphical terms can be used to depict formulae in regular logic. Fix a regular calculus $\rc{P}$, a context $\Gamma\in\cont P$ and suppose $\omega\colon \ord n_1+ \dots + \ord n_s \to \ord n_\mathrm{out}$ is a morphism of $\mathbb{W}$, that is, by \cref{rem.canonical_decomp}, canonically a cospan of the form
\[ \ord n_{1}+\dots+\ord n_{s} \xrightarrow{[\omega_{1},\dots,\omega_{s}]}\ord n_{\omega}\xleftarrow{\omega_{\mathrm{out}}}\ord n_{\mathrm{out}}\ .\]
By the supply of $\mathbb{W}$ this induces a morphism $\omega\colon\bigotimes\Gamma\tpow{n_{i}}\iso\Gamma^{\otimes(\sum n_{i})}\to\Gamma^{\otimes n_{\mathrm{out}}}$. While we will not dwell on the translation, a graphical term $(\theta_1,\dots,\theta_s; \omega)$ represents the following regular formula in free variables $x_{(\mathrm{out},1)},\,\ldots,\,x_{(\mathrm{out},{n}_{\mathrm{out}})}$.
\[
\bigexists_{\substack{i \in \{1,\dots,s,\omega\} \\ j \in \ord{n}_i}} x_{(i,j)} \left[
\bigwedge_{i'\in\{
1,\dots s\}} \theta_{i'}(x_{(i',1)},\, \ldots,\, x_{(i',n_{i'})}) \quad\wedge\quad \bigwedge_{\substack{i'\in\{1,\,\ldots,\,s,\mathrm{out}\} \\j' \in \ord{n}_{i'}
}} \big(x_{(i',j')} = x_{(\omega,\omega_{i'}{(j')})}\big)\right]
\]
For example, given the supplied morphism $\omega\colon\Gamma\tpow{3}\otimes\Gamma\tpow{3}\otimes\Gamma\tpow{4}\to\Gamma\tpow{6}$ of \cref{ex.wiring_diagram} and predicates $\theta_{1}\in P(\Gamma\tpow{3})$, $\theta_{2}\in P(\Gamma\tpow{2})$, and $\theta_{3}\in P(\Gamma\tpow{4})$, the graphical term
\begin{diagram*}[][][penetration=0, inner WD, pack size=9pt, link size=2pt, scale=2, baseline=(out)]
\node[packs] at (-1.5,-1) (f) {$\theta_{3}$};
\node[packs] at (0,1.9) (g) {$\theta_{1}$};
\node[packs] at (1.5,-1) (h) {$\theta_{2}$};
\node[outer pack, inner sep=34pt] at (0,.2) (out) {};
\node[link] at ($(f)!.5!(h)$) (link1) {};
\node[link] at (-2.4,-.25) (link2) {};
\node[link] at ($(f.75)!.5!(g.-135)$) (link3) {};
\draw[fr_out] (out.270) to node[auto]{$x'$} (link1);
\draw[fr_out] (out.190) to node[near end, above]{$z''$} (link2);
\draw[fr_out] (out.155) to node[auto, inner sep=2pt]{$y$} (link3);
\draw[fr_out] (out.-35) to node[auto]{$x'$} (h.-30);
\draw[to_out,fr_out] (out.15) to[out=-165,in=-110] node[auto,swap, near start]{$z$} node[auto,swap,near end,inner sep=1pt]{$z'$} (out.70);
\draw (f.30) to[out=0,in=130] (link1);
\draw (f.-30) to[out=0,in=-130] (link1);
\draw (h.180) to (link1);
\draw (g.-60) to node[auto,inner sep=2pt]{$\tilde x$} (h.120);
\draw (f.45) to node[auto,swap, inner sep=2pt]{$\tilde y$} (g.-105);
\draw (f.75) to (link3);
\draw (g.-135) to (link3);
\end{diagram*}
would represent the following formula after simplification.
\[
\psi(y,z,z',x,x',z'') = \exists\, \tilde{x},\tilde{y}, \left[\theta_1(\tilde{x},\tilde{y},y) \wedge \theta_2(x',\tilde{x},x) \wedge \theta_3(y,\tilde{y},x',x') \wedge (z=z')\right]
\]
\end{remark}
Before we address the \emph{calculus} portion of our graphical notation for regular calculi, let us turn our attention to a final class of examples of our graphical notation: bare regular calculi.
\begin{example}\label{ex.graphical_terms_in_bare}
Recall that a regular calculus $\rc P$ is bare if its category of contexts $\cont P$ is of the form $\cont P=\bigsqcup_{J}\mathbb{W}$. Given our notation of \cref{sec.depict_ww,sec.wds_for_ww_suppliers} for wiring diagrams in $\mathbb{W}$ and the result of \cref{prop.barebare}, we see that a graphical term $(\theta_{1,i_{1}},\,\ldots\,\theta_{k,i_{k}};\omega)$ in the regular calculus $(\bigsqcup_{J}\mathbb{W},P)$ is precisely a collection of wiring diagrams for $\mathbb{W}$ each of whose IO shells has been annotated by predicates $\theta_{j,i_{j}}\in P(\ord n_{j,i_{j}})$ where $j,i_{j}\in J$. A somewhat typical example might therefore be
\begin{diagram*}
\node(diagram) {
\begin{tikzpicture}[inner WD]
\node[pack] (theta1) {$\theta_{1,i}$};
\node[pack, below=.3 of theta1] (theta2) {$\theta_{2,i}$};
\coordinate (helper) at ($(theta1)!.5!(theta2)$);
\node[link, right=2 of helper] (dot R) {};
\draw (theta1.east) to[out=0, in=120] (dot R);
\draw (theta2.east) to[out=0, in=-120] (dot R);
\draw (dot R) -- +(1, 0) -- node[link](end){} +(1,0);
\node[outer pack, fit=(theta1) (theta2) (end)] (outer) {};
\end{tikzpicture}
};
\node(w)[right=2cm of diagram.east, anchor=west] {$\theta_{1,j}$, $\theta_{2,j}\in P(\ord 1_{j})$ for $j\in J$ and $\omega=\mu\mathbin{\fatsemi}\epsilon\colon \ord 1_{j}+\ord 1_{j}\to \ord 0_{j}$.};
\coordinate (m) at ($(diagram.east)!0.5!(w.west)$);
\node[anchor=base] at (m|-w.base) {where};
\end{diagram*}
As there are no labelled morphisms decorating its graphical terms and instead only wires, the regular calculus $(\bigsqcup_{J}\mathbb{W},P)$ is in this visual sense considered \emph{bare}.
\end{example}
\section{Reasoning with graphical terms}
Now that we have understood the graphical notation, it is time to attend to the calculus of manipulations it supports. Let $\rc{P}$ be a regular calculus. The following basic rules for reasoning with graphical terms express the $2$-functoriality and monoidal structure of the po-functor $P\colon \cont P\to\poCat{Poset}$.
\begin{proposition} \label{prop.diagrams_basic}
Let $(\theta_1,\dots,\theta_s;\omega)$ be a graphical term, where $\theta_i \in P(\Gamma_i)$.
\begin{propenum}
\item\label{prop.diagrams_basic_mono} (Monotonicity) Suppose $\theta_i \vdash \theta_i'$ for some $i$. Then
\[
\church{(\theta_1,\dots,\theta_i,\dots,\theta_s; \omega)} \vdash \church{(\theta_1,\dots,\theta_i',\dots,\theta_s; \omega)}.
\]
\item (Breaking) Suppose $\omega \leq \omega'$ in $\cont P$. Then
\[
\church{(\theta_1,\dots,\theta_s; \omega)} \vdash
\church{(\theta_1,\dots,\theta_s; \omega')}.
\]
\item\label{prop.diagrams_basic_nesting} (Nesting) Suppose $\theta_i = \church{(\theta'_1,\dots,\theta'_\ell; \omega')}$ for some $1\leq i\leq s$. Then
\begin{align*}
&\church{(\theta_{1}\,,\ldots,\, \theta_{i-1}, \church{(\theta'_1,\dots,\theta'_\ell; \omega')}, \theta_{i+1}, \ldots, \theta_{s};\omega)} =
\church{(\theta_1,\dots, \theta_{i}\ldots, \theta_s; \omega)}\\
=\, &\church{(\theta_1,\dots,\theta_{i-1},\theta'_1, \dots,
\theta'_\ell,\theta_{i+1},\dots,\theta_s; (\textstyle\bigotimes_{1\leq j< i}\Gamma_{j} \otimes\, {\omega'} \otimes \textstyle\bigotimes_{i< j\leq s}\Gamma_{j} )\mathbin{\fatsemi} \omega)}
\end{align*}
\end{propenum}
\end{proposition}
\begin{proof}
By examining $\church{-}$ of \cref{def.graphical_term} we may reason as below.
\begin{enumerate}[label=(\roman*)]
\item This claim follows from the monotonicity of the map $\boxplus^{s}\mathbin{\fatsemi} P(\omega)$.
\item This claim follows from the $2$-functoriality of $P$.
\item This claim follows from the monoidal structure and $1$-functoriality of $P$. By using the symmetry of $\cont P$, without loss of generality we may assume that $i=s$. In this case, to prove the desired equality it is sufficient to demonstrate the commutativity of the following diagram.
\begin{diagram*}[][][node distance=2cm]
\node(1)[]{$\prod\limits_{j=1}^{s-1}P(\Gamma_j) \times \prod\limits_{j=1}^\ell P(\Gamma'_j)$};
\node(2)[below of= 1]{$\prod\limits_{j=1}^{s-1}P(\Gamma_j) \times P\left(\bigotimes_{j=1}^\ell\Gamma'_j\right)$};
\node(3)[below of= 2]{$\prod\limits^{s}_{j=1}P(\Gamma_{j})$};
\node(4)[right=0.75cm of 2]{$P\left(\bigotimes^{s-1}_{j=1}\Gamma_{j}\otimes \bigotimes^{\ell}_{j=1}\Gamma'_{j}\right)$};
\node(5)[below of= 4]{$P\left(\bigotimes_{j=1}^{s}\Gamma_{j}\right)$};
\node(6) at ($(4)-(1)+(4)$){$P(\Gamma_{\mathrm{out}})$};
\draw[a](1)to node[la,left]{$\id\times\boxplus^{\ell}$}(2);
\draw[a](2)to node[la,left]{$\id\times P(\omega)$}(3);
\draw[a](1)to node[la,auto]{$\boxplus^{(s-1)+\ell}$}(4);
\draw[a](2)to node[la,above]{$\boxplus^{s}$}(4);
\draw[a](4)to node[la,over,inner sep=3pt]{$P(\bigotimes^{s-1}_{j=1}\Gamma_{j}\otimes \omega')$}(5);
\draw[a](3)to node[la,below]{$\boxplus^{s}$}(5);
\draw[a](4)to node[la,auto]{$P((\bigotimes^{s-1}_{j=1}\Gamma_{j}\otimes \omega')\mathbin{\fatsemi} \omega)$}(6);
\draw[a](5)to node[la,below]{$P(\omega)$}(6);
\end{diagram*}
In the above diagram, the upper triangle commutes by coherence laws for $\boxplus$, the square commutes by naturality of $\boxplus$, and the right hand triangle commutes by functoriality of $P$. \qedhere
\end{enumerate}
\end{proof}
\begin{example}
\cref{prop.diagrams_basic} is perhaps more quickly grasped through a graphical example of these facts in action. Suppose we have the entailment
\begin{center}
\begin{tikzpicture}[unoriented WD, font=\small, pack size=7pt, baseline=(P1.south)]
\def\angle{-65};
\node (P1) {\begin{tikzpicture}[inner WD,baseline=(theta.base)]
\node[pack] (theta) {$\theta_1$};
\draw (theta.180) -- +(180:2pt);
\draw (theta.0) -- +(0:2pt);
\draw (theta.\angle) -- +(\angle:2pt);
\end{tikzpicture}};
\node[right=3 of P1] (P2) {
\begin{tikzpicture}[inner WD,baseline=(xi1.base)]
\node[pack] (xi1) {$\xi_1$};
\node[pack, right=1 of xi1] (xi2) {$\xi_2$};
\node[link] at ($(xi1.east)!.5!(xi2.west)$) (dot) {};
\node[outer pack, fit=(xi1) (xi2)] (outer) {};
\draw[fr_out] (outer) -- (xi1.west);
\draw (xi1.east) -- (dot);
\draw (dot) -- (xi2);
\draw[to_out] (xi2) -- (outer);
\draw[to_out] (dot) -- (outer.\angle);
\end{tikzpicture}};
\node at ($(P1.east)!.5!(P2.west)$) {$\vdash$};
\end{tikzpicture}
\end{center}
Then using monotonicity, nesting, and then breaking we may deduce the entailment
\begin{center}
\begin{tikzpicture}[unoriented WD, font=\small, baseline=(P1.195)]
\node (P1) {
\begin{tikzpicture}[inner WD]
\node[pack] (theta1) {$\theta_1$};
\node[pack, right=1.5 of theta1] (theta2) {$\theta_2$};
\node[pack] at ($(theta1)!.5!(theta2)+(0,-2)$) (theta3) {$\theta_3$};
\node[outer pack, inner xsep=2pt, inner ysep=1pt, fit=(theta1) (theta2) (theta3.-30)] (outer) {};
\node[link] at ($(theta2.30)!.5!(outer.30)$) (dot1) {};
\node[link, left=.1 of theta3.west] (dot2) {};
\draw (theta2) -- (dot1);
\draw[to_out] (dot1) to[bend right] (outer.20);
\draw[to_out] (dot1) to[bend left] (outer.45);
\draw (dot2) -- (theta3);
\draw[to_out] (theta1) -- (outer);
\draw[to_out] (theta3) -- (outer);
\draw (theta1) -- (theta3);
\draw (theta1) -- (theta2);
\draw (theta2) -- (theta3);
\end{tikzpicture}
};
\node[right=2 of P1] (P2) {
\begin{tikzpicture}[inner WD]
\node[pack] (xi1) {$\xi_1$};
\node[pack, right=1 of xi1] (xi2) {$\xi_2$};
\node[link] at ($(xi1.east)!.5!(xi2.west)$) (dot) {};
\draw (xi1.east) -- (dot);
\draw (dot) -- (xi2);
\node[outer pack, inner xsep=0, inner ysep=0, fit=(xi1) (xi2)] (outerxi) {};
\node[pack, right=1.5 of xi2] (theta2) {$\theta_2$};
\node[pack, below=.6 of xi2] (theta3) {$\theta_3$};
\node[outer pack, inner xsep=1pt, inner ysep=2pt, fit=(xi1.west) (theta2) (theta3.-10)] (outer) {};
\node[link] at ($(theta2.30)!.5!(outer.30)$) (dot1) {};
\node[link, left=.1 of theta3.west] (dot2) {};
\draw (theta2) -- (xi2.east);
\draw (dot) -- (theta3);
\draw[to_out] (dot1) to[bend right] (outer.20);
\draw[to_out] (dot1) to[bend left] (outer.35);
\draw (dot2) -- (theta3);
\draw[to_out] (xi1) -- (outer);
\draw[to_out] (theta3.south) -- (theta3|-outer.south);
\draw (theta2) -- (theta3);
\draw (theta2) -- (dot1);
\end{tikzpicture}
};
\node[right=2 of P2] (P3) {
\begin{tikzpicture}[inner WD]
\node[pack] (xi1) {$\xi_1$};
\node[pack, right=1 of xi1] (xi2) {$\xi_2$};
\node[link] at ($(xi1.east)!.5!(xi2.west)$) (dot) {};
\draw (xi1.east) -- (dot);
\draw (dot) -- (xi2);
\node[pack, right=1 of xi2] (theta2) {$\theta_2$};
\node[pack, below=.6 of xi2] (theta3) {$\theta_3$};
\node[outer pack, inner xsep=1pt, inner ysep=1pt, fit=(theta1) (theta2) (theta3.-20)] (outer) {};
\node[link] at ($(theta2.30)!.5!(outer.30)$) (dot1) {};
\node[link, left=.1 of theta3.west] (dot2) {};
\draw (theta2) -- (xi2.east);
\draw (dot) -- (theta3);
\draw[to_out] (dot1) to[bend right] (outer.20);
\draw[to_out] (dot1) to[bend left] (outer.35);
\draw (dot2) -- (theta3);
\draw[to_out] (xi1) -- (outer);
\draw[to_out] (theta3) -- (outer);
\draw (theta2) -- (theta3);
\draw (theta2) -- (dot1);
\end{tikzpicture}
};
\node[right=2 of P3] (P4) {
\begin{tikzpicture}[inner WD]
\node[pack] (xi1) {$\xi_1$};
\node[pack, right=1 of xi1] (xi2) {$\xi_2$};
\node[link] at ($(xi1.east)!.5!(xi2.west)+(.5,-1)$) (dotc) {};
\draw (xi1.east) -- (xi2.west);
\node[pack, right=1 of xi2] (theta2) {$\theta_2$};
\node[pack, below=.6 of xi2] (theta3) {$\theta_3$};
\node[outer pack, inner xsep=1pt, inner ysep=1pt, fit=(theta1) (theta2) (theta3.-20)] (outer) {};
\node[link] at ($(theta2.30)!.5!(outer.30)$) (dot1) {};
\node[link, left=.1 of theta3.west] (dot2) {};
\draw (theta2) -- (xi2.east);
\draw (dotc) -- (theta3);
\draw[to_out] (dot1) to[bend right] (outer.20);
\draw[to_out] (dot1) to[bend left] (outer.35);
\draw (dot2) -- (theta3);
\draw[to_out] (xi1) -- (outer);
\draw[to_out] (theta3) -- (outer);
\draw (theta2) -- (theta3);
\draw (theta2) -- (dot1);
\end{tikzpicture}
};
\node (imp1) at ($(P1.east)!.5!(P2.west)$) {$\vdash$};
\node[above=-.5 of imp1] {(i)};
\node (imp2) at ($(P2.east)!.5!(P3.west)$) {$=$};
\node[above=-.5 of imp2] {(iii)};
\node (imp3) at ($(P3.east)!.5!(P4.west)$) {$\vdash$};
\node[above=-.5 of imp3] {(ii)};
\end{tikzpicture}
\end{center}
We'll see many further examples in \cite{grl2}, where we prove that we can construct a regular category from a regular calculus.
\end{example}
The nesting rule of \cref{prop.diagrams_basic_nesting} has two particularly important cases.
\begin{example}[Wiring diagrams as predicates]\label{ex.wd_as_preds}
Let $\omega\colon I\to \Gamma$ be a morphism in $\cont P$. Observe that we have the equalities
\begin{diagram*}[,][6]
\node(1)[anchor=east]{$\church{(;\omega)}=\big(1$};
\node(2)[right=2cm of 1]{$P(I)$};
\node(3)[right=2cm of 2]{$P(\Gamma)\big)$};
\node(4)[anchor=east] at ($(1.east)-(0,0.75cm)$){$=\big(1$};
\node(5)[] at ($(2)-(0,0.75cm)$){$P(\Gamma)$};
\node(6)[anchor=west] at ($(3.west)-(0,0.75cm)$){$P(\Gamma)\big)=\church{(\church{(;w)};\Gamma)}$};
\draw[a](1)to node[la,above]{$\const{true}$}(2);
\draw[a](2)to node[la,above]{$P(\omega)$}(3);
\draw[a](4)to node[la,above]{$\const{true}\mathbin{\fatsemi} P(\omega)$}(5);
\draw[a](5)to node[la,above]{$P(\Gamma)$}(6);
\end{diagram*}
so that we are justified in equating the following two graphical terms
\begin{diagram*}[.][pred]
\node(wd){\begin{tikzpicture}[inner WD]
\node(w)[oshellr,syntax]{$\omega$};
\node(outer)[outer pack,fit={(w)},inner sep=7pt]{};
\draw[to_out](w.east) -- (outer);
\end{tikzpicture}};
\node(pred)[right=1cm of wd]{\begin{tikzpicture}[inner WD]
\node(p)[pack]{$\church{(;\omega)}$};
\node(outer)[outer pack,fit={(p)},inner sep=7pt]{};
\draw[to_out](p.south) -- (outer);
\end{tikzpicture}};
\node at ($(wd)!0.45!(pred)$) {$=$};
\end{diagram*}
Moreover, every regular calculus has a rich stock of such morphisms $I\to \Gamma$. In \cref{lemma.name_unfolding_iso} we exhibited an isomorphism $\cont P(\bigotimes \Gamma_{i},\Gamma_{\mathrm{out}})\iso \cont P(I,\bigotimes\Gamma_{i}\otimes \Gamma_{\mathrm{out}})$ mediated by taking the name $\omega\mapsto \name\omega$ of a morphism. In this way we may view arbitrary wiring diagrams $\omega\colon \bigotimes\Gamma_{i}\to \Gamma_{\mathrm{out}}$ in $\cont P$ as graphical terms $\church{(\church{(;\name\omega)};\bigotimes \Gamma_{i}\otimes \Gamma_{\mathrm{out}})}$ of the above-right form.
Note, however, that in general the above merely constitutes a map of wiring diagrams into predicates which preserves representation. It is not necessarily the case that all predicates $\theta\in P(\Gamma)$ may be realised as $\church{(\church{(;\omega)};\Gamma)}$ for some wiring diagram $\omega$. Nevertheless, in \cref{sec.prd_graphical_terms} we will see that there is a large class of regular calculi in which wiring diagrams and predicates do coincide.
\end{example}
\begin{example}[Exterior conjunction]
Let $\Gamma_{1}$ and $\Gamma_{2}$ be contexts, and let $\theta_{1}\in P(\Gamma_{1})$ and $\theta_{2}\in P(\Gamma_{2})$ be predicates. Observe that we have the equalities
\[ \theta_{1}\boxplus \theta_{2} = \church{(\theta_{1}\boxplus \theta_{2};\Gamma_{1}\otimes \Gamma_{2})} =\church{(\theta_{1},\theta_{2};\Gamma_{1}\otimes \Gamma_{2})}\]
of elements of $P(\Gamma_{1}\otimes \Gamma_{2})$ so that we are justified in equating, for example, the following two graphical terms.
\begin{diagram*}
\node (P1) {
\begin{tikzpicture}[inner WD]
\node[pack] (theta1) {$\theta_1$};
\node[pack, below=.4 of theta1] (theta2) {$\theta_2$};
\node[outer pack, inner ysep=0pt, fit=(theta1) (theta2)] (outer) {};
\draw (theta1.0) -- (theta1.0-|outer.east);
\draw (theta1.90) -- (outer.north);
\draw (theta1.180) -- (theta1.180-|outer.west);
\draw (theta2.270) -- (outer.south);
\end{tikzpicture}
};
\node[right=1cm of P1] (P2) {
\begin{tikzpicture}[inner WD, pack size=6pt]
\node[pack] (rho) {$\theta_1\boxplus\theta_2$};
\draw (rho.0) -- +(0:2pt);
\draw (rho.90) -- +(90:2pt);
\draw (rho.180) -- +(180:2pt);
\draw (rho.270) -- +(270:2pt);
\end{tikzpicture}
};
\node at ($(P1.east)!.5!(P2.west)$) {$=$};
\pgfresetboundingbox
\useasboundingbox (P1.130) rectangle (P2.-40);
\end{diagram*}
Under the interpretation of graphical terms as formulae in regular logic suggested by \cref{rem.graphical_regular}, this process of vertical merging of graphical terms corresponds to the logical conjunction of the formulae they represent.
\end{example}
In \cref{prop.meet_sl} we saw how a regular calculus endows each poset $P(\Gamma)$ with the structure of a meet-semilattice. As we will now see, this structure permits an intuitive graphical interpretation. In the following proposition, the graphical terms on right are illustrative examples of the equalities stated on the left.
\begin{proposition} \label{prop.diagrams_meet}
For all contexts $\Gamma\in\ob\cont P$ and predicates $\theta_{1}$, $\theta_{2}\in P(\Gamma)$, we have
\begin{propenum}
\item\label{prop.diagrams_meet_true} (True is removable) $\church{(\const{true}_{\Gamma};\Gamma)} = \church{ (;\eta_\Gamma) }$, \hfill $
\begin{tikzpicture}[unoriented WD, font=\small,baseline=(true)]
\node (P1) {
\begin{tikzpicture}[inner WD]
\node[pack] (true) {$\const{true}$};
\draw (true.0) -- +(0:2pt);
\draw (true.120) -- +(120:2pt);
\draw (true.240) -- +(240:2pt);
\end{tikzpicture}
};
\node[right=3 of P1] (P2) {
\begin{tikzpicture}[inner WD, shorten >=-2pt]
\coordinate (helper);
\node[link] (dot0) at ($(helper)+(0:5pt)$) {};
\node[link] (dot120) at ($(helper)+(120:5pt)$) {};
\node[link] (dot240) at ($(helper)+(240:5pt)$) {};
\node[outer pack, surround sep=8pt, fit=(helper)] (outer) {};
\draw (dot0) -- (outer.0);
\draw (dot120) -- (outer.120);
\draw (dot240) -- (outer.240);
\end{tikzpicture}
};
\node at ($(P1.east)!.5!(P2.west)$) {$=$};
\end{tikzpicture}
$
\item (Meets-are-merges)
$
\church{ (\theta_1\wedge\theta_2;\Gamma)} = \church{
(\theta_1,\theta_2;\mu_\Gamma) }.
$
\hfill
$\begin{tikzpicture}[unoriented WD,baseline=(theta)]
\node (P1) {
\begin{tikzpicture}[inner WD]
\node[pack] (theta1) {$\theta_1$};
\node[pack, below=.3 of theta1] (theta2) {$\theta_2$};
\coordinate (helper) at ($(theta1)!.5!(theta2)$);
\node[link, right=2 of helper] (dot R) {};
\node[link, left=2 of helper] (dot L) {};
\draw (theta1.east) to[out=0, in=120] (dot R);
\draw (theta2.east) to[out=0, in=-120] (dot R);
\draw (theta1.west) to[out=180, in=60] (dot L);
\draw (theta2.west) to[out=180, in=-60] (dot L);
\draw (dot R) -- +(5pt, 0);
\draw (dot L) -- +(-5pt, 0);
\end{tikzpicture}
};
\node[right=3 of P1] (P2) {
\begin{tikzpicture}[inner WD]
\node[pack] (theta) {$\theta_1\wedge\theta_2$};
\draw (theta.180) -- +(180:2pt);
\draw (theta.0) -- +(0:2pt);
\end{tikzpicture}
};
\node at ($(P1.east)!.5!(P2.west)$) {$=$};
\end{tikzpicture}
$
\end{propenum}
\end{proposition}
\begin{proof}
These equations are simply the definitions of $\const{true}$
and meet; see \eqref{eqn.meet_sl} and \eqref{eqn.graph_term_reps}.
\end{proof}
\begin{example}[Discarding]\label{lemma.dotting_off}
Note that \cref{prop.diagrams_meet_true} and the monotonicity of diagrams
(\cref{prop.diagrams_basic_mono}) further imply that for all $\theta \in P(\Gamma)$
we have $\theta \vdash \church{
(;\eta_\Gamma) }$:
\[
\begin{tikzpicture}[unoriented WD, font=\small, baseline=(phi)]
\node (P1) {
\begin{tikzpicture}[inner WD, shorten >=-2pt]
\node[pack] (phi) {$\theta$};
\draw (phi.0) -- +(0:2pt);
\draw (phi.180) -- +(180:2pt);
\draw (phi.270) -- +(270:2pt);
\end{tikzpicture}
};
\node[right=3 of P1] (P2) {
\begin{tikzpicture}[inner WD, shorten >=-2pt]
\node[pack, fill=white, white] (phi) {$\theta$};
\node[outer pack, fit=(phi)] (outer) {};
\node[link] at ($(outer.180) - (180:5pt)$) (dot180) {};
\node[link] at ($(outer.0) - (0:5pt)$) (dot0) {};
\node[link] at ($(outer.270) - (270:5pt)$) (dot270) {};
\draw (dot0) -- (outer.0);
\draw (dot180) -- (outer.180);
\draw (dot270) -- (outer.270);
\end{tikzpicture}
};
\node at ($(P1.east)!.5!(P2.west)$) {$\vdash$};
\end{tikzpicture}
\]
\end{example}
\section{Graphical terms in the regular calculus of a regular theory}
\label{sec.graph_terms_regul}
In this section we will revisit \cref{con.rgcalc_rgtheory} which associates to a regular theory $(\Sigma,\mathbb T)$ a regular calculus $(\ccat{C}_{\mathbb T},F_{\mathbb T})$. In particular we will illustrate graphically the highlighted formulae of \cref{ex.rgth_inhabited,ex.rgth_preorder,ex.rgth_monoid,ex.rgth_cat}.
\subparagraph{Theory of an inhabited sort} In \cref{ex.rgth_inhabited} we outlined the theory of an inhabited sort. In the associated regular calculus, the graphical terms are simply wiring diagrams whose IO shells are annotated by formulae whose arity is no greater than the number of wires. The single axiom of the theory, $\top\vdash\exists_{s}\top$, is rendered as the below entailment of graphical terms. Note that $\top$ is the blank graphical term in the empty context and we have made use of the equation $\exists_{s}\top=F_{\mathbb T}(\eta;\epsilon)(\top)$ for the consequent.
\begin{diagram*}
\node(B)[]{
\begin{tikzpicture}[inner WD]
\node[outer pack, minimum size=20pt]{};
\end{tikzpicture}
};
\node(A)[right=2cm of B]{
\begin{tikzpicture}[inner WD]
\node(1)[link]{};
\node[outer pack, minimum size=20pt,anchor=center] at (1){};
\end{tikzpicture}
};
\node(At)[below=0.5 of A]{$\exists_{s}\top$};
\node(Bt)at (At-|B){$\top$};
\path (B) -- node[midway]{$\leq$} (A);
\end{diagram*}
In fact this inequality is an equality, for we always have $\exists_{s}\top\leq\top$, and so the theory of an inhabited sort allows us to freely erase or introduce dots; compare \eqref{eqn.extra_law} of $\mathbb{W}$. The outer circles were rendered above to emphasise that these are graphical terms, but from here we will suppress this detail in keeping with \cref{rem.suppress}.
\subparagraph{Theory of a pre-order} In \cref{ex.rgth_preorder} we outlined the theory of a pre-order. Here we wish to draw attention to the two axioms of the theory, reflexivity $x\vdash xPx$ and transitivity $\exists_{y}[xPy\wedge yPz]\vdash xPz$, and their renditions as the below-left and below-right graphical terms in the associated regular calculus, respectively. Note in particular that the former contributes to the regular calculus genuine, strict inequalities between graphical terms.
\begin{diagram*}
\node(A)[]{
\begin{tikzpicture}[inner WD]
\node(1)[minimum size=0, inner sep=0]{};
\draw(1) -- ++(-2,0) -- ++(2,0);
\node(2)[pack] at ($(1)+(6,0)$) {$P$};
\draw(2.west) -- ++(-0.5, 0);
\draw(2.east) -- ++(+0.5, 0);
\path(1) -- node(a)[midway]{$\leq$} ($(2)-(1,0)$);
\end{tikzpicture}
};
\node(B)[right=3cm of A.east,anchor=west]{
\begin{tikzpicture}[inner WD]
\node(1)[pack]{$P$};
\node(2)[below= of 1,pack]{$P$};
\coordinate(l) at ($(1)!0.5!(2)$){};
\draw(1) -- (2);
\draw(1.north) -- ++(0, +0.5);
\draw(2.south) -- ++(0, -0.5);
\node(3)[right=4 of l,pack]{$P$};
\draw(3.north) -- ++(0,+0.5);
\draw(3.south) -- ++(0,-0.5);
\path(l-|2.east) -- node[midway]{$\leq$} (3);
\end{tikzpicture}
};
\end{diagram*}
\subparagraph{Theory of a monoid} In \cref{ex.rgth_monoid} we outlined the theory of a monoid, and explored the corresponding regular calculus. Here in particular we wish to highlight the utility of graphical terms by drawing a graphical term below-right which represents the formula for commutativity of two variables below-left.
\begin{diagram*}[][][]
\node(A)[]{$\exists_{m,m'}\big[\star(x,y,m)\wedge\star(y,x,m')\wedge m=m'\big]$};
\node(C)[right= of A]{
\begin{tikzpicture}[inner WD]
\node(1)[pack]{$\star$};
\node(2)[pack,right= of 1]{$\star$};
\draw(1.east)--(2.west);
\node(l1)[left= of 1,link]{};
\draw(l1)-- ++(-1.0,0);
\node(l2)[right= of 2,link]{};
\draw(l2)-- ++(+1.0,0);
\draw(l1) to[out=-45,in=270] (1);
\draw(l2) to[out=-135,in=270] (2);
\draw(l1) to[out=45,in=120] (2);
\draw(l2) to[out=135,in=60] (1);
\end{tikzpicture}
};
\end{diagram*}
\subparagraph{Theory of a category} In \cref{ex.rgth_cat} we outlined the theory of regular a category, and produced a formula on pairs $f,g$ of morphisms corresponding to the property of the pair being mutually inverse $[f\circ g=\id]\wedge [g\circ f=\id]$. Rendered as a graphical term in the associated regular calculus this formula might look as follows.
\begin{diagram*}[][][inner WD]
\node(l1)[link]{};
\node(c1)[pack]at ($(l1)+(5,+2)$) {$\comp$};
\node(c2)[pack]at ($(l1)+(5,-2)$) {$\comp$};
\node(l2) at ($(c1)+(5,-2)$)[link]{};
\node(id1)[pack,above=of c1]{$\id$};
\node(la1)[link,above= of id1]{};
\node(cod1)[pack,left= of la1]{$\cod$};
\node(dom1)[pack,right= of la1]{$\dom$};
\node(id2)[pack,below=of c2]{$\id$};
\node(la2)[link,below= of id2]{};
\node(dom2)[pack,left= of la2]{$\dom$};
\node(cod2)[pack,right= of la2]{$\cod$};
\draw(c1.north) -- (id1.south);
\draw(id1.north) -- (la1);
\draw(cod1.east) -- (la1) -- (dom1.west);
\draw(l1) to[out=60,in=180] (cod1);
\draw(l1) to[out=45,in=180] (c1);
\draw(l1) to[out=-45,in=180] (c2);
\draw(l1) to[out=-60,in=180] (dom2);
\draw(c2.south) -- (id2.north);
\draw(id2.south) -- (la2);
\draw(cod2.west) -- (la2) -- (dom2.east);
\draw(l2) to[out=120,in=0] (dom1);
\draw(l2) to[out=135,in=0] (c1);
\draw(l2) to[out=-135,in=0] (c2);
\draw(l2) to[out=-120,in=0] (cod2);
\draw(l1) -- node[above]{$f$} ++(-3,0);
\draw(l2) -- node[above]{$g$} ++(+3,0);
\end{diagram*}
\section{Graphical terms in the regular calculus of predicates}\label{sec.prd_graphical_terms}
We wish to highlight, as a special case, regular calculi of the form $\poCat{Prd}\ccat{R}$ where $\ccat{R}$ is a prerelational po-category. Recall the definition of graphical terms (\cref{def.graphical_term}) and observe that a graphical term in such a regular calculus comprises the data of a wiring diagram $\omega\colon r_{1}\otimes\cdots\otimes r_{s}\to r_{\mathrm{out}}$ in $\ccat{R}$ as well as \emph{morphisms} $\{\theta_{i}\colon I\to r_{i}\}_{i\in\ord s}$ of $\ccat{R}$.
Note that the right ajax structure on $\ccat{R}(I,-)$ of \eqref{eqn.prd_right_ajax} has in particular $\const{true}=\id_{I}$. As such, a graphical term $(\theta_{1},\,\ldots,\,\theta_{s};\omega)$ with $s>0$ represents the same predicate as the graphical term $(;\sigma\mathbin{\fatsemi}(\bigotimes \theta_{i})\mathbin{\fatsemi}\omega)$, namely $\church{(\theta_{1},\,\ldots,\,\theta_{s};\omega)}=\sigma\mathbin{\fatsemi} (\bigotimes \theta_{i})\mathbin{\fatsemi}\omega$, where $\sigma\colon I\to I\tpow{s}$ is the appropriate symmetry. As such, we have succeeded in obtaining an equality\footnote{Recall that graphical terms inherit an equality relation under taking of representations, $\church{-}$} of the graphical terms $(\theta_{1},\,\ldots,\,\theta_{s};\omega)=(;\sigma\mathbin{\fatsemi}(\bigotimes \theta_{i})\mathbin{\fatsemi}\omega)$, where the latter is merely the data of a wiring diagram $I\to r_{\mathrm{out}}$.
Of course we may read the equality $(\theta_{1},\,\ldots,\,\theta_{s};\omega)=(;\sigma\mathbin{\fatsemi}(\bigotimes \theta_{i})\mathbin{\fatsemi}\omega)$ ``the other way''. Given any wiring diagram $\omega\colon I\to r_{\mathrm{out}}$ -- that is, a graphical term of the form $(;\omega)$ -- we may construct the graphical term $(\omega;\id_{r_{\mathrm{out}}})$ such that $(;\omega)=(\omega;\id_{r_{\mathrm{out}}})$. Graphically this is the observation that any wiring diagram with empty domain $I$ may be equivalently re-drawn as a predicate.
Recall however that in the presence of the name-unfolding isomorphism of \cref{lemma.name_unfolding_iso}, the distinction between domain and codomain is fluid. With this final piece we are ready to observe that the data of graphical terms in $\poCat{Prd}\ccat{R}$ is precisely the data of wiring diagrams in $\ccat{R}$, in the following sense.
\begin{lemma}\label{lemma.graphical_terms_in_prd}
Let $\ccat{R}$ be a prerelational po-category and fix $s\in\mathbb{N}$ and objects $r_{1},\,\ldots,\,r_{s},\, r_{\mathrm{out}}$ of $\ccat{R}$. Then, with graphical terms considered in $\poCat{Prd}\ccat{R}$,
\begin{lemenum}
\item the set of wiring diagrams $\omega\colon\bigotimes r_{i}\to r_{\mathrm{out}}$ in $\ccat{R}$ is in bijection with the set of graphical terms of the form $(;\omega'\colon I\to (\bigotimes r_{i})\otimes r_{\mathrm{out}})$, mediated by the assignment $\omega\mapsto(;\name{\omega})$,
\item\label{lemma.graphical_terms_in_prd_ii} the set of graphical terms $(\theta_{1},\,\ldots,\,\theta_{s};\omega\colon\bigotimes r_{i}\to r_{\mathrm{out}})$ and the set of graphical terms of the form $(;\omega'\colon I\to r_{\mathrm{out}})$ admit the following opposed functions between them which preserve the represented predicate
\begin{align*} (\theta_{1},\,\ldots,\,\theta_{s};\omega)&\mapsto(;\sigma\mathbin{\fatsemi}(\textstyle\bigotimes\theta_{i})\mathbin{\fatsemi}\omega)\\
(\omega';\id_{r_{\mathrm{out}}})&\mapsfrom(;\omega')
\end{align*}
where $\sigma\colon I\to I\tpow{s}$ is the appropriate symmetry of $\ccat{R}$.\hfill$\qed$
\end{lemenum}\vspace{-\topsep}
\end{lemma}
The first bijection above is a strengthening of the graphical phenomenon we identified in \cref{ex.wd_as_preds}. The second correspondence, although not strictly a bijection of sets\footnote{but rather an isomorphism of setoids}, lends itself readily to graphical understanding.
\begin{example}[Graphical terms as wiring diagrams]
For a given graphical term, the combination of the correspondence of \cref{lemma.graphical_terms_in_prd_ii} above and nesting (\cref{prop.diagrams_basic_nesting}) establishes that any IO shell containing a predicate may be re-drawn as a morphism of $\ccat{R}$ with empty domain $I$ without altering the represented predicate. For example, the following two graphical terms represent the same predicate.
\begin{diagram*}[][][baseline=(A.base)]
\node(A){
\begin{tikzpicture}[inner WD]
\node(t1)[pack]{$\theta_{1}$};
\draw(t1.east) -- +(1, 0) node(l1)[link]{};
\draw(l1) to[out=-90, in=180] +(1, -1.5) -- +(2, 0) node(f2)[oshellr,syntax,anchor=140]{$f_{2}$};
\draw(f2.east) -- +(1, 0);
\node(t2)[pack] at (t1|-f2.220) {$\theta_{2}$};
\draw(t2.east) -- (t2-|f2.220);
\draw(l1) to[out=+90, in=180] +(1, +1.5) -- +(1,0) node(l2)[link]{};
\node(f1)[oshellr,syntax,anchor=west] at (l2-|f2.west) {$f_{1}$};
\draw(f1.east) -- +(1, 0);
\draw(l2) -- (f1.west);
\node(t3)[pack, above of=l2, anchor=south]{$\theta_{3}$};
\draw(t3.south) -- (l2);
\end{tikzpicture}
};
\node(B)[right=3cm of A.east, anchor=west]{
\begin{tikzpicture}[inner WD]
\node(t1)[oshellr,syntax]{$\theta_{1}$};
\draw(t1.east) -- +(1, 0) node(l1)[link]{};
\draw(l1) to[out=-90, in=180] +(1, -1.5) -- +(2, 0) node(f2)[oshellr,syntax,anchor=140]{$f_{2}$};
\draw(f2.east) -- +(1, 0);
\node(t2)[oshellr,syntax] at (t1|-f2.220) {$\theta_{2}$};
\draw(t2.east) -- (t2-|f2.220);
\draw(l1) to[out=+90, in=180] +(1, +1.5) -- +(1,0) node(l2)[link]{};
\node(f1)[oshellr,syntax,anchor=west] at (l2-|f2.west) {$f_{1}$};
\draw(f1.east) -- +(1, 0);
\draw(l2) -- (f1.west);
\node(t3)[oshelld,syntax, above of=l2, anchor=south]{$\theta_{3}$};
\draw(t3.south) -- (l2);
\end{tikzpicture}
};
\path (A) --node[midway]{$=$} (B);
\end{diagram*}
\end{example}
\section{The syntactic po-category of a regular calculus: a sketch}\label{sec.syn}
As we have seen, graphical terms provide an effective way to reason in regular calculi. It is thus of interest to consider forming, from a given regular calculus $\rc P$, a po-category $\poCat{Syn}\rc P$ whose \emph{objects} are the graphical terms of $\rc P$. In this way we may study the collection of representations for predicates at once -- that is, we may study the \emph{syntax} of the regular calculus.
In what follows we will sketch the ``syntactic po-category'' construction, but we will choose here to defer the full details to the companion paper. Consider then the following collections of data which together form the objects and morphisms of the \define{syntactic po-category} $\poCat{Syn}\rc P$ of the regular calculus $\rc P$.
\begin{equation}\label{eqn.syntactic_def}
\begin{cases}
\ob\poCat{Syn}\rc{P} &\coloneqq\{(\Gamma,p)\mid \Gamma\in\ob\cont{P}, p\in P(\Gamma)\}\\
\poCat{Syn}\rc{P}\big( (\Gamma_1,p_1), (\Gamma_2,p_2)\big) &\coloneqq\{\theta_{12}\in P(\Gamma_1\otimes \Gamma_2)\mid\theta_{12}\leq p_1\boxplus p_2\}
\end{cases}
\end{equation}
Of course to claim that these data form a po-category we must provide various additional structures and prove properties thereof. While it is possible to continue our sketch in the language of supplied morphisms and right ajax po-functors -- that is, semantically -- we will instead make use of the tools of graphical regular logic. Thus, in pictures, the to-be po-category $\poCat{Syn}\rc P$ has:
\begin{itemize}
\item objects $(\Gamma,p)$ represented by graphical terms
\begin{tikzpicture}[inner WD, baseline=(g)]
\node[pack] (rho) {$p$};
\draw (rho.south) to node(g)[anchor=west]{$\Gamma$} +(0,-0.5);
\end{tikzpicture}
\item morphisms $\theta_{12}\colon(\Gamma_1,p_1)\to(\Gamma_2,p_2)$ represented by graphical terms
\begin{tikzpicture}[inner WD, pack size=4pt, baseline=(g.base)]
\node[pack] (theta) {$\theta_{12}$};
\draw (theta.west) to[pos=1] node(g)[left=-3pt] {$\Gamma_1$} +(-2pt, 0);
\draw (theta.east) to[pos=1] node[right=-2pt] {$\Gamma_2$} +(2pt, 0);
\end{tikzpicture}
together with an entailment
\begin{tikzpicture}[inner WD, baseline=(g.base)]
\node (P1) {\begin{tikzpicture}[inner WD]
\node[pack, pack size=6pt] (theta) {$\theta_{12}$};
\draw (theta.west) to[pos=1] node(g)[left=-3pt] {$\Gamma_1$} +(-2pt, 0);
\draw (theta.east) to[pos=1] node[right=-2pt] {$\Gamma_2$} +(2pt, 0);
\end{tikzpicture}};
\node (P2) [right=3 of theta] {
\begin{tikzpicture}[inner WD]
\node[pack] (p1) {$p_1$};
\node[pack, right=of p1] (p2) {$p_2$};
\node[outer pack, fit=(p1) (p2)] (outer) {};
\draw[shorten >=-2pt] (p1.west) to[pos=1] node[left] {$\Gamma_1$} (outer.west);
\draw[shorten >=-2pt] (p2.east) to[pos=1] node[right] {$\Gamma_2$} (outer.east);
\end{tikzpicture}
};
\node[font=\normalsize] at ($(P1.east)!.5!(P2.west)$) {$\vdash$};
\end{tikzpicture}
\vspace{-.2in}
\item the identity on $(\Gamma,p)$ represented by the graphical term
\begin{tikzpicture}[inner WD, baseline=(g.base)]
\node[pack] (phi) {$p$};
\node[link, below=3pt of phi] (dot) {};
\node[outer pack, fit=(phi) (dot)] (outer) {};
\draw (phi) -- (dot);
\draw (dot) to[pos=1] node(g)[left] {$\Gamma$} (dot-|outer.west);
\draw (dot) to[pos=1] node[right] {$\Gamma$} (dot-|outer.east);
\end{tikzpicture}
\item the composite $\theta_{12}\mathbin{\fatsemi}\theta_{23}$ represented by the graphical term
\begin{tikzpicture}[inner WD, pack size=4pt, baseline=(theta.base)]
\node[pack] (theta) {$\theta_{12}$};
\node[pack, right=2 of theta] (theta') {$\theta_{23}$};
\node[outer pack, fit=(theta) (theta')] (outer) {};
\draw (outer.east) to (theta'.east) node[right=1] {$\scriptstyle \Gamma_3$};
\draw (theta) to node[above=-.3] {$\scriptstyle \Gamma_2$} (theta');
\draw (theta.west) node[left=1] {$\scriptstyle \Gamma_1$} to (outer.west);
\end{tikzpicture}
\end{itemize}
With the induced poset structure on the homs, in the companion paper \cite{grl2} we prove that the collections $\poCat{Syn}\rc P$ with the composition and identities above indeed form a po-category.
\begin{example}
In \cref{ex.reg_calc_meet_sl} we established that $\wedge$-semilattices $L$ are equivalently regular calculi $(1,L)$. By unwinding the syntactic po-category construction for such a regular calculus we see that $\poCat{Syn}(1,L)$ is the po-category whose objects are the elements of $L$, whose hom posets $\poCat{Syn}(1,L)(l,l')$ are the down-sets $\mathop\downarrow\{l\wedge l'\}$, whose composition is meet, and whose identities are given by the top element of each hom poset.
\end{example}
The syntactic po-category moreover inherits from $\cont P$ a canonical symmetric monoidal structure where the monoidal product of objects, the monoidal product of morphisms, and the braiding correspond respectively to the following graphical terms -- for details see \cite{grl2}.
\begin{diagram*}
\node(prod_obj){\begin{tikzpicture}[inner WD, baseline=(theta2.north)]
\node[pack] (theta1) {$p_1$};
\node[pack, right=.4 of theta1] (theta2) {$p_2$};
\node[inner ysep=5pt, fit=(theta1) (theta2)] (outer) {};
\draw (theta1.south) -- (theta1.south|-outer.south);
\draw (theta2.south) -- (theta2.south|-outer.south);
\end{tikzpicture}};
\node(prod_mor)[right=1 of prod_obj]{ \begin{tikzpicture}[inner WD, baseline=(theta2.north)]
\node[pack] (theta1) {$\theta_1$};
\node[pack, below=.4 of theta1] (theta2) {$\theta_2$};
\node[inner ysep=0pt, fit=(theta1) (theta2)] (outer) {};
\draw (theta1.0) -- (theta1.0-|outer.east);
\draw (theta1.180) -- (theta1.180-|outer.west);
\draw (theta2.0) -- (theta2.0-|outer.east);
\draw (theta2.180) -- (theta2.180-|outer.west);
\end{tikzpicture}};
\node(braiding)[right=1 of prod_mor]{\begin{tikzpicture}[inner WD, baseline=(dot2.north)]
\node[link] (dot1) {};
\node[link,below=.4 of dot1] (dot2) {};
\node[pack,above=.3 of dot1,inner sep=1pt] (phi1) {$p_1$};
\node[pack,below=.3 of dot2,inner sep=1pt] (phi2) {$p_2$};
\node[inner xsep=3pt, inner ysep=-3pt, fit=(phi1) (phi2)] (outer) {};
\draw (dot1) -- (phi1.south);
\draw (dot2) -- (phi2.north);
\draw (dot1) -- (dot1-|outer.west);
\draw (dot1) -- (dot2-|outer.east);
\draw (dot2) -- (dot2-|outer.west);
\draw (dot2) -- (dot1-|outer.east);
\end{tikzpicture}};
\end{diagram*}
With this symmetric monoidal structure we may induce, from the supply of $\mathbb{W}$ in $\cont P$, a supply of $\mathbb{W}$ in $\poCat{Syn}\rc P$. This work appears in \cite{grl2}, but we summarise the results here. Recall that $\epsilon$, $\delta$, $\eta$, and $\mu$ are the generating morphisms of $\mathbb{W}$; see \eqref{eqn.generating_wires}. For an object $(\Gamma,p)\in\ob\poCat{Syn}\rc P$, the supplied morphisms corresponding to these generators are the following graphical terms.
\begin{equation*}
\begin{tikzpicture}[unoriented WD]
\node (P1) {
\begin{tikzpicture}[inner WD]
\node[pack,inner sep=1pt] (phi1) {$p$};
\node[outer pack, minimum size=7ex, fit=(phi1)] (outer) {};
\draw (phi1.west) -- (outer.180);
\end{tikzpicture}
};
\node[below=.1 of P1] {$\epsilon_{(\Gamma,p)}$};
\node[right=3 of P1] (P2) {
\begin{tikzpicture}[inner WD]
\node[link] (dot1) {};
\node[pack,above=.3 of dot1,inner sep=1pt] (phi1) {$p$};
\node[outer pack, minimum size=7ex, fit=(dot1)] (outer) {};
\draw (dot1) -- (phi1.south);
\draw (dot1) -- (outer.180);
\draw (dot1) -- (outer.30);
\draw (dot1) -- (outer.-30);
\end{tikzpicture}
};
\node[below=.1 of P2] {$\delta_{(\Gamma,p)}$};
\node[right=3 of P2] (P3) {
\begin{tikzpicture}[inner WD]
\node[pack,inner sep=1pt] (phi1) {$p$};
\node[outer pack, minimum size=7ex, fit=(phi1)] (outer) {};
\draw (phi1.east) -- (outer.0);
\end{tikzpicture}
};
\node[below=.1 of P3] {$\eta_{(\Gamma,p)}$};
\node[right=3 of P3] (P4) {
\begin{tikzpicture}[inner WD]
\node[link] (dot1) {};
\node[pack,above=.3 of dot1,inner sep=1pt] (phi1) {$p$};
\node[outer pack, circle, minimum size=7ex, fit=(dot1)] (outer) {};
\draw (dot1) -- (phi1.south);
\draw (dot1) -- (outer.0);
\draw (dot1) -- (outer.150);
\draw (dot1) -- (outer.210);
\end{tikzpicture}
};
\node[below=.1 of P4] {$\mu_{(\Gamma,p)}$};
\end{tikzpicture}
\end{equation*}
In fact there is even more structure present, $\poCat{Syn}\rc P$ is a relational po-category. As a corollary of our main theorem we prove \cref{cor.vague_eqv} in the companion: there is a suitably natural equivalence of relational po-categories $\ccat{R}\eqv\poCat{Syn}\poCat{Prd}\ccat{R}\eqv\poCat{Syn}(\bigsqcup_{J}\mathbb{W},F)$ for some ajax po-functor $F$ derived from $\ccat{R}$. By using the equivalence $\CCat{RlPoCat}\eqv\CCat{RgCat}$ between relational po-categories and regular categories of \cref{thm.carboni_equivalence} we may therefore observe that we have obtained equivalences $\cat R\eqv\Funr{LAdj}\poCat{Syn}\rc P$ for some regular calculus $\rc P$ derived from $\cat R$.
Of interest here to the classical categorical logician is that we may work in regular categories using our graphical syntax for regular calculi. As this equivalence is one of regular categories, we may in particular compute limits graphically. For example, given two parallel left adjoints $h,k$ in $\Funr{LAdj}\poCat{Syn}\rc P$ we may compute the equaliser $\operatorname{eq}(h,k)\to\dom(h)$ and the image sub-object $\operatorname{im}(h)\rightarrowtail\cod(h)$ by the below-left and below-right graphical terms respectively.
\begin{diagram*}
\node(A){
\begin{tikzpicture}[inner WD]
\node(1)[funcr,syntax]{$h$};
\node(2)[funcr,syntax,below= of 1]{$k$};
\coordinate(3) at ($(1)!0.5!(2)+(2,0)$);
\draw[out=0,in=90] (1) to (3) to[out=270,in=0] (2);
\node(l)[link] at ($(1)!0.5!(2)-(2,0)$){};
\draw[out=180,in=90] (1) to (l) to[out=270,in=180] (2);
\draw(l) -- ++(-1,0);
\end{tikzpicture}
};
\node(B)[right=3cm of A.east,anchor=west]{
\begin{tikzpicture}[inner WD]
\node(1)[funcr,syntax]{$h$};
\draw(1.east) -- ++(1,0);
\draw(1.west) -- ++(-1,0) node[link]{};
\end{tikzpicture}
};
\end{diagram*}
We conclude this section with some remarks about the nature of our syntactic po-category construction.
\begin{remark}
Observe that $\poCat{Poset}$ is a sub-$2$-category of $\CCat{Cat}$. Instead of our bespoke construction of the syntactic po-category above, it is tempting to consider some appropriate po-categorical variant of a monoidal Grothendieck construction -- perhaps as developed in \cite{moeller2018monoidal} or \cite{Buckley2013}. Indeed, at the level of objects it would seem that there is a coincidence between $\poCat{Syn}\rc P$ and the total space of a Grothendieck-type construction $\int\rc P$.
However, it presently appears to the authors that any so-attempted recasting of $\poCat{Syn}$ is doomed to failure. In a putative Grothendieck construction, consider the pair of objects $(I,\const{true})$ and $(\Gamma,p)$. In order for the total space to supply $\mathbb{W}$, we would require the presence of morphisms $\widehat\epsilon\colon(\Gamma,p)\to(I,\const{true})$ and $\widehat\eta\colon(I,\const{true})\to(\Gamma,p)$. However, in general we have only $P(\epsilon)(p)\vdash\const{true}$ and $p\vdash P(\eta)(\const{true})$, and thus there appears to be no uniform way to select the direction of the inequalities for morphisms in $\int\rc P$.
In this way, some form of `symmetrisation' of domain and codomain becomes necessary, considerations of which result in our $\poCat{Syn}\rc P$.
\end{remark}
Despite the fact that in later sections and the companion paper we shall realise $\poCat{Syn}$ as a highly-structured $2$-functor, it fails to mediate any form of categorical equivalence. That is, the following example stands to establish that a regular calculus is \emph{more} than the data of its graphical terms. Of course this is completely expected from the syntax-semantics point of view, and stands in directly analogy with other means of describing theories: different logical axioms can result in the same theory or have the same semantics. For an example of the former, for the theory of a pre-order we see that the formulations $x P y\wedge y P z\vdash_{x,y,z} x Pz$ and $\exists_{y}[x P y\wedge y P z]\vdash_{x,z} x Pz$ of transitivity imply the same theory. For an example of the latter, a finite product sketch with $A=B\times C$ is semantically equivalent with one in which $A=B\times C$ and $A=C\times B$. In this way, the appropriate notion of ``same-ness'' for regular calculi is a form of ``Morita equivalence'' as detected by $\poCat{Syn}$; see \cref{thm.model_equiv} for a theorem from this perspective.
\begin{example}[$\poCat{Syn}$ identifies distinct regular calculi]
Recall \cref{ex.reg_calc_meet_sl}, that is, that $\wedge$-semilattices $L$ are equivalently regular calculi $L\colon1\to\poCat{Poset}$. Consider then that we may form the degenerate regular calculus $(\ccat{C},\ccat{C}\To{!}1\To{L}\poCat{Poset})$ for any $\ccat{C}$ supplying $\mathbb{W}$. The syntactic po-categories of these degenerate regular calculi are equally degenerate: in $\poCat{Syn}(\ccat{C},\ccat{C}\To{!}1\To{L}\poCat{Poset})$ there is a (natural) isomorphism $(c,p)\iso(c',p)$ for all elements $p\in L$ and objects $c,c'\in\ob\ccat{C}$, viz, $p$ itself.
It may be checked that given $\ccat{C}$ and $\ccat{D}$ inequivalent symmetric monoidal po-categories supplying $\mathbb{W}$, the regular calculi $(\ccat{C},!_{\ccat{C}}\mathbin{\fatsemi} L)$ and $(\ccat{D},!_{\ccat{D}}\mathbin{\fatsemi} L)$ are inequivalent. However, the po-functor $\poCat{Syn}(\ccat{C},!_{\ccat{C}}\mathbin{\fatsemi} L)\to\poCat{Syn}(\ccat{D},!_{\ccat{D}}\mathbin{\fatsemi} L)$ which sends $(c,p)\mapsto (I^{\ccat{D}},p)$ and $\theta\mapsto\theta$ mediates an equivalence (with evident inverse) of symmetric monoidal po-categories supplying $\mathbb{W}$. Thus $\poCat{Syn}$ cannot mediate a $2$-dimensional equivalence of $2$-categories.
\end{example}
\begin{remark}
In fact this example is part of a more general class. In \cref{prop.syn_is_relational} below we record a result of the companion paper: whenever there is a morphism of regular calculi $(F,F^{\sharp})\colon\cont P\to\cont Q$ such that $F$ is essentially surjective (\cref{def.po_equiv}) and $F^{\sharp}$ is an isomorphism then $\poCat{Syn}(F,F^{\sharp})\colon\poCat{Syn}\cont P\to\poCat{Syn}\cont Q$ mediates an equivalence. This result corresponds to our understanding above: the essential surjectivity of $F$ says roughly that $\cont P$ has enough contexts to model those of $\cont Q$, and the invertibility of $F^{\sharp}$ says that we may bijectively translate relations across contexts from $\cont P$ to $\cont Q$ without altering semantics. At worst then, $\cont P$ contains somehow redundant copies of contexts of $\cont Q$, but without a difference in semantics.
From this the above example may be derived. Observe that for any po-category $\ccat{C}$ supplying $\mathbb{W}$, the canonical morphism $(!_{\ccat{C}},\id_{L})\colon(\ccat{C},!_{\ccat{C}}\mathbin{\fatsemi} L)\to(1,L)$ of regular calculi satisfies these conditions. Thus we may conclude that $\poCat{Syn}(\ccat{C},!_{\ccat{C}}\mathbin{\fatsemi} L)\eqv\poCat{Syn}(1,L)$.
\end{remark}
\section{Morphisms of regular calculi \texorpdfstring{\&}{and} graphical terms}\label{sec.graphical_morphisms}
Our notions of $1$- and $2$-morphisms of regular calculi, \cref{def.reg_calc_morphisms}, interact well with graphical terms and indeed preserve all of the desired structure. We have just seen the sense in which graphical terms in a regular calculus are meaningfully the objects of a syntactic po-category \eqref{eqn.syntactic_def}, so we now elucidate the manner in which morphisms of regular calculi $\rc P\to \rc Q$ act on graphical terms.
Given a morphism $(F,F^{\sharp})\colon \rc P\to\rc Q$ of regular calculi and a collection of contexts $\{\Gamma_{i}\}_{i\in\{1,\,\ldots,\,s,\,\mathrm{out}\}}$ in $\cont P$, the monoidal $2$-naturality of $F^{\sharp}$ renders commutative the following diagram.
\begin{diagram*}
\node(1)[]{$\prod P(\Gamma_{i})$};
\node(s)[right=2cm of 1.east]{};
\node(2)[right=2cm of s,anchor=west]{$P(\bigotimes \Gamma_{i})$};
\node(3)[right=2cm of 2,anchor=west]{$P(\Gamma_{\mathrm{out}})$};
\node(4)[below= of 1]{$\prod P'F(\Gamma_{i})$};
\node(5)[anchor=base] at (s|-4.base) {$P'(\bigotimes' F(\Gamma_{i}))$};
\node(6)[below= of 2]{$P'F(\bigotimes \Gamma_{i})$};
\node(7)[below= of 3]{$P'F(\Gamma_{\mathrm{out}})$};
\draw[a](1)to node[la,above]{$\boxplus$}(2);
\draw[a](2)to node[la,above]{$P(\omega)$}(3);
\draw[a](1)to node[la,left]{$\prod F^{\sharp}_{\Gamma_{i}}$}(4);
\draw[a](3)to node[la,right]{$F^{\sharp}_{\Gamma_{\mathrm{out}}}$}(7);
\draw[a](4)to node[la,below]{$\boxplus'$}(5);
\draw[a](5)to node[la,below]{$P'\varphi$}(6);
\draw[a](6)to node[la,below]{$P'F(\omega)$}(7);
\end{diagram*}
Thus, given a graphical term $(\theta_{1},\,\ldots,\,\theta_{s};\omega)$ of $\rc P$ where $\theta_{i}\in P(\Gamma_{i})$, we see that we obtain the graphical term $(F^{\sharp}_{\Gamma_{1}}(\theta_{1}),\,\ldots,\,F^{\sharp}_{\Gamma_{s}}(\theta_{s});\varphi\mathbin{\fatsemi} F(\omega))$ of $\rc Q$ with the property \[\church{(F^{\sharp}_{\Gamma_{1}}(\theta_{1}),\,\ldots,\,F^{\sharp}_{\Gamma_{s}}(\theta_{s});\varphi\mathbin{\fatsemi} F(\omega))}=F^{\sharp}_{\Gamma_{\mathrm{out}}}(\church{(\theta_{1},\,\ldots,\,\theta_{s};\omega)})\ .\]
The fact that $\varphi$ is the strongator of the supply-preserving strong symmetric monoidal po-functor $F$ affords us an easy graphical understanding of this action. First we replace all the predicates $\theta_{i}\in P(\Gamma_{i})$ in IO shells with the predicates $F^{\sharp}_{\Gamma_{i}}(\theta_{i})$, and then when $\omega$ is composed of tensors of morphisms $\omega=\bigotimes\omega_{i}$, we may ``pull $\varphi$ through the tensors'' in $F(\omega)$ and preserve wiring as we go. These principles are illustrated by the following example.
\begin{diagram*}
\node(G1) {
\begin{tikzpicture}[inner WD]
\node(t1)[pack]{$\theta_{1}$};
\draw(t1.east) -- +(1, 0) node(l1)[link]{};
\draw(l1) to[out=-90, in=180] +(1, -1.5) -- +(2, 0) node(f2)[oshellr,syntax,anchor=140]{$f_{2}$};
\draw(f2.east) -- +(1, 0);
\node(t2)[pack] at (t1|-f2.220) {$\theta_{2}$};
\draw(t2.east) -- (t2-|f2.220);
\draw(l1) to[out=+90, in=180] +(1, +1.5) -- +(1,0) node(l2)[link]{};
\node(f1)[oshellr,syntax,anchor=west] at (l2-|f2.west) {$f_{1}$};
\draw(f1.east) -- +(1, 0);
\draw(l2) -- (f1.west);
\node(t3)[pack, above of=l2, anchor=south]{$\theta_{3}$};
\draw(t3.south) -- (l2);
\end{tikzpicture}
};
\node(F1)[right=2cm of G1.east,anchor=west] {
\begin{tikzpicture}[inner WD]
\node(t1)[pack]{$F^{\sharp}\theta_{1}$};
\draw(t1.east) -- +(1, 0) node(l1)[link]{};
\draw(l1) to[out=-90, in=180] +(1, -2) -- +(2, 0) node(f2)[funcr,syntax,anchor=150,inner sep=7pt]{$\varphi$};
\draw(f2.east) -- +(1, 0) node(f2p)[oshellr,syntax,anchor=west, shrink]{$Ff_{2}$};
\draw(f2p.east) -- +(1, 0);
\node(t2)[pack] at (t1|-f2.210) {$F^{\sharp}\theta_{2}$};
\draw(t2.east) -- (t2-|f2.210);
\draw(l1) to[out=+90, in=180] +(1, +2) -- +(1,0) node(l2)[link]{};
\node(f1)[oshellr,syntax,anchor=west, shrink] at (l2-|f2.west) {$Ff_{1}$};
\draw(f1.east) -- (f1.east-|f2p.east) -- +(1, 0);
\draw(l2) -- (f1.west);
\node(t3)[pack, above of=l2, anchor=south]{$F^{\sharp}\theta_{3}$};
\draw(t3.south) -- (l2);
\end{tikzpicture}
};
\path (G1) --node[midway]{$\xmapsto{(F,F^{\sharp})}$} (F1);
\end{diagram*}
This description suggests that morphisms of regular calculi preserve the connectivity, wiring, and compositionality of graphical terms, and so all the structure present in our syntactic po-categories. Indeed, as we prove in the companion paper \cite{grl2}, such a morphism $(F,F^{\sharp})\colon\rc P\to\rc Q$ of regular calculi induces a symmetric monoidal supply preserving po-functor $\poCat{Syn}(F,F^{\sharp})\colon\poCat{Syn}\rc P\to\poCat{Syn}\rc Q$. In this fashion we may prove that $\poCat{Syn}$ forms a $2$-functor $\poCat{Syn}\colon\CCat{RgCalc}\to\CCat{RlPoCat}$, and observe moreover that it sends certain morphisms of regular calculi to equivalences.
\begin{proposition}[{\cite{grl2}}]\label{prop.syn_is_relational}
The syntactic po-category construction is the on-objects component of a $2$-functor $\poCat{Syn}\colon\CCat{RgCalc}\to\CCat{RlPoCat}$. Moreover, if $(F,F^{\sharp})\colon\cont P\to\cont Q$ is a morphism of regular calculi such that $F$ is essentially surjective and $F^{\sharp}$ is an isomorphism then $\poCat{Syn}(F,F^{\sharp})$ is an equivalence.
\end{proposition}
\chapter{Preview of the next paper}\label{chap.preview}
Now that we have seen all of the major players in the theory of regular calculi, let us give an overview of the results and details developed in the companion paper \cite{grl2}.
To begin with, in the companion we give the full account of the construction of the syntactic po-category $\poCat{Syn}\rc P$ of a regular calculus $\rc P$ and prove that this extends to a $2$-functor to relational po-categories, subsuming the outlines of \cref{sec.syn,sec.graphical_morphisms} here.
Using this and the $2$-functor $\poCat{Prd}$ of \cref{sec.prd} here, we turn our attention to rigorously stating and proving our main theorem which connects regular calculi to relational po-categories, and so to regular categories.
\begin{theorem}\label{thm.main}
The $2$-functors $\poCat{Syn}\colon\CCat{RgCalc}\to\CCat{RlPoCat}$ and $\poCat{Prd}\colon\CCat{RlPoCat}\to\CCat{RgCalc}$ are involved in a bi-adjunction $\poCat{Syn}\dashv_{\mathrm{bi}}\poCat{Prd}$. Moreover, the co-unit of the bi-adjunction is part of an adjoint equivalence so that this bi-adjunction is pseudo-reflection of $\CCat{RlPoCat}$ into $\CCat{RgCalc}$.
\end{theorem}
This main theorem allows us to deduce almost all of the rest of our results. For instance, we use it to give a $2$-dimensional account of the regular completion of a category with finite limits.
\begin{theorem}\label{thm.reg_lex}
The forgetful $2$-functor $U\colon\CCat{RgCat}\to\CCat{FLCat}$ is right bi-adjoint to the composite
\[ \CCat{FLCat}\To{\Funr{Span^{po}}}\CCat{PrlPoCat}\To{\poCat{Prd}}\CCat{RgCalc}\To{\poCat{Syn}}\CCat{RlPoCat}\To{\Funr{LAdj}}\CCat{RgCat}\]
where $\CCat{FLCat}$ is the category of categories with finite limits.
\end{theorem}
As a corollary of this result, we are able to show that taking the regular category of left adjoints in a relational po-category is in fact bi-represented.
\begin{corollary}\label{cor.ladj_birep}
The $2$-functor $\Funr{LAdj}\colon\CCat{RlPoCat}\to\CCat{RgCat}$ is bi-represented by the relational po-category $\poCat{Syn}\poCat{Prd}\mathbb{W}$, that is, there is a pseudo-natural adjoint equivalence of the $2$-functors
\[ \CCat{RlPoCat}(\poCat{Syn}\poCat{Prd}\mathbb{W},-)\ensuremath{\mathrel{\overset{ae}\eqv}}\Funr{LAdj}\colon\CCat{RlPoCat}\to\CCat{RgCat}\ .\]
\end{corollary}
Our main theorem also affords us an important ``bare-ification'' result for regular calculi: every relational po-category is equivalent to the syntactic po-category of a bare regular calculus. As such, we are entitled to reason in an arbitrary relational po-category using graphical terms with \emph{no} kites.
\begin{corollary}\label{cor.vague_eqv}
Let $\ccat{R}$ be a relational po-category, then there are pseudo-natural equivalences in $\CCat{RlPoCat}$
\[
\ccat{R} \eqv \poCat{Syn}\poCat{Prd}\ccat{R}\eqv \poCat{Syn}\big(\textstyle\bigsqcup_{\ob\ccat{R}}\mathbb{W},\ s^{\sqcup}\mathbin{\fatsemi} \ccat{R}(I,-)\big)\ .
\]
\end{corollary}
With these tools, in the companion we then return to the topic of regular theories and their connection to and realisations as regular calculi. We connect our work on regular calculi to the classical devices of categorical regular logic by first proving the following theorem.
\begin{theorem}\label{thm.syn_equiv}
Given a regular theory $(\Sigma,\mathbb T)$, there is an equivalence of categories between the syntactic regular category $\cat C^{\mathrm{reg}}_{\mathbb T}$ associated to the regular theory and the regular category of left adjoints $\Funr{LAdj}\poCat{Syn}(\ccat{C}_{\mathbb T},F_{\mathbb T})$ in the syntactic po-category of the regular calculus associated to the regular theory (\cref{con.rgcalc_rgtheory}).
\end{theorem}
To address the question of models of a regular calculus, we first make the following definition.
\begin{definition}\label{def.models}
The $2$-category of models of a regular calculus $\rc P$ in a relational po-category $\ccat{R}$ is defined to be $\rc P\model(\ccat{R})\coloneqq\CCat{RlPoCat}(\poCat{Syn}\rc P,\ccat{R})$. In this way, $\rc P\model(\cdot)$ extends to a $2$-functor $\rc P\model(\cdot)\colon\CCat{RlPoCat}\to\CCat{Cat}$.
\end{definition}
Using this, our main theorem, \cref{thm.main}, and \cref{thm.syn_equiv} we are able to prove that models in our sense suitably agree with and therefore generalise the classical notion of models of a regular theory.
\begin{theorem}\label{thm.model_equiv}
Let $(\Sigma, \mathbb T)$ be a regular theory. There is an equivalence of categories \[(\ccat{C}_{\mathbb T},F_{\mathbb T})\text-\operatorname{Mod}(\poCat{Rel}\cat R)\eqv \mathbb T\text-\operatorname{Mod}(\cat R)\] pseudo-natural in regular categories $\cat R$.
\end{theorem}
\printbibliography
\end{document}
|
\section{Conclusions and future work} \label{Section:Conclusions}
In this work, we established a formal framework for probabilistic unclean graph databases, which allowed us to study several basic problems within them. The formalism defined in Section~\ref{section:definitions} expands the ideas described in \cite{de2018formal} addressing some particular difficulties that arise when dealing with data-graphs. We also described a set of restrictions over the initial model analogous to those found in other works of the area (mainly the usual subset and superset restrictions considered when dealing with inconsistent databases as in \cite{bienvenu2012complexity} or \cite{barcelo2017data}) and show that they still hold a reasonable expressive power to capture real life examples. We defined for unclean data-graphs the problems of \decisionProblem{Data cleaning} and \decisionProblem{Probabilistic Query Answering} that strongly relate with two fundamental questions in the context of unclean databases developed in \cite{de2018formal}. First, given an observed database and a model for the evolution of inconsistencies: Which is the most likely original state of the database? And second, given the same prior knowledge: Which is the probability that a certain condition or constraint holds in the original state of the database?
In Sections~\ref{section:dataCleaning} and~\ref{section:PQA} we studied specific restrictions of the problems and determined that the complexity of the subset and superset versions of \decisionProblem{Data cleaning} and \decisionProblem{Probabilistic Query Answering} is hard in both cases. Nevertheless, we provided additional restrictions for which the problem becomes tractable. Regarding the study of \decisionProblem{PQA} in probabilistic graph databases, we considered \decisionProblem{Global PQA} and \decisionProblem{Existential PQA} over data-graphs, and studied this using the \Gregxpath syntax to express conditions or constraints over data-graphs. It follows from the proofs given in Section~\ref{section:PQA},
that the results obtained can be generalized to similar graph query languages, as long as they are (1) expressive enough to define some simple path conditions on every node and (2) simple enough to compute the set of answers given a data-graph instance in polynomial time.
Following a similar idea, we could also define the \emph{update-PUGD} restriction over the PUGD model, in which we consider the possibility of modification of labels and data values in edges and nodes, respectively, as well as edge or node deletions or additions. This could also be seen as an analogous restriction already studied in the literature of inconsistency reasoning: namely, the symmetric difference repairs studied for example in \cite{afrati2009repair}. In the update case we have a much less restrictive environment than those studied in Sections~\ref{section:dataCleaning} and~\ref{section:PQA}, and
therefore the results obtained in this work suggest that this generalization is a much more difficult problem to address.
We observed in Section~\ref{section:constraints} that the concept of soft and hard constraints from the area of database repairing can be lifted to our framework of EMGDs and PUGDs. We do so by defining (weighted) restriction sets, which modify the distribution of a probabilistic data-graph by reducing the probability of those data-graphs which violate the constraints in the restriction set.
In this work we barely mentioned the complexity considerations of asking for constraint satisfaction, as this was not the focus of our study, but we hope to explore this topic in the future.
Some questions remain open in the context of this work. First, we did not prove a tight upper bound for the \decisionProblem{PQA} problem when considering only subset PUGDs. The best upper bound we show is \textsc{PSPACE}, while the lower bound is \textsc{PP}, so there might be some room for improvement. Second, we did not find restrictions over the realization model $\aRealizationModel$ that could make the problems easy, without trivializing it. In most cases, the interaction between really simple probabilistic databases $\aProbabilisticDatabase$ and realization models $\aRealizationModel$ already make the problems hard, so one possible way to advance would be to consider instances of the \decisionProblem{Data Cleaning} and \decisionProblem{PQA} problem where the distribution defined by $\aProbabilisticDatabase$ is uniform and the weight of the reasoning is put into $\aRealizationModel$.
Regarding the imposition of syntactical limitations in order to reduce complexity,
subsets of $\Gregxpath$ could be explored to lower the complexity of the \decisionProblem{PQA} problems. Nonetheless, observe that the results of hardness use a really narrow set of tools from the \Gregxpath grammar, and therefore serious limitations on the complexity of the expression might have to be imposed. Considering expressions from weaker navigation languages (such as those mentioned in \cite{barcelo2012expressive}) might be a good place to start.
We also left open to explore the possibility of semantic restrictions in the universe $\aUniverseOfGraphs$ of data-graphs, such as considering only data-trees, DAGs, or other type of structures in order to investigate whether these restrictions provide a lowering of the complexity for our problems. Similarly, we can study the trade-offs involved in further restrictions on the probabilistic data-graphs $\aProbabilisticDatabase$ and the noisy observers $\aRealizationModel$.
It would also be interesting to consider an alternative but natural definition of the semantics of a given formula $\aFormulaNodeOrPath$ over a data-graph, such that it returns the proportion of nodes or pairs of nodes that satisfy the formula.
\section{Adding constraints}\label{section:constraints}
While a particular EMGD $(\aProbabilisticDatabase, \aRealizationModel)$ can be used to encode information about our epistemic understanding, sometimes it makes sense to restrict our space of possible worlds in particular ways. Depending on the field of application, these restrictions could arise from obtaining a deeper theoretical understanding, as a way to explore multiple hypotheses, from practical limitations on the world-models that can be effectively analyzed, or even from philosophical considerations such as the anthropic principle.
The concept of hard and soft constraints from the area of database repairing \cite{barcelo2017data, tenCate:2012, arenas1999consistent} can be incorporated to our framework of graph databases, for both the problems of Data cleaning and PQA. We can consider a \defstyle{set of constraints} of the form $$\aRestrictionSet = \{\aPathExpression_1, \dots, \aPathExpression_m, \aNodeExpression_1, \dots, \aNodeExpression_n\},$$ where $m,n \ge 0$, and $\{\aPathExpression_i\}_{1 \le i \le m}$ and $\{\aNodeExpression_i\}_{1 \le i \le n}$ are sets of path expressions and node expressions, respectively (note that they can be empty sets).
\begin{definition}[Restrictions and consistency]
Let $\aGraph$ be a data-graph and $\aRestrictionSet = \aRestrictionSetPaths \union \aRestrictionSetNodes$ a set of \defstyle{restrictions} (also called \defstyle{constraints}), where $\aRestrictionSetPaths$ consists of path expressions and $\aRestrictionSetNodes$ of node expressions.
We say that $\aGraph$ \defstyle{violates} $\aNodeExpression$, if $\exists \aNode \in V_\aGraph$ such that $\aNode \not\in \semantics{\aNodeExpression}$. Otherwise we say that $\aGraph$ \defstyle{satisfies} $\aNodeExpression$, which we notate by $\aGraph \models \aNodeExpression$.
Similarly, we say that $\aGraph$ violates $\aPath$ if $\exists \aNode,\aNodeb \in V_\aGraph$ such that $(\aNode,\aNodeb) \not\in \semantics{\aPath}$, and otherwise we say that $\aGraph$ \defstyle{satisfies} $\aPath$ ($\aGraph \models \aPath$).
If for all $\aFormulaNodeOrPath \in \aRestrictionSet$, $\aGraph$ does not violate $\aFormulaNodeOrPath$, we say that $\aGraph$ is \defstyle{consistent} w.r.t.\ $\aRestrictionSet$, and we note this as $\aGraph \models \aRestrictionSet$.
We say that $\aGraph,\aNode$ \defstyle{violates} $\aNodeExpression$, if $\aNode \not\in \semantics{\aNodeExpression}$.
Similarly, we say that $\aGraph, \aNode, \aNodeb$ violates $\aPath$ if $(\aNode,\aNodeb) \not\in \semantics{\aPath}$.
\end{definition}
Naturally, this notion of consistency is useful for dividing the set of data-graphs into two groups: those who satisfy the constraints and those who do not. However, this stark division can be generalized.
Given $\aFormulaNodeOrPath \in \aRestrictionSet$, we can assign a weight $0 \le w_\aFormulaNodeOrPath < 1$ to it. If we do so for all formulas in the set, we call the resulting set $\aRestrictionSetWithWeights \eqdef \{(\aFormulaNodeOrPath,w_\aFormulaNodeOrPath) \mid \aFormulaNodeOrPath \in \aRestrictionSet\}$ a \defstyle{weighted restriction set} (which we may also think of as a function with domain $\aRestrictionSet$).
A value of $0$ can be thought of as a \defstyle{hard (exclusionary) constraint}: we want our clean models to never satisfy the formula. On the other hand, a value between $0$ and $1$ is a \defstyle{soft constraint}: the models that satisfy the formula are penalized in their probability, but not absolutely.
Given $\aRestrictionSetWithWeights$ and a decision problem $\decisionProblem{M}$ (or a function problem $\functionProblem{M}$) that depends on $(\aProbabilisticDatabase, \aRealizationModel)$, the associated problem \defstyle{$\problemWithConstraints{\decisionProblem{M}}{\aRestrictionSetWithWeights}$} (resp. \defstyle{$\problemWithConstraints{\functionProblem{M}}{\aRestrictionSetWithWeights}$}) is the same problem but over $(\aProbabilisticDatabase_\aRestrictionSetWithWeights, \aRealizationModel)$, which is a an EMGD where the weights given by $\aRestrictionSetWithWeights$ have been incorporated and normalized from the original values of $\aProbabilisticDatabase$.
More precisely:
\[
\aProbabilisticDatabase_\aRestrictionSetWithWeights(\aGraph)
\eqdef \frac{\aRestrictionSetWithWeights(\aGraph)}{\sum_{\aGraphb \in \aProbabilisticDatabase} \aRestrictionSetWithWeights(\aGraphb)}
\eqdef \frac{\aProbabilisticDatabase(\aGraph) \times \prod_{\aFormulaNodeOrPath \in \aRestrictionSet} \aRestrictionSetWithWeights(\aGraph, \aFormulaNodeOrPath)}{\sum_{\aGraphb \in \aProbabilisticDatabase} \aRestrictionSetWithWeights(\aGraphb)}
\]
where we notated:
\[
\aRestrictionSetWithWeights(\aGraph) \eqdef \aProbabilisticDatabase(\aGraph) \times \prod_{\aFormulaNodeOrPath \in \aRestrictionSet} \aRestrictionSetWithWeights(\aGraph, \aFormulaNodeOrPath)
\]
and where $\aRestrictionSetWithWeights(\aGraph, \aFormulaNodeOrPath)$ is defined as:
\[ \aRestrictionSetWithWeights(\aGraph, \aFormulaNodeOrPath) \eqdef \begin{cases}
\aRestrictionSetWithWeights(\aFormulaNodeOrPath) & \mbox { if $\aGraph \models \aFormulaNodeOrPath$}\\
1 & \mbox{ otherwise}
\end{cases}
\]
Note that incorporating $\aRestrictionSetWithWeights$ does not introduce modifications in $\aRealizationModel$, but it does change $\aRealizationModel^{-1}$.
\subsection{Complexity considerations}
Given an EMGD $(\aProbabilisticDatabase, \aRealizationModel)$ and a weighted restriction set $\aRestrictionSetWithWeights$, how much harder are the problems of Data cleaning and PQA over $\aProbabilisticDatabase_\aRestrictionSetWithWeights$?
Observe that, since $\aProbabilisticDatabase$ could contain an infinite number of data-graphs of non-zero probability, explicitly calculating $\aProbabilisticDatabase_\aRestrictionSetWithWeights$ might not be possible.
\begin{example}
Assume a very simple case where $\aRestrictionSet$ consists of a single node or a single path expression, i.e. $\aRestrictionSet = \{\aFormulaNodeOrPath\}$. Furthermore, assume that we know the total probability mass $p_\aFormulaNodeOrPath$ of data-graphs in $\aProbabilisticDatabase$ where $\aFormulaNodeOrPath$ holds. Then if follows from a simple calculation that:
$$\aProbabilisticDatabase_\aRestrictionSetWithWeights(\aGraph) = \frac{\aProbabilisticDatabase(\aGraph) \times \aRestrictionSetWithWeights(\aGraph, \aFormulaNodeOrPath)}{p_\aFormulaNodeOrPath \times \aRestrictionSetWithWeights(\aFormulaNodeOrPath) + (1 - p_\aFormulaNodeOrPath) }$$
Indeed, this follows from the definition of $\aProbabilisticDatabase_\aRestrictionSetWithWeights(\aGraph)$ and from writing the denominator as follows:
\begin{align*}
\sum_{\aGraphb \in \aProbabilisticDatabase} \aRestrictionSetWithWeights(\aGraphb) &= \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \models \aFormulaNodeOrPath} \aRestrictionSetWithWeights(\aGraphb) + \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \not\models \aFormulaNodeOrPath} \aRestrictionSetWithWeights(\aGraphb)\\
& = \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \models \aFormulaNodeOrPath} (\aProbabilisticDatabase(\aGraphb) \times \aRestrictionSetWithWeights(\aGraphb, \aFormulaNodeOrPath)) + \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \not\models \aFormulaNodeOrPath} (\aProbabilisticDatabase(\aGraphb) \times \aRestrictionSetWithWeights(\aGraphb, \aFormulaNodeOrPath)) \\
& = \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \models \aFormulaNodeOrPath} (\aProbabilisticDatabase(\aGraphb) \times \aRestrictionSetWithWeights(\aFormulaNodeOrPath)) + \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \not\models \aFormulaNodeOrPath} (\aProbabilisticDatabase(\aGraphb) \times 1) \\
& = \aRestrictionSetWithWeights(\aFormulaNodeOrPath) \times \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \models \aFormulaNodeOrPath} \aProbabilisticDatabase(\aGraphb)+ \sum_{\aGraphb \in \aProbabilisticDatabase\, \mid\, \aGraphb \not\models \aFormulaNodeOrPath} \aProbabilisticDatabase(\aGraphb) \\
& = \aRestrictionSetWithWeights(\aFormulaNodeOrPath) \times p_\aFormulaNodeOrPath + (1-p_\aFormulaNodeOrPath)
\end{align*}
\end{example}
Note that the same idea of this example can be generalized for more complex sets of constraints, as long as we know the probability masses in $\aProbabilisticDatabase$ for each combination the Boolean values of the formulas in $\aRestrictionSet$ (e.g., if $\aRestrictionSet = \{\aNodeExpression, \aNodeExpressionb\}$, we ask for $p_{\aNodeExpression, \aNodeExpressionb}, p_{\lnot\aNodeExpression, \aNodeExpressionb}, p_{\aNodeExpression, \lnot\aNodeExpressionb}, p_{\lnot\aNodeExpression, \lnot\aNodeExpressionb}$, respectively meaning the total probability of data-graphs satisfying both formulas, of those satisfying the second but not the first, and so on).
\section{Data cleaning} \label{section:dataCleaning}
We recall that one of the intuitive ideas behind an epistemic model of a graph database
$(\aProbabilisticDatabase,\aRealizationModel)$ is that $\aProbabilisticDatabase$ represents a known distribution of `possible worlds', while $\aRealizationModel$ represents how each of them could be `observed' in a noisy manner with the possible introduction of errors. A probabilistic unclean graph database $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$ adds a concrete observation (namely $\aGraph'$) to our model, allowing us to actually reason about the `underlying reality' given our knowledge of possible worlds and the limits or deficiencies of our observation method. A central question that arises naturally in this interpretation, already mentioned in Example~\ref{example:BibliographicalRIntroducesAuthorship}, is the functional problem of \functionProblem{Data cleaning}, which aims to find the most probable world (a graph database in $\aProbabilisticDatabase$) given our epistemic model of reality and our observation.
More precisely, we define:
\begin{definition} [\functionProblem{Data cleaning}]
We define:
\begin{center}
\fbox{\begin{minipage}{30em}
\textsc{Problem}: Most likely data-graph given $\aRestrictionSet$ and $\aRealizationModel$
\textsc{Input}: A PUGD $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$
\textsc{Output}: A data-graph $I_m \in \aProbabilisticDatabase$ such that the probability of $\aRealizationModel^{-1}(\aGraph')(I_m)$ is maximum.
\end{minipage}}
\end{center}
\end{definition}
Recall that this data-graph with maximal probability exists because, by definition of $\aRealizationModel$, for a fixed $\aGraph'$, $\aRealizationModel^{-1}(\aGraph')$ is a Probabilistic Data-graph (and therefore there is a finite number of graphs with probability greater than any $\epsilon>0$).
We also define a decision problem related to data cleaning, \decisionProblem{data cleaning lower bound}:
\begin{center}
\fbox{\begin{minipage}{30em}
\textsc{Problem}: Is there a sufficiently likely clean data-graph given an unclean observation?
\textsc{Input}: A PUGD $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$ and a bound $b \in [0,1] \cap \Q$
\textsc{Output}: Decide whether there exists a data-graph $I \in \aProbabilisticDatabase$ such that $\aRealizationModel^{-1}(\aGraph')(I) > b$.
\end{minipage}}
\end{center}
Note that an algorithm for solving the data cleaning problem can be used to solve the decision version by checking whether the data-graph $I_m$ found by the algorithm satisfies $\aProbabilisticDatabase(I_m) \times \aRealizationModel(I_m)(\aGraph') > b$.
\subsection{Data cleaning in subset-PUGD}
Suppose that we have a subset-PUGD $\aPUGD = (\aProbabilisticDatabase, \aRealizationModel, \aGraph')$ and we want to find the data-graph $I_m \in \aProbabilisticDatabase$ that maximizes $\aRealizationModel^{-1}(\aGraph')(I_m)$. Based on the definition of $\aRealizationModel$ we know that the set of data-graphs $I$ such that $\aRealizationModel(I)(\aGraph')>0$ is finite (and noted as $\aRealizationModel^{-1}(\aGraph')$). Nevertheless, notice that this set could be of any arbitrary size, so generating it could be expensive. Therefore we explore some conditions under which we can bound the size of this set.
To begin with, let us consider that there are no node deletions. This implies that the preimage of $\aGraph'$ under $\aRealizationModel$ is a subset of the set of data-graphs that contains the same set of nodes of $\aGraph'$, but may have more edges. Since $|\Sigma_e|$ is finite and fixed, the number of possible graphs in $\aRealizationModel^{-1}(\aGraph')$ is bounded by $(2^{|\Sigma_e|})^{n^2}$. This bound is for the worst possible case, when $\aGraph'$ does not have any edges, and therefore we need to consider all the graphs obtained by adding some set of edges to every possible pair of nodes from $\aGraph'$. For each possible pair ($n^2$ pairs) there are $2^{|\Sigma_e|}$ possible subsets of edge labels to pick, and each election is independent from the others.
We prove that this restricted version of the input already makes the decision problem intractable:
\begin{theorem}
\label{teo:subsetNpComplete}
The problem \decisionProblem{data cleaning lower bound} is \textsc{NP-complete} when considering no node deletions and a $\Sigma$-uniform $\aRealizationModel$ such that $\aRealizationModel^{-1}$ is efficiently computable.
\end{theorem}
\begin{proof}
The problem is in \textsc{NP}: when given a certificate data-graph $I$ it suffices to check that $\aRealizationModel^{-1}(\aGraph')(I) > b$. This last inequality can be checked in $poly(|I|+|\aGraph'|)$ by definition of $\aRealizationModel^{-1}$, and given the restrictions on $\aRealizationModel$ of not allowing node deletions we know that any valid certificate will satisfy $|I| = O(poly(\aGraph'))$.
For the hardness, let us reduce \decisionProblem{SAT} to the problem of \decisionProblem{data cleaning lower bound} with $|\Sigma_e|>1$ and a $\Sigma$-uniform $\aRealizationModel$. Given a formula $\aFormula$ of $n$ variables $x_i$ (that can be evaluated in $poly(|\aFormula|)$ given the $n$ assignments of its variables) we will define a PUGD $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$ and a bound $b$ such that $\aFormula$ is satisfiable if and only if there exists a data-graph $I_m$ such that $\aRealizationModel^{-1}(\aGraph')(I_m) > b$. The PUGD $\aPUGD$ and the bound $b$ will be computable in $poly(|\aFormula|)$ time.
For this proof, consider $\{e_1,e_2\} \subseteq \Sigma_e$ and $\{c\} \subseteq \Sigma_n$. We prove the hardness of the problem only when there are at least two different edge labels.
We define $\aGraph'$ as the data-graph with $n+2$ nodes with data value $c$ and no edges. That is, $\aGraph' = (\{v_i$ $|$ $1\leq i \leq n\} \cup \{\top, null\}, L, D)$ where $L(w,z) = \emptyset$ and $D(w) = c$ $\forall w, z \in V_{\aGraph'}$.
The function $\aProbabilisticDatabase$ is defined in the following way:
\[ \mathcal{I}(I) = \begin{cases}
\frac{1+eval(I)}{2^{n+1}} & \text{if }valid(I,e_1) \\
\frac{1-eval(I)}{2^{n+1}} & \text{if }valid(I,e_2) \\
0 & \text{otherwise}.
\end{cases}
\]
The predicate $valid(I,e)$ is true when $V_I = V_{\aGraph'}$, there is an edge $(\top,e,null)$ and all the rest of the edges are of the form $(v_i,e,\top)$ for some $1\leq i\leq n$. The expression $eval(I)$ evaluates to 1 if the formula $\aFormula$ is satisfied by the assignment to the variables $x_i$ defined through the graph $I$, and to 0 otherwise. The variable $x_i$ is set as $\top$ if $(v_i,e,\top) \in E_I$ for some $e \in \{e_1,e_2\}$. Note that $\aProbabilisticDatabase$ is a probabilistic database:
\begin{align*}
\sum_{I \in \aProbabilisticDatabase} \aProbabilisticDatabase(I) &=
\sum_{e \in \{e_1,e_2\}} \sum_{I : valid(I,e)} \aProbabilisticDatabase(I) =
\sum_{I:valid(I,e_1)} \left(\frac{1+eval(I)}{2^{n+1}} + \frac{1-eval(I)}{2^{n+1}} \right)
\\
&=\sum_{I:valid(I,e_1)} \frac{2}{2^{n+1}} = 2^n (\frac{1}{2^n}) = 1.
\end{align*}
The second equality follows because there is a data-graph $I$ that satisfies $valid(I,e_1)$ if and only if there exists a data-graph $I'$ that satisfies $valid(I',e_2)$. More precisely, the edges of these two data-graphs coincide except for the edge labels (that is, they have the same `topology' if we ignore the labels of the edges) and so there is a bijection between the set of data-graphs that satisfy $valid(I,e_1)$ and $valid(I,e_2)$, and these sets are disjoint.\footnote{Observe that if we do not include the edge between $\top$ and $null$ this sets would not be disjoint because of the data-graph $I = \aGraph'$, which would map into itself through the funcion that changes edge labels.}.
We now define $b = \frac{1}{2^{n+1}}$ and $\aRealizationModel(I)(\aGraph') = 1$ for every data-graph $I$. Notice that this $\aRealizationModel$ is $\Sigma$-uniform, and that $\aRealizationModel^{-1}(\aGraph')(I)$ can be computed efficiently. This follows since:
$$
\aRealizationModel^{-1}(\aGraph')(I) = \frac{\aProbabilisticDatabase(I) \times \aRealizationModel(I)(\aGraph')}{\sum \aProbabilisticDatabase(I') \times \aRealizationModel(I')(\aGraph')} = \frac{\aProbabilisticDatabase(I)}{\sum \aProbabilisticDatabase(I')} = \aProbabilisticDatabase(I)
$$
Next, we prove that $\aFormula$ is satisfiable if and only if there exists a graph $I \in \aProbabilisticDatabase$ such that $\aRealizationModel^{-1}(\aGraph')(I) > b$.
$\implies$) Let $\nu:var(\aFormula) \to \{\top,\bot\}$ be a valuation over the variables of $\aFormula$ that satisfies $\aFormula$.
We define the data-graph $I = (\{v_i$ $|$ $1\leq i \leq n\} \cup \{\top\}, L, D)$ where, if $\nu(x_i) = \top$, then $D(v_i)=c$ and $L(v_i,\top) = \{e_1\}$ for each $i=1, \ldots, n$.
Otherwise, $L(w,z) = \emptyset$ for every pair $w,z \in V_I$.
It follows from the previous discussion that $\aRealizationModel^{-1}(\aGraph')(I) = \frac{1+1}{2^{n+1}} \times 1 = \frac{1}{2^n} > \frac{1}{2^{n+1}} = b$.
$\impliedby$) Let $I$ be a data-graph $I \in \aProbabilisticDatabase$ that satisfies $\aRealizationModel^{-1}(\aGraph')(I) > \frac{1}{2^{n+1}}$. Hence, every edge label of $I$ must be $e_1$. We can define a valuation $\nu$ over $var(\aFormula)$ based on $I$ as:
$$\nu(x_i) = \top \iff L(v_i,\top) = \{e_1\}$$
for $1\leq i \leq n$. Since $I$ satisfies $eval(I)$ (otherwise it is not possible that $\aRealizationModel^{-1}(\aGraph')(\aProbabilisticDatabase) > b$), we know that $\nu$ satisfies $\aFormula$.
\end{proof}
We can obtain a polynomial bound on the cardinality of $\aRealizationModel^{-1}(\aGraph')$ by bounding the number of edges that can be added by a constant $k_e$. That is, we assume that $\aRealizationModel(I)(\aGraph') = 0$ if $|E(I) \setminus E(\aGraph')| > k_e$. Taking this restriction into account, we can show that $|\aRealizationModel^{-1}(\aGraph')|$ is bounded by a polynomial on $n$:
$$
|\aRealizationModel^{-1}(\aGraph')| = \sum_{\substack{i=0 \\ i\leq missing(\aGraph')}}^{k_e} \binom{missing(\aGraph')}{i} \leq k_e (missing(\aGraph'))^{k_e} = O (n^{2 k_e}),
$$
where $missing(\aGraph') = |\Sigma_e| n^2 - |E_{\aGraph'}|$ counts the number of edges missing from $\aGraph'$, and therefore the ones that could have been subtracted.
\begin{theorem}
\decisionProblem{Data cleaning} of subset-PUGD can be done in polynomial time in the size of the observed data-graph $\aGraph'$ if there are no node deletions (i.e. $\aSubsetRealizationModelNode(c) = 0$ $ \forall c \in \Sigma_n$) and the number of edges deleted is bounded by a constant.
\end{theorem}
Consider now the subset-PUGD problem allowing also node deletions.
Recall that $|\Sigma_e|$ is finite and fixed. Given that data-graphs may model problems where the data labels for each node may be countable, we cannot assume that $|\Sigma_n|$ is finite. Suppose first that, even though $\Sigma_n$ may be countable, the possible data labels for each node is uniformly bounded by some constant $\delta$. In other words, the data labels may be infinite, but each node only allows a finite number of options, and this options are bounded by the same constant $\delta$.
Furthermore, suppose as in the previous case that the numbers of edges deleted by $\aRealizationModel$ is bounded by $k_e$, and also consider the number of nodes deleted is bounded by $k_n$. This is, both the number of edges and nodes that can be deleted to obtain $\aGraph'$ are bounded by constants.
\begin{theorem}
Let only a constant number of node deletions $k_n$ and edge deletions $k_e$ are allowed and the number of valid data labels permitted for each node is uniformly bounded by a constant $\delta$, then \emph{data cleaning of subset-PUGD} can be done in polynomial time in the data-bound $\delta$.
\end{theorem}
\begin{proof}
The number of graphs that can be obtained by adding $i$ nodes and at most $k_e$ edges to $\aGraph'$ is:
\[
h(i,\aGraph',k_e)
=
\delta^i \sum_{j=0}^{\min(k_e, f(G',i))} \binom{f(\aGraph',i)}{j}
\]
\[
\leq
\delta^i k_e (f(\aGraph',i))^{k_e}
\]
\[
\leq
\delta^i k_e(|\Sigma_e|(n+i)^2)^{k_e},
\]
where
\[
f(\aGraph',i) = |\Sigma_e|(n+i)^2 - |E(\aGraph')| \leq |\Sigma_e|(n+i)^2.
\]
denotes the maximum number of edges that can be added to $\aGraph$.
In the formula $\delta^i$ is used to count all the possible data values that we can assign to the nodes added. The sum then counts all the possible ways to select at most $k_e$ edges from those not present in the graph defined by $\aGraph'$ and the new nodes added. Now, we sum over all possible values of $i$ (that is, $0 \leq i \leq k_n)$ to obtain a bound on the size of the preimage of $\aRealizationModel$ over $\aGraph'$:
\begin{align*}
|\aRealizationModel^{-1}(\aGraph')|
&\leq \sum_{i=0}^{k_n} h(i,\aGraph',k_e)
\leq
\sum_{i=0}^{k_n} \delta^i k_e (|\Sigma_e|(n+i)^2)^{k_e} \\
& \leq
\delta^{k_n} k_e(|\Sigma_e|(n+k_n)^2)^{k_e} \sum_{i=0}^{k_n} 1 \\
&=\delta^{k_n} k_e(|\Sigma_e|(n+k_n)^2)^{k_e} (k_n+1)
=
\mathcal{O}(n^{2k_e})
\end{align*}
Therefore we can do an exhaustive search in polynomial time.
\end{proof}
\subsection{Data cleaning in superset-PUGD}
Let us consider the problem of superset-PUGD version of the Data Cleaning problem. This is, given a superset-PUGD $\aPUGD = (\aProbabilisticDatabase, \aRealizationModel, \aGraph')$, we want to find the data-graph $I_m \in \aProbabilisticDatabase$ that maximizes $\aRealizationModel^{-1}(\aGraph')(I)$.
Recall that $\aGraph'$ is the observed unclean data-graph that may have loops and as many edges as distinct possible labels. Notice that, in this case, and disregarding whether we consider node and/or edge deletions, every graph in $I_m \in \aRealizationModel^{-1}(\aGraph')$ is necessarily a subgraph of $\aGraph'$. Therefore, the set of data-graphs $\aRealizationModel^{-1}(\aGraph')$ is finite, and its size is bounded by $\sum_{i=0}^n \binom{n}{i} 2^{i^{2}\times|\Sigma_e|} \leq 2^{n^{2}\times|\Sigma_e|+n}$ in the worst case scenario.
Moreover, if we only consider edge deletions, then $|\aRealizationModel^{-1}(\aGraph')|$ is bounded by $2^{n^{2}\times|\Sigma_e|}$.
The bound on the size of $\aRealizationModel^{-1}(\aGraph')$ depends solely on how many nodes and distinct labels the observed unclean data-graph has, although this quantity may be --at first glance-- exponential.
Since every data-graph from $\aRealizationModel^{-1}(\aGraph')$ has its size bounded by $|\aGraph'|$ we can conclude that:
\begin{theorem}
If $\aRealizationModel^{-1}(\aGraph')(I)$ can be computed efficiently, then the superset-PUGD Data Cleaning bound problem is in \textsc{NP}.
\end{theorem}
\begin{proof}
Given a bound $b$, if there exists a graph $I$ such that $b > \aRealizationModel^{-1}(\aGraph')(I)$, then $|I| \leq \aGraph'$, and therefore our algorithm only has to guess $I$ and compute $\aRealizationModel^{-1}(\aGraph')(I)$.
\end{proof}
This complexity bound is better than the one from the subset-PUGD case because we did not have to bound the size of the preimage of $\aRealizationModel^{-1}$ \footnote{It is actually bounded by the problem definition itself}.
We can get a lower bound in a similar restricted scenario as in the subset case:
\begin{theorem}
The problem is \textsc{NP-complete} when $\aRealizationModel^{-1}$ is efficiently computable, node additions are not allowed and $\aRealizationModel$ is $\Sigma$-uniform.
\end{theorem}
\begin{proof}
Since node additions are not allowed, the preimage of an observed graph $\aGraph$ through $\aRealizationModel$ has to be a subset of the graphs built with the same set of nodes from $\aGraph'$ and a subset of its edges.
As in the subset case, we will reduce 3-SAT to the superset-PUGD data cleaning problem under the stated conditions. Given a 3-SAT formula $\phi$ of $n$ variables and $m$ clauses we define the PUGD $\aPUGD = (\aProbabilisticDatabase, \aRealizationModel, (V,L,D))$ and the bound $b$ in the following way:
\begin{align*}
& V = \{x_i : 1 \leq i \leq n\} \cup \{\top,null\}\\
& L(x_i,\top) = \{e_1,e_2\} \text{ for } 1\leq i \leq n \\
& L(v,w) = \emptyset \text{ for any other case } \\
& D(v) = c \text{ for any } v \in V \\
& \aRealizationModel(I)(\aGraph') = 1 \text{ for any } I \subseteq \aGraph \\
& b = \frac{1}{2^{n+1}}
\end{align*}
Finally, the function $\aProbabilisticDatabase$ is defined as in Theorem~\ref{teo:subsetNpComplete} as:
\[ \mathcal{I}(I) = \begin{cases}
\frac{1+eval(I)}{2^{n+1}} & \text{if }valid(I,e_1) \\
\frac{1-eval(I)}{2^{n+1}} & \text{if }valid(I,e_2) \\
0 & \text{otherwise}.
\end{cases}
\]
It follows analogously as in the subset-PUGD case that any graph $I$ that satisfies $\aRealizationModel^{-1}(\aGraph')(I) > b$ encodes an assignment of the $n$ variables from $\phi$ that satisfies the formula (and vice versa). Also note that the construction of $\aPUGD$ and $b$ can be done in time $poly(n)$.
\end{proof}
\section{Definitions}\label{Section:Definitions}
Fix a finite set of edge labels $\Sigma_e$ and a countable set (either finite or infinite enumerable) of data values $\Sigma_n$ (sometimes called data labels), which we assume non-empty and with $\Sigma_e \cap \Sigma_n = \emptyset$. A (finite) \defstyle{data-graph} $G$ is a tuple $(V,L_e,\dataFunction)$ where $V$ is a finite set of nodes, $L_e$ is a mapping from $V \times V$ to $\parts{\Sigma_e}$ defining the edges of the graph, and $\dataFunction$ is a function mapping the nodes from $V$ to data values from $\Sigma_n$. We note as $E(\aGraph)$ the set $\{(v,e,w)$ $|$ $v,w \in V_\aGraph, e \in L_\aGraph(v,w)\}$, which can be understood as the set of edges of the graph.
\begin{definition}
We say that $\aProbabilisticDatabase$ is a \defstyle{probabilistic data-graph} if it is a probability distribution over a set of data-graphs.
In particular, we will consider
$\aProbabilisticDatabase: \aUniverseOfGraphs \to [0,1]$, where $\aUniverseOfGraphs$ is the set of all data-graphs over $\Sigma_e, \Sigma_n$, and $$\sum_{\aGraph \in \aUniverseOfGraphs}\aProbabilisticDatabase(\aGraph) = 1.$$
For simplicity, we will sometimes abuse the notation and write $\aGraph \in \aProbabilisticDatabase$ to refer to $\aGraph \in \aUniverseOfGraphs$ such that $\aProbabilisticDatabase(\aGraph)>0$
\end{definition}
In Figure~\ref{figure:exampleDataGraph} we show an example of a data-graph with 4 nodes $V = \{v_1,v_2,v_3,v_4\}$ where the data values and the edges are:
\noindent
\vspace{-0.3cm}
\begin{center}
\begin{minipage}[t]{.25\textwidth}
\raggedright
\begin{flalign*}
D(v_1) &= \esDato{Alice}&&\\
D(v_2) &= \esDato{Dave}&&\\
D(v_3) &= \esDato{Carl}&&\\
D(v_4) &= \esDato{Bob} &&
\end{flalign*}
\end{minipage}
\hspace{1cm}
\noindent
\begin{minipage}[t]{.25\textwidth}
\raggedleft
\begin{align*}
L(v_1,v_1) &= \emptyset \\
L(v_1,v_2) &= \{\esLabel{friend}\} = L(v_2, v_1) \\
L(v_1,v_4) &= \{\esLabel{blocked}\} \\
L(v_4,v_1) &= \{\esLabel{follows}\} \\
& \vdots
\end{align*}
\end{minipage}
\end{center}
\input{image_example-graph.tex}
The probability distribution $\aProbabilisticDatabase$ can represent different types of uncertainty, such as the details of a concrete database, our priors about the relative frequency of different databases within a certain class, etc. $\aProbabilisticDatabase(G)$ denotes the probability that the state of an uncertain data-graph is precisely $\aGraph$. This model is actually a generalization of the relational \textit{Tuple-Independent probabilistic Database} (TID), where each tuple might either exist or not based upon an associated probability. Given $\aProbabilisticDatabase$ of finite support, we can derive the probability of a certain fact being true over $\aProbabilisticDatabase$ by simply evaluating the fact in each data-graph $\aGraph$ from $\aProbabilisticDatabase$ and then multiplying by the probability $\aProbabilisticDatabase(\aGraph)$.
\begin{definition} [EMGD]
We define a \defstyle{Epistemic Model of a Graph Database (EMGD)} as a pair ($\aProbabilisticDatabase,\aRealizationModel$), where:
\begin{itemize}
\item \aProbabilisticDatabase is a probabilistic data-graph (over the universe of data-graphs $\aUniverseOfGraphs$ given by $\Sigma_e$ and $\Sigma_n$).
\item \aRealizationModel, referred as the \defstyle{noisy observer} (or also as the \defstyle{realization model}), is a function which maps data-graphs into probabilistic data-graphs such that for any $(\Sigma_e,\Sigma_n)$ data-graph $\aGraph'$, if there exists $\aGraph \in \aProbabilisticDatabase$ such that $\aRealizationModel(\aGraph)(\aGraph')>0$ then:
$$\aRealizationModel^{-1}(\aGraph') \eqdef \{\aGraph'' \in \aProbabilisticDatabase \mid \aRealizationModel(\aGraph'')(\aGraph') >0) \} \mbox{ is finite.}$$
In other words, for any data-graph of positive probability, its preimage via $\aRealizationModel$ is finite.
\end{itemize}
Greatly abusing the notation for convenience and readability, we will also notate by \defstyle{$\aRealizationModel^{-1}$} the
function whose domain is $\aUniverseOfGraphs$ and whose image for each data-graph $\aGraph'$ is a particular a probabilistic data-graph $\aProbabilisticDatabase_{\aGraph'}$. This $\aProbabilisticDatabase_{\aGraph'}$ is such that $\aProbabilisticDatabase_{\aGraph'}(\aGraph) = \aRealizationModel^{-1}(\aGraph')(\aGraph)$, which returns the probability that the data-graph $\aGraph$ was the clean data-graph which transformed via the noisy observer $\aRealizationModel$ into the observed $\aGraph'$ (considering implicitly the effect of $\aProbabilisticDatabase$). More precisely, using the intuition of Bayes' theorem:
$$\aRealizationModel^{-1}(\aGraph')(\aGraph) \eqdef \frac{\aRealizationModel(\aGraph)(\aGraph') \times \aProbabilisticDatabase(\aGraph)}{\sum_{\aGraphb \in \aProbabilisticDatabase} (\aProbabilisticDatabase(\aGraphb)\times \aRealizationModel(\aGraphb)(\aGraph'))}.$$
\end{definition}
\begin{remark}
$\aRealizationModel$ does not necessarily depend on $\aProbabilisticDatabase$ by definition, but in practical use cases it will probably do. Meanwhile, $\aRealizationModel^{-1}$ depends profoundly on $\aProbabilisticDatabase$ and $\aRealizationModel$, as $\aRealizationModel^{-1}$ depends on \emphstyle{all} data-graphs that have non-zero probability over $\aProbabilisticDatabase$.
\end{remark}
As is usually the case, we are going to assume that $\aProbabilisticDatabase$, $\aRealizationModel$, and $\aRealizationModel^{-1}$ are all computable functions over adequate representations of their domains and codomains.
When proving complexity bounds for problems related to EMGDs (or rather PUGDs, defined later in this work) we will be concerned with the set of probabilistic databases and realization models that can be computed efficiently. This means that, given $\aGraph$, we consider $\aProbabilisticDatabase$'s and $\aRealizationModel$'s such that we can compute $\aProbabilisticDatabase(\aGraph)$ and $\aRealizationModel(\aGraph)$ in time $poly(|\aGraph|)$. We also ask for the set $\aRealizationModel^{-1}(\aGraph')$ to be computable in polynomial time given its input and output: that is, we can compute it in $poly(|\aGraph'| +|\aRealizationModel^{-1}(\aGraph')|)$
We also note that defining an arbitrary efficiently computable $\aProbabilisticDatabase$ and $\aRealizationModel$ does not imply that $\aRealizationModel^{-1}(\aGraph')(\aGraph)$ can be computed in polynomial time. Nonetheless, in most proofs shown here this happens because the normalization factor (the denominator in the definition of $\aRealizationModel^{-1}(\aGraph')(\aGraph)$) can be deduced analytically, and does not need to be computed by enumerating all the graphs from $\aRealizationModel^{-1}(\aGraph')$.
We can think of $\aProbabilisticDatabase$ as a collection of guesses over the real state of a database, and $\aRealizationModel$ as a noisy process which allows us to `observe' a potentially distorted version of each of them; each possibility in $\aProbabilisticDatabase$ is transformed via $\aRealizationModel$ into many, each of them with their particular probability of occurrence.
\begin{example}\label{example:toySocialNetworkAndNoisyObserverOverGraph}
A particular example of the previous possible interpretation of an EMGD is shown in Figure~\ref{figure:exampleARealizationModel}, where a toy social network is incompletely observed via the `deletion' of a single edge. Here, the data values (in $\Sigma_n$) are name identifiers $\esDato{Alice}, \esDato{Bob}, \esDato{Carl}$, and the possible edge-labels (in $\Sigma_e$) are $\esLabel{friend}$ and $\esLabel{follows}$. In this case, one possible action of $\aRealizationModel$ over the data-graph $\aGraph$ is leaving it unchanged, while the other shown possibility is the deletion of a $\esLabel{friend}$ edge in the direction from Bob to Alice. Assuming that the representation in the figure is complete, we have that $\aRealizationModel(\aGraph)(\aGraph) + \aRealizationModel(\aGraph)(\aGraph') = 1$ (as $\aRealizationModel(\aGraph)$ is a probabilistic data-graph).
\input{figure_exampleARealizationModel}
\end{example}
Another possible interpretation for EMGDs is that $\aProbabilisticDatabase$ represents idealized databases over certain domain, and $\aRealizationModel$ transforms each idealized version into many possible error-prone implementations. For example, $\aProbabilisticDatabase$ could represent bibliographical databases, and $\aRealizationModel$ could produce databases where some type of data-duplication errors have been made.
Following \cite{rekatsinas2017holoclean}, we define an unclean graph database in the following way:
\begin{definition} [PUGD]
We define a \defstyle{Probabilistic Unclean Graph Database (PUGD)} as a triple $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$, where:
\begin{itemize}
\item ($\aProbabilisticDatabase,\aRealizationModel$) is a EMGD.
\item $\aGraph'$ is the observed or unclean data-graph.
\end{itemize}
\end{definition}
Intuitively, $\aGraph'$ is the current observable state of the database, which we think is not correct because of errors introduced due to data entry mistakes, data integration problems or other kind of operations that could have polluted the original data-graph~$\aGraph$. We have some knowledge of which was the original state of the data-graph prior to these issues, which is modeled by the distribution $\aProbabilisticDatabase$; and we also have some idea of the possible ways these original data-graphs from $\aProbabilisticDatabase$ could have become unclean, which is modeled by $\aRealizationModel$.
\begin{example}\label{example:SocialNetwork}
Let $\Sigma_e = \{ \esLabel{friend\_of} \}$, $\Sigma_n$ be a set of common names.
Consider $S$ the set of finite data-graphs over $\Sigma_e, \Sigma_n$ that have at least 2 nodes and no more than $10^7$ nodes.
Let $\aProbabilisticDatabase : \aUniverseOfGraphs \to [0,1]$ be a uniform distribution of probabilities among all graphs of $S$: having $\aProbabilisticDatabase(\aGraph) = 0$ if $\aGraph \in \aUniverseOfGraphs \backslash S$, and otherwise with $\aProbabilisticDatabase(\aGraph)$ given by some prior over the typical form of social networks.
We can use $\aRealizationModel$ to model the probability of erroneously removing a friendship relation between two people. For example, and without entering into the details, it could be that the probability of this error is a fixed number $0<p<1$ for each edge, and it is independent for each edge; therefore, bigger databases would more often be observed with more errors (that is, more errors will probably be introduced when starting from a larger clean database and `observing' it with $\aRealizationModel$).
In this context, the `unclean' observation of a graph $\aGraph'$, the image via $\aRealizationModel$ of some $\aGraph \in \aProbabilisticDatabase$, gives us valuable information.
\end{example}
A natural problem in the context of the previous example would be that of trying to deduce the original `clean' social graph $\aGraph$ that derived in our observation. But unless $\aProbabilisticDatabase$ or $\aGraph'$ have a very peculiar form, it would usually be impossible to determine this with certainty, as there are many graphs which can lead to the same one via the removal of edges. However, we can determine the most probable candidate. We will explore this problem, called `data cleaning', in Section~\ref{section:dataCleaning}.
However, one must ask about the meaning of later conducting a query over this `clean' candidate database, and how to measure the uncertainty of this query given that we do not know whether graph found by data cleaning was the `original' one. The approach used to better quantify this uncertainty corresponds to detecting all graphs that could possibly have resulted in our observation, and analyzing the query over them, aggregating the results by giving more weight to those graphs that are more probable given our observation. This is the approach we call Probabilistic Query Answering, which we will explore in Section~\ref{section:PQA}.
\begin{example}\label{example:BibliographicalRIntroducesAuthorship}
Let $\Sigma_e = \{\esLabel{author\_of}, \esLabel{name}, \esLabel{title}\}$, and let $\Sigma_n$ be the set of words with spaces over the English alphabet.
We can consider an
EMGD $(\aProbabilisticDatabase,\aRealizationModel)$, where $\aProbabilisticDatabase$ assigns non-zero probability to all data-graphs that form a viable representation of bibliographical data, and where $\aRealizationModel$ can introduce errors by adding outgoing $\esLabel{author\_of}$ edges from any node that has no incoming edges. For example, $\aRealizationModel$ could erroneously assign co-authorship of the same book, or it could (in a more blatant error) introduce authorship between a person an a name, or between a person and another person.
If we have a PUGD $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$, the observed $\aGraph'$ might not coincide with the initial clean data-graph, but, if we want to answer this question (see Data-cleaning in Section~\ref{section:dataCleaning}), the nature of $\aRealizationModel$ lets us know that the true answer must be a subgraph of $\aGraph'$ where the only changes, if any, are the removal of $\esLabel{author\_of}$ edges. Furthermore, the information on $\aProbabilisticDatabase$ can also be used: for example, we automatically know that any $\esLabel{author\_of}$ edges between different authors are an error introduced by $\aRealizationModel$.
\end{example}
In principle, $\aRealizationModel$ can represent any possible transformation between data-graphs. In order to restrict the generality of $\aRealizationModel$ while still preserving a good amount of expressiveness useful for real applications we will consider realization models that either only delete or only add data in the data-graphs (either nodes or edges).
These two notions can be related to two common semantics for repairing databases: the superset-based and subset-based repairs. Considering any of these two restrictions is an usual practice when reasoning over uncertainty or inconsistencies \cite{afrati2009repair,barcelo2017data, chomicki2005minimal, Cate:2015}.
In order to formalize the first notion, we can define the \defstyle{{subset}-EMGD}:
\begin{definition}
A \defstyle{subset-EMGD} is a pair $(\aProbabilisticDatabase, \aRealizationModel)$ such that the noisy observer $\aRealizationModel$ satisfies that $\aGraph' \subseteq \aGraph$ for every $\aGraph'$ such that $\aRealizationModel(\aGraph)(\aGraph') > 0$.
\end{definition}
In a way, we can think that the $\aRealizationModel$ of a subset-EMGD makes node or edge deletions.
However, using this definition implies some difficulties when facing the complexity of reasoning around unclean databases.
Mainly, that $\aRealizationModel(\aGraph)$ could depend arbitrarily on $\aGraph$, potentially making $\aRealizationModel$ quite complex.
In order to avoid this problem, we are going to consider a simplified version. We first consider $\aRealizationModel$ such that the probabilities it assigns depend only on the data values and labels in the original graph and those that are removed. That is:
\begin{definition} \label{def:noisyObserverMultisetsLabels}
We say that $(\aProbabilisticDatabase, \aRealizationModel)$ (or simply $\aRealizationModel$) is \defstyle{subset-$\Sigma$-uniform} if $(\aProbabilisticDatabase, \aRealizationModel)$ is a subset-EMGD and it satisfies that given $\aGraph, \aGraph', \aGraphb, \aGraphb'$, if:
\begin{itemize}
\item the multiset of data values of $V_{\aGraph}$ and of $V_{\aGraphb}$ are the same, and the multiset of labels of $Im(L_{\aGraph})$ and $Im(L_{\aGraphb})$ are the same, and
\item the multiset of labels of $Im(L_{\aGraph}) \backslash Im(L_{\aGraph'})$ is equal to that of $Im(L_{\aGraphb}) \backslash Im(L_{\aGraphb'})$, and
\item the multiset of data values from $V_{\aGraph} \backslash V_{\aGraph'}$ is equal to that of $V_\aGraphb \backslash V_{\aGraphb'}$.
\end{itemize}
then $\aRealizationModel(\aGraph)(\aGraph') = \aRealizationModel(\aGraph)(\aGraphb') = \aRealizationModel(\aGraphb)(\aGraph') = \aRealizationModel(\aGraphb)(\aGraphb')$.
\end{definition}
These conditions put serious restrictions on the potential of $\aRealizationModel$. Observe, for example, that we lose the ability to select interesting topology conditions based on paths or certain subgraphs. Nonetheless, they will be useful for studying lower bounds of the problems we define later.
Another problem is related with the size of the data-graphs in $\aRealizationModel(\aGraph)$. Even though we asked for this set to be finite, the data-graphs inside it could be of any arbitrary size. An easy solution to this would be to restrict the number of elements that $\aRealizationModel$ can delete by some value that may or may not depend on the observed graph $\aGraph'$. For example, we could consider the subset realization models $\aRealizationModel$ such that $\aRealizationModel(\aGraph)(\aGraph')>0$ implies\footnote{By $|\aGraph \setminus \aGraph'|$ we refer to the sum of $|V_{\aGraph} \setminus V_{\aGraph'}|$ and $|Im(L_{\aGraph})\setminus Im(L_{\aGraph'})|$.} that $|\aGraph \setminus \aGraph'| < f(|\aGraph'|)$ for some function $f$. As we will see, the complexity of reasoning around PUGDs is naturally related to the bound $f$ we use.
To formalize the second previously discussed notion, we define \defstyle{superset-EMGD} as follows:
\begin{definition}
A \defstyle{superset-EMGD} is a pair $(\aProbabilisticDatabase, \aRealizationModel)$ such that the noisy observer $\aRealizationModel$ satisfies that $\aGraph' \supseteq \aGraph$ for every $\aGraph'$ such that $\aRealizationModel(\aGraph)(\aGraph') > 0$.
\end{definition}
\begin{definition} \label{def:noisyObserverMultisetsLabelsSupersetVersion}
We say that $(\aProbabilisticDatabase, \aRealizationModel)$ (or simply $\aRealizationModel$) is \defstyle{superset-$\Sigma$-uniform} if $(\aProbabilisticDatabase, \aRealizationModel)$ is a superset-EMGD and it satisfies that given $\aGraph, \aGraph', \aGraphb, \aGraphb'$, if:
\begin{itemize}
\item the multiset of data values of $V_{\aGraph}$ and of $V_{\aGraphb}$ are the same, and the multiset of labels of $Im(L_{\aGraph})$ and $Im(L_{\aGraphb})$ are the same, and
\item the multiset of labels of $Im(L_{\aGraph'}) \backslash Im(L_{\aGraph})$ is equal to that of $Im(L_{\aGraphb'}) \backslash Im(L_{\aGraphb})$, and
\item the multiset of data values from $V_{\aGraph'} \backslash V_{\aGraph}$ is equal to that of $V_\aGraphb' \backslash V_{\aGraphb}$.
\end{itemize}
then $\aRealizationModel(\aGraph)(\aGraph') = \aRealizationModel(\aGraph)(\aGraphb') = \aRealizationModel(\aGraphb)(\aGraph') = \aRealizationModel(\aGraphb)(\aGraphb')$.
\end{definition}
In other words, in this case the noisy observer can add nodes or edges to the original clean graph.
\begin{remark}
Observe that if $\aRealizationModel$ is superset-$\Sigma$-uniform, then $\aRealizationModel^{-1}(\aGraph')$ is always finite. Indeed, that preimage consists of all subgraphs of $\aGraph'$, which is a finite set.
\end{remark}
It is common to expect databases to follow some semantic structure related to the world they are representing. This structure can be enforced by defining integrity constraints that limit the shape and data contained in the database. In the case of data-graphs, a common way of defining these constraints is through \textit{path constraints} defined in some specific language. Given some constraint $\aFormula$ we would want our database to be consistent with respect to $\aFormula$, given some definition of consistency based on the semantics of $\aFormula$.
In this work queries will be defined with the \Gregxpath language. These formulas can capture both nodes and pair of nodes based upon expressions that define paths along the graph with the common regular expression operators, as well as having means of comparing data values and some extra utilities.
\textit{Path expressions} of \Gregxpath are given by:
\begin{center}
$\aPath, \aPathb$ = $\epsilon$ $|$ $\labelComodin$ $|$ $\aLabel$ $|$ $\aLabel^{-}$ $|$ $\expNodoEnCamino{\aFormula}$ $|$ $\aPath$ . $\aPathb$ $|$ $\aPath \pathUnion \aPathb$ $|$ $\aPath \pathIntersection \aPathb$ $|$ $\aPath^{*}$ $|$ $\pathComplement{\aPath}$ $|$
$\aPath^{n,m}$
\end{center}
where $\aLabel$ iterates over all labels from $\Sigma_e$ and $\aFormula$ is a \textit{node expression} defined by the following grammar:
\begin{center}
$\aFormula, \aFormulab$ = $\neg \aFormula \mid \aFormula \wedge \aFormulab$ $|$ $\comparacionCaminos{\aPath}$ $|$ $\esDatoIgual{\aData}$ $|$ $\esDatoDistinto{\aData}$ $|$ $\comparacionCaminos{\aPath = \aPathb}$ $|$ $\comparacionCaminos{\aPath \neq \aPathb}$ $|$
$\aFormula \vee \aFormulab$
\end{center}
where $\aPath$ and $\aPathb$ are path expressions (and so they are defined by mutual recursion) and $c$ iterates over $\Sigma_n$. If we only allow the Kleene star to be applied to labels and their inverses (i.e. $a^-$) we have a subset of \Gregxpath called \Gcorexpath. The semantics of these languages are defined in \cite{libkin2016querying} in a similar fashion as the usual regular languages for navigating graphs \cite{barcelo2013querying}, while adding some extra capabilities such as the complement of a path expression $\pathComplement{\aPath}$ and data tests. The $\comparacionCaminos{\aPath}$ operator is the usual one for \textit{nested regular expressions} or NREs used in \cite{barcelo2012relative}. Given a data-graph $\aGraph=(V,L,D)$, the semantics are:
\begin{itemize}[leftmargin=.6in,label={}]
\itemsep0em
\item $\semantics{\epsilon}_\aGraph = \{(v,v)$ $|$ $v \in V\}$
\item $\semantics{\labelComodin}_\aGraph = \{(v,w)$ $|$ $v,w \in V, L(v,w) \neq \emptyset\}$
\item $\semantics{\aLabel}_\aGraph = \{ (v,w)$ $|$ $\aLabel \in L(v,w)\}$
\item $\semantics{\aLabel^-}_\aGraph = \{ (w,v)$ $|$ $\aLabel \in L(v,w)\}$
\item $\semantics{\aPath^*}_\aGraph = $ the reflexive transitive closure of $\semantics{\aPath}_\aGraph$
\item $\semantics{\aPath . \aPathb}_\aGraph = \semantics{\aPath}_\aGraph \circ \semantics{\aPathb}_\aGraph$
\item $\semantics{\aPath \pathUnion \aPathb}_\aGraph = \semantics{\aPath}_\aGraph \cup \semantics{\aPathb}_\aGraph$
\item $\semantics{\aPath \pathIntersection \aPathb}_\aGraph = \semantics{\aPath}_\aGraph \cap \semantics{\aPathb}_\aGraph$
\item $\semantics{\pathComplement{\aPath}}_\aGraph = V \times V \setminus \semantics{\aPath}_\aGraph$
\item $\semantics{\expNodoEnCamino{\aFormula}}_\aGraph = \{(v,v)$ $|$ $v \in \semantics{\aFormula}_\aGraph$\}
\item $\semantics{\aPath^{n,m}}_\aGraph = \bigcup\limits_{k=n}^m(\semantics{\aPath}_\aGraph)^k$
\item $\semantics{\comparacionCaminos{\aPath}}_\aGraph = \pi_1(\semantics{\aPath}_\aGraph) = \{v \mid \exists w\in V\, (v,w) \in \semantics{\aPath}_\aGraph \} $
\item $\semantics{\lnot \aFormula}_\aGraph = V \setminus \semantics{\aFormula}_\aGraph$
\item $\semantics{\aFormula \wedge \aFormulab}_\aGraph = \semantics{\aFormula}_\aGraph \cap \semantics{\aFormulab}_\aGraph$
\item $\semantics{\aFormula \vee \aFormulab}_\aGraph = \semantics{\aFormula}_\aGraph \cup \semantics{\aFormulab}_\aGraph$
\item $\semantics{\esDatoIgual{c}} = \{v \in V$ $|$ $D(v) = c$\}
\item $\semantics{\esDatoDistinto{c}} = \{v \in V$ $|$ $D(v) \neq c$\}
\item $\semantics{\comparacionCaminos{\aPath = \aPathb}}_\aGraph = \{v \in V$ $|$ $\exists v', v'' \in V$, $(v,v') \in \semantics{\aPath}_\aGraph$, $(v,v'') \in \semantics{\aPathb}_\aGraph$, $D(v') = D(v'')$\}
\item $\semantics{\comparacionCaminos{\aPath \neq \aPathb}}_\aGraph = \{v \in V$ $|$ $\exists v', v'' \in V$, $(v,v') \in \semantics{\aPath}_\aGraph$, $(v,v'') \in \semantics{\aPathb}_\aGraph$, $D(v') \neq D(v'')$\}
\end{itemize}
We use $\aPath \entoncesCamino \aPathb$ to denote the path expression $\aPathb \pathUnion \pathComplement{\aPath}$, and $\aNodeExpression \entoncesNodo \aNodeExpressionb$ to denote the node expression $\aNodeExpressionb \lor \neg \aNodeExpression$. We also note a label $\aLabel$ as $\down_\esLabel{\aLabel}$ in order to easily distinguish the `path' fragment of the expressions. For example, the expression \esLabel{son\_of} [\esDatoIgual{Maria}] \esLabel{sister\_of} will be noted as $\down_\esLabel{son\_of} [\esDatoIgual{Maria}] \down_\esLabel{sister\_of}$.
Naturally, the expression $\aPath \cap \aPathb$ can be rewritten as $\overline{\overline{\aPath} \cup \overline{\aPathb}}$ while preserving the semantics, and something similar happens with the operators $\wedge$ and $\vee$ for the case of node expressions using the $\lnot$ operator. In this grammar all these operators are defined because there are some natural restrictions of the language which have the same grammar except for some ``negative'' productions (like $\overline{\aPath}$) in which this simulations wouldn't be possible, and therefore we need to have them into the definition. Considering only the ``positive'' fragment of a language usually allows to obtain better complexity bounds in reasoning problems. This happens mainly because of the \textit{monotony} property: intuitively a set of expressions from a language $\mathcal{L}$ over structures $\mathcal{H}$ is said to satisfy monotony if whenever $H \in \mathcal{H}$ satisfies $\nu \in \mathcal{L}$ then $H'$ satisfies $\nu$ for every $H \subseteq H'$.
In our case, \Gposregxpath is defined as the subset of \Gregxpath expressions that don't use $\overline{\aPath}$ nor $\lnot \aFormula$. Thus, in \Gposregxpath we will not be able to `simulate' the $\cap$ operator unless it is present in the original \Gregxpath grammar. Also it can be proven that the set of \Gposregxpath node expressions satisfy monotony in the following sense: if $v \in \semantics{\nu}_\aGraph$ for $\nu \in \Gposregxpath$ then $v \in \semantics{\nu}_{\aGraph'}$ for every data-graph $\aGraph' \supseteq \aGraph$
\section{Introduction}
There is an increasing interest on graph databases as a means to adequately handle both the topology of the data and the data itself, which is specially useful for applications that involve analysis over social networks and other types of semi-structured data \cite{anand2010techniques,arenas2011querying,fan2012graph}. In these settings the structure of the database is queried through navigational languages such as \textit{regular path queries} or RPQs~\cite{barcelo2013querying} that can capture pairs of nodes connected by a specific kind of path. More expressive query languages have been defined with a tradeoff on evaluation complexity, nevertheless, RPQs and its most common extensions (C2RPQs and NREs ~\cite{barcelo2012relative}) do not treat data values in the nodes of the graph as first class citizens in the language leaving behind any possible interaction with that value source of data. For this reason, query languages have been defined for the case of data-graphs (i.e. graph databases where data lies both in the paths and in the nodes themselves), such as REMs and \Gregxpath~\cite{libkin2016querying}.
When accessing or querying a data repository we expect to obtain data that comply, at least to a certain extent, to the semantics of the domain, either in terms of quality or by satisfying a series of integrity constraints. This is a major challenge for any non-trivial data-driven application. Specifically for data-graphs, integrity constraints can be expressed in graph databases through \textit{path constraints}~\cite{abiteboul1999regular,buneman2000path}.
In the literature, two main approaches have been thoroughly explored, on the one hand data cleaning, or data repairing, focuses on frameworks that allow to identify inconsistencies caused by incorrect, missing, and duplicated data, and to restore it to a {\em clean} state, i.e., a state that satisfies the imposed constraints. On the other hand, repairing the data may not always be the best option or it may actually be impossible, e.g, we may not have the permission to actually change it. In these cases a different approach is to develop the means (theory and algorithms) to obtain consistent answers from inconsistent databases without changing the data itself. The field of Consistent Query Answering, first defined for relational databases~\cite{arenas1999consistent, lian2010consistent} has lately been applied to semi-structured data such as graph databases~\cite{barcelo2017data}.
One of the main drawbacks of traditional approaches to repairing and consistent query answering is that they do not allow to represent the fact that when we observe the {\em unclean} data, in general, we may not know exactly how the data was corrupted as a variety of factors may have been involved.
For this reason, inspired by the work in~\cite{de2018formal, rekatsinas2017holoclean}, in this paper we propose a probabilistic framework for repairing and querying data-graphs, assuming an epistemic model that describes the distribution over all possible ways in which the clean (uncertain) data-graph could have been polluted. A probabilistic epistemic model allows to incorporate uncertain domain knowledge to the framework formalizing
both a set of possible original {\em clean} databases and a noisy process that yields for each one potentially distorted versions which is application dependent. Clearly, this framework allows a more fined-grain treatment of inconsistency. In this work particularly, we study probabilistic data cleaning and query answering for data-graphs focusing only on two types of transformations that can be applied to the data-graph, by either removing or adding nodes and edges.
The specific contributions of this work are the following:
\begin{itemize}
\item We define the notion of a probabilistic unclean graph database model (PUGD) based on the observed data-graph and an Epistemic Model of the Graph Database (EMGD), composed of a probabilistic distribution over all possible clean data-graphs and a realization model that represents the noisy process which allows us to ‘observe’ a potentially distorted version of each of the clean versions.
\item We restrict the (epistemic model) EMGD to subset-EMGD and superset-EMGD considering only realization models that either only delete or only add data in the data-graphs. These notions relate to two common semantics for repairing databases, namely subset and superset repairs \cite{tenCate:2012}.
\item We study the problem of Data Cleaning, this is, given a probabilistic unclean graph database $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$, find the most probable graph database in $\aProbabilisticDatabase$ given our epistemic model of reality and our observation.
\item We also study the problem of Probabilistic Query Answering, which computes the probability of a $\aFormulaNodeOrPath$ over $\aProbabilisticDatabase$, given a probabilistic unclean graph database $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$.
\item Finally, we observe that the concepts of hard and soft integrity constraints \cite{arenas1999consistent, barcelo2017data, tenCate:2012} can be easily incorporated in our framework.
\end{itemize}
This work is organized as follows. In Section~\ref{Section:Definitions} we introduce the necessary preliminaries and notation for the syntax and semantics for our probabilistic model. In Section~\ref{section:dataCleaning} we define the problem of probabilistic data cleaning and study its complexity for both subset-PUGD and superset-PUGD. Later, in Section~\ref{section:PQA} we study the problem of probabilistic query answering for PUGDs again for the subset and superset versions of the epistemic model.
In Section~\ref{section:constraints} we briefly discuss the generalization of our framework to
consider soft and hard constraints.
Finally, Section~\ref{Section:Conclusions} we discuss related work and final remarks.
\section{Probabilistic query answering} \label{section:PQA}
\begin{definition}
Given $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$ and a node or path expression $\aFormulaNodeOrPath$, we define \defstyle{the probability of $\aFormulaNodeOrPath$ over $\aPUGD$} (or over $\aProbabilisticDatabase$ given $\aGraph'$ and $\aRealizationModel$) as:
$$\sum_{\aGraph \in \aProbabilisticDatabase} A_{\aGraph, \aFormulaNodeOrPath} \times \aRealizationModel^{-1}(\aGraph')(\aGraph),$$
where $A_{\aGraph, \aFormulaNodeOrPath}$ is the binary value of ``$\aFormulaNodeOrPath$ holds over all nodes (or pair of nodes) in $\aGraph$''.
\end{definition}
\begin{definition} [\functionProblem{Global PQA} over data-graphs]
We define the probabilistic query answering problem for graph-databases:
\begin{center}
\fbox{\begin{minipage}{30em}
\textsc{Problem}: The probability that a node [path] query $\aFormulaNodeOrPath$ holds over all nodes [pairs of nodes].
\textsc{Input}: A PUGD $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$
\textsc{Output}: The probability of $\aFormulaNodeOrPath$ over $\aProbabilisticDatabase$, given $\aGraph'$ and $\aRealizationModel$.
\end{minipage}}
\end{center}
\end{definition}
\begin{definition} [\functionProblem{Existential PQA} over data-graphs]
We could define in an alternative way the probabilistic query answering for graph-databases:
\begin{center}
\fbox{\begin{minipage}{30em}
\textsc{Problem}: The probability that a node [path] query $\aFormulaNodeOrPath$ holds over \emph{at least} one node [pair of nodes].
\textsc{Input}: A PUGD $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$
\textsc{Output}: The probability that $\aFormulaNodeOrPath$ holds over at least one node [pair of nodes] in the data-graphs from $\aProbabilisticDatabase$, given $\aGraph'$ and $\aRealizationModel$.
\end{minipage}}
\end{center}
However, note that this problem is the dual of global PQA: given a node expression $\aNodeExpression$ [a path expression $\aPathExpression$], the answer to the Global PQA problem for that formula coincides with 1 minus the answer to the Existential PQA problem for $\lnot \aNodeExpression$ [$\complementoCamino{\aPathExpression}$], and vice-versa:
Calling $A_{\aGraph, \aNodeExpression}$ the binary value of ``$\aNodeExpression$ holds over all nodes in $\aGraph$'' and $E_{\aGraph, \aNodeExpression}$ the binary value of ``$\aNodeExpression$ holds over some node in $\aGraph$'', we have:
\begin{align*}
\sum_{\aGraph \in \aProbabilisticDatabase} A_{\aGraph, \aNodeExpression} \times \aRealizationModel^{-1}(\aGraph')(\aGraph) &= \sum_{\aGraph \in \aProbabilisticDatabase} (1 - E_{\aGraph, \lnot \aNodeExpression}) \times \aRealizationModel^{-1}(\aGraph')(\aGraph) \\
&= \sum_{\aGraph \in \aProbabilisticDatabase} \aRealizationModel^{-1}(\aGraph')(\aGraph) - \sum_{\aGraph \in \aProbabilisticDatabase} E_{\aGraph, \lnot \aNodeExpression} \times \aRealizationModel^{-1}(\aGraph')(\aGraph)
\end{align*}
And therefore:
$$ \sum_{\aGraph \in \aProbabilisticDatabase} A_{\aGraph, \aNodeExpression} \times \aRealizationModel^{-1}(\aGraph')(\aGraph) = 1 - \sum_{\aGraph \in \aProbabilisticDatabase} E_{\aGraph, \lnot \aNodeExpression} \times \aRealizationModel^{-1}(\aGraph')(\aGraph) $$
We also define a decision problem related to \functionProblem{Global PQA}, namely \functionProblem{Global PQA bound}:
\begin{center}
\fbox{\begin{minipage}{30em}
\textsc{Problem}: Global PQA bound.
\textsc{Input}: A PUGD $\aPUGD = (\aProbabilisticDatabase,\aRealizationModel,\aGraph')$ and a bound $b$.
\textsc{Output}: Decide if the probability of $\aFormulaNodeOrPath$ over $\aProbabilisticDatabase$, given $\aGraph'$ and $\aRealizationModel$, is above $b$.
\end{minipage}}
\end{center}
\end{definition}
\input{example_globalPQAPathExpressionToySocialNetwork}
\begin{example} [Global PQA\textemdash probability that an (incompletely) observed network has no point of attack]
Suppose that an attacker observes data-graph $\aGraph'$ representing an incomplete diagram of a computer network. Assume they want to query whether all the nodes of the (full) network are secure (assuming for simplicity that this is represented by them having the data value \esDato{secure}, and they have a model of possible networks ($\aProbabilisticDatabase$) and of the limitations of their data-gathering methods ($\aRealizationModel$).
Here, it makes sense for the semantics of the probability in the PQA to represent the proportion of clean data-graphs were \emph{all} nodes are secure (i.e., the (weighted) proportion of $\aGraph \in \aProbabilisticDatabase$ such that for all $\aNode \in \aGraph$ we have that $\aGraph, \aNode \models \esDatoIgual{secure}$). The relative proportion of secure or not secure nodes should not be reflected in the calculated probability: the only thing that matters is the existence of at least a single point of attack. In other words, we are talking of the global PQA problem.
\end{example}
\subsection{PQA in subset-PUGD}
\input{PQASubsetPUGD.tex}
\subsection{PQA in superset-PUGD}
\input{PQASupersetPUGD}
\section{Supersets problems - basic definitions}
\begin{definition}
A \defstyle{superset-EMGD} is a pair $(\aProbabilisticDatabase, \aRealizationModel)$ such that the noisy observer $\aRealizationModel$ satisfies that $\aGraph' \supseteq \aGraph$ for every $\aGraph'$ such that $\aRealizationModel(\aGraph)(\aGraph') > 0$.
\end{definition}
\begin{definition} \label{def:noisyObserverMultisetsLabelsSupersetVersion}
We say that $(\aProbabilisticDatabase, \aRealizationModel)$ (or simply $\aRealizationModel$) is \defstyle{superset-$\Sigma$-uniform} if $(\aProbabilisticDatabase, \aRealizationModel)$ is a superset-EMGD and it satisfies that given $\aGraph, \aGraph', \aGraphb, \aGraphb'$, if:
\begin{itemize}
\item the multiset of data values of $V_{\aGraph}$ and of $V_{\aGraph'}$ are the same, and the multiset of labels of $Im(L_{\aGraph})$ and $Im(L_{\aGraph'})$ are the same, and
\item the multiset of labels of $Im(L_{\aGraphb}) \backslash Im(L_{\aGraph})$ is equal to that of $Im(L_{\aGraphb}) \backslash Im(L_{\aGraph'})$, and
\item the multiset of data values from $V_{\aGraphb} \backslash V_{\aGraph}$ is equal to that of $V_\aGraphb \backslash V_{\aGraph'}$.
\end{itemize}
then $\aRealizationModel(\aGraph)(\aGraphb) = \aRealizationModel(\aGraph)(\aGraphb') = \aRealizationModel(\aGraph')(\aGraphb) = \aRealizationModel(\aGraph')(\aGraphb')$.
\end{definition}
In other words, in this case the noisy observer can add nodes or edges to the original clean graph.
\begin{remark}
Observe that if $\aRealizationModel$ is superset-$\Sigma$-uniform, then $\aRealizationModel^{-1}(\aGraphb)$ is always finite. Indeed, that preimage consists of all subgraphs of $\aGraphb$, which is finite.
\end{remark}
\widesanti{Genial que tengamos esto. No creo que haga falta hacer un ejemplo paticular para el caso de superset. Lo único que me parece copado pensar es la definición de $\Sigma$-uniform. Al principio me parecía la restricción piola, pero en todas las demos las $\aRealizationModel$ terminan siendo incluso más simples (las demos que tenemos hasta ahora, con esto de superset vamos a tener más casos para ver). No digo de sacarla, pero si seguimos logrando cotas con versiones más débiles estaría bueno encuadrar esas versiones bajo algún nombre (muchas son literal $\aRealizationModel(H)(H') = 1 \iff H'=\aGraph$, que es probablemente lo más simple posible dado $\aGraph$. De todas formas esto depende también de como contextualicemos la $\aRealizationModel^{-1}$ (ver comentarios de más arriba)))}
\widesergio{14/7 Observar que en superset ya automáticamente tenemos una cota para la inversa de $\aRealizationModel$} |
\section{Introduction}
\label{sect:intro}
Photoacoustic tomography (PAT) is a non-invasive hybrid imaging modality based on the photoacosutic (PA) effect. By using laser excitation and acoustic detection, PAT takes advantage of the high contrast imaging present in purely optical modalities while maintaining the great resolution given by ultrasonic detection \cite{xu2006}. The PAT technique relies upon the generation of ultrasonic acoustic waves (or PA waves) induced by the illumination of biological tissue with short-pulsed non-ionizing laser light. The absorbed energy leads to a rapid increase in temperature and to the formation of pressure waves due to a transient thermoelastic expansion of illuminated object. The PA waves travel through the tissue and are sensed by a series of ultrasonic transducers placed at a number of fixed positions around the sample. The received signals can then be analyzed and processed to recover the initial pressure distribution. Optical absorption is closely related to several important physiological properties, such as oxygen saturation and hemoglobin concentration. Thus, in past few decades, PAT has been successful in obtaining high fidelity images of vascular anatomy in small animals and functional images of blood oxygenation \cite{zhang2009, lutzwieler2013}.
Unfortunately, the PAT inverse problem (the reconstruction of the 2-D or 3-D initial pressure distribution from the PA signals) is typically ill-posed. This means that even small errors in the measurements or inaccuracies in the modeling of the system can lead to an erroneous solution or a significantly distorted image \cite{tian2020}. These modeling errors have different sources and can occur due to several reasons: non-homogeneous speed of sound, the presence of noise in the measurements, sensor responses not sufficiently broadband or not omni-directional, limited-view detection, undersampling in space or time due to hardware constraints, etc \cite{hauptmann2020}. Furthermore, algorithms require precise information of the geometry of the experimental setup used for detecting the PA waves. Therefore, inaccuracies in the knowledge of the position, shape and size of the ultrasonic transducers can also lead to faulty modeling \cite{,tian2020,Sahlstrom2020}.
Previously, some studies have analyzed how acoustic heterogeneities \cite{tick2019, matthews2018, dean2012, jin2006} and finite transducer size \cite{mitsuhashig2014,rosenthal2011} affect the solution of the inverse problem. The effect of different discretizations of the forward problem on image reconstruction was briefly discussed in \cite{rosenthal2010}. More recently, Salhström et al. \cite{Sahlstrom2020} studied the usefulness of using a Bayesian framework for taking into account some of the inherent modeling errors in PAT when approaching the inverse problem.
Several reconstruction algorithms have been employed in the estimation of the initial pressure distribution \cite{lutzwieler2013, Sahlstrom2020}. A distinction between analytic inversion formulas and algebraic inversion procedures, both in time- and frequency-domain, can be made \cite{rosenthal2013}. Analytic reconstruction techniques use the mathematical model of the PA pressure propagation in order to perform the exact inversion of mathematical operator of the forward problem \cite{Minghua_Xu_Wang_2002, and_Wang_2002a, and_Wang_2002b}. Algebraic reconstruction techniques differ conceptually from analytic formulas as the inversion is typically performed from a numerical approximation to the forward or inverse problem. The most relevant approaches of each group are the Universal Back-Projection \cite{xu2005} and model-based algorithms \cite{guo2010}, respectively.
The back-projection (BP) algorithms are based on approximate analytical inversion formulas, and they are closely related with the (inverse) spherical Radon transform \cite{tick2016}. Due to this, BP algorithms are limited to a specific reconstruction geometry (e.g. cylindrical \cite{burgholzer2007}) and require that the PA signals be densely sampled along the detection surface \cite{rosenthal2013}. Furthermore, practical implementations of these closed-form solutions may lead to the appearance of significant streak-type artifacts, blurring and negative values in the reconstructed images \cite{xu2005, haltmeier2010}.
In contrast, the matrix model-based (MM) image reconstruction algorithms are based on a discrete representation of the acoustic forward model describing the propagation of pressure and not on a particular analytical solution like BP type approaches. This allows the construction of a matrix, which jointly represents the forward operator and the PAT measurement system and which is strongly dependent on the characteristics of the experimental setup. With this approach, the image reconstruction is performed by numerically minimizing the error (usually quantified by a quadratic loss function) between the measured acoustic signals and those predicted using the acoustic forward model \cite{rosenthal2013}. Typically, the minimization is done through a least squares based inversion of the model matrix. As this matrix could be ill-conditioned, regularization techniques such as Tikhonov regularization are typically used to stabilize the numerical inversion \cite{tick2016}. MM frameworks tend to be more computationally intensive than BP algorithms, since they require calculating and manipulating large matrices \cite{ding2016}. Nevertheless, they can be more versatile given that they can be applied to arbitrary measurement geometries and that many additional linear effects can be added to the model \cite{rosenthal2010}.
The accuracy or fidelity of the model used for representing an experimental setup is crucial for the success of MM algorithms \cite{rosenthal2011}. Since the construction of the matrix depends heavily on certain parameters of a given measurement geometry, care must be taken in the characterization of the PAT system to ensure good image reconstruction quality \cite{tian2020}. However, in typical measurement setups, eliminating all the sources that give rise modeling errors can be a difficult task \cite{Sahlstrom2020}. For example, temperature variations during the measurement process may cause slight drifts in the speed of sound of a medium \cite{tick2019} or the measurement setup may be built in such a way that the exact determination of transducers locations could result challenging \cite{Sahlstrom2020}.
Hence, it is of interest to analyze the behaviour of MM algorithms when some of the measurement setup parameters suffer from uncertainties in their values. In particular, such an analysis may show if some MM schemes prove to be more robust than others, or if some lead to better overall image quality when faced with uncertain parameters.
In this paper we analyze how uncertain knowledge of sensor locations in a typical experimental setup can lead to unsatisfactory image reconstruction quality due to modeling errors. Two common MM algorithms are tested: one working in the time domain \cite{paltauf2002} and another in the frequency domain \cite{provost2009}. Through simulations, uncertainty is added to the model matrices in a controlled manner, allowing us to analyze qualitatively and quantitatively the quality of the resulting reconstructed images. In particular, it is shown that it is advantageous to use a time domain over a frequency domain model when the knowledge of the sensor positions becomes moderately uncertain. In other words, the time domain model proves to be more robust to modeling errors than the frequency domain model when the value of a key parameter in the measurement setup like the transducer locations, is not accurately known.
This same behaviour is then observed when attempting to reconstruct an image from experimental measurements from a PAT setup in which the sensor positions are only known to $\sim1\%$ of their nominal value. Moreover, we discuss how uncertanties in the speed of sound in an acoustically homogeneous medium leads to a similar behaviour for the frequency domain MM method.
The remainder of this paper is organized as follows. Sections \ref{subsec:met_forward} and \ref{subsec:met_inverse} introduce the forward and inverse acoustic problems in PAT. Sections \ref{subsec:metsim} and \ref{subsec:metexp} explain the methodology for the simulations and experimental reconstructions. Then, in section \ref{s:results}, the obtained results are presented. In section \ref{s:disc}, a theoretical sensitivity analysis of the mathematical operator for the direct problem is provided. Finally, the conclusions of this paper are in section \ref{s:conclu}.
\section{Methods}
\label{s:met}
\subsection{Forward problem}
\label{subsec:met_forward}
According to PA theory, for an acoustically non-absorbing homogeneous medium the acoustic pressure $p(\mathbf{r},t)$ at position $\mathbf{r} \in\mathbb{R}^3$ and time $t$ originated from the optical absorption of a sample excited by an electromagnetic pulse $\delta(t)$, satisfies the homogeneous wave equation \cite{burgholzer2007}:
\begin{equation}
\left(\frac{\partial^2}{\partial t^2} - v_s^2 \, \nabla^2 \right) p(\mathbf{r},t) = 0
\label{eq:waveeq}
\end{equation}
\noindent with the initial conditions,
\begin{equation*}
p(\mathbf{r},0) = p_0(\mathbf{r})\, \text{,} \quad \left(\partial p /\partial t\right)(\mathbf{r},0)= 0
\label{eq:waveeq_cond_ini}
\end{equation*}
\noindent where $p_0(\mathbf{r})$ is the initial OA pressure and $v_s$ represents the speed of sound in the medium. Moreover, if thermal and acoustic confinement are fulfilled, that is, when the laser pulse duration is short enough that the heat conduction and acoustic propagation into neighboring regions of the illuminated region can be neglected, the initially induced pressure $p_0(\mathbf{r})$ is proportional to the total absorbed optical energy density \cite{lutzwieler2013}.
By solving the wave equation using the Green’s function in free space, the forward solution of the PA pressure detected by an ideal point-detector at position $\mathbf{r_d}$ placed on a surface $S$ surrounding the volume of interest $V$, can be written as \cite{biomedicalOptics}:
\begin{equation}
p_d(\mathbf{r_d},t)=\frac{1}{4\pi\,v_s^2} \frac{\partial}{\partial t}\iiint_{V} \, p_0(\mathbf{r}) \frac{\delta\left(t-|\mathbf{r_d}-\mathbf{r}|/v_s\right)}{|\mathbf{r_d}-\mathbf{r}|} d^3\mathbf{r}
\label{eq:fo_time}
\end{equation}
\medskip
The goal of the PAT inverse problem is to reconstruct the initial pressure $p_0(\mathbf{r})$ from the measured PA signals $p_d(\mathbf{r_d},t)$ at different positions $\mathbf{r_d}$ on a surface $S$ surrounding the volume of interest \cite{lutzwieler2013}.
Taking the Fourier transform with respect to $v_s\,t$, the forward operator (\ref{eq:fo_time}) can be rewritten in the frequency domain as \cite{xu2006}:
\begin{equation}
\bar{p}_d(\mathbf{r_d},k) = -i\, k \iiint_V \, p_0(\mathbf{r}) \, \frac{\exp\left(-i\,k\, |\mathbf{r_d}-\mathbf{r}|\right)}{4\pi|\mathbf{r_d}-\mathbf{r}|}\, d^3\mathbf{r}
\label{eq:fo_frec}
\end{equation}
\noindent where $k=\omega/v_s$ , $\omega$ is the angular frequency equal to $2\pi f$, and $f$ is the signal temporal frequency.
\subsection{Acoustic inverse problem}
\label{subsec:met_inverse}
Analytical inversion formulas were used for image reconstruction tasks in most PA systems in the past \cite{dean2013,rosenthal2013}. The reconstruction approach most favored in several articles are BP type algorithms, due to their simple implementation and applicability to most practical imaging scenarios \cite{rosenthal2013}. As mentioned in section \ref{sect:intro}, one of the most prominent formulations of the BP approach is the universal back-projection algorithm \cite{xu2005}. In a homogeneous medium with a constant $v_s$, the universal back-projection formula directly links $p_0(\mathbf{r})$ to $p_d(\mathbf{r_d},t)$ on the detection surface $S$ that encloses the PA source \cite{xu2005}:
\begin{equation}
p_0(\mathbf{r})=\int_{\Omega_s} b\left(\mathbf{r_d},t=|\mathbf{r_d}-\mathbf{r}|/v_s\right) \, \frac{d\Omega_s}{\Omega_s}
\label{eq:ubp}
\end{equation}
\medskip
\noindent where $b(\mathbf{r_d},t)=2\,p_d(\mathbf{r_d},t)-2\,t\,\partial p_d(\mathbf{r_d},t)/\partial t$ is the back-projection term related to the measurement at position $\mathbf{r_d}$, $\Omega_s$ is the solid angle of the whole surface $S$ with respect to the reconstruction point inside $S$, $d\Omega_s = dS \,\cos \theta_s/|\mathbf{r_d}-\mathbf{r}|$ and $\theta_s$ denotes the angle between the outwards pointing unit normal of $S$ and $(\mathbf{r_d}-\mathbf{r})$. The above formula is exact for cylindrical, planar and spherical geometries, and assumes point detectors with infinite bandwidth and uniform angular response \cite{cui2021,rosenthal2013,rosenthal2010}. However, in practice, the transducers can not be considered point-like, are band limited and do not have flat angular responses. In these non-ideal imaging scenarios, the analytical inverse formulas significantly deviate from reality, generating imaging artifacts (e.g. appearance of negative values that have no physical interpretation) and distorted images \cite{rosenthal2013,rosenthal2010}.
A different reconstruction approach is based on the MM algorithms. In this technique, the forward solution is represented by a matrix equation, which is used for solving the inverse problem. One of the advantages of this approach is that any linear effect in the system may be considered. Therefore, any spatio-temporal detection response that can be modeled or measured, may be taken into account in the inversion process \cite{paltauf2018,rosenthal2013,paltauf2002}. In the ideal case of point detectors and a homogeneous lossless acoustic medium, the model matrix in the time domain may be calculated by discretizing the integral relation in (\ref{eq:fo_time}) \cite{rosenthal2013}:
\begin{equation}
\mathbf{p_d}=\mathbf{A} \, \mathbf{p_0}
\label{eq:mbt}
\end{equation}
\medskip
\noindent where $\mathbf{p_d} \in\mathbb{R}^{N_d \cdot N_t\times 1}$ is a column vector representing the measured pressures at a set of detector locations $\mathbf{r_d}_l$ ($l=1 \ldots N_d$) and time instants $t_k$ ($k=1 \ldots N_t$); $\mathbf{p_0} \in\mathbb{R}^{N\times 1}$ is a column vector representing the values of the initial acoustic pressure on the imaging region grid; and $\mathbf{A} \in\mathbb{R}^{N_d \cdot N_t\times N}$ is the model matrix. The $j$-th element ($j=1 \ldots N$) in $\mathbf{p_0}$ contains the average value of the initial pressure within a volume element of size $\Delta V$ at position $\mathbf{r}_j$. Once the discrete formulation has been established, the inverse problem is reduced to the algebraic problem of inverting (\ref{eq:mbt}). The matrix $\mathbf{A}$ can be written as the multiplication of two matrices $\mathbf{A^{oa} \, A^s}$ where $\mathbf{A^s}$ represents the point response function of the imaging system for an ideal point-like sensor and $\mathbf{A^{oa}}$ is a time derivative operator. The matrix $\mathbf{A^s}$ is defined as \cite{paltauf2018, wang2011,dean2012,guo2010}:
\begin{equation}
A^s_{lkj} = \frac{1}{4\pi v_s^2}\frac{\Delta V}{\Delta t^2} \frac{d(t_k,\mathbf{r}_j)}{|\mathbf{r_d}_l - \mathbf{r}_j|}
\label{eq:Gs1}
\end{equation}
\begin{equation}
d(t_k,\mathbf{r}_j) = \begin{cases}
1 & \text{si } |t_k - \frac{|\mathbf{r_d}_l - \mathbf{r}_j|}{v_s}| < \Delta t/2 \\
0 & \text{else }
\end{cases}
\label{eq:Gs2}
\end{equation}
\medskip
\noindent where $\Delta t$ is the time step at which $p_d(\mathbf{r_d},t)$ are sampled and $\mathbf{A^{oa}}$ is a square matrix with replications of the basic ``N-shape'' signals grouped in columns along its main diagonal \cite{paltauf2018}.
In the frequency domain, the matrix representation of the forward model is similar as the time domain model described above \cite{provost2009,meng2012_2}. From (\ref{eq:fo_frec}) we can obtain:
\begin{equation}
\mathbf{\bar{p}_d}=\mathbf{K} \, \mathbf{p_0}
\label{eq:mbf}
\end{equation}
\medskip
\noindent where $\mathbf{\bar{p}_d} \in\mathbb{C}^{N_d \cdot N_f \times 1}$ is a column vector representing the measured pressures at $\mathbf{r_d}_l$ positions ($l=1 \ldots N_d$) and frequency samples $f_p$ ($p=1 \ldots N_f$); and $\mathbf{K} \in\mathbb{C}^{N_d \cdot N_f\times N}$ is the model matrix in the frequency domain which can be written as \cite{provost2009,meng2012_2}:
\begin{equation}
K_{lpj} = -i\,k_p \, \Delta V\, \frac{\exp\left(ik_p\,|\mathbf{r_d}_l - \mathbf{r}_j|\right)}{4\pi|\mathbf{r_d}_l - \mathbf{r}_j|}
\label{eq:Km}
\end{equation}
\noindent where $k_p = 2\,\pi \,f_p/v_s$.
In the case of a finite-size detector, the spatial impulse response (SIR) of the sensor must be taken into account \cite{rosenthal2011}. For the numerical calculation of the SIR, the area of the detector is divided into surface elements (treated as point detectors) which are then added up. Secondly, in \eqref{eq:Gs1} and \eqref{eq:Km}, a weight factor to take into account the size of surface elements and the directional sensitivity of the detectors is included \cite{paltauf2018}.
As already mentioned above, once a model matrix for the ideal case is constructed, the MM approach allows the model matrix to be modified in order to refine the forward model, such as to include the frequency or time response of the detector or the time/frequency properties of the illumination pulse \cite{paltauf2018,rosenthal2013}. Furthermore, the MM approach enables imposing constraints on the PA source to regularize the solution of the inverse problem. A typical regularization scheme is the so called Tikhonov regularization, which involves a square error minimization criterion coupled with a term weighing the $\ell_2$ norm of the solution \cite{provost2009,dean2012}:
\begin{equation}
\mathbf{\hat{p}_0} = \min_{p_0} \: || \mathbf{A} \, \mathbf{p_0} - \mathbf{p_d} ||_{\ell_2}^2 + \alpha \, ||\mathbf{p_0}||_{\ell2}^2
\label{eq:regTik}
\end{equation}
\medskip
\noindent where $\alpha\geq 0$ is the regularization parameter. Given a fixed $\alpha$, the solution to \eqref{eq:regTik} is unique and is called the Tikhonov-regularized pseudo-solution \cite{provost2009}:
\begin{equation}
\mathbf{\hat{p}_0} = (\mathbf{A}^H\,\mathbf{A} + \alpha \mathbf{I})^{-1} \mathbf{A}^H\,\mathbf{p_d}
\label{eq:solTik}
\end{equation}
\medskip
\noindent where $\mathbf{I}$ is the identity matrix and $H$ denotes the conjugate transpose operator. The value of the regularization parameter $\alpha$ can affect the solution and must be carefully chosen. For example, higher regularization tends to over-smooth the image while lower $\alpha$ values amplifies the noise in the images. In order to obtain the inverse equations for the frequency domain, $\mathbf{A}$ and $\mathbf{p_d}$ should be replaced by $\mathbf{K}$ and $\mathbf{\bar{p}_d}$ on (\ref{eq:regTik}) and (\ref{eq:solTik}).
\subsection{Simulations}
\label{subsec:metsim}
\begin{figure
\centering
\includegraphics[width = 0.35\textwidth]{Figure1.pdf}
\caption{Simulation geometry of the PA data acquisition setup.}
\label{fig:simulsetup}
\end{figure}
The effect of uncertainties in ultrasound sensor locations was studied using the simulation geometry of the PA data acquisition setup shown in Fig. \ref{fig:simulsetup} with increasing levels of uncertainty. The setup presented in Fig. \ref{fig:simulsetup} represents a two-dimensional PAT system implemented with a sensor rotating around the imaging region where the sample, uniformly illuminated, is placed. This single-detector based PAT system was demonstrated to be very useful for proof-of-concept studies due to its simplicity, low cost, and effectiveness \cite{tian2020,sharma2019}. A square imaging region having a size of 30 mm x 30 mm and a resolution of 64 x 64 pixels was used and the sensor (black circle) was placed on a circle of 22.5 mm radius. The PA signals were detected over 120 locations placed equidistantly around this circumference. Although large area detectors are typically used in practice, for simplicity, we assumed the sensor to be point-like with a bandwidth of 20 MHz. For data collection, the time step $\Delta t$ was 50 ns with $N_t=N_f=$ 420 samples. The speed of sound was set to $v_s=$ 1500 m/s and the medium was assumed homogeneous with no absorption or dispersion of sound. The transducer frequency response was modeled using a band-pass filter with upper and lower cutoff frequencies 0.1 MHz and 20 MHz, respectively. In order to simulate data with uncertainties in sensor locations, the ultrasound sensor locations $\mathbf{r_d}$ were perturbed as illustrated in Fig. \ref{fig:simulsetup}. These altered sensor locations were drawn from generalized uniform distributions for the radial values. The simulation conducted can be summarized into three steps:
\begin{enumerate}
\item Construct $\mathbf{A}$ and $\mathbf{K}$ with a radial value $R=$ 22.5 mm. Generate $\mathbf{p_d}$ and $\mathbf{\bar{p}_d}$ from a phantom $\mathbf{p_0}$ using these model matrices.
\item Construct $\mathbf{A_N}$ and $\mathbf{K_N}$ using a radial value in the uniform interval $\left[R - X\%,\, R + X\%\right]$, where $X$ represents the uncertainty introduced in the radius of the circumference of the sensor location. $\mathbf{A_N}$ and $\mathbf{K_N}$ are the matrices assumed by the reconstruction algorithm to be the real ones.
\item Solve (\ref{eq:regTik}) employing the data obtained in step 1 and the model matrices constructed in step 2 with an iterative method to approximate the solution, using the \textit{lsqr} algorithm provided in the Python module \textit{scipy.sparse.linalg}.
\end{enumerate}
In order to perform a statistical analysis, for each value of $X$, we performed 50 reconstructions with different radial sensor location randomly obtained within the range $\left[R - X\%,\, R + X\%\right]$.
As the initial pressure distribution $\mathbf{p_0}$, we used the standard Shepp–Logan numerical phantom \cite{shepp1974}. In all cases, Gaussian uncorrelated noise with zero mean and standard deviation of 1$\%$ of the maximum simulated peak amplitude, resulting in a signal to noise ratio (SNR) of 40 dB, was added to the data. Moreover, for comparison, we also reconstructed the image using the BP approach based on \eqref{eq:ubp} \cite{xu2005}. The simulations were carried out in the Python framework.
\subsection{Experimental setup}
\label{subsec:metexp}
\begin{figure
\centering
\includegraphics[width = 0.46\textwidth]{Figure2.pdf}
\caption{Experimental setup.}
\label{fig:expsetup}
\end{figure}
In order to obtain experimental PA signals for comparing the performance between the different reconstruction algorithms under study, we used a 2-D PAT system whose capabilities have been demonstrated in a previous work \cite{insabella2020b}. A diagram of the experimental setup is presented in Fig. \ref{fig:expsetup}. A polymer piezoelectric sensor and a sample were immersed in a vessel filled with deionized water. The water temperature was measured with a calibrated thermocouple to determine the speed of sound ($v_s=1479$ m/s). A Nd:YAG laser with second harmonic generation (Continuum Minilite I, 532 nm), 5 ns pulse duration, 10 Hz repetition rate and pulse energy less than 1 mJ, was the light source. A diverging lens adapted the diameter of the laser beam to a size larger than the sample, trying to achieve homogeneous illumination. The ultrasonic detector was fixed and pointed to the center of the rotating sample stage using an $xyz$ translation stage and it was oriented along the $z$ axis, perpendicular to the image plane $xy$. The distance between the sensor and the center of the rotating sample was (8.35 $\pm$ 0.05) mm (a position uncertainty of about 1\%). The phantom was fixed to a rotatory stage (Newport PR50CC) and rotated 360$^{\circ}$ in 3$^\circ$ steps. The sensor output was amplified with a transimpedance amplifier (FEMTO HCA-100MHz-50K-C), digitized by an oscilloscope (Tektronix TDS 2024) and processed on a personal computer. The PA signals were recorded and averaged 16 times. A pyroelectric detector (Coherent J-10MB-LE) measured the laser pulse energy.
The ultrasonic sensor used in this experiment is an integrating line detector with an active detection area of approximately 0.7 mm x 24 mm. The detection system (sensor + amplifier) has a pressure sensitivity of 1.6 mV/Pa, a noise equivalent pressure (NEP) of 0.4 mPa/$\sqrt{\text{Hz}}$ and a bandwidth of 20 MHz \cite{insabella2020b}.
The sample consists of an ink pattern laser (three capital letters reading UBA) printed on a transparent film embedded in agarose gel. The agarose gel was prepared with 2.5$\%$ (w/v) agarose in distilled water. First, a cylindrical base of the agarose gel with a diameter of 14 mm and a height of $\sim$20 mm was prepared. Then, the object (ink pattern on transparent film) was placed in the middle of the cylinder and fixed with a few drops of the gel. Finally, another layer of gel with a thickness of $\sim$1 mm was formed on top of the sample object. A picture of the sample is shown in Fig. \ref{fig:experiment}.a.
\section{Results}
\label{s:results}
\subsection{Simulations}
\label{subsec:simul}
Following the steps detailed in section \ref{subsec:metsim}, we simulated radial uncertainties $X\%$ in the range between 0.01 $\%$ and 10 $\%$. Figs. \ref{fig:sim_01percent}, \ref{fig:sim_1percent} and \ref{fig:sim_5percent} show the averaged images obtained by the MM approaches for three different radial uncertainties values in the range under study. As a benchmark, the reconstructed images using the BP method based on (\ref{eq:ubp}) are also presented.
\begin{figure
\centering
\includegraphics[width = 0.46\textwidth]{Figure3.pdf}
\caption{Average images obtained for $X=0.1\%$. (a) BP. (b) TDMM. (c) FDMM.}
\label{fig:sim_01percent}
\end{figure}
\begin{figure
\centering
\includegraphics[width = 0.46\textwidth]{Figure4.pdf}
\caption{Average images obtained for $X=1\%$. (a) BP. (b) TDMM. (c) FDMM. }
\label{fig:sim_1percent}
\end{figure}
\begin{figure
\centering
\includegraphics[width = 0.46\textwidth]{Figure5.pdf}
\caption{Average images obtained for $X=5\%$. (a) BP. (b) TDMM. (c) FDMM.}
\label{fig:sim_5percent}
\end{figure}
\begin{figure
\centering
\includegraphics[width = 0.45\textwidth]{Figure6.pdf}
\caption{Correlation value between the reconstructed image and the original image as a function of the uncertainty $X$.}
\label{fig: PSgraph}
\end{figure}
\begin{figure*
\centering
\includegraphics[height=4.9cm]{Figure7.pdf}
\caption{Images obtained employing the different reconstruction algorithms using 120 locations with $\sim1\%$ sensor position uncertainty. The phantom (ink pattern in agarose gel) reads UBA and its size is 3.0 mm x 2.5 mm. The calculated PC values are: (b) PC$=0.612$; (c) PC$=0.723$; and (d) PC$=0.123$.}
\label{fig:experiment}
\end{figure*}
As it can be seen in Fig. \ref{fig:sim_01percent}, the phantom can be perfectly recognized in the reconstructed images for $X= 0.1 \%$. However, as the value of $X$ increases, the images obtained with the frequency domain model matrix (FDMM) method are more affected than reconstructions obtained by the time domain model matrix (TDMM) and BP approaches. For example, in the image obtained by the frequency approach for $X= 1 \%$ (Fig. \ref{fig:sim_1percent}.c) almost no appreciable detail of the original image can be appreciated. On the other hand, the images reconstructed using time domain methods (Figs. \ref{fig:sim_1percent}.a and \ref{fig:sim_1percent}.b) are comparable to that achieved for $X= 0.1 \%$. For higher values of $X$, the images obtained with the time domain approaches are also affected as shown on Fig. \ref{fig:sim_5percent}. However, even for $X= 5 \%$, we are still able to observe a diffuse version of the phantom reconstructed with TDMM and BP (Figs. \ref{fig:sim_5percent}.a and \ref{fig:sim_5percent}.b), while, for the FDMM approach, it is not possible to observe any vestige of the original image (Fig. \ref{fig:sim_5percent}.c).
In order to carry out a quantitative performance analysis of the reconstructed images, we use the Pearson Correlation (PC) coefficient as an evaluation metric. The PC is a measure of linear correlation between two entities, here, the original image and the reconstructed image \cite{prakash2014}. It has a value between -1 and 1, with either extreme signifying perfect correlation (or anti-correlation) and a value of 0 meaning no correlation between the images.
Fig. \ref{fig: PSgraph} shows the calculated PC values versus the simulated radial uncertainties $X\%$ (on a logarithmic scale) for the three reconstruction methods. It can be observed that the value of the PC remains relatively constant up to an uncertainty of 0.3 \% for the time domain methods (TDMM and BP) and begins to fall sharply after $X\%$ exceeds 1\%. In the case of the FDMM algorithm, the PC is approximately constant up to 0.03 \% and starts to decrease rapidly for $X\%$ values greater than 0.1 \%.
\subsection{Experiment}
\label{subsec:experiment}
In order to verify the results of the simulations, we carried out measurements with the 2-D PAT system detailed in section \ref{subsec:metexp}. As the sensor of the experimental setup is not a point detector, we made modifications to the reconstruction algorithms used in the simulations. In the MM approach, for the numerical calculation of the SIR of the detector, its area is divided into surface elements which are then added up. Moreover, in order to take into account the size of surface elements and the directional sensitivity of the detectors, a weight factor is included in (\ref{eq:Gs1}) and (\ref{eq:Km}) \cite{paltauf2018}.
On the other hand, based on \cite{burgholzer2007}, we implemented a temporal BP algorithm for integrating line detectors. The reconstructed images from the measured pressures with a radial location uncertainty of $\sim$1 \% are presented in Fig. \ref{fig:experiment}. As it is mentioned in \ref{subsec:experiment}, the location uncertainty arises from $xyz$ translation stage used to position the sensor.
The reconstructions presented on Fig. \ref{fig:experiment} show a behavior similar to that found in section \ref{subsec:simul} where the time domain methods obtain a better quality image than FDMM. With TDMM (Fig. \ref{fig:experiment}.c) and BP (Fig. \ref{fig:experiment}.b) it is possible to distinguish the capital letters UBA in the center of the image. However, with FDMM, it is not possible to see any detail of the original phantom (Fig. \ref{fig:experiment}.d). The failure of the latter method can be attributed to the uncertainty in the sensor position. The time domain approaches also show some artifacts characteristic of the methods themselves. For example the BP (Fig. \ref{fig:experiment}.b) show some streak type artifacts caused by the finite number of angles used in the reconstruction \cite{rosenthal2013}. On the other hand, the fuzzy edges seen in the image corresponding to TDMM (Fig. \ref{fig:experiment}.c) are due to the effect of the Tikhonov regularization \cite{provost2009}. The calculated PC values of this images, indicated in the caption of Fig. \ref{fig:experiment}, are in agreement with the values presented in Fig. \ref{fig: PSgraph}.
\section{Discussion}
\label{s:disc}
As observed via simulations in section \ref{subsec:simul}, the FDMM approach showed higher sensitivity to slight variations in sensor positions than time domain methods. We attributed this sensitivity to the failure of the FDMM approach in recovering an image using experimentally obtained data in section \ref{subsec:experiment}. This results may indicate that when using a traditional pulsed laser experimental setup as the one used in section \ref{subsec:metexp}, time domain reconstruction techniques may be preferred over frequency domain model based approaches.
A possible explanation accounting for the greater sensitivity of the FDMM algorithm comes from looking at the entries of the matrix $\mathbf{K}$ given by \eqref{eq:Km}. One of the terms in \eqref{eq:Km} is a complex exponential whose argument is $i\,\frac{f_p\,2\,\pi}{v_s}\,|\mathbf{r_d} - \mathbf{r}|$. In a typical PAT setting, the frequency values $f_p$ range from a few hundred kHz to several tens of MHz. Therefore, any small error in the value of the sensor position $\mathbf{r_d}$ is amplified through its multiplication with $f_p$, leading to a drastic change in the value of the high frequency complex exponential (or phasor). This phenomenon can be shown through a simple numerical example. Taking $f_p = 1\,\text{MHz}$, $v_s = 1500\,\text{m/s}$ and $\mathbf{r} =( 0\, \hat{x} + 0\, \hat{y} + 0\, \hat{z})\,\text{mm}$, if a detector is placed on a point with coordinates $\mathbf{r_d} = 22.500 \, \hat{x} \, \text{mm}$, the complex exponential evaluates to $1.00 + i\,0.00$. Now suppose that, due to a measurement error, the detector position is considered to be at $\widetilde{\mathbf{r_d}} = 22.275 \, \hat{x}\,\text{mm}$ (only a 1\% difference with respect to $\mathbf{r_d}$). Using $\widetilde{\mathbf{r_d}}$, the value of the complex exponential becomes $0.59 - i\,0.81$.
Finally, the term $\frac{\exp\left(i\,f_p\,2\,\pi\,|\mathbf{r_d} - \mathbf{r}|\right)}{|\mathbf{r_d} - \mathbf{r}|}$ in \eqref{eq:Km} computes to roughly $(44.44 + i\,0.00)\, \text{m}^{-1}$ when using $\mathbf{r_d}$ and to $(26.4-i\,36.3)\, \text{m}^{-1}$ when using $\widetilde{\mathbf{r_d}}$, a substantial difference. The frequency domain matrix $\mathbf{K}$ is highly dense (i.e not sparse) and every entry in it depends upon the value of the complex exponential present in \eqref{eq:Km}. As a consequence, any tiny error inside the high frequency phasor affects every value in $\mathbf{K}$, leading to a matrix that does not accurately represent the system at hand. This effect only gets worse with higher values of $f_p$, meaning that higher bandwidth systems will be much more sensitive to uncertainties in $\mathbf{r_d}$. Nonetheless, even for typical detectors operating between 1-5 MHz, this simple numerical example shows that the error generated by this effect is noticeable.
On the other hand, the time domain model given by \eqref{eq:mbt}, whose characteristic matrix $\mathbf{A}$ is defined by \eqref{eq:Gs1}, is also affected by this problem but in a less extreme manner due to the lack of an exponential term. Instead, the term involving the detector position is $\frac{d(t_k,\mathbf{r})}{|\mathbf{r_d} - \mathbf{r}|}$ where $d(t_k,\mathbf{r})$ is a discrete representation of the delta term in \eqref{eq:fo_time} and can only be either 1 or 0, which does not drastically affect the numerical value of the matrix entry. Using the same numerical example as before, this term evaluates to $44.4\, \text{m}^{-1}$ when using $\mathbf{r_d}$ and $44.9 \, \text{m}^{-1}$ when using $\widetilde{\mathbf{r_d}}$, only a small difference. Slight changes in $\mathbf{r_d}$ also correspond to a slight shift in the position of the matrix entries in $\mathbf{A}$ due to the $d(t_k,\mathbf{r})$ term. This tiny shifting of the matrix entries is the cause of the blurring observed in the TDMM reconstructions in Section \ref{subsec:simul}, since the pressures measured at the detectors are interpreted as originating from different (but nearby) pixel in the image.
Incidentally, note that since the speed of sound $v_s$ is present in the same complex exponential (for the FDMM algorithm) and delta term (for the TDMM one) as $\mathbf{r_d}$, any uncertainties in this parameter will have a similar effect upon the matrices and lead to similar problems.
We can provide some more analysis regarding the sensitivity of frequency and time domain models. Both models are linear ones as equations (\ref{eq:mbt}) and (\ref{eq:mbf}) show. In the following we will denote with $\mathbf{M}$ the discretized forward operator present in those equations (either matrix $\mathbf{A}$ or $\mathbf{K}$). The presence of uncertainties in the measurement system implies the existence of a deviation in operator $\mathbf{M}$ with respect to a nominal situation. In more precise terms, the measured pressures at the detector locations can be written as:
\begin{eqnarray}
\mathbf{p}_\mathbf{d}&=&\mathbf{M} \,\mathbf{p}_{\mathbf{0}}+\mathbf{n}\nonumber\\
&=&\mathbf{M_N} \, \mathbf{p}_{\mathbf{0}}+\Delta\mathbf{M} \,\mathbf{p}_{\mathbf{0}}+\mathbf{n}
\label{eq:nominal_model}
\end{eqnarray}
\noindent where $\mathbf{M}$ is the actual discretized forward operator, $\mathbf{M_N}$ is the nominal model, assumed by the reconstruction algorithm to be the true one, $\Delta\mathbf{M}$ is the deviation from that nominal model induced by the uncertainties of the measurement system and $\mathbf{n}$ is a additive measurement noise (i.e., white noise in the detectors). From previous equation, it can be seen that the uncertainties of the measurement system can be interpreted as an additional source of noise, $\Delta\mathbf{M} \, \mathbf{p}_{\mathbf{0}}$, which is dependent of the ground truth $\mathbf{p}_\mathbf{0}$ in the forward model. This additional perturbation makes the reconstruction of the image harder. A simple characterization of the influence of this quantity is the norm of this term: $\|\Delta\mathbf{M} \, \mathbf{p}_{\mathbf{0}}\|$. Clearly, the precise influence of this quantity depends not only on $\Delta\mathbf{M}$, but also on the specific details of $\mathbf{p}_{\mathbf{0}}$. For example, if $\mathbf{p}_{\mathbf{0}}$, is aligned with a particular bad direction of $\Delta\mathbf{M}$, the influence of that term will be significant. From a practical point of view, it is better to have a quantitative measure of the influence of that term but also independent of the true value of $\mathbf{p}_\mathbf{0}$. In that sense, we can look for a worst-case criterion given by:
\begin{equation}
\|\Delta\mathbf{M}\|_2\equiv \sup_{\mathbf{p}_\mathbf{0}\neq\mathbf{0}}\frac{\|\Delta\mathbf{M}\mathbf{p}_{\mathbf{0}}\|}{\|\mathbf{p}_\mathbf{0}\|}
\label{eq:spectral_norm}
\end{equation}
Equation (\ref{eq:spectral_norm}) is the mathematical definition of the spectral norm of matrix $\Delta\mathbf{M}$ and requires obtaining its singular value decomposition in order to be computed \cite{Golub_Loan_1996}.
Finally, using these definitions, a quantitative measure of the effect of the uncertainties, with respect to the case of perfect knowledge of the measurement system (i.e., $\mathbf{M}=\mathbf{M}_N$) can be given by:
\begin{equation}
\delta \mathbf{M} \triangleq \frac{\|\Delta\mathbf{M}\|_2 }{\|\mathbf{M}\|_2 }=\frac{\|\mathbf{M} - \mathbf{M_N}\|_2 }{\|\mathbf{M}\|_2 }
\label{eq:final_measure_uncertain_v2}
\end{equation}
The idea behind this metric is to allow us to understand what happens to model matrices $\mathbf{A}$ and $\mathbf{K}$ as the percentage error $X\%$ in the detectors position $\mathbf{r_d}$ increases. The term $\|\mathbf{M} - \mathbf{M_N}\|_2$ in \eqref{eq:final_measure_uncertain_v2} quantifies the difference between the true forward operator $\mathbf{M}$ (the one that correctly models the system) and the operator $\mathbf{M_N}$ which contains the modeling errors (and it is assumed to be true by the reconstruction algorithm). Thus, if $\delta \mathbf{M}$ is low over increasing values of $X\%$, it would mean that $\mathbf{M_N}$ can accurately model the system even in the face of uncertainty in its transducers position. On the other hand if $\delta \mathbf{M}$ where to increase rapidly, it would mean that the model is sensitive to slight changes in $\mathbf{r_d}$ since the difference term $\|\mathbf{M} - \mathbf{M_N}\|_2$ grows quickly. Intuitively, this means that using $\mathbf{M_N}$ to invert data would lead to a poor quality and highly distorted image for even relatively low values of $X\%$.
\begin{figure
\centering
\includegraphics[width = 0.45\textwidth]{Figure8.pdf}
\caption{$\delta$ metric (Eq. \eqref{eq:final_measure_uncertain_v2}) for TDMM and FDMM model matrices $\mathbf{A}$ and $\mathbf{K}$ as a function of the uncertainty in the transducer positions.}
\label{fig:Errormat}
\end{figure}
Computing the $\delta$ metric is straightforward through simulation. First, the true model $\mathbf{A}$ or $\mathbf{K}$ is constructed (via \eqref{eq:Gs1} or \eqref{eq:Km}) by using a fixed value for the detector position $\mathbf{r_d}$. Then $\mathbf{A}_{\mathbf{N}}$ or $\mathbf{K}_{\mathbf{N}}$ is calculated with a corresponding detector position $\mathbf{r_d} + X\%$, were $X\%$ is the percent error manually introduced to represent faulty modeling of the system. Finally, using the corresponding matrices, the $\delta$ metric is computed according to \eqref{eq:final_measure_uncertain_v2}.
Fig. \ref{fig:Errormat} shows a graph of $\delta \mathbf{A}$ and $\delta \mathbf{K}$ for increasing values of $X\%$ plotted on a logarithmic scale.
As it can be clearly seen $\delta \mathbf{K}$ increases faster than $\delta \mathbf{A}$, indicating that the matrix entries present in $\mathbf{K}_\mathbf{N}$ are vastly different from those present in $\mathbf{K}$, even when the percent difference in the value in $\mathbf{r_d}$ is small. This means that even a slight difference between the experimentally measured and true values of $\mathbf{r_d}$ will lead to a matrix that can not accurately represent the system, and consequently, will not be able to reconstruct an good quality image. As mentioned before, we attribute this behaviour to the high frequency phasor present in \eqref{eq:Km}, whose value changes drastically given tiny changes in the value of $\mathbf{r_d}$ used to construct the FDMM matrix. Fig. \ref{fig:Errormat} was obtained using the same bandwidth, $v_s$ and $\mathbf{r_d}$ values as the simulations in Section \ref{subsec:simul}
Two other phenomena can be observed in Fig. \ref{fig:Errormat}. First, the variations in the value of $\delta \mathbf{K}$ after it reaches its peak value are due to the oscillatory nature of the complex exponential that make up the entries in $\mathbf{K}$. Secondly, both $\delta \mathbf{A}$ and $\delta \mathbf{K}$ will seem to decrease and converge to the same value when $X\%$ increases sufficiently. This can be easily explained recalling the definition for the $\delta$ metric \eqref{eq:final_measure_uncertain_v2} and the expressions for the matrix entries given in \eqref{eq:Gs1} and \eqref{eq:Km}. Both \eqref{eq:Gs1} and \eqref{eq:Km} contain a term of the form $1/|\mathbf{r_d} + \mathbf{r}|$, which diminishes in value as $\mathbf{r_d}$ increases, making every matrix entry smaller in absolute value. Thus, as $X\%$ increases, $\mathbf{r_d} + X\%$ increases and the coefficients in $\mathbf{M_N}$ get smaller. As a result the value of $\|\mathbf{M} - \mathbf{M_N}\|_2$ will approach $\| \mathbf{M} \|_2$ making $\delta M$ approach 1 for large enough $X\%$.
It is interesting to note that, when uncertainties in the speed of sound of the medium are present, a qualitatively similar behaviour as in Fig. \ref{fig:Errormat} is obtained. Once again, as occurs with $\mathbf{r_d}$, this is due to the presence of the speed of sound $v_s$ in the high frequency complex exponential in \eqref{eq:Km}.
The authors acknowledge other articles in which the FDMM has been successfully applied. In \cite{provost2009, meng2012, meng2012_3} a setup with a pulsed laser is employed and image is obtained using a similar process as the one described in \ref{s:met}, but no detail is given on the relative uncertainties in the measurement systems. As shown in section \ref{subsec:simul}, if the uncertainty in the transducer position is low enough, little modeling error is introduced and both the time domain and frequency domain approaches are able to reconstruct good quality images.
Other articles exist in which the FDMM has been applied successfully, but contain some noteworthy differences. For example, in \cite{mohajerani2014}, a frequency domain PAT system is presented in which the sample is excited by an amplitude modulated continuous-wave laser operating at different frequencies, allowing for the retrieval of phase and amplitude measurements.
This fundamentally changes the data processing and image reconstruction steps, and may be the key to a more robust setup. But again, no mention of the uncertainties in the system is given and a specialized setup different from the usual pulsed laser approach is used.
\section{Conclusions}
\label{s:conclu}
In this paper, the robustness of time and frequency domain model-based reconstruction algorithms was compared and analyzed. It was shown that in the presence of modeling errors caused by uncertain transducer positions in a typical PAT setup, the time domain approach proved to be more robust than the frequency domain one, which was unable to recover an image even when the uncertainty was moderately low.
The robustness of the different approaches was first studied through simulations, where modeling errors were deliberately added to the matrices used for reconstruction. This allowed for the qualitative and quantitative analysis of the obtained images and clearly showed the deterioration present in the frequency domain reconstructions. The experimental reconstructions done using a typical pulsed laser PAT setup with a $\sim1\%$ uncertainty in its transducers positions showed similar results.
The analysis provided in Section \ref{s:disc} supports the conclusion of the frequency domain approach being less robust and points to a high frequency complex exponential as a possible source for its sensitivity. The model mismatch caused by uncertainties can be interpreted as an additional source of noise, causing distorted reconstructions. All in all, the results indicate that when using a standard pulsed laser PAT setup, time domain reconstruction approaches should be used in favor of frequency domain ones, due to their robust characteristics.
Even though the frequency domain approach suffers from severe robustness issues when model uncertainties are present, it has proven useful when employing more advanced image reconstruction algorithms, such as the application of the compressed-sensing paradigm PAT image reconstruction \cite{provost2009}. Such cases would require experimental setups with very low working uncertainty to obtain good quality images. This opens the path analysing more closely the frequency model's limitations and studying ways to improve its robustness, such as using a Bayesian approach \cite{Sahlstrom2020} for solving the inverse problem. Another possible alternative, more suited for the frequency domain approach, is to study the implementation of a signal processing based calibration approach in order to estimate the uncertainties in the measurement setup in a similar way to the procedures typically used in MIMO-OFDM communication systems \cite{Minn_2010}.
\subsection* {Acknowledgments}
This work was supported by the ANPCyT (grants PICT 2016-2204 and PICT 2018-04589) and by the UBA (UBACyT grants 20020190100032BA).
|
\section{Introduction}
According to weak value approaches to particle presence, non-zero weak values show where a quantum particle has been, even when this would normally be impossible due to the pre- and post-selected states not being eigenstates of spatial projection operators \mbox{\cite{Danan2013Asking,Vaidman2014Tracing}}. These weak values are obtained via weak measurement and pre- and post-selection \mbox{\cite{Aharonov1988Weak}}, of sub-ensembles from ensembles of particles. However, in these approaches, particle presence is attributed to individual particles.
A version of this is the `weak trace' approach, according to which particles leave traces along the paths they traverse, due to interaction with the environment, but such a trace is left even when this interaction with the environment is vanishing. Such a non-zero weak value is taken to show that an individual particle travelled along that path \mbox{\cite{Vaidman2013Past}}, in this vanishing limit of interaction (a weak trace).
This approach claims there is a weak trace in a region when any operator formed by the product of the spatial projection operator for that region with an operator for a property of the particle (e.g. spin) has a non-zero weak value. This can happen even if the weak value of the spatial projection operator itself is zero. Weak traces are supposed to give us more information about the particle than von Neumann measurement of the spatial projection operator would allow. The weak trace approach even allows discontinuous particle trajectories, hinting at a mechanism by which seemingly disconnected events can affect one another, such as in counterfactual communication \mbox{\cite{Salih2013Protocol}}. Therefore, this approach has been controversial \cite{Li2013CommVaidman,Vaidman2013LiPastReply,Englert2017Revisited,Griffiths2016Path,Vaidman2017CommentGriffith,Griffiths2017Reply,Salih2018CommentPath,Griffiths2018Reply,Hashmi2016TSVF,Vaidman2016PastInterferom}.
Another approach connecting weak values and particle presence, due to Aharonov et al, requires that the spatial projection operator itself for a location has a non-zero weak value, for us to say the particle was present at that location. In the quantum Cheshire cat \mbox{\cite{Aharonov2013CheshireCats}}, these approaches come apart since the weak value of the spatial projection operator on a path can be zero, but the weak value of the spin operator times the spatial projection operator non-zero --- in these cases, the weak trace approach asserts the particle \emph{was} on the path with a zero weak value for the spatial projection operator, contrary to what Aharonov et al claim.
In this paper, we show how one obtains the weak value of an operator. We then give a case where the weak value of an operator containing the spatial projection operator is unexpectedly non-zero (and so, by weak values approaches, a particle has left an unexpected trace in its environment). After this, we look specifically at Vaidman's weak trace approach, and show it is inconsistent. Peleg and Vaidman argue that the weak trace approach is justified because there is some non-vanishing local interaction between the quantum particle and its environment along whatever path it travels \mbox{\cite{Peleg2018CommentRevisited}}. This implies the weak value of the spatial projection operator along a path should always be non-zero whenever a particle travels along that path, and so the particle was not present if the weak value of the spatial projection operator is zero. However, there are cases where the weak trace approach asserts a particle was present on a path, despite the spatial projection operator for that path having weak value zero (e.g. in the quantum Cheshire cat set-up).
We then point out four major problems more generally with weak value approaches.
First, even weak measurements disturb a system, so any approach relying on such a perturbation to determine the location of a quantum particle will only describe this disturbed system, not a hypothetical undisturbed state. We highlight this using the case of a balanced interferometer tuned to have destructive interference (i.e. no light exiting at its dark port) where such a perturbation changes the nature of the system. The unperturbed state is the vacuum, while the perturbed state has light present. While the measurement effect can be made arbitrarily small, this is not the same as removing it entirely. Further, we show Peleg's and Vaidman's attempt to respond to this line of argument by saying there are no completely unperturbed systems, as weak interactions are ubiquitous in nature \mbox{\cite{Peleg2018CommentRevisited}}, contradicts their association of presence with non-zero weak values of some operation multiplied by the spatial projection operator, when the weak value of the spatial projection operator is zero.
Secondly, even assuming no disturbance, there is no reason to associate the non-zero weak value of an operator containing the spatial projection operator with the classical idea of `particle presence'. Indeed, in some situations, just taking the weak value gives features inconsistent with classical ideas associated with a particle being present (e.g., giving discontinuous particle trajectories, or particles not being in coarse-grainings of their location).
Thirdly, weak values are only measurable over ensembles, and so to infer properties of individual particles from values of them is problematic at best.
Finally, weak values approaches to the path of a particle are not useful---they do not give us any new physics beyond standard quantum theory \mbox{\cite{Vaidman1996WeakElReality}}, to explain the causes of counterintuitive quantum effects (or even the paradoxes the approach itself creates). They simply assume a connection between particle presence and weak values without contributing anything testable.
\section{Weak Values}\label{WeakVals}
We first go over the derivation of a weak value of an operator, and how it is argued this relates to the trace left by a particle on its environment.
Aharonov et al defined the weak value $O_w$ of an operator $\hat{O}$ \cite{Aharonov1988Weak}, where
\begin{equation}\label{Eq:WeakVal}
O_w=\av{\hat{O}}_w=\frac{\bra{\psi_f}\hat{O}\ket{\psi_i}}{\braket{\psi_f}{\psi_i}}
\end{equation}
As this weak value increases as $\braket{\psi_f}{\psi_i}$ goes to zero, weak value protocols are used in metrology to amplify signals from delicate results so they can be observed experimentally \cite{Dixon2009WeakAmp,Jordan2014AdvWeak,Harris2017Weak}. This is at the expense of postselection reducing success probability. Weak values however also lead to a range of paradoxes \cite{Aharonov2005Paradoxes}.
To derive Eq. \mbox{\ref{Eq:WeakVal}}, we first couple our initial system $\ket{\psi_i}$ to our initial pointer state, $\ket{\phi}$, by weakly measuring them with the probe Hamiltonian $\hat{H}=\lambda\hat{O}\otimes\hat{P_d}/T$ for small coupling constant $\lambda$ and state-probe interaction time $T$. This produces the state
\begin{equation}
\begin{split}
\ket{\psi_w}&=e^{-\frac{i\hat{H}T}{\hbar}}\ket{\psi_i}\otimes\ket{\phi}\\
&=e^{-\frac{i\lambda}{\hbar}\hat{O}\otimes\hat{P_d}}\ket{\psi_i}\otimes\ket{\phi}
\end{split}
\end{equation}
where $\hat{P}_d$ is the momentum of that pointer. We then strongly measure this weak-measured state $\ket{\psi_w}$ with the operator
\begin{equation}
\hat{F}_1=\ketbra{\psi_f}{\psi_f}\otimes\hat{I}_d
\end{equation}
Assuming $\hat{P}_d$ has Gaussian distribution around 0 with low variance (so $\hat{X}_d$, the position of the pointer, has Gaussian distribution with high variance),
we can say
\begin{equation}\label{Eq:FirstOrderApprox}
\begin{split}
&e^{-\frac{i\lambda}{\hbar}\hat{O}\otimes\hat{P_d}}=\sum_{k=0}^\infty\big(-\frac{i\lambda}{\hbar}\hat{O}\otimes\hat{P_d}\big)^k/k!\\
&\;=\mathds{1}-\frac{i\lambda}{\hbar}\hat{O}\otimes\hat{P_d}+\mathcal{O}(\lambda^2)\approx \mathds{1}-\frac{i\lambda}{\hbar}\hat{O}\otimes\hat{P_d}
\end{split}
\end{equation}
so this strong measurement gives the result
\begin{equation}
\begin{split}
&\ketbra{\psi_f}{\psi_f}e^{-\frac{i\lambda}{\hbar}\hat{O}\otimes\hat{P_d}}\ket{\psi_i}\otimes\ket{\phi}\\
&\approx\ketbra{\psi_f}{\psi_f}(\mathds{1}-\frac{i\lambda}{\hbar}\hat{O}\otimes\hat{P_d})\ket{\psi_i}\otimes\ket{\phi}\\
&=\ket{\psi_f}\otimes\braket{\psi_f}{\psi_i}(\mathds{1}-\frac{i\lambda}{\hbar}O_w\hat{P}_d)\ket{\phi}\\
&\approx\ket{\psi_f}\otimes\braket{\psi_f}{\psi_i}e^{-\frac{i\lambda}{\hbar}O_w\hat{P}_d}\ket{\phi}
\end{split}
\end{equation}
This means the position of our initial pointer state $\ket{\phi}$, acting as our ``readout needle", shifts, with measurement of the pointer position giving a read-out value $(\langle x\rangle - a)$. For an initial Gaussian pointer state, this shift causes the average position to move from $\langle x \rangle$ to $(\langle x \rangle - \overline{a})$ due to the application of the effective operator $\mathds{1}-i\lambda O_w\hat{P}_d/\hbar$. $a$ is distributed over a wide range of values for many repeats, but the distribution of ``$a$"s will be a Gaussian centred on (the real part of) $\lambda O_w$. Peculiarly, this weak value $O_w$ can be very far from any of the eigenvalues of $\hat{O}$, or even imaginary \cite{Aharonov1988Weak,Tamir2013Intro,Dressel2014UnderstandingWeak}. This is odd, given this weak value appears in exactly the place in the equation that an eigenvalue of $\hat{O}$ would for a Von Neumann measurement (where the variance of $\hat{X}_d$ on the measured state is vanishingly small). This led to the belief that the weak value $O_w$ represented some fundamental value of the operator $\hat{O}$ \textbf{between} measurements.
Due to their interrelatedness, we can graphically show where the spatial projection operator has a non-zero weak value through the Two-State Vector Formalism (TSVF). Aharonov et al developed the TSVF \cite{Aharonov1964TSVF}, which considers both the backwards and forwards-evolving quantum states, rather than just the forward as in standard quantum mechanics. (A similar intuition led Watanabe to develop the Double Inferential-state Vector Formalism \cite{Watanabe1955DIVF}.) For some operator $\hat{O}$, the forwards travelling initial state $\ket{\psi_i}$, and the backwards travelling final state $\bra{\psi_f}$, the TSVF gives out a conditional probability amplitude $\bra{\psi_f}\hat{O}\ket{\psi_i}$, where $\bra{\psi_f}\ket{\psi_i}$ is referred to as the Two-State Vector. This conditional probability amplitude is of the same form as the numerator of the weak value, and so we can use the TSVF to graphically plot where the weak value of an operator containing the spatial projection operator is nonzero. This involves plotting the forward-evolving state (possible paths the particle could have travelled via from its original position) and the backwards-evolving state (possible locations the particle could have come from to reach its final position). Weak value approaches assert a quantum particle is present wherever these states visibly overlap (as shown in Fig. \ref{fig:TSVFDiag}).
\begin{figure}
\centering
\includegraphics[width=\linewidth]{TSVFDiagW.png}
\caption{Nested-interferometer model, used by supporters of weak value approaches to discuss cases where the Two-State Vector Formalism contradicted ``common-sense" continuous-path approaches \cite{Vaidman2013Past}.}
\label{fig:TSVFDiag}
\end{figure}
The nested-interferometer model, which we show in Fig. \ref{fig:TSVFDiag}, was pointed out as a case where the Two-State Vector Formalism contradicted ``common-sense" continuous-path approaches \cite{Vaidman2013Past}. The inner interferometer is balanced such that a photon entering from arm $D$ exits to detector $D3$. Consequently, the outer interferometer is unbalanced, so there is an equal probability of a photon introduced from the source ending in $D1$ or $D2$. When the photon ends at $D2$, common sense would tell you it must have travelled via path $A$. This is as, had it travelled via $D$ into the inner interferometer, it could not have exited onto path $E$, or reached $D2$. However, supporters of the weak trace approach claim, while the photon never travelled paths $D$ or $E$, it travelled along paths $B$ and $C$ as well as along path $A$.
\section{Vaidman's Weak Trace Approach}\label{Vaidman}
Consider a case where the weak value of either the spatial projection operator, or an operator formed by multiplying the spatial projection operator with the operator for some property of a particle, along a certain path was non-zero for a given $\ket{\psi_i}$ and $\bra{\psi_f}$. Vaidman argues we should interpret this as meaning a particle was on that path and left a weak trace along it, on two grounds \mbox{\cite{Vaidman2013Past}}. Firstly, if it were a strong measurement, an operator containing the spatial projection operator having non-zero eigenvalues would determine if a quantum particle was present. Secondly, if a particle was present in a region, he argues, it should have non-zero interaction with the local environment, due to decoherence, and so leave a ``trace'' along its local path. While this interaction would necessarily be weak, as we can see by the environmental coupling not collapsing position superpositions in the way a projective measurement would, supporters of the weak trace approach claim it would still be in principle detectable (i.e. only the first order in the expansion of the weak coupling constant), and so would be equivalent to the weak value of an operator containing the spatial projection operator along that path (which is also only to first order in $\lambda$).
However, this approach is inconsistent because there are cases where it asserts a particle was present on a path, despite the spatial projection operator for that path having weak value zero (e.g. Aharonov's quantum Cheshire cat). This contradicts Peleg's and Vaidman's claim that the weak trace is justified because there is some non-vanishing local interaction, and so a quantum particle interacts with the environment along whatever path it travels \mbox{\cite{Peleg2018CommentRevisited}}. This claim implies the weak value of the spatial projection operator along a path should always be non-zero whenever a particle travels along that path, and so the particle was not present if the weak value of the spatial projection operator is zero. This contradicts Vaidman's claims in \mbox{\cite{Vaidman2013Past}} that a particle is present along a path even when the weak value of the spatial projection operator along that path is zero, so long as the weak value of some other operator multiplied by the spatial projection operator is nonzero (e.g. the weak value of the spin along the right-hand path in the quantum Cheshire cat set-up).
\section{Issues with the weakness assumption}
The process of weak measurement and strong postselection given above leads to a small shift in the position $X_d$ of our pointer system $\ket{\phi(x)}$, with an average value
\begin{equation}
\overline{a}=\lambda \text{Re}(O_w) << \Delta X_d
\end{equation}
Performing the measurement on a pre-and-postselected ensemble of $N$ particles allows measurement of the shift to precision $\Delta X_d/\sqrt{N}$. Therefore, supporters of the weak values approach claim, as long as $N>(\Delta X_d)^2\mathcal{O}(\lambda^{-2})$, the presence of the particle is revealed. This is still however a coupling--- so long as $\lambda\neq0$, there is measurement. This is as it must be, because Busch's Theorem says we cannot gain information about the state of the system without disturbing it \mbox{\cite{Busch2009Limit}}.
This is at odds with the ``weakness assumption'' about weak measurement---that if we measure weakly enough, we can effectively see how the system behaves when it is not measured. As argued above, weak interaction is still interaction and so disturbs the system (albeit negligibly so for some purposes). As long as there is a non-zero coupling, as in all weak value experiments, the `negligibly' small disturbance is still a disturbance, and there are differences in observable effects between small coupling and no coupling (see \mbox{\cite{Hance2022DQCCDetectable}} for an example of such a difference).
Popular interpretations of weak values claim that they exist in the absence of measurement and perturbation \mbox{\cite{Aharonov2017Disappearing,Aharonov2018Mirage,Vaidman1996WeakElReality}}. However, this is clearly not the case \mbox{\cite{Svensson2013Weak,Wiesniak2014CriticismAsking,Salih2015CommentAsking,svensson2015non,Ben2017CommSvenNonRep,Kastner2017DemystifWeakVals,Svensson2017ReCommNonRep,Ipsen2022Disturbance,Hance2022DQCCDetectable}}. To defend against this point Peleg and Vaidman say, ``in a hypothetical world with vanishing interaction of the photon with the environment, Vaidman's definition is not applicable, but in the real world there is always some non-vanishing local interaction. Unquestionably, there is an unavoidable interaction of the photon with the mirrors and beam splitters of the interferometer" \cite{Peleg2018CommentRevisited}.
This statement contradicts the approach taken practically in obtaining spatial weak values. If the purpose of weak value approaches is to identify the non-vanishing interaction a quantum particle has with its environment, to trace its path, then a definition of this trace which neglects some of these non-vanishing local interactions is only telling us part of the story. Proponents of weak value approaches only consider the first-order trace, as per the approximation given in Eq. \mbox{\ref{Eq:FirstOrderApprox}}, due to the supposedly vanishing nature of the terms of second-order or higher in $\lambda$. Despite this, Peleg and Vaidman specifically say that these environmental terms provide non-vanishing local interaction\mbox{\cite{Peleg2018CommentRevisited}}. A typical photon interaction with a mirror for instance transfers of order $10^{-33}$ of its energy per reflection (for a 1500nm-wavelength photon and a 1 gram mirror), a suitably weak interaction that is mostly safely ignored. Hence even the higher order terms are many orders of magnitude larger than the strongest perturbation from the environment in optics experiments, and ignoring higher-order terms, while invoking the weak trace naturally left by the quantum object on its surroundings, is misleading.
It could be argued that this is more an objection to the experimental methods used to detect spatial weak values than the application of the concept itself---however, it is from consideration of these experimental methods that supporters of weak value approaches claim we should neglect the higher-order terms (due to their comparative undetectability)---so therefore, if we are instead looking more theoretically at the non-vanishing local interactions, we should associate presence with all terms, rather than just the first order terms in $\lambda$.
\section{Do weak values reveal particle presence?}
The classical conception of a particle presence --- being present at a certain place at a certain time --- can be characterised as follows:
\renewcommand{\labelenumi}{\roman{enumi}).}
\begin{enumerate}
\item Every particle is located in space at all times.
\item Particles cannot be on more than one path simultaneously.
\item Particle trajectories are continuous (or at least as continuous as space is) so particles cannot get from one place to another without passing through the space in between.
\item Particles interact with other objects/fields local to their location.
\item If a particle is on a path at a given time, and that path is within some region, then the particle is also located in that region at that time.
\item If a particle's property is at a location, the particle must be at that location too.
\end{enumerate}
In standard quantum mechanics, when a particle is in an eigenstate of a position operator, then it is attributed a position. However, quantum particles can be in superpositions of position eigenstates. Whether they then have positions at all, or even have two positions at once, is contentious and interpretation-dependent. To require that there is always some location at every time where conditions (i) and (ii) are satisfied is to advocate a hidden variable approach to quantum mechanics, where the hidden variable is the particle's location. However, despite quantum tunnelling and other phenomena, condition (iii) still applies for quantum particles in superpositions of position evolving according to the Schr\"odinger equation, since the evolution of the superposition is simply the superposition of the classical evolutions of the classical components of the superposition, each of which has a continuous trajectory. (iv-vi) are also compatible with standard quantum mechanics.
Particles that are not in eigenstates of position can nonetheless have nonzero weak values for the spatial projection operator at a given location (or some composite operator formed of this multiplied by some other operator). Perhaps a condition for a particle being present at a location in a quantum context is that it has a non-zero weak value for an operator containing the spatial projection operator at that location. Whenever a particle satisfies condition (ii), it also satisfies this condition, as its forwards and backwards-travelling states always overlap. However, this can only be a necessary condition for particle presence in the sense of (i) to (vi) above, rather than a sufficient condition, as there are cases where:
\renewcommand{\labelenumi}{\arabic{enumi}).}
\begin{enumerate}
\item A particle has a non-zero weak value of an operator containing the spatial projection operator at a location, but has no continuous path to/from this location.
\item A particle has a non-zero weak value of an operator containing the spatial projection operator at a location, but we cannot say for certain that the particle would interact with other objects/fields in that location, as any such effect would cause a breakdown in interference and so change the system.
\item A particle has a non-zero weak value of an operator containing the spatial projection operator at a location, but a weak value of zero for an operator containing the spatial projection operator at a coarse-graining of that location (e.g. having a non-zero weak value of an operator containing the spatial projection operator on path $B$, and a non-zero weak value of an operator containing the spatial projection operator on path $C$, but a weak value of zero for an operator containing the spatial projection operator for the space composed of paths $B$ and $C$).
\item A particle has a weak value of zero for the spatial projection operator along a certain path, yet has a non-zero weak value for some other operator (e.g. the spin operator in a given direction) multiplied by the spatial projection operator along that path (e.g. in the quantum Cheshire cat set-up).
\end{enumerate}
Point (2) could be challenged by Peleg's and Vaidman's claim that quantum objects leave a ubiquitous weak trace wherever they travel, which would have an effect on the environment there. However, Dziewior et al \mbox{\cite{Dziewior2019Polzn}} show there are locations where ``no effect on local external systems can be observed, even when the forward-evolving wavefunction did not vanish'', and further claim that certain key cases where an effect on local systems (``coupling to the external degrees of freedom'') can be thought of ``as being due to misalignment of the interferometer.'' This reinforces the idea that there is a difference between a non-zero weak trace and presence in the setup as claimed, rather than a misaligned version (with a different actual state to that claimed through the ``weakness' assumption).
Using the two-state vector formalism and weak value tools, we can quantitatively analyse the nested interferometer set-up to show Point (3). We first define the forwards-travelling initial vector and backwards-travelling final vector by the paths via which they evolve:
\begin{equation}
\begin{split}
\ket{\psi_i}=\frac{\sqrt{2}\ket{A}+\ket{B}+\ket{C}}{2}\\
\bra{\psi_f}=\frac{\sqrt{2}\bra{A}+\bra{B}-\bra{C}}{2}
\end{split}
\end{equation}
Using these, and defining the spatial projection operator for arm $A$ as $\hat{P}_A=\proj{A}$ (similarly for $B$ and $C$), we get the weak values Vaidman does in \cite{Vaidman2013Past}:
\begin{equation}
\begin{split}
\av{\hat{P}_A}_w=&\frac{\bra{\psi_f}\proj{A}\ket{\psi_i}}{\braket{\psi_f}{\psi_i}}=1\\
\av{\hat{P}_B}_w=&\frac{\bra{\psi_f}\proj{B}\ket{\psi_i}}{\braket{\psi_f}{\psi_i}}=\frac{1}{2}\\
\av{\hat{P}_C}_w=&\frac{\bra{\psi_f}\proj{C}\ket{\psi_i}}{\braket{\psi_f}{\psi_i}}=-\frac{1}{2}
\end{split}
\end{equation}
If we want to see however if the particle was present in the inner interferometer as a whole (as in either arm $B$ or $C$), we define $\hat{P}_{BC}=\proj{B}+\proj{C}$, as we are allowed to do, given projectors in standard quantum mechanics are linear. We find
\begin{equation}
\begin{split}
\av{\hat{P}_{BC}}_w&=\frac{\bra{\psi_f}(\proj{B}+\proj{C})\ket{\psi_i}}{\braket{\psi_f}{\psi_i}}\\
&=\frac{1}{2}-\frac{1}{2}=0
\end{split}
\end{equation}
If we assumed a non-zero weak trace implies particle presence, this would mean the photon was never in the inner interferometer (made up of arms $B$ or $C$) overall. (Vaidman explicitly says that weak values obey the sum rule, and so allows us to say $\av{\hat{P}_{BC}}_w$ must equal $\av{\hat{P}_{B}}_w+\av{\hat{P}_{C}}_w$ \cite{Vaidman1996WeakElReality,Ben2017Improved}.) This seems incoherent, given, taken separately, it claims the particle was in arm B, and was in arm C. (Aharonov et al consider a similar scenario in their three-box experiment, and also discuss this idea of negative weak value in the equivalent of arm C cancelling the positive weak value in the equivalent of arm B \cite{Aharonov2017Disappearing}.) This illustrates the importance of the sign of the weak values, which the weak trace approach, and the graphical form of the TSVF (as in Fig.\ref{fig:TSVFDiag}), neglect.
For weak value experiments aiming to distinguish which-path information, we can link this to the Visibility-Distinguishability Inequality \cite{Jaeger1995Complementarity,Englert1996Fringe},
\begin{equation}\label{Eq:DistVis}
\mathcal{D}^2+\mathcal{V}^2\leq1
\end{equation}
where $\mathcal{D}$ is the distinguishability of which path the light travelled, and $\mathcal{V}$ is the fringe-visibility at the output of the interferometer. Therefore, doing anything which would increase the distinguishability between the two paths (e.g. placing different tags on $B$ and $C$) will affect the interference pattern at the $BCE$ beamsplitter. Given perfect interference is required to ensure all light that enters the inner interferometer from $D$ exits into $D3$, anything causing distinguishability between paths $B$ and $C$ will allow light to leak through onto $E$, and cause a trace to show in $D2$. Therefore, it will never be showing what would happen in an unperturbed system.
Given Eq. \mbox{\ref{Eq:DistVis}}, $\av{\hat{P}_{BC}}_w$ is a far better measure of whether light reaching $D2$ was ever in the undisturbed inner interferometer. This is as measuring $\av{\hat{P}_{BC}}_w$ does not cause distinguishability between paths $B$ and $C$ in the inner interferometer, and so does not affect the interference pattern required to output all inner interferometer light into $D3$. $\av{\hat{P}_{BC}}_w$ being 0 provides support for a `common-sense' path (i.e. light only travelling via $A$ to reach $D2$) in an unperturbed system.
An argument against looking at the weak value of the projector over multiple paths (e.g. $\hat{P}_{BC}$) specific to Vaidman's weak trace approach is that the weak trace is only defined as the effect of the particle on the environment local to it---and combining the two weak values for these two paths is not a local operation, so does not indicate anything about the weak trace. However, this still goes against classical ideas of presence, given it shows that the presence in each of the two arms must have opposite sign, so they can cancel when added, and so implies that being realist about weak values implying presence means being realist in some way about `negative' presence. This negative presence is yet another way the `presence' asserted by weak values approaches differs from classical presence.
Point (4) seems to support Vaidman's weak trace approach to particle presence, whereby a non-zero weak value of any operator corresponding to a property of the particle, multiplied by the spatial projection operator along a path, implies the particle was present along that path. However, as we saw in Section \mbox{\ref{Vaidman}}, taking this to imply presence even when the weak value of the spatial projection operator alone is zero, contradicts Peleg and Vaidman's claim that the weak trace comes from the necessarily nonzero interaction between a particle and its environment, given if this were the case, we would expect a particle to necessarily have a nonzero weak value for the spatial projection operator for any path along which it travels. This is a dilemma for weak value approaches---they either contradict the rationale for weak values indicating particle presence, or imply that a particle's properties are not necessarily in the same location as that particle.
From these four points, we see that the ``nonzero weak value for an operator containing the spatial projection operator'' condition can be satisfied while the conditions for a particle being present at a location are not satisfied. Therefore, it is not a sufficient condition for particle presence in the standard sense of the term.
\section{Weak values are only defined over ensembles}
It is important to note that advocates of the weak value approach to particle presence use it to attribute presence to individual particles. This is despite the experimental results of measurements of weak values necessarily being produced by ensembles, because weak values are obtained by post-selection. It is non-obvious that we can go from facts about ensembles to facts about individual elements of that ensemble. As an example, we see that one can have a weak value of the velocity of electrons in a given set-up that is greater than the speed of light in a vacuum \cite{Rohrlich2002Cherenkov,Solli2004Superluminal,Ahnert2004Superluminal}. However, this does not necessarily mean than any of the individual electrons in the ensemble used to obtain this weak value travelled at superluminal speeds---as otherwise this would be a violation of special relativity. Instead, similar to how group velocities of wavepackets can be superluminal despite the phase velocity (or velocity of the components) being below $c$, this superluminal weak value for speed is a fact only about the ensemble, rather than any of its constituents. Similarly, there is no reason to infer facts about the presence of individual quantum particles from weak values of operators containing the spatial projection operator.
\section{Usefulness of weak value approaches to particle presence}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{TSVFSalih.png}
\caption{Two-State Vector Formalism analysis of Salih et al's polarisation-based single-outer-cycle protocol for counterfactual communication. Bob communicates with Alice by turning off/on his switchable mirrors to determine whether the photon goes to $D1$ or $D0$ respectively. We specify the polarisation, given it determines direction of travel through the polarising beamsplitters (PBSs). The forwards- and backwards-travelling states do not overlap anywhere on the inner interferometer chain when there is a detection at $D0$, meaning, by weak value approaches, the particle detected at $D0$ was never at Bob. This means Bob's ability to communicate with Alice is not explained by weak value approaches any more than it is by standard quantum mechanics.}
\label{fig:Salih}
\end{figure}
Supporters of weak value approaches argue they provide more (interpretational) information about the underlying state of the system than standard quantum mechanics. They claim that an issue with Wheeler's ``common-sense" approach to particle trajectories \cite{Wheeler1978Past}, is that it is entirely operational: not telling us anything about the underlying mechanisms at work, just the final result \cite{Vaidman2013Past}. Similarly, Vaidman claims ``the von Neumann description of the particle alone is not sufficient to explain the weak trace'' \mbox{\cite{Vaidman2013Past}}---implying weak value approaches provide something more than the standard von Neumann approach does. We however turn this criticism back on these supporters---weak value approaches do not tell us anything about the underlying system either, beyond standard quantum mechanics.
A key motivation for weak value approaches was to explain how the results of interference are affected by changes to disconnected regions. They were intended to show that phenomena such as Wheeler's Delayed Choice, or Salih et al's Counterfactual Communication protocol \cite{Salih2013Protocol,Hance2021Quantum} and related effects \cite{Salih2016Qubit,Salih2018Paradox,Hance2020CFGI,Salih2021EFQubit,Salih2020DetTele} aren't as ``spooky" as they appear \cite{Vaidman2014SalihCommProtocol,Vaidman2015Counterfactuality,Vaidman2016Counterfactual,Vaidman2019Analysis}. Salih et al however have shown that in their communication protocol, the weak value of the spatial projection operator (or any compound operator including the spatial projection operator) at Bob is always zero when Alice receives the quantum particle (see Fig.\ref{fig:Salih}) \cite{Salih2018Laws}. Aharonov and Vaidman have also given an altered protocol for counterfactual communication where the weak value of the spatial projection operator (or any compound operator including the spatial projection operator) at Bob is zero \cite{Aharonov2019Modification,Wander2021ApprCF}. These both show that weak value approaches do not explain this phenomenon.
While suggesting a time-symmetry to quantum processes through the TSVF, weak value approaches do not imply any new physics beyond standard quantum theory \cite{Vaidman1996WeakElReality}, to explain the causes of counterintuitive quantum effects when there is a nonzero weak value of the spatial projection operator. These approaches simply assume the particle was present wherever the weak value of an operator containing the spatial projection operator is nonzero. Hence they add a claim of particle presence but contributes nothing testable to our ontology. This label confuses matters by oversimplifying a complex concept: what it means for a specific particle to be sequentially ``present" at a two specific places in quantum field theory. It also causes a number of paradoxes by itself, such as the discontinuous photon trajectories in the nested interferometer set-up above.
\section{Conclusion}
Not only do weak value approaches to particle presence give incoherent results (claiming that a particle can be in $B$, or in $C$, but not in $B$ or $C$), but they also rest on the faulty assumption that weak coupling is equivalent to no coupling. Regardless of disturbance, a non-zero weak value of an operator containing the spatial projection operator is insufficient for particle presence.
\textit{Acknowledgements -} We thank Hatim Salih and Sophie Inman for useful discussions. This work was supported by the University of York's EPSRC DTP grant EP/R513386/1, and the Quantum Communications Hub funded by EPSRC grants EP/M013472/1 and EP/T001011/1.
\bibliographystyle{unsrturl}
|
\section{Introduction}
\label{sec.intro}\ghost{sec.intro}
The properties of mixed anion compounds strongly depend on the degree of order or disorder of the anions.
Anionic ordering can cause heteroleptic coordinations or low-dimensional structure
which in turn modify electronic properties.
This intriguing possibility to modify a material has attracted significant attention~\cite{2018KAG,2018KUR_AST,2018OSH_UND,2020OSH_TWO}.
A typical example are ABX$_3$ perovskites.
For example, SrTaO$_2$N and BaTaO$_2$N have a high dielectric constant
due to the O/N anion order~\cite{2004KIM_CHA,2007PAG_LOC,2010YAN_ANI}.
In addition, an oxyhydride SrVO$_2$H shows the two-dimensional
electronic conduction and compression anisotropy
due to the O/H anion order~\cite{2017YAM_THE}.
\vspace{2mm}
Multiple anions and cations are heterolepltically coordinated and
their concentrations typically obey valence charge neutrality conditions.
When differences in ionic radii, electronegativity or polarizability are large,
materials tend to exhibit ionic ordering~\cite{2013JPA,2006AF}.
However,
although O$^{2-}$ and F$^{-}$ are neighboring anions in the periodic table,
the oxyfluorides exhibit an anionic ordering depending on the structure.
While simple cubic perovskites
(SrFeO$_2$F~\cite{2014THO},
BaFeO$_2$F~\cite{2011BER},
PbScO$_2$F~\cite{2008KAT},
BaScO$_2$F~\cite{1998NEE},
AgFeOF$_2$~\cite{2018TAK},
BaInO$_2$F~\cite{2019KAT},
and AgTiO$_2$F~\cite{2020INA})
adopt disordered configurations,
a variety of Ruddlesden-Popper
type layered perovskites
(Sr$_2$CuO$_2$F$_2$~\cite{1994AI},
Sr$_2$FeO$_3$F~\cite{1999SIM,2001HEC},
Ba$_2$InO$_3$F~\cite{1995NEE},
Ba$_2$ScO$_3$F~\cite{1996NEE},
Sr$_2$MnO$_3$F~\cite{2016SU},
Sr$_3$Fe$_2$O$_{5-x}$F$_y$~\cite{1999SIM,2011TSU})
exhibit ordered configurations
of F$^-$~\cite{1998DU}.
In layered perovskites, multiple distinct anion sites are present and support anion order.
This manner can be interpreted by considering \tadd{Pauling's second rule},
where a stable ionic structure is arranged to compensate the valence of
an anion by the bond strength of coordinated cations. Thus, F$^-$ prefers more open site compared to O$^{2-}$,
leading to O/F anion order.
\tadd{
On the other hand, there are some perovskites with relatively simple structures
showing the partial anion ordering, such as PbFeO$_2$F~\cite{2005INA,2012KAT},
Pb$_2$Ti$_2$O$_{5.4}$F$_{1.2}$~\cite{2016OKA}, PbOF$_2$~\cite{2019INA},
and Pb$_2$Ti$_4$O$_9$F$_2$~\cite{2015OKA}.
Their anion order is considered the consequence of the Jahn-Teller distortion relevant
to the 6$s^2$ lone pair~\cite{2012KAT,2016OKA,2015OKA}.
}
\vspace{2mm}
To investigate the role played by the 6$s^2$ lone pair,
we concentrate on the oxyfluorides containing Pb or Bi~\cite{2019INA,2015OKA,2016OKA}.
For an ideal analysis, completely isostructural oxyfluoride and oxide (as a reference) would be desirable,
though such a pair of compounds are rarely found.
Pyrochlore oxyfluoride and oxide,
Pb$_2$Ti$_2$O$_{5.4}$F$_{1.2}$
and Bi$_2$Ti$_2$O$_7$~\cite{2004HEC},
would have the right stoichiometry,
but these compounds belong to different space groups
[Pb$_2$Ti$_2$O$_{5.4}$F$_{1.2}$ ($Fd$-43$m$)
and Bi$_2$Ti$_2$O$_7$($Fd$-3$m$)].
Another candidate would be
PbOF$_2$ and Bi$_2$O$_3$, since
PbOF$_2$ adopts a similar structure
to $\beta$-Bi$_2$O$_3$~\cite{2009HUL},
but they also belong to different space groups
[PbOF$_2$~($P4_{2}/nmc$) and $\beta$-Bi$_2$O$_3$~($P$-42$_{1}c$)].
\vspace{2mm}
Eventually, we found an isostructural pair
Pb$_2$Ti$_4$O$_9$F$_2$~\cite{2015OKA} and oxide Bi$_2$Ti$_4$O$_{11}$~\cite{1995KAH}.
Both Pb$^{2+}$ and Bi$^{3+}$ have the same electron configuration and a 6$s^2$ lone pair.
Bi$_2$Ti$_4$O$_{11}$ undergoes antiferroelectic-paraelectric transitions from $C2/c$ to $C2/m$,
while Pb$_2$Ti$_4$O$_9$F$_2$ does not.
The high-temperature paraelectric phase of
Bi$_2$Ti$_4$O$_{11}$~($C2/m$) adopts
the same space group symmetry as Pb$_2$Ti$_4$O$_9$F$_2$~\cite{1995KAH}.
Thus, the system provides ideal platform
to investigate the role of 6$s^2$ lone pair
on O/F anion arrangement.
\vspace{2mm}
An earlier work by Oka et al. found that F$^-$ selectively occupies
the closest anionic site to Bi/Pb site in {Pb$_2$Ti$_4$O$_9$F$_2$}.~\cite{2015OKA}.
The anionic site was named site-6.
They observed that the fluorine occupation elongates the distance
between the site-6 and Bi/Pb site.
Oka et al. concluded that the elongation caused the disappearance
of antiferroelectric transition in {Pb$_2$Ti$_4$O$_9$F$_2$}.
\vspace{2mm}
In this paper, to understand what is unique at site-6,
we carefully \tadd{analyzed and compared}
\tadd{{Pb$_2$Ti$_4$O$_9$F$_2$} and {Bi$_2$Ti$_4$O$_{11}$}.}
\tadd{
The anion configuration around Bi is more distorted than that of Pb,
where the O$^{2-}$ at the site-6 is significantly closer to Bi than the other surrounding oxygen ions.
We established that this difference between {Pb$_2$Ti$_4$O$_9$F$_2$} and {Bi$_2$Ti$_4$O$_{11}$} attributed to that 6$s^2$ lone-pair more strongly
hybridizes with oxygen than fluorine at the site-6.
}
\tadd{
Here, the question is why site-6 is preferred by fluorine ions in {Pb$_2$Ti$_4$O$_9$F$_2$}. To answer this question,
}
we performed ab initio calculations.
Though the total energy is surely stabilized by the site-6 occupation,
it is not fully stabilized due to the electrostatic
\tadd{electrostatic energy}, being an exception of \tadd{Pauling's second rule}.
Our structural and bonding nature analyses showed that
the site-6 occupation weakens the anti-bonding orbital,
between Pb and site-6 anion, distributing on the other side of the site-6 across Pb.
The change of the anti-bonding orbital reduces the steric effect
that hampers Pb to bond with the site-2 oxygen on the other side of the site-6.
This can be the reason of the deviation from \tadd{Pauling's second rule}.
\vspace{2mm}
We organize the rest of the paper as follows.
In Section \ref{sec.methods}, we describe the details of the synthesis
and synchrotron X-ray powder diffraction (SXRD) and the calculation conditions.
In Section \ref{sec.results1},
we compare and discuss the geometric and electronic structures of {Pb$_2$Ti$_4$O$_9$F$_2$} with the isostructural oxide,
the high-temperature (HT) phase of {Bi$_2$Ti$_4$O$_{11}$}, and the low-temperature (LT) phase of {Bi$_2$Ti$_4$O$_{11}$}.
In Section \ref{sec.results2}, we show the total energies and electrostatic energies
for the possible five anion orders to show that their lowest energies structures are inconsistent.
This is not explained by \tadd{Pauling's second rule}.
We discuss the reason focusing on the orbital hybridizations.
We summarize our work in Section \ref{sec.conc}.
\section{Methods}
\label{sec.methods}\ghost{sec.methods}
The powder samples of {Bi$_2$Ti$_4$O$_{11}$} and {Pb$_2$Ti$_4$O$_9$F$_2$} were prepared by solid-state reaction.
as previously reported~\cite{2015OKA,1995KAH}.
Bi$_2$Ti$_4$O$_{11}$ was synthesized
from a stoichiometric mixture of Bi$_2$O$_3$ (99.9\%, Rare Metallic Co.) and
TiO$_2$ (99.9\%, Rare Metallic Co.) powders. The pelletized mixture was treated at 1273 K
for 12 h in the air.
Pb$_2$Ti$_4$O$_9$F$_2$ was synthesized from
\tadd{
a mixture of PbO (99.9~\%, Rare Metallic Co.),
PbO$_2$ (99.9~\%, Rare Metallic Co.),
PbF$_2$ (99.9~\%, Rare Metallic Co.),
and TiO$_2$ (rutile, 99.9~\%, Rare Metallic Co.)
}
powders were weighed to be \tadd{10~mol\% fluorine-rich to compensate the loss of fluorine}
during the reaction~\cite{2012KAT}. The pelletized mixture was sealed in an evacuated Pyrex tube
and treated at 823 K for 12 h.
\vspace{2mm}
SXRD
patterns were collected with a large Debye–Scherrer camera
installed at beamline BL02B2 of SPring-8
using a glass capillary and a solid state detector \cite{2017KAW}.
The crystallographic parameters were
refined by the Rietveld method using the RIETAN-FP program~\cite{2007IZU}.
The electron density distributions were \tadd{estimated} by the maximum entropy method (MEM)
using the Dysnomia program~\cite{2013MOM}. The crystal structures and electron density mappings
are visualized using VESTA program~\cite{2011MOM}.
\vspace{2mm}
The density functional theory (DFT) calculations were
performed with Quantum Espresso~\cite{2009GIA}.
\tadd{PBE~\cite{PBE} and BLYP~\cite{BLYP1,BLYP2} semi-local functionals were employed.}
The core orbitals were described by projector augmented wave (PAW) method~\cite{Kresse1999}.
\tadd{PAW pseudopotentials were taken from the pslibrary~\cite{2014ADC}.}
The valence orbitals were expanded with plane waves.
The cutoff energy was 100 Ry and the $k$-point mesh was 7$\times$7$\times$5 for the unit cell.
The LOBSTER code~\cite{2020NEL} was used to perform
the Crystal Orbital Hamiltonian Population (COHP) analyses based on the DFT results
with pbeVaspFit2015 basis set~\cite{2013MAI,2016MAI}.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\hsize]{position-eps-converted-to.pdf}
\caption{
\label{fig.position}\ghost{fig.position}
Local coordination environments around Pb/Bi for
Pb$_2$Ti$_4$O$_9$F$_2$ and Bi$_2$Ti$_4$O$_{11}$.
The annotations mean the bond lengths (\AA) between Pb/Bi-O/F.
}
\end{figure*}
\section{Comparison between $\bm{\mathrm{Bi_{2}Ti_{4}O_{11}}}$ and $\bm{\mathrm{Pb_2Ti_4O_9F_2}}$}
\label{sec.results1}\ghost{sec.results1}
Production of single phase for both Bi$_{2}$Ti$_{4}$O$_{11}$
and Pb$_2$Ti$_4$O$_9$F$_2$ samples were confirmed by SXRD technique.
\tadd{The SXRD patterns and results of Rietveld refinement
for the LT~(300~K) and HT~(600~K) phases of Bi$_2$Ti$_4$O$_{11}$
and Pb$_2$Ti$_4$O$_9$F$_2$ (300~K) are given in the supporting information.
}
Bi$_2$Ti$_4$O$_{11}$ undergoes structural
transition from low-temperature (LT) antiferroelectric
(Space group: $C2/c$) to high-temperature (HT)
paraelectric phase (Space group: $C2/m$) at 506~K~\cite{1995KAH}.
As previously reported, Pb$_2$Ti$_4$O$_9$F$_2$
is isostructural with the HT paraelectric phase of
Bi$_2$Ti$_4$O$_{11}$ and does not show structural
transition down to 5~K~\cite{2015OKA}.
\vspace{2mm}
The local coordination environments around Bi/Pb
in Bi$_2$Ti$_4$O$_{11}$ and Pb$_2$Ti$_4$O$_9$F$_2$
are shown in Figure~\ref{fig.position}.
The numbers at the anions indicate the site indices.
We provide the experimental crystal structures in the CIF format as the supporting information,
as one can distinguish the anion sites with their own visualizing softwares.
The shortest Bi/Pb-O/F6 bond lengths are largely different
between {Bi$_2$Ti$_4$O$_{11}$} and {Pb$_2$Ti$_4$O$_9$F$_2$}.
Pb-F6 of {Pb$_2$Ti$_4$O$_9$F$_2$} is significantly longer than Bi-O6 of the HT phase of {Bi$_2$Ti$_4$O$_{11}$}.
Here, O6 (F6) indicates an oxygen (fluorine) ion in the site-6.
The elongated bond length reduces the degree of anisotropy
of the ligand configurations around Bi/Pb as shown in Figure \ref{fig.position}.
The antiferroelectric transition in {Bi$_2$Ti$_4$O$_{11}$} is known to be associated with
displacements of the Bi ion from the centrosymmetric position along the $b$ axis~\cite{1995KAH,1995NIS}.
Therefore,
the disappearance of antiferroelectric transition for Pb$_2$Ti$_4$O$_9$F$_2$
would be related to the reduced anisotropic character around Pb$^{2+}$ ion.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\hsize]{density.eps}
\caption{
\label{fig.density}\ghost{fig.density}
\tadd{
Electron density distributions obtained by MEM for Pb$_2$Ti$_4$O$_9$F$_2$ and Bi$_2$Ti$_4$O$_{11}$.
The isosurface levels for 2 $e$/\AA$^3$, 4 $e$/\AA$^3$ and 12 $e$/\AA$^3$ are shown in green, yellow and red colors, respectively.
}
}
\end{figure*}
\begin{figure}[htbp]
\centering
\includegraphics[width=\hsize]{2ddense.eps}
\caption{
\label{fig.2ddense}\ghost{fig.2ddense}
\tadd{
Local coordination environments around Pb/Bi for Pb$_2$Ti$_4$O$_9$F$_2$ and Bi$_2$Ti$_4$O$_{11}$.
The numbers on the anion sites represent the sites summarized in Table S1, S2 and S3
in the supporting information.
}
}
\end{figure}
\vspace{2mm}
The changes in local coordination environment around Bi$^{3+}$
or Pb$^{2+}$ were further investigated.
Figure \ref{fig.density} shows the electron density distribution obtained by MEM.
The differences of covalent character of the Bi-O6 bonds
between the HT and LT phases of Bi$_2$Ti$_4$O$_{11}$ are obvious.
This anisotropic change in covalency around Bi$^{3+}$
is associated with the anti-parallel
displacement of Bi$^{3+}$ ion along the $b$ axis in the LT phase,
suggesting $6s^2$ lone pair electrons at play.
Therefore, we infer that the absence of antiferroelectric
transition in Pb$_2$Ti$_4$O$_9$F$_2$ is
a consequence of the reduction on the interaction between the Pb-$6s$ and anion-$2p$ electrons.
Figure~\ref{fig.2ddense} shows a 2-dimensional electron
density mapping sliced along Bi/Pb-O/F6 bond obtained by MEM.
The electron density associated to the covalent character of Bi-O6 is observed
in Bi$_2$Ti$_4$O$_{11}$ at the HT phase.
On the other hand, the covalency between Pb and F6 is less pronounced
in Pb$_2$Ti$_4$O$_9$F$_2$.
These results suggest that O/F anion order in Pb$_2$Ti$_4$O$_9$F$_2$ occurs
to reduce the interaction between $6s$ and $2p$ electrons.
We confirmed that the
\vspace{2mm}
We verified the experimental conclusion theoretically
performing a COHP analysis to the hybrid orbital of Pb/Bi-6$s$ and O/F-2$p$.
The COHP analysis quantifies the degree of (anti-)bonding strength of
the hybrid orbital evaluating the integrated COHP (ICOHP).
Positive (negative) ICOHP indicates anti-bonding (bonding) nature.
The absolute value indicates a degree of (anti-)bonding strength.
We obtained ICOHP(Pb-6$s$ -- F6-2$p$) = -0.032 and ICOHP(Bi-6$s$ -- O6-2$p$) = -0.289:
Pb-6$s$ and F6-2$p$ make a significantly
weaker bonding hybrid orbital than Bi-6$s$ and O6-2$p$.
\tadd{
The stronger bonding hybridization between Bi-6$s$ and O6-2$p$
makes Bi easily displaces towards the site 6 than Pb.
This explains the relatively high temperature of the antiferroelectric
transition ($C2/m \to C2/c$ in Figure \ref{fig.density}) of {Bi$_2$Ti$_4$O$_{11}$}.
}
\vspace{2mm}
\tadd{
We also confirmed that DFT and MEM give qualitatively similar results
as shown in the supporting information.
}
\begin{table}[htbp]
\begin{center}
\caption{
\label{tab.TotalEnergy}\ghost{tab.TotalEnergy}
Comparisons of the total energies
fo the different site occupations by F$^-$ of
of Pb$_2$Ti$_4$O$_9$F$_2$.
The F-in-site6 is taken as the reference
for the energy deviations, $\Delta E$,
given in the unit of [eV/f.u.].
}
\begin{tabular}{ccc}
\hline
& $\Delta E$ (PBE) & $\Delta E$ (BLYP) \\
\hline
F-in-site1 & 1.019 & 0.965 \\
F-in-site2 & 0.935 & 0.889 \\
F-in-site4 & 0.321 & 0.308 \\
F-in-site5 & 0.061 & 0.066 \\
F-in-site6 & 0.000 & 0.000 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.495\hsize]{pbe.eps}
\includegraphics[width=0.495\hsize]{blyp.eps}
\caption{
\label{fig.static}\ghost{fig.static}
Comparision of electrostatic energies and total energies calculated by DFT.
}
\end{figure*}
\section{Reason of the F-in-site6 preference in $\bm{\mathrm{Pb_2Ti_4O_9F_2}}$}
\label{sec.results2}\ghost{sec.results2}
We performed DFT ab initio calculations to evaluate the energetics
of the F$^-$ occupation patterns of {Pb$_2$Ti$_4$O$_9$F$_2$}.
\tadd{
We calculated the energy of each pattern with PBE~\cite{PBE} and BLYP~\cite{BLYP1,BLYP2}
functionals as shown in Table~\ref{tab.TotalEnergy}.
Here, for example, F-in-site2 indicates the pattern that fluorine selectively occupies the site 2.
F-in-site 3 is not considered because the multiplicity of site 3 is 1 per f.u.,
but there are 2 fluorine atoms per f.u.
Both functionals reproduce the experimental F-in-site6 preference.
The two results are very close to each other and qualitatively identical.
Therefore, the DFT result would not significantly alter
depending on the choice of functional.
}
To understand the origin of F-in-site6 stabilization,
we compared
\tadd{electrostatic energies}
\tadd{
using PBE and BLYP functionals, shown as the vertical axis values in Figure~\ref{fig.static}.
To get the electrostatic energies, we used ultrasoft pseudopotentials provided in the pslibrary~\cite{2014ADC},
since we could not get the electrostatic energies from the quantum espresso output with PAW pseduopotential.
This is because the one-center term includes both electrostatic and exchange-correlation energies~\cite{1994PEB}.
We confirmed that the relative energies of fluorine occupation patterns are identical
between PAW and ultrasoft pseudopotentials within 1~meV/f.u.
Figure~\ref{fig.static} shows that PBE and BLYP qualitatively agree with each other.
}
The total energies are roughly proportional to the electrostatic energies,
consistent with a previous systematic study on NdNiO$_2$F by simulations~\cite{2019KUR}.
Meanwhile, the F-in-site6 cannot be fully explained
by \tadd{Pauling's second rule} (structural stabilization via electrostatic gain),
requiring us to investigate further the origin of the F-in-site6 stabilization.
\tadd{
(In the supporting information, we discussed the electrostatic energies approximated
by Madelung potentials with different point charge models. )
}
\vspace{2mm}
\tadd{
Later in this paper, we use only PBE functional because PBE and BLYP agree with
each other about the total and electrostatic energies.
}
\begin{figure*}[p]
\centering
\includegraphics[width=\hsize]{pdos.eps}
\caption{
\label{fig.pdos}\ghost{fig.pdos}
Partial density of states predicted by DFT.
The arrows indicate the peaks accompanied by
\tadd{the anti-bonding hybrid orbitals
explained by the lone-pair model.}
}
\end{figure*}
\begin{table}[htbp]
\begin{center}
\caption{
\label{tab.integral}\ghost{tab.integral}
Energy differences (F-in-site5 - F-in-site6) of the valence orbital channels.
}
\begin{tabular}{cr}
\hline
Orbital channel & Energy (eV) \\
\hline
F-$2s$ & +0.098 \\
F-$2p$ & -0.062 \\
O-$2s$ & -0.206 \\
O-$2p$ & \textbf{+1.288} \\
Pb-$6s$ & +0.820 \\
Pb-$6p$ & +0.078 \\
Ti-$3s$ & +0.005 \\
Ti-$3p$ & +0.009 \\
Ti-$3d$ & +0.095 \\
Ti-$4s$ & +0.028 \\
\hline
Total & +2.15 \\
\hline
\end{tabular}
\end{center}
\end{table}
\vspace{2mm}
We focused on the the orbital energies to explain the F-in-site6 stabilization.
\tadd{
One can understand qualitatively the stability
of each orbital channel analyzing its integrated partial density of states (PDOS).
In Kohn-Sham DFT, the kinetic energy is given as
the sum of the orbital eigen values minus a contribution
that largely cancels out if the density in and around
the orbitals in the F atoms is similar~\cite{2004RMM}.
The integrated PDOS corresponds to the sum of
the orbital eigen values of the orbital channel.
}
Figure \ref{fig.pdos} shows the PDOS for each F$^-$ occupation pattern in {Pb$_2$Ti$_4$O$_9$F$_2$}.
Except for F-in-site6, a peaky state with \tadd{Pb-6$s$ component} exists
at the upper edge of the valence band, as indicated by the arrow.
\tadd{
This peaky state signals the orbital hybridization described by the revised lone-pair (RLP) model~\cite{2011WAL}.
This model proposes that the anti-bonding orbital hybridized by Pb-6$s$ and O-2$p_{z}$ is stabilized
by further \tadd{hybridizing} with the Pb-6$p$ orbital.
The peaky state corresponds to the anti-bonding orbital of this hybridization.
This anti-bonding orbital distributes near Pb in opposite to the oxygens.
This charge distribution would function as a steric effect for interactions between the Pb and other anions.
}
\vspace{2mm}
\tadd{The PDOS result} indicates that the hybridization occurred mainly between Pb and O6.
The energy of orbital $\alpha$ is given as the integration of $\varepsilon \cdot D\left(\varepsilon\right)$,
where $D_\alpha\left(\varepsilon\right)$ is the PDOS of the orbital $\alpha$ at the energy $\varepsilon$.
We performed the PDOS integration for each orbital in the valence band range, [-10eV:0eV]
($\varepsilon$=0~eV is the Fermi energy).
Then, we compared the orbital energy differences between F-in-site5 and F-in-site6,
as shown in Table \ref{tab.integral}:
the orbital energy of O-2$p$ was significantly higher for F-in-site5 than F-in-site6.
\begin{table}[htbp]
\begin{center}
\caption{
\label{tab.intoxy}\ghost{tab.intoxy}
Energy differences (F-in-site5 - F-in-site6) of the 2$p$ channels of the anions.
}
\begin{tabular}{cc}
\hline
$\alpha$ (orbital) & $\int \varepsilon \cdot \Delta D_\alpha (\varepsilon)~d\varepsilon$ \\
\hline
O1-2$p$ & +0.119 \\
O2-2$p$ & \textbf{+0.335} \\
O3-2$p$ & +0.015 \\
O4-2$p$ & -0.064 \\
O5/6-2$p$+F6/5-2$p$ & +0.191 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure*}[htbp]
\centering
\includegraphics[width=\hsize]{pbo2.eps}
\caption{
\label{fig.pbo2}\ghost{fig.pbo2}
The crystal structures around the Pb and O2 of the F-in-site5 and F-in-site6.
}
\end{figure*}
\vspace{2mm}
The 2$p$ orbital energy difference of each oxygen between F-in-site5 and F-in-site6 is shown in Table~\ref{tab.intoxy}.
(For sites 5 and 6, we compared O5-2$p$+F6-2$p$ for F-in-site5 and F5-2$p$+O6-2$p$ for F-in-site6.)
The result shows that 2$p$ orbital energy of O2 differs the most.
O2 is located on the other side of O/F6 across the Bi/Pb as shown in Figure~\ref{fig.pbo2}.
The figure shows the distance and ICOHP between Pb and O2 and the values indicate weaker binding strength for the F-in-site5.
This would be attributed to that the Pb-O6 hybridization for F-in-site5 is stronger than the Pb-F6 hybridization for F-in-site6.
The hybridization makes an anti-bonding orbital distributing in the other side of the O/F6 across the Bi/Pb, i.e., between the Bi/Pb and O2.
Therefore, the anti-bonding orbital caused by the 6$s^2$ lone pair would hamper the binding between Bi/Pb and O2 as a steric effect.
Since the hybridization is stronger for F-in-site5 than F-in-site6, the steric effect would be also stronger for F-in-site5.
This is a significant origin of the F-in-site6 preference in {Pb$_2$Ti$_4$O$_9$F$_2$}, which cannot be fully explained by \tadd{Pauling's second rule}.
\section{Conclusion}
\label{sec.conc} \ghost{sec.conc}
To understand the cause of the F$^-$ ordering,
preferably occupying a specific site, Bi$_2$Ti$_4$O$_{11}$
and Pb$_2$Ti$_4$O$_9$F$_2$ were simultaneously investigated
by SXRD technique and ab initio calculations.
In {Bi$_2$Ti$_4$O$_{11}$}, the tetragonal instability and the accompanying
anisotropic coordination result in an antiferroelectric transition.
Meanwhile, Pb$_2$Ti$_4$O$_9$F$_2$ does not undergo an antiferroelectric transition down to 5~K,
because F$^-$ selectively occupies the closest anion site (site-6) to Bi/Pb,
and then F$^-$ elongates the Bi/Pb-O/F6 distance and reduces the distortion around the Bi/Pb.
We found that the binding hybridization between Pb-6$s$ and F6-2$p$ is
significantly weaker than Bi-6$s$ and O6-2$p$, explaining the longer distance of Pb-F6 than Bi-O6.
Ab initio calculations were performed to analyze the F-in-site6 preference.
Though the F-in-site6 is surely observed experimentally to be the most stable configuration,
it is not fully explained by the \tadd{electrostatic energy},
being an exception of Pauling's second rule.
To understand the reason, we focus on the orbital energies.
The PDOS analysis shows that the O2-2$p$ is significantly
more stable for F-in-site6 than F-in-site5.
From the structural and COHP analyses, we conclude that
the anti-bonding hybrid orbital between Pb and O/F6,
explained by the RLP model~\cite{2011WAL}, hampers the binding
between Pb and O2 more strongly for the F-in-site5 than F-in-site6.
\vspace{2mm}
\section*{Acknowledgments}
The synchrotron radiation experiments were performed at the BL02B2 of SPring-8 with the approval
of the Japan Synchrotron Radiation Research Institute (JASRI) (Proposal No.~2016A1157 and No.~2018A1227).
The ab initio calculations were performed with the computational resources
of the Research Center for Advanced Computing Infrastructure (RCACI) at JAIST.
T.I. and F.A.R. acknowledge support from US Department of Energy, Office of Science, Basic Energy Sciences,
Materials Sciences and Engineering Division.
This work was also partially supported by a Grant-in-Aid for Scientific
Research on Innovative Area “Mixed Anion (Project, JP17H05489, JP19H04706)” (JSPS),
and a Grant-in-Aids for Scientific Research (C)
(Project JP16K05731 and 21K04659).
\tadd{
K.H. is grateful for financial support from
MEXT-KAKENHI (JP16H06439, JP19K05029, JP19H05169, and JP21K03400),
and the Air Force Office of Scientific Research
(Award Numbers: FA2386-20-1-4036).
}
\tadd{
R.M. is grateful for financial supports from
MEXT-KAKENHI (21K03400 and 19H04692),
from the Air Force Office of Scientific Research
(AFOSR-AOARD/FA2386-17-1-4049;FA2386-19-1-4015),
and from JSPS Bilateral Joint Projects (with India DST).
}
|
\section{Introduction and literature survey}
\label{sec:introduction}
Several applications in biotechnology, bio-process engineering, polymer processing and others involve
bubbly flows in rheologically complex liquids.
It is then of core relevance to understand the rise behavior of bubbles in viscoelastic liquids, since this determines in particular the residence (i.e., the contact) time.
Since the pioneering paper by Astarita and Apuzzo \cite{Astarita1965} it has been known that for single bubbles rising in quiescent viscoelastic liquids, the terminal bubble rise velocity
as a function of the bubble volume
may exhibit a jump discontinuity at a critical volume such that a supercritical bubble rises up to an order of magnitude faster than a subcritical one.
We call this \emph{the bubble rise velocity jump discontinuity}.
At this critical bubble size, the terminal rise velocity can experience a discontinuous increase by up to an order of magnitude, depending on the material properties of the liquid.
Astarita and Apuzzo reported this jump in the bubble rise velocity to be accompanied by a change in the bubble shape from a convex to a “teardrop”-shaped surface. As a possible
explanation, the authors of \cite{Astarita1965} suggested that the jump of the bubble rise velocity corresponds to a change of the interfacial mobility from rigid to free.
This switch in the boundary conditions at the bubble surface
would then result in a transition similar to the one from the Stokes to the Hadamard-Rybczynski regime for a Newtonian ambient fluid. They also showed that, for a shear-thinning liquid, the ratio of the steady
rise velocities just above and below the critical volume can be much larger than the well-known factor of 1.5 in the Newtonian case.
Some time later, the suggested explanation was supported by
investigations of the mass transfer coefficients
of single carbon dioxide bubbles in an aqueous polyethylene oxide solution by Calderbank et al.\ \cite{Calderbank1970}.
Moreover, experiments with glass spheres moving in viscoelastic liquids by Leal et al. \cite{Leal1971} showed no discontinuity of the settling velocity as a function of the sphere volume, hence also supporting the relevance of the surface mobility for the velocity jump to occur. Furthermore, a numerical analysis based on the creeping flow equations for shear-thinning liquids, but without elastic effects, indicated that the height of the experimentally observed velocity jump cannot be explained without accounting for elastic effects.
Acharya et al.\ suggested in \cite{Acharya1977} that polymer molecules act as surface-active agents (surfactants), generating surface stresses which dampen the interfacial
motion, in particular in the downstream part of the bubble. In this case, a partial cleansing of the surface could be responsible for the rapid velocity change, and the authors claim that this is likely to occur much more abruptly in the case of a viscoelastic liquid rather than for Newtonian fluids. Analyzing the available data, they also derived a criterion for the critical bubble size, saying that a certain Bond number is of order of unity,
\begin{equation}
\textnormal{Bo} = \frac{\rho_l g R_c^2}{\sigma} \approx 1
\end{equation}
(liquid density $\rho_l$, gravitational acceleration $g$, critical bubble radius $R_c$, and surface tension $\sigma$ against the gas in the bubble). This turned out inappropriate for predicting the discontinuity; note that Bo can be unity for Newtonian fluids, while no rise velocity jump discontinuity has ever been observed in this case. Despite this, the study of surfactant effects has been continued later by others, as will be briefly mentioned below.
Liu et al.\ \cite{Liu1995} studied the formation of a cusp at the rear end of bubbles rising in viscoelastic liquids. More precisely, the bubble shape displays a conical point at the rear end with a locally concave profile in the vertical section, i.e.\ a “teardrop”-like shape. They related the cusp formation and the associated increase in rise velocity to a critical capillary number, viz.\
\begin{equation}
\textnormal{Ca} = \frac{\eta_0 U_T}{\sigma} \approx 1
\label{eq:capnumber}
\end{equation}
(zero-shear viscosity $\eta_0$ and terminal bubble rise velocity $U_T$). Recall that the capillary number expresses the ratio between viscous forces and surface tension forces, while elastic forces are not included. The authors of \cite{Liu1995} have shown that data published previously by several authors also seem to correlate with a critical capillary number of order one. They also noted that the presence of a cusp is not sufficient for the appearance of a velocity jump. This has been confirmed, for instance, by De Kee et al.\ \cite{DeKee1990,Dekee1986}, Rodrigue and Blanchet \cite{Rodrigue2002} and Pilz and Brenn \cite{Pilz2007}, where, despite a clearly visible cusp at the rear end, in some cases no jump appeared.
Extensive investigations of the effect of surfactants on the rise velocity jump discontinuity of bubbles in polymer solutions can be found in \cite{Rodrigue2002} - \cite{Rodrigue1998}. Rodrigue et al.\ \cite{Rodrigue1998} suggested that an appropriate jump criterion must represent elastic, viscous and capillary forces acting on the bubble. Their jump criterion states that a quantity $\alpha = \textnormal{Ca} \, \textnormal{De} / \textnormal{Ma}$ ($\textnormal{Ca}$ - capillary, $\textnormal{De}$ - Deborah, $\textnormal{Ma}$ - Marangoni number) changes its value from $< 1$ to $> 1$ upon transition from sub- to supercritical bubble state \cite{Rodrigue1998}. In the formulation of the quantity $\alpha$, the capillary number accounts for the shear-rate dependence of the dynamic viscosity of the polymer solution. The Deborah number accounts for the liquid elasticity by virtue of the first normal stress difference, which is also formulated as shear-rate dependent. In Rodrigue et al.\ \cite{Rodrigue1996} it is furthermore pointed out that polymer molecules act as surfactants, since the presence of a polymer alters the surface tension of the liquid against the gas phase. The authors proposed that the discontinuity is the result of an instability at the gas-liquid interface caused by an imbalance of forces, and that the origin of the instability should be related to normal forces which, under certain conditions, may extract surfactant as well as polymer molecules from the bubble surface, leaving a zone of different interfacial and rheological properties. Therefore, a Marangoni stress due to a surface tension gradient appears. According to this hypothesis, Rodrigue et al.\ \cite{Rodrigue1998} gave a physical interpretation of the discontinuity as follows. At the rear stagnation point, where the local strain is large, causing strong curvature and local deformation, high normal stresses are developed. Polymer and/or surfactant molecules are stretched along the liquid streamlines, and therefore induce a change of the fluid properties. The jump could be the result of the normal forces acting in the vicinity of the bubble, removing molecules from the bubble surface or causing a sudden change in the interfacial conditions. In the work of Rodrigue and Blanchet \cite{Rodrigue2002} it is shown that the jump can be eliminated by using surfactant concentrations above the critical association concentration (CAC). From their observations they concluded that the origin of the jump is most likely related to a change in interfacial conditions due to an imbalance in surface tension gradient and elastic forces at the gas-liquid interface.
Since the presence of a surfactant introduces additional complexity in the nonlinear behavior underlying an already complicated phenomenon, and the rise velocity jump phenomenon appears in systems without surfactant, we do not go into further details on this topic.
Soto et al.\ \cite{Soto2006} investigated the rise behavior of single air bubbles in solutions of a hydrophobically modified alkali soluble polymer (HASE). These liquids exhibit a nearly constant shear viscosity over a wide range of shear rates, so that velocity enhancements due to shear thinning and due to elasticity can be distinguished. They compared the mean shear rate $\bar{{\dot \gamma}} = 2 U_T / d$ calculated from the terminal bubble rise velocity $U_T$ and the sphere-equivalent bubble diameter $d$ with data obtained from shear experiments. The mean shear rate, which can be attributed to the rise velocity jump, is approximately the same as the shear rate where the first normal stress difference $N_1$ became detectable in the shear experiments. The authors concluded that the jump can be directly related to elasticity, since the shear viscosity around the “jump shear rate” can be regarded as constant. In their opinion, the presence of significant normal stresses causes a change of the bubble shape, which finally leads to an abrupt drag reduction. They were able to classify the data of their rise velocity experiments by introducing the non-dimensional group
\begin{equation}
\Pi = \frac{N_1 d}{\sigma},
\label{eq:group}
\end{equation}
which relates elastic forces to interfacial forces, the latter represented by $\sigma / d$. The jump appears at a critical value of $\Pi_{crit} \approx 0.25$ for their test liquids.
Flow visualization measurements by Funfschilling and Li \cite{Funfschilling2001} (Particle-Image Velocimetry (PIV) and birefringence) showed flow fields around bubbles rising in aqueous polyacrylamide solutions which were very different from those in a Newtonian glycerol solution. For polyacrylamide solutions, the flow field around the bubble can be divided into three distinct zones: an upward flow in front of the bubble, similar to that in the Newtonian case; a downward flow in the central wake, as described earlier as a negative wake (see, e.g., Hassager \cite{Hassager1979}); finally, a hollow cone of upward flow enclosing the region of the negative wake. The birefringence visualization qualitatively revealed a butterfly-like spatial distribution of shear stresses around the bubble within any plane going through the axis of symmetry.
Employing PIV measurements, Kemiha et al.\ \cite{Kemiha2006} observed a dependency of the opening angle of the hollow cone on the bubble Reynolds number, which converges to an asymptotic value for large bubble Reynolds numbers, whereas the evolution of the angle seems to depend on the nature of the fluid. Experiments carried out with glass spheres revealed similar flow patterns, thus providing evidence that the formation of the negative wake is not a consequence of the interface deformation (cusped bubble shape). Kemiha et al.\ \cite{Kemiha2006} concluded from their observations that the formation of a negative wake is governed by the liquid rheology (i.e., its viscoelasticity), which was supported by numerical simulations of a sedimenting sphere in a viscoelastic liquid with the Lattice-Boltzmann method. The numerically calculated flow field agreed qualitatively with the measured data.
Herrera-Velarde et al.\ \cite{Herrera-Velarde2003} found that the flow configuration changes drastically at the critical bubble volume, and the flow situation described as a negative wake appears for bubble volumes larger than the critical one. They noticed further that the size of the containment affects the magnitude of the jump, but not the critical bubble volume at which the jump occurs. This observation is in agreement with the PIV measurements by Soto and co-workers \cite{Soto2006}.
Pilz and Brenn studied the rise of individual air bubbles in viscoelastic solutions of different polymers in various solvents \cite{Pilz2007}. Varying the bubble volume and measuring the steady bubble rise velocity, they showed that the jump discontinuity of the bubble rise velocity did not occur in all the liquids. The concentration of flexible polymers in the solutions needed to be high enough to enable the jump, and solutions of rigid, rod-like polymers exhibited a bend rather than a jump in the curve of the rise velocity as a function of the bubble volume. These authors developed a criterion accounting for this fact. From their experiments they also derived an equation for the non-dimensional critical bubble volume at the jump discontinuity.
A breakthrough in the numerical description of the bubble motion may be seen in the paper by Pillapakkam et al.\ \cite{Pillapakkam2007}. In contrast to the simulations known from earlier literature, this paper presents a fully three-dimensional simulation of the flow field without any assumption, of axisymmetry for instance (see, e.g., Noh et al.\ \cite{Noh1993}, Frank and Li \cite{Frank2005}, Kemiha et al.\ \cite{Kemiha2006}, or Málaga and Rallison \cite{Malaga2007}). The constitutive equation in their calculations was the Oldroyd-B model with constant shear viscosity. The gas-liquid interface was captured by a level set method, using a constant value for the interfacial tension. The calculations revealed a rise velocity jump, showing increasing magnitude of the velocity enhancement with increasing “polymer concentration factor”, and the appearance of a negative wake for supercritical bubbles. Their calculations also showed an asymmetry in the cusp for bubbles with larger volumes, which was observed experimentally, e.g., by Liu et al.\ \cite{Liu1995} and Soto et al.\ \cite{Soto2006}. The paper \cite{Pillapakkam2007} demonstrates that the rise velocity jump is possible without any variations of interfacial tension or shear viscosity, but the only requirement seems to be three-dimensional flow. The authors conclude that the asymmetric cusped bubble shape, the additional vortex ring due to the negative wake, and the change in the velocity field due to the change in the bubble shape contribute to the velocity jump at a critical bubble volume in a certain parameter range. Theoretical investigations of the bubble shape revealed that the formation of a cusp- or corner-shaped trailing end can result in an upward force, since the integral of the surface tension force over the bubble surface is no longer zero. However, an estimation of the magnitude of that additional force with the para\-meters used in their simulations revealed that this force is too small to explain the strong increase of the rise velocity. On the other hand, it turned out that the additional upward force is approximately the same for different values of the polymer concentration parameter. Therefore, the authors of \cite{Pillapakkam2007} conclude that the global modification in the liquid velocity field around the bubble is presumably the reason for the steep increase of the bubble rise velocity. This is consistent with their calculations, since both the jump magnitude and the global modification of the velocity field in the ambient fluid are more substantial for higher values of the polymer concentration parameter.
Fraggedakis et al.\ \cite{Fraggedakis2016} numerically analyzed the velocity jump discontinuity by means of a dynamical systems approach. Using an axisymmetric arbitrary Lagrangian-Eulerian mixed finite element method, they computed quasi-steady solutions of the
governing equations for a single rising bubble in
a viscoelastic liquid. In particular, they studied the
setup of the experimental results from \cite{Pilz2007},
where the ambient liquid was approximated with an exponential Phan-Thien Tanner (EPTT) constitutive equation. The approach of \cite{Fraggedakis2016} is based on a continuation method, which allows to compute a manifold of quasi-steady state solutions for varying bubble radius. For two different polymer concentrations, very
good agreement with the experimental data from \cite{Pilz2007} has been obtained for the subcritical rise velocities, both concerning the rise velocity and the bubble shape. As the bubble radius approaches the critical value, the arc of simulated quasi-steady solutions in the diameter-velocity diagram exhibits a vertical tangent, i.e.\ infinite slope, and continues backward with increasing velocity until another turning point with infinite slope is reached and the arc of solutions continues further, with still increasing velocity but now also with increasing volume. This $S$-shaped arc of solutions is typical for dynamical systems which admit two stable as well as an unstable solution for a certain parameter range. Here, this parameter is the bubble diameter. Therefore, for increasing bubble diameter, the lower stable arc comes to an abrupt end at a certain critical value. For a slightly larger bubble, the only stable solution displays a significantly larger rise velocity.
The authors of \cite{Fraggedakis2016} reported ``numerical difficulties when trying to continue on the upper branch of the solution family to higher velocities''. As possible reasons, the experimentally observed deviation from axisymmetry, also reported for the 3D numerical simulations in \cite{Pillapakkam2007},
and the increasing Weissenberg number which induces numerical problems, are mentioned.
In \cite{Niethammer2019}, an extended Volume-of-Fluid (VOF) method with a novel scheme for the second-order accurate finite-volume discretization of the interface stress term is employed to perform direct numerical simulations of the transient motion of a single bubble rising in a quiescent viscoelastic liquid. The rise velocity jump discontinuity and the negative wake are captured, with a quantitative match to the experimental data from \cite{Pilz2007}.
A similar VOF method has been used in \cite{Yuan2021} to simulate the same setup, i.e.\ single rising bubbles corresponding to the experiments from \cite{Pilz2007}, as well as different rise regimes concerning the Galilei number, including pulsating bubbles. The authors of \cite{Yuan2020} carried out three-dimensional numerical simulations with that same setup to study the stability of the bubbles during their rise in viscoelastic liquids characterized by the Weissenberg number and a viscosity ratio of the solvent to the solution. The influence of the fluid viscoelastic properties on the rise velocity jump discontinuity was pointed out. At high elasticity, for certain sets of the E\"otv\"os and Galilei numbers, vortex shedding in the wake region was found to yield a pulsating rise velocity. It was found that polymer stretching at the downstream end of the bubble may force the bubble to break up.
The investigation of the present paper identifies the molecular mechanism behind the rise velocity jump discontinuity for single bubbles rising in a viscoelastic liquid.
The present work combines three different approaches: experimental, numerical and theoretical. Put together, this--for the first time--gives a complete and clear picture of the underlying \emph{self-amplifying} mechanism leading not only to higher rise velocities, but explaining the abruptness of the change once the bubble volume exceeds a critical value.
Moreover, it sheds new light on the relation between the characteristic features mentioned above, i.e.\ the velocity jump, the negative wake and the cusped shape.
It also shows the intimate relationship of this phenomenon to the 3D flow kinematics.
The paper is organized as follows:
Section~\ref{sec:experimentalmethod} presents the experimental methods used for the investigations of bubbles rising in viscoelastic liquids, including the flow field in the liquid.
Section~\ref{sec:numericalmethod} details the numerical methods used for the simulations. In Section~\ref{sec:experimentalresults} we present the results from the experiments of this study, in particular the bubble rise velocities and the velocity and vorticity fields in the liquid. Section~\ref{sec:numericalresults} presents and validates the results from the numerical simulations in comparison with the experiments. The results from a theoretical analysis of the transport and conformation of polymer molecules in the flow around a bubble are presented and discussed in Section~\ref{sec:theory-results} in the context of the molecular processes behind the formation of the bubble rise velocity jump discontinuity. The findings are discussed and the conclusions are drawn in Section~\ref{sec:discussconclusions}.
\section{Experimental methods}
\label{sec:experimentalmethod}
The experiments for measuring the terminal rise velocity of single bubbles in viscoelastic liquids
were carried out in an apparatus as shown in Fig.~\ref{fig:setup-PIV}. The apparatus and the experimental techniques were described in detail in \cite{Pilz2007}. Single air bubbles with controllable and well-defined volume in the range between $2$ and $1000~\textnormal{mm}^3$ were produced in the bubble-generating chamber by collecting individual bubbles detached from the end of a capillary. The resulting single bubbles were then allowed to rise in the liquid to a position half a meter above the location of their production to exclude any influence from the initial deformed shape of the bubble on the rise velocity. The viscoelastic liquids were solutions of poly(acrylamide)s and poly(ethyleneoxide) in water, in a $70\!:\!30$~wt.~\% mixture of glycerol with water, and in ethylene glycol. We limit the present study to the flexible polymers, excluding the hydrolyzed poly(acrylamide)s \cite{Pilz2009}.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.7]{figure/FIGPIVsketch.png
\caption{\small Experimental setup for investigating the flow field around bubbles in
viscoelastic liquids.}
\label{fig:setup-PIV}
\end{figure}
\subsection{Viscoelastic liquid characterization}
\label{sec:relax}
An important aspect in the dynamics of bubble motion in viscoelastic polymer solutions is the relaxation behavior of the polymer molecules in the solution, which is briefly reviewed in the following, since we will use characteristic time scales of the relaxation behavior in our explanation of the bubble rise velocity jump discontinuity phenomenon.
Larson points out in his textbook \cite{Larson1999} that the stress in any polymeric system, dilute or concentrated, depends on the conformation of the polymer molecules. The contribution to the stress increases as the orientation and degree of stretch is increased. The total stress in a fluid element results from the individual conformations of all molecules in that element. A description of the stress arising from molecular theories must therefore include the distribution of conformations that results from deformation \cite{Larson1999}.
In order to describe the molecular process related to diffusion or relaxation, which changes the conformation for a flexible polymer, the polymer coil is usually modeled as a series of beads connected by springs \cite{Larson1999}. Although such a bead-and-spring chain does not represent the chemical structure of the polymer (Bird et al.\ \cite{Bird1987a,Bird1987b}), it is able to mirror the forces acting on the molecule, namely viscous and elastic forces \cite{Larson1999}, including a high number of degrees of freedom, as one expects from a flexible polymer \cite{Bird1987a,Bird1987b}. The relaxation of the entire molecule, i.e., the slowest relaxation, can be modeled by a single elastic dumbbell, which consists of only two spheres connected by a linear spring.
Yarin \cite{Yarin1993} constructed a system of constitutive equations for topological restrictions based on micromechanical polymer models by studying the mechanism of stress relaxation after a step strain according to Doi and Edwards \cite{DoiEdwards1986} and Doi \cite{Doi1980}.
The analysis shows that the polymer relaxation time may be represented as
\begin{equation}
\lambda \approx \frac{N_0^2 b^2 \zeta}{3 \pi^2 k_B T}.
\label{eq:timeconstant-sim}
\end{equation}
Herein, $N_0$ represents the number of Rouse segments in the entire chain, $b^2$ stands for the mean squared length of the Rouse segment at equilibrium, $\zeta$ is the friction factor (including solvent friction and interaction with the wall of a reptation tube \cite{Yarin1993}), $k_B$ the Boltzmann constant, and $T$ the temperature. Equation (\ref{eq:timeconstant-sim}) mirrors the statement given in Yarin (\cite{Yarin1993}, p.\ 363) that, under strong deformation (i.e., inside the liquid filament generated with a filament-stretching elongational rheometer, as used in our characterization of the polymer solutions), the relaxation processes which are faster than the reptational diffusion appear to be very important.
The paper \cite{Pilz2007} was -- to the best of our knowledge -- the first to account for liquid properties obtained from elongational rheometry in investigations of the critical bubble size at the rise velocity jump discontinuity. Our hypothesis is that the elongational behavior of the polymer solution plays a significant role in the rise of bubbles, since the liquid flow field around the bubbles exhibits considerable elongational components. Our rheometric characterization of the polymer solutions included both the shear viscosity at the first Newtonian plateau and the relaxation time obtained from filament stretching experiments carried out by means of the elongational rheometer described in Stelter et al.\ \cite{Stelter2000}. The device enables one to measure the diameter decrease of a self-thinning liquid filament with time after inducing a step strain within a liquid bridge between two plates.
In Stelter et al.\ \cite{Stelter2000} it was shown that the self-thinning of filaments of semi-dilute solutions of flexible polymers can be subdivided into two regimes. In the first regime, the filament diameter decreases exponentially with time, and in the second regime it decreases linearly with time.
In the first regime, the polymer solution exhibits viscoelastic behavior. The filament diameter decreases according to the law
\begin{equation}
d_f (t) = d_{f0} ~\exp \left( - t / 3 \lambda_E \right) ~,
\label{eq:filthin}
\end{equation}
where $d_f$ is the filament diameter, $d_{f0}$ is its initial value (at time $t = 0$), and $\lambda_E$ is the relaxation time, which is equal to the one defined in Eq.\ \eqref{eq:timeconstant-sim}, i.e.\ $\lambda_E \equiv \lambda $.
In the second regime, a maximum of the polymer extension is reached, and the polymer solution exhibits Newtonian-like behavior, which allows the steady terminal elongational viscosity $\eta_{E}^\infty$ to be determined from the relation
\begin{equation}
d_f (t) = d_{f,0} - \frac{\sigma}{\eta_{E}^\infty } t~,
\label{eq:termvisc}
\end{equation}
where $\sigma$ denotes the surface tension.
It should be pointed out that, in the elasticity-dominated regime, the rate of stretching within the filament,
\begin{equation}
\dot \epsilon = \frac{\partial u_z}{\partial z} = - \frac{2}{d_f} \frac{{\rm d}\, d_f}{{\rm d}\, t}
\label{eq:thinning}
\end{equation}
is constant during the self-thinning process. We show this by combining Eqs.\ \eqref{eq:filthin} and \eqref{eq:thinning} to obtain
\begin{equation}
\dot \epsilon = \frac{2}{3 \lambda_E}.
\label{eq:thinconst}
\end{equation}
The relaxation time $\lambda_E$ obtained from the viscoelastic regime of the filament self-thinning can be related to the longest time constant of the Zimm spectrum obtained from shear rheometric measurements \cite{Anna2001}.
It is therefore emphasized that, regardless of the mechanical properties of the polymer, the polymer concentration or the solvent quality, the dependency of the filament diameter on time (i.e., Eq.\ \eqref{eq:filthin}) does not change, and the relaxation time $\lambda_E$ calculated from the measured evolution of $d_f$ does not depend on any particular assumptions on the macromolecular model and polymer/polymer interactions \cite{Stelter2002}. This parameter $\lambda_E$ plays an essential role in the prediction of the critical bubble size at the rise velocity jump discontinuity \cite{Pilz2009}.
\subsection{PIV for flow field measurements}
\label{sec:PIV}
For investigating the flow field around the bubble, Particle-Image Velocimetry (PIV) was used as the measuring technique. The PIV system employed was the Dantec FlowMap 1500 system (Dantec Dynamics, Skovlunde, Denmark), together with a Dantec 80C60 HiSense camera ($1280 \times 1024$ pixel) and a NewWave Gemini double-cavity Nd:YAG laser ($120~\textnormal{mJ}/\textnormal{pulse}$).
A special light-sheet optics consisting of a spherical lens ($600~\textnormal{mm}$ focal length), a cylindrical lens ($-10~\textnormal{mm}$ focal length), and a prism was used to illuminate the seeding particles in plane sections of the flow. The light sheet with a thickness of approximately $5~\textnormal{mm}$ was placed in a symmetry plane of the test column, which contains the trajectory of the bubble center.
The system was calibrated using a Dantec calibration target, a $100 \times 100~\textnormal{mm}$ white plate with a square grid of black dots of spacing $2.5~\textnormal{mm}$ and size $1.5~\textnormal{mm}$.
The essential difficulty when measuring with PIV in a liquid field with bubbles is the strong reflection of the irradiated laser light from the surfaces of the bubbles, as shown in Fig.~\ref{fig:PIV-reflexion-bubblesurface}.
\begin{figure}[bt]
\centering
\includegraphics[scale=0.5]{figure/figPIVReflexion.png
\caption{\small Scattered light from the bubble surface.}
\label{fig:PIV-reflexion-bubblesurface}
\end{figure}
The bright zone on the images hinders an accurate velocity measurement in the vicinity of a part of the bubble surface, which must be avoided.
A remedy is the use of fluorescent particles as tracers in the liquid phase, and of an optical filter, which lets only the fluorescent light pass and removes the Mie-scattered laser light reflected from the bubble surface.
Fluorescent particles were produced for the present investigations by saturating polyamide particles ($20~\mathrm{\upmu m}$ diameter, PSP20, Dantec Dynamics, Skovlunde; Denmark) with the dye Eosin
$(\mathrm{C}_{20}\mathrm{H}_{6}\mathrm{Br}_{4}\mathrm{Na}_{2}\mathrm{O}_{5})$
and drying them such that the dye became insoluble in the polymer solutions. For this purpose, the polyamide particles were soaked in an alcoholic Eosin solution for $24$ hours
and then washed with distilled water, filtered, and dried in air.
Washing and filtering were repeated seven times.
A quantity of $0.4~\textnormal{g}$ particles was added to $40~\textnormal{ml}$ of the polymer solution and dispersed using an ultrasonic bath.
This concentrated solution was then stirred into $7.5$ liters of the aqueous polymer solution.
An additional mirror opposite from the light sheet probe helped to reflect light from the light sheet into the shadow from the bubbles.
As a result, pictures were obtained with clearly visible particle images and sufficiently visible bubble contours to determine the bubble volume by the image processing method described in section~\ref{sec:imageprocessing-PIV} below; cf.~\cite{Pilz2009}.
The double images from one PIV recording were evaluated using a cross-correlation
technique and an interrogation area size of $32 \times 32$ pixel with 50~\% overlap.
The time between the two pulses was 1--100 ms, depending on flow situation and setup.
Finally, a range validation, a correlation peak ratio validation,
and a moving average validation were applied to reject invalid vectors
(Dantec Flow Manager software, v.\ 4.60.28).
Maps of $79 \times 63$ vectors were obtained.
The results from these investigations are presented below in section~\ref{sec:PIV-velocityfields}.
\subsection{Image Processing}
\label{sec:imageprocessing-PIV}
The shape of the bubble and its position relative to the picture frame were extracted from each picture using image processing tools provided by Adobe Photoshop and Matlab.
In order to preserve the contour of the bubble, images of the seeding particles had to be
removed from the frame. For this purpose, filter options for contrast enhancement and masking tools provided by Adobe Photoshop were applied first.
Furthermore, the centroid $C$ of the meridional area was determined, and the bubble contour points were filtered out. The coordinates of the bubble contour points were transformed into a local coordinate system $(x_{1}, x_{2}, x_{3})$ with right-handed orientation and the $x_{1}$ axis pointing towards the camera.
Furthermore, the origin of the coordinate system coincides with the centroid $C$ of the bubble's meridional area.
The two in-plane axes $(x_{2}, x_{3})$ represent the principal axes of the bubble's
meridional area computed via image processing tools with Matlab.
Since, for the volume range of interest, axial symmetry around the direction of motion of the bubbles can be assumed
and, furthermore, the bubbles rise along a straight vertical path,
the axis of symmetry (denoted $x_{3}$ in Fig.~\ref{fig:image-processing-PIV})
points in the direction of motion.
\begin{figure}[ht!]
\centering
\includegraphics[scale=1]{figure/FIGBubbleBC_windws_full-2.pdf
\caption{\small Extraction of the boundary curve.}
\label{fig:image-processing-PIV}
\end{figure}
The volume of the air bubble and the arc length of the contour
are obtained from an approximation of the meridional curve $M$,
which represents one half of the previously extracted boundary curve by means of piecewise
linear functions through two adjacent points $(z_{i}, \varpi_{i})$ and $(z_{i + 1}, \varpi_{i + 1})$
of the boundary curve.
The separation of the boundary curve into left-hand and right-hand sides was carried out
after calculation of the corresponding polar angles $\vartheta_{i}$ for each
data point of the curve.
The data points were sorted in ascending order, based on the angles $\vartheta_{i}$,
where negative values of $\vartheta_{i}$ refer to the left-hand side (LHS) of the bubble contour,
and positive values refer to the right-hand side (RHS) (see Fig. \ref{fig:image-processing-PIV}).
A partial volume $\mathrm{\Delta} V_{i}$ of the air bubble is then computed
form data of one half of the boundary curve according to
\begin{equation}
\label{eq:partial-volume}
\mathrm{\Delta} V_{i} =
\frac{\pi}{3} \left( \varpi_{i}^{2} + \varpi_{i} \varpi_{i + 1}
+ \varpi_{i + 1}^{2} \right)(z_{i + 1} - z_{i}) \; \text{.}
\end{equation}
These partial bubble volumes were then added up to obtain the total volume.
Similarly, the arc lengths $l_{i}$ corresponding to the discrete points $(z_{i}, \varpi_{i})$
of the meridian curve (i.e.\ one half of the boundary curve)
were calculated in accordance with
the approximation of the boundary curve by
\begin{equation}
\label{eq:arc-length-PIV}
l_{i} = l_{i - 1} + \sqrt{\left(z_{i} - z_{i - 1} \right)^{2}%
+ \left(\varpi_{i} - \varpi_{i - 1} \right)^{2}}~.
\end{equation}
The bubble rise velocity was obtained from the relative displacement of the bubble
contour curve within two consecutive images and checked by comparison with the maximal velocity measured at the boundary curve by PIV. For depicting liquid-phase velocity fields, the bubble rise velocity was subtracted from the instantaneous velocities measured by PIV.
Before doing this, the PIV data had to be transformed
into the local coordinate system $(x_{1}, x_{2}, x_{3})$,
since the centroid of the bubble meridian plane did not necessarily coincide
with the reference mark (i.e.\ the ``zero marker'') of the calibration target \cite{Pilz2009}.
\section{Numerical method}
\label{sec:numericalmethod}
Direct numerical simulations (DNS) of bubbles rising in a viscoelastic fluid are carried out using an extended volume of fluid (VOF) method, introduced in Niethammer et al. \cite{Niethammer2019, Niethammer2019b}. Numerical experiments in \cite{Niethammer2019, Niethammer2019b, Niethammer2019c} demonstrate that the extended VOF method captures characteristic flow phenomena of single bubbles rising in a quiescent viscoelastic fluid, such as the rise velocity jump discontinuity and the negative wake. It is noteworthy that good quantitative agreement of the steady-state rise velocities was achieved between experimental measurements \cite{Pilz2007} and three-dimensional DNS \cite{Niethammer2019}. Besides the velocity jump discontinuity and the negative wake, the present study provides a detailed insight into the local stress and conformation tensor distributions in the neighborhood of the rising bubble. Moreover, the spectrum of the conformation tensor, especially the leading eigenvalues and corresponding eigenvectors, is investigated to analyze the orientation of the polymer molecules; cf.\ Appendix~A about the meaning of the eigenvalues of the conformation tensor for the orientation of the polymer molecules.
The extended VOF method employs a one-field formulation based on the volume-averaged two-phase flow equations. The reader is referred to \cite{Niethammer2019} for the derivation of the volume-averaged local instantaneous bulk equations and interface jump conditions. In section
\ref{sec:vofmodel}, we provide a summary of the extended VOF model.
\subsection{Governing equations}
\label{subsec:governing-equations}
Assume the two-phase fluid system to fill the domain ${\CCV \subset \mathbb{R}^3}$ under consideration, such that the two bulk phases $\CCV^{\phaseOne}(t)$ and $\CCV^{\phaseTwo}(t)$ are separated by the two-dimensional surface ${\Sigma(t) \in \mathbb{R}^3}$, i.e.\ $\CCV = {\CCV^{\phaseOne}(t) \cup \CCV^{\phaseTwo}(t) \cup \Sigma(t)}$. Let us further denote the interface unit normal pointing towards the phase $\CCV^{\phaseOne}$ as $\vec{n}_{\Sigma}$.
Assuming isothermal flow with constant mass densities in the bulk phases, the balance equations for mass and momentum in their local formulation in the bulk read
\begin{align}
\label{contieq00}
\nabla \dprod \velocity &= 0 \quad \textnormal{in} \ \CCV \setminus \Sigma, \\
\label{monentumeq00}
\rho {\partial_t \velocity} + \rho \left(\velocity \dprod \nabla\right) \velocity &= \nabla \dprod \ST + \rho \vec{b} \quad \textnormal{in} \ \CCV \setminus \Sigma,
\end{align}
where $\velocity$ is the velocity, $\rho$ is the density and the term $\rho \vec{b}$ represents body forces with the mass specific density $\vec{b}$. The term $\nabla \dprod \ST$ represents contributions from contact forces, where the second rank tensor $\ST$ is the Cauchy stress.
At the interface $\Sigma$, additional jump conditions are needed in order to obtain a complete system. We define the jump $\jump{ \PhiT }$ of a quantity $\PhiT$ at $\Sigma$ as
\begin{align}
\label{defjump}
\jump{ \PhiT } (\vec{x}) = \lim\limits_{h \rightarrow 0^{+}} \left(\PhiT (\vec{x} + h\vec{n}_{\Sigma}) - \PhiT (\vec{x} - h\vec{n}_{\Sigma})\right), \quad \vec{x} \in \Sigma.
\end{align}
Then, assuming no-slip at $\Sigma$,
the interfacial jump conditions are
\begin{align}
\label{jcontieqVV00}
\jump{ \velocity } &= 0 \quad \textnormal{on} \ \Sigma, \\
\label{jmomentumeqVV00}
{\jump{- \ST}} \dprod \vec{n}_{\Sigma} &= \sigma \kappa \vec{n}_{\Sigma}
\quad \textnormal{on} \ \Sigma, \\
\label{snd00}
\velocity \dprod \vec{n}_\Sigma &= V_\Sigma \quad \textnormal{on} \ \Sigma,
\end{align}
where $\sigma$ denotes the surface tension and the local curvature (more precisely, twice the mean curvature) is defined as ${\kappa = \nabla_{\Sigma} \dprod (-\vec{n}_{\Sigma})}$ with the surface gradient operator $\nabla_\Sigma$.
The kinematic condition (\ref{snd00}) for the interface normal velocity is satisfied in absence of phase change, with $V_\Sigma$ denoting the speed of normal displacement of $\Sigma$.
Note that in the momentum transmission equation (\ref{jmomentumeqVV00}), we assume a constant surface tension $\sigma$ and neglect the effect of any intrinsic interfacial rheology, i.e.\ assume zero surface viscosities. This is a good assumption as long as local microstructures along the interface, e.g.\ interfacial polymer monolayers, have a minor impact compared to the bulk-fluid stresses acting across the interface.
We split the stress $\ST$ into an isotropic part, a viscous solvent component and an elastic polymer contribution, i.e.\
\begin{align}
\label{stressdef00}
\ST = -p \, \IT + \stressS + \stressP,
\end{align}
where $p$ is the pressure. The viscous stress obeys Newton's law, i.e.\
\begin{align}
\label{stressdef01}
\stressS = \svisc \big( \grad{\velocity} + \trans{\grad{\velocity}} \big),
\end{align}
where $\svisc$ is the solvent viscosity.
The polymer stress is directly related to a conformation tensor $\CT$, representing the average configuration of the polymer molecules in the macroscopic fluid element \cite{Bird1987a, Larson1988}. In particular, the conformation tensor is related to the second moment of the distribution function of the molecular configuration \cite{Larson1988}. By its definition, this implies that $\CT$ is a symmetric and positive definite (spd, for short) tensor of second rank; cf.\ also Appendix~\ref{asec:conftens}.
Throughout this paper, we exploit the fact that every spd tensor is diagonalizable with real, positive eigenvalues.
The polymer stress $\stressP$ is a linear function of the polymer conformation tensor $\CT$, viz.\
\begin{align}
\label{stressdef02}
\stressP = \frac{\pvisc}{\lambda (1-\zeta)} \left( h_{0} \IT + h_{1} \CT\right),
\end{align}
where $\pvisc$ is the polymer viscosity, $\lambda$ is the relaxation time and $\zeta$ is a material parameter (sometimes called ``slip parameter"). Note that (\ref{stressdef02}) can be derived from a closed form of the Kramers' expression for the polymer stress tensor \cite{Bird1987b, Larson1988}. The coefficients $h_{j}$ depend on the molecular model and the closure approximation. For most rheological models, the scalar coefficients $h_{j}$ are either functions of the first invariant of $\CT$, i.e.\ $h_{j} = \hat{h}_{j}(\tr{\CT})$, or constants (cf.\ Table \ref{tab:constitutiveEx}).
To obtain a closed system, additional constitutive equations for $\CT$ are required. In this work, we restrict our considerations to partial differential conformation tensor constitutive equations of the form
\begin{align}
\label{constCEq00}
\partial_t \CT + \left( \velocity \dprod \nabla \right) \CT - \LT \dprod \CT - \CT \dprod \trans{\LT}
=
\frac{1}{\lambda} \cvec{P}\left(\CT \right) \quad \textnormal{in} \ \CCV \setminus \Sigma
\end{align}
with $\LT = \trans{\grad \velocity} - \frac{\zeta}{2} \left({\grad \velocity} + \trans{\grad \velocity}\right)$ and the material parameter ${\zeta} \in [0, 2]$ characterizing the degree of non-affine response of the polymer chains to an imposed deformation. For ${\zeta} = 0$, the motion becomes affine, and in this case the left-hand side of (\ref{constCEq00}) reduces to the upper-convected derivative, first proposed by Oldroyd~\cite{Oldroyd1950}. The relaxation is introduced by a specification of the tensor function $\cvec{P} \left(\CT \right)$. We impose the restriction that $\cvec{P} \left(\CT \right)$ is a real analytic tensor function of $\CT$, such that
\begin{equation}
\label{defisotropicf}
\OT \dprod \cvec{P} \left(\CT \right) \dprod \trans{\OT} = \cvec{P} \big( \OT \dprod \CT \dprod \trans{\OT} \big)
\end{equation}
for every orthogonal tensor $\OT$. In this case, $\cvec{P} \left(\CT \right)$ is referred to as an isotropic tensor-valued function of the second-rank spd tensor $\CT$. It is well-known \cite{RivlinEricksen1955, Rivlin1955} that every real analytic and isotropic tensor-valued function $\cvec{P} \left(\CT \right)$ of a second-rank tensor $\CT$ has a representation of the form
\begin{equation}
\label{Pfunc}
\cvec{P}\left(\CT \right) = g_{0} \IT + g_{1} \CT + g_{2} {\CT}^2,
\end{equation}
where $g_{0}$, $g_{1}$ and $g_{2}$ are isotropic invariants (isotropic scalar functions) of $\CT$ and hence can be expressed as functions of its three scalar principal invariants, i.e.\ $g_{i} = \hat{g}_{i}(I_{1}, I_{2}, I_{3}), \ i = 0,1,2$. The principal invariants $I_{1}, I_{2}, I_{3}$ are
\begin{equation}
I_{1} = \textnormal{tr} \: \CT, \ I_{2} = \frac{1}{2}
\left[ (\textnormal{tr} \: \CT)^2 - \tr{{\CT}^2} \right]\!, \ I_{3} = \det \CT.
\end{equation}
Here, we only use models for which $g_{2} = 0$, i.e.\ the quadratic term in (\ref{Pfunc}) vanishes. Table \ref{tab:constitutiveEx} summarizes the model-specific expressions for the constitutive equations used in the present work.
\begin{table}[h!]
\small
\caption{\small Model-dependent scalar-valued tensor functions for the generic constitutive equation (\ref{constCEq00}) and (\ref{Pfunc}).}
\label{tab:constitutiveEx}
\begin{center}
\begin{tabular}{@{}lcccccc@{}}
\toprule
constitutive model & \ ${\zeta}$ \ & \ $g_{0}$ \ & \ $g_{1}$ \ & \ $g_{2}$ \ & \ $h_{0}$ \ & \ $h_{1}$ \ \\
\midrule
Maxwell/Oldroyd-B & $0$ & $1$ & $-1$ & $0$ & $-1$ & $1$ \\
LPTT & $\in \left[0, 2 \right]$ & $1 + \frac{\varepsilon}{1 - {\zeta}} \left( \tr \CT - 3 \right) $ & $ - g_{0} $ & $0$ & $-1$ & $1$ \\
EPTT & $\in \left[0, 2 \right]$ & $ \exp \left[ \frac{\varepsilon}{1 - {\zeta}} \left( \tr \CT - 3 \right) \right] $ & $ - g_{0} $ & $ 0 $ & $-1$ & $1$ \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\subsection{Stabilization}
\label{subsec:stab}
Because of the high Weissenberg number problem (HWNP)~\cite{Joseph1985, Keunings1986}, a numerical stabilization is required to increase the robustness of the numerical method at moderate and high degrees of fluid elasticity. Fattal and Kupferman~\cite{Fattal2004} showed that a logarithmic change of variables in the conformation tensor equation circumvents the HWNP. Balci et al.~\cite{Balci2011} proposed a square root conformation tensor representation that does not require any diagonalization of the conformation tensor. Afonso et al.~\cite{Afonso2011} generalized the conformation representations with generic kernel transformation functions. In the present work, we use the unified mathematical and numerical stabilization framework, proposed by Niethammer et al.~\cite{Niethammer2018} for the general conformation tensor constitutive equations (\ref{constCEq00}) and (\ref{Pfunc}). This generic framework has been validated~\cite{Niethammer2018, Niethammer2019c} and extended to non-isothermal~\cite{Meburger2020} and two-phase flows~\cite{Niethammer2019}. For a detailed description of the generic stabilization framework and its validation in computational benchmarks, we refer to~\cite{Niethammer2018, Niethammer2019b}. Below, we give a brief summary of the stabilization framework used in the present work.
The generic constitutive equation is obtained by introducing the real analytic tensor function $\cvec{F}(\CT)$ of the spd second-rank tensor $\CT$, such that
\begin{equation}
\label{diagonalizef0}
\OT \dprod \cvec{F}(\CT) \dprod \trans{\OT}
=
\cvec{F}(\OT \dprod \CT \dprod \trans{\OT})
\end{equation}
for every orthogonal tensor ${\OT}$, i.e.\ $\cvec{F}(\CT)$ is an \textit{isotropic} tensor-valued function of $\CT$. If $\CT$ is governed by the generic constitutive equation (\ref{constCEq00}), then $\cvec{F}(\CT)$ satisfies
\begin{equation}
\label{evolutionConf3}
\DDt {\cvec{F}(\CT)}
=
{2 \BT \dprod {\UpsilonT} \dprod \CT}
+
\OmegaT \dprod {\cvec{F}(\CT)} - {\cvec{F}(\CT)} \dprod \OmegaT
+
\frac{1}{\lambda} \UpsilonT \dprod \cvec{P} \left(\CT \right),
\end{equation}
where we exploit the representation $\CT = \QT \dprod \DT \dprod \trans{\QT}$ with the diagonal tensor $\DT$ containing the three real eigenvalues $d_1, d_2, d_3$ and the orthogonal tensor $\QT$ containing the corresponding set of eigenvectors. Then, $\cvec{F}(\CT) = \cvec{F}(\QT \dprod \DT \dprod \trans{\QT}) = \QT \dprod \cvec{F}(\DT) \dprod \trans{\QT}$. The deformation terms in the convective derivative are decomposed into the first three terms on the right-hand side of (\ref{evolutionConf3}), containing the tensors $\BT$ and $\OmegaT$, which are functions of the velocity gradient. This local decomposition of the velocity gradient was first proposed in~\cite{Fattal2004} and is in general necessary to transform the conformation tensor equation (\ref{constCEq00}) into an evolution equation for the analytic tensor function $\cvec{F}(\CT)$. The functional relationship of $\BT$ and $\OmegaT$ with $\LT$ is given in Appendix \ref{asec:lddt}.
The generic constitutive equation (\ref{evolutionConf3}) falls back to the classical logarithm or root conformation representations (LCR or RCR), depending on the choice of $\cvec{F}(\CT)$ and $\UpsilonT$.
Table~\ref{tab:JJ}
\begin{table}[htbp]
\caption{\small Function $\cvec{F}(\CT)$ and $\UpsilonT$ terms in (\ref{evolutionConf3}) for the conformation tensor representation, the $k$-th root of $\CT$ representation, and the logarithm of $\CT$ to base $a$ representation.}
\renewcommand{\arraystretch}{1.35}
\begin{center}
\linespread{1.15}
\begin{tabular}{rlll}
\toprule
\ & $\cvec{F}(\CT)$ & \ $\UpsilonT$ \ \\
\midrule
conformation tensor \ & $\CT$ & \ $\IT$ \ \\
$k$-th root of $\CT$ \ & $\RT = \QT \dprod \DT^{\frac{1}{k}} \dprod \trans{\QT}$ & \ $\frac{1}{k} \RT^{1-k}$ \ \\
logarithm of $\CT$ to base $a$ \ & $\ST = \QT \dprod \loga{(\DT)} \dprod \trans{\QT}$ & \ $\frac{1}{\operatorname{ln}(a)} a^{-\ST}$ \ \\
\bottomrule
\end{tabular}
\end{center}
\renewcommand{\arraystretch}{1.0}
\label{tab:JJ}
\end{table}
shows the corresponding terms for specific conformation representations. In the present work, we choose the RCR, given by
\begin{align}
\label{evolutionConfRTb}
\DDt \RT
&=
\frac{2}{k} \BT \dprod \RT
+
\OmegaT \dprod {\RT} - {\RT} \dprod \OmegaT
+
\frac{1}{k \lambda} \left(g_0 \RT^{1-k} + g_1 \RT + g_2 \RT^{1+k}\right).
\end{align}
In preliminary studies we tried root stabilization with $k=2,3, 4$, from which $k=4$ turned out to be the optimal choice concerning stability and convergence.
We also tried the logarithmic transformation
but did not observe significant differences.
Therefore, all results shown below have been obtained using the $4^{\rm th}$ order root.
\subsection{The Volume of Fluid (VOF) model}
\label{sec:vofmodel}
The VOF method \cite{Hirt1981} is an interface capturing method, widely used for sharp interface direct numerical simulations (DNS). Due to its inherent mass conservation and robustness in handling strong morphological changes, the VOF method is an appropriate choice for the simulation of rising bubbles in a viscoelastic fluid. It was first shown in Niethammer et al.~\cite{Niethammer2019} that the VOF method is capable to capture the bubble rise velocity jump discontinuity. Recently, the VOF results for the velocity jump have been confirmed by Yuan et al.~\cite{Yuan2021}.
In the VOF method, the fluid interface is captured by a phase indicator function. We employ the approach of an algebraic VOF, where the propagation of the fluid interface is achieved implicitly by solving a scalar transport equation for the volume fraction field. The volumetric phase fraction $\alpha$ is defined over the entire domain as the volume average of the phase indicator function over a mesh cell and is constrained to satisfy $0 \leq \alpha \leq 1$. We let $\alpha = 1$ in the viscoelastic phase, hence $\alpha = 0$ in the gas phase of the bubble, and the interface is located in mesh cells, where the volume fraction has values $0 < \alpha < 1$.
The extended VOF equations for viscoelastic fluids are derived from the local instantaneous volume equations and the interfacial jump conditions from section~\ref{subsec:governing-equations} by successively applying a volume averaging procedure and the homogeneous mixture model as a closure. The definitions of volume-averaged quantities are found in the appendix \ref{asec:vadefinitions} and for a detailed derivation of the extended VOF equations we refer to \cite{Niethammer2019}. As a result, we obtain a single set of equations, valid for both phases throughout the computational domain, which is shown below. Introducing the definition for the mixture density
\begin{align}
\label{mixdensity00}
\rho_m = \alpha {\rho}^{\phaseOne} + \left(1 - \alpha \right) {\rho}^{\phaseTwo},
\end{align}
the mixture velocity
\begin{align}
\label{mixvelocity00}
{\velocity}_m = \frac{\alpha \overline{\velocity}^{\phaseOne} + \left(1 - \alpha \right) \overline{\velocity}^{\phaseTwo}}{\rho_m}
\end{align}
and the mixture stress
\begin{align}
\label{mixstress00}
{\ST}_m = \alpha \overline{\ST}^{\phaseOne} + \left(1 - \alpha \right) \overline{\ST}^{\phaseTwo},
\end{align}
the mass and momentum equations can be written as
\begin{align}
\label{mixContieq00}
\nabla \dprod {\velocity}_m &= 0, \\
\label{mixUeq00}
\partial_t {\rho}_m {\velocity}_m
+ ({\velocity}_m \dprod \nabla) {\rho}_m {\velocity}_m
&= {\nabla \dprod {\ST}_m}
+ {\rho}_m \vec{b}
+ \vec{f}_\sigma,
\end{align}
where $\vec{f}_\sigma$ represents the surface tension force.
We employ the widely used continuous surface force (CSF) model proposed by Brackbill et al.~\cite{Brackbill1992} as a closure for the surface tension force.
The scalar transport equation for the volume fraction field reads
\begin{align}
\label{vofraceq02b}
\partial_t \alpha + \left({\velocity}_m \dprod \nabla\right) \alpha = 0. \end{align}
Note that the discretization of the advection term in (\ref{vofraceq02b}) requires special numerical schemes that simultaneously keep the interface sharp and produce monotonic profiles of the volume fraction between its bounds $0 \leq \alpha \leq 1$. In the present work, we use the multi-dimensional flux corrected transport (FCT) method, introduced by Boris and Book \cite{Boris1973} and extended to multiple dimensions by Zalesak \cite{Zalesak1979}. The FCT method is a multi-dimensional flux limiting strategy that applies a local weighting procedure between an unconditionally bounded low-order upwind differencing scheme and an anti-diffusive higher-order correction scheme, such that the higher-order flux is used to the greatest extent possible, while keeping the solution bounded between some local extrema. For the higher-order corrections, we use the inter-Gamma scheme \cite{Jasak1995} which introduces a certain amount of downwind differencing to maintain the sharpness of the interface.
The mixture stress tensor ${\ST}_m$ in the momentum balance (\ref{mixUeq00}) is split according to
\begin{align}
\label{mixstress01}
{\ST}_m = -\grad p + {\nabla \dprod \stressSm} + {\nabla \dprod \stressPm},
\end{align}
where the mixture solvent stress is defined as
\begin{align}
\label{closestress04a}
{\stressSm} &= {\smvisc} \big(\grad{{\velocity}_m} + \trans{\grad{{\velocity}_m}}\big)
\end{align}
with mixture solvent viscosity
\begin{align}
\label{mixviscosity00}
{\smvisc} = \alpha {\eta}_{s}^{\phaseOne} + \left(1 - \alpha \right) {\eta}_{s}^{\phaseTwo}.
\end{align}
Using the reformulation
\begin{align}
\label{closestress04b}
{\nabla \dprod \stressSm} = \nabla \dprod \left({\smvisc} \grad{{\velocity}_m}\right) + \grad{{\velocity}_m} \dprod \nabla {\smvisc}
\end{align}
for the solvent stress divergence in (\ref{mixstress01}) is beneficial due to the elliptic (Laplacian-type) term, which is discretized implicitly in this work.
For the closure of the polymer stress term in (\ref{mixstress01}), we introduce the definitions
\begin{align}
\label{mixpstress00}
{\stressPm} &= \alpha \overline{\stressE}_{p}^{\phaseOne} + \left(1 - \alpha \right) \overline{\stressE}_{p}^{\phaseTwo},\\
\label{mixG00}
G_m &= \alpha G^{\phaseOne} + \left(1 - \alpha \right) G^{\phaseTwo}, \\
\label{mixC00}
{\CT}_m &= \frac{\alpha G^{\phaseOne} \overline{\CT}^{\phaseOne} + \left(1 - \alpha \right) G^{\phaseTwo} \overline{\CT}^{\phaseTwo}}{G_m},
\end{align}
where $G^{\pm} = \frac{\eta_{p}^{\pm}}{\lambda^{\pm} \left(1 - \zeta^{\pm} \right)}$.
Inserting (\ref{mixpstress00})--(\ref{mixC00}) into the volume averaged form of (\ref{stressdef02}), we obtain the relation between the mixture polymer stress and the mixture conformation tensor as
\begin{align}
\label{mixstresstoC00}
{\stressPm} = G_m \left({\CT}_m - \IT\right).
\end{align}
Equation (\ref{mixstresstoC00}) is valid for two viscoelastic phases. However, we know a priori that the polymer stress is zero in the gas phase of the bubble. For the special case that only one phase is viscoelastic, eq.\ (\ref{mixstresstoC00}) reduces to
\begin{align}
\label{mixstresstoC01}
{\stressPm} =
\alpha \overline{\stressE}_{p}^{\phaseOne}
= \alpha G^{\phaseOne} \big(\overline{\CT}^{\phaseOne} - \IT\big).
\end{align}
Furthermore, the conformation tensor equation can be written as
\begin{align}
\label{constCEqva02}
&\partial_t \overline{\CT}^\phaseOne + \left( {\velocity}_m \dprod \nabla \right) \overline{\CT}^\phaseOne - \alpha {\LT}_m \dprod \overline{\CT}^\phaseOne - \alpha \overline{\CT}^\phaseOne \dprod \trans{({\LT}_m)} =
{\frac{\alpha}{\lambda^{\phaseOne}} \cvec{P}\big(\overline{\CT}^\phaseOne \big)}.
\end{align}
The derivation of (\ref{constCEqva02}) from (\ref{constCEq00}) is found in \cite{Niethammer2019}. To arrive at a more robust form, we transform (\ref{constCEq00}) using the change-of-variable representation (\ref{evolutionConf3}) to obtain
\begin{align}
\nonumber
&\partial_t {\cvec{F}(\overline{\CT}^\phaseOne)} + \left( {\velocity}_m \dprod \nabla \right) {\cvec{F}(\overline{\CT}^\phaseOne)} \\
= \; &
\label{constCEqva03}
\alpha \Big(
{2 {\BT}_m \dprod \overline{\UpsilonT}^\phaseOne \dprod \overline{\CT}^\phaseOne }
+
{\OmegaT}_m \dprod {\cvec{F}(\overline{\CT}^\phaseOne)} - {\cvec{F}(\overline{\CT}^\phaseOne)} \dprod {\OmegaT}_m
+
\frac{1\;}{\lambda^\phaseOne} \overline{\UpsilonT}^\phaseOne \dprod \cvec{P} (\overline{\CT}^\phaseOne ) \Big).
\end{align}
The change-of-variable representation (\ref{constCEqva03}) is the volume-averaged counterpart for the mixture model of the generic change-of-variable representation (\ref{evolutionConf3}) for a single-phase system. It is valid only if the first phase is viscoelastic, while the second is not. The application of a certain tensor function $\cvec{F}(\overline{\CT}^\phaseOne)$ can affect the numerical stability, as described in section~\ref{subsec:stab}. According to the benchmark results \cite{Niethammer2018}, an appropriate choice is the 4th-root tensor function, which is also used in the present work.
Equations (\ref{mixstresstoC01}) and (\ref{constCEqva03}) are used as a closure for the polymer stress term in (\ref{mixstress01}). Note that the polymer stress divergence in (\ref{mixstress01}) can be decomposed using the product rule as
\begin{align}
\label{closestress07}
\nabla \dprod \left(\alpha \overline{\stressE}_{p}^\phaseOne\right) = \alpha \nabla \dprod \overline{\stressE}_{p}^\phaseOne + \overline{\stressE}_{p}^\phaseOne \dprod \nabla \alpha.
\end{align}
This decomposition is used to separate pure interface contributions in the second term on the right-hand side of (\ref{closestress07}) from the remainder before the numerical discretization is applied. A discretization scheme for the interface contribution has been proposed in \cite{Niethammer2019}.
\subsection{Numerical setup}
\label{subsec:numericalsetup}
The flow curve and the shear stress as a function of the shear rate for the aqueous $0.8$~wt.~\% Praestol $2500$ (P2500) solution in simple shear flow are depicted in Fig.~\ref{fig:viscosityfit}. The experimental data show the shear-thinning behavior at shear rates above 1 $s^{-1}$. This behavior cannot be captured with simple two-parameter constitutive equations like the Oldroyd-B model. The Oldroyd-B model does not account for shear thinning, but uses a constant shear viscosity over $\dot{\gamma}$. The four-parameter exponential Phan-Thien Tanner (EPTT) model, in contrast, shows a shear-thinning behavior, which corresponds better to the measurements, as can be seen from the solid line in Fig.~\ref{fig:viscosityfit}. The EPTT parameter fit is identical to our previous work \cite{Niethammer2019}. Note that the fluid relaxation time $\lambda$ is estimated from measurements in elongational flow \cite{Pilz2007}. Therefore, only the remaining EPTT parameters, $\pvisc$, $\svisc$, $\epsilon$ and $\zeta$ are adjusted to the data from shear rheometry in Fig.~\ref{fig:viscosityfit}. In simple shear flow, a change of $\epsilon$ and $\zeta$ results in a different slope of the flow curve, but does not shift the curve along the shear-rate coordinate. Hence, when keeping $\lambda$ fixed, the viscosity curve must have a steeper slope to achieve good agreement with the experiments. The parameters are listed in Table \ref{tab:parameters}.\!\!\!\!
\begin{table}[h!]
\small
\begin{center}
\caption{\small Parameters for the polymer solution used for the investigation.}
\label{tab:parameters}
\begin{tabular}{@{}llllllll@{}}
\toprule
$w$ $[\%]$ \ & $\rho^\phaseOne$ $[\textnormal{kg}/\textnormal{m}^3]$ \ & $\eta_p^\phaseOne$ $[\textnormal{Ns}/\textnormal{m}^2]$ \ & $\eta_s^\phaseOne$ $[\textnormal{Ns}/\textnormal{m}^2]$ \ & $\lambda^\phaseOne$ $[\textnormal{s}]$ \ & $\sigma$ $[\textnormal{N}/\textnormal{m}]$ \ & $\epsilon^\phaseOne$ \ & $\zeta^\phaseOne$ \ \\
\midrule
$0.8$ & $1000.9$ & $1.483$ & $0.03$ & $0.203$ & $0.07555$ & $0.05$ & $0.12$ $\!$ \\
\bottomrule
\end{tabular}
\end{center}
\end{table}\
\input{figure/fig_viscosityfit.tex}\
Regarding the flow and shear stress curves in Fig.~\ref{fig:viscosityfit}, an important question is how shear thinning behavior affects the velocity jump discontinuity. Results from other researchers \cite{Pillapakkam2007} suggest that the velocity jump occurs also for rheological models with shear rate-independent viscosities. This means that shear thinning cannot be a necessary condition for the appearance of the velocity jump discontinuity. In the preparation of this work we made similar observations, employing the Oldroyd-B model. However, to obtain quantitative agreement of the bubble rise velocity with the experiments, account for shear-thinning was shown to be necessary \cite{Niethammer2019}. Therefore, the EPTT equations are used for all the simulations of the present study. The potential of this material model to exhibit shear-banding (cf.~\cite{Divoux2016}) for shear stresses between 5 Pa and 8 Pa (Fig. \ref{fig:viscosityfit} right) does not affect the bubble dynamics, since the shear stress $\tau_{r\theta}$ with effect for the $z$-momentum of the bubble, occurring along the lateral parts of the bubble, is well below 5 Pa. This was shown in Fig. 14 of our earlier paper \cite{Niethammer2019}.
The simulation is initiated by placing a spherical bubble in the center of a cubic computational domain. The size of the three-dimensional computational domain is defined with respect to the initial bubble diameter. The domain length $L$ equals 21.7 times the initial bubble diameter $d_b$. This ratio of $L/d_b$ is sufficient to eliminate wall effects in the flow field around the bubble for the mesh shown in Fig.~\ref{fig:meshdetail} \cite{Niethammer2019}. Local adaptive mesh refinement is used to increase
spatial resolution in the vicinity of the bubble interface, while less resolution is provided far away from the bubble to save computational resources. In this way, the best mesh quality is provided in the most important region near the bubble interface, where the mesh cells are uniform and well-spaced around the bubble, such that throughout the simulation the deforming fluid interface is always located inside the region of finest resolution. The resolution on the finest refinement level equals 108 cells per initial bubble diameter. The total number of mesh cells is $1.58$ million.
\begin{figure}[b!t]
\centering{
\includegraphics[width=0.99\textwidth]{figure/fig_meshdetail_s-figure0.pdf}
}
\caption{\small Visualization of the three-dimensional computational mesh M1. Left: Cutting planar mesh slice through the center of the computational domain. Right: Detail of the mesh in the vicinity of the bubble. The initial bubble shape is represented in gray.}
\label{fig:meshdetail}
\end{figure}
The system is solved in a moving reference frame, co-moving with the bubble's center of mass, such that the bubble remains in the center of the computational domain throughout the simulation. The use of a moving reference frame on a static mesh saves computational costs that otherwise would be caused by a re-meshing procedure. Due to the moving reference frame, the velocity boundary conditions of the computational domain must be adjusted relative to the motion of the reference frame at each time step. This is achieved by defining a Dirichlet boundary condition for the velocity on each side of the domain, except on the bottom. The Dirichlet boundary condition is set with the time-dependent frame velocity $\velocity_F$. The boundary values are updated in every time step of the simulation. Zero normal gradient boundary conditions are applied for the pressure, the polymer stress and the volumetric phase fraction. On the bottom face of the cube, different boundary conditions are employed. Since we expect an outflow there, zero normal gradient boundary conditions are applied for the velocity, the polymer stress and the volumetric phase fraction. For the pressure, a Dirichlet boundary condition is used with a constant value.
Due to the moving reference frame, an additional force term must be considered in the momentum balance, which results from the different acceleration of the bubble in the two reference frames. Denoting the frame acceleration with $\vec{a}_F$, the body force in the non-inertial frame becomes
\begin{align}
{\rho} \vec{b} = {\rho} \left(\vec{g} - \vec{a}_F\right),
\end{align}
where the $\vec{g}$ is the gravitational acceleration and the frame acceleration $\vec{a}_F$ is given by
\begin{align}
\vec{a}_F = \frac{d {\velocity}_F}{d t}.
\end{align}
The velocity ${\velocity}_F$ is computed in every time step of the simulation based on the relative movement of the center of mass of the bubble. To dampen small oscillations in the velocity field, a PD controller is applied for the computation of ${\velocity}_F$.
\section{Experimental results}
\label{sec:experimentalresults}
\subsection{Velocity fields around the bubbles}
\label{sec:PIV-velocityfields}
Velocity fields in the liquid phase around the rising bubbles were measured with PIV. For doing this, the liquid phase was seeded with fluorescent particles, as detailed in Sec. \ref{sec:PIV}. Upon formation of the bubbles from individual small bubbles, as detailed in Sec. \ref{sec:experimentalmethod}, it was noted that, for producing bubbles of a certain volume, more individual small bubbles were needed in the setup for the PIV measurements than in the bubble rise experiments (BRE) of \cite{Pilz2007}. For the BRE in \cite{Pilz2007}, 12 single bubbles were needed to form a bubble of $42.4~\textnormal{mm}^3$.
In the experiments for the PIV measurements, 12 single bubbles created a bubble of only $37.7~\textnormal{mm}^3$. According to the equation \cite{Brauer1971}
\begin{equation}
d = \left( \frac{3 \Phi \sigma d_{cap}}{\rho_l g} \right)^{1/3},
\label{eq:brauer1971}
\end{equation}
which predicts the size $d$ of gas bubbles formed by a quasi-static process at the end of a capillary, as used in the present experiments, this indicates that the surface tensions in the two sets of experiments were not the same. Equation (\ref{eq:brauer1971}) states that the ratio of the single-bubble volume in the PIV measurements (subscript PIV) to the one from the bubble rise experiments (subscript BRE) \cite{Pilz2007} is related to the ratio of the respective surface tensions according to \cite{Pilz2009}
\begin{equation}
\label{eq:diam-ratio-it}
\left(\frac{d_{\mathrm{PIV}}}{d_{\mathrm{BRE}}}\right)^{3} =
\frac{V_{\mathrm{PIV}}}{V_{\mathrm{BRE}}} =
\frac{\sigma_{\mathrm{PIV}}}{\sigma_{\mathrm{BRE}}} \; \text{,}
\end{equation}
provided that all the other parameters are identical.
Thus, when inserting the bubble volumes and the surface tension given in
\cite{Pilz2007} into (\ref{eq:diam-ratio-it}),
it follows that the interfacial tension in the PIV measurements amounted to only
$0.067~\textnormal{J}/{\textnormal{m}^2}$. The reduced surface tension may have been due to some Eosin that may have escaped from the tracer particles into the liquid. Inserting this value for the surface
tension into the correlation for the critical bubble volume at the jump given in \cite{Pilz2007},
a critical bubble volume of $38.8~\textnormal{mm}^3$ results.
\begin{figure}[t]
\centering
\includegraphics[scale=1.0]{figure/velocityvolume_C_piv-figure0.pdf}
\caption{\small Terminal bubble rise velocity vs.\ bubble volume in an aqueous
$0.8$~wt.~\% P2500 solution obtained from the BRE and the PIV experiments.}
\label{fig:JumpPIV-P2500}
\end{figure}
Terminal bubble rise velocities measured in the BRE of \cite{Pilz2007} and in the PIV experiments are depicted in Fig. \ref{fig:JumpPIV-P2500}. Comparison of the data shows that the jump is clearly visible in both data sets, and that the rise velocities from both sets of experiments are in the same order of magnitude. The critical bubble volumes are $46~\textnormal{mm}^3$ and $37~\textnormal{mm}^3$ in the BRE and the PIV experiments, respectively. The latter is close to the value of $38.8~\textnormal{mm}^3$ predicted due to the reduced surface tension. The lower surface tension may also be the reason that for the bubble rise velocities from the PIV experiments are slightly higher than in the BRE in \cite{Pilz2007}.
The velocity fields in the liquid phase around two different bubbles rising in an aqueous $0.8$~wt.~\% Praestol $2500$ solution measured with PIV are depicted in Fig.~\ref{fig:vorticity-P2500-subsupcritical}.
Fig.~\ref{fig:vorticity-P2500-subsupcritical}(a) shows a plot
of the instantaneous velocity vectors around a rising
bubble with a volume of $34.3~\textnormal{mm}^3$, which is just below the critical value for this setup.
The terminal bubble rise velocity is $24~\textnormal{mm}/\textnormal{s}$. In Fig.~\ref{fig:vorticity-P2500-subsupcritical}(b), the velocity field around a bubble with the volume $V = 37.7~\textnormal{mm}^3$ is shown. This bubble volume is above the critical value, and the terminal rise velocity is $110~\textnormal{mm}/\textnormal{s}$. The velocity fields correspond to the sub- and supercritical states, respectively, the supercritical field showing the negative wake.
\begin{figure}[t!]
\centering{
\begin{tabular}{@{}cc@{}}
\includegraphics[width=0.49\textwidth]{figure/P2500M22_FIG_VorticityC.pdf} &
\includegraphics[width=0.49\textwidth]{figure/P2500M21_2_FIG_VorticityC.pdf} \\%{P2500M21_2_FIG_Vorticity.png}
(a) & (b) \\
\end{tabular}
}
\caption{\small Instantaneous velocity field around a bubble with (a) subcritical volume
($U_{\mathrm{T}} = 24~\textnormal{mm}/\textnormal{s}$, $V = 34.3~\textnormal{mm}^3$) and (b) supercritical volume
($U_{\mathrm{T}} = 110~\textnormal{mm}/\textnormal{s}$, $V = 37.7~\textnormal{mm}^3$), rising in an aqueous $0.8$~wt.~\% solution of Praestol $2500$. The colors depict the vorticity $\omega_{1}$ defined in (\ref{eq:vorticity-outofplane-PIV}).}
\label{fig:vorticity-P2500-subsupcritical}
\end{figure}
The contours show the vorticity component $\omega_{1}$ corresponding to the measured velocity vectors, defined as
\begin{equation}
\label{eq:vorticity-outofplane-PIV}
\omega_{1} = \frac{\partial u_{3}}{\partial x_{2}} - \frac{\partial u_{2}}{\partial x_{3}}\; \text{.}
\end{equation}
The spatial derivatives in (\ref{eq:vorticity-outofplane-PIV}) were obtained using
the numerical gradient operation provided by the Matlab package.
\subsection{Transport of polymer molecules along the bubble contour}
\label{sec:transport-time}
In the literature discussed in the introduction, a possible explanation for the occurrence of the velocity jump phenomenon was proposed,
which is based on the assumption that the phenomenon is due to a change of
dominance between capillary and elastic forces.
If a given polymer solution has the potential for the discontinuity to occur,
the time scale of elastic stress relaxation in the liquid,
as compared to the time it takes to
transport the relaxing liquid portions downstream along the bubble contour,
decides about the sub- or supercritical state of the bubble:
the (hoop) stress set free due to the polymer relaxation generates
an additional elastic force on the bubble surface.
The vertical component of this force represents either a push upwards if the relaxation takes very long, or, if the relaxation is fast, an additional resistance,
depending on the position along the
bubble contour where the polymers have substantially relaxed and set free the hoop stress.
In order to evaluate the times of liquid transport down the bubble contour,
the transport time $\xi$ of a material element was estimated
by numerical evaluation of the integral
\begin{equation}
\label{eq:time-scale-bubblecontour}
\xi(l) = \int_{0}^{l} \frac{\mathrm{d} \tilde{l}}{\| \vec{u} \|}\; \text{,}
\end{equation}
where $\vec{u}$ represents the vector of the fluid velocity relative to the bubble, evaluated
along the bubble contour in the meridional plane, and $l$ is the arc length measured
from the north pole of the bubble (see Fig.~\ref{fig:image-processing-PIV}).
The front stagnation point is identified by $l = 0$, while the maximum value of $l$
is reached at the rear stagnation point and depends on the bubble volume and shape.
Fig.~\ref{fig:velocity-arclength} shows the distribution of $\| \vec{u} \|$ over
the arc length $l$ as measured by PIV along the supercritical bubble in Fig.~\ref{fig:vorticity-P2500-subsupcritical}b. The symmetry is reasonable. The velocity maxima are located in the equatorial plane of the bubble.
\begin{figure}[t]
\centering{
\includegraphics[width=0.8\textwidth]{figure/P2500M21_2_FIG_VelocityB.pdf
}
\caption{\small Norm of the velocity vectors versus arc length evaluated from the measured velocity field of an air bubble rising in an aqueous $0.8$~wt.~\% solution of Praestol $2500$ ($U_{\mathrm{T}} = 110~\textnormal{mm}/\textnormal{s}$, $V = 37.7 ~\textnormal{mm}^3$). The corresponding velocity field is shown in Fig.~\ref{fig:vorticity-P2500-subsupcritical}.}
\label{fig:velocity-arclength}
\end{figure}
\begin{figure}[h!t]
\centering{
\begin{tabular}{@{}cc@{}}
\includegraphics[width=0.49\textwidth]{figure/P2500M22_FIG_TimeC.pdf} &
\includegraphics[width=0.49\textwidth]{figure/P2500M21_2_FIG_TimeC.pdf} \\
(a) & (b) \\
\end{tabular}
}
\caption{\small Time $\xi$ elapsed for a fluid particle to travel the arc length $l$ along the contour of (a) a subcritical bubble ($U_{\mathrm{T}} = 24~\textnormal{mm}/\textnormal{s}$, $V = 34.3~\textnormal{mm}^3$) and (b) a supercritical bubble ($U_{\mathrm{T}} = 110~\textnormal{mm}/\textnormal{s}$, $V = 37.7~\textnormal{mm}^3$) in an aqueous $0.8$~wt.~\% solution of Praestol $2500$, compared to the relaxation time $\lambda_E = 0.207~\textnormal{s}$ of the solution.}
\label{fig:timescale-P2500-subcritical}
\end{figure}
The values of $\xi(l)$ defined by (\ref{eq:time-scale-bubblecontour}) are determined from the PIV velocity data along the half contours in the bubble's meridional plane, identified as the left-hand side (LHS) and right-hand side (RHS) of the contour, for check of symmetry.
The transport times for the flow along the bubble contour are compared to the
corresponding relaxation times $\lambda_{\mathrm{E}}$ obtained from elongational rheometry of the test liquids \cite{Pilz2009}.
Figure \ref{fig:timescale-P2500-subcritical}(a) shows results obtained for the
bubble with the subcritical volume of $34.3~\textnormal{mm}^3$ in Fig. \ref{fig:vorticity-P2500-subsupcritical}a, rising in an aqueous solution of $0.8$~wt.~\% Praestol $2500$. It is clearly visible that the relaxation process is already finished at the upper part of the bubble.
For the bubble with the supercritical volume of $37.7~\textnormal{mm}^3$ in Fig. \ref{fig:vorticity-P2500-subsupcritical}b, however, the transport time down the bubble and the relaxation time of the polymer solution appear to be of the same order of magnitude, as shown in Fig.~\ref{fig:timescale-P2500-subcritical}(b). In the supercritical case, therefore, the relaxation of polymeric stress in the liquid persists to the downstream part of the bubble \cite{Pilz2009}, providing a push and leading to the higher rise velocity.
\section{Numerical results}
\label{sec:numericalresults}
Numerical results of bubbles rising in an aqueous solution of $0.8$~wt.~\% Praestol $2500$ are presented in sections \ref{subsec:steadyjump} to \ref{subsec-conformation}. Transient, three-dimensional simulations were performed for various bubble volumes in, both, the subcritical and the supercritical state. The numerical setup described in section \ref{subsec:numericalsetup} was used for all simulations. In section \ref{subsec:steadyjump}, we first show a comparison of the steady-state rise velocities between experiment and simulation. In section \ref{subsec:negativewake}, the velocity field around the bubble is presented to examine the negative wake effect.
While similar results obtained with the same numerical method have already been reported in \cite{Niethammer2019}, we include these short sections for providing a complete picture.
The main numerical results originate from a local polymer stress and conformation tensor analysis, which is presented in sections \ref{subsec:stressdistribution} and \ref{subsec-conformation}, providing new insights into the molecular orientation and extension for the subcritical and the supercritical states of bubble motion.
\subsection{Terminal bubble rise velocity}
\label{subsec:steadyjump}
The validity and accuracy of the numerical method is verified by comparing the terminal rise velocities of bubbles of different volumes. A comparison of the terminal rise velocities as a function of the bubble volume is shown in Fig.~\ref{fig:velvol}.
\input{figure/fig_velvol.tex}
The simulation results obtained with the VOF method show good quantitative agreement with the experiments of Pilz and Brenn \cite{Pilz2007}. The terminal rise velocities of the supercritical bubble volumes are nearly identical to the experimental measurements. The subcritical bubble rise velocities show a slightly steeper increase with bubble volume than the experiments. It is worth to note that the critical bubble volume at about $V_{crit} = 46~\textnormal{mm}^3$ is quantitatively correctly captured by the simulations.
\subsection{Negative wake}
\label{subsec:negativewake}
Figure \ref{fig:negwake} shows the velocity vector field in a vertical cutting plane through the center of the rising bubble for three different bubble volumes when the terminal rise velocity is reached.
\input{figure/fig_negwake.tex}
The direction of the fluid velocity vectors of the subcritical and supercritical volumes differ significantly in the bubble wake. The flow field of the smallest bubble with $V = 30~\textnormal{mm}^3$ is directed upwards into the direction of motion of the rising bubble. This agrees with the flow behavior known from rising bubbles in Newtonian fluids. However, for the two supercritical bubbles with $V = 50~\textnormal{mm}^3$ and $V = 70~\textnormal{mm}^3$, the liquid velocity behind the bubble is pointing in the opposite direction to the bubble motion. The spatial extension of the negative velocities in the bubble wake can be determined from the color scheme, where blue represents negative velocities with respect to the vertical coordinate direction. The negative velocity is observed in an almost cone-shaped region below the bubble's trailing end. This is consistent with experimental findings of the negative wake in non-Newtonian fluids. For example, the PIV measurements in Fig.~\ref{fig:vorticity-P2500-subsupcritical}~(b) show negative fluid velocities in the bubble wake, too. Comparing the two supercritical bubbles with $V = 50~\textnormal{mm}^3$ and $V = 70~\textnormal{mm}^3$ suggests that, above the supercritical volume, the flow field is only subject to minor changes. In particular, the spatial extension of the negative wake is very similar for both bubble volumes.
\subsection{Stress distribution}
\label{subsec:stressdistribution}
The components of the stress tensor are transformed into spherical coordinates $\left(r, \theta, \phi \right)$, which is convenient for the analysis in the vicinity of the bubble interface. The origin of the spherical coordinate system is placed in the bubble’s center of mass. We use standard notation: $r$ is the radial distance from the origin, $\theta$ is the polar angle, with zero value in the vertical upward direction, and $\phi$ is the azimuthal angle. Since the bubble interface is approximately spherical in the upper half, the components $\stressE_{\theta \theta}$ and $\stressE_{\phi \phi}$ represent the tangential stress components, with the hoop stress $\stressE_{\phi \phi}$ acting in the circumferential direction (i.e., the $\phi$-direction) in a horizontal cutting plane through the bubble interface.
\begin{figure}[t!]
\centering{
\includegraphics[width=\textwidth]{figure/fig_tau_PP_sub.pdf}\vspace{-22pt}
}
\caption{\small Stress tensor component ${\tau}_{\phi\phi}$ (in Pa) in horizontal cutting planes at different heights (a) - (e). Subcritical bubble volume $V = 30~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution. The bubble interface is represented by the $0.5$ iso-surface of the volume fraction field (white) in (b) - (d). The color scale is adapted to the maximum stress in (b).}
\label{fig:tau_PP_sub}
\end{figure}
\begin{figure}[h!]
\centering{
\includegraphics[width=\textwidth]{figure/fig_tau_PP_super.pdf}\vspace{-22pt}
}
\caption{\small Stress tensor component ${\tau}_{\phi\phi}$ (in Pa) in horizontal cutting planes at different heights (a) - (e). Supercritical bubble volume $V = 60~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution. The bubble interface is represented by the $0.5$ iso-surface of the volume fraction field (white) in (b) - (d). The color scale is adapted to the maximum stress in (b).}
\label{fig:tau_PP_super}
\end{figure}
The stress component $\stressE_{\phi \phi}$ is displayed in Figs.~\ref{fig:tau_PP_sub} and \ref{fig:tau_PP_super} for a subcritical and a supercritical bubble volume, respectively. The stress field is shown in five horizontal cutting planes (a) to (e) at different heights. The planes (a) and (e) are located slightly above and below the bubble, respectively. The planes (b) to (d) intersect the bubble, where (c) represents the equatorial cutting plane\footnote{Note that the plane passing through this center point is not precisely the same as the plane in which the intersection with the bubble has maximum diameter. But the distance between these planes, measured by the angle between the lines passing through the center and the intersection point between the bubble surface and the respective plane, is less than $3^\circ$ in all cases.}. In the horizontal cutting planes (a), we see the formation of a hoop stress $\stressE_{\phi \phi}$ near the north pole. As the polar angle increases, the hoop stress decreases, which implies a relaxation of the polymer in the circumferential direction as it flows around the bubble; cf.~(b)~to~(d). Comparing the subcritical and the supercritical states, the hoop stress is significantly larger in the super- than in the subcritical state of the bubble, which can be seen from the color scales. The regions in the liquid affected by the stress are furthermore significantly larger in the super- than in the subcritical state. Moreover, the stress persists further downstream along the bubble, strongly present down to plane (d) in the supercritical state, while it has relaxed almost completely in the plane (c) for the subcritical state. The hoop stress for further bubble volumes and the stress components $\tau_{rr}$ and $\tau_{\theta\theta}$ are presented in Appendix \ref{asec:stresstensor}.
\subsection{Conformation tensor analysis}
\label{subsec-conformation}
The conformation tensor is a key element in a group of material models for polymeric substances. It is employed for the modeling of polymeric flows by formulating the relationship between the molecular polymer conformation and the bulk stress. The conformation is defined as the ensemble-averaged dyadic product of the end-to-end vector of the polymer molecule,
resulting from the respective placement of the rotatable Rouse segments along the macromolecule, with itself. This tensor, often denoted as $\CT$, can be interpreted as a tensorial measure of the molecular orientation \cite{Guenette1992}; see Appendix~\ref{asec:conftens} for more details. The tensor is symmetric and positive definite (spd). Its trace is the ensemble-average of the squared length of the molecules.
Values of the trace increasing in time therefore represent a stretching of the macromolecules. But note that the trace of $\CT$ does not describe the spatial orientation of the molecules.
\begin{figure}[b!]
\centering{
\includegraphics[width=\textwidth]{figure/fig_tr_C_sub.pdf}\vspace{-18pt}
}
\caption{\small Trace of the conformation tensor $\CT$ in horizontal cutting planes at different heights (a) - (e). Subcritical bubble volume $V = 30~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution. The bubble interface is represented by the $0.5$ iso-surface of the volume fraction field (white) in (b) - (d). The color scale is adapted to the maximum $\tr{\CT}$ values in (b).}
\label{fig:tr_C_sub}
\end{figure}
\begin{figure}[htbp!]
\centering{
\includegraphics[width=\textwidth]{figure/fig_tr_C_super.pdf}\vspace{-18pt}
}
\caption{\small Trace of the conformation tensor $\CT$ in horizontal cutting planes at different heights (a) - (e). Supercritical bubble volume $V = 60~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution. The bubble interface is represented by the $0.5$ iso-surface of the volume fraction field (white) in (b) - (d). The color scale is adapted to the maximum $\tr{\CT}$ values in (b).}
\label{fig:tr_C_super}
\end{figure}
Figures~\ref{fig:tr_C_sub} and \ref{fig:tr_C_super} depict the spatial distribution of the trace of $\CT$ for a sub- and a supercritical bubble, respectively.
The stretching of the molecules in the biaxial elongational flow around the upper bubble pole and in the uniaxial elongational flow around the downstream end of the bubble is clearly seen in the images. For the subcritical bubble, the flow causes molecular elongation in the biaxial flow along the upper hemisphere of the bubble
(Fig.~\ref{fig:tr_C_sub}~(a) and (b)), which relaxes before the molecules pass the equatorial plane (Fig.~\ref{fig:tr_C_sub}~(c))
and persists in this relaxed state (Fig.~\ref{fig:tr_C_sub}~(d)). A new elongation appears in the flow field around the lower bubble pole and in the wake flow downstream of it (Fig.~\ref{fig:tr_C_sub}~(e)).
In contrast, the flow around the supercritical bubble leads to a stronger stretching of the polymer molecules which also persists below the bubble equator (Fig.~\ref{fig:tr_C_super}~(a) to (e)) and down to the lower pole.
Consequently, the polymer molecules get strongly elongated near the upper pole and never fully relax during their traveling along the bubble surface.
Around the downstream pole, the trace of the conformation tensor does not show strong differences between the sub- and the supercritical states. This indicates that the elongation state in the lower pole region is caused by a different mechanism than the elongation in the upper bubble hemisphere.
\begin{figure}[t!]
\centering{
\includegraphics[width=\textwidth]{figure/fig_C_PP_sub.pdf}\vspace{-28pt}
}
\caption{\small Conformation tensor component ${C}_{\phi\phi}$ in horizontal cutting planes at different heights (a) - (e). Subcritical bubble volume $V = 30~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution. The bubble interface is represented by the $0.5$ iso-surface of the volume fraction field (white) in (b) - (d). The color scale is adapted to the maximum ${C}_{\phi\phi}$ values in (b).}
\label{fig:C_PP_sub}
\end{figure}
\begin{figure}[h!]
\centering{
\includegraphics[width=\textwidth]{figure/fig_C_PP_super.pdf}\vspace{-28pt}
}
\caption{\small Conformation tensor component ${C}_{\phi\phi}$ in horizontal cutting planes at different heights (a) - (e). Supercritical bubble volume $V = 60~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution. The bubble interface is represented by the $0.5$ iso-surface of the volume fraction field (white) in (b) - (d). The color scale is adapted to the maximum ${C}_{\phi\phi}$ values in (b).}
\label{fig:C_PP_super}
\end{figure}
Inspection of the conformation tensor component ${C}_{\phi\phi}$ in Figs. \ref{fig:C_PP_sub} and \ref{fig:C_PP_super} shows that the elongation in the azimuthal angular (i.e., circumferential) direction is dominant around the upper bubble hemisphere, and, again, that this elongation survives far down the supercritical bubble (Fig.~\ref{fig:C_PP_super}). The data downstream from the bubble's rear pole show that the molecular elongation there is not due to the azimuthal component. It is even smaller there than in the surrounding liquid bulk, where the azimuthal orientation occurs randomly. The very small values of the tensor component ${C}_{\phi\phi}$ in this region shows that the elongation which is present there (cf.\ the results on the conformation tensor trace discussed above) must result from stretching in another direction,
which turns out to be the radial one; the latter conclusion is supported from the supplementary results on ${C}_{rr}$ given in Appendix~\ref{asec:conformationtensor}.
The conformation and orientation of the polymer molecules in the flow around the bubble can be quantified by a spectral analysis of the conformation tensor. For this purpose, the eigenvectors of the tensor with their corresponding eigenvalues are studied, as depicted in Figs.~\ref{fig:ev_sub} and \ref{fig:ev_super} for the sub- and supercritical bubble, respectively.
{\begin{figure}[t!]
\centering{
\includegraphics[width=0.993\textwidth]{figure/fig_EV_sub.pdf}\vspace{-21.5pt}
}
\caption{\small Line elements visualizing the direction of the eigenvectors which belong to the largest eigenvalues ${d_{\max}}$ of the conformation tensor in horizontal cutting planes at different heights (a) - (e), colored corresponding to the value of ${d_{\max}}$. Subcritical bubble volume $V\,=\,30~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution.}
\label{fig:ev_sub}
\end{figure}
\begin{figure}[h!]
\centering{
\includegraphics[width=0.993\textwidth]{figure/fig_EV_super.pdf}\vspace{-21.5pt}
}
\caption{\small Line elements visualizing the direction of the eigenvectors which belong to the largest eigenvalues ${d_{\max}}$ of the conformation tensor in horizontal cutting planes at different heights (a) - (e), colored corresponding to the value of ${d_{\max}}$. Supercritical bubble volume $V~=~60~\textnormal{mm}^3$ in an aqueous $0.8$~wt.~\% Praestol $2500$ solution.}
\label{fig:ev_super}
\end{figure}}
The figures display, locally in every mesh cell, the direction of the eigenvector which corresponds to the largest eigenvalue in this position. This is indicative for the mean polymer molecular orientation, since the square root of the eigenvalues is a measure for the length of the polymer end-to-end vector projected on the direction given by the corresponding eigenvector; cf.\ Appendix~\ref{asec:conftens}. From this analysis, it is seen that the formation of large eigenvalues is more pronounced in the supercritical than in the subcritical case, and that the large eigenvalues persist further downstream along the supercritical bubble. While the predominantly circumferential orientation of the eigenvectors, which is characteristic of planes (a) - (c) in both cases, is lost in plane (d) for the subcritical bubble, it persists for the supercritical one. Even in plane (e), where the eigenvectors are re-oriented in a predominantly radial direction for the subcritical bubble, the supercritical case still exhibits a marked circumferential component. This different behavior of the polymer molecular conformation in the sub- and the supercritical cases will be further analyzed in the following section by theoretical means.
\section{Theoretical results}
\label{sec:theory-results}
The analysis of the local conformation tensor from the DNS results as given in subsection~\ref{subsec-conformation} shows that, predominantly, the polymer molecules get oriented and stretched in a circumferential direction as they move along the bubble's upper hemisphere. This is a purely kinematic effect, since the distance on neighboring longitudinal curves grows with the polar angle $\theta$ as $\sin \theta $. The stretching of the polymer molecules is mediated via friction between the solvent and the polymer molecules. A part of the kinetic energy is stored in the polymer molecules due to their entropic elasticity.
The release of the stored energy takes place with a time delay which is intimately related to the relaxation time of the polymer molecules.
The retraction of the polymer particles generates a hoop stress, which acts on the bubble surface.
Depending on the ratio of the polymer relaxation time to the time scale of advection, this hoop stress is either applied above or below the bubble's equator. In the former case this leads to a force against the bubble rise, while the hoop stress pushes the bubble upwards in the latter case.
In order to understand whether this perception is correct and explains the abruptness of the change in rise velocity at a critical bubble state, we further investigate the transport of the polymer molecules around the bubble. Since the numerical simulations revealed the relevance of the circumferential component of the conformation tensor, we study the evolution of this component in more detail by theoretical means.
\subsection{Polymer orientation-deformation evolution}
The key process in the flow of the polymeric liquid around the rising bubble is the transport and deformation of the polymer molecules. The ensemble-average state of conformation of individual polymer
molecules is described by the orientation-deformation tensor $\pmb L$, which satisfies the transport equation \cite{Yarin1993}
\begin{equation}
\frac{{\rm d} \pmb{L}}{{\rm d}t} = \pmb{\nabla} \pmb{u} \cdot \pmb{L} + \pmb{L} \cdot \pmb{\nabla} \pmb{u}^{\sf T} - \frac{2}{\lambda} \left( \pmb{L} - \frac{1}{3} N_{01} b_1^2 \pmb{I}\right),
\label{eq:conform}
\end{equation}
where $\lambda$ is the stress relaxation time, and $N_{01}$ and $b_1$ are the number and length of the Kuhn segments constituting the polymer molecule. The flow around the upstream hemisphere of the rising bubble is a biaxial straining flow with the dominant straining component oriented in the direction of the azimuthal angle $\phi$ of the spherical coordinate system. Consequently, as already underlined by the numerical results shown above, it is advisable to analyze the transport of the orientation-deformation tensor component $L_{\phi \phi}$. Under the reasonable assumption of an axisymmetric, non-swirling flow (both w.r.\ to the vertical axis through the bubble center), the differential equation for the azimuthal conformation tensor component decouples from the other components and reads
\begin{equation}
\frac{{\rm d}L_{\phi\phi}}{{\rm d}t}=
2 \, \left( \frac{u_r}{r} + \frac{u_\theta}{r} \cot \theta \right) L_{\phi\phi} - \frac{2}{\lambda} \left(L_{\phi\phi} -\frac{1}{3} N_{01} b_1^2 \right)
\label{eq:conform_phi}
\end{equation}
where $u_r$ and $u_\theta$ denote the radial and polar angular velocity components in spherical coordinates, respectively.
We add the assumption of a solenoidal flow field, which is appropriate for polymer solutions with small variations in the mixture density. Since we are interested in the quasi-steady situation of a bubble rising at fixed terminal velocity and with constant shape, we switch to the inertial frame with the bubble center of mass as its origin.
Then the velocity is given via a Stokes stream function $\psi = \psi(r,\theta)$ according to
\begin{equation}
u_r = - \, \frac{1}{r^2 \sin \theta}\, \partial_\theta \psi, \qquad
u_\theta = \frac{1}{r \sin \theta}\, \partial_r \psi.
\label{eq:Velocity_from_streamfunction}
\end{equation}
In order to arrive at a closed form of the ordinary differential equation (ODE) (\ref{eq:conform_phi}), we further assume that the bubble shape resembles a half-sphere in the upper half space (i.e., for $0\leq \theta \leq \pi / 2$) and that the Hadamard-Rybczynski flow provides a reasonable approximation of the velocity field at least in the upper half space. More precisely, we assume the
functional dependence of the flow (in the upper half space) to be governed by
that of the Hadamard-Rybczynski solution, with the rise velocity $U_{\mathrm{T}}$ taken as the actual terminal velocity of the bubble (and not determined as a part of the Hadamard-Rybczynski solution). This is an admissible flow field, since the quasi-steady two-phase Stokes problem in a co-moving frame with fixed interface is a linear problem. As visible in Fig. \ref{fig:velocity_at_surface}, these assumptions are indeed sensible, since the numerically computed flow field resembles this analytical solution even quantitatively. This corresponds to the Reynolds numbers $\leq O(0.1)$ of the bubble rise in the experiments; cf.\ also
Figure~\ref{fig:velocity-arclength}.
\begin{figure}[htbp!]
\centering{
\includegraphics[width=0.49\textwidth]{figure/V30_UT_fit.pdf}
\includegraphics[width=0.49\textwidth]{figure/V60_UT_fit.pdf}
}
\caption{\small Velocity component $u_{\theta}$ near the bubble interface as a function of the polar angle. Note that the interface is represented as an iso-surface of the volume fraction field and therefore does not correspond to the exact position of the interface, which cannot be determined precisely in the algebraic VOF method.}
\label{fig:velocity_at_surface}
\end{figure}
With these assumptions, the stream function in the form normalized by the bubble rise velocity $U_{\mathrm{T}}$ and radius $R$, i.e.\ $\psi^* := \psi / (U_{\mathrm{T}} R^2 / 2)$, and with the normalized radial coordinate $r^* := r/R$ reads
\begin{equation}
\psi^* (r^* ,\theta ) = r^* (r^* -1) {\sin}^2 \theta.
\label{eq:HR-streamfunction}
\end{equation}
With these notations, the velocity components normalized by $U_{\mathrm{T}}$ read
\begin{equation}
u^*_r (r^*, \theta) = \Big( \frac{1}{r^*} -1 \Big) \cos \theta, \qquad u^*_\theta (r^*, \theta) = \Big( 1 - \frac{1}{2 r^*} \Big) \sin \theta.
\label{eq:hadaryb}
\end{equation}
Around the upstream bubble hemisphere, the radial velocity component in the liquid is negative, and the polar angular component is positive, indicating a flow towards the bubble directed in the negative $z$ direction at large distances from the bubble. Insertion of \eqref{eq:hadaryb}
into \eqref{eq:conform_phi} and normalizing the time variable by $R/U_{\mathrm{T}}$ and the circumferential conformation tensor component by $N_{01} b^2_1 / 3$,
yields the dimensionless form of the circumferential conformation evolution as
\begin{equation}\label{eq:Lphiphi_norm}
\frac{{\rm d}L^*_{\phi\phi}}{{\rm d}t^*}=
\frac{\cos \theta}{(r^*)^2} L^*_{\phi\phi}
- \frac{1}{{\rm De}_c} \big( L^*_{\phi\phi} -1 \big),
\end{equation}
where the dimensionless tensor $\pmb{L}^*$ is identical to the tensor $\pmb{C}$ in Section~\ref{sec:numericalmethod}.
In eq.\ \eqref{eq:Lphiphi_norm} we have used the convective Deborah number, which we define as
\begin{equation}
{\rm De}_c = \frac{\lambda U_{\mathrm{T}}}{2 R}.
\label{eq:def_Deborah}
\end{equation}
To finally obtain a closed form of the ODE, one replaces the time variable by either $\theta$ or $r^*$ and then eliminates the other remaining variable
using \eqref{eq:HR-streamfunction} along the streamlines $\psi^* = {\rm constant}$. We choose $\theta$ as the new independent variable, let $y(\theta):=L^*_{\phi\phi}(t^* (\theta))$
and obtain, employing
\begin{equation}
\frac{{\rm d}\theta}{{\rm d} t^*} = \frac{u^*_\theta}{r^*}
=\frac{1}{r^*} \Big( 1-\frac{1}{2 r^*} \Big) \sin \theta,
\end{equation}
the ODE
\begin{equation}
y'(\theta)=
\frac{2}{\sqrt{4 \psi^* + {\sin}^2 \theta}} \,
\Big( \cos \theta \, y(\theta)- \frac{r^* (\theta)^2}{{\rm De}_c} (y(\theta) -1) \Big),
\label{eq:conform_phi_theta}
\end{equation}
where, along the streamline, the normalized radial coordinate $r^*$ depends
on $\theta$ according to
\begin{equation}
r^* (\theta) = \frac{1}{2} + \sqrt{\frac{1}{4} + \frac{\psi^*}{\sin^2 \theta}}.
\label{eq:streamline}
\end{equation}
Equation \eqref{eq:conform_phi_theta} is a linear, inhomogeneous ordinary
differential equation of first order and with variable coefficients.
Since the coefficient functions are continuously differentiable (hence, in particular, locally Lipschitz continuous), the initial value problems associated to \eqref{eq:conform_phi_theta} are uniquely solvable.
Since $r^* (\theta)$ is bounded away from $\theta = 0$ and $\theta = \pi$, it is also clear that solutions exist globally, i.e.\ on the full interval $(0,\pi)$.
Moreover, any solution starting in some $\theta_0 \in (0, \pi)$ with an initial value $y_0 > 0$,
stays positive on $(\theta_0, \pi)$, which follows from
the fact that the right-hand side in \eqref{eq:conform_phi_theta}
is positive if a solution reaches zero, i.e.\ $y(\theta)=0$ implies $y' (\theta)>0$.
An analytical representation of the general solution can -- in principle -- be obtained by the variation-of-constants-formula. But since the primitives which appear in the integration are not given by elementary functions, this does not help to analyze the behavior w.r.\ to the Deborah number.
\begin{figure}[b!]
\centering{
\includegraphics[width=\textwidth]{figure/Fig_de_variation.png} \\
\includegraphics[width=\textwidth]{figure/Fig_psi_variation.png}\vspace{-18pt}
}
\caption{\small Top row: vector fields of the ODE \eqref{eq:conform_phi_theta} for different values of ${\rm De}_c >0$ at $\psi^* = 0$, together with the curve of critical points (maxima) of the right-hand side of \eqref{eq:conform_phi_theta}. Bottom row:
vector fields of the ODE \eqref{eq:conform_phi_theta} for ${\rm De}_c =1$ and different values of $\psi^* > 0$, corresponding to minimal $r^*$ of $1.05$, $1.1$ and $1.2$.}
\label{fig:de_variation}
\end{figure}
Instead, to get a first impression of the influence of especially the Deborah number, the top row of plots in Fig.~\ref{fig:de_variation} shows the vector fields associated with the ODE \eqref{eq:conform_phi_theta} along the streamline $\psi^* = 0$ for different values of ${\rm De}_c >0$. In addition, the dashed curve shows all points $(\theta, y)$, in which the right-hand side of \eqref{eq:conform_phi_theta} vanishes. Hence, a solution which passes through a point on this curve has a horizontal tangent there; in fact, it has a local maximum at this point. Consequently, solutions increase as long as they run below the dashed line, and they decrease while they run above it. Note, however, that this curve
itself is not a solution. The line element plots clearly indicate that solutions which start at some (small) $\theta_0 >0$ with a value of $y_0 \approx 1$, say, reach a maximum which becomes higher for larger values of ${\rm De}_c$. This is also reflected by the dashed curves of critical points of the ODE, which are given as
\begin{equation}\label{maxima-curve}
y(\theta ;{\rm De}_c) = \frac{1}{1- {\rm De}_c \cos \theta }
\quad \mbox{for } \theta \in (\theta_0 ({\rm De}_c), \pi ),
\end{equation}
where $\theta_0 ({\rm De}_c)=0$ for ${\rm De}_c \leq 1$ and
$\theta_0 ({\rm De}_c)=\arccos ({\rm De}_c^{-1})$ for ${\rm De}_c > 1$.
In the latter case and for $\theta \in (0, \theta_0)$, every solution
is strictly increasing there for all values of $y>0$.
Equation \eqref{maxima-curve} indicates that ${\rm De}_c =1$ is a critical value for the Deborah number, since the possible maxima are all finite for ${\rm De}_c < 1$, while arbitrarily large maxima are -- in principle -- possible in case of ${\rm De}_c > 1$.
Concerning the choice of the initial value as $y_0 \approx 1$, recall that the dimensionless orientation-deformation tensor ${\bf L}^*$ equals the identity tensor if the polymer molecules are in a relaxed and unoriented state.
\begin{figure}[t!]
\centering{
\includegraphics[width=\textwidth]{figure/Fig_de_var_solutions.png
}
\caption{\small Solutions of \eqref{eq:conform_phi_theta} for different Deborah numbers (${\rm De}_c =0.5$, $0.8$, $1.0$, $1.2$ plotted in blue, orange, green, red, respectively) and $\psi^* =0$. All solutions start at $y=1$, but at different initial angles: $\theta_0 = 0.001$ (left), $\theta_0 = 0.01$ (right).}
\label{fig:de_variation_solutions}
\end{figure}
\noindent
To come to rigorous conclusions about the influence of the Deborah number, note that in the complete continuum mechanical model from subsection~\ref{subsec:governing-equations}, the forces on the bubble are determined by the pressure distribution and the viscoelastic stresses directly at the bubble surface. While the transport equation \eqref{eq:conform} of the conformation tensor alone does not account
for any back-effect onto the flow field, the limiting case of $\psi^* \to 0+$, corresponding to the streamline running along the bubble surface, nevertheless is the most relevant case. At this point it is important to observe that the right-hand side in \eqref{eq:conform_phi_theta} depends continuously on $\psi^* \geq 0$ and, of course, $r^* (\theta) \to 1$ in the limit as $\psi^* \to 0+$.
The dependence on $\psi^* \geq 0$ is depicted in the bottom row of
Figure~\ref{fig:de_variation} in the critical case ${\rm De}_c =1$: increasing $\psi^*$ corresponds to streamlines passing the bubbles at larger distances and the polymer molecules on more distant streamlines get less stretched in the circumferential direction, experiencing kinematic stretching only at a much weaker intensity.
The solutions of \eqref{eq:conform_phi_theta} for $\psi^* >0$ approach the solutions of the limiting ODE, where the latter reads
\begin{equation}
y'(\theta)=
\frac{2 \cos \theta}{\sin \theta} \, y(\theta)
- \frac{2}{{\rm De}_c \, \sin \theta} (y(\theta) -1).
\label{eq:conform_phi_theta_surface}
\end{equation}
The general solution of \eqref{eq:conform_phi_theta_surface} is
\begin{equation}
y(\theta) = y_p(\theta) + C \frac{\sin^2\theta}{\tan^{2/{\rm De}_c} \big( \theta / 2 \big)}
\end{equation}
with arbitrary $C\in \mathbb{R}$ and the particular solution
\begin{equation}
y_p(\theta)= \frac{1}{2} \sin^2 \theta + \frac{1}{1+ {\rm De}_c} \left(\frac{1-\cos\theta}{2}\right)^2 + \frac{1}{1- {\rm De}_c} \left(\frac{1+ \cos\theta}{2}\right)^2
\end{equation}
in case ${\rm De}_c \neq 1$, and
\begin{equation}
y_p(\theta)= \frac{1}{2} \sin^2 \theta + \frac{1}{2} \left(\frac{1-\cos\theta}{2}\right)^2 + \frac{1}{2} \ln (\tan (\frac{\theta}{2}))
\end{equation}
in case ${\rm De}_c = 1$.
Now observe that
\begin{equation}
y_p(0+) = \frac{1}{1- {\rm De}_c}
\quad \mbox{as $\theta \to 0+$ in case } {\rm De}_c \neq 1.
\end{equation}
In case ${\rm De}_c >1$, this implies $y(0+) = 1/(1- {\rm De}_c)$
independently of $C\in \mathbb{R}$. On the other hand, in case ${\rm De}_c <1$ the solution has the same limit as $\theta \to 0+$ only if $C=0$,
while for $C\neq 0$ the solution diverges to $\pm \infty$ as $\theta \to 0+$, where the sign is given by ${\rm sgn}(C)$.
While the behavior differs depending on the value
of the Deborah number, it is not clear how to formulate reasonable initial value problems which allow to directly assess the qualitative behavior of prototypical solutions. This is also supported by the sample solutions shown in Fig.~\ref{fig:de_variation_solutions} (left, middle): the solutions starting with $y_0=1$ but at different $\theta_0>0$ close to zero, strongly depend on $\theta_0$. This is
explained by the fact that the upper pole is an accumulation point
of the velocity field, i.e.\ solutions for smaller $\theta_0$ stay
significantly longer close to this pole.
In order to clarify whether the value of ${\rm De}_c$ impacts the solutions of \eqref{eq:conform_phi_theta_surface} in a way explaining the rise velocity jump, the crucial question is under which conditions the ODE \eqref{eq:conform_phi_theta_surface} has solutions starting at $0< \theta_0 \approx 0$ from values $y_0 \approx 1$ and reach the bubble's equator (i.e., $\theta = \pi /2$) in a significantly extended state, which is still relaxing, such that forces from the contracting polymer molecules will accelerate the bubble.
Evaluating the right-hand side of \eqref{eq:conform_phi_theta_surface} at $\theta = \pi /2$ yields
\begin{equation}\label{eq:9}
y' \left(\frac{\pi}{2}\right)= -\, \frac{1}{2\, {\rm De}_c}
\big(y(\frac{\pi}{2})-1\big).
\end{equation}
This shows that the polymer molecules passing by the bubble's equator are relaxing ($y' <0$) if and only if they are stretched in the circumferential direction ($y >1$), and the relaxation is the stronger, the larger the extension in the circumferential direction, i.e.\ the larger $y>1$. Remarkably, this relation is independent of ${\rm De}_c$. Note also that the same conclusion holds in the general case of \eqref{eq:conform_phi_theta}. However, we keep focussing on the limiting case $\psi^* \to 0+$, since the solutions of \eqref{eq:conform_phi_theta} converge to those of \eqref{eq:conform_phi_theta_surface} for fixed Deborah number; cf.~Fig.~\ref{fig:de_variation_solutions} (right).
To answer the question under which conditions the ODE \eqref{eq:conform_phi_theta_surface} allows for solutions which reach a value significantly above $y=1$ at $\theta = \pi/2$,
we solve \eqref{eq:conform_phi_theta_surface} backwards, starting at $\theta = \pi/2$. We therefore consider $u(\theta)= y(\pi /2 -\theta)-1$ for $\theta \in [0,\pi/2)$ with $u(0)=u_0$, i.e.\ we study the initial value problem
\begin{equation}\label{eq:12}
u'(\theta)=2 \frac{1/{\rm De}_c - \sin\theta}{\cos\theta} u(\theta) - 2 \frac{\sin\theta}{\cos\theta}, \quad u(0)=u_0.
\end{equation}
For ${\rm De}_c \neq 1$, the initial value problem \eqref{eq:12} has the unique solution
\begin{equation}\label{eq:15}
u(\theta)= \frac{(1+\sin\theta)^2}{4({\rm De}_c^{-1} -1)}
- \frac{(1-\sin\theta)^2}{4({\rm De}_c^{-1} +1)}
+ C_0 \, \cos^2 \theta \Big( \frac{1+\sin\theta}{1-\sin\theta} \Big)^{{\rm De}_c^{-1}}
\end{equation}
with
\begin{equation}\label{eq:16}
C_0 = u_0 + \frac{1}{4({\rm De}_c^{-1} +1)} - \frac{1}{4({\rm De}_c^{-1} -1)}.
\end{equation}
To analyze the behavior of the solution $u(\cdot)$ of \eqref{eq:12} as $\theta$ approaches $\pi/2$, we rewrite \eqref{eq:15} as
\begin{equation}\label{eq:15b}
u(\theta)= \frac{(1+\sin\theta)^2}{4({\rm De}_c^{-1} -1)}
- \frac{(1-\sin\theta)^2}{4({\rm De}_c^{-1} +1)}
+ C_0 \, (1+\sin \theta)^2 \Big( \frac{1+\sin\theta}{1-\sin\theta} \Big)^{{\rm De}_c^{-1}-1}.
\end{equation}
Now, as $\theta \to \frac{\pi}{2}-$, all terms converge to a finite value, except for the last factor, which tends to zero in case of ${\rm De}_c >1$, while it diverges to $+ \infty$ in case of ${\rm De}_c <1$. Hence, the qualitative behavior of $u(\cdot)$ for $\theta \to \frac{\pi}{2}-$ is as follows:
\vspace{12pt}\\
{\bf Case ${\rm De}_c <1$. } We must distinguish two subcases:
\begin{equation}
u(\theta)\to - \infty \mbox{ as } \theta \to \frac{\pi}{2}-
\; \mbox{ if }
u_0 < \frac{1}{4({\rm De}_c^{-1} -1)} - \frac{1}{4({\rm De}_c^{-1} +1)}
\end{equation}
and
\begin{equation}
u(\theta)\to + \infty \mbox{ as } \theta \to \frac{\pi}{2}-
\; \mbox{ if }
u_0 > \frac{1}{4({\rm De}_c^{-1} -1)} - \frac{1}{4({\rm De}_c^{-1} +1)}.
\end{equation}
\begin{figure}[t!]
\centering{
\includegraphics[width=0.8 \textwidth]{figure/y_at_equator_reduced.png}
}
\caption{\small Blue curve: Maximum reachable circumferential conformation component at the bubble equator as a function of the Deborah number in case ${\rm De}_c <1$. Other curves: circumferential conformation component at the bubble equator as a function of the Deborah number for initial values of $y_0=1$ (orange), $y_0=3$ (green) and $y_0= 5$ (red)
and initial polar angles of $\theta_0 =0.01$ (lowest), $\theta_0 =0.001$ (middle) and $\theta_0 = 0.0001$ (highest).}
\label{fig:ymax-plot}
\end{figure}
\noindent
In the first subcase, the limit of $-\infty$ of the backward solution $u(\cdot)$ implies for the original ODE that every $y_0=u_0 + 1$ with such a $u_0$ is reachable at the bubble's equator from an appropriate (possibly large) initial value $y_0$ and initial $\theta_0 >0$ close to zero. In the second subcase, the respective values $y_0=u_0 + 1$ are
not reachable at all. In other words, in case of ${\rm De}_c <1$, there exists a maximal value $y_{\rm max}$, depending on ${\rm De}_c$ according to
\begin{equation}\label{eq:ymax}
y_{\rm max} ({\rm De}_c) = 1 + \frac{1}{4({\rm De}_c^{-1} -1)} - \frac{1}{4({\rm De}_c^{-1} +1)} = \frac{1- 0.5 \, {\rm De}_c^{2}}{1- {\rm De}_c^{2}},
\end{equation}
which is the maximum reachable circumferential conformation component at this value of the Deborah number. The relation \eqref{eq:ymax} is plotted as the blue curve in Fig.~\ref{fig:ymax-plot}.
For comparison, the other curves shown in this figure visualize the values reachable from a certain set of initial values. \vspace{12pt}\\
{\bf Case ${\rm De}_c >1$: } Then
\begin{equation}
u(\theta)\to \frac{1}{4({\rm De}_c^{-1} -1)} - \frac{1}{4({\rm De}_c^{-1} +1)} \, < \, 0.
\end{equation}
In this case, arbitrary large values of $y$ are reachable at the bubble equator for appropriate (correspondingly large) initial values $y_0$ at initial $\theta_0 >0$ close to zero. \\
This analysis substantiates and confirms the impression from the vector field plots in Fig.~\ref{fig:de_variation}
and leads us to the following conclusion:
in case of ${\rm De}_c <1$, i.e.\ for fast rel\-ax\-ation of the polymer molecules compared to the convective time scale, the normalized circum\-ferential conformation component $L_{\phi \phi}^*$ cannot attain a value significantly larger than 1 at the equator, starting at a value of about 1 near the upper bubble pole. In turn, in case of
${\rm De}_c >1$, if $L_{\phi \phi}^*$ is just slightly larger than 1 at small polar angle $\theta$, then $L_{\phi \phi}^*$ can reach values significantly above 1 at the bubble's equator.
Thus, at small Deborah number the major part of the relaxation takes place before the molecules reach the bubble's equator, while, at high Deborah number, a strong relaxation persists until below the equator. In the latter case, the bubble experiences an upward pointing force due to a hoop stress below its equator, which leads to an increased rise velocity. Since this effect is self-amplifying, a velocity jump arises at a critical Deborah number, where the latter is given by ${\rm De}_{\rm crit} =1$ in the idealized scenario considered above.
\section{Discussion and conclusions}
\label{sec:discussconclusions}
The aim of this study is to identify the molecular mechanism behind the formation of sub- or supercritical terminal rise velocities of individual gas bubbles in polymer solutions, providing a physical explanation of the mechanism which leads to a jump discontinuity, i.e.\ to an abrupt increase of the terminal rise velocity for a slight increase in bubble size at a critical volume.
For this purpose, and in order to allow for a cross-validation of our findings, we combine three different approaches: (i) an experimental study, providing visual information on the bubble shape complemented by PIV-based measurements of the local velocity fields, (ii) a computational approach, employing fully resolved numerical simulations of the continuum physical governing equations employing appropriate rheological models which account for the orientation-deformation state of the polymer molecules via a conformation tensor and (iii) a theoretical investigation, which focusses on understanding the evolution of the most relevant conformation tensor component, which turns out to be the circumferential one. We first summarize the main findings from the different approaches.
Ad (i). While the experiments confirm the occurrence of the rise velocity jump discontinuity at a certain critical bubble volume (cf.\ Fig.~\ref{fig:JumpPIV-P2500}), which depends in particular on the liquid material properties, the results presented in Section~\ref{sec:experimentalresults} go considerably beyond the existing knowledge on integral quantities. Indeed, the PIV data not only display
the negative wake in the velocity field below supercritical bubbles (cf.\ Fig.~\ref{fig:vorticity-P2500-subsupcritical}), but the local velocity information is analyzed regarding Lagrangian transport times of the polymer molecules. This ana\-ly\-sis shows that polymer molecules, starting from the region at the upper pole and traveling along (or close to) the bubble surface in a time interval longer than the relaxation time, cannot reach the equator in a stretched state (cf.\ Fig.~\ref{fig:timescale-P2500-subcritical}(a)), while they evidently reach the lower hemisphere stretched in the supercritical case, in which the transport and stress relaxation times are similar
(cf.\ Fig.~\ref{fig:timescale-P2500-subcritical}(b)).
The crucial point in this experimental finding is that the relaxation time is, hence, comparable to bubble-related hydrodynamic transport times, but clearly different ratios of the relaxation to the convection time scales distinguish the sub- from the supercritical bubbles. This shows that the rise velocity jump discontinuity is related to the interplay of convection and relaxation and their characteristic time scales.
Ad (ii).
The numerical approach enables simulations which solve the initial-boundary value problem for the buoyancy-driven bubble motion in a polymeric liquid and capture the complete fields of the velocity within the liquid and the gas phase as well as the polymer orientation-deformation tensor as functions of space and time. The simulation results therefore include all the information needed for a physical explanation of the terminal bubble rise velocity, including the genesis of the jump discontinuity. In particular, our numerical approach delivers the correct rise velocities, yields a negative wake in case of supercritical bubbles, and also realistic bubble shapes \cite{Niethammer2019}, although a true cusped shape cannot be accurately reproduced, as with most numerical two-phase flow methods. But the latter is not problematic, since the cusped shape is known to be of minor importance for the jump discontinuity of the rise velocity (cf.\ the survey in Section~\ref{sec:introduction}); we shall briefly come back to this point below.
With fully resolved numerical simulations at hand, it is possible to analyze the local field quantities, in particular the stress acting on the bubble surface. This approach was already employed in \cite{Pillapakkam2007} and in further papers since then, including our work \cite{Niethammer2019}. Up to now, the focus in all contributions was completely on the radial and the polar angular normal stress components.
The corresponding results, such as the ones shown in Figs.~\ref{fig:tau_TT_sub}--\ref{fig:tau_RR_super} in Appendix~\ref{asec:stresstensor}, show a peculiar behavior:
the polar angular component displays a significant difference between the sub- and the supercritical cases, with a stress boundary layer building
up in the region around the upper bubble pole, but either relaxing completely along the upper hemisphere, or persisting far beyond the bubble equator and with higher values. The radial normal stress component does not show much difference between the two cases and is localized around the bubble's lower pole.
While this is obviously of interest, it does not help much to explain the underlying mechanism causing such a different behavior.
To this end, the interplay between the local flow and the orientation-deformation state of the polymer molecules needs to be understood. For this purpose, the conformation tensor is the most relevant quantity and, for this reason, the analysis of the local trace of the conformation tensor has indeed already been started in \cite{Pillapakkam2007}. The corresponding results from our present simulations, displayed in Figs.~\ref{fig:tr_C_sub} and \ref{fig:tr_C_super}, show a marked difference between the sub- and the supercritical states, as discussed in Section~\ref{subsec-conformation}. %
But since the trace of the conformation tensor represents the average elongation (length) of the polymer molecules, directional information is not included. To extract information on the polymer orientation, spectral information about the conformation tensor is required. More precisely, as explained in detail in Appendix~\ref{asec:conftens}, the
eigenvector belonging to the largest of the three (positive) eigenvalues gives the direction of the main polymer orientation, with the square root of the eigenvalue measuring the average component of the polymer end-to-end vectors in this direction.
This allows the polymer molecule's average orientation to be visualized
by plotting line elements in the direction of the eigenvectors, colored according to the eigenvalue, as displayed in Figs.~\ref{fig:ev_sub} and \ref{fig:ev_super}.
The simulations show that, during their passage through the upper pole region, the polymer molecules get oriented mainly in the circumferential direction.
While this observation about the polymer molecular conformation at the bubble's upper pole region was also already made in \cite{Noh1993}, the new and central finding is that the circumferential stretching and the ratio between the transport time of the polymer molecules along the bubble surface and their relaxation time scale is the key for understanding the rise velocity jump discontinuity. In fact, this molecular orientation-deformation state either relaxes before the molecules pass the equator, or it persists far downstream beyond the bubble's equator.
This strongly supports the hypothesis that the rise velocity jump discontinuity is related to the interplay of convection and relaxation and their characteristic time scales.
Furthermore, these results underline the crucial importance of the circumferential conformation tensor component.
\begin{figure}[b!]
\centering{
\includegraphics[width=0.7 \textwidth]{figure/KinematicEffect.png}
}
\caption{\small Illustration of the kinematic effect due to diverging (upper pole region) and converging (lower pole region, not shown) trajectories (red and blue) and the evolution of the circumferential distance (orange) $d(\theta)$ of particles moving along neighboring longitudinal trajectories.}
\label{fig:kinematic_effect}
\end{figure}
The reason for the polymer molecules to get oriented and stretched in the circumferential direction near the upper pole is a purely kinematic one. As illustrated in Fig.~\ref{fig:kinematic_effect}, trajectories going down along the upper bubble hemisphere, starting from initially nearby positions with a slight offset in the azimuthal angle, diverge until their maximum distance at the equator, before they converge again. This qualitative behavior, especially the divergence on the upper hemisphere and the resulting orientation and stretching of the polymer molecules in the circumferential direction, is independent of the bubble volume.
What is influenced by the bubble volume and, hence, by the (transient) rise velocity, is how far the stretched polymer molecules travel along the bubble contour before relaxation is complete. This, of course, decides on where the elastic energy stored in the elongated molecules will be returned (partly) to the fluid, and hence to the bubble: above or below the equator.
Ad (iii). With a velocity field given, the transport and reorientation-deformation of individual polymer molecules is described by the transport-evolution equation \eqref{eq:conform}. Moreover, under realistic symmetry assumptions, the transport-evolution equation for the circumferential component decouples from the remaining system, allowing for a thorough theoretical analysis in this case. Motivated by both experimental and numerical results on the velocity field close to the bubble's upper hemisphere, we employ the Hadamard-Rybczynski flow field around a fluid sphere translating in a liquid as the velocity field for use with the transport-evolution equation. The resulting differential equation is written as an ODE in the polar angle $\theta$ as the independent variable. From a backward solution analysis, it is rigorously shown that ${\rm De}_c =1$ is a critical Deborah number: For ${\rm De}_c <1$, the maximum values of $C_{\phi \phi}$ are bounded and moderate in value, while arbitrary large values for $C_{\phi \phi}$ are possible in the opposite case ${\rm De}_c >1$. Quantitatively, if ${\rm De}_c < 0.9$, say, $C_{\phi \phi}$ at the bubble's equator is bounded to values around 2, while values of 10 or higher are reached there in case of ${\rm De}_c >1$. This ODE analysis in fact shows that for ${\rm De}_c <1$, polymer relaxation is fast and completed above the bubble's equator, while relaxation is slower than convection and, thus, persists below the equator in case of ${\rm De}_c >1$.
It is worth noting that this ODE analysis does not contain any back-coupling of the polymer dynamics to the flow fields. The latter is accounted for in the numerical simulations and, there, causes a self-amplification.
To explain this feedback mechanism, we consider the transient
rise of a bubble:
If the transport of the polymer molecule is fast enough such that the relaxation persists until slightly below the equator, the resulting hoop stress accelerates the bubble. Then the subsequent polymer molecules unload their elastic energy a bit further below the bubble's equator, where the force component in the rise direction will be larger due to the different normal direction of the bubble surface. This further accelerates the bubble, and so on. This process amplifies itself until, if this state is reached, the relaxation persists even downstream from the bubble. Then, with increasing distance of the relaxing polymer molecules from the bubble, the upward pushing force vanishes. The mechanism as just explained is illustrated by the schematics in Fig.~\ref{fig:hoop-stress}.
\begin{figure}[t!]
\centering{
\includegraphics[width=0.96 \textwidth]{figure/hoop-stress-reduced.png}
}
\caption{\small Illustration of the kinematic orientation and stretching of the polymer molecules and the persistence of the relaxation phase for a supercritical bubble, leading to a resulting upward force induced by the hoop stress.}
\label{fig:hoop-stress}
\end{figure}
In order to conclude our investigation, let us first recall that the rise velocity jump discontinuity is usually explained as a consequence of either the appearance of a negative wake below the rising bubble, or the formation of a cusp at the bubble's trailing edge, or both features together. The published work on this topic, however, does not give a clear picture.
Even more, none of the existing papers provides a root cause for the existence of a critical bubble volume at which the rise velocity undergoes a true jump, i.e.\ an abrupt increase for only a tiny--almost infinitesimal--change in the bubble volume.
Some exception is the work by Fraggedakis et al.\ \cite{Fraggedakis2016}, which studies the rise velocity jump phenomenon by means of a dynamical systems approach. Employing an FEM-based ALE-type numerical method for axisymmetric two-phase flow in combination with a continuation method, they are able to detect a curve in the bubble diameter-rise velocity-plane, corresponding to a family of quasi-steady state solutions under the variation of the bubble diameter as a parameter. With this approach they show the existence of multiple steady states of different stability properties: in some range of parameters, three possible rise velocities for the same bubble diameter are found. Two stable ones, corresponding to the slow and the fast bubble, and an unstable rise at an intermediate velocity. These three solutions are placed on an S-shaped curve, where the turning points indicate a change in the stability behavior.
The authors conclude that an increase of the bubble diameter on the arc with low rise velocity ends at such a turning point, and that a further increase of the bubble diameter then leads to the higher rise velocity, i.e.\ to the point on the highest arc of the solution curve, since this is again a stable solution. In contrast to the literature prior to \cite{Fraggedakis2016}, this study provides a dynamical systems theoretical justification for the appearance of a jump. But it still does not provide a physical mechanism that explains the occurrence of arcs of solutions with different rise velocities.
With the present findings, this mechanism is now identified to be the self-amplifying bubble acceleration due to a vertical hoop stress component.
To confirm and test the proposed mechanism, it is important that it can explain the occurrence of the two stable, hypothetical solutions at different rise velocities for a certain parameter regime, as well as another hypothetical solution which is unstable. This is indeed possible: If a subcritical bubble, with a volume not too far below the critical volume, would by accelerated to an artificial, higher rise velocity by an external force, the relaxation of the polymer molecules, being oriented and stretched in the azimuthal direction, would take place below the bubble's equator, thus further accelerating the bubble. Then--in this thought experiment--if the artificial external force would fade, the bubble still could keep a higher velocity if the relaxation still continues below the equator. This proves the possible existence of a second stable solution with a larger rise velocity, although it will not occur in real experiments. But even the third, intermediate and unstable solution can be explained along the same line: if the bubble would be brought to an intermediate rise velocity at which the relaxation-induced effects of de- and acceleration are just balanced, the bubble could--in principle--also hold its velocity. Of course, any disturbance slightly changing the bubble velocity will induce either a fall back to the lower or an acceleration to the faster rise mode. Here, again, the self-amplifying character of the mechanism is necessary.
Consequently, the mechanism which we propose as the root cause for the bubble rise velocity jump discontinuity, also explains the stability behavior of the dynamical system composed of a buoyancy-driven bubble rising in a viscoelastic liquid.
Furthermore, the formation of a ring of circumferentially oriented and
stretched polymer molecules at the upper pole region can also help to explain the other characteristic features, i.e.\ the cusped shape and the negative wake of supercritical bubbles.
Indeed, since the supercritical case is correlated with the presence of significant hoop stress below the bubble's equator, this hoop stress
not only transmits an upward force to the bubble, but also pushes the bubble surface towards the axis of symmetry. The fraction of the horizontal force component increases, the more the surface normal turns into a horizontal direction. This is another type of self-amplification, which can lead to a surface normal field that is almost horizontal, uniformly in all azimuthal directions. This produces the shape of a cusp and, thus, shows that the formation of a cusped shape is another consequence of the unloading of circumferentially stretched polymer molecules below the bubble equator. In particular,
the cusped shape is not the cause of the velocity jump discontinuity, but rather another result of the same underlying mechanism.
To see the relation with the negative wake of supercritical bubbles, note that, as visible in Fig.~\ref{fig:ev_super}~(e), the ring of circumferentially oriented, stretched polymer molecules persists even until below the bubble, where it still contracts and, thus, pushes the liquid towards the axis of symmetry. A part of the liquid then moves upwards, the rest moves downwards, creating a negative wake.
Note that this explanation requires the found hoop stress generated by the circumferential orientation of the polymer molecules.
\begin{figure}[t!]
\centering{
\includegraphics[width=0.68 \textwidth]{figure/Illustration-comparison.png}
}
\caption{\small Illustration of the different evolution of the polymer deformation state for sub- and supercritical bubble volume.}
\label{fig:mechanism_comparison}
\end{figure}
The experimental, numerical and theoretical investigations presented above show that the states of steady motion of sub- and a supercritical bubbles in a viscoelastic polymer solution differ concerning the hoop stress set free by the relaxation of polymer molecules oriented and elongated in the flow along the bubble contour. Starting from a stagnation state at the upper pole of the bubble, the molecules are stretched in the polar and azimuthal directions of a biaxial elongational flow along the bubble. The molecules tend to relax in order to maximize the entropy of their state of conformation. The related relaxation time competes with the transport time down the bubble contour. Subcritical bubbles exhibit slow molecule transport, so that the stress relaxation takes place along the upstream hemisphere of the bubble. In contrast, the polymer molecules are transported fast along supercritical bubbles, so that the stress relaxation takes place along the downstream part of the bubble contour. In the two cases, the hoop stress set free hinders or enhances the motion, respectively. The two different scenarios are illustrated in Fig.~\ref{fig:mechanism_comparison}. Let us finally note that the relevance of the circumferential orientation of the polymer molecules found also explains the need for 3D, or at least axisymmetric, simulations to capture the rise velocity jump discontinuity.
\section*{Acknowledgements}
This work was partly supported from the German Research Foundation (DFG) -- Project-ID 265191195 -- SFB 1194. Funding was furthermore provided by the Austrian Science Fund (FWF) under project number P 17624-N07. This support is gratefully acknowledged.
Support of the PIV experiments by Prof. J. Woisetschl{\"a}ger at TU Graz is gratefully acknowledged.
Calculations for this research were conducted on the Lichtenberg high performance computer of the Technische Universit{\"a}t Darmstadt.
\bibliographystyle{unsrturl}
|
\section{Introduction}
In this paper we introduce and study normalizers and centralizers of subnormal subsystems of saturated fusion systems. In doing so, we generalize and unify some concepts which were known before: Aschbacher defined centralizers of normal subsystems as well as normalizers and centralizers of components of fusion systems (cf. \cite[Chapter~6]{Aschbacher:2011} and \cite[Sections~2.1 and 2.2]{AschbacherFSCT}). The latter construction was generalized by Oliver \cite{Oliver:NormalizerComponents} to define normalizers of products of components. All these concepts are special cases of the normalizers and centralizers introduced in this paper. However, the existence of centralizers of normal subsystems (and the concrete construction of such centralizers given in \cite{Henke:2018}) is used in the proof of our results.
\smallskip
Our approach uses the fact that there is an (essentially unique) regular locality associated to every saturated fusion system. Here regular localities are group-like structures introduced by Chermak. Recently, Chermak and the author of this paper \cite{Chermak/Henke} established a kind of dictionary which allows us to translate between concepts in saturated fusion systems and in regular localities. In particular, subnormal subsystems of saturated fusion systems correspond to partial subnormal subgroups of regular localities.
\smallskip
In the present paper, we investigate first ``regular normalizers'' and centralizers of partial subnormal subgroups of regular localities. Only afterwards, this is used to define and study normalizers and centralizers of subnormal subsystems. Along the way, it becomes automatically transparent how one can translate between regular localities and fusion systems with regards to the concepts we will be concerned with. Generally, some flexibility in moving between fusion systems and localities may lead to simplifications in Aschbacher's program to give a new treatment of the classification of finite simple groups. Before we state our results on regular localities, we collect some definitions, notation and background results. The reader who is only interested in the fusion-theoretic results is invited to go immediately to Subsection~\ref{SS:FusionResults}.
\subsection{Some background} Throughout this paper we write homomorphisms on the right hand side of the argument. The reader is referred e.g. to \cite[Section~3]{Chermak/Henke} for a detailed summary of the required definitions and results concerning partial groups and localities. As in this reference and in Chermak's original papers \cite{Chermak:2013,Chermak:2015}, we write $\W(\L)$ for the set of words in a set $\L$, $\emptyset$ for the empty word, and $v_1\circ v_2\circ\cdots\circ v_n$ for the concatenation of words $v_1,\dots,v_n\in\W(\L)$. Recall that a partial group consists of a set $\L$, a ``product'' $\Pi\colon \D\rightarrow \L$ defined only on the words in some subset $\D$ of $\W(\L)$, and an involutory bijection $\L\rightarrow \L,f\mapsto f^{-1}$ called an ``inversion map'', subject to certain group-like axioms. If $\L$ is a partial group with product $\Pi\colon \D\rightarrow \L$ then, given $(x_1,\dots,x_n)\in\D$, we write also $x_1x_2\cdots x_n$ for $\Pi(x_1,\dots,x_n)$. Moreover, we have the following important definitions.
\begin{itemize}
\item A subset $\H\subseteq\L$ is called a \emph{partial subgroup} of $\L$ if $\Pi(w)\in\H$ for every $w\in\W(\H)\cap \D$ and $h^{-1}\in\H$ for every $h\in\H$.
\item A partial subgroup $\H$ of $\L$ is called a \emph{subgroup} of $\L$ if $\W(\H)\subseteq\D$. Observe that every subgroup of $\L$ forms an actual group. We call a subgroup $\H$ of $\L$ a $p$-subgroup if it is a $p$-group.
\item Given $f\in\L$, we write $\D(f):=\{x\in\L\colon (f^{-1},x,f)\in\D\}$ for the set of elements $x\in\L$ for which the \emph{conjugate} $x^f:=\Pi(f^{-1},x,f)$ is defined. This gives us a conjugation map $c_f\colon\D(f)\rightarrow \L$ defined by $x\mapsto x^f$.
\item We say that a partial subgroup $\N$ of $\L$ is a \emph{partial normal subgroup} of $\L$ (and write $\N\unlhd\L$) if $x^f\in\N$ for every $f\in\L$ and every $x\in \D(f)\cap \N$.
\item We call a partial subgroup $\H$ of $\L$ a \emph{partial subnormal subgroup} of $\L$ (and write $\H\subn\L$) if there is a sequence $\H_0,\H_1,\dots,\H_k$ of partial subgroups of $\L$ such that
\[\H=\H_0\unlhd\H_1\unlhd\cdots\unlhd \H_{k-1}\unlhd \H_k=\L.\]
\item If $f\in\L$ and $\X\subseteq \D(f)$ set $\X^f:=\{x^f\colon x\in\X\}$. For every $\X\subseteq\L$, call
\[N_\L(\X):=\{f\in\L\colon \X\subseteq\D(f)\mbox{ and }\X^f=\X\}\]
the \emph{normalizer} of $\X$ in $\L$.
\item For $\X\subseteq \L$, we call
\[C_\L(\X):=\{f\in\L\colon \X\subseteq\D(f),\;x^f=x\mbox{ for all }x\in\X\}\]
the \emph{centralizer} of $\X$ in $\L$.
\item For any two subsets $\X,\Y\subseteq\L$, their \emph{product} can be naturally defined by
\[\X\Y:=\{\Pi(x,y)\colon x\in\X,\;y\in\Y,\;(x,y)\in\D\}.\]
\end{itemize}
A \emph{locality} is a triple $(\L,\Delta,S)$ such that $\L$ is a partial group, $S$ is maximal among the $p$-subgroups of $\L$, and $\Delta$ is a set of subgroups of $S$ subject to certain axioms, which imply in particular that $N_\L(P)$ is a subgroup of $\L$ for every $P\in\Delta$. If $R\leq S$ and $f\in\L$ with $R\subseteq\D(f)$ and $R^f\subseteq S$, then $R^f$ is a subgroup of $S$ and $c_f|_R$ turns out to be an injective group homomorphism from $R$ to $R^f$. Moreover, for every $f\in\L$, the subset
\[S_f:=\{x\in S\colon x\in\D(f),\;x^f\in S\}\]
is a subgroup of $S$ and indeed an element of $\Delta$. In particular, $c_f|_{S_f}$ is an injective group homomorphism $S_f\rightarrow S$.
\begin{definition}
If $\H$ is a partial subgroup of $\L$ and $T=S\cap\H$, then $\F_T(\H)$ denotes the fusion system over $T$ generated by the maps of the form $c_h\colon R\rightarrow R^h$ where $h\in\H$ and $R\leq T$ with $R^h\leq T$. Equivalently, $\F_T(\H)$ is generated by all the maps $c_h|_{S_h\cap \H}$ with $h\in\H$.
\end{definition}
Notice that we have in particular $\F_S(\L)$ defined for every locality $(\L,\Delta,S)$. We say that $(\L,\Delta,S)$ is a locality \emph{over} $\F$ if $\F=\F_S(\L)$.
\smallskip
Below we will consider situations in which we are given a partial subgroup $\X$ of $\L$ such that, for some $p$-subgroup $S_0$ and some set $\Gamma$ of subgroups of $S_0$, the triple $(\X,\Gamma,S_0)$ is a locality. Notice, that we may then also form the fusion system $\F_{S_0}(\X)$ according to the definition above, even if $S_0$ is not a subgroup of $S$.
\smallskip
Chermak \cite{ChermakIII} defined a locality $(\L,\Delta,S)$ over a saturated fusion system $\F$ to be \emph{regular} if the set $\Delta$ equals a certain set $\delta(\F)$ of subgroups of $S$ and $N_\L(P)$ is a group of characteristic $p$ for every $P\in\Delta$. Here a finite group $G$ is \emph{of characteristic $p$} if $C_G(O_p(G))\leq O_p(G)$.
\smallskip
We will list now some theorems which will be crucial in our approach. The first one follows essentially from the existence and uniqueness of centric linking systems. For a detailed proof see e.g. \cite[Theorem~A]{Henke:2015} and \cite[Lemma~10.4]{Henke:Regular}.
\begin{bth}\label{bT:0}
Let $\F$ be a saturated fusion system. Then there exists a regular locality $(\L,\Delta,S)$ over $\F$, which is unique up to an isomorphism which restricts to the identity on $S$.
\end{bth}
The next theorem was first observed by Chermak \cite[Theorem~C]{ChermakIII}. A proof can also be found in \cite[Theorem~10.16(e), Corollary~10.19]{Henke:Regular}.
\begin{bth}\label{bT:1}
Let $(\L,\Delta,S)$ be a regular locality and $\H\subn\L$. Then $\E:=\F_{S\cap\H}(\H)$ is saturated and $(\H,\delta(\E),S\cap \H)$ is a regular locality over $\E$. If $\H\unlhd\L$, then $C_\L(\H)\unlhd\L$.
\end{bth}
As a consequence of the above theorem, for every partial normal subgroup $\H$ of $\L$, the centralizer $C_\L(\H)$ can be given the structure of a regular locality.
\smallskip
Building on Background Theorem~\ref{bT:1}, it is easy to define a notion of components of regular localities. If $(\L,\Delta,S)$ is regular, then we write $\Comp(\L)$ for the set of components of $\L$. The product $E(\L)$ of components of $\L$ forms a partial normal subgroups of $\L$. Similarly, there is a natural notion of components of fusion systems introduced by Aschbacher \cite[Chapter~9]{Aschbacher:2011} such that there is a normal subsystem $E(\F)$ of $\F$ which is a central product of the components. Again, we write $\Comp(\F)$ for the set of components of $\F$.
\begin{bth}[{\cite[Proposition~7.1(a), Lemma~7.3(a),(b)]{Chermak/Henke}}]\label{bT:2}
The map \[\hat{\Psi}\colon \{\H\colon\H\subn\L\}\rightarrow\{\E\colon\E\subn\F\},\H\mapsto \F_{S\cap\H}(\H)\] is well-defined and a bijection. It restricts to a bijection from the set of partial normal subgroups of $\L$ to the set of normal subsystems of $\F$, and to a bijection from $\Comp(\L)$ to $\Comp(\F)$. Moreover,
\[\hat{\Psi}(E(\L))=E(\F).\]
\end{bth}
\subsection{Normalizers and centralizers in regular localities}
Suppose $(\L,\Delta,S)$ is a regular locality and $\H\subn\L$. The normalizer $N_\L(\H)$ as defined above is not the right subset to consider if we want to find a partial subgroup of $\L$ in which $\H$ is normal. Indeed, $\H$ is in many cases not even contained in $N_\L(\H)$. We will however find that there is another partial subgroup $\bN_\L(\H)$ of $\L$ which one can also reasonably think of as a ``normalizer of $\H$''.
\smallskip
One can observe relatively easily that $\tT:=S\cap E(\L)$ is an element of $\Delta$. In particular, $G:=N_\L(\tT)$ is a finite group of characteristic $p$. We show in Lemma~\ref{L:BasisNLtT} that $G$ acts on $\L$ via conjugation. More precisely, for every $g\in G$, we have $\L\subseteq\D(g)$ (i.e. conjugation by $g$ is defined on all elements of $\L$), and this gives an action of the group $G$ on the set $\L$. It turns also out that the conjugation map $c_g\colon \L\rightarrow \L$ is for every element $g\in G$ an automorphism of $\L$. As $G$ acts on $\L$, the set-wise stabilizer
\[N_G(\X):=\{g\in G\colon \X^g=\X\}\]
of $\X$ in the group $G$ is for every subset $\X$ of $\L$ a subgroup of $G$. Set $N_S(\X):=S\cap N_G(\X)=S\cap N_\L(\X)$.
\smallskip
If we consider now a partial subnormal subgroup $\H$ of $\L$ as above then, similarly as in finite groups, every component of $\L$ is either contained in $\H$ or in the centralizer of $\H$. Hence, it is reasonable to think that the product $E(\L)$ of components should be contained in $\bN_\L(\H)$. There is moreover a Frattini Argument for localities (cf. \cite[Corollary~3.11]{Chermak:2015}) which implies that
\[\L=E(\L)N_\L(\tT)=E(\L)G.\]
Therefore, the following definition seems natural.
\begin{defn}
Let $(\L,\Delta,S)$ be a regular locality over $\F$, set $\tT:=E(\L)\cap S$ and $G:=N_\L(\tT)$. For every partial subnormal subgroup $\H$ of $\L$, the subset
\[\bN_\L(\H):=E(\L)N_G(\H)\]
is called the \emph{regular normalizer} of $\H$ in $\L$.
\end{defn}
If $\H\unlhd\L$, then $\H^g=\H$ for every $g\in G$. So we have in this case that $N_G(\H)=G$ and thus $\bN_\L(\H)=E(\L)G=\L$ as one would expect. In the general case, the properties of $\bN_\L(\H)$ are summarized in the following theorem.
\begin{Th}\label{T:MainbNLH}
Let $(\L,\Delta,S)$ be a regular locality and $\H\subn\L$. Then the regular normalizer $\bN_\L(\H)$ is a partial subgroup of $\L$. Moreover, setting
\[\tT:=E(\L)\cap S,\;G:=N_\L(\tT),\;T:=\H\cap S,\mbox{ and }\E:=\F_T(\H),\]
the following hold for every $S_0\in\Syl_p(N_G(\H))$:
\begin{itemize}
\item [(a)] There exists a set $\Gamma$ of subgroups of $S_0$ such that $(\bN_\L(\H),\Gamma,S_0)$ is a regular locality. In particular, the fusion system $\F_0:=\F_{S_0}(\bN_\L(\H))$ is well-defined and saturated.
\item [(b)] $\H\unlhd\bN_\L(\H)$. If $S_0$ is chosen such that $T\leq S_0$, then $T=\H\cap S_0$ and $\E\unlhd\F_0$.
\item [(c)] $\Comp(\bN_\L(\H))=\Comp(\L)$, $E(\bN_\L(\H))=E(\L)$ and $\tT=E(\L)\cap S_0$. In particular, $\Comp(\F)=\Comp(\F_0)$ and $E(\F)=E(\F_0)$.
\item [(d)] $S\cap \bN_\L(\H)=N_S(\H)$.
\item [(e)] Suppose $\X$ is a partial subgroup of $\L$ such that $\H\unlhd\X$ and $(\X,\Gamma_\X,S_\X)$ is a locality for some $p$-subgroup $S_\X$ of $\X$ with $T\leq S_\X$, and some set $\Gamma_\X$ of subgroups of $S_\X$. Then $\X\subseteq \bN_\L(\H)$.
\item [(f)] $C_\L(T)\subseteq \bN_\L(\H)$ and
\[C_\L(\H)=C_{\bN_\L(\H)}(\H)\unlhd \bN_\L(\H).\]
\end{itemize}
\end{Th}
With the notation as above, Theorem~\ref{T:MainbNLH}(a) says that $\bN_\L(\H)$ gives rise to a saturated fusion system, which is however not necessarily a subsystem of $\F:=\F_S(\L)$ (as $S_0$ might not be a subgroup of $S$). It follows from Background Theorem~\ref{bT:1} and parts (a),(b),(f) of Theorem~\ref{T:MainbNLH} that $C_\L(\H)$ can be given the structure of a regular locality. Thus, $C_\L(\H)$ gives also rise to a saturated fusion system, which again is not necessarily a subsystem of $\F$. One can see here that it might be an advantage to work with regular localities rather than with fusion systems in some contexts. Nevertheless, we want to have some more fusion-theoretic results in place as well. We therefore define normalizers and centralizers of subnormal subsystems of saturated fusion systems now.
\begin{defn}\label{D:NFECFE}
Let $\F$ be a saturated fusion system over $S$ and let $\E$ be a subnormal subsystem of $\F$ over $T\leq S$. Fix a regular locality $(\L,\Delta,S)$ over $\F$ and $\H\subn\L$ with $T=S\cap \H$ and $\E=\F_T(\H)$ (which exist by Background Theorems~\ref{bT:0} and \ref{bT:2}). We call then
\[N_\F(\E):=\F_{N_S(\H)}(\bN_\L(\H))\]
the \emph{normalizer of $\E$ in $\F$} and
\[C_\F(\E):=\F_{C_S(\H)}(C_\L(\H)).\]
the \emph{centralizer of $\E$ in $\F$}.
\end{defn}
Recall that a regular locality $(\L,\Delta,S)$ over $\F$ is by Background Theorem~\ref{bT:0} essentially unique. Moreover, for a given regular locality $(\L,\Delta,S)$ over $\F$, a partial subnormal subgroup $\H$ as above is also unique by Background Theorem~\ref{bT:2}. This can be used to show that $N_\F(\E)$ and $C_\F(\E)$ do not depend on the choice of $(\L,\Delta,S)$ and $\H$. Alternatively, we prove some characterizations of $N_\F(\E)$ and $C_\F(\E)$ in terms of fusion systems, which imply that $N_\F(\E)$ and $C_\F(\E)$ are well-defined.
\subsection{Normalizers and centralizers of subnormal subsystems}\label{SS:FusionResults}
We will now investigate the properties of $N_\F(\E)$ and $C_\F(\E)$ as defined above. The reader is referred to \cite[Chapter~I]{Aschbacher/Kessar/Oliver:2011} for an introduction to the theory of fusion systems. Throughout this subsection we assume the following hypothesis:
\smallskip
\noindent \textbf{$\F$ is a saturated fusion system over $S$ and $\E$ is a subnormal subsystem of $\F$ over $T\leq S$.}
\smallskip
We will need the following definitions.
\vspace{0.5 cm}
\begin{defn}~
\begin{itemize}
\item If $Q,R$ are $p$-groups, $\mD$ is a fusion system over $R$ and $\phi\colon R\rightarrow Q$ is an injective group homomorphism, then write $\mD^\phi$ for the fusion system over $R\phi$ with
\[\Hom_{\mD^\phi}(P\phi,Q\phi)=\{\phi^{-1}\psi\phi\colon \psi\in\Hom_{\mD}(P,Q)\}\mbox{ for all }P,Q\leq R.\]
\item By $\Aut(\F)$ we denote the set of automorphisms $\alpha$ of $S$ with $\F^\alpha=\F$.
\item Set $N_S(\E):=\{s\in N_S(T)\colon c_s|_T\in\Aut(\E)\}$.
\item Every element of $\E^\F:=\{\E^\phi\colon \phi\in\Hom_\F(T,S)\}$ is called an \emph{$\F$-conjugate} of $\E$.
\item The subsystem $\E$ is said to be \emph{fully $\F$-normalized}, if
\[|N_S(\E)|\geq |N_S(\tE)|\mbox{ for all }\tE\in\E^\F.\]
\item Similarly, $\E$ is called \emph{fully $\F$-centralized} if
\[|C_S(\E)|\geq |C_S(\tE)|\mbox{ for all }\tE\in\E^\F.\]
\end{itemize}
\end{defn}
It is shown in Lemma~\ref{L:ConjugatesSubnormal} below that every $\F$-conjugate of $\E$ is again a subnormal subsystem of $\F$. Note that there is always a fully $\F$-normalized $\F$-conjugate and a fully $\F$-centralized $\F$-conjugate of $\E$.
\smallskip
If $\F$ is constrained, then $\F$ has a model, i.e. a group of characteristic $p$ realizing $\F$ (cf. \cite[Theorem~III.5.10]{Aschbacher/Kessar/Oliver:2011}). Moreover, if $\F$ is constrained and $G$ is a model for $\F$, then it is shown in \cite[Lemma~2.9]{Chermak/Henke} that there exists a unique subnormal subgroup $H$ of $G$ with $T=H\cap S$ and $\E=\F_T(H)$. This is used to formulate the following proposition. It gives a characterization of $N_\F(\E)$, which one could take alternatively as a definition of this subsystem.
\begin{Th}\label{T:FusionNormalizerGenerate}~
\begin{itemize}
\item [(a)] $N_\F(\E)$ is a subsystem of $\F$ over $N_S(\E)$.
\item [(b)] Fix $T_0,\tT\leq S$ such that
\[E(\E)\mbox{ is a fusion system over $T_0$ and $E(\F)$ is a fusion system over $\tT$.}\]
Then $N_\F(\tT)$ is constrained and $N_\E(T_0)$ is subnormal in $N_\F(\tT)$.
\item [(c)] Choose a model $G$ for $N_\F(\tT)$ and $H\subn G$ with $H\cap S=T$ and $\F_T(H)=N_\E(T_0)$. Then $N_S(H)=N_S(\E)$ and
\[N_\F(\E)=\<\,E(\F)N_S(\E)\, ,\,\F_{N_S(H)}(N_G(H))\,\>.\]
\end{itemize}
\end{Th}
For the definition of $E(\F)N_S(\E):=(E(\F)N_S(\E))_\F$, which is used in part (c) of the theorem above, the reader is referred to Definition~\ref{D:ProductER}. Some important properties of $N_\F(\E)$ are summarized in the following theorem.
\begin{Th}\label{T:FusionNormalizer}
\begin{itemize}
\item [(a)] $\E$ is a subsystem of $N_\F(\E)$, which is $N_\F(\E)$-invariant.
\item [(b)] Every saturated subsystem $\mD$ of $\F$ with $\E\unlhd\mD$ is contained in $N_\F(\E)$.
\item [(c)] If $\E$ is fully $\F$-normalized, then $N_\F(\E)$ is saturated, $\E$ is normal in $N_\F(\E)$, and
\[E(N_\F(\E))=E(\F).\]
\item [(d)] If $X,Y\leq N_S(\E)$ and $\alpha\in\Hom_\F(X,Y)$ with $T\leq X\cap Y$ and $\phi|_T\in\Aut(\E)$, then $\phi$ is a morphism in $N_\F(\E)$. In particular, $C_\F(T)=C_{N_\F(\E)}(T)$.
\item [(e)] If $\F$ is constrained, $G$ is a model for $\F$ and $H\subn G$ with $T=S\cap H$ and $\F_T(H)=\E$, then $N_\F(\E)=\F_{N_S(H)}(N_G(H))$.
\end{itemize}
\end{Th}
If $\E$ is fully $\F$-normalized, then notice that $N_\F(\E)$ is by parts (b) and (c) of Theorem~\ref{T:FusionNormalizer} the largest saturated subsystem of $\F$ containing $\E$ as a normal subsystem. In Lemma~\ref{L:EfullyNormalized} we give some equivalent conditions for $\E$ to be fully $\F$-normalized. In particular, given a regular locality $(\L,\Delta,S)$ and $\H\subn\L$ with $\H\cap S=T$ and $\F_S(\H)=\E$, we show that $N_S(\E)=N_S(\H)$; moreover, setting $G=N_\L(E(\L)\cap S)$ as before, $\E$ is fully $\F$-normalized if and only if $N_S(\E)=N_S(\H)$ is a Sylow $p$-subgroup of $N_G(\H)$.
\smallskip
When proving Theorem~\ref{T:FusionNormalizer}, we obtain along the way a nice notion of a product subsystem $\E R=(\E R)_\F$ for every subgroup $R$ of $N_S(\E)$. The existence of such a product was known before in the case that $\E$ is normal (cf. \cite[Chapter~8]{Aschbacher:2011} or \cite{Henke:2013}). For arbitrary $\E$, a precise description of the product subsystem $\E R=(\E R)_\F$ is given in Definition~\ref{D:ProductER}. We have the following theorem.
\begin{Th}\label{T:MainER}
Let $R\leq N_S(\E)$. Then the following hold.
\begin{itemize}
\item [(a)] The subsystem $\E R=(\E R)_\F$ is saturated and the unique saturated subsystem $\mD$ of $\F$ over $TR$ with $O^p(\mD)=O^p(\E)$. Moreover, $\E\unlhd \E R$.
\item [(b)] If $(\L,\Delta,S)$ is a regular locality over $\F$ and $\H\subn\L$ with $T=S\cap\H$ and $\F_T(\H)=\E$, then $\E R=\F_{TR}(\H R)$.
\end{itemize}
\end{Th}
With the above results about the normalizer $N_\F(\E)$ in place, it is also relatively easy to study the centralizer $C_\F(\E)$. First of all, we can show that there is a reasonable notion of a centralizer of $\E$ in $S$.
\begin{Th}\label{T:CentralizerEinS}
The set $\{R\leq S\colon \E\subseteq C_\F(R)\}$ has (with respect to inclusion) a unique maximal member $C_S(\E)$. Moreover, $C_\F(\E)$ is a fusion system over $C_S(\E)$.
\end{Th}
Next we want to list some of the properties of the centralizer of $\E$ in $\F$. For that we will use the following definition.
\begin{defn}\label{D:CommuteSubsystems}
Given two saturated subsystems $\F_1$ and $\F_2$ of $\F$ over $S_1$ and $S_2$ respectively, we say that $\F_1$ commutes with $\F_2$, if $\F_i\subseteq C_\F(S_{3-i})$ for each $i=1,2$.
\end{defn}
If $(\L,\Delta,S)$ is a regular locality over $\F$ and $\H\subn\L$ with $S\cap\H=T$ and $\E=\F_T(\H)$, then the subgroup $C_S(\E)$ from above is actually equal to $C_S(\H)$. Moreover, we show in Lemma~\ref{L:EFullyCentralized} that $\E$ is fully centralized if and only if $C_S(\H)$ is a maximal $p$-subgroup of $C_\L(\H)$.
\smallskip
Centralizers of normal subsystems were defined by Aschbacher \cite[Chapter~6]{Aschbacher:2011}. We show that the centralizer $C_\F(\E)$ defined above coincides with the centralizer subsystem defined by Aschbacher if $\E$ is normal in $\F$. An alternative construction of centralizers of normal subsystems was given in \cite{Henke:2018}. It allows us to give an easy description of $C_\F(\E)$ if $\E$ is fully centralized (cf. Theorem~\ref{T:FusionCentralizer}(c) below).
\begin{Th}\label{T:FusionCentralizer}
\begin{itemize}
\item [(a)] $C_\F(\E)$ is a subsystem of $N_\F(\E)$ over $C_S(\E)$. Moreover, $C_\F(\E)$ commutes with $\E$, and every saturated subsystem of $\F$ which commutes with $\E$ is contained in $C_\F(\E)$.
\item [(b)] $C_\F(\E)$ is weakly $N_\F(\E)$-invariant in the sense of Definition~\ref{D:WeaklyInvariant} below. If $\E$ is fully $\F$-normalized, then $C_\F(\E)=C_{N_\F(\E)}(\E)$ is normal in $N_\F(\E)$.
\item [(c)] Suppose $\E$ is fully centralized. Then $C_\F(\E)$ is saturated and thus the largest saturated subsystem of $\F$ that commutes with $\E$. Moreover,
\[C_\F(\E)=\<O^p(\Aut_{C_\F(T)}(P))\colon P\leq C_S(\E)\>_{C_S(\E)}.\]
\end{itemize}
\end{Th}
For more information on the centralizer $C_\F(\E)$, the reader is referred to Subsection~\ref{SS:Centralizer}. It seems difficult to obtain a neat description of this subsystem if $\E$ is not fully centralized. However, the subsystem $C_\F(\E)$ is uniquely determined by the properties stated in Theorem~\ref{T:FusionCentralizer}(c) and Lemma~\ref{L:MoveToFullyCentralized}.
\section{Preliminaries}
\textbf{Throughout this section let $\F$ be a fusion system over a $p$-group $S$.}
\smallskip
We adapt the terminology and notation from \cite[Chapter~I]{Aschbacher/Kessar/Oliver:2011}, except that we write homomorphisms on the right hand side. We use moreover the following notation.
\begin{notation}
\begin{itemize}
\item Given $T\leq S$, we write $\F|_T$ for the full subcategory of $\F$ whose objects are all the subgroups of $T$.
\item If $T$ is any $p$-group and $\Phi$ a set of injective group homomorphisms between subgroups of $T$, then $\<\Phi\>_T$ denotes the fusion system over $T$ generated by $\Phi$, i.e. the smallest fusion system over $T$ containing every morphism in $\Phi$.
\end{itemize}
\end{notation}
\subsection{Invariant subsystems} As in \cite[Definition~I.6.1]{Aschbacher/Kessar/Oliver:2011}, we say that a subsystem $\E$ of $\F$ over $T$ is \emph{$\F$-invariant} if $T$ is strongly closed, $\E^\alpha=\E$ for every $\alpha\in\Aut_\F(T)$ and the Frattini condition holds, i.e. for every $P\leq T$ and every $\phi\in\Hom_\F(P,T)$, there exists $\phi_0\in\Hom_\E(P,T)$ and $\alpha\in\Aut_\F(T)$ with $\phi=\phi_0\alpha$. There is also a different notion of $\F$-invariant subsystems in the literature. To distinguish the two, we introduce the following definition.
\begin{definition}\label{D:WeaklyInvariant}
We call a subsystem $\E$ of $\F$ over $T$ \emph{weakly $\F$-invariant}, if $T$ is strongly closed and, for all $P\leq Q\leq T$, all $\phi\in\Hom_\E(P,Q)$ and all $\alpha\in\Hom_\F(Q,T)$, we have
\[\phi^\alpha:=(\alpha|_P)^{-1}\phi\alpha \in\Hom_\E(P\alpha,Q\alpha).\]
\end{definition}
If $\F$ is saturated, it is shown in \cite[Proposition~I.6.4]{Aschbacher/Kessar/Oliver:2011} that $\E$ is $\F$-invariant if and only if it is weakly $\F$-invariant. In general, the two notions are different, but we have the following lemma.
\begin{lemma}\label{L:Finvariant}
Let $\F$ be a fusion system over $S$ and $\E$ a subsystem over $\F$ over $T\leq S$.
\begin{itemize}
\item [(a)] If $\E$ is $\F$-invariant, then $\E$ is weakly $\F$-invariant.
\item [(b)] Suppose $T$ is strongly $\F$-closed and $\E^\alpha=\E$ for every $\alpha\in \Aut_\F(T)$. Then $T$ is $\F$-invariant if and only if $\F|_T=\<\E,\Aut_\F(T)\>_T$.
\end{itemize}
\end{lemma}
\begin{proof}
This is shown in the proof of \cite[Proposition~I.6.4]{Aschbacher/Kessar/Oliver:2011} if $\F$ is saturated. The arguments do not use the assumption that $\F$ is saturated.
\end{proof}
\subsection{Commuting Subsystems}
As in Definition~\ref{D:CommuteSubsystems}, we say that two saturated subsystems $\F_i$ over $S_i$ ($i=1,2$) commute in $\F$ if $\F_i\subseteq C_\F(S_{3-i})$. Notice that this definition makes also sense if $\F$ is not saturated. If $\F_1$ and $\F_2$ commute then $[S_1,S_2]=1$ and thus $S_1S_2$ is a subgroup of $S$. We set in this case
\[\F_1*\F_2:=\<\phi\in\Hom_\F(P_1P_2,S_1S_2)\colon P_i\leq S_i,\;\phi|_{P_i}\in\Hom_{\F_i}(P_i,S_i)\mbox{ for each }i=1,2\>_{S_1S_2}.\]
The following lemma could be concluded from the results in \cite{Chermak/Henke}, but for simplicity we write out part of the argument.
\begin{lemma}\label{L:CentralProduct}
Let $\F_1$ and $\F_2$ be saturated subsystems of $\F$ over $S_1$ and $S_2$ respectively such that $\F_1$ and $\F_2$ commute in $\F$. Then $\mD:=\F_1*\F_2$ is a saturated subsystem of $\F$ with $\mD|_{S_i}=\F_i$, $\F_i\unlhd \mD$ and $\F_i\subseteq C_\F(S_{3-i})$ for each $i=1,2$.
\end{lemma}
\begin{proof}
By \cite[Proposition~3.3]{Henke:2018}, $\mD=\F_1*\F_2$ is a saturated subsystem of $\F$. Moreover, by \cite[Lemma~3.2]{Henke:2018}, $S_1\cap S_2\leq Z(\F_1)\cap Z(\F_2)$.
\smallskip
Let $P_i\leq S_i$ and $\phi\in\Hom_\F(P_1P_2,S_1S_2)$ with $\phi|_{P_i}\in\Hom_{\F_i}(P_i,S_i)$ for $i=1,2$. Fix now $i\in\{1,2\}$. Notice that $\hat{P}_i:=(P_1P_2)\cap S_i=P_iZ$, where $Z:=P_{3-i}\cap S_i\leq S_1\cap S_2\leq Z(\F_1)\cap Z(\F_2)$. As $Z\leq P_{3-i}\cap Z(\F_{3-i})$ and $\phi|_{P_{3-i}}$ is a morphism in $\F_{3-i}$, we have in particular $\phi|_Z=\id_Z$. As $Z\leq Z(\F_i)$ and $\phi|_{P_i}$ is a morphism in $\F_i$, it follows that $\phi|_{\hat{P}_i}$ is a morphism in $\F_i$. This property implies that $S_i$ is strongly closed in $\mD$ and
\[\mD|_{S_i}=\F_i.\]
In particular, $\F_i$ is $\F$-invariant. By assumption, $\F_i$ is saturated and thus weakly normal in $\mD$.
\smallskip
As $\F_i\subseteq C_\F(S_{3-i})$, every morphism $\phi_i\in\Hom_{\F_i}(P_i,S_i)$ extends to a morphism
\[\phi\in \Hom_\F(P_iS_{3-i},S_1S_2)\]
with $\phi|_{S_{3-i}}=\id_{S_{3-i}}$. By definition of $\mD$, $\phi$ is then a morphism in $\mD$. This shows that $\F_i\subseteq C_{\mD}(S_{3-i})$. Using this property for $P_i:=S_i$, it follows moreover that the pair $(\F_i,\mD)$ satisfies the extension condition for normal subsystems (cf. \cite[Definition~I.6.1]{Aschbacher/Kessar/Oliver:2011}). Hence, $\F_i$ is normal in $\F$.
\end{proof}
\subsection{Two lemmas on fusion systems}
\begin{lemma}\label{L:CentricRadicalIntersection}
If $\E$ is a subnormal subsystem of $\F$ over $T\leq S$ and $R\in\F^{cr}$, then $R\cap T\in\E^{cr}$.
\end{lemma}
\begin{proof}
If $\E$ is normal, this is shown in \cite[Lemma~1.20(d)]{AOV1} (or alternatively in \cite[Lemma~4.6]{Grazian/Henke}). The result follows thus by induction on the subnormal length of $\E$.
\end{proof}
\begin{lemma}\label{L:EE=EF}
Let $\E$ be a normal subsystem of $\F$ over $T\leq S$ with $E(\E)=E(\F)$. If $P\leq S$ with $P\cap T\in\delta(\E)$, then $P\in\delta(\F)$.
\end{lemma}
\begin{proof}
Since the set $\delta(\E)$ is overgroup-closed in $T$, we have $(PO_p(\F))\cap T\in\delta(\E)$. Hence, by \cite[Lemma~4.9(a)]{Grazian/Henke}, $PO_p(\F)\in\delta(\F)$. Using Background Theorem~\ref{bT:0}, it follows now from \cite[Lemma~10.6]{Henke:Regular} that $P\in\delta(\F)$.
\end{proof}
\subsection{Products of subnormal subsystems with $p$-groups}
If $\F$ is saturated and $\E$ is a normal subsystem of $\F$ over $T\leq S$, then a product subsystem $\E R$ was defined by Aschbacher \cite[Chapter~8]{Aschbacher:2011} for every subgroup $R\leq S$. A new construction of this product subsystems was given in \cite{Henke:2013}, which we generalize in the following definition.
\begin{definition}\label{D:ProductER}
Let $\E$ be a subsystem of $\F$ over $T\leq S$.
\begin{itemize}
\item For $P\leq S$ set
\[\Ac_{\F,\E}(P):=\<\phi\in\Aut_\F(P)\colon \phi\mbox{ $p^\prime$-element },\;[P,\phi]\leq P\cap T,\;\phi|_{P\cap T}\in\Aut_\E(P\cap T)\>.\]
\item If $\F$ is saturated and $\E$ is subnormal in $\F$, then for every $R\leq N_S(\E)$ set
\[\E R:=(\E R)_\F:=\<\Ac_{\F,\E}(P)\colon P\leq TR,\;P\cap T\in\E^{cr}\>_{TR}.\]
\end{itemize}
\end{definition}
\textbf{For the remainder of this section let $\F$ be a saturated fusion system over a $p$-group $S$.}
\begin{prop}\label{P:ERnormal}
Suppose $\E$ is a normal subsystem of $\F$ over $T\leq S$. Then $N_S(\E)=S$. Moreover, for every $R\leq S$, the subsystem $\E R=(\E R)_\F$ is saturated and
\[\E R=\<\Ac_{\F,\E}(P)\colon P\leq TR,\;P\cap T\in\E^c\>_{TR}.\]
Moreover, $\E\unlhd \E R$ and $\E R$ is the unique saturated subsystem $\mD$ of $\F$ over $TR$ with $O^p(\mD)=O^p(\E)$.
\end{prop}
\begin{proof}
As $\E\unlhd\F$, we have $T\unlhd S$ and $\Aut_S(T)\leq \Aut_\F(T)\leq \Aut(\E)$. Hence, $N_S(\E)=S$. Set now
\[\G:=\<\Ac_{\F,\E}(P)\colon P\leq TR,\;P\cap T\in\E^c\>_{TR}.\]
Then by \cite[Theorem~1]{Henke:2013}, $\G$ is saturated and the unique saturated subsystem $\mD$ of $\F$ over $TR$ with $O^p(\mD)=O^p(\E)$. By \cite[Lemma~2.33]{Chermak/Henke}, $\E\unlhd\G$. Hence, it remains to argue that $\G$ equals $\E R$. Clearly, $\E R$ is contained in $\G$. As $\G$ is saturated, it follows from Alperin's Fusion Theorem (cf. \cite[Theorem~I.3.6]{Aschbacher/Kessar/Oliver:2011}) that $\G$ is generated by all the automorphism groups $\Aut_{\G}(P)$, where $P\in \G^{cr}$ is fully $\G$-normalized. For every $P\in\G^{cr}$, Lemma~\ref{L:CentricRadicalIntersection} gives $P\cap T\in\E^{cr}$, and \cite[Lemma~4.7]{Henke:2013} implies therefore that $O^p(\Aut_\G(P))=\Ac_{\F,\E}(P)$ is contained in $\E R$. If $P\in\G^{cr}$ is fully $\G$-normalized, then $P$ is fully $\G$-automized and hence $\Aut_\G(P)=\Aut_{TR}(P)O^p(\Aut_\G(P))$ is contained in $\E R$. This shows that $\G\subseteq\E R$ and thus $\G=\E R$ as required.
\end{proof}
The following lemma can be seen as a Frattini argument for fusion systems.
\begin{lemma}\label{L:FusionSystemFrattini}
Let $\E$ a normal subsystem of $\F$ over $T\leq S$. Then
\[\F=\<\E S,N_\F(T)\>.\]
More precisely, for every subgroup $P\leq S$ and every morphism $\phi\in\Hom_\F(P,S)$, there exists $\psi\in\Hom_{\E P}(P,TP)$ and $\alpha\in\Hom_{N_\F(T)}(P\psi,S)$ such that $\phi=\psi\alpha$.
\end{lemma}
\begin{proof}
We use throughout that, for every subgroup $R\leq S$, the subsystem $\E R$ defined above coincides by Proposition~\ref{P:ERnormal} with the equally denoted subsystem defined in \cite{Henke:2013}. It was first shown by Aschbacher \cite[1.3.2]{AschbacherFSCT} that $\F=\<\E S,N_\F(T)\>$; a proof using localities is given in \cite[Lemma~4.10]{Grazian/Henke}. We use the property $\F=\<\E S,N_\F(T)\>$ now to prove the second part of the claim. Fix $P\leq S$ and $\phi\in\Hom_\F(P,S)$.
\smallskip
By the definition of $\E S=(\E S)_\F$, every morphism in $\E S$ is a product of morphisms each of which is either a restriction of a morphisms in $\Ac_{\F,\E}(Q)$ for some $Q\leq S$ with $Q\cap T\in\E^{cr}$, or a morphism induced by conjugation with an element of $S$. Notice that morphisms induced by conjugation with elements of $S$ are in particular morphisms in $N_\F(T)$. Since $\F=\<\E S,N_\F(T)\>$, it follows that $\phi$ can be factorized as the product of morphisms, each of which is either in $N_\F(T)$ or a restriction of an automorphism in $\Ac_{\F,\E}(Q)$ for some $Q\leq S$ with $Q\cap T\in\E^{cr}$. By induction on the length of such a product, we may assume that $\phi$ can be factorized as
\[\phi=\psi\alpha\gamma_0\]
where $\psi\in\Hom_{\E P}(P,TP)$, $\alpha\in\Hom_{N_\F(T)}(P\psi,S)$ and either
\begin{itemize}
\item [(i)] $\gamma_0$ is a morphism in $N_\F(T)$; or
\item [(ii)] $\gamma_0$ is a restriction of a morphism $\gamma\in\Ac_{\F,\E}(Q)$ for some $Q\leq S$ with $Q\cap T\in\E^{cr}$ and $P\psi\alpha\leq Q$.
\end{itemize}
In case (i), $\alpha\gamma_0$ is a morphism in $N_\F(T)$ and the claim holds. Hence, we may assume that (ii) holds. It follows from the definition of $\E P=(\E P)_\F$ that $TP=T(P\psi)$. Moreover, $\alpha$ extends to $\hat{\alpha}\in\Hom_\F(T(P\psi),T(P\psi\alpha))$ with $T\hat{\alpha}=T$. Replacing $Q$ by $Q\cap T(P\psi\alpha)$ and $\gamma$ by $\gamma|_{Q\cap T(P\psi\alpha)}$, we may assume that $Q\leq T(P\psi\alpha)$. Notice that
\[\delta:=\hat{\alpha}\gamma\hat{\alpha}^{-1}\in\Ac_{\F,\E}(Q\hat{\alpha}^{-1})\]
with
\[\phi=\psi\delta|_{P\psi}\hat{\alpha}.\]
By \cite[Lemma~2.6]{Chermak/Henke}, we have $(Q\hat{\alpha}^{-1})\cap T=(Q\cap T)\hat{\alpha}^{-1}\in\E^{cr}$. Moreover, $Q\hat{\alpha}^{-1}\leq T(P\psi)=TP$. Hence, $\delta$ is a morphism in $\E P$ and thus $\psi\delta|_{P\psi}$ is a morphism in $\E P$. As $\hat{\alpha}$ is a morphism in $N_\F(T)$ the assertion follows.
\end{proof}
\subsection{Some lemmas on localities}
\begin{lemma}\label{L:FrattiniSplitting}
Let $(\L,\Delta,S)$ be a locality, $\N\unlhd\L$ and $T:=\N\cap S$. Then for every $g\in\L$ there exist $n\in\N$ and $f\in N_\L(T)$ such that $(n,f)\in\D$, $g=nf$ and $S_g=S_{(n,f)}$.
\end{lemma}
\begin{proof}
This is \cite[Lemma~3.13]{Grazian/Henke}.
\end{proof}
\begin{lemma}\label{L:AutLFusion}
Let $(\L,\Delta,S)$ and $(\tL,\tDelta,\tS)$ be localities, and suppose $\alpha\colon \L\rightarrow \tL$ is an isomorphism of partial groups. Let moreover $\H$ be a partial subgroup of $\L$. Then the following hold:
\begin{itemize}
\item [(a)] If $S\alpha=\tS$, then $(S\cap \H)\alpha=\tS\cap (\H\alpha)$.
\item [(b)] If $(S\cap \H)\alpha=\tS\cap (\H\alpha)$, then $\alpha|_{S\cap \H}$ induces an isomorphism from $\F_{S\cap \H}(\H)$ to $\F_{\tS\cap (\H\alpha)}(\H\alpha)$.
\end{itemize}
\end{lemma}
\begin{proof}
As $\alpha$ is injective, $S\alpha=\tS$ implies $(S\cap \H)\alpha=(S\alpha)\cap (\H\alpha)=\tS\cap (\H\alpha)$. So assume from now on just that $(S\cap \H)\alpha=\tS\cap (\H\alpha)$. Let $h\in \H$. It is sufficient to show that
\begin{equation}\label{E:Conjbyalpha}
\alpha^{-1}(c_h|_{S_h\cap \H})\alpha=c_{h\alpha}|_{\tS_{h\alpha}\cap (\H\alpha)}
\end{equation}
To prove this let $x\in S_h\cap \H$, i.e. $x\in\D(h)$ and $x^h\in S$. As $\alpha$ is a homomorphism of partial groups, we have then $x\alpha\in\tilde{\D}(h\alpha)$ (where $\tilde{\D}$ denotes the domain of the product on $\tL$). Moreover, since $x^h\in S\cap \H$, we have
\[(x\alpha)^{h\alpha}=(x^h)\alpha\in (S\cap \H)\alpha=\tS\cap (\H\alpha).\]
In particular, as $x\alpha\in (S\cap \H)\alpha=\tS\cap (\H\alpha)$, it follows $x\alpha\in \tS_{h\alpha}$, which proves
\[(S_h\cap \H)\alpha\leq \tS_{h\alpha}\cap (\H\alpha).\]
It follows from \cite[Lemma~2.18]{Henke:2020} that $\alpha^{-1}$ is also an isomorphism of partial groups. Hence, using the property we just proved with $\alpha^{-1}$ in place of $\alpha$, we can conclude that $(\tS_{h\alpha}\cap (\H\alpha))\alpha^{-1}\leq S_h\cap \H$ and thus $\tS_{h\alpha}\cap (\H\alpha)=(S_h\cap \H)\alpha$. As $(x\alpha)^{h\alpha}=(x^h)\alpha=(x\alpha)(\alpha^{-1}c_h\alpha)$, property \eqref{E:Conjbyalpha} follows.
\end{proof}
\begin{corollary}\label{C:AutLFusion}
Let $(\L,\Delta,S)$ be a regular locality, $\H\subn\L$ and $\alpha\in\Aut(\L)$ with $(\H\cap S)\alpha\leq S$. Then $(\H\cap S)\alpha=(\H\alpha)\cap S$ and $\alpha$ induces an isomorphism from $\F_{S\cap \H}(\H)$ to $\F_{S\cap (\H\alpha)}(\H\alpha)$.
\end{corollary}
\begin{proof}
It is a consequence of Background Theorem~\ref{bT:1} and the definition of a locality that $S\cap \H$ is a maximal $p$-subgroup of $\H$. Notice that $\alpha\in\Aut(\L)$ induces an isomorphism from $\H$ to $\H\alpha$. Hence, $(S\cap \H)\alpha$ is a maximal $p$-subgroup of $\H\alpha$. As $(S\cap \H)\alpha\leq S\cap (\H\alpha)$ by assumption, it follows $(S\cap \H)\alpha=S\cap (\H\alpha)$. The assertion follows now from Lemma~\ref{L:AutLFusion}(b).
\end{proof}
A linking locality is a locality $(\L,\Delta,S)$ such that $\F_S(\L)$ is saturated, $\F_S(\L)^{cr}\subseteq\Delta$ and $N_\L(P)$ is of characteristic $p$ for every $P\in\Delta$. If $\F$ is saturated, then the set $\F^s$ of \emph{subcentric subgroups} is the largest possible object set of a linking locality. It consists of the subgroups $P\leq S$ such that, for every fully $\F$-normalized $\F$-conjugate $Q$ of $P$, the normalizer $N_\F(Q)$ is constrained (cf. \cite[Proposition~3.1]{Henke:2015}).
\smallskip
For every saturated fusion system $\F$, the set $\delta(\F)$ contains $\F^{cr}$ (see e.g. \cite[Lemma~10.4]{Henke:Regular}). Hence, every regular locality is a linking locality. For every linking locality $(\L,\Delta,S)$, there is the \emph{generalized Fitting subgroup} $F^*(\L)$ defined (cf. \cite[Definition~3]{Henke:Regular}).
\smallskip
We set $Z(\L)=C_\L(\L)$ for every partial group $\L$. A linking locality $(\L,\Delta,S)$ is called \emph{quasisimple} if $\L=O^p(\L)$ (i.e. $\L\neq \N S$ for every proper partial normal subgroup $\N$ of $\L$) and $\L/Z(\L)$ is simple (i.e. $\L/Z(\L)$ has precisely two partial normal subgroups).
\smallskip
Suppose now that $(\L,\Delta,S)$ is a regular locality. By Background Theorem~\ref{bT:1}, a partial subnormal subgroup $\K$ of $\L$ can be regarded as a regular locality. It is called a \emph{component} if it is quasisimple. As mentioned already in the introduction, the product $E(\L)$ of components of $\L$ forms a partial normal subgroup of $\L$. The \emph{generalized Fitting subgroup} $F^*(\L)$ equals $E(\L)O_p(\L)$ (cf. \cite[Theorem~4(a),(b),(c)]{Henke:Regular}).
\begin{lemma}\label{L:BasisNLtT}
Let $(\L,\Delta,S)$ be a regular locality. Set
\[\tT:=E(\L)\cap S\mbox{ and }T^*:=F^*(\L)\cap S.\]
Then $\tT$ and $T^*$ are elements of $\Delta$, and $G=N_\L(\tT)=N_\L(T^*)$. Moreover, $\L\subseteq \D(f)$ for every $f\in G$, and $G$ acts on $\L$ via conjugation.
\end{lemma}
\begin{proof}
By \cite[Lemma~11.9]{Henke:Regular}, we have $E(\L)\unlhd\L$ and $F^*(\L)=\tT O_p(\L)$. Therefore $G=N_\L(\tT)=N_\L(T^*)$. By \cite[Theorem~2]{Henke:Regular}, $F^*(\L)$ is centric in $\L$. Hence, by \cite[Lemma~10.11(c)]{Henke:Regular}, we have
\[\L\subseteq\D(f)\mbox{ and }c_f\in\Aut(\L)\]
for all $f\in N_\L(T^*)$. Notice that $x^{\One}=x$ for all $x\in\L$. Moreover, if $f,g\in N_\L(T^*)$, we have $u=(g^{-1},f^{-1},x,f,g)\in\D$ via $(S_x\cap T^*)^{fg}$. Hence, $x^{fg}=\Pi((fg)^{-1},x,(fg))=\Pi(u)=(x^f)^g$. Thus, $N_\L(T^*)$ acts on $\L$ via conjugation.
\end{proof}
\subsection{Some group theoretic lemmas}
Throughout this section let $G$ be a finite group.
\begin{lemma}\label{L:OpGinNormalizer}
If $H\subn G$, then $O_p(G)\leq N_G(O^p(H))$.
\end{lemma}
\begin{proof}
Let $G$ be a minimal counterexample. Notice that $H\neq G$ as otherwise the assertion clearly holds. As $H\subn G$, there exists thus $N\unlhd G$ with $H\subn N$ and $N\neq G$. Then $N$ is not a counterexample and hence $[O^p(H),O_p(N)]\leq O^p(H)$. Notice that
\[[O_p(G),H]\leq [O_p(G),N]\leq O_p(G)\cap N=O_p(N).\]
Hence, by a property of coprime action (cf. \cite[8.2.7(b)]{KS}), we have
\[[O_p(G),O^p(H)]=[O_p(G),O^p(H),O^p(H)]\leq [O_p(N),O^p(H)]\leq O^p(H)\]
and so the assertion holds.
\end{proof}
Recall from the introduction that $G$ is of \emph{characteristic $p$} if $C_G(O_p(G))\leq O_p(G)$.
\begin{lemma}\label{L:MS}
Let $G$ be a group of characteristic $p$. Then the following hold:
\begin{itemize}
\item[(a)] $N_G(P)$ is of characteristic $p$ for all non-trivial $p$-subgroups $P \leq G$.
\item[(b)] Every subnormal subgroup of $G$ is of characteristic $p$.
\item[(c)] If $H\leq G$ with $O_p(G)\leq H$, then $H$ is of characteristic $p$.
\end{itemize}
\end{lemma}
\begin{proof}
See \cite[Lemma 1.2(a),(c)]{MS:2012b} for parts (a) and (b). If $H$ is as in (c), then $O_p(G)\leq O_p(H)$ and hence $C_H(O_p(H))\leq C_G(O_p(G))\leq O_p(G)\leq O_p(H)$.
\end{proof}
\begin{lemma}\label{L:CharpNGH}
Let $G$ be of characteristic $p$ and $H\subn G$. Then $N_G(H)$ is of characteristic $p$.
\end{lemma}
\begin{proof}
As $H\subn G$, it follows from Lemma~\ref{L:MS}(b) that $H$ is of characteristic $p$. By Lemma~\ref{L:OpGinNormalizer}, we have $O_p(G)\leq N_G(O^p(H))$. Hence, $N_G(O^p(H))$ has characteristic $p$ by Lemma~\ref{L:MS}(c). Notice that $N_G(H)\leq N_G(O^p(H))$ as $O^p(H)$ is characteristic in $H$. Moreover, $H\subn N_G(O^p(H))$. Hence, replacing $G$ by $N_G(O^p(H))$, we can and will assume from now on that $O^p(H)\unlhd G$.
\smallskip
Let $T\in\Syl_p(H)$. Since $H\unlhd N_G(H)$, it is a consequence of \cite[Lemma~4.3]{Grazian/Henke} that $N_G(H)$ is of characteristic $p$ if and only if $N_{N_G(H)}(T)$ is of characteristic $p$. However, as $O^p(H)\unlhd G$ and $H=O^p(H)T$, we have $N_{N_G(H)}(T)=N_G(T)$. By Lemma~\ref{L:MS}(a), $N_G(T)$ has characteristic $p$.
\end{proof}
\begin{lemma}\label{L:NormalizerConstrained}
Let $\F$ be a constrained fusion system. Suppose $G$ and $\tG$ are models for $\F$. Let moreover $H\subn G$ and $\tH\subn \tG$ such that $\F_{S\cap H}(H)=\F_{S\cap \tH}(\tH)$. Then
\[\F_{N_S(H)}(N_G(H))=\F_{N_S(\tH)}(N_G(\tH)).\]
\end{lemma}
\begin{proof}
As $G$ and $\tG$ are models, it follows from the \cite[Theorem~III.5.10]{Aschbacher/Kessar/Oliver:2011} that there exists an isomorphism $\alpha\colon G\rightarrow \tG$ which restricts to the identity on $S$. If $\Delta$ is the set of all subgroups of $S$, then $(G,\Delta,S)$ and $(\tG,\Delta,S)$ are localities. Hence, it is a special case of Lemma~\ref{L:AutLFusion} that
\begin{equation}\label{E:FX}
S\cap X=S\cap (X\alpha)\mbox{ and }\F_{S\cap X}(X)=\F_{S\cap X}(X)^\alpha=\F_{S\cap (X\alpha)}(X\alpha)\mbox{ for all }X\leq S.
\end{equation}
Notice that $H\alpha\subn\tG$. Moreover \eqref{E:FX} together with our assumption gives $\F_{S\cap \tH}(\tH)=\F_{S\cap H}(H)=\F_{S\cap (H\alpha)}(H\alpha)$. By \cite[Lemma~2.9]{Chermak/Henke}, $\tH$ is the unique subnormal subgroup of $\tG$ realizing $\F_{S\cap \tH}(\tH)$. Hence, $\tH=H\alpha$. It follows $N_G(H)\alpha=N_{\tG}(\tH)$. Hence, using \eqref{E:FX} now for $X=N_G(H)$, we obtain $\F_{N_S(H)}(N_G(H))=\F_{N_S(\tH)}(N_G(\tH))$.
\end{proof}
\section{Normalizers and centralizers}
Throughout this section we assume the following hypothesis.
\begin{hypo}\label{H:main}
Let $(\L,\Delta,S)$ be a regular locality and $\H\subn\L$. Set
\[\tT:=E(\L)\cap S,\;G:=N_\L(\tT),\]
\[T:=\H\cap S,\;\E:=\F_T(\H),\;T_0:=E(\H)\cap S,\]
\[\M:=\left(\prod_{\K\in\Comp(\L)\backslash\Comp(\H)}\K\right)\mbox{ and }T_0^\perp:=\M\cap S.\]
\end{hypo}
We use above and throughout this section that, by Background Theorem~\ref{bT:1}, $\E$ is saturated and $(\H,\delta(\E),T)$ is a regular locality. In particular $E(\H)\unlhd\H$ and thus $T_0$ is well-defined.
\subsection{Some basic results}
\begin{lemma}\label{L:BasicT0perp}
\begin{itemize}
\item [(a)] $E(\H)$ and $\M$ are partial normal subgroups of $E(\L)$ with $E(\L)=E(\H)\M$. Moreover, $\tT=T_0T_0^\perp$, $\H\subseteq C_\L(\M)$ and $\M\subseteq C_\L(\H)$.
\item [(b)] We have $\Comp(\M)=\Comp(\L)\backslash\Comp(\H)$, i.e. $\Comp(\L)$ is the disjoint union of $\Comp(\M)$ and $\Comp(\H)$.
\item [(c)] $\H\M$ is a partial subnormal subgroup of $\L$ with $\H\unlhd \H\M$ and $\M\unlhd \H \M$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a,c)} We refer in this proof to the notion of an (internal) central product, which was defined in \cite[Definition~4.1]{Henke:Regular}. However, we will not need to go into the details of the definition.
\smallskip
By \cite[Remark~11.2(b)]{Henke:Regular}, we have $\Comp(\H)\subseteq\Comp(\L)$. By \cite[Lemma~4.5]{Henke:Regular}, forming the internal central product is associative. Hence, it follows from \cite[Theorem~11.18(a)]{Henke:Regular} that $E(\L)$ is a central product of $E(\H)$ and $\M$, and from \cite[Theorem~11.18(c)]{Henke:Regular} that $\H\M$ is subnormal in $\L$ and a central product of $\H$ and $\M$. In particular, by \cite[Lemma~4.9]{Henke:Regular}, $E(\H)$ and $\M$ are normal in $E(\L)=E(\H)\M$, and $\H$ and $\M$ are normal in $\H\M$. It is moreover a consequence of \cite[Lemma~4.8]{Henke:Regular} that $\H\subseteq C_\L(\M)$ and $\M\subseteq C_\L(\H)$.
\smallskip
By Background Theorem~\ref{bT:1}, $E(\L)$ can be regarded as a regular locality with Sylow subgroup $\tT$. Hence, it follows from \cite[Theorem~5.1]{Chermak:2015} (applied with $E(\L)$ in place of $\L$) that $\tT=(E(\H)\cap \tT)(\M\cap\tT)=T_0T_0^\perp$.
\smallskip
\textbf{(b)} Part (a) yields that $\M\unlhd E(\L)\unlhd\L$ is subnormal in $\L$. In particular, $\Comp(\M)$ is well-defined by Background Theorem~\ref{bT:1}. Using \cite[Remark~11.2]{Henke:Regular}, one sees that $\Comp(\L)\backslash\Comp(\H)\subseteq\Comp(\M)$ and $\Comp(\M)\subseteq\Comp(\L)$. If $\K\in\Comp(\M)\cap \Comp(\H)$, then part (a) yields that $\K=Z(\K)$, which contradicts $\K$ being quasisimple. Hence, $\Comp(\M)\subseteq\Comp(\L)\backslash\Comp(\H)$ and the assertion holds.
\end{proof}
By Lemma~\ref{L:BasisNLtT}, $G$ is a subgroup of $\L$ which acts on $\L$ via conjugation. Thus, for every subset $\X$ of $\L$, the set-wise stabilizer
\[N_{G}(\X):=\{f\in G\colon \X^f=\X\}\]
is a subgroup of $G$. In particular, $N_G(\H)$ is a subgroup of $G$. As $S\leq G$, it follows that $N_S(\H):=S\cap N_G(\H)$ is a subgroup of $S$.
\begin{lemma}\label{L:NinNLtTofH}
The following hold:
\begin{itemize}
\item [(a)] $N_\H(\tT)=N_\H(T_0)\subn G$.
\item [(b)] $N_\F(\tT)$ is constrained and $G$ is a model for $N_\F(\tT)$. Moreover, $N_\H(\tT)\subn G$ is a model for $N_\E(T_0)$. In particular,
\[N_\E(T_0)\subn N_\F(\tT).\]
\item [(c)] $N_G(\H)=N_G(N_\H(\tT))=N_G(N_\H(T_0))$ is of characteristic $p$.
\item [(d)] $N_{E(\L)}(\tT)\subseteq N_{G}(\H)$.
\item [(e)] $N_{G}(\H)\subseteq N_{G}(E(\H))=N_{G}(T_0)=N_{G}(\M)=N_{G}(T_0^\perp)$.
\item [(f)] If $g\in G$ with $T_0^g\leq T$, then $g\in N_G(E(\H))$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} Using Lemma~\ref{L:BasicT0perp}(a), one can observe that $\H\subseteq C_\L(\M)\subseteq C_\L(T_0^\perp)$ and then $N_\H(T_0)\subseteq N_\H(T_0T_0^\perp)=N_\H(\tT)$. As $T_0\leq \tT$, we have $T_0=E(\H)\cap \tT$. Thus, $E(\H)\unlhd\H$ implies that $T_0$ is normalized by $N_\H(\tT)$. This proves (a).
\smallskip
\textbf{(b)} As $\tT\in\Delta$ by Lemma~\ref{L:BasisNLtT}, the group $G$ is of characteristic $p$. By \cite[Lemma~3.10(b)]{Grazian/Henke}, we have $N_\F(\tT)=\F_S(G)$, i.e. $N_\F(\tT)$ is constrained and $G$ is a model for $N_\F(\tT)$. The same argument with $(\H,\delta(\E),T)$ and $\E$ in place of $(\L,\Delta,S)$ and $\F$ yields that $N_\H(T_0)$ is a model for $N_\E(T_0)$. By (a), $N_\H(T_0)=N_\H(\tT)$. If $\H=\H_0\unlhd\H_1\unlhd\cdots\unlhd\H_n=\L$ is a subnormal series of $\H$ in $\L$, then $N_\H(\tT)=N_{\H_0}(\tT)\unlhd N_{\H_1}(\tT)\unlhd\cdots \unlhd N_{\H_n}(\tT)=G$ is a subnormal series of $N_\H(\tT)$ in $G$, i.e. $N_\H(\tT)\subn G$. In particular, $N_\E(T_0)\subn N_\F(\tT)$ by \cite[Proposition~I.6.2]{Aschbacher/Kessar/Oliver:2011}. This proves (b).
\smallskip
\textbf{(f)} Let $g\in G$ with $T_0^g\leq T$ and $\K\in \Comp(\H)$. Recall that $c_g\in\Aut(\L)$ by Lemma~\ref{L:BasisNLtT}
and so $\K^g\in\Comp(\L)$ by \cite[Lemma~11.12]{Henke:Regular}. Notice moreover that $\K\cap S\leq T_0$ and thus $(\K\cap S)^g\leq T_0^g\leq T$. If $\K^g$ were not a component of $\H$, then Lemma~\ref{L:BasicT0perp}(a) would imply that $\K^g\subseteq \M\subseteq C_\L(\H)\subseteq C_\L(T)$ and thus $(\K\cap S)^g$ were abelian. As $\K$ is quasisimple and $\K\cap S\cong (\K\cap S)^g$, this would contradict \cite[Lemma~7.10]{Henke:Regular}. Hence, $\K^g\in\Comp(\H)$. As $\K$ was arbitrary, this yields $E(\H)^g=E(\H)$.
\smallskip
\textbf{(c)} Using parts (a) and (b), Lemma~\ref{L:CharpNGH} yields that $N_{G}(N_\H(\tT))=N_G(N_\H(T_0))$ is of characteristic $p$. Hence, it remains only to show that $N_{G}(\H)=N_{G}(N_\H(\tT))$.
\smallskip
Clearly, $N_G(\H)\leq N_G(\H\cap G)=N_G(N_\H(\tT))$. Fix now $g\in N_G(N_\H(\tT))$. Then $T_0^g\leq N_\H(\tT)\cap \tT\leq \H\cap S=T$. Hence, by (f), we have $E(\H)^g=E(\H)$. It follows from the Frattini Lemma \cite[Corollary~3.11]{Chermak:2015} applied with $(\H,\delta(\E),T)$ in place of $(\L,\Delta,S)$ that $\H=E(\H)N_\H(T_0)=E(\H)N_\H(\tT)$. Using again that $c_g\in\Aut(\L)$ by Lemma~\ref{L:BasisNLtT}, we can conclude that $\H^g=E(\H)^gN_\H(\tT)^g=E(\H)N_\H(\tT)=\H$. Hence, $N_{G}(N_\H(\tT))\leq N_{G}(\H)$ and so equality holds.
\smallskip
\textbf{(d)} By Lemma~\ref{L:BasicT0perp}(a), $E(\H)\unlhd E(\L)=E(\H)\M$ and $\M\subseteq C_\L(\H)\subseteq C_\L(T_0)\subseteq N_\L(T_0)$. In particular, $N_{E(\L)}(\tT)$ normalizes $T_0=E(\H)\cap \tT$. Using the Dedekind argument for localities \cite[Lemma~1.10]{Chermak:2015} and part (a), we conclude that $N_{E(\L)}(\tT)\subseteq N_{E(\L)}(T_0)=N_{E(\H)}(T_0)\M=N_{E(\H)}(\tT)\M$ and thus $N_{E(\L)}(\tT)=N_{E(\H)}(\tT)N_\M(\tT)$. As $N_{E(\H)}(\tT)\subseteq N_\H(\tT)\subseteq N_G(N_\H(\tT))$ and $\M\subseteq C_\L(\H)$, this implies $N_{E(\L)}(\tT)\subseteq N_{G}(N_\H(\tT))$. Now (d) follows from (c).
\smallskip
\textbf{(e)} By Lemma~\ref{L:BasisNLtT}, we have $c_g\in\Aut(\L)$ for every $g\in G$. Hence, for $g\in N_G(\H)$, $c_g$ restricts to an automorphism of $\H$. It follows now from \cite[Lemma~11.12]{Henke:Regular} that $G$ acts on $\Comp(\L)$, and that $N_G(\H)$ acts on $\Comp(\H)$. In particular, $N_G(\H)\leq N_G(E(\H))$. Similarly, $N_G(E(\H))$ acts on $\Comp(\H)=\Comp(E(\H))$ (cf. \cite[Remark~11.2]{Henke:Regular}). Hence, $N_G(E(\H))$ acts also on $\Comp(\L)\backslash \Comp(\H)$, which implies $N_G(E(\H))\leq N_G(\M)$. Notice that $N_G(E(\H))\leq N_G(E(\H)\cap \tT)=N_G(T_0)$. By (f), we have $N_G(T_0)\leq N_G(E(\H))$. We know thus that
\[N_G(E(\H))=N_G(T_0)\leq N_G(\M).\]
By Lemma~\ref{L:BasicT0perp}(a),(b), $\M\unlhd E(\L)\unlhd\L$ is subnormal in $\L$, $\Comp(\H)=\Comp(\L)\backslash\Comp(\M)$ and $\M=E(\M)$. Hence, applying the same property with $\M$ in place of $\H$, we obtain $N_G(\M)=N_G(T_0^\perp)\leq N_G(E(\H))$. This shows (e).
\end{proof}
\begin{lemma}\label{L:NLT}
Let $f\in N_\L(T)$. Then $\H\subseteq \D(f)$. Moreover, there exist $n\in E(\L)$ and $g\in G$ such that $(n,g)\in\D$, $f=ng$ and $\H^f=\H^g$.
\end{lemma}
\begin{proof}
By Lemma~\ref{L:FrattiniSplitting}, there exist $n\in E(\L)$ and $g\in G$ such that $(n,g)\in\D$, $f=ng$ and $S_f=S_{(n,g)}$. As $f\in N_\L(T)$, we have then $T\leq S_f\leq S_n$. By Lemma~\ref{L:BasicT0perp}(c), $\H\M$ is a partial subnormal subgroup of $\L$ with $\H\unlhd\H\M$. Notice that $E(\L)=E(\H)\M\subseteq \H\M$ by Lemma~\ref{L:BasicT0perp}(a). So $T^n\subseteq S\cap\H=T$ and thus $n\in N_{E(\L)}(T)\subseteq N_{\H\M}(T)$. As $\H\M\subn\L$, Background Theorem~\ref{bT:1} yields that $\H\M$ can be given the structure of a regular locality with Sylow subgroup $S\cap(\H\M)$. Hence, by \cite[Theorem~10.16(f)]{Henke:Regular} applied with $\H\M$ in place of $\L$, we have $N_{\H\M}(T)\subseteq N_{\H\M}(\H)$. In particular, $n\in N_\L(\H)$. Fixing $h\in\H$, it follows $(n^{-1},h,n)\in\D$ and thus $Q:=S_{(n^{-1},h,n)}\in\Delta$ by \cite[Corollary~2.7]{Chermak:2015}. As $(\L,\Delta,S)$ is regular, we have $\Delta=\delta(\F)$ and \cite[Lemma~10.2]{Henke:Regular} yields $Q\cap F^*(\L)\in\Delta$. By Lemma~\ref{L:BasisNLtT}, $g\in N_\L(F^*(\L)\cap S)$ and then $v=(g^{-1},n^{-1},h,n,g)\in\D$ via $(Q\cap F^*(\L))^g$. By \cite[Lemma~1.4(f)]{Chermak:2015}, $g^{-1}n^{-1}=f^{-1}$. Hence, using the ``associativity axiom'' in partial groups, we can conclude that $(f^{-1},h,f)\in\D$ and $h^f=\Pi(v)=(h^n)^g$. Since this holds for an arbitrary element $h\in\H$, this shows $\H\subseteq \D(f)$ and $\H^f=(\H^n)^g=\H^g$, where the last equality uses $n\in N_\L(\H)$. This proves the assertion.
\end{proof}
\subsection{The regular normalizer and the centralizer in $\L$}
Recall from the introduction that
\[\bN_\L(\H):=E(\L)N_{G}(\H)\]
is called the \emph{regular normalizer} of $\H$ in $\L$.
\smallskip
In this subsection we will prove Theorem~\ref{T:MainbNLH} and some related results. Before we do so let us consider the special case that $\F$ is constrained.
\begin{remark}\label{R:Constrained}
Let $\F$ be constrained. Then by \cite[Lemma~11.6]{Henke:Regular}, $E(\L)=1$ and $\L$ is a model for $\F$. In particular, $\H$ is a subnormal subgroup of $\L$, $\tT=1$ and $\bN_\L(\H)=N_\L(\H)$.
\end{remark}
\begin{theorem}\label{T:bNLH}
The regular normalizer $\bN_\L(\H)$ is a partial subgroup of $\L$. Moreover, fixing $S_0\in\Syl_p(N_{G}(\H))$ and setting
\[\Delta_0:=\{P\leq S_0\colon P\cap\tT\in E(\F)^s\},\]
the following hold:
\begin{itemize}
\item [(a)] The triple $(\bN_\L(\H),\Delta_0,S_0)$ is a linking locality and the fusion system $\F_0:=\F_{S_0}(\bN_\L(\H))$ is saturated. Moreover, $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$ and
\[\delta(\F_0):=\{P\leq S_0\colon PO_p(\bN_\L(\H))\in\Delta_0\}.\]
\item [(b)] $\H\unlhd\bN_\L(\H)$. If $T\leq S_0$, then $T=\H\cap S_0$ and $\E\unlhd\F_0$.
\item [(c)] $\Comp(\bN_\L(\H))=\Comp(\L)$, $E(\bN_\L(\H))=E(\L)$ and $\tT=E(\L)\cap S_0$. In particular, $\Comp(\F)=\Comp(\F_0)$ and $E(\F)=E(\F_0)$.
\item [(d)] $E(\H)\unlhd\bN_\L(\H)$, and $\M\unlhd\bN_\L(\H)$. In particular, $E(\E)\unlhd \F_0$ and $\F_{T_0^\perp}(\M)\unlhd \F_0$.
\item [(e)] $N_{\bN_\L(\H)}(\tT)=N_{G}(\H)$. In particular, $S\cap \bN_\L(\H)=N_S(\H)$.
\end{itemize}
\end{theorem}
\begin{proof}
\textbf{(a)} Set $T^*:=F^*(\L)\cap S$ and recall from Lemma~\ref{L:BasisNLtT} that $G=N_\L(T^*)$. So $H:=N_{G}(\H)\leq N_\L(T^*)$. By Lemma~\ref{L:NinNLtTofH}(d), we have
\[N_{E(\L)}(T^*)=N_{E(\L)}(\tT)\leq H.\]
In particular, $\tT$ is contained in $H$ and then a normal $p$-subgroup of $H$. Therefore $\tT\leq S_0$. By Lemma~\ref{L:NinNLtTofH}(c), $H$ is of characteristic $p$. By Background Theorem~\ref{bT:2}, we have $\F_{S\cap E(\L)}(E(\L))=E(\F)$ and by \cite[Lemma~7.22]{Chermak/Henke}, $\delta(E(\F))=E(\F)^s$. Hence,
\[\Delta_0=\{P\leq S_0\colon P\cap E(\L)\in\delta(\F_{S\cap E(\L)}(E(\L)))\}.\]
It follows now from \cite[Theorem~6.1]{Grazian/Henke} (applied with $E(\L)$ in place of $\N$) that $\bN_\L(\H)$ is a partial subgroups of $\L$, $(\bN_\L(\H),\Delta_0,S_0)$ is a linking locality, $\F_0:=\F_{S_0}(\bN_\L(\H))$ is saturated and $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$, where
\[\delta(\F_0)=\{P\leq S_0\colon PO_p(\bN_\L(\H))\in\Delta_0\}.\footnote{The reader might want to observe that the arguments to prove \cite[Theorem~6.1]{Grazian/Henke} become actually easier in the case that $\N$ equals $E(\L)$.} \]
This shows (a).
\smallskip
We will use now throughout that $\F_0$ is well-defined and saturated, and that $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$.
\smallskip
\textbf{(e)} By the Dedekind Lemma, $N_{\bN_\L(\H)}(\tT)=N_{E(\L)}(\tT)N_{G}(\H)$. Hence, Lemma~\ref{L:NinNLtTofH}(d) implies that $N_{\bN_\L(\H)}(\tT)=N_{G}(\H)$. As $\tT\unlhd S$, it follows in particular that $S\cap \bN_\L(\H)=S\cap N_{\bN_\L(\H)}(\tT)=S\cap N_G(\H)=N_S(\H)$.
\smallskip
\textbf{(b)} If $\H$ is a partial normal subgroup of $\L$, then $\E=\F_T(\H)\unlhd \F_0$ by (a) and Background Theorem~\ref{bT:1}. Thus, it is sufficient to show $\H\unlhd\bN_\L(\H)$.
\smallskip
We prove first that $\H$ is contained in $\bN_\L(\H)$. It follows from the Frattini Lemma \cite[Corollary~3.11]{Chermak:2015} applied with $(\H,\delta(\E),T)$ in place of $(\L,\Delta,S)$ that $\H=E(\H)N_\H(T_0)$. As $E(\H)\subseteq E(\L)$ by Lemma~\ref{L:BasicT0perp}(a) and $N_\H(T_0)=N_\H(\tT)\subseteq N_{G}(\H)$ by Lemma~\ref{L:NinNLtTofH}(a), it follows $\H\subseteq\bN_\L(\H)$.
\smallskip
As $\H$ is a partial subgroup of $\L$, it is also a partial subgroup of $\bN_\L(\H)$. Let $f\in\bN_\L(\H)$ and $h\in \H$ such that $h\in\D(f)$. We need to show that $h^f\in\H$.
\smallskip
Clearly $E(\L)\unlhd \bN_\L(\H)$ as $E(\L)\unlhd\L$ is contained in $\bN_\L(\H)$. Hence, by Lemma~\ref{L:FrattiniSplitting} applied with $(\bN_\L(\H),\delta(\F_0),S_0)$ in place of $(\L,\Delta,S)$, there exists $x\in E(\L)$ and $g\in N_{\bN_\L(\H)}(\tT)$ such that $(x,g)\in\D$, $f=xg$ and
\[(S_0)_f=(S_0)_{(x,g)}.\]
It follows from \cite[Lemma~1.4(f), Proposition~2.6(b)]{Chermak:2015} that $(g^{-1},x^{-1})\in\D\cap \W(\bN_\L(\H))$ via $(S_0)_{f^{-1}}=(S_0)_f^f=(S_0)_{(x,g)}^{xg}$ and $f^{-1}=g^{-1}x^{-1}$. So $u:=(g^{-1},x^{-1},h,x,g)\in\D$ via $(S_0)_{(f^{-1},x,f)}$ and, by the axioms of a partial group, $h^f=\Pi(f^{-1},h,f)=\Pi(u)=(h^x)^g$.
\smallskip
By Lemma~\ref{L:BasicT0perp}(a),(b), $E(\L)=E(\H)\M\subseteq \H\M$, and $\H\M$ is a partial subgroup of $\L$ with $\H\unlhd \H\M$. In particular, $h^x\in\H$. By (e), $g\in N_{G}(\H)$. Thus, it follows $h^f=(h^x)^g\in\H$ as required.
\smallskip
\textbf{(c)} As every component of $\L$ is subnormal in $\L$ and contained in the partial subgroup $\bN_\L(\H)$ of $\L$, it is straightforward to see that every component of $\L$ is subnormal in $\bN_\L(\H)$ (cf. \cite[Lemma~3.7(a)]{Henke:Regular}). Thus $\Comp(\L)\subseteq\Comp(\bN_\L(\H))$.
\smallskip
Suppose now that there is a component $\K\in\Comp(\bN_\L(\H))\backslash\Comp(\L)$. As $E(\L)\unlhd\bN_\L(\H)$, it follows from \cite[Lemma~11.16]{Henke:Regular} (applied with $(\bN_\L(\H),E(\L))$ in place of $(\L,\N)$ and combined with \cite[Theorem~10.16(e)]{Henke:Regular}) that
\[\K\subseteq C_{\bN_\L(\H)}(E(\L))\subseteq C_{\bN_\L(\H)}(\tT)\subseteq N_{\bN_\L(\H)}(\tT).\]
By (e) and Lemma~\ref{L:NinNLtTofH}(c), $N_{\bN_\L(\H)}(\tT)=N_{G}(\H)$ is a group of characteristic $p$. As $\K$ is a partial subnormal subgroup of $\bN_\L(\H)$, it follows that $\K$ is a partial subnormal subgroup of $N_{\bN_\L(\H)}(\tT)$ and thus by Lemma~\ref{L:MS}(b) of characteristic $p$. However, by \cite[Lemma~7.10]{Henke:Regular}, the fact that $\K$ is a quasisimple locality implies $O_p(\K)=Z(\K)$. This contradicts $\K$ being of characteristic $p$ and shows thus that
\[\Comp(\L)=\Comp(\bN_\L(\H)).\]
In particular, $E(\L)=E(\bN_\L(\H))$. Using (a), it follows thus from Background Theorem~\ref{bT:2} that $\Comp(\F)=\Comp(\F_0)$ and $E(\F)=E(\F_0)$.
\smallskip
\textbf{(d)} As $\H\unlhd\bN_\L(\H)$, it follows from \cite[Lemma~11.13]{Henke:Regular} that $E(\H)\unlhd\bN_\L(\H)$. By Lemma~\ref{L:BasicT0perp}(a), $\M\unlhd E(\L)$. Moreover, by (e) and Lemma~\ref{L:NinNLtTofH}(e), $N_{\bN_\L(\tT)}(\tT)=N_G(\H)$ normalizes $\M$. As $E(\L)\unlhd\bN_\L(\H)$ with $S_0\cap E(\L)=\tT$, it follow now from \cite[Corollary~3.13]{Chermak:2015} that $\M\unlhd \bN_\L(\H)$.
\smallskip
It is a consequence of Background Theorem~\ref{bT:2} (applied with $(\N,\delta(\E),T)$ in place of $(\L,\Delta,S)$) that $E(\E)=\F_{T_0}(E(\H))$. As $T_0=E(\H)\cap\tT=E(\H)\cap S_0$ and $T_0^\perp=\M\cap \tT=\M\cap S_0$, Background Theorem~\ref{bT:2} (applied with $(\bN_\L(\H),\delta(\F_0),S_0)$ in place of $(\L,\Delta,S)$) gives now also $E(\E)\unlhd \F_0$ and $\F_{T_0^\perp}(\M)\unlhd\F_0$.
\end{proof}
\begin{lemma}\label{L:X}
Let $\X$ be a partial subgroup of $\L$ such that $\H\unlhd\X$ and $(\X,S_\X,\Gamma_\X)$ is a locality for some $p$-subgroup $S_\X$ of $\X$ with $T\leq S_X$ and some set $\Gamma_\X$ of subgroups of $S_X$. Then $\X\subseteq \bN_\L(\H)$.
\end{lemma}
\begin{proof}
It is a consequence of Background Theorem~\ref{bT:1} (and the definition of a locality) that $T$ is a maximal $p$-subgroup of $\H$. As $T\leq S_\X\cap\H$, it follows thus that $T=S_\X\cap \H$. The Frattini argument for localities \cite[Corollary~3.11]{Chermak:2015} (applied with $(\X,S_\X,\Delta_\X)$ in place of $(\L,\Delta,S)$) gives therefore $\X=\H N_\X(T)$. By Theorem~\ref{T:bNLH} $\bN_\L(\H)$ is a partial subgroup of $\L$ with $\H\subseteq \bN_\L(\H)$. Thus it remains to show that $N_\X(T)\subseteq \bN_\L(\H)$.
\smallskip
Let $f\in N_\X(T)$. By Lemma~\ref{L:NLT}, we have $\H\subseteq\D(f)$ and there exist $n\in E(\L)$ and $g\in G$ such that $(n,g)\in\D$, $f=ng$ and $\H^f=\H^g$. As $\H\unlhd \X$, it follows $\H^g=\H^f=\H$ and thus $g\in N_G(\H)$. Hence, $f=ng\in E(\L)N_G(\H)=\bN_\L(\H)$. This completes the proof.
\end{proof}
\begin{lemma}\label{L:finNLtT}
Let $f\in N_{G}(T)$. Then $f\in N_{G}(\H)$ if and only if $c_f|_T\in\Aut(\E)$. In particular, \[N_S(\E)=N_S(\H)=N_S(N_\H(\tT))=N_S(N_\E(T_0)).\]
\end{lemma}
\begin{proof}
Recall that $c_f\in\Aut(\L)$ by Lemma~\ref{L:BasisNLtT}. In particular, as $\H\subn\L$, we have $\H^f\subn\L$. Moreover, setting $\alpha:=c_f|_T$, it follows from Lemma~\ref{C:AutLFusion} that $S\cap \H^f=T^f=T$ and
\[\F_T(\H^f)=\E^\alpha.\]
In particular, if $\H^f=\H$, then $\alpha\in\Aut(\E)$. On the other hand, if $\alpha\in\Aut(\E)$, then $\F_T(\H^f)=\E^\alpha=\E=\F_T(\H)$. Since the map $\hat{\Psi}$ from Background Theorem~\ref{bT:2} is injective, we see thus that $\alpha\in\Aut(\E)$ implies $\H^f=\H$. This shows $f\in N_{G}(\H)$ if and only if $c_f|_T\in\Aut(\E)$. In particular,
\begin{equation}\label{E:NSE}
N_S(\E)=N_S(\H).
\end{equation}
By Lemma~\ref{L:NinNLtTofH}(b), $G$ is a model for $N_\F(\tT)$ and $N_\H(\tT)\subn G$ is a model for $N_\E(T_0)$. It follows from Lemma~\ref{L:MS}(a) that $(G,\delta(N_\F(\tT)),S)$ is a regular locality over $N_\F(\tT)$. Thus, applying \eqref{E:NSE} with $(G,N_\H(\tT),N_\E(T_0))$ in place of $(\L,\H,\E)$, we obtain $N_S(N_\H(\tT))=N_S(N_\E(T_0))$. Lemma~\ref{L:NinNLtTofH}(c) implies that $N_S(\H)=N_S(N_\H(\tT))$. Hence the assertion holds.
\end{proof}
\begin{lemma}\label{L:CLTinbNLH}
\begin{itemize}
\item [(a)] $C_\L(T)\subseteq \bN_\L(\H)$.
\item [(b)] $C_\L(\H)=C_{\bN_\L(\H)}(\H)\unlhd \bN_\L(\H)$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} As $T$ is a $p$-subgroup of $N_G(\H)$, we may choose $S_0\in\Syl_p(N_G(\H))$ with $T\leq S_0$. Set $\F_0:=\F_{S_0}(\bN_\L(\H))$. By Theorem~\ref{T:bNLH}(a),(b), $\F_0$ is saturated and $\E\unlhd\F_0$. Let $f\in C_\L(T)$. By Lemma~\ref{L:FrattiniSplitting}, there exists $x\in E(\L)$ and $g\in G$ such that $(x,g)\in\D$, $f=xg$ and $S_{(x,g)}=S_f$. Notice that $E(\L)\subseteq \bN_\L(\H)$, so $c_x|_T\in\Aut_{\F_0}(T)\leq \Aut(\E)$. As $(c_x|_T)\circ (c_g|_T)=c_f|_T=\id\in \Aut(\E)$, it follows $c_g|_T\in\Aut(\E)$. Hence, Lemma~\ref{L:finNLtT} yields $g\in N_{G}(\H)$ and thus $f=xg\in E(\L)N_{G}(\H)=\bN_\L(\H)$.
\smallskip
\textbf{(b)} Using (a), we see that $C_\L(\H)\subseteq C_\L(T)\subseteq \bN_\L(\H)$ and thus $C_\L(\H)=C_{\bN_\L(\H)}(\H)$. Recall that $\bN_\L(\H)$ forms a regular locality by Theorem~\ref{T:bNLH}(a). Hence, part (b) follows from Background Theorem~\ref{bT:1} applied with $\bN_\L(\H)$ in place of $\L$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{T:MainbNLH}]
Parts (a),(b),(c) follows from Theorem~\ref{T:bNLH}(a),(b),(c), and part (d) is stated in Theorem~\ref{T:bNLH}(e). Part (e) is stated in Lemma~\ref{L:X}, and part (f) in Lemma~\ref{L:CLTinbNLH}.
\end{proof}
\subsection{The normalizer of $\E$}
We continue to assume Hypothesis~\ref{H:main} and assume in addition for the remainder of this section the following hypothesis.
\begin{hypo}\label{H:S0F0}
Let $S_0$ be a Sylow $p$-subgroup of $N_{G}(\H)$ with $T\leq S_0$. Set
\[\F_0:=\F_{S_0}(\bN_\L(\H)).\]
\end{hypo}
We will use throughout that, by Theorem~\ref{T:bNLH}, $\F_0$ is well-defined and saturated, $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$, $E(\L)=E(\bN_\L(\H))\unlhd \bN_\L(\H)$, $\tT=E(\L)\cap S_0$, $E(\F)=E(\F_0)\unlhd \F_0$, $\H\unlhd\bN_\L(\H)$, $T=S_0\cap \H$ and $\E\unlhd\F_0$. We use moreover throughout that
\[N_S(\E)=N_S(\H)\]
by Lemma~\ref{L:finNLtT}. Recall that
\[N_\F(\E)=\F_{N_S(\H)}(\bN_\L(\H)),\]
so in particular, $N_\F(\E)$ is a fusion system over $N_S(\H)=N_S(\E)$.
\smallskip
In this subsection, we will prove Theorem~\ref{T:FusionNormalizerGenerate}, Theorem~\ref{T:FusionNormalizer}(a),(b),(d) and Theorem~\ref{T:MainER}, as well as some related results concerning in particular the connection between $N_\F(\E)$ and $\F_0$.
\begin{lemma}\label{L:EFR}
For every $R\leq N_S(\E)\cap S_0$, we have
\[E(\F)R:=(E(\F)R)_\F=\F_{\tT R}(E(\L)R)=(E(\F)R)_{\F_0}\]
and $(E(\L)R,\delta(E(\F)R),\tT R)$ is a regular locality over $E(\F)R$. In particular, $E(\F)R\subseteq N_\F(\E)$.
\end{lemma}
\begin{proof}
By Background Theorem~\ref{bT:2}, we have $E(\F)=\F_{S\cap E(\L)}(E(\L))=\F_{\tT}(E(\L))=\F_{S_0\cap E(\L)}(E(\L))$. Hence, by \cite[Theorem~D(a)]{Chermak/Henke} applied twice (once with $\L$ and once with $\bN_\L(\H)$ in place of $\L$), we have
\[(E(\F)R)_\F=\F_{\tT R}(E(\L)R)=(E(\F)R)_{\F_0}\]
and $(E(\L)R,\delta(E(\F)R),\tT R)$ is a regular locality over $E(\F)R$. As $\tT R\leq N_S(\E)$ and $E(\L)R\leq \bN_\L(\H)$, it follows that $E(\F)R\subseteq N_\F(\E)$.
\end{proof}
For every $R\leq N_S(\E)$ we write from now on usually $E(\F)R$ for $(E(\F)R)_\F$.
\begin{lemma}\label{L:FrattiniApply}
Let $X\leq N_S(\E)$ and $\phi\in\Hom_\F(X,S)$. Then there exist
\[\psi\in\Hom_{E(\F)X}(X,X\tT)\subseteq \Hom_{N_\F(\E)}(X,X\tT)\]
and $\alpha\in\Hom_{N_\F(\tT)}(X\psi,S)$ such that $\phi=\psi\alpha$. If $X\leq S_0$, then $\psi$ is a morphism in $\F_0$.
\end{lemma}
\begin{proof}
The existence of $\psi\in\Hom_{E(\F)X}(X,\tT X)$ and $\alpha\in\Hom_{N_\F(\tT)}(X\psi,S)$ with $\phi=\psi\alpha$ follows from Lemma~\ref{L:FusionSystemFrattini} applied with $E(\F)$ in place of $\E$. The assertion follows now from Lemma~\ref{L:EFR}.
\end{proof}
\begin{lemma}\label{L:NFEbNLH}
Set $\B:=\F_{N_S(\H)}(N_G(\H))$. Then the following hold:
\begin{itemize}
\item [(a)] We have
\[N_\F(\E)=\<E(\F)N_S(\E), \B\>.\]
If $N_S(\E)\leq S_0$, then also
\[N_\F(\E)=\F\cap \F_0=\F_0|_{N_S(\E)}.\]
\item [(b)] If $P\leq N_S(\E)$ and $\phi\in\Hom_{N_\F(\E)}(P,N_S(\E))$, then there exist
\[\psi\in\Hom_{E(\F)P}(P,\tT P)\mbox{ and }\alpha\in\Hom_{\B}(P\psi,N_S(\E))\]
such that $\phi=\psi\alpha$.
\item [(c)] $\E$ is a $N_\F(\E)$-invariant subsystem of $N_\F(\E)$.
\end{itemize}
\end{lemma}
\begin{proof}
As $N_S(\E)=N_S(\H)$ is a $p$-subgroup of $N_G(\H)$, there exists always a Sylow $p$-subgroup of $N_G(\H)$ containing $N_S(\E)$. Hence, we may assume throughout the proof that $N_S(\E)\leq S_0$.
\smallskip
\textbf{(a,b)} As $N_S(\E)\leq S_0$ and $\E\unlhd\F_0$, we have $S_0\cap S=N_S(\E)=N_S(\H)$. Notice that $N_\F(\E)=\F_{N_S(\H)}(\bN_\L(\H))$ is contained in $\F=\F_S(\L)$ and in $\F_0=\F_{S_0}(\bN_\L(\H))$. Hence,
\[N_\F(\E)\subseteq \F\cap \F_0\subseteq \F_0|_{N_S(\E)}.\]
It follows moreover from Lemma~\ref{L:EFR} that $E(\F)N_S(\E)=\F_{N_S(\E)}(E(\L)N_S(\E))\subseteq \F_0$. Clearly, $\B$ is contained in $\F_0$. Hence,
\[\G:=\<E(\F)N_S(\E),\B\>\subseteq \F_0\cap \F\subseteq \F_0|_{N_S(\E)}.\]
Let now $P,Q\leq N_S(\E)$ and $\phi\in\Hom_{\F_0}(P,Q)$. Notice that any morphism in $\F_0|_{N_S(\E)}$ is of this form. As $E(\F)\unlhd\F_0$, it follows from Lemma~\ref{L:FusionSystemFrattini} that there exists $\psi\in\Hom_{(E(\F)P)_{\F_0}}(P,\tT P)$ and $\alpha\in \Hom_{N_{\F_0}(\tT)}(P\psi,N_S(\E))$ such that $\phi=\psi\alpha$. By Lemma~\ref{L:EFR},
\[(E(\F)P)_{\F_0}=(E(\F)P)_{\F}\subseteq (E(\F)N_S(\E))_\F\subseteq \G\]
and
\[(E(\F)P)_{\F_0}\subseteq N_\F(\E).\]
By Lemma~\ref{L:BasisNLtT} applied with $\bN_\L(\H)$ in place of $\L$, we have $\tT\in\delta(\F_0)$. Hence, by \cite[Lemma~3.10(b)]{Grazian/Henke} and Lemma~\ref{T:bNLH}(d),
\[N_{\F_0}(\tT)=\F_{S_0}(N_{\bN_\L(\H)}(\tT))=\F_{S_0}(N_{G}(\H)).\]
As $P\psi$ and $P\psi\alpha=P\phi$ are contained in $N_S(\E)=N_S(\H)$, it follows that $\alpha$ is a morphism in $\B=\F_{N_S(\H)}(N_G(\H))$. Since $N_\F(\E)\subseteq \F_0|_{N_S(\E)}$, this proves (b). Observe that $\B\subseteq \G\cap N_\F(\E)$ and so the argument shows also that a morphism $\phi$ as above is a morphism in $\G$ and in $N_\F(\E)$. Hence, $\F_0|_{N_S(\E)}$ is contained in $\G$ and in $N_\F(\E)$. This completes the proof of (a).
\smallskip
\textbf{(c)} Using (a) one sees that $\E\subseteq \F\cap \F_0=N_\F(\E)$. As $\E$ is normal in $\F_0$, $T$ is strongly $N_\F(\E)$-closed. Moreover, $\E^\alpha=\E$ for every $\alpha\in \Aut_{\F_0}(T)$ and thus also for every $\alpha\in\Aut_{N_\F(\E)}(T)$. By Lemma~\ref{L:Finvariant}(b) remains to show that every $N_\F(\E)$-morphism between subgroups of $T$ is in $\<\E,\Aut_\F(T)\>_T$. Fix $f\in \bN_\L(\H)$. We need to show that $c_f|_{T\cap S_f}\in \<\E,\Aut_\F(T)\>_T$.
\smallskip
By Lemma~\ref{L:FrattiniSplitting} applied with $(\bN_\L(\H),\delta(\F_0), S_0)$ and $\H$ in place of $(\L,\Delta,S)$ and $\N$, there exist $h\in\H$ and $g\in N_{\bN_\L(\H)}(T)$ such that $(h,g)\in\D$, $f=hg$ and $(S_0)_{(h,g)}=(S_0)_f$. As $T\leq S_0$ and $T$ is strongly closed in $\F_0$, we have in particular $T\cap S_{(h,g)}=T\cap S_f$. Hence, $c_f|_{T\cap S_f}$ is the composition of $c_h|_{T\cap S_f}$ and $c_g|_{(T\cap S_f)^h}$. Here $c_h|_{T\cap S_f}$ is a morphism in $\E$, and $c_g|_{(T\cap S_f)^h}$ extends to $c_g|_T$, which is an element of $\Aut_{N_\F(\E)}(T)$. This proves (c).
\end{proof}
\begin{proof}[Proof of Theorem~\ref{T:FusionNormalizerGenerate}]
As already observed above, $N_\F(\E)$ is by definition a fusion system over $N_S(\H)=N_S(\E)$, i.e. (a) holds. As throughout, $T_0$ and $\tT$ are defined as in Hypothesis~\ref{H:main}. Background Theorem~\ref{bT:2} gives then however that $E(\E)$ is a fusion system over $T_0$ and $E(\F)$ is a fusion system over $\tT$. By Lemma~\ref{L:NinNLtTofH}(b), $N_\F(\tT)$ is constrained, $G$ is a model for $N_\F(\tT)$, and $N_\H(\tT)\subn G$ is a model for $N_\E(T_0)\subn N_\F(\tT)$. In particular, part (b) holds.
\smallskip
If $\tG$ is now an arbitrary a model for $N_\F(\tT)$ and $\tH\subn \tG$ with $\F_{S\cap \tH}(\tH)=N_\E(T_0)$, then it follows from Lemma~\ref{L:NormalizerConstrained} that $\F_{N_S(\tH)}(N_{\tG}(\tH))=\F_{N_S(N_\H(\tT))}(N_G(N_\H(\tT)))$. As $N_G(\H)=N_G(N_\H(\tT))$ by Lemma~\ref{L:NinNLtTofH}(c), we can conclude for such $\tG$ and $\tH$ that $N_S(\tH)=N_S(N_\H(\tT))=N_S(\H)=N_S(\E)$ and
\[\F_{N_S(\tH)}(N_{\tG}(\tH))=\F_{N_S(\H)}(N_G(\H)).\]
Hence, part (c) follows from the first statement in Lemma~\ref{L:NFEbNLH}(a).
\end{proof}
\begin{lemma}\label{L:inAutEinNFE}
Let $X,Y\leq N_S(\E)$ with $T\leq X\cap Y$ and $\phi\in\Hom_\F(X,Y)$ with $\phi|_T\in\Aut(\E)$. Then $\phi$ is a morphism in $N_\F(\E)$.
\end{lemma}
\begin{proof}
By Lemma~\ref{L:FrattiniApply}, there exists $\psi\in\Hom_{N_\F(\E)}(X,\tT X)$ and $\alpha\in\Hom_{N_\F(\tT)}(X\psi,S)$ such that $\phi=\psi\alpha$. As $\E$ is $N_\F(\E)$-invariant by Lemma~\ref{L:NFEbNLH}(c), we have $T\psi=T$ and $\E^\psi=\E$. Hence, $T=T\psi\leq X\psi$, $T\alpha=T\psi\alpha=T\phi=T$ and $\E^\alpha=(\E^\psi)^\alpha=\E^\phi=\E$. By Lemma~\ref{L:NinNLtTofH}(b), $G$ is a model for $N_\F(\tT)$, i.e. there exists $g\in G$ with $\alpha=c_g|_{X\psi}$. As $c_g|_T=\alpha|_T\in\Aut(\E)$, it follows from Lemma~\ref{L:finNLtT} that $g\in N_G(\H)$. As $X\psi\leq \tT X\leq N_S(\E)=N_S(\H)$, it follows $X\psi\alpha=(X\psi)^g\leq S\cap N_G(\H)=N_S(\H)$. So $\alpha$ is a morphism in $N_\F(\E)=\F_{N_S(\H)}(\bN_\L(\H))$. Hence, $\phi=\psi\alpha$ is in $N_\F(\E)$.
\end{proof}
\begin{corollary}\label{C:CFT}
We have $C_\F(T)=C_{N_\F(\E)}(T)$.
\end{corollary}
\begin{proof}
Clearly $C_{N_\F(\E)}(T)\subseteq C_\F(T)$. If $P,Q\leq C_S(T)$ and $\phi\in\Hom_{C_\F(T)}(P,Q)$, the $\phi$ extends to $\hat{\phi}\in \Hom_\F(PT,QT)$ with $\hat{\phi}|_T=\id_T$. As $\id_T\in\Aut(\E)$, Lemma~\ref{L:inAutEinNFE} yields that $\hat{\phi}$ is a morphism in $N_\F(\E)$. Hence, $\phi=\hat{\phi}|_P$ is a morphism in $C_{N_\F(\E)}(T)$.
\end{proof}
\begin{corollary}\label{C:CFTCF0T}
If $N_S(\E)\leq S_0$, then $C_\F(T)=C_{\F_0}(T)|_{C_S(T)}$.
\end{corollary}
\begin{proof}
As $N_S(\E)\leq S_0$, Lemma~\ref{L:NFEbNLH}(a) gives that $N_\F(\E)=\F_0|_{N_S(\E)}$. In particular, $N_\F(\E)\subseteq\F_0$ and so Corollary~\ref{C:CFT} implies
\[C_\F(T)=C_{N_\F(\E)}(T)\subseteq C_{\F_0}(T)|_{C_S(T)}.\]
On the other hand, if $P,Q\leq C_S(T)$ and $\phi\in\Hom_{C_{\F_0}(T)}(P,Q)$, then $\phi$ extends to $\hat{\phi}\in\Hom_{\F_0}(PT,QT)$ with $\hat{\phi}|_T=\id_T$. Using Lemma~\ref{L:finNLtT}, one observes that $PT$ and $QT$ are contained in $TC_S(T)\leq N_S(\H)=N_S(\E)$. Hence, it follows that $\hat{\phi}$ is a morphism in $\F_0|_{N_S(\E)}=N_\F(\E)$. This yields that $\phi$ is a morphism in $C_{N_\F(\E)}(T)=C_\F(T)$ and proves thus that $C_{\F_0}(T)|_{C_S(T)}\subseteq C_\F(T)$.
\end{proof}
Our next goal is to show that every saturated subsystem $\mD$ of $\F$ with $\E\unlhd\mD$ is contained in $N_\F(\E)$. We first state some technical lemmas which are used in the proof.
\begin{lemma}\label{L:QcapT0phi}
Let $T_0^\perp\leq Q\leq N_S(\E)$ and $\phi\in\Hom_\F(Q,S)$ such that $T_0^\perp\phi=T_0^\perp$. Then
\[(Q\cap T_0)\phi\leq T_0.\]
\end{lemma}
\begin{proof}
We may choose $S_0$ such that $N_S(\E)\leq S_0$. By Lemma~\ref{L:FrattiniApply}, there exist then $\psi\in\Hom_{\F\cap\F_0}(Q,\tT Q)$ and $\alpha\in\Hom_{N_\F(\tT)}(Q\tT,S)$ with $\phi=\psi\alpha$. Theorem~\ref{T:bNLH}(d) implies that $T_0$ and $T_0^\perp$ are strongly closed in $\F_0$. Thus,
\[(Q\cap T_0)\psi\leq T_0\]
and $T_0^\perp\psi=T_0^\perp$. This yields $T_0^\perp\alpha=T_0^\perp\psi\alpha=T_0^\perp\phi=T_0^\perp$. By Lemma~\ref{L:NinNLtTofH}(b), $G$ realizes $N_\F(\tT)$. Hence, $\alpha$ is realized as a conjugation homomorphism by an element of $G$ which then lies in $N_G(T_0^\perp)$ as $T_0^\perp\alpha=T_0^\perp$. Recall from Lemma~\ref{L:NinNLtTofH}(e) that $N_{G}(T_0^\perp)=N_{G}(T_0)$. Hence, it follows that $T_0\alpha=T_0$ and thus $(Q\cap T_0)\phi=(Q\cap T_0)\psi\alpha\leq T_0\alpha=T_0$.
\end{proof}
\begin{lemma}\label{L:NormalizeT0perp}
Let $Q\in \E^{cr}$ and $\phi\in\Hom_\F(Q\tT,S)$ with $Q\phi=Q$. Then $T_0\phi=T_0$ and $T_0^\perp\phi=T_0^\perp$.
\end{lemma}
\begin{proof}
If $\K\in\Comp(\H)$, then $\F_{S\cap \K}(\K)\subn \E$ by Background Theorem~\ref{bT:2}. Hence, by Lemma~\ref{L:CentricRadicalIntersection}, $Q\cap \K\in \F_{S\cap \K}(\K)^{cr}$. In particular, $Q\cap \K\not\leq Z(\K)$ as $Z(\K)\neq \K\cap S$ by \cite[Lemma~7.10]{Henke:Regular}. On the other hand, if $\K\in \Comp(\L)\backslash\Comp(\H)$, then $\H\subseteq C_\L(\K)$ by \cite[Lemma~3.5, Lemma~11.17]{Henke:Regular}. Hence, as $Q\leq T\subseteq\H$, we have in this case $Q\cap \K\subseteq \H\cap\K\subseteq C_\K(\K)=Z(\K)$. This shows
\[\Comp(\H)=\{\K\in\Comp(\L)\colon \K\cap Q\not\leq Z(\K)\}.\]
As $\tT$ is strongly $\F$-closed, $\phi$ is a morphism in $N_\F(\tT)=\F_S(G)$. Hence, there exists $f\in G$ with $\phi=c_f|_{Q\tT}$. Then conjugation by $f$ induces by Lemma~\ref{L:BasisNLtT} an automorphism of $\L$ and thus permutes by \cite[Lemma~11.12]{Henke:Regular} the components of $\L$. Moreover, for every $\K\in\Comp(\L)$, $c_f$ induces an isomorphism from $\K$ to $\K^f$ so that $Z(\K)^f=Z(\K^f)$. As $Q^f=Q\phi=Q$, we have also $(\K\cap Q)^f=\K^f\cap Q$. So for every $\K\in\Comp(\L)$, we have $\K\cap Q\not\leq Z(\K)$ if and only if $\K^f\cap Q\not\leq Z(\K^f)$. Therefore, $f$ permutes the elements of $\Comp(\H)$ via conjugation and thus also the elements of $\Comp(\L)\backslash \Comp(\H)$. This implies $E(\H)^f=E(\H)$ and $\M^f=\M$. As $\phi=c_f|_{Q\tT}$, the assertion follows.
\end{proof}
\begin{lemma}\label{L:GCollect}
Let $R\leq S_0$ and set
\[\G:=(\E R)_{\F_0}\]
Then the following hold:
\begin{itemize}
\item [(a)] $\G$ is saturated, $\E\unlhd \F$, $O^p(\G)=O^p(\E)$ and $(\H R,\delta(\G),TR)$ is a regular locality over $\G$ with $E(\H R)=E(\H)$. In particular, $E(\G)=E(\E)$. Moreover, if $R\leq N_S(\E)$, then $\G=\F_{T R}(\H R)\subseteq \F$.
\item [(b)] If $P\leq TR$ with $P\cap T\in\delta(\E)$, then $P\in\delta(\G)$.
\item [(c)] Suppose $P\in\delta(\G)$. Then $N_{\H R}(P)$ and $N_\H(P)$ are subgroups of $\L$. In particular,
\[\Aut_\H(P):=\{c_h\colon h\in N_\H(P)\}\]
is a subgroup of $\Aut(P)$. Moreover, $O^p(\Aut_\G(P))=O^p(\Aut_\H(P))$.
\item [(d)] If $P\leq TR\cap N_S(\E)$ such that $P\cap T\in\E^{cr}$, then $P\cap T\in\delta(\E)$ and thus $P\in\delta(\G)$. Moreover, \[O^p(\Aut_\G(P))=O^p(\Aut_\H(P))=\Ac_{\F_0,\E}(P)=\Ac_{N_\F(\E),\E}(P).\]
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} We use as before that $\F_0$ is well-defined and saturated, $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$ with $T=S_0\cap \H$, $\H\unlhd\bN_\L(\H)$ and $\E=\F_T(\H)\unlhd\F_0$. By Proposition~\ref{P:ERnormal}, $\G$ is saturated, $\E\unlhd \G$, $O^p(\G)=O^p(\E)$ and our notion of a product of a normal subsystem with a $p$-subgroup coincides with the one used in \cite{Chermak/Henke}. Thus, it follows from \cite[Theorem~D(a)]{Chermak/Henke} that $\G$ is saturated and $(\H R,\delta(\G),TR)$ is a regular locality over $\G$. In particular, if $R\leq N_S(\E)$, then
\[\G=\F_{T P}(\H P)\subseteq \F_S(\L)=\F.\]
For an arbitrary subgroup $R\leq S_0$, $\H$ is a partial normal subgroup of $\H P$ of $p$-power index. Thus, by \cite[Lemma~11.14]{Henke:Regular}, $E(\H P)=E(\H)$. It follows now from Background Theorem~\ref{bT:2} applied twice (once with $(\H,T,\E)$ and once with $(\H P,TP,\G)$ in place of $(\L,S,\F)$) that $E(\E)=\F_{T_0}(E(\H))=E(\G)$. This proves (a).
\smallskip
\textbf{(b)} As $E(\E)=E(\G)$, part (b) follows from Lemma~\ref{L:EE=EF} applied with $\G$ in place of $\F$.
\smallskip
\textbf{(c)} Let $P\in\delta(\G)$. As $(\H R,\delta(\G),TR)$ is a regular locality over $\G$, it follows that $N_{\H R}(P)$ is a subgroup of $\H R$ and thus of $\L$. Moreover,
\[\Aut_\G(P)=\{c_g\colon g\in N_{\H R}(P)\}.\]
Since $\H\unlhd \H R$, $N_\H(P)=\H\cap N_{\H R}(P)$ is a normal subgroup of $N_{\H R}(P)$. In particular, $\Aut_\H(P)$ is a subgroup of $\Aut(P)$. By \cite[Lemma~6.1]{Henke:2020} (applied with $(\H P,\delta(\G),TP)$ in place of $(\L,\Delta,S)$), we have $O^p(N_{\H P}(P))=O^p(N_\H(P))$. Observe that every $p^\prime$-element of $\Aut_\H(P)$ can be realized by conjugation with a $p^\prime$-element of $N_\H(P)$, and similarly every $p^\prime$-automorphism of $\Aut_\G(P)$ can be realized by conjugation with a $p^\prime$-element of $N_{\H P}(P)$. This implies
\[O^p(\Aut_\H(P))=\{c_h\colon h\in O^p(N_\H(P))\}=\{c_h\colon h\in O^p(N_{\H R}(P))\}=O^p(\Aut_\G(P)).\]
Hence (c) holds.
\smallskip
\textbf{(d)} By \cite[Lemma~10.4]{Henke:Regular}, $\E^{cr}\subseteq \delta(\E)$. Moreover, if $P\leq TR$ with $P\cap T=\E^c$, then \cite[Lemma~4.7]{Henke:2013} gives $O^p(\Aut_\G(P))=\Ac_{\F_0,\E}(P)$. Hence, supposing $P$ is a subgroup of $TR$ with $P\cap T\in\E^{cr}$, it follows from (b) and (c) that $O^p(\Aut_\H(P))=O^p(\Aut_\G(P))=\Ac_{\F_0,\E}(P)$. Assume now in addition that $P\leq N_S(\E)$. Then $\Ac_{\F_0,\E}(P)\leq \Aut_\H(P)\leq \Aut_\F(P)$. As $N_\F(\E)=\F\cap\F_0$ by Lemma~\ref{L:NFEbNLH}(b), it follows $\Ac_{\F_0,\E}(P)=\Ac_{N_\F(\E),\E}(P)$. This proves (d).
\end{proof}
We use from now on without further reference that, for all $P\leq N_S(\E)$ with $P\cap T\in\E^{cr}$, the normalizer $N_\H(P)$ is a subgroup of $\L$ and
\[\Aut_\H(P):=\{c_h\colon h\in N_\H(P)\}\]
is a subgroup of $\Aut(P)$. Notice that this is true by Lemma~\ref{L:GCollect}(b),(c) (applied e.g. for a Sylow subgroup $S_0$ of $N_G(\H)$ containing $N_S(\H)=N_S(\E)$, and with $R=N_S(\E)$).
\begin{lemma}\label{L:AutHPsubnormal}
Let $P\leq N_S(\E)$ such that $P\cap T\in\E^{cr}$. Then $\Aut_\H(P)\subn \Aut_{N_\F(T_0^\perp)}(P)$.
\end{lemma}
\begin{proof}
By Lemma~\ref{L:BasicT0perp}(a), $\H\subseteq C_\L(\M)\subseteq C_\L(T_0^\perp)$. In particular,
\[\Aut_\H(P)\leq \Aut_{N_\F(T_0^\perp)}(P).\]
As $\F_{T_0}(E(\H))=E(\E)\unlhd \E$ by Background Theorem~\ref{bT:2}, it follows from Lemma~\ref{L:CentricRadicalIntersection} and \cite[Lemma~10.4]{Henke:Regular} that
\[P_0:=P\cap T_0\in E(\E)^{cr}\subseteq \delta(E(\E)).\]
By Lemma~\ref{L:BasicT0perp}(a), $E(\H)$ and $\M$ are partial normal subgroups of $E(\L)$ with $E(\L)=E(\H)\M$ and $E(\H)\subseteq \H\subseteq C_\L(\M)$. The latter property implies by \cite[Lemma~3.5]{Henke:Regular} that $E(\H)$ commutes with $\M$ in the sense of \cite[Definition~2]{Henke:Regular}. Thus, it follows from \cite[Lemma~10.15(d)]{Henke:Regular} that
\[X:=P_0T_0^\perp\in\delta(\F_{S\cap E(\L)}(E(\L))).\]
By Background Theorem~\ref{bT:2}, we have $\F_{S\cap E(\L)}(E(\L))=E(\F)$ and by \cite[7.22]{Chermak/Henke}, $\delta(E(\F))\subseteq\delta(\F)$. \footnote{Alternatively, a direct argument involving Lemma~10.2, Lemma~10.6, Lemma~10.15(d) and Lemma~11.5 in \cite{Henke:Regular} could be given to show $\delta(\F_{S\cap E(\L)}(E(\L)))\subseteq\delta(\F)$.} We can conclude that \[X\in\Delta=\delta(\F).\]
In particular, $N_\L(X)$ is a group.
\smallskip
It follows from Lemma~\ref{L:NinNLtTofH}(e) that $P\leq N_S(\E)=N_S(\H)\leq N_S(T_0^\perp)=N_S(T_0)$. Thus, $P\leq N_S(P_0)$ and
\[P\leq N_S(X).\]
Notice moreover that
\[N_\H(P)\leq N_\H(X)\]
as $P_0=P\cap E(\H)$ is normalized by $N_\H(P)$ and $\H\subseteq C_\L(T_0^\perp)$.
\smallskip
Let now $\H=\H_0\unlhd\H_1\unlhd\cdots\unlhd\H_n=\L$ be a subnormal series of $\H$ in $\L$. Then
\[N_\H(P)=N_{N_\H(X)}(P)=N_{N_{\H_0}(X)}(P)\unlhd N_{N_{\H_1}(X)}(P)\unlhd\cdots \unlhd N_{N_{\H_n}(X)}(P)=N_{N_\L(X)}(P)\]
is a subnormal series of $N_\H(P)$ in $N_{N_\L(X)}(P)$. This leads to a subnormal series of $\Aut_\H(P)$ in
\[\Aut_{N_\L(X)}(P)=\{c_g|_P\colon g\in N_{N_\L(X)}(P)\}.\]
Thus, it remains only to argue that $\Aut_{N_\F(T_0^\perp)}(P)\leq \Aut_{N_\L(X)}(P)$. By \cite[Lemma~3.10(b)]{Grazian/Henke}, $N_\F(X)=\F_{N_S(X)}(N_\L(X))$ and thus
\[\Aut_{N_\L(X)}(P)=\Aut_{N_\F(X)}(P).\]
Let $\phi\in\Aut_{N_\F(T_0^\perp)}(P)$. Then $\phi$ extends to $\hat{\phi}\in\Hom_\F(PT_0^\perp,S)$ with $T_0^\perp\hat{\phi}=T_0^\perp$. By Lemma~\ref{L:QcapT0phi}, we have $P_0\hat{\phi}\leq T_0$ and thus $P_0\hat{\phi}\leq P\cap T_0=P_0$. This implies $X\hat{\phi}=X$ and proves thus $\Aut_{N_\F(T_0^\perp)}(P)\leq \Aut_{N_\F(X)}(P)=\Aut_{N_\L(X)}(P)$ as required.
\end{proof}
\begin{lemma}\label{L:ConjAcbyNFE}
Let $P\leq N_S(\E)$ and $\chi\in\Hom_{N_\F(\E)}(P,N_S(\E))$. Then
\[\chi^{-1}\Ac_{\F,\E}(P)\chi=\Ac_{\F,\E}(P\chi)\mbox{ and }\chi^{-1}\Ac_{N_\F(\E),\E}(P)\chi=\Ac_{N_\F(\E),\E}(P\chi).\]
In particular, $\Ac_{\F,\E}(P)=\Ac_{N_\F(\E),\E}(P)$ if and only if $\Ac_{\F,\E}(P\chi)=\Ac_{N_\F(\E),\E}(P\chi)$.
\end{lemma}
\begin{proof}
Recall that $\E$ is $N_\F(\E)$-invariant by Lemma~\ref{L:NFEbNLH}(c). In particular, $\E$ is weakly $N_\F(\E)$-invariant by Lemma~\ref{L:Finvariant}(a). We can conclude that $(P\cap T)\chi=(P\chi)\cap T$ and
\[\chi^{-1}\Aut_\E(P\cap T)\chi=\Aut_\E((P\chi)\cap T).\]
Since the map $\Aut_\F(P)\rightarrow \Aut_\F(P\chi),\delta\mapsto \chi^{-1}\delta\chi$ is an isomorphism which restricts to an isomorphism from $\Aut_{N_\F(\E)}(P)\rightarrow \Aut_{N_\F(\E)}(P)$, the assertion is now easy to check.
\end{proof}
\begin{lemma}\label{L:Findchi}
Let $P\leq N_S(\E)\cap S_0$ such that $P\cap T\in\E^{cr}$. Then there exists $\chi\in\Hom_{N_\F(\E)}(P,N_S(\E))$ such that $P\chi\leq S_0$ and the following hold:
\begin{itemize}
\item[(i)] $O_p(N_{\Aut_\F(P\chi)}((P\chi)\cap T))\leq \Aut_{N_\F(T_0^\perp)}(P\chi)$;
\item[(ii)] $(P\chi)\cap T$ is fully normalized in $(\E(P\chi))_{\F_0}$ and an element of $\E^{cr}$.
\end{itemize}
\end{lemma}
\begin{proof}
To ease notation set $Q:=P\cap T$. We show first
\begin{eqnarray}\label{E:Getphi}
\mbox{There exists $\phi\in\Hom_\F(P,S)$ such that $N_{\Aut_S(P\phi)}(Q\phi)\in\Syl_p(N_{\Aut_\F(P\phi)}(Q\phi))$}.
\end{eqnarray}
For the proof fix $\beta\in\Hom_\F(P,S)$ such that $P\beta$ is fully $\F$-normalized. Then by the Sylow axiom \cite[Proposition~I.2.5]{Aschbacher/Kessar/Oliver:2011}, we have $\Aut_S(P\beta)\in\Syl_p(\Aut_\F(P\beta))$. As $H:=N_{\Aut_\F(P\beta)}(Q\beta)$ is a subgroup of $\Aut_\F(P\beta)$, there exists $\gamma\in\Aut_\F(P\beta)$ such that $H^\gamma\cap \Aut_S(P\beta)\in\Syl_p(H^\gamma)$. Note that $P\beta\gamma=P\beta$ and $H^\gamma=N_{\Aut_\F(P\beta)}(Q\beta\gamma)$. Hence \eqref{E:Getphi} holds for $\phi:=\beta\gamma$.
\smallskip
Let now $\phi$ be as in \eqref{E:Getphi}. Then in particular,
\[O_p(N_{\Aut_\F(P\phi)}(Q\phi))\leq \Aut_S(P\phi)\leq \Aut_{N_\F(\tT)}(P\phi).\]
By Lemma~\ref{L:FrattiniApply}, there exist $\psi\in\Hom_{N_\F(\E)}(P,\tT P)\cap \Hom_{\F_0}(P,\tT P)$ and $\alpha\in\Hom_{N_\F(\tT)}(P\psi,S)$ such hat $\phi=\psi\alpha$. Notice that we have an isomorphism from $N_{\Aut_\F(P\phi)}(Q\phi)$ to $N_{\Aut_\F(P\psi)}(Q\psi)$ defined by $\delta\mapsto \alpha\delta\alpha^{-1}$. Hence,
\begin{equation}\label{E:OpinNtT}
O_p(N_{\Aut_\F(P\psi)}(Q\psi))=\alpha\,O_p(N_{\Aut_\F(P\phi)}(Q\phi))\alpha^{-1}\leq \Aut_{N_\F(\tT)}(P\psi).
\end{equation}
Since $\E\unlhd\F_0$ and $\psi$ is a morphism in $\F_0$, it follows from $Q\in\E^{cr}$ and \cite[Lemma~2.6]{Chermak/Henke} that $Q\psi\in\E^{cr}$. Hence, using \eqref{E:OpinNtT} and applying Lemma~\ref{L:NormalizeT0perp} with $Q\psi$ in place of $Q$, we can conclude that
\begin{equation}\label{E:OpinNT0perp}
O_p(N_{\Aut_\F(P\psi)}(Q\psi))\leq \Aut_{N_\F(T_0^\perp)}(P\psi).
\end{equation}
Observe now that $P\psi\leq N_S(\E)=N_S(\H)\leq N_S(T_0^\perp)$ by Lemma~\ref{L:finNLtT} and Lemma~\ref{L:NinNLtTofH}(e). Moreover, by Lemma~\ref{L:BasicT0perp}(a), $\H\subseteq C_\L(\M)\subseteq C_\L(T_0^\perp)$. As $N_\L(T_0^\perp)$ is by \cite[Lemma~5.5]{Henke:2015} a partial subgroup of $\L$, it follows $\H (P\psi)\subseteq N_\L(T_0^\perp)$. Using Lemma~\ref{L:GCollect}(a), we see now that
\[\G:=(\E (P\psi))_{\F_0}=\F_{T(P\psi)}(\H(P\psi))\subseteq N_\F(T_0^\perp)\]
is saturated. Thus, by \cite[Lemma~I.2.6(c)]{Aschbacher/Kessar/Oliver:2011}, there exists $\rho\in\Hom_\G(N_{T(P\psi)}(Q\psi),T(P\psi))$ such that $Q\psi\rho$ is fully $\G$-normalized. Notice that $Q\psi\unlhd P\psi$ and thus $P\psi\rho$ is defined. Moreover, by definition of $\G$, we have $T(P\psi)=T(P\psi\rho)$ and thus $\G=(\E (P\psi\rho))_{\F_0}$. Since $\E\unlhd \F_0$ and $\psi$ and $\rho$ are morphisms in $\F_0$, we have that $(P\psi\rho)\cap T=Q\psi\rho$ is fully $\G$-normalized and moreover an element of $\E^{cr}$ by \cite[Lemma~2.6]{Chermak/Henke}. Since $\rho$ is a morphism in $\G$ and $\G\subseteq N_\F(T_0^\perp)$, it follows from \eqref{E:OpinNT0perp} that
\[O_p(N_{\Aut_\F(P\psi\rho)}((P\psi\rho)\cap T))=\rho^{-1}O_p(N_{\Aut_\F(P\psi)}(Q\psi))\rho\leq \Aut_{N_\F(T_0^\perp)}(P\psi\rho).\]
Hence, the assertion follows for $\chi=\psi\rho$.
\end{proof}
\begin{lemma}\label{L:CrucialLemma}
Let $P\leq N_S(\E)\cap S_0$ such that $P\cap T\in\E^{cr}$. Then
\[\Ac_{\F,\E}(P)=\Ac_{N_\F(\E),\E}(P)=\Ac_{\F_0,\E}(P)=O^p(\Aut_\H(P)).\]
\end{lemma}
\begin{proof}
Set $\G:=(\E P)_{\F_0}$. By Lemma~\ref{L:GCollect}, $\G$ is saturated, $(\H P,\delta(\G),TP)$ is a regular locality over $\G$, $P\in\delta(\G)$ and
\begin{equation}\label{E:AcEqualities}
O^p(\Aut_\G(P))=\Ac_{N_\F(\E),\E}(P)=\Ac_{\F_0,\E}(P)=O^p(\Aut_\H(P)).
\end{equation}
Thus, it remains to show $\Aut_{N_\F(\E),\E}(P)=\Ac_{\F,\E}(P)$. Set
\[A:=N_{\Aut_\F(P)}(P\cap T).\]
Replacing $P$ by a suitable $N_\F(\E)$-conjugate, Lemma~\ref{L:ConjAcbyNFE} and Lemma~\ref{L:Findchi} allow us to assume that
\begin{equation}\label{E:Assumption}
P\cap T\mbox{ is fully $\G$-normalized and }O_p(A)\leq \Aut_{N_\F(T_0^\perp)}(P).
\end{equation}
Clearly, $\Ac_{N_\F(\E),\E}(P)\leq \Ac_{\F,\E}(P)$, so it remains to show the converse inclusion. For the proof let $\phi\in\Ac_{\F,\E}(P)$ be a $p^\prime$-automorphism. By \eqref{E:AcEqualities}, it remains to show that $\phi\in O^p(\Aut_\H(P))$.
\smallskip
Observe that $\phi_0:=\phi|_{P\cap T}\in\Aut_\E(P\cap T)$ is a $p^\prime$-automorphism. Set $\hat{P}:=PC_{TP}(P\cap T)$. We show first that
\begin{equation}\label{E:psihat}
\mbox{there exists $\beta\in\Aut_\G(\hat{P})$ such that $\beta|_{P\cap T}=\phi_0$.}
\end{equation}
Notice that $\phi_0$ is a morphism in $\G$ and $C_{TP}(P\cap T)\leq N_{\phi_0}^\G$ (where we use the notation introduced in \cite[Definition~I.2.2]{Aschbacher/Kessar/Oliver:2011}). Using e.g. \cite[Remark~2.1]{Henke:2013}), observe furthermore that
\[\phi_0^{-1}\Aut_P(P\cap T)\phi_0=\Aut_{P\phi}(P\cap T)=\Aut_P(P\cap T)\]
and hence $P\leq N_{\phi_0}^{\G}$. Therefore, $\hat{P}\leq N_{\phi_0}^\G$. Since $\G$ is saturated and $P\cap T$ is by \eqref{E:Assumption} fully $\G$-normalized, it follows that $\phi_0$ extends to $\beta\in\Hom_\G(\hat{P},TP)$ (cf. \cite[Proposition~I.2.5]{Aschbacher/Kessar/Oliver:2011}). As $\phi$ is a $p^\prime$-automorphism and $[P,\phi]\leq P\cap T$, using a property of coprime action (see e.g. \cite[8.2.7(a)]{KS}), we can conclude that $P=(P\cap T)C_P(\phi)$. As $\phi|_{P\cap T}=\phi_0=\beta|_{P\cap T}$ and $P\cap T\unlhd P$, it is a consequence of \cite[Lemma~2.4]{Henke:2013} that
\[[C_P(\phi),\beta]:=\<x^{-1}(x\beta)\colon x\in C_P(\phi)\>\leq C_{TP}((P\cap T)\beta)=C_{TP}(P\cap T)\leq \hat{P}.\]
Notice that $\beta$ normalizes $P\cap T$ and thus also $C_{TP}(P\cap T)$. As $\hat{P}=(P\cap T)C_P(\phi)C_{TP}(P\cap T)$, we can conclude that $\hat{P}\beta=\hat{P}$. Hence, \eqref{E:psihat} holds.
\smallskip
We choose now $\beta$ as in \eqref{E:psihat}. Since $\phi_0$ is a $p^\prime$-element, replacing $\beta$ by a suitable power of itself, we may assume that $\beta$ is a $p^\prime$-element of $\Aut_\G(\hat{P})$. As $\G=(\E P)_{\F_0}$, it follows from \cite[Lemma~4.7]{Henke:2013} that $\beta\in O^p(\Aut_\G(\hat{P}))\leq \Ac_{\F_0,\E}(\hat{P})$. In particular,
\[[\hat{P},\beta]\leq \hat{P}\cap T.\]
As above, it follows from \cite[Lemma~2.4]{Henke:2013} that $[C_P(\phi),\beta]\leq C_{TP}(P\cap T)$. Hence,
\[[C_P(\phi),\beta]\leq C_T(P\cap T)\leq P\cap T \]
since $P\cap T\in\E^c$. As $P=(P\cap T)C_P(\phi)$ and $(P\cap T)\beta=P\cap T$, this yields $P\beta=P$ and thus, by \eqref{E:AcEqualities},
\[\psi:=\beta|_P\in O^p(\Aut_\G(P))=O^p(\Aut_\H(P))=\Ac_{N_\F(\E),\E}(P).\]
\smallskip
Notice that $\psi|_{P\cap T}=\phi_0=\phi|_{P\cap T}$ and thus $(\phi\psi^{-1})|_{P\cap T}=\id_{P\cap T}$. Moreover, $[P,\psi]\leq P\cap T$ and $[P,\phi]\leq P\cap T$, so $[P,\phi\psi^{-1}]\leq P\cap T$. By a property of coprime action (see e.g. \cite[8.2.7(b)]{KS}), $C_A(P/P\cap T)\cap C_A(P\cap T)$ is a normal $p$-subgroup of $A$ and thus contained in $O_p(A)$. Hence, $\phi\psi^{-1}\in O_p(A)$ and thus
\[\phi\in X:=O_p(A)O^p(\Aut_\H(P)).\]
Notice that $\Aut_\H(P)\leq A$ normalizes $O_p(A)$ and therefore $X$ is a group with $O_p(A)\leq O_p(X)$. By Lemma~\ref{L:AutHPsubnormal}, $\Aut_\H(P)\subn \Aut_{N_\F(T_0^\perp)}(P)$ and thus $O^p(\Aut_\H(P))\subn \Aut_{N_\F(T_0^\perp)}(P)$. By assumption \eqref{E:Assumption}, we have $O_p(A)\leq \Aut_{N_\F(T_0^\perp)}(P)$. Hence, $O^p(\Aut_\H(P))$ is subnormal in $X$. By Lemma~\ref{L:OpGinNormalizer}, this implies that $O_p(A)\leq O_p(X)$ normalizes $O^p(\Aut_\H(P))$. In particular, $O^p(\Aut_\H(P))$ is normal $X$. As $O^p(\Aut_\H(P))$ has $p$-power index in $X$, this implies $O^p(X)=O^p(\Aut_\H(P))$. As $\phi$ is a $p^\prime$-automorphism of $P$, it follows $\phi\in O^p(X)=O^p(\Aut_\H(P))$ as required.
\end{proof}
\begin{prop}\label{P:DinNFE}
Let $\mD$ be a saturated subsystem of $\F$ with $\E\unlhd\mD$. Then $\mD\subseteq N_\F(\E)$.
\end{prop}
\begin{proof}
Let $R\leq S$ such that $\mD$ is a fusion system over $R$. As $\E\unlhd\mD$, we have $R\leq N_S(\E)$. Lemma~\ref{L:FusionSystemFrattini} gives
\[\mD=\<(\E R)_{\mD}, N_{\mD}(T)\>.\]
Notice that we may choose $S_0$ always such that $N_S(\E)\leq S_0$. For $P\leq TR$ with $P\cap T\in\E^{cr}$, it follows thus from Lemma~\ref{L:CrucialLemma} that $\Ac_{\mD,\E}(P)\leq \Ac_{\F,\E}(P)\leq \Aut_{N_\F(\E)}(P)$. Hence, by definition of $(\E R)_{\mD}$, we have $(\E R)_{\mD}\subseteq N_\F(\E)$.
\smallskip
A morphism in $N_{\mD}(T)$ is the restriction of a morphism $\phi\in\Hom_{\mD}(X,Y)$ where $X,Y\leq R$ with $T\leq X\cap Y$ and $T\phi=T$. As $\E\unlhd \mD$, for any such $\phi$, we have $\phi|_T\in\Aut(\E)$. Thus $\phi$ is a morphism in $N_\F(\E)$ by Lemma~\ref{L:inAutEinNFE}. This shows $N_{\mD}(T)\subseteq N_\F(\E)$ and therefore $\mD\subseteq N_\F(\E)$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{T:MainER}]
Choose $S_0$ such that $N_S(\E)=N_S(\H)\leq S_0$. It follows then from Lemma~\ref{L:CrucialLemma} and the definitions of $\E R:=(\E R)_\F$ and $(\E R)_{\F_0}$ that $(\E R)_\F=(\E R)_{\F_0}$. Hence, by Lemma~\ref{L:GCollect}(a), $\E R$ is saturated, $\E\unlhd \E R$, $O^p(\E R)=O^p(\E)$ and $(\H R,\delta(\E R),TR)$ is a regular locality over $\E R$.
\smallskip
Fix now a saturated subsystem $\mD$ of $\F$ over $T R$ with $O^p(\mD)=O^p(\E)$. If $\E\unlhd \mD$, then it follows from Proposition~\ref{P:DinNFE} that $\mD$ is contained in $N_\F(\E)\subseteq \F_0$ and is thus equal to $(\E R)_{\F_0}=\E R$ by Proposition~\ref{P:ERnormal} applied with $\F_0$ in place of $\F$. Hence,
\begin{equation}\label{E:ReductionEunlhdD}
\mD=\E R\mbox{ if }\E\unlhd\mD.
\end{equation}
Using this property with $(O^p(\E),T)$ in place of $(\E,R)$ (along with the fact that $O^p(\E)=O^p(O^p(\E))$ by \cite[Lemma~2.29]{Chermak/Henke}) we can conclude the following statement.
\begin{equation}\label{E:ApplyOupperpE}
\mbox{Every saturated subsystem $\mD'$ of $\F$ over $T$ with $O^p(\mD')=O^p(\E)$ equals $(O^p(\E) T)_\F$.}
\end{equation}
As $T\unlhd TR$, by \cite[Theorem~I.7.4]{Aschbacher/Kessar/Oliver:2011}, there is a unique normal subsystem $\E'$ of $\mD$ over $T$ of $p$-power index. Then $O^p(\E')=O^p(\mD)=O^p(\E)$ by \cite[Lemma~2.29]{Chermak/Henke} and assumption. Hence, applying \eqref{E:ApplyOupperpE} twice, we obtain $\E=(O^p(\E)T)_\F=\E'\unlhd\mD$. By \eqref{E:ReductionEunlhdD}, this implies $\mD=\E R$.
\end{proof}
\subsection{Conjugates of $\E$} In this subsection we will look at properties of $\F$-conjugates of $\E$ and complete the proof of Theorem~\ref{T:FusionNormalizer}. Many results we prove below are not needed in the proofs of the main theorems, but seem interesting on their own.
\smallskip
We continue to assume Hypotheses~\ref{H:main} and \ref{H:S0F0}. Recall from the introduction that
\[\E^\F:=\{\E^\phi\colon \phi\in\Hom_\F(T,S)\}\]
denotes the set of \emph{$\F$-conjugates of $\E$}.
\begin{lemma}\label{L:ConjEandT}
We have $\E^\F=\E^{N_\F(\tT)}$ and $T^\F=T^{N_\F(\tT)}$.
\end{lemma}
\begin{proof}
As $T\leq N_S(\E)$, it follows from Lemma~\ref{L:FrattiniApply} that there exists $\psi\in\Hom_{N_\F(\E)}(T,\tT T)$ and $\alpha\in\Hom_{N_\F(\tT)}(T\psi,S)$ such that $\phi=\psi\alpha$. By Lemma~\ref{L:NFEbNLH}(c), $\E$ is $N_\F(\E)$-invariant. Hence, $T\psi=T$ and $\E^\psi=\E$, which yields $T\phi=T\psi\alpha=T\alpha\in T^{N_\F(\tT)}$ and similarly $\E^\phi=(\E^\psi)^\alpha=\E^\alpha\in\E^{N_\F(\tT)}$. The assertion follows.
\end{proof}
\begin{lemma}\label{L:ConjH}~
\begin{itemize}
\item[(a)] Let $g\in G$ such that $T^g\leq S$. Then $\H^g\subn \L$ with $\H^g\cap S=T^g$. Moreover, setting $\phi:=c_g|_T$, we have $\E^\phi=\F_{T^g}(\H^g)\subn \F$.
\item[(b)] For every $g\in G$, the conjugation automorphism $c_g\in\Aut(\L)$ restricts to isomorphisms from $N_G(\H)$ to $N_G(\H^g)$ and from $\bN_\L(\H)$ to $\bN_\L(\H^g)$. For $S_0\in\Syl_p(N_G(\H))$ we have $S_0^g\in\Syl_p(N_G(\H^g))$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} By Lemma~\ref{L:BasisNLtT}, $c_g\in\Aut(\L)$ and thus $\H^g\subn\L$. It follows moreover from Corollary~\ref{C:AutLFusion} that $\E^\phi=\F_{T^g}(\H^g)$. By Background Theorem~\ref{bT:2}, we have $\F_{T^g}(\H^g)\subn\F$.
\smallskip
\textbf{(b)} Using again $c_g\in\Aut(\L)$, one observes easily that $N_G(\H)c_g=N_G(\H^g)$, i.e. $c_g$ induces an isomorphism from $N_G(\H)$ to $N_G(\H^g)$, which then takes a Sylow $p$-subgroup of $N_G(\H)$ to a Sylow $p$-subgroup of $N_G(\H^g)$. By \cite[Lemma~11.12]{Henke:Regular}, we have also $E(\L)c_g=E(\L)$. As $\bN_\L(\H)=E(\L)N_G(\H)$ and $\bN_\L(\H^g)=E(\L)N_G(\H^g)$ the assertion follows.
\end{proof}
\begin{lemma}\label{L:ConjugatesSubnormal}
Every $\F$-conjugate of $\E$ is a subnormal subsystem of $\F$.
\end{lemma}
\begin{proof}
As $N_\F(\tT)=\F_S(G)$, this follows from Lemma~\ref{L:ConjEandT} and Lemma~\ref{L:ConjH}.
\end{proof}
Recall that we call $\E$ \emph{fully $\F$-normalized}, if
\[|N_S(\E)|\geq |N_S(\tE)|\mbox{ for all }\tE\in\E^\F.\]
Notice that $\E^\F=(\E')^\F$ for every $\E'\in\E^\F$. Hence, we can always choose a fully $\F$-normalized $\F$-conjugate of $\E$.
\begin{definition}
The subsystem $\E$ is called \emph{fully $\F$-automized} if
\[\Aut_S(T)\cap \Aut(\E)\in\Syl_p(\Aut_\F(T)\cap \Aut(\E)).\]
\end{definition}
We will prove below that $\E$ is fully $\F$-normalized if and only if $T$ is fully $\F$-centralized and $\E$ is fully $\F$-automized.
\begin{lemma}\label{L:EFullyAutomized} Suppose $N_S(\E)\leq S_0$. Then the following hold.
\begin{itemize}
\item [(a)] $\Aut_{S_0}(T)$ is a Sylow $p$-subgroup of $\Aut_{\F_0}(T)$ and of
\[\Aut_{\F_0}(T)\cap \Aut_\F(T)=\Aut_{N_\F(\E)}(T)=\Aut_\F(T)\cap\Aut(\E).\]
\item [(b)] The following properties are equivalent.
\begin{itemize}
\item [(i)] $\E$ is fully automized;
\item [(ii)] $T$ is fully $N_\F(\E)$-automized.
\item [(iii)] $\Aut_{N_S(\E)}(T)=\Aut_{S_0}(T)$.
\end{itemize}
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} As $\F_0:=\F_{S_0}(\bN_\L(\H))$ is saturated and $T\unlhd S_0$, we have $\Aut_{S_0}(T)\in\Syl_p(\Aut_{\F_0}(T))$. Notice moreover that $\Aut_{S_0}(T)\leq \Aut_\F(T)$ since $\F=\F_S(\L)$. Hence, $\Aut_{S_0}(T)$ is also a Sylow $p$-subgroup of $\Aut_{\F_0}(T)\cap \Aut_\F(T)$. It follows from Lemma~\ref{L:NFEbNLH}(b) that
\[\Aut_\F(T)\cap \Aut_{\F_0}(T)=\Aut_{\F\cap\F_0}(T)=\Aut_{N_\F(\E)}(T).\]
By Lemma~\ref{L:NFEbNLH}(c), $\E$ is $N_\F(\E)$ invariant and so $\Aut_{N_\F(\E)}(T)\leq \Aut_\F(T)\cap \Aut(\E)$. The converse inclusion follows from Lemma~\ref{L:inAutEinNFE}. So (a) holds.
\smallskip
\textbf{(b)} Notice that $\Aut_S(T)\cap \Aut(\E)=\Aut_{N_S(\E)}(T)$. Hence, by (a), property (i) is equivalent to $\Aut_{N_S(\E)}(T)\in\Syl_p(\Aut_{N_\F(\E)}(T))$, i.e. to (ii). Assume now $N_S(\E)\leq S_0$. Then $\Aut_{N_S(\E)}(T)\leq \Aut_{S_0}(T)$. So (ii) is by (a) equivalent to $\Aut_{N_S(\E)}(T)=\Aut_{S_0}(T)$.
\end{proof}
\begin{lemma}\label{L:Tfullycentralized}~
\begin{itemize}
\item [(a)] Suppose $C_S(T)\leq S_0$. Then $T$ is fully $\F$-centralized if and only if $C_S(T)=C_{S_0}(T)$.
\item [(b)] If $T$ is fully centralized and $N_S(\E)\leq S_0$, then $C_\F(T)=C_{\F_0}(T)$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} It follows from Lemma~\ref{L:finNLtT} that $C_G(T)\leq X:=N_G(\H)$ and thus $C_G(T)=C_X(T)\unlhd N_X(T)$. Notice that $S_0$ is a Sylow $p$-subgroup of $X$ which is contained in $N_X(T)$. Thus, $S_0$ is also a Sylow $p$-subgroup of $N_G(T)$. As $C_G(T)\unlhd N_X(T)$, it follows that $C_{S_0}(T)=S_0\cap C_G(T)\in\Syl_p(C_G(T))$. Since $C_S(T)\leq C_{S_0}(T)$ by assumption, we can conclude that $C_S(T)\in\Syl_p(C_G(T))$ if and only if $C_S(T)=C_{S_0}(T)$. Since, $N_\F(\tT)=\F_S(G)$, the condition $C_S(T)\in\Syl_p(C_G(T))$ is by \cite[Proposition~I.5.4]{Aschbacher/Kessar/Oliver:2011} equivalent to $T$ being fully $N_\F(\tT)$-centralized. It follows from Lemma~\ref{L:ConjEandT} that this is in turn equivalent to $T$ being fully $\F$-centralized.
\smallskip
\textbf{(b)} Suppose now that $T$ is fully centralized or equivalently that $C_S(T)=C_{S_0}(T)$. Assume furthermore $N_S(\E)\leq S_0$. By Lemma~\ref{L:NFEbNLH}(a), we have $N_\F(\E)=\F_0|_{N_S(\E)}$. Using Corollary~\ref{C:CFT}, we can in particular conclude that $C_\F(T)=C_{N_\F(\E)}(T)\subseteq C_{\F_0}(T)$. Moreover, if $P,Q\leq C_{S_0}(T)=C_S(T)$ and $\phi\in\Hom_{C_{\F_0}(T)}(P,Q)$, then $\phi$ extends to $\hat{\phi}\in\Hom_{\F_0}(PT,QT)$ with $\hat{\phi}|_T=\id_T$. Notice that $PT$ and $QT$ are contained in $N_S(\E)$. Hence, $\hat{\phi}$ is a morphism in $\F_0|_{N_S(\E)}=N_\F(\E)$. Thus, $\phi=\hat{\phi}|_P$ is a morphism in $N_{N_\F(\E)}(T)=C_\F(T)$. This proves (b).
\end{proof}
\begin{lemma}\label{L:EfullyNormalized}
The following properties are equivalent:
\begin{itemize}
\item [(i)] $\E$ is fully $\F$-normalized;
\item [(ii)] $N_S(\H)\in\Syl_p(N_{G}(\H))$;
\item [(iii)] $\E$ is fully $\F$-automized and $T$ is fully centralized.
\end{itemize}
If so, then $N_\F(\E)$ is saturated, $\E\unlhd N_\F(\E)$, $E(N_\F(\E))=E(\F)$ and $(\bN_\L(\H),\delta(N_\F(\E)),N_S(\E))$ is a regular locality over $N_\F(\E)$.
\end{lemma}
\begin{proof}
Recall that $N_S(\E)=N_S(\H)$ and $N_\F(\E)=\F_{N_S(\H)}(\bN_\L(\H))$. In particular, if (ii) holds, then Theorem~\ref{T:bNLH}(a),(b),(c) applied with $S_0=N_S(\H)$ gives that $N_\F(\E)$ is saturated, $\E\unlhd N_\F(\E)$, $E(N_\F(\E))=E(\F)$ and $(\bN_\L(\H),\delta(N_\F(\E)),N_S(\E))$ is a regular locality over $N_\F(\E)$. Thus, it remains only to prove the equivalence of properties (i)-(iii). For the proof choose $S_0\in\Syl_p(N_G(\H))$ such that $N_S(\E)=N_S(\H)\leq S_0$.
\smallskip
\textbf{(ii)$\Longleftrightarrow$(iii)} Observe that
\begin{equation}\label{E:Inequality}
|\Aut_{N_S(\E)}(T)|\cdot |C_S(T)|=|N_S(\E)|\leq |S_0|=|\Aut_{S_0}(T)|\cdot |C_{S_0}(T)|.
\end{equation}
The inequality here is an equality if and only if $N_S(\E)=S_0$, i.e. if and only if (ii) holds. We have moreover $\Aut_{N_S(\E)}(T)\leq \Aut_{S_0}(T)$ with equality if and only if $\E$ is fully automized (see Lemma~\ref{L:EFullyAutomized}(b)). As $C_S(T)\leq N_S(\E)\leq S_0$, we have moreover $C_S(T)\leq C_{S_0}(T)$. Lemma~\ref{L:Tfullycentralized}(a) gives that $C_S(T)=C_{S_0}(T)$ if and only if $T$ is fully centralized. Hence, the inequality in \eqref{E:Inequality} is also an equality if and only if (iii) holds.
\smallskip
\textbf{(i)$\Longrightarrow$(ii)} To prove this implication by contraposition, assume (ii) is false, i.e. $N_S(\E)<S_0$. As $S\in\Syl_p(G)$ and $S_0$ is a $p$-subgroup of $G$, it follows from Sylow's Theorem that there exists $g\in G$ with $S_0^g\leq S$. Notice that $T\leq N_S(\E)\leq S_0$. So in particular, $T^g\leq S$ and $\phi=c_g|_T$ is a morphism in $\F$. By Lemma~\ref{L:ConjH}(a), we have $\E^\phi=\F_{T^g}(\H^g)$. By Lemma~\ref{L:ConjH}(b), $S_0^g\leq N_G(\H^g)$ and thus $S_0^g\leq N_S(\H^g)=N_S(\E^\phi)$ by Lemma~\ref{L:finNLtT}. Hence, $|N_S(\E)|<|S_0|=|S_0^g|\leq |N_S(\E^\phi)|$ and therefore $\E$ is not fully $\F$-normalized.
\smallskip
\textbf{(ii)$\Longrightarrow$(i)} Assume (ii) and fix $\E'\in \E^\F$ such that $\E'$ is fully $\F$-normalized. It is sufficient to show that $|N_S(\E')|\leq |N_S(\E)|$. By Lemma~\ref{L:ConjEandT} we have $\E'\in\E^{N_\F(\tT)}$. As $N_\F(\tT)=\F_S(G)$, there exists thus $g\in G$ with $T^g\leq S$ such that, setting $\phi=c_g|_T$, we have $\E'=\E^\phi$. Then Lemma~\ref{L:ConjH}(a) gives $\E'=\F_{T^g}(\H^g)$. By Lemma~\ref{L:ConjH}(b), $c_g$ induces an isomorphism from $N_G(\H)$ to $N_G(\H^g)$. By Lemma~\ref{L:finNLtT}, $N_S(\E')=N_S(\H^g)$ is a $p$-subgroup of $N_G(\H^g)$. So (ii) implies $|N_S(\E')|\leq |N_S(\H)|=|N_S(\E)|$ as required,
\end{proof}
\begin{proof}[Proof of Theorem~\ref{T:FusionNormalizer}]
Part (a) is a restatement of Lemma~\ref{L:NFEbNLH}(c), part (b) was shown as Proposition~\ref{P:DinNFE}, part (c) follows from Lemma~\ref{L:EfullyNormalized}, and part (d) was proved as Lemma~\ref{L:inAutEinNFE} and Corollary~\ref{C:CFT}. If $\F$ is constrained, then by Remark~\ref{R:Constrained}, $\L$ is a model for $\F$, $\H$ is a subnormal subgroup of $\L$ realizing $\E$, $\bN_\L(\H)=N_\L(\H)$ and thus $N_\F(\E)=\F_{N_S(\H)}(N_\L(\H))$. Part (e) follows thus from Lemma~\ref{L:NormalizerConstrained}.
\end{proof}
We conclude this section with two lemmas which make it easier to move to fully $\F$-normalized conjugates of $\E$.
\begin{lemma}\label{L:MapNSE}
Let $\beta\in\Hom_\F(N_S(\E),S)$. Then $N_S(\E)\beta\leq N_S(\E^\beta)$.
\end{lemma}
\begin{proof}
If $s\in N_S(\E)$, then $c_s|_T\in\Aut(\E)$. So using \cite[Remark~2.1]{Henke:2013}, one can observe that
\[c_{s\beta}|_{T\beta}=\beta^{-1}(c_s|_T)\beta\in\Aut(\E^\beta)\]
and thus $s\beta\in N_S(\E^\beta)$.
\end{proof}
\begin{lemma}
\begin{itemize}
\item [(a)] Let $\hat{\E}\in\E^\F$ such that $\hat{\E}$ is fully $\F$-normalized. Then there exists $\alpha\in\Hom_\F(N_S(\E),S)$ such that $\E^\alpha=\hat{\E}$.
\item [(b)] Fix $\alpha\in\Hom_\F(N_S(\E),S)$ such that $\E^\alpha$ is fully normalized. Then for all $P,Q\leq N_S(\E)$,
\[\Hom_{N_\F(\E)}(P,Q)=\{\phi\in\Hom_\F(P,Q)\colon \alpha^{-1}\phi\alpha\in\Hom_{N_\F(\E^\alpha)}(P\alpha,Q\alpha)\}.\]
In particular, if $\E$ is also fully normalized, then $N_S(\E)\alpha=N_S(\E^\alpha)$ and $\alpha$ induces an isomorphism from $N_\F(\E)$ to $N_\F(\E^\alpha)$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} Let $\hat{\E}$ be a subsystem over $\hat{T}$. As $\hat{\E}\in\E^\F$, there exists $\phi\in\Hom_\F(T,S)$ such that $\hat{T}=T\phi$ and $\hat{\E}=\E^\phi$. Notice that $(\Aut_\F(T)\cap \Aut(\E))^\phi=\Aut_\F(\hat{T})\cap \Aut(\hat{\E})$. In particular, $\Aut_{N_S(\E)}(T)^\phi$ is a $p$-subgroup of $\Aut_\F(\hat{T})\cap \Aut(\hat{\E})$. As $\hat{\E}$ is fully normalized, Lemma~\ref{L:EfullyNormalized} gives that $\hat{T}$ is fully centralized and $\Aut_S(\hat{T})\cap \Aut(\E)\in\Syl_p(\Aut_\F(\hat{T})\cap \Aut(\hat{\E}))$. Hence, by Sylow's Theorem there exists $\gamma\in\Aut_\F(\hat{T})\cap\Aut(\hat{\E})$ such that $\Aut_{N_S(\E)}(T)^{\phi\gamma}\leq \Aut_S(\hat{T})$. This means that $N_S(\E)\leq N_{\phi\gamma}$. Hence, by the extension axiom \cite[Proposition~I.2.5]{Aschbacher/Kessar/Oliver:2011}, $\phi\gamma$ extends to $\alpha\in\Hom_\F(N_S(\E),S)$. As $\gamma\in\Aut(\hat{\E})$, we have $\E^\alpha=\E^{\phi\gamma}=\hat{\E}^\gamma=\hat{\E}$. This proves (a).
\smallskip
\textbf{(b)} For the proof of (b) let $\alpha\in\Hom_\F(N_S(\E),S)$ be arbitrary such that $\E^\alpha$ is fully normalized. By Lemma~\ref{L:MapNSE}, $N_S(\E)\alpha\leq N_S(\E^\alpha)$. If $\E$ is fully normalized, then $|N_S(\E)\alpha|=|N_S(\E)|\geq |N_S(\E^\alpha)|$ and so $N_S(\E)\alpha=N_S(\E^\alpha)$.
\smallskip
By Lemma~\ref{L:BasicT0perp}(a), we have $\tT=T_0T_0^\perp\subseteq TC_S(\H)\leq N_S(\H)=N_S(\E)$. As $\tT=E(\L)\cap S$ is strongly $\F$-closed, it follows that $\tT\alpha=\tT$ and $\alpha$ is a morphism in $N_\F(\tT)=\F_S(G)$. Thus, we may pick $g\in G$ such that $\alpha=c_g|_{N_S(\E)}$.
\smallskip
Recall that $N_S(\E)=N_S(\H)$. By Lemma~\ref{L:ConjH}(a), we have $\E^\alpha=\F_{T^g}(\H^g)$ and $T^g=\H^g\cap S$. In particular, $N_S(\E^\alpha)=N_S(\H^g)$ by Lemma~\ref{L:finNLtT}. As stated in Lemma~\ref{L:ConjH}(b), $c_g$ induces an isomorphism from $N_G(\H)$ to $N_G(\H^g)$. As $\E^\alpha$ is fully normalized, it follows from Lemma~\ref{L:EfullyNormalized} that $N_S(\E^\alpha)=N_S(\H^g)\in\Syl_p(N_G(\H^g))$. Hence, $N_S(\E^\alpha)^{g^{-1}}$ is a Sylow $p$-subgroup of $N_G(\H)$. This allows us to choose $S_0$ such that
\[S_0=N_S(\E^\alpha)^{g^{-1}}.\]
By Lemma~\ref{L:MapNSE}, we have $N_S(\E)^g=N_S(\E)\alpha\leq N_S(\E^\alpha)$ and thus $N_S(\E)\leq S_0$. Now by Lemma~\ref{L:ConjH}(b), $c_g$ induces an isomorphism from $\bN_\L(\H)$ to $\bN_\L(\H^g)$ with $S_0c_g=N_S(\E^\alpha)=N_S(\H^g)$. Hence, using Theorem~\ref{T:bNLH}(a), it follows from Lemma~\ref{L:AutLFusion} that $c_g$ induces an isomorphism from $\F_0:=\F_{S_0}(\bN_\L(\H))$ to $N_\F(\E^\alpha)=\F_{N_S(\H^g)}(\bN_\L(\H^g))$.
\smallskip
If $\E$ is fully normalized, then Lemma~\ref{L:EfullyNormalized} implies $N_S(\E)=S_0$ and thus $\F_0=N_\F(\E)$, i.e. $\alpha$ induces an isomorphism from $N_\F(\E)$ to $N_\F(\E^\alpha)$. In the general case, Lemma~\ref{L:NFEbNLH}(a) gives $\F_0\cap \F=N_\F(\E)$. Given $\phi\in\Hom_\F(P,Q)$, we have thus $\phi\in\Hom_{N_\F(\E)}(P,Q)$ if and only if $\phi\in\Hom_{\F_0}(P,Q)$, and this is the case if and only if $\alpha^{-1}\phi\alpha=(c_g|_{S_0})^{-1}\phi (c_g|_{S_0})\in\Hom_{N_\F(\E^\alpha)}(P\alpha,Q\alpha)$. This proves the assertion.
\end{proof}
\subsection{The centralizer of $\E$ and fully centralized conjugates of $\E$}\label{SS:Centralizer}
We continue to assume Hypotheses~\ref{H:main} and \ref{H:S0F0}. In particular, $S_0$ is a Sylow $p$-subgroup of $N_G(\H)$ containing $T$.
\smallskip
If $\E$ is normal in $\F$, then it was first shown by Aschbacher that there is a unique largest element of the set
\[\{R\leq S\colon \E\subseteq C_\F(R)\},\]
which we denote by $C_S(\E)=C_S(\E)_\F$ (cf. \cite[Chapter~6]{Aschbacher:2011} or \cite[Theorem~1(a)]{Henke:2018}). We extend this result now to the case that $\E$ is subnormal in $\F$. As $\F_0$ is saturated and $\E\unlhd \F_0$, we can use that the subgroup $C_{S_0}(\E)_{\F_0}$ is defined as the largest element of the set $\{R\leq S_0\colon \E\subseteq C_{\F_0}(R)\}$. Notice that the following proposition implies Theorem~\ref{T:CentralizerEinS}.
\begin{prop}\label{P:CSE}
The subgroup $C_S(\H)$ is with respect to inclusion a maximal element of the set $\m{X}:=\{R\leq S\colon \E\subseteq C_\F(R)\}$. If $C_S(\H)\leq S_0$, then $C_S(\H)=S\cap C_{S_0}(\E)_{\F_0}$.
\end{prop}
\begin{proof}
By \cite[Lemma~3.5]{Henke:Regular}, we have $\H\subseteq C_\L(C_S(\H))$, which implies easily that $\E=\F_T(\H)\subseteq C_\F(C_S(\H))$. Hence, $C_S(\H)\in\X$.
\smallskip
Choose now $S_0$ such that $C_S(\H)\leq S_0$. Recall from Theorem~\ref{T:bNLH}(a),(b) that $\F_0$ is saturated and $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$ with $\H\unlhd\bN_\L(\H)$, $T=S_0\cap \H$ and $\E\unlhd\F_0$. Hence, it follows from \cite[Proposition~4]{Henke:2015} that $C_{S_0}(\E)_{\F_0}=C_{S_0}(\H)$. As $C_S(\H) \leq S_0$ by assumption, we can conclude that $C_S(\H)=S\cap C_{S_0}(\H)=S\cap C_{S_0}(\E)_{\F_0}$.
\smallskip
Choose now $S_0$ such that $N_S(\E)\leq S_0$. We have then $N_\F(\E)\subseteq \F_0$ by Lemma~\ref{L:NFEbNLH}(a). Let $R\in\X$. Then $\E$ and $\F_R(R)$ commute in the sense of Definition~\ref{D:CommuteSubsystems}. Hence, by Lemma~\ref{L:CentralProduct}, $\mD:=\E*\F_R(R)$ is a saturated subsystem of $\F$ over $T R$ with $\E\unlhd\mD$ and $\E\subseteq C_{\mD}(R)$. Proposition~\ref{P:DinNFE} yields now $\mD\subseteq N_\F(\E)\subseteq \F_0$. In particular, $R\leq N_S(\E)\leq S_0$. As $\E\subseteq C_\mD(R)\subseteq C_{\F_0}(R)$, it follows $R\leq C_{S_0}(\E)_{\F_0}$. Hence, $R\leq S\cap C_{S_0}(\E)_{\F_0}=C_S(\H)$. This proves the assertion.
\end{proof}
\begin{definition}
We write $C_S(\E)$ or $C_S(\E)_\F$ for $C_S(\H)$, i.e. for the maximal element of the set
\[\{R\leq S\colon \E\subseteq C_\F(R)\}.\]
We call $C_S(\E)=C_S(\E)_\F$ the \emph{centralizer of $\E$ in $S$}.
\end{definition}
The reader might want to observe that
\[C_S(\E)=C_S(\H)\leq N_S(\H)=N_S(\E)\]
by Lemma~\ref{L:finNLtT}.
\smallskip
Recall from the introduction that $C_\F(\E)$ is defined as
\[C_\F(\E):=\F_{C_S(\H)}(C_\L(\H)).\]
We call $C_\F(\E)$ the \emph{centralizer of $\E$ in $\F$}.
\smallskip
Note that $C_\F(\E)$ depends a priori not only on $\F$ and $\E$, but also on $\L$. However, the properties we show below imply that $C_\F(\E)$ is indeed independent of the choice of $\L$. Alternatively, this can be seen more directly using Background Theorems~\ref{bT:0} and \ref{bT:2} combined with Lemma~\ref{L:AutLFusion}.
\smallskip
The reader is referred to \cite[Definition~I.7.1]{Aschbacher/Kessar/Oliver:2011} for the definition of the focal subgroup $\foc(\F)$ and the hyperfocal subgroup $\hyp(\F)$. We extend the definition of the focal subgroup to fusion systems which are not saturated. Namely, for every fusion system $\G$ over a $p$-group $X$, we set
\[\foc(\G):=\<x^{-1}(x\phi)\colon x\in X,\;\phi\in\Hom_\F(\<x\>,X)\>.\]
It is always true that $\hyp(\F)\leq \foc(\F)$. For every subgroup $R\leq S$ with $\hyp(\F)\leq R$, it is shown in \cite[Theorem~I.7.4]{Aschbacher/Kessar/Oliver:2011} that
\begin{equation}\label{E:DefineFR}
\F_R:=\<O^p(\Aut_\F(P))\colon P\leq R\>_R
\end{equation}
is the unique saturated subsystem of $\F$ over $R$ of $p$-power index. This concept will play a role in obtaining a description of $C_\F(\E)$ if $\E$ is fully centralized.
\smallskip
If $\E$ is normal in $\F$, then a subsystem $C_\F(\E)$ was first defined by Aschbacher \cite[Chapter~6]{Aschbacher:2011}, and a new construction of $C_\F(\E)$ was given in \cite{Henke:2018}. However, if $\E$ is normal, then by Background Theorem~\ref{bT:1}, we have $\H\unlhd\L$. Hence, \cite[Proposition~6.7]{Chermak/Henke} states in this case that the subsystem $C_\F(\E)$ defined in \cite{Aschbacher:2011,Henke:2018} equals $\F_{C_S(\H)}(C_\L(\H))$ and coincides thus with the centralizer $C_\F(\E)$ defined in this paper. Using this, the following theorem follows from \cite[Proposition~1, Theorem~2]{Henke:2018} and the construction of $C_\F(\E)$ given in \cite{Henke:2018}.
\begin{theorem}\label{T:Henke2018}
Suppose $\E\unlhd\F$. Then $\hyp(C_\F(T))\leq \foc(C_\F(T))\leq C_S(\E)$. Moreover,
\[C_\F(\E)=C_\F(T)_{C_S(\E)}\]
is normal in $\F$. Every saturated subsystem of $\F$ which commutes with $\E$ is contained in $C_\F(\E)$. \qed
\end{theorem}
We use here the notation introduced in \eqref{E:DefineFR} for $C_\F(T)$ in place of $\F$. Notice that this is possible, as $T$ is in the situation above fully centralized and thus $C_\F(T)$ is saturated.
\begin{proof}
By Lemma~\ref{L:EfullyNormalized}, $N_\F(\E)$ is saturated, $\E\unlhd N_\F(\E)$ and $(\bN_\L(\H),\delta(N_\F(\E)),N_S(\E))$ is a regular locality over $N_\F(\E)$. By Lemma~\ref{L:CLTinbNLH}, we have $C_\L(\H)=C_{\bN_\L(\H)}(\H)$ and thus $C_\F(\E)=C_{N_\F(\E)}(\E)$ by definition. It follows from Lemma~\ref{L:inAutEinNFE} that $C_\F(T)=C_{N_\F(\E)}(T)$. Hence, the assertion follows from Theorem~\ref{T:Henke2018}.
\end{proof}
\begin{lemma}\label{L:CF0ECollect}
The centralizer $C_{\F_0}(\E)$ is normal in $\F_0$, $\foc(C_{\F_0}(T))\leq C_{S_0}(\E):=C_{S_0}(\E)_{\F_0}$ and
\[C_{\F_0}(\E)=\F_{C_{S_0}(\H)}(C_\L(\H))=C_{\F_0}(T)_{C_{S_0}(\E)}.\]
Moreover, $(C_\L(\H),\delta(C_{\F_0}(\E)),C_{S_0}(\H))$ is a regular locality over $C_{\F_0}(\E)$. In particular, $C_{S_0}(\H)=C_{S_0}(\E)_{\F_0}$ is a maximal $p$-subgroup of $C_\L(\H)$.
\end{lemma}
\begin{proof}
As $\F_0$ is saturated and $\E\unlhd\F_0$, it follows from Theorem~\ref{T:Henke2018} that $C_{\F_0}(\E)=C_{\F_0}(T)_{C_{S_0}(\E)}$ is normal in $\F_0$. Recall that $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$ and $C_\L(\H)=C_{\bN_\L(\H)}(\H)$ by Lemma~\ref{L:CLTinbNLH}. Hence, the remaining part of the assertion follows from Background Theorem~\ref{bT:2} and the definition $C_{\F_0}(\E)$.
\end{proof}
Recall that $\E$ is called \emph{fully $\F$-centralized} if
\[|C_S(\E)|\geq |C_S(\tE)|\mbox{ for all }\tE\in\E^\F.\]
\begin{lemma}\label{L:EFullyCentralized}~
\begin{itemize}
\item [(a)] Choose $S_0$ such that $C_S(\H)=C_S(\E)\leq S_0$. Then the following are equivalent:
\begin{itemize}
\item [(i)] $\E$ is fully $\F$-centralized;
\item [(ii)] $C_S(\H)=C_{S_0}(\H)$;
\item [(iii)] $C_S(\H)$ is a maximal $p$-subgroup of $C_\L(\H)$.
\end{itemize}
\item [(b)] Let $\hat{\E}\in\E^\F$ such that $\hat{\E}$ is fully $\F$-centralized. Then there exists $\alpha\in\Hom_\F(TC_S(\E),S)$ such that $\E^\alpha=\hat{\E}$.
\item [(c)] If $\alpha\in\Hom_\F(TC_S(\E),S)$, then $\alpha$ is a morphism in $N_\F(\tT)$ and $C_S(\E)\alpha\leq C_S(\E^\alpha)$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a,b)} Choose $S_0$ such that $C_S(\H)\leq S_0$. By Lemma~\ref{L:CF0ECollect}, $C_{S_0}(\H)\geq C_S(\H)$ is a maximal $p$-subgroup of $C_\L(\H)$. Hence, (ii) and (iii) are equivalent. Thus, it remains only to show that (i) and (ii) are equivalent and that (c) holds.
\smallskip
\textit{Step~1:} We show that (i) implies (ii). Notice that $TC_{S_0}(\H)$ is a $p$-subgroup of $N_G(\H)$. As $S\in\Syl_p(G)$, there exists thus $g\in G$ with $(TC_{S_0}(\H))^g\leq S$. As $T^g\leq S$, $\phi=c_g|_T$ is an $\F$-morphism and $T\phi=T^g\in T^\F$. By Lemma~\ref{L:ConjH}(a), we have $T^g=\H^g\cap S$ and $\E^\phi=\F_{T^g}(\H^g)$. In particular, $C_S(\E^\phi)=C_S(\H^g)$. Using Lemma~\ref{L:ConjH}(b), one observes that $C_{S_0}(\H)^g\leq C_S(\H^g)$. Thus, if $C_S(\H)<C_{S_0}(\H)$, then $|C_S(\E)|=|C_S(\H)|<|C_S(\H^g)|=|C_S(\E^\phi)|$ and $\E$ is thus not fully centralized. So Step~1 is complete.
\smallskip
\emph{Step~2:} We show that (b) holds and (ii) implies (i). For the proof fix $\hat{\E}\in \E^\F$ fully centralized (and notice that such $\hat{\E}$ always exists). By Lemma~\ref{L:ConjEandT}, there is $\phi\in\Hom_{N_\F(\tT)}(T,S)$ such that $\E^\phi=\hat{\E}$. As $N_\F(\tT)=\F_S(G)$, there exists $g\in G$ with $\phi=c_g|_T$. By Lemma~\ref{L:ConjH}(a), we have $\F_{T^g}(\H^g)=\hat{\E}$ and thus $C_S(\hat{\E})=C_S(\H^g)$. Fix $\hat{S}_0\in \Syl_p(N_G(\H^g))$ such that
\[T^gC_S(\H^g)\leq \hat{S}_0.\]
As $T$ is a maximal $p$-subgroup of $\H$, the conjugate $T^g=T\phi$ is a maximal $p$-subgroup of $\H^g$ and thus equal to $S\cap \H^g$. By Theorem~\ref{T:bNLH}(b) applied with $\H^g$ in place of $\H$, we have moreover $\H^g\cap \hat{S}_0=T^g=S\cap \H^g$. Step~1 applied with $(\tE,\H^g)$ in place of $(\E,\H)$ gives \[C_S(\H^g)=C_{\hat{S}_0}(\H^g).\]
Notice that $S_0^g$ is a Sylow $p$-subgroup of $N_G(\H^g)$ as $c_g$ induces an isomorphism from $N_G(\H)$ to $N_G(\H^g)$ by Lemma~\ref{L:ConjH}(b). Hence, there exists $x\in N_G(\H^g)$ such that $S_0^{gx}=\hat{S}_0$. Observe that $\H^g=\H^{gx}$. Moreover, $T^{gx}\leq S_0^{gx}\cap \H^g=\hat{S}_0\cap \H^g=T^g$ and thus $T^{gx}=T^g\leq S$. Thus, replacing $\phi$ and $g$ by $c_{gx}$ and $gx$, we may assume that $S_0^g=\hat{S}_0$. Hence, using $c_g\in\Aut(\L)$, we may conclude that
\begin{equation}\label{E:CSHg}
C_S(\hat{\E})=C_S(\H^g)=C_{\hat{S}_0}(\H^g)=C_{S_0}(\H)^g\geq C_S(\H)^g=C_S(\E)^g.
\end{equation}
So $\alpha:=c_g|_{TC_S(\E)}$ is an element of $\Hom_{N_\F(\tT)}(TC_S(\E),S)$ with $\E^\alpha=\hat{\E}$. This proves (b). If we assume in addition that $C_S(\H)=C_{S_0}(\H)$, then we obtain from \eqref{E:CSHg} that $C_S(\E)^g=C_S(\H)^g=C_{S_0}(\H)^g=C_S(\H^g)=C_S(\hat{\E})$. As $\hat{\E}$ was chosen to be fully centralized, it follows then that $\E$ is fully centralized. This completes Step~2 and thus the proof of (a) and (b).
\smallskip
\textbf{(c)} By Lemma~\ref{L:BasicT0perp}(a), we have $\tT=T_0T_0^\perp\subseteq TC_S(\H)=TC_S(\E)$. As $\tT$ is strongly $\F$-closed, it follows that a morphism $\alpha\in\Hom_\F(TC_S(\E),S)$ is a morphism in $N_\F(\tT)$. As $N_\F(\tT)=\F_S(G)$, there exists thus $g\in G$ with $\alpha=c_g|_{TC_S(\E)}$. By Lemma~\ref{L:ConjH}(a), we have $S\cap \H^g=T^g$ and $\E^\alpha=\F_{T^g}(\H^g)$. In particular, $C_S(\E^\alpha)=C_S(\H^g)$. Using Lemma~\ref{L:ConjH}(b), we can conclude that $C_S(\E)\alpha=C_S(\E)^g=C_S(\H)^g\leq S\cap C_G(\H^g)=C_S(\H^g)=C_S(\E^\alpha)$.
\end{proof}
\begin{prop}\label{P:CFEDescribe}
If $\E$ is fully centralized, then $C_\F(\E)$ is saturated, $\foc(C_\F(T))\leq C_S(\E)$ and
\[C_\F(\E)=\<O^p(\Aut_{C_\F(T)}(P))\colon P\leq C_S(\E)\>_{C_S(\E)}.\]
If in addition $S_0$ is chosen such that $C_S(\E)\leq S_0$, then $C_\F(\E)=C_{\F_0}(\E)$ is normal in $\F_0$.
\end{prop}
\begin{proof}
Choose first $S_0$ such that $C_S(\E)\leq S_0$. Assuming that $\E$ is fully $\F$-centralized, part (a) gives $C_{S_0}(\E)_{\F_0}=C_{S_0}(\H)=C_S(\H)=C_S(\E)$. Hence, using the definition of $C_\F(\E)$ and Lemma~\ref{L:CF0ECollect}, we can conclude that
\begin{equation}\label{E:Foc}
\foc(C_{\F_0}(T))\leq C_S(\E),
\end{equation}
and that
\begin{equation}\label{E:CFE}
C_\F(\E)=\F_{C_S(\H)}(C_\L(\H))=\F_{C_{S_0}(\H)}(C_\L(\H))=C_{\F_0}(\E)=C_{\F_0}(T)_{C_S(\E)}
\end{equation}
is saturated and normal in $\F_0$. It remains to show that $\foc(C_\F(T))\leq C_S(\E)$ and $C_\F(\E)=\<O^p(\Aut_{C_\F(T)}(P))\colon P\leq C_S(\E)\>_{C_S(\E)}$. For the proof we may choose $S_0$ such that $N_S(\E)\leq S_0$. By Corollary~\ref{C:CFTCF0T}, we have then
\[C_\F(T)=C_{\F_0}(T)|_{C_S(T)}.\]
In particular, if $x\leq C_S(T)$ and $\phi\in\Hom_{C_\F(T)}(x,C_S(T))$, then $\phi$ is a morphism in $C_{\F_0}(T)$ and $x^{-1}(x\phi)\in\foc(C_{\F_0}(T))\leq C_S(\E)$ by \eqref{E:Foc}. This shows $\foc(C_\F(T))\leq C_S(\E)$.
\smallskip
As $C_S(\E)\leq C_S(T)$, it follows moreover that $\Aut_{C_{\F_0}(T)}(P)=\Aut_{C_\F(T)}(P)$ for every $P\leq C_S(\E)$. Hence, using \eqref{E:DefineFR} and \eqref{E:CFE}, we can conclude that
\begin{eqnarray*}
C_\F(\E)&=&C_{\F_0}(T)_{C_S(\E)}=\<O^p(\Aut_{C_{\F_0}(T)}(P))\colon P\leq C_S(\E)\>_{C_S(\E)}\\
&=&\<O^p(\Aut_{C_\F(T)}(P))\colon P\leq C_S(\E)\>_{C_S(\E)}.
\end{eqnarray*}
This proves the assertion.
\end{proof}
\begin{lemma}\label{L:CentralizerTFullyCentralized}~
\begin{itemize}
\item [(a)] Suppose $T$ is fully $\F$-centralized. Then $\E$ is fully $\F$-centralized, $\foc(C_\F(T))\leq C_S(\E)$ and $C_\F(\E)=C_\F(T)_{C_S(\E)}$.
\item [(b)] If $\E$ is fully $\F$-normalized, then $C_\F(\E)=C_{N_\F(\E)}(\E)\unlhd N_\F(\E)$.
\end{itemize}
\end{lemma}
\begin{proof}
\textbf{(a)} Assume $T$ is fully $\F$-centralized and choose $S_0$ such that $C_S(T)\leq S_0$. Lemma~\ref{L:Tfullycentralized}(a) gives then that $C_S(T)=C_{S_0}(T)$. In particular, $C_S(\H)=C_{C_S(T)}(\H)=C_{C_{S_0}(T)}(\H)=C_{S_0}(\H)$. Hence, by Lemma~\ref{L:EFullyCentralized}(a), $\E$ is fully centralized. Using that $C_\F(T)$ is saturated, part (a) follows now from Proposition~\ref{P:CFEDescribe} and \eqref{E:DefineFR}.
\smallskip
\textbf{(b)} If $\E$ is fully normalized, then Lemma~\ref{L:EfullyNormalized} yields $N_S(\E)=N_S(\H)\in\Syl_p(N_G(\H))$. Hence, we may assume that $S_0=N_S(\E)$ and thus $\F_0=N_\F(\E)$. The assertion follows then from Lemma~\ref{L:CF0ECollect}.
\end{proof}
We now want to consider the properties of $C_\F(\E)$ if $\E$ is not necessarily fully centralized or fully normalized. The reader might want to recall Definition~\ref{D:WeaklyInvariant}.
\begin{lemma}\label{L:CFEWeaklyInvariant}
Suppose $C_S(\H)=C_S(\E)\leq S_0$. Then
\[C_\F(\E)=\F\cap C_{\F_0}(\E)=C_{\F_0}(\E)|_{C_S(\E)}.\]
In particular, $C_\F(\E)$ is weakly $N_\F(\E)$-invariant.
\end{lemma}
\begin{proof}
By Lemma~\ref{L:CF0ECollect}, $\C:=C_{\F_0}(\E)=\F_{C_{S_0}(\H)}(C_\L(\H))$ is a normal subsystem of $\F_0$. In particular, $\C$ is weakly $\F_0$-invariant, which implies that $\F\cap \C$ is weakly $(\F\cap \F_0)$-invariant. By Lemma~\ref{L:NFEbNLH}(a), we have $N_\F(\E)=\F\cap \F_0$ if $S_0$ is chosen such that $N_S(\E)\leq S_0$. Hence, it is sufficient to prove that $C_\F(\E)=\F\cap \C=\C|_{C_S(\E)}$.
\smallskip
By Proposition~\ref{P:CSE}, both $C_\F(\E)$ and $\F\cap \C$ are fusion systems over
\[C_S(\H)=C_S(\E)=S\cap C_{S_0}(\E)_{\F_0}=S\cap C_{S_0}(\H).\]
Notice that $C_\F(\E)=\F_{C_S(\H)}(C_\L(\H))$ is contained in $\C=\F_{C_{S_0}(\H)}(C_\L(\H))$ and thus in $\F\cap \C$. Hence
\[C_\F(\E)\subseteq \C\cap \F\subseteq \C|_{C_S(\E)}\]
and it remains to show that $\C|_{C_S(\E)}\subseteq C_\F(\E)$.
\smallskip
Recall from Lemma~\ref{T:bNLH}(a),(b),(c) that $(\bN_\L(\H),\delta(\F_0),S_0)$ is a regular locality over $\F_0$ with $\H\unlhd\bN_\L(\H)$ and $\Comp(\L)=\Comp(\bN_\L(\H))$. Since $C_\L(\H)=C_{\bN_\L(\H)}(\H)$ by Lemma~\ref{L:CLTinbNLH}, it follows from \cite[Lemma~11.16, Theorem~10.16(e)]{Henke:Regular} (applied with $(\bN_\L(\H),\H)$ in place of $(\L,\N)$) that
\[\Comp(C_\L(\H))=\Comp(\bN_\L(\H))\backslash\Comp(\H)=\Comp(\L)\backslash\Comp(\H).\]
Thus, $E(C_\L(\H))=\M$ by definition of $\M$. Here $\M$ is subnormal in $\L$ as $\M\unlhd E(\L)\unlhd\L$ by Lemma~\ref{L:BasicT0perp}(a). Thus, it follows from Background Theorem~\ref{bT:1} and the definition of a locality that $T_0^\perp=\M\cap S$ is a maximal $p$-subgroup of $\M$. As $T_0^\perp\leq C_S(\H)\leq C_{S_0}(\H)$ by Lemma~\ref{L:BasicT0perp}(a), it follows that $T_0^\perp=C_{S_0}(\H)\cap \M$. Hence, $E(\C)=\F_{T_0^\perp}(\M)$ by Background Theorem~\ref{bT:2} applied with $(C_\L(\H),C_{S_0}(\H),\C)$ in place of $(\L,S,\F)$.
\smallskip
Let now $P,Q\leq C_S(\E)=C_S(\H)$ and $\phi\in\Hom_{\C}(P,Q)$. We need to show that $\phi$ is a morphism in $N_\F(\E)$. By Lemma~\ref{L:FusionSystemFrattini} (applied with $\C$ and $E(\C)$ in place of $\F$ and $\E$), there exist $\psi\in\Hom_{E(\C)P}(P,PT_0^\perp)$ and a morphism $\alpha\colon P\psi\rightarrow Q$ in $N_{\C}(T_0^\perp)$ such that $\phi=\psi\alpha$. By \cite[Theorem~D(a)]{Chermak/Henke}, we have $E(\C)P=\F_{T_0^\perp P}(\M P)$. Notice that $T_0^\perp P\leq C_S(\H)$ and $\M P\subseteq C_\L(\H)$. Hence, $E(\C)P$ is contained in $C_\F(\E)=\F_{C_S(\H)}(C_\L(\H))$ and $\psi$ is a morphism in $C_\F(\E)$. In particular, $P\psi\leq C_S(\H)$.
\smallskip
By Lemma~\ref{L:BasisNLtT} applied with $C_\L(\H)$ in place of $\L$, we have $T_0^\perp=\M\cap S_0=E(C_\L(\H))\cap S_0\in\delta(\C)$. Hence, by \cite[Lemma~3.10(b)]{Grazian/Henke}, $N_{\C}(T_0^\perp)$ is realized by the group $N_{C_\L(\H)}(T_0^\perp)$. In particular, $\alpha=c_g|_{P\psi}$ for some $g\in N_{C_\L(\H)}(T_0^\perp)\subseteq C_\L(\H)$. As $P\psi$ and $Q$ are contained in $C_S(\H)=C_S(\E)$, it follows that $\alpha$ is a morphism in $C_\F(\E)$. Hence, $\phi=\psi\alpha$ is a morphism in $C_\F(\E)$. This shows $\C|_{C_S(\E)}\subseteq C_\F(\E)$ and thus $C_\F(\E)=\F\cap \C=\C|_{C_S(\E)}$ as required.
\end{proof}
\begin{lemma}\label{L:CommuteWithE}
Every saturated subsystem of $\F$ which commutes with $\E$ is contained in $C_\F(\E)$.
\end{lemma}
\begin{proof}
Let $\G$ be a saturated subsystem of $\F$ which commutes with $\E$. Then by Lemma~\ref{L:CentralProduct}, $\mD:=\E *\G$ is a saturated subsystem of $\F$ such that $\E\unlhd\mD$, and moreover $\E$ and $\G$ commute in $\mD$. Hence, by Proposition~\ref{P:DinNFE}, $\mD\subseteq N_\F(\E)$. Choose now $S_0$ such that $N_S(\E)\leq S_0$. Then $\mD\subseteq N_\F(\E)\subseteq \F_0$. As $\E$ and $\G$ commute in $\mD$, they commute also in $\F_0$. Since $\E\unlhd \F_0$, it follows now from Theorem~\ref{T:Henke2018} that $\G\subseteq C_{\F_0}(\E)$. Using Lemma~\ref{L:CFEWeaklyInvariant}, we can conclude that $\G\subseteq \F\cap C_{\F_0}(\E)=C_\F(\E)$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{T:FusionCentralizer}]
We have $C_\L(\H)\subseteq C_\L(T)$ and, by \cite[Lemma~3.5]{Henke:Regular}, $\H\subseteq C_\L(C_S(\H))$. As $\E=\F_T(\H)$ and $C_\F(\E)=\F_{C_S(\H)}(C_\L(\H))$, it follows that $C_\F(\E)\subseteq C_\F(T)$ and $\E\subseteq C_\F(C_S(\H))$, i.e. $\E$ and $C_\F(\E)$ commute in $\F$. Lemma~\ref{L:CLTinbNLH} implies that $C_\F(\E)$ is contained in $N_\F(\E)$, and clearly $C_S(\E)$ is a fusion system over $C_S(\H)=C_S(\E)$.
Now part (a) follows from Lemma~\ref{L:CommuteWithE}.
\smallskip
Part (b) follows from Lemma~\ref{L:CentralizerTFullyCentralized}(b) and Lemma~\ref{L:CFEWeaklyInvariant}. Part (c) is a consequence of (a) and Proposition~\ref{P:CFEDescribe}.
\end{proof}
The following lemma together with Lemma~\ref{L:EFullyCentralized}(b),(c) yields a description of $C_\F(\E)$ in terms of the fusion systems $\E$ and $\F$.
\begin{lemma}\label{L:MoveToFullyCentralized}
Fix $\alpha\in\Hom_\F(TC_S(\E),S)$ such that $\E^\alpha$ is fully $\F$-centralized. Then for all $P,Q\leq C_S(\E)$,
\[\Hom_{C_\F(\E)}(P,Q)=\{\phi\in\Hom_\F(P,Q)\colon \alpha^{-1}\phi\alpha\in\Hom_{C_\F(\E^\alpha)}(P\alpha,Q\alpha)\}.\]
In particular, if $\E$ is also fully $\F$-centralized, then $C_S(\E)\alpha=C_S(\E^\alpha)$ and $\alpha$ induces an isomorphism from $C_\F(\E)$ to $C_\F(\E^\alpha)$.
\end{lemma}
\begin{proof}
Let $\alpha\in\Hom_\F(TC_S(\E),S)$ such that $\E^\alpha$ is fully centralized. By Lemma~\ref{L:EFullyCentralized}(c), $\alpha$ is a morphism in $N_\F(\tT)=\F_S(G)$ and $C_S(\E)\alpha\leq C_S(\E^\alpha)$. In particular, there exists $g\in G$ such that $\alpha=c_g|_{TC_S(\E)}$.
\smallskip
By Lemma~\ref{L:ConjH}(a), we have $\E^\alpha=\F_{T^g}(\H^g)$ and $T^g=\H^g\cap S$. Fix $\hat{S}_0\in\Syl_p(N_G(\H^g))$ such that $T^gC_S(\H^g)\leq N_S(\H^g)=N_S(\E^\alpha)\leq \hat{S}_0$. As $\E^\alpha$ is fully centralized, it follows from Lemma~\ref{L:EFullyCentralized}(a) that $C_S(\H^g)=C_{\hat{S}_0}(\H^g)$. As $G$ acts on $\L$, one can observe that $c_g$ induces an isomorphism from $N_G(\H)$ to $N_G(\H^g)$. Thus, $\hat{S}_0^{g^{-1}}$ is a Sylow $p$-subgroup of $N_G(\H)$ containing $T$. Hence, we may choose $S_0$ such that
\[S_0:=\hat{S}_0^{g^{-1}}.\]
Notice that $C_S(\H)^g=C_S(\E)\alpha\leq S\cap N_G(\H^g)=N_S(\H^g)\leq \hat{S}_0$ and thus $C_S(\H)\leq S_0$. Moreover,
\[C_{S_0}(\H)^g=C_{S_0^g}(\H^g)=C_{\hat{S}_0}(\H^g)=C_S(\H^g).\]
Set $\hat{\F}_0:=\F_{\hat{S}_0}(\bN_\L(\H^g))$. By Proposition~\ref{P:CFEDescribe}, $C_{\hat{\F}_0}(\E^\alpha)=C_\F(\E^\alpha)$. Lemma~\ref{L:CF0ECollect} gives now that $(C_\L(\H),\delta(C_{\F_0}(\E)),C_{S_0}(\H))$ is a regular locality over $C_{\F_0}(\E)$, and $(C_\L(\H^g),\delta(C_\F(\E^\alpha)),C_S(\H^g))$ is a regular locality over $C_\F(\E^\alpha)$. Moreover, $\tilde{\alpha}:=c_g\in\Aut(\L)$ induces an isomorphism from $C_\L(\H)$ to $C_\L(\H^g)$ which takes $C_{S_0}(\H)$ to $C_S(\H^g)$. Thus, it follows from Lemma~\ref{L:AutLFusion} that
\begin{equation}\label{E:ConjCF0E}
C_{\F_0}(\E)^{\tilde{\alpha}}=\F_{C_{S_0}(\H)}(C_\L(\H))^{\tilde{\alpha}}=\F_{C_{S_0}(\H)^g}(C_\L(\H)^g)=\F_{C_S(\H^g)}(C_\L(\H^g))=C_\F(\E^\alpha).
\end{equation}
Let now $P,Q\leq C_S(\H)=C_S(\E)$ and $\phi\in\Hom_\F(P,Q)$. As $\phi$ is a morphism in $\F$, Lemma~\ref{L:CFEWeaklyInvariant} yields that $\phi$ is a morphism in $C_\F(\E)$ if and only if it is a morphism in $C_{\F_0}(\E)$. By \eqref{E:ConjCF0E}, this is the case if and only if $\alpha^{-1}\phi\alpha=\tilde{\alpha}^{-1}\phi\tilde{\alpha}\in \Hom_{C_\F(\E^\alpha)}(P\alpha,Q\alpha)$. This yields the first part of the assertion.
\smallskip
If $\E$ is fully centralized, then $|C_S(\E)|\geq |C_S(\E^\alpha)|$ and hence $C_S(\E)\alpha=C_S(\E^\alpha)$. Moreover, $C_{\F_0}(\E)=C_\F(\E)$ by Proposition~\ref{P:CFEDescribe}. Thus, by \eqref{E:ConjCF0E}, $\alpha$ induces an isomorphisms from $C_\F(\E)$ to $C_\F(\E^\alpha)$.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction} \label{sec:intro}
Multi-domain dialogue state tracking (DST) has been an important challenge of conversational artificial intelligence~\cite{ram2018conversational}.
Compared to single-domain DST, the belief state of a multi-domain DST contains slot types across multiple domains.
It causes difficulty to the traditional DST models that assume the complete ontology is available because a complete ontology becomes hard to obtain~\cite{wu2019global} when there are many domains,
To tackle these challenges, many of the mainstream approaches for DST formulate this as a span prediction task~\cite{xu-hu-2018-end, wu-etal-2019-transferable, kumar2020ma}.
Instead of predicting a predefined value in the ontology, the span-based DST models predict slot values by extracting a span from the conversation between the user and the system.
By directly extracting spans as slot values, the DST models are able to handle unseen slot values and are potentially transferable to different domains.
However, the span-based DST models can only deal with the slot values that are explicitly expressed as a sequence of tokens and fall short in dealing with coreference ("I'd like a restaurant in the same area") and implicit choice ("Any of those is ok")
The current state-of-the-art DST model TripPy~\cite{heck-etal-2020-TripPy} is designed to deal with these problems.
In addition to extracting values directly from the user utterance, TripPy maintains two additional memories on the fly and uses them to deal with the coreference and implicit choice challenges.
The design details of TripPy will be further illustrated in Section~\ref{subsec:TripPy}
Despite the success of these span-based models, one concern is that they often consider the slots conditionally independently.
For example, in TripPy, the conversation is first encoded by BERT~\cite{devlin-etal-2019-bert}, and then the extracted features are used by n modules for the n slot types independently.
That implies, conditioning the features extracted by BERT, the slots are predicted independently.
This may increase the difficulty of this task because these slot types are indeed not independent of each other.
If the value of one slot type in a domain is present in an utterance, then it is more likely that the value of another slot type of the same domain is present.
In particular, in this work, we focus on the cases where modeling the joint probability of the slots could possibly be helpful.
We notice that there are some slot types that share the same data type.
The data types of both the slot type "hotel-people" and the slot type "restaurant-book people" are both integers.
Therefore, when an integer is present in an utterance, a DST model needs to predict which slot(s) this integer is for.
Given the fact that some slots co-occur more often than others, modeling the slot types jointly may be helpful when making this kind of prediction.
Based on this intuition, we group the slot types into 4 groups by their data type: time, place name, integer.
We then analyze the current state-of-the-art model TripPy~\cite{heck-etal-2020-TripPy,MehriDialoGLUE2020}.
Our analysis shows that TripPy is indeed often confused with slot types in the same group.
We then investigate whether modeling the relation between slot types explicitly can alleviate this problem.
By modeling the slot types jointly, information from the slots that the model is more confident with may reduce the confusion for the other slots.
Specifically, to achieve this goal, we propose to model the joint probability of slots with MRF and LSTM models.
They can be plugged into any DST model that models the slots conditionally independently.
In this work, we choose the current state-of-the-art DST model, TripPy pre-trained with DialoGLUE \cite{MehriDialoGLUE2020}.
We experiment on the most widely used multi-domain DST dataset MultiWoZ 2.1 \cite{zang2020MultiWoZ,eric2019MultiWoZ}.
The results show that both the MRF and the LSTM models are able to mitigate the confusion, and can improve the performance consistently.
We push the state-of-the-art performance on MultiWoZ 2.1 from 58.7 to 61.3.
\section{Background}
\subsection{Problem Formulation}
The dialogue state tracking (DST) task is to predict values of slot types for each turn in task-oriented dialogue. Specifically, dialogue consists of a sequence of user and system utterances.
The input of the task is a dialogue $X = \{(U_1, M_1), \cdots, (U_T, M_T)\}$,
where $U_t$ and $M_t$ are utterances by user and system respectively, and $T$ is the total number of turns in this dialogue.
The output of the task is the dialogue state at each time step.
A dialogues state consists of $N$ slot types.
The goal of a DST task is to predict the slot values of those $N$ slot types at each time step.
The slot value $S^{(t)}_{s}$ of the s-th slot type at time step $t$ might be a span in the previous conversation of simply \textit{none}.
We denote the state at time step $t$ as $S^{(t)} = \{S^{(t)}_{1}, S^{(t)}_2, \cdots, S^{(t)}_N \}$.
\subsection{TripPy: Triple Copy Strategy for DST} \label{subsec:TripPy}
As described in Section~\ref{sec:intro}, in addition to extract values from the user utterance, TripPy maintains two memories to tackle the coreference and the implicit choice problems in the span-based DST model.
Specifically, we use \textit{system inform memory} to remember the slots which the model previously informed, and \textit{DS memory} to store the seen slots in the dialogue history.
At dialogue turn $t$, TripPy predicts the belief state $S^{(t)}$ with the following steps:
1) The model first uses a pretrained BERT model to encode the dialogue history until the current time step $t$.
2) For each slot type $s$, it uses a classifier to predict the \textit{slot class} $C^{(t)}_s$.
For slot types whose possible values are "yes" or "no", possible values of $C^{(t)}_s$ are \{\textit{none}, \textit{dontcare}, \textit{yes}, \textit{no}\}.
For the other slot types, possible values of $C^{(t)}_s$ are \{\textit{none}, \textit{dontcare}, \textit{span}, \textit{inform}, \textit{refer}\}.
These slot classes determine which source the slot value should be copied from.
3) The first two slot classes express special cases.
\textit{none} denotes that the slot does not take a value in this turn, and \textit{dontcare} states that any value is acceptable for this slot.
If $C^{(t)}_s$ is \textit{span}, then TripPy uses a span prediction model to copy a span of text from the dialogue history as the prediction of a slot type.
On the other hands, if $C^{t}_s$ is \textit{inform} and \textit{refer}, the model will instead copy the value from \textit{system inform memory} and \textit{DS memory} respectively.
The \textit{system inform memory} allows the model to solve the implicit choice issue and the \textit{DS memory} helps the model solve coreference problems.
\subsection{Experimental Setup}
We train and test our model on MultiWoZ 2.1 \cite{eric2019MultiWoZ}, which is the most challenging dataset for DST and is widely used for the evaluation of multi-domain DST models.
MultiWoZ 2.1 is comprised of over 10000 dialogues in 5 domains and has 30 different slots with over 45000 possible values.
We compute the joint goal accuracy (JGA) on all test samples to evaluate our models.
The JGA is the ratio of dialog turns in the dataset for which all slots have been filled with the correct value according to the ground truth.
Our baseline is the TripPy model in \cite{MehriDialoGLUE2020}, which is additionally pretrained with more dialogue tasks.
To reproduce their result, we follow the setting of their hyperparameters and successfully achieve JGA 58.0, which is close to the reported result of 58.7 in the paper.
\section{Analysis}
\subsection{Accuracy per Copy Class}
\begin{table*}[]
\tiny
\centering
\begin{tabular}{l|ccccccc|cc}
\toprule
Slot Type & none & don't care & span & inform & refer & true & false & Oracle & $\Delta$ \\
\midrule
taxi-leaveAt & .99 (7218) & - (0) & .88 (131) & .91 (23) & .00 (2) & . - (0) & . - (0) & .581 & .001 \\
taxi-destination & .98 (7159) & - (0) & .98 (108) & .94 (18) & .61 (89) & . - (0) & . - (0) & .587 & .007\\
taxi-departure & .98 (7169) & - (0) & .89 (100) & 1.0 (21) & .60 (84) & . - (0) & . - (0) & .587 & .007\\
taxi-arriveBy & .99 (7278) & .50 (2) & .85 (55) & 1.0 (10) & .76 (29) & . - (0) & . - (0) & .581 & .001\\
restaurant-book\_people & .99 (7001) & - (0) & .99 (319) & 1.0 (41) & .85 (13) & . - (0) & . - (0) & .583 & .002\\
restaurant-book\_day & 1.0 (7011) & - (0) & .99 (286) & 1.0 (53) & .92 (24) & . - (0) & . - (0) & .581 & .001\\
restaurant-book\_time & .99 (6979) & - (0) & .98 (328) & 1.0 (67) & . - (0) & . - (0) & . - (0) & .583 & .003\\
restaurant-food & .99 (6801) & .45 (20) & 1.0 (422) & .98 (131) & . - (0) & . - (0) & . - (0) & .593 & .013\\
restaurant-pricerange & .99 (6906) & .78 (9) & .98 (343) & .98 (113) & 1.0 (3) & . - (0) & . - (0) & .587 & .006\\
restaurant-name & .96 (6926) & .00 (2) & .93 (162) & 1.0 (283) & .00 (1) & . - (0) & . - (0) & .598 & .018\\
restaurant-area & .98 (6881) & .71 (17) & .97 (314) & .96 (137) & .88 (25) & . - (0) & . - (0) & .594 & .014\\
hotel-book\_people & .99 (7031) & - (0) & .99 (300) & .97 (35) & .62 (8) & . - (0) & . - (0) & .581 & .001\\
hotel-book\_day & 1.0 (7035) & - (0) & .98 (289) & 1.0 (40) & .70 (10) & . - (0) & . - (0) & .582 & .002\\
hotel-book\_stay & .99 (7015) & - (0) & .99 (314) & 1.0 (45) & . - (0) & . - (0) & . - (0) & .582 & .001\\
hotel-name & .96 (6884) & .00 (1) & .98 (170) & 1.0 (319) & . - (0) & . - (0) & . - (0) & .59 & .009\\
hotel-area & .97 (7028) & .45 (29) & 1.0 (205) & .98 (92) & .65 (20) & . - (0) & . - (0) & .595 & .014\\
hotel-parking & .97 (7204) & .33 (6) & . - (0) & . - (0) & . - (0) & .86 (154) & .50 (10) & .587 & .007\\
hotel-pricerange & .98 (6986) & .67 (21) & .97 (271) & 1.0 (91) & .60 (5) & . - (0) & . - (0) & .59 & .01 \\
hotel-stars & .99 (7094) & .50 (6) & .98 (210) & .91 (64) & . - (0) & . - (0) & . - (0) & .584 & .004\\
hotel-internet & .98 (7200) & 1.0 (2) & . - (0) & . - (0) & . - (0) & .89 (169) & .67 (3) & .588 & .007\\
hotel-type & .93 (7135) & .00 (5) & . - (0) & . - (0) & . - (0) & .59 (80) & .85 (154) & .612 & .031\\
attraction-type & .98 (6835) & .38 (8) & .98 (430) & 1.0 (101) & . - (0) & . - (0) & . - (0) & .595 & .015\\
attraction-name & .98 (7006) & .00 (2) & .97 (137) & 1.0 (229) & . - (0) & . - (0) & . - (0) & .59 & .01 \\
attraction-area & .98 (6946) & .39 (18) & .98 (248) & .96 (135) & .93 (27) & . - (0) & . - (0) & .593 & .013\\
train-book\_people & .98 (7040) & - (0) & .96 (314) & 1.0 (8) & .42 (12) & . - (0) & . - (0) & .591 & .011\\
train-leaveAt & .97 (7125) & .50 (2) & .96 (228) & .68 (19) & . - (0) & . - (0) & . - (0) & .612 & .032\\
train-destination & .98 (6693) & - (0) & .99 (503) & 1.0 (176) & .00 (2) & . - (0) & . - (0) & .586 & .006\\
train-day & .99 (6791) & - (0) & 1.0 (481) & 1.0 (86) & .94 (16) & . - (0) & . - (0) & .585 & .005\\
train-arriveBy & .99 (7034) & .75 (4) & .98 (322) & .75 (12) & .00 (2) & . - (0) & . - (0) & .592 & .012\\
train-departure & .98 (6749) & - (0) & .99 (478) & 1.0 (146) & .00 (1) & . - (0) & . - (0) & .587 & .007\\
\midrule
sum & 3928 & 77 & 168 & 73 & 29 & 116 & 37 \\
\bottomrule
\end{tabular}
\caption{Accuracy of each slot type by ground truth class. The number in the parenthesis is the number of occurrence of the class for the slot type. The oracle column is the performance calculated when presuming the slot value is predicted correctly.}
\label{tab:slot-accuracy}
\end{table*}
\begin{table}[]
\centering
\begin{tabular}{l|ccc}
\toprule
& Other & FN & FP \\
\midrule
taxi-leaveAt & 13 & 7 & 25 \\
taxi-destination & 10 & 32 & 13 \\
taxi-departure & 14 & 37 & 13 \\
taxi-arriveBy & 1 & 15 & 19 \\
restaurant-book\_people & 1 & 5 & 37 \\
restaurant-book\_day & 1 & 3 & 16 \\
restaurant-book\_time & 3 & 4 & 12 \\
restaurant-food & 3 & 13 & 48 \\
restaurant-pricerange & 4 & 8 & 57 \\
restaurant-name & 9 & 7 & 45 \\
restaurant-area & 3 & 19 & 57 \\
hotel-book\_people & 2 & 6 & 34 \\
hotel-book\_day & 1 & 7 & 8 \\
hotel-book\_stay & 2 & 2 & 17 \\
hotel-name & 6 & 3 & 72 \\
hotel-area & 2 & 24 & 94 \\
hotel-parking & 11 & 20 & 88 \\
hotel-pricerange & 5 & 13 & 38 \\
hotel-stars & 4 & 10 & 24 \\
hotel-internet & 1 & 18 & 85 \\
hotel-type & 12 & 49 & 220 \\
attraction-type & 2 & 11 & 104 \\
attraction-name & 21 & 5 & 31 \\
attraction-area & 0 & 25 & 53 \\
train-book\_people & 1 & 19 & 37 \\
train-leaveAt & 3 & 14 & 57 \\
train-destination & 6 & 4 & 12 \\
train-day & 1 & 2 & 17 \\
train-arriveBy & 3 & 9 & 24 \\
train-departure & 3 & 6 & 8 \\
\midrule
& 148 & 397 & 1365 \\
\bottomrule
\end{tabular}
\caption{Number of incorrect slot value predictions in different cases of \textit{none} prediction. FN (false negative)/FP (false positive) means that the ground truth class is \textit{none}/is not \textit{none}, but the model predicts the class incorrectly.}
\label{tab:none-accuracy}
\end{table}
\iffalse
\begin{table*}[]
\tiny
\centering
\begin{tabular}{l|ccccccc}
\toprule
Slot Type & ALL & ALL-SW & Context-1 & Context-2 & Verb & Slot Value & ALL-SV \\
\midrule
taxi-leaveAt & 3.4/8.2 & 4.4/16.3 & 1.8/2.1 & 3.6/6.4 & 28.5/$>100$ & 3.2/3.2 & 2.7/6.2 \\
taxi-destination & 3.3/16.4 & 3.9/68.8 & 1.8/2.0 & 3.4/5.8 & 6.2/$>100$ & 5.6/13.4 & 2.9/11.3 \\
taxi-departure & 3.2/18.1 & 4.0/72.4 & 2.2/2.6 & 3.4/4.3 & 7.6/$>100$ & 5.4/29.4 & 2.8/7.8 \\
taxi-arriveBy & 2.7/19.2 & 3.0/64.8 & 1.6/1.5 & 2.9/5.2 & 5.4/$>100$ & 3.0/3.1 & 2.3/13.9 \\
restaurant-book\_people & 4.9/8.4 & 6.9/14.0 & 1.5/1.7 & 3.0/4.2 & 22.0/$>100$ & 3.4/3.5 & 4.3/7.2 \\
restaurant-book\_day & 4.8/7.6 & 6.8/12.7 & 1.7/2.0 & 3.4/3.8 & 25.9/$>100$ & 2.7/2.7 & 4.4/6.8 \\
restaurant-book\_time & 5.2/7.9 & 6.9/12.0 & 1.6/1.7 & 3.5/4.6 & 28.0/$>100$ & 2.6/2.9 & 4.1/6.0 \\
restaurant-food & 6.6/9.3 & 10.7/16.8 & 5.0/5.6 & 5.2/6.7 & 63.5/$>100$ & 5.8/5.9 & 4.9/6.6 \\
restaurant-pricerange & 5.0/9.5 & 7.6/17.1 & 3.4/5.2 & 4.4/7.6 & 46.1/$>100$ & 2.2/2.0 & 4.7/9.3 \\
restaurant-name & 4.5/16.1 & 5.6/34.5 & 3.6/4.3 & 4.6/7.9 & 11.3/$>100$ & 5.6/7.2 & 3.8/9.5 \\
restaurant-area & 4.6/7.9 & 7.6/16.2 & 1.3/1.4 & 1.5/1.8 & 22.4/$>100$ & 2.3/2.2 & 4.1/7.0 \\
hotel-book\_people & 5.3/7.8 & 6.5/10.8 & 2.1/2.4 & 4.1/5.6 & 16.7/$>100$ & 3.4/3.5 & 4.8/7.2 \\
hotel-book\_day & 5.2/7.9 & 6.4/10.7 & 2.4/2.7 & 3.2/4.4 & 18.1/$>100$ & 2.7/2.7 & 4.6/6.9 \\
hotel-book\_stay & 5.4/8.1 & 6.8/10.5 & 2.4/2.5 & 4.0/4.7 & 18.2/$>100$ & 2.6/2.6 & 4.9/7.1 \\
hotel-name & 4.3/10.7 & 5.1/16.3 & 2.7/3.0 & 4.1/6.3 & 11.0/$>100$ & 3.4/3.4 & 3.7/8.1 \\
hotel-area & 4.6/9.5 & 7.2/18.7 & 1.2/1.9 & 1.6/2.3 & 13.6/$>100$ & 2.4/2.4 & 4.1/7.9 \\
hotel-parking & NA/NA & NA/NA & NA/NA & NA/NA & NA/NA & NA/NA & NA/NA \\
hotel-pricerange & 4.8/9.4 & 6.5/15.1 & 3.3/4.0 & 4.1/9.3 & 16.8/$>100$ & 2.2/2.3 & 4.6/9.1 \\
hotel-stars & 4.7/8.6 & 6.8/14.1 & 3.0/3.2 & 4.8/5.7 & 14.0/$>100$ & 1.9/2.0 & 4.5/8.0 \\
hotel-internet & NA/NA & NA/NA & NA/NA & NA/NA & NA/NA & NA/NA & NA/NA \\
hotel-type & 1.0/NA & 1.0/NA & 1.0/NA & 1.0/NA & 1.0/NA & 1.0/NA & 1.0/NA \\
attraction-type & 5.9/9.8 & 10.7/22.9 & 3.5/4.4 & 5.4/8.5 & 25.1/$>100$ & 4.2/3.9 & 5.1/8.4 \\
attraction-name & 3.7/13.3 & 4.4/25.6 & 3.4/3.7 & 4.3/5.5 & 11.8/$>100$ & 3.8/3.8 & 3.1/8.8 \\
attraction-area & 4.1/7.7 & 7.1/18.3 & 1.2/1.3 & 1.5/1.6 & 13.7/$>100$ & 2.2/2.2 & 3.9/7.0 \\
train-book\_people & 4.8/8.4 & 5.9/12.3 & 2.5/2.5 & 4.2/4.8 & 19.1/$>100$ & 3.7/3.7 & 4.0/7.2 \\
train-leaveAt & 4.3/6.6 & 6.1/10.8 & 1.3/1.3 & 2.9/4.3 & 10.3/$>100$ & 2.9/3.1 & 3.3/4.9 \\
train-destination & 4.9/6.0 & 8.0/9.9 & 1.6/1.7 & 3.3/3.8 & 25.0/$>100$ & 2.3/2.3 & 4.5/5.5 \\
train-day & 5.0/6.8 & 8.0/11.2 & 1.6/1.7 & 3.8/4.5 & 64.0/$>100$ & 2.7/2.7 & 4.3/5.6 \\
train-arriveBy & 4.1/6.1 & 5.9/10.4 & 1.3/1.2 & 2.2/2.7 & 7.6/$>100$ & 2.8/2.9 & 3.2/4.7 \\
train-departure & 4.8/5.6 & 7.5/8.9 & 1.8/1.7 & 3.5/3.5 & 32.2/$>100$ & 2.4/2.3 & 4.6/5.2 \\
\bottomrule
\end{tabular}
\caption{Perplexity (PPL) of the LSTM language model trained on different parts of the data. The PPL computed on training and development set is separated by / (the left one is the training PPL). NA indicates there is no enough data to train a language model. We use $> 100$ to represent the PPL which is larger than $100$ to save the space. }
\label{tab:ppl}
\vspace{-4mm}
\end{table*}
\begin{table*}[]
\tiny
\centering
\begin{tabular}{l|cc|ccccccc}
\toprule
Slot Type & Acc. & Majority & ALL & ALL-SW & Context-1 & Context-2 & Verb & Slot Value & ALL-SV \\
\midrule
taxi-leaveAt & 87.2 & 86.6 & 2 & 6 & 5 & 3 & 7 & 1 & 4 \\
taxi-destination & 83.9 & 81.4 & 2 & 4 & 6 & 5 & 7 & 3 & 1 \\
taxi-departure & 81.2 & 80.8 & 2 & 6 & 4 & 3 & 7 & 5 & 1 \\
taxi-arriveBy & 86.9 & 85.0 & 5 & 6 & 4 & 3 & 7 & 1 & 2 \\
restaurant-book\_people & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
restaurant-book\_day & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
restaurant-book\_time & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
restaurant-food & 99.4 & 99.4 & 3 & 5 & 4 & 6 & 7 & 2 & 1 \\
restaurant-pricerange & 99.2 & 99.2 & 3 & 4 & 5 & 6 & 7 & 2 & 1 \\
restaurant-name & 91.4 & 91.1 & 4 & 5 & 3 & 6 & 7 & 2 & 1 \\
restaurant-area & 96.7 & 96.7 & 3 & 4 & 6 & 5 & 7 & 1 & 2 \\
hotel-book\_people & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
hotel-book\_day & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
hotel-book\_stay & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
hotel-name & 95.3 & 95.3 & 4 & 5 & 3 & 6 & 7 & 1 & 2 \\
hotel-area & 98.3 & 98.3 & 2 & 4 & 1 & 6 & 7 & 3 & 5 \\
hotel-parking & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
hotel-pricerange & 97.7 & 97.6 & 3 & 5 & 6 & 2 & 7 & 4 & 1 \\
hotel-stars & 98.2 & 98.2 & 4 & 1 & 3 & 2 & 7 & 5 & 6 \\
hotel-internet & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
hotel-type & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
attraction-type & 97.6 & 97.8 & 2 & 3 & 6 & 5 & 7 & 4 & 1 \\
attraction-name & 98.6 & 95.9 & 4 & 5 & 6 & 3 & 7 & 2 & 1 \\
attraction-area & 97.0 & 97.0 & 3 & 2 & 5 & 1 & 7 & 6 & 4 \\
train-book\_people & NA & NA & NA & NA & NA & NA & NA & NA & NA\\
train-leaveAt & 97.1 & 95.3 & 4 & 6 & 2 & 5 & 7 & 1 & 3 \\
train-destination & 97.9 & 97.9 & 1 & 2 & 6 & 5 & 7 & 4 & 3 \\
train-day & 99.4 & 99.5 & 2 & 1 & 6 & 4 & 7 & 5 & 3 \\
train-arriveBy & 98.1 & 97.0 & 2 & 5 & 6 & 4 & 7 & 1 & 3 \\
train-departure & 98.2 & 97.9 & 2 & 1 & 5 & 6 & 7 & 3 & 4 \\
\midrule
Mean & 95.8 & 95.3 & 2.7 & 3.7 & 4.8 & 4.4 & 7.0 & 3.0 & 2.4 \\
\bottomrule
\end{tabular}
\caption{Logistic regression results and the feature importance ranks. The model uses PPLs of language models to predict whether the DST model will make mistake. Majority column shows the accuracy of simply predicting the most frequent class. If the feature importance rank is lower, such feature is more important. NA indicates there is no enough data to train the logistic regression model.}
\label{tab:ppl-regression}
\end{table*}
\fi
We analyze the accuracy of each slot by the ground truth copy strategy to use on the validation set.
For each slot type, we also calculate the joint goal accuracy presuming it is predicted correctly.
From the results in Table~\ref{tab:slot-accuracy} we can have a few observations.
1) The class \textit{none} takes most of the cases. Therefore, it is important to predict whether the value is \textit{none} correctly.
2) On the other hand, improving other classes may be less effective, since the number of mistakes is much smaller.
The above results encourage us to inspect the prediction of the class \textit{none} more closely.
Here we focus on the cases where the model is to decide whether it is going to give a value to a slot type.
We ignore cases where the value of the slot type in the previous turn is not none.
We calculate the number of incorrect value predictions when \textit{none} is predicted incorrectly.
From Table~\ref{tab:none-accuracy}, we can see that many incorrect predictions are resulted from incorrect \textit{none} prediction.
\subsection{Relation Between Slots} \label{subsec:incorrect_freq}
\begin{figure*}
\centering
\includegraphics[width=0.70\linewidth]{figures/violation.pdf}
\caption{The value of $p_1 + p_2$ as described in section \ref{subsec:incorrect_freq}.}
\label{fig:violation}
\end{figure*}
We further examine the model's behavior when \textit{none} is not predicted correctly.
We check the frequency of the following cases of incorrect prediction:
1) $p_1$: when the ground truth class is \textit{none}, the cases where the predicted slot value is the ground truth value of another slot type.
2) $p_2$: when the predicted class is \textit{none}, cases of the slot value that is the ground truth value of another slot type.
We plot the frequency $p_1 + p_2$ in Figure~\ref{fig:violation}.
From the figure, we can see that the model is often confused by similar slot types.
For example, the value of the slot type "taxi-departure" and the value of the slot type "taxi-ariveBy" is time.
It may be the reason why the model often fails to give non-\textit{none} value to the correct slot type.
\iffalse
\subsection{Text Similarity between Training and Validation Data}
For each slot, we are interested in its input text similarity between the training and validation data.
If there is an dataset bias for a certain slot, the model might be able to exploit it and easily predict the correct answer.
For example, if the verb $w$ always occur with the slot $s$ and its slot value $v$, the model can use $w$ as an important feature while the verb $w$ might be totally not related to the slot $s$ in both semantic and syntactic aspects.
Similarly, if the slot value $v$ of the slot $s$ is always the same across training and validation data, then the DST model don't need to consider slot $s$ when there is no $v$ in the sentence, which is not the desired behavior.
Let the input sentence be $\{w_1, w_2, \dots, w_n\}$ which contains slot value $\{w_i, w_{i+1} \dots, w_j\}$ where $0 \leq i < j \leq n $, we consider following subsets of the input text whose similarity might be important for the model performance:
\begin{itemize}[noitemsep]
\item \textbf{ALL}: The entire input sentence.
\item \textbf{ALL-SW}: The input sentence without stop words.
\item \textbf{Context-1}: The word $w_{i-1}$ before the slot value.
\item \textbf{Context-2}: The two words $\{w_{i-2}, w_{i-1}\}$ before the slot value.
\item \textbf{Verb}: The verbs in the input sentence.
\item \textbf{Slot Value}: The slot value $\{w_i, \dots, w_j \}$.
\item \textbf{ALL-SV}: The input sentence with slot value replaced with mask token.
\end{itemize}
We train LSTM language models on these different subsets of texts in training data, and report the training and validation perplexity (PPL) on the same subset of texts.
If the training and validation PPL are similar, then we can say the text distribution is very similar.
The results is presented in Table~\ref{tab:ppl}.
Different to our initial guess that verbs is an important indicator of the slots, the results show that it is hard to use only verbs to train a language model and gives us very high PPL on both training and validation data.
On the other hand, there are no much PPL differences on \textbf{Context-1}, \textbf{Context-2} and \textbf{Slot value}, which indicates that for a certain slot, the slot values and its context are very similar in both training and validation data.
We can also observe that when we use the whole input sentence (\textbf{ALL}), certain slots such as \textit{train-leaveAt} and \textit{train-departure} has very similar training and validation PPL.
This might be a reason that model performs better on these slots.
The observation further motivates us to ask "can we use these PPL to predict whether the DST model will make mistakes?"
If we can, which subset of the text is the most important one?
To answer this question, we uses the prediction of DST model on the validation data to train a logistic regression model.
For each input text, the logistic model takes the PPL of different subsets of text from corresponding language models as input, and predict whether the DST model will make mistake.
The accuracy and the importance of each text subset measured by feature weights (the lower is more important) are in Table~\ref{tab:ppl-regression}.
We also show the accuracy of only predicting the most frequent class as baseline (Majority).
Although there is a serious data imbalance problem, we can still observe the PPL is useful for predicting whether the model can detect slots correctly.
Surprisingly, the PPL of \textbf{Slot Value}, \textbf{Context-1} and \textbf{Context-2} are not the most important features, but \textbf{ALL} and \textbf{ALL-SV} have the highest feature weights.
It shows that while the text similarity between train and validation data is important, the DST model might not only rely on some simple features of the input text to do the detection.
\fi
\begin{figure*}[h!]
\centering
\includegraphics[width=0.5\linewidth]{figures/mrf.pdf}
\caption{The value of $p_1 + p_2$ (Section~\ref{subsec:incorrect_freq}) predicted by the TripPy + MRF model.}
\label{fig:mrf_violation}
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[width=0.5\linewidth]{figures/lstm_violation.pdf}
\caption{The value of $p_1 + p_2$ (Section~\ref{subsec:incorrect_freq}) predicted by the TripPy + LSTM model.}
\label{fig:lstm_violation}
\end{figure*}
\section{Proposed Method}
We conjecture some mistakes made by TripPy maybe because it models the slots conditionally independently.
Based on Table~\ref{tab:none-accuracy}, the model does not predict \textit{none} very accurately, and Figure~\ref{fig:violation} indicates that it is because the model is often confused by similar slot types.
In the original design, given a dialogue $X$, TripPy models the class prediction independently as follow:
\begin{equation}
P(C^{(t)}_1, C^{(t)}_2, \cdots, C^{(t)}_N |X) = \prod_{s = 1}^{N} P(C^{(t)}_s | X).
\end{equation}
TripPy ignores the potential correlation between different slots.
Thus, for all $s$, $P(C^{(t)}_s | X)$ needs to predict the class correctly at the same time.
It may be difficult since there are numerous slots.
To alleviate this problem, we propose to directly model the joint probability $P(C^{(t)}_1, C^{(t)}_2, \cdots, C^{(t)}_N |X)$.
If $c_1, c_2, \cdots, c_N$ are the correct classes, modeling the joint probability does not require
$P(C^{(t)}_s = c_s) > P(C^{(t)}_s = c'_s)$ for all $s$ and all incorrect assignments $\{c'_s\}_{s=1}^N$.
Instead, to predict the belief state correctly, the model is only required to ensure
$P(C^{(t)}_1=c_1, C^{(t)}_2=c_2, \cdots, C^{(t)}_N=c_N |X) > P(C^{(t)}_1 = c'_1, C^{(t)}_2 = c'_2, \cdots, C^{(t)}_N = c'_N |X)$.
In this way, information from the other states could possibly help the prediction of difficult slot types.
For example, if the model has high confidence that the utterance contains the value of slot "restaurant-name", the number in the utterance is more likely the value of slot "restaurant-book people".
Therefore, confusion shown in Figure~\ref{fig:violation} could possibly be alleviated.
\subsection{Modeling the Joint Probability with Graphical Models}
One approach that models the joint probability is to use a Markov random field (MRF).
This approach is parameter-efficient and incurs little computation overhead for inference.
However, having a large number of vertices in an MRF makes training intractable.
Therefore, we follow the intuition above, grouping the slots by their domains, and model the joint distribution within each group respectively.
Since a large portion of the mistakes made by the original TripPy model is due to the incorrect prediction of whether a slot value is present in an utterance, we further decompose the probability of a class as
\begin{equation}
\begin{split}
&P(C^{(t)}_s = c) \\
&= P(C^{(t)}_s = c | C^{(t)}_s \ne \text{"\textit{none}"}) P(C^{(t)}_s \ne \text{"\textit{none}"} )
\end{split}
\end{equation}
Let $C'^{(t)}_s$ be the random variable indicating whether $C^{(t)}_s = \text{"\textit{none}"}$.
Modeling only the $P(\{ C'^{(t)}_s \}_{s=1}^{N})$ part jointly makes the training and inference of the MRF model feasible: given input $x$, for each domain $d$ and the assignment of slots $c_1, c_2, \cdots, c_N$ for the $N$ slots in $d$, the probability of the assignment is
\begin{align}
\begin{split}
& P(c_1, c_2, \cdots, c_N | x) \\
&= \frac{1}{Z(x)} \Phi_{d}(\{C'^{(t)}_s = c'_s\}_{s=1}^N) \\
& \prod_{s=1}^N \Phi_s(c_s, x) P(C^{(t)}_s = c_s | C'^{(t)}_s = c'_s),
\end{split}
\label{eq:potential}
\end{align}
where $\Phi_{d}$ map each assignment $c'_1, c'_2, \cdots, c'_N$ to a learnable scalar, and $\Phi_s(c'_s, x)$ is a scalar resulted from applying a linear layer over the pooled BERT output.
Since the number of slot types in a domain is at most 10, and all of the variables are binary, the partition function $Z_i(x)$ can be calculated by enumerating all the possible assignments of $c'_1, c'_2, \cdots, c'_N$.
Therefore, we can train this model with the plain MLE loss.
\subsection{Modeling with LSTM models}
Another approach we explore is to use the LSTM model~\cite{schmidhuber1997long}.
Similar to language modeling, we model the joint probability by predicting the value of each slot sequentially:
\begin{equation}
\begin{split}
&P(C^{(t)}_1, C^{(t)}_2, \cdots, C^{(t)}_N |X) \\ \
&= \prod_{s = 1}^{N} P( C^{(t)}_s | X, C^{t}_1, \dots, C^{t}_{s-1}),
\end{split}
\label{eq:lstm}
\end{equation}
For the implementation of this model, we first encode the dialogue history $X$ with BERT and extract its CLS token representation $h_X$.
The original TripPy model will feed $h_X$ into a linear layer $f_{s}$ to predict slot class for each slot.
\begin{equation}
P(C^{(t)}_s|X) = Softmax(f_{s}(h_X))
\end{equation}
Here, for all $s$, we additionally train another linear layer $f^{lstm}_{s}$ to transform $h_X$ and feed it into a LSTM layer:
\begin{equation} \label{eq:lstm}
h_s = LSTM(f^{lstm}_{s}(h_X), h_{s-1})
\end{equation}
We then concatenate $h_s$ and $h_x$ to model the conditional probability:
\begin{equation} \label{eq:predict_lstm}
P(C^{(t)}_s | X, C^{(t)}_1, \dots, C^{(t)}_{s-1}) = f_{s}([h_X;h_s])
\end{equation}
where $[x;y]$ denotes the concatenation of vectors $x$ and $y$.
We name this model TripPy+LSTM.
\section{Results and Discussion}
\begin{table}[]
\centering
\begin{tabular}{l c c}
\toprule
Model & Dev. & Test \\
\midrule
TripPy~\cite{MehriDialoGLUE2020} & & 58.7 \\
TripPy (reproduced) & 61.3 & 58.0 \\
+ MRF & 59.3 & 60.1 \\
+ LSTM & 62.8 & 61.3 \\
\bottomrule
\end{tabular}
\caption{Joint goal accuracy on MultiWoZ 2.1.}
\label{tab:main_result}
\end{table}
Our main results are shown in Table~\ref{tab:main_result}.
Both MRF and LSTM modules have a consistent improvement over the baseline model on the test set of MultiWoZ 2.1, which proves the effectiveness of our proposed approaches.
As a more fine-grained analysis, we plot the frequency of incorrect prediction as described in Section~\ref{subsec:incorrect_freq}.
Compared to Figure~\ref{fig:violation}, Figure~\ref{fig:mrf_violation} and Figure~\ref{fig:lstm_violation} show that both TripPy+MRF and TripPy+LSTM can relieve the confusion between slot types that have the same data type.
This is coherent to our intuition for modeling the slot types jointly.
The results also show that TripPy+LSTM performs better than TripPy+MRF.
It may be because TripPy+LSTM models all the assignment $\{ C_s \}_{s=1}^N$ of the classes jointly, while TripPy+MRF only models $\{ C'_s \}_{s=1}^N$ jointly.
Also, TripPy+LSTM has the potential to model the joint distribution across the domains.
However, this also sacrifices efficiency.
Since TripPy+LSTM predicts the slots sequentially, it cannot predict the slots in parallel.
TripPy+MRF could be a better choice when latency is an important issue.
\section{Related Work}
\label{sec:related_work}
\citet{williams-etal-2013-dialog} is the first benchmark for DST models.
After that, \citet{henderson-etal-2014-second} proposed the DSTC2 dataset which includes more linguistic phenomena.
\citet{budzianowski-etal-2018-multiwoz} proposed MultiWoZ, which is the first large-scale human-human multi-domain task-oriented DST dataset.
Its successive versions \cite{eric2019MultiWoZ,zang-etal-2020-MultiWoZ} cleanse the noisy conversations in it.
Since then, it has become the standard DST benchmark dataset.
Early approaches for DST rely on a list candidate slot values \cite{mrksic-etal-2017-neural,Liu2017,ramadan-etal-2018-large,rastogi2017scalable}.
Some recent approaches formulate it as a sequence generation task.
TRADE \cite{wu-etal-2019-transferable} and MA-DST \cite{kumar2020ma} generates slot values with a pointer generator with a copy mechanism.
SOM-DST \cite{kim-etal-2020-efficient} uses a BERT model to encode the dialogue. To improve the efficiency, the state from the previous turn instead of the whole dialogue history is fed into the model.
SUMBT \cite{lee-etal-2019-sumbt} utilizes a BERT model \cite{devlin-etal-2019-bert} to encode both the dialogue history and candidate slot values.
In \citet{gao-etal-2019-dialog} they formulate DST as a question answering task.
DS-DST \cite{zhang-etal-2020-find} adopts a dual strategy: they predict slot values from candidates but also predict slot values by copying from the dialogue.
TripPy \cite{heck-etal-2020-TripPy} predicts slot values by copying text span from three different sources.
In \citet{MehriDialoGLUE2020}, they improve the performance of TripPy by pre-training the BERT with multiple dialogue tasks.
Simple-TOD \cite{hosseini2020simple} and SOLOIST \cite{peng2020soloist} utilize GPT-2 for end-to-end slot value and response generation.
In \citet{SHIYANG2020CoCoCC}, they investigate the generalization capability of DST models for unseen slot values.
MinTL \cite{lin2020mintl} augments data by substituting slot values in the utterances and modifying the syntax structure.
\section{Conclusion}
In this work, we explore the relation between slot types.
Our contributions are two-fold.
1) We identify a problem with the current state-of-the-art model TripPy.
We observe that it is often confused by slots that share the same type of value.
2) We propose TripPy+MRF and TripPy+LSTM that aim at mitigating this problem.
One can choose one of them based on the trade-off between efficiency and accuracy.
Our preliminary results show that both of them are able to alleviate the confusion.
Our findings in this work lead future researchers in a promising direction to improve the performance of multi-domain DST.
\section*{Acknowledgments}
We would like to thank Graham Neubig for his in-depth discussions. We are also thankful to the anonymous reviewers for their comments on the paper.
\section{Introduction}
\label{sec:introduction}
\paragraph{Problem Statement}
Tracking dialogue states is the core component to ensure a robust manner in dialog systems. In order to fulfill a user goal, a dialogue system must be capable of extracting meaning and intent from the user input, and be able to keep and up-date this information over the continuation of the dialog \cite{young2010hidden}. This task is called dialog state tracking (DST). For example, in order to recommend a restaurant to a user, the systemn eeds to know their preferences in terms of price,location, etc. These concepts are encapsulated in an ontology, where dialogue domain (e.g., restaurant or hotel), slot (e.g., price range or location), and value (e.g. cheap or expensive) are defined. Solving this information extraction task is prerequisite for forming a belief over the dialog state
The traditional methods, which have been widely used in most commercial implementations, often adopt hand-crafted rules to select the most likely result\cite{goddeau1996form}. However, these rule-based systems are prone to frequent errors as the most likely result is not always the desired one\cite{williams2014web}.
In this survey, we first conduct a comprehensive study of the traditional as well as recent architectures and techniques used to model this task. We further transition to explore commonly used methods for Multi-domain dialog state tracking. Furthermore, we deep dive into benchmark data sets and state-of-the art models used in this task.
\section{Related Work}
\label{sec:related_work}
\cite{williams-etal-2013-dialog} is the first benchmark for DST models.
After that, \cite{henderson-etal-2014-second} proposed the DSTC2 dataset which includes more linguistic phenomena.
\cite{ramadan2018large} proposed MultiWoz, which is the first large scale human-human multi-domain task-oriented DST dataset.
Its successive versions \cite{eric2019multiwoz,zang-etal-2020-multiwoz} cleanse the noisy conversations in it.
Since then, it has become the standard DST benchmark dataset.
Early approaches for DST rely on a list candidate slot values \cite{mrksic-etal-2017-neural,Liu2017,ramadan-etal-2018-large,rastogi2017scalable}.
Some recent approaches formulate it as a sequence generation task.
TRADE \cite{wu-etal-2019-transferable} and MA-DST \cite{kumar2020ma} generates slot values with a pointer generator with a copy mechanism.
SOM-DST \cite{kim-etal-2020-efficient} uses a BERT model to encode the dialogue. To improve the efficiency, the state from the previous turn instead of the whole dialogue history is fed into the model.
SUMBT \cite{lee-etal-2019-sumbt} utilizes a BERT model \cite{devlin-etal-2019-bert} to encode both the dialogue history and candidate slot values.
In \cite{gao-etal-2019-dialog} they formulate DST as a question answering task.
DS-DST \cite{zhang-etal-2020-find} adopts a dual strategy: they predict slot values from candidates but also predict slot slot values by copying from the dialogue.
Trippy \cite{heck-etal-2020-trippy} predicts slot values by copying text span from three different sources.
In \cite{MehriDialoGLUE2020}, they improve the performance of Trippy by pre-training the BERT with multiple dialogue tasks.
Simple-TOD \cite{hosseini2020simple} and SOLOIST \cite{peng2020soloist} utilize GPT-2 for end-to-end slot value and response generation.
In \cite{SHIYANG2020CoCoCC}, they investigate the generalization capability of DST models for unseen slot values.
MinTL \cite{lin2020mintl} augments data by substituting slot values in the utterances and modifying the syntax structure.
\section{Datasets}
\label{sec:modeling}
\subsection{Benchmarking Pretrained Models}
\section{Results \& Discussion}
\label{sec:results_discussion}
\input{emnlp2020-templates/tables/benchmarking}
\input{emnlp2020-templates/tables/pos}
|
\section{Introduction}
Faraday rotation enables studies on the cosmic magnetism.
As linearly polarised radiation travels along a line-of-sight, its plane of polarisation rotates through an angle $RM \lambda^{2}$, where
\begin{equation}\label{eq:RM}
RM = 0.812\int_{\rm observer}^{\rm source}n_{\rm e}(r)B_{\parallel}(r)dr,
\end{equation}
where RM is in units of $\rm rad\,m^{-2}$, $n_{\rm e}$ is electron density in $\rm cm^{-3}$, $B_{\parallel}$ is the magnetic field strength along the line-of-sight in $\mu G$, and r is a path length in $\rm pc$.
Extragalactic compact sources, e.g., radio galaxies and quasars, are often used to probe RM at a pinpoint location on the sky plane.
The integration is between the source and the observer along the line-of-sight; any patches of magneto-ionic media that polarised radiation is transmitted through on the way to the observer cause Faraday rotation and influence the resulting RM.
In practice, it is common to have multiple Faraday rotating regions along a line of sight within a telescope beam. The RM of such 'Faraday complex' sources is expressed using the Faraday depth ($\phi$), a parameter that describes the Faraday rotation at individual Faraday screens (\citealt{Burn_1966});
\begin{equation}\label{eq:faraday_rotation}
\phi(X) = 0.812\int_{\rm observer}^{\rm X}n_{\rm e}(r)B_{\parallel}(r)dr,
\end{equation}
where $X$ is a certain position along the line-of-sight and $\phi$ is a function of $X$. In this study, we use RM and $\phi$ interchangeably. Details will be described in Section \ref{sec:atca}.
The `RM grid' technique probes a contrast in the overall distribution of RM on and off magneto-ionic structures, assuming RM sources are mostly in the background of target objects. For this statistical approach, a sufficient number of polarised sources is required in the region of interest.
Thus far, the RM source density of current radio surveys (e.g., $1 \,\rm deg^{-2}$, \citealt{Taylor_2009}) limits this approach on studies of the large-scale Galactic magnetic fields or several square-degree-size extended objects.
Upcoming radio surveys using next generation radio telescopes are expected to provide immensely denser RM grid (e.g., $25\,\rm deg^{-2}$, \citealt{Anderson_2021}).
Ever-improving RM grids are enabling new measurements of the magnetic field strength and structure in a range of objects including galaxy clusters (e.g. \citealt{Anderson_2021}), individual resolved galaxies (e.g. \citealt{Gaensler_2005}; \citealt{Mao_2008}), Galactic objects (e.g. \citealt{Harvey-Smith_2011}) and high velocity clouds (e.g. \citealt{McClure-Griffiths_2010}; \citealt{Hill_2013}).
There is ample evidence that we are surrounded by the multi-phase circumgalactic medium (CGM) in the Milky Way halo (see \citealt{Putman_2012} and references therein).
Understanding the nature of the CGM is key to understanding the evolution of galaxies, as it connects the pristine intergalactic medium (IGM) to the star-forming interstellar medium (ISM).
The Milky Way provides a unique environment for studying the complex structures of the CGM on a spatially resolved scale, which is hard to achieve by studies of distant galaxies.
Although most baryons belonging to the Milky Way are distributed in the Galactic disc, observations have reported the presence of extraplanar gas clouds at high Galactic altitude moving at a high relative velocity with respect to Galactic rotation (\citealt{Muller_1963}; \citealt{Putman_2002}; \citealt{McClure-Griffiths_2009}; \citealt{Saul_2012}). They are so-called High-Velocity Clouds (HVCs) and Intermediate-Velocity Clouds (IVCs).
As HVCs travel through the halo, they hydrodynamically interact with their surroundings. Ram pressure pushes and strips low-density structures to the opposite direction of the cloud motion, Kelvin-Helmholtz and Rayleigh-Taylor instabilities
develop turbulent mixing layer across the cloud-wind interface, which is susceptible to the cloud stripping (\citealt{Jones_1994}; \citealt{Schiano_1995}).
Taken all together, it is easy to conclude that moving clouds dissipate in a short timescale and if so, one would expect to find HVCs dominated in the warm ionized phase (\citealt{Heitsch_2009}).
However, a significant mass of observed HVCs are in cold phase HI and they are often suggested as a source of cold gas fueling the star-formation of the Milky Way (\citealt{Putman_2012}).
This indicates additional forces that stabilize HVCs so that the clouds survive their journey through the Galactic halo and deliver gas to the Galactic disc.
The magnetic field is proposed as one of the possible sources that provide stability of HVCs (\citealt{Konz_2002}; \citealt{Santillan_2004}; \citealt{Kwak_2011}; \citealt{McCourt_2015}; \citealt{Banda-Barragan_2016};\citealt{Banda-Barragan_2018}; \citealt{Gronnow_2017}; \citealt{Gronnow_2018}).
Magnetic fields can be dragged along and amplified following the motion of plasma they are embedded in.
This dragging effect leads to the field `draping' around the moving clouds when the relative velocity between the ambient medium and the clouds is high enough to overcome the tension of the ambient magnetic field (\citealt{Dursi_2008}).
Hence, the magnetic field draping is particularly efficient around HVCs due to the high relative velocity and the weak magnetic fields of the Galactic halo.
There have been attempts to observationally constrain the magnetic field strength associated with the Miky Way HVCs using the RM grid (\citealt[hereafter, McG10]{McClure-Griffiths_2010}; \citealt{Hill_2013}; \citealt{Kaczmarek_2017}; \citealt{Betti_2019}).
In this paper, we present a high-density RM grid towards the Magellanic Leading Arm/Antlia SNR regions. We combine it with the Galactic diffuse polarised emission in order to test
whether the Magellanic HVC or the Antlia SNR is the source of Faraday rotation.
The paper is structured as follows: in Section \ref{sec:method}, we present an overview of observation data and the Faraday RM synthesis technique.
In Section \ref{sec:confusion}, we introduce properties of two gaseous objects-- the Antlia SNR and the Magellanic Leading Arm-- overlapping closely with each other.
Then we explore the compact-source RM (Section \ref{Sec:RM}) and the diffuse continuum emission (Section \ref{sec:diffuse}) in the region.
The discussions on our results are presented in Section \ref{sec:discussion}, including Section \ref{sec:b_los} where we estimate the magnitude of the line-of-sight magnetic field of Antlia SNR based on the observation data presented.
Our conclusions are summarized in Section \ref{sec:summary}.
\section{Methodology}\label{sec:method}
\subsection{Compact-source rotation measure}\label{sec:rm_cat}
For constructing the RM grid covering our field-of-interest, we use two publicly available RM catalogues in addition to new radio continuum observations with the Australia Telescope Compact Array (ATCA). A summary of observational specifications are presented in the following subsections.
\subsubsection{NVSS and S-PASS/ATCA RM catalogue}
First, we have a RM catalogue from the NRAO Very Large Array Sky Survey (NVSS, \citealt{Condon_1998}) RM catalogue (\citet{Taylor_2009}) that covers the sky above a declination of $-40^{\circ}$ with the RM source density of $1\,\rm deg^{-2}$ on average.
Although it has a significant role in the search for magnetised HVCs (\citetalias{McClure-Griffiths_2010}; \citealt{Hill_2013}), it should be noted that the RMs are derived from limited coverage in frequency domain: $42\,\rm MHz$-wide bands centred at $1364.9\,\rm MHz$ and $1435.1\,\rm MHz$. Yet, the $n\phi$-ambiguity tests performed by \citet{Ma_2019} has shown that the NVSS RM catalog is mostly reliable for sources located out of the Galactic plane and in the $\phi$ range used for this study.
For the southern sky below a declination of $-1^{\circ}$, there is the S-band polarisation All-Sky Survey (S-PASS, \citealt{Carretti_2019}). S-PASS/ATCA (\citealt{Schnitzeler_2019}) catalogue is derived from a follow-up observations of sources selected from S-PASS using ATCA. It provides the first wide-band ($1.3-3.1\,\rm GHz$) polarimetry data of compact sources with the average polarised source number density of $0.2\,\rm deg^{-2}$.
\subsubsection{ATCA observations}\label{sec:atca}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{catalog_matching.pdf}
\caption{
Top: Comparison of RMs matched between different catalogues using sky coordinates (red: ATCA $\times$ NVSS, blue: ATCA $\times$ S-PASS). The size of the circles represents the angular separation between matched pairs. The grey line shows the one-to-one relation where RMs from the two catalogues are identical.
Bottom: The histogram of the separation between the matched pairs.
}
\label{fig:rm_matching}
\end{figure}
In order to increase the polarised source density, we performed follow up observations on 737 fields in the region $10h:00m:00s < \alpha< 13h:30m:00s$ and $-52.3^{\circ} < \delta < -32^{\circ}$ with ATCA. Each source was visited $\approx6$ times over a 12-hour scan and the total observation time is on average 1.5 minutes per source.
The frequency range used for our analysis is $2\,\rm GHz$ wide continuum band from $1.1\,\rm GHz$ to $3.1\,\rm GHz$, and the spectral resolution is $\approx 1\,\rm MHz$. The angular diameter of the largest beam at the lowest frequency is $13.2''$. Radio bright sources were identified using {\sc Aegean} source finder (\citealt{Hancock_2012}; \citealt{Hancock_2018}) based on Stokes $I$ clean image stacked over the entire 2GHz-wide band. The total number of sources detected in the field is about 3000, including both polarised and unpolarised sources.
We used the {\sc miriad} software package (\citealt{Sault_1995}) provided by Australia Telescope National Facility (ATNF) for data reduction and imaging. The software is particularly designed for processing radio interferometry data observed with ATCA.
Prior to imaging, we performed flagging using the \texttt{mirflag} task in {\sc miriad} which immediately flags channels whose amplitudes deviate more than 14 times the median deviation from the channel median.
For calibration of data, we used \texttt{1934-638} as a flux calibrator and \texttt{1104-445}, \texttt{1206-399}, and \texttt{1215-457} as phase calibrators, respectively for each night. The calibration process determines tables of bandpass functions, antenna gains, and polarisation leakage and corrects the observed visibility to get the ideal sky intensity distribution of radio sources.
Imaging of data was performed separately for each polarisation (Stokes $I, Q, U,$ and $V$) and each chunk of channels, where the channel width ($\delta \lambda^{2}$ in the equation \ref{eq:phi_max} below) is determined by estimating the maximum value of observable Faraday depth:
\begin{equation}\label{eq:phi_max}
\left\| \phi_{\rm max}\right\| \approx \sqrt{3}/\delta \lambda^{2}.
\end{equation}
The targeted sky is at relatively high galactic latitude where we do not normally expect to detect extremely high Faraday rotation compared to the Galactic disc as both the magnetic field strength and ionized gas density is relatively low. Therefore, we used a channel width of $20\,\rm MHz$ which limits the measurable Faraday depth to around $\left\| \phi_{\rm max} \right\| \approx 750\, \rm rad\,\rm m^{-2}$.
In order to match the spatial resolution throughout the channels, the clean images of the frequency slices were smoothed by the largest beam among them before being stacked into a cube. At the end of this imaging stage, we have Stokes $I, Q, U, V$ parameters of the identified sources as a function of frequency.
The RM of each source was estimated using the Faraday RM synthesis technique (e.g., \citealt{Burn_1966}; \citealt{Brentjens_2005}; \citealt{Heald_2009}; \citealt{Mao_2010}.)
We use {\sc rm tools 1d} software
(\citealt{Purcell_2020}) provided by Canadian Initiative for Radio Astronomy Data Analysis (CIRADA).
The idea of RM synthesis is to bring the complex polarised surface brightness $\mathcal{P}$, defined below using Stokes $Q$ and $U$, to the Faraday depth ($\phi$) domain so that one can interpret the changes in $\phi$ along the line-of-sight.
\begin{equation}\label{eq:Stokes_P}
\mathcal{P}(\lambda^{2})=Q+i\,U.
\end{equation}
A simple approach is to introduce the Faraday dispersion function which is a Fourier conjugate of $\mathcal{P}(\lambda^{2})$.
\begin{equation}
\mathcal{F}(\phi)={\frac{1}{\pi}}\int_{-\infty}^{\infty} \mathcal{P}(\lambda^{2})e^{-2i\phi\lambda^{2}}d(\lambda^{2}).
\end{equation}
However, there is an incompleteness in $\lambda^{2}$ sampling in the following inevitable reasons:
\begin{itemize}
\item Mathematically, $\lambda^{2}$ can only be positive.
\item Stokes $Q$ and $U$, therefore $\mathcal{P}(\lambda^{2})$, are measured within a bandwidth of a telescope and discretized to a finite number of channels.
\item Some channels are flagged during the data reduction process.
\end{itemize}
The limited sampling of $\lambda^{2}$ space results in sidelobes in ``observed $\mathcal{F}(\phi)$'' (hereafter, $\tilde{\mathcal{F}}(\phi)$) which is distinct from the ideal $\mathcal{F}(\phi)$. The Rotation measure spread function (RMSF), $R(\phi)$ is introduced to describe the discrepancy.
\begin{equation}
\tilde{\mathcal{F}}(\phi)=\mathcal{F}(\phi)*R(\phi)
\end{equation}
The \texttt{RM-clean} algorithm (\citealt{Heald_2009}) was introduced in order to uncover physically meaningful signal from a dirty Faraday spectrum with noise and sidelobes; a clean Faraday spectrum is obtained by deconvolving the dirty spectrum with an RMSF. The resolution of the Faraday spectra, $\delta\phi$, i.e., FWHM of the RMSF, and the largest detectable scale in $\phi$ space, $\phi_{\rm max-scale}$, are defined as follows:
\begin{equation}
\delta \phi \approx 2\sqrt{3}/ (\lambda_{\rm max}^{2}-\lambda_{\rm min}^{2}),
\end{equation}
\begin{equation}
\phi_{\rm max-scale} \approx \pi/\lambda_{\rm min}^{2}.
\end{equation}
Given the bandwidth of our data $0.009\,\rm m^{2}<$ $\lambda^{2}<0.074\,\rm m^{2}$, the estimated FWHM is $\sim53\,\rm rad\,m^{-2}$ and the maximum measurable scale is $\sim349\, \rm rad\,m^{-2}$.
We determined the Faraday depth at the strongest peak in a clean Faraday spectrum and adopted it as the Faraday depth of a source. Therefore, the Faraday depth and RM are interchangeable and the Faraday complexity is not in the scope of this study.
For the catalogue of RM sources used for our analysis, we adopted thresholds as follows.
\begin{enumerate}
\item The number of channels used as an input for the RM synthesis is larger than 40 so that a certain level of $\lambda^{2}$ coverage is achieved.
\item The lower limit of $\left\| \phi_{\rm max}\right\|$ is set to $300\,\rm rad\,\rm m^{-2}$ to ensure reasonably small separation between the channels in $\lambda^{2}$ space after flagging (see equation \ref{eq:phi_max}).
\item The signal-to-noise of the peaks in the Faraday spectra identified using the \texttt{RM-clean} algorithm is larger than 7.
\item The observed $\left|\phi\right|$ is smaller than $\left\| \phi_{\rm max} \right\| \approx 750\, \rm rad\,\rm m^{-2}$. This criterion rejects the artificial peaks that appear close to the lower and upper limit of the Faraday spectra.
\end{enumerate}
The resulting 210 sources that match the above criteria are presented in Table \ref{tab:source}.
In order to test the consistency of the observed RMs, we performed a matching between the catalogues. The pairing is based on the sky coordinates of sources.
The top panel of Figure \ref{fig:rm_matching} shows the RMs of matched pairs between ATCA $\times$ NVSS (red) and ATCA $\times$ S-PASS (blue). The size of the symbols corresponds to the angular separation between the matched pairs. The bottom panel presents the histogram of the separation.
Total 33 and 24 pairs were identified, respectively, with the maximum separation between sources limited to 1 arcmin. We found that most of the matched pairs have approximately identical RMs.
\subsection{S-PASS Galactic diffuse polarisation}
S-PASS provides maps of polarimetric data of the southern sky ($\rm Dec<-1^{\circ}$). The observation of the survey was performed with the S-band ($2.2-3.6\,\rm GHz$) receiver of the Parks radio telescope.
The FWHM of the beam for the final Stokes $I, Q, U,$ and $V$ images is $10.75'$, which is several times enhanced resolution compared with previous continuum surveys of a similar kind (e.g., \citealt{Reich_2001}). This frequency range and the angular resolution enable detailed studies of magnetism at the Galactic disc and the disc-halo interface.
We refer interested readers to \citet{Carretti_2019} for further description of the survey.
In this paper (specifically, Section \ref{sec:diffuse}), we used the Stokes $I$ map to examine the continuum emission at our region of interest and the Stokes $Q$ and $U$ maps to identify depolarised features arising on top of the diffuse polarised emission from the Galactic interstellar medium.
\section{Potential for confusion from sources along the line-of-sight}\label{sec:confusion}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{hi+halpha.pdf}
\caption{
The distribution of High-velocity HI clouds ($v_{\rm LSR} = [200-300] \,\rm km/s$; blue contour) overlaid on top of the Galactic $H\alpha$ emission ($v_{\rm LSR} = [-100, 80 ]\,\rm km/s$).
The green dotted circle is a schematic drawing of the Antlia SNR with an angular diameter of $24^{\circ}$ (\citealt{McCullough_2002}.)
}
\label{fig:hi+halpha}
\end{figure*}
In this paper, we focus on the Magellanic Leading arm region where the compact-source RM distribution was earlier studied by \citetalias{McClure-Griffiths_2010}.
There are two gaseous objects which are large in solid angle and closely overlap with each other: the Magellanic Leading Arm and the Antlia SNR. Figure \ref{fig:hi+halpha} shows the distribution of high-velocity HI emission (integrated between $200-300\,\rm km/s$; blue contours) from the Galactic All Sky Survey (GASS; \citealt{McClure-Griffiths_2009}) and the Antlia SNR bright in $H\alpha$ composite image at lower velocity range generated by \citet{Finkbeiner_2003} using the Wisconsin H-Alpha Mapper (WHAM; \citealt{Haffner_2003}), the Virginia Tech Spectral line Survey (VTSS), and the Southern $H\alpha$ Sky Survey Atlas (SHASSA; \citealt{Gaustad_2001}). In the following paragraphs, we briefly introduce some known properties of these objects.
The Magellanic Leading Arm is a stream of material tidally stripped out from the Magellanic System during the interaction between the Large Magellanic Cloud, the Small Magellanic Cloud, and the Milky Way (e.g., \citealt{Nidever_2008}; \citealt{Besla_2012}; \citealt{Lucchini_2020}).
A network of large high-velocity complexes, namely, LA $\rm I-IV$, and associated cloudlets are found in HI emission. Their distribution is extended from the Magellanic system to the high latitude sky beyond the Galactic disc.
In the field studied in this paper, only LA II and LA III are present.
Given the positive galactic latitude of the LA II and the LA III, they are often considered to have passed through the Galactic midplane (\citealt{McClure-Griffiths_2008}).
Recently, a young stellar association was discovered near the tip of the LA II using \textit{Gaia} DR2 (\textit{Price-Whelan 1}; \citealt{Price-Whelan_2019}; \citealt{Nidever_2019}).
The \textit{PW 1} star cluster is located at $28.7\,\rm kpc$ from the sun. Its estimated age is comparable with the time since the traversing of the Leading Arm through the Galactic disc, which makes the compression of the materials during the interaction a favourable explanation for the formation of the star cluster.
\citetalias{McClure-Griffiths_2010} reported the morphological agreement between the structures in the RM map and the distribution of HVCs in the field.
They suggested that the magnetic field associated with the Leading Arm II is reinforced by relatively strong magnetic fields of the disc that it penetrated before moving into the halo.
Thestrength of the coherent line-of-sight magnetic field estimated from the RM grid is $B_{\parallel}\gtrsim 6\,\rm \mu G$.
This simple calculation is valid under an assumption that the structures appearing in the RM and the HI emission are physically associated, in other words, if the HVCs are the dominant source of Faraday rotation along the sight-lines.
The Antlia supernova remnant (SNR) is located at $(l,b)=(276.5^{\circ}, +19^{\circ})$ and has a large angular diameter of $24^{\circ}$ (\citealt{McCullough_2002}).
Because such high galactic latitude and size are not common among known Galactic SNRs, whether the object is a supernova-driven remnant or not had been suspected since its first discovery by \citet{McCullough_2002}. Only recently, it is confirmed that the remnant reveals shock-driven emission regions in UV and optical lines supporting the SNR origin (\citealt{Fesen_2021}).
This SNR is bright in $H\alpha$ but weak in radio continuum, suggesting that it is a relatively evolved system.
The distance to the SNR is not well constrained, but the large angular size and features arising from interacting with nearby ISM (e.g., Gum Nebula) locate it around $60-340\,\rm pc$ away within the Galactic disc (\citealt{McCullough_2002}). Note its striking morphological coincidence with the Magellanic Leading Arms presented in Figure \ref{fig:hi+halpha}.
\section{Results} \label{sec:results}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{hi+rm.pdf}
\caption{
Raw RMs from NVSS ($+$ symbol; \citealt{Taylor_2009}), S-PASS/ATCA($\times$ symbol; \citealt{Schnitzeler_2019}), and our observations with ATCA (\protect\rotatebox[origin=c]{180}{$\text{\sffamily Y}$} symbol). The HI contour and the Antila SNR diagram are the same as Figure \ref{fig:hi+halpha}, coloured in black for clarity. The dashed boxes are regions discussed in the text.
}
\label{fig:hi+rm}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{stokes_i.pdf}
\caption{
The Stokes $I$ image from S-PASS data. The Antlia SNR diagram (green dotted circle) and the white dashed boxes are the same as Figure \ref{fig:hi+rm} above.
}
\label{fig:stokes_i}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{p.pdf}
\caption{
The linear polarisation intensity calculated from the Stokes $Q$ and $U$ of S-PASS data.}
\label{fig:p}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{grad_p.pdf}
\caption{
The normalized polarisation gradient ($|\nabla \mathcal{P}|/|\mathcal{P}|$) of S-PASS data.
}
\label{fig:p_grad}
\end{figure*}
\subsection{Structures in the compact-source RM map} \label{Sec:RM}
To construct the RM grid of the field, we combine 3 RM catalogues as explained in Section \ref{sec:rm_cat}.
We achieved the maximum RM source density of $\approx2\,\rm sources/deg^{2}$. Figure \ref{fig:hi+rm} shows interesting features standing out in the RM grid. We separate the field into several regions labeled in Figure \ref{fig:hi+rm} for convenience in the description.
\begin{enumerate}[start=1,label={Region \Roman*:},wide = 0pt, leftmargin = 4em]
\item Below galactic latitude $b \lesssim 10^{\circ}$, especially in the south-eastern side of the field, there are large RMs with high fluctuations.
The Faraday rotation towards this region is dominated by high density and strongly magnetized ISM cells distributed along the Galactic disc.
\item Toward the south-western corner of the field, there is a group of positive RM sources ($\phi_{\rm obs}>100\,\rm rad\,m^{-2}$) associated with the north-eastern edge of the Gum Nebula which is clearly visible in the $H\alpha$ emission map in Figure \ref{fig:hi+halpha}. \citet{Purcell_2015} constrained the electron density ($n_{\rm e} = 1.4\pm0.4\,\rm cm^{-3}$) and the magnetic field strength ($B=3.9^{+4.9}_{-2.2}\,\rm\mu G$) in this region from the observed RM distribution together with a simple geometric model of a magneto-ionized spherical shell.
\item There are positive RMs extending along $l\sim 265^{\circ}$ up to $b\sim 25^{\circ}$.
This pillar of positive RMs seems to extend from the RM distribution of the Gum Nebula in Region II, but also aligns well with the western edge of the Antlia SNR and LA III. Also, it was earlier reported by \citet{Reynoso_1997} that there is a vertical HI structure in this region that is possibly related with blown-out ISM from the Galactic disc to the halo (i.e., a galactic chimney).
Due to the complexity in the region, it is difficult to identify which object is a dominant Faraday rotator along the line-of-sight.
\item We also notice a group of negative RMs ($\phi_{\rm obs}<-100\,\rm rad\,m^{-2}$) near the western boundary of the field. Interestingly, there is no corresponding radio continuum nor $H\alpha$ emission detected nor previously reported magneto-ionized object in this region.
\item At the eastern edge of the Antlia SNR where it overlaps with the LA II, there is a group of RMs close to zero surrounded by negative RMs. This is the region studied by \citetalias{McClure-Griffiths_2010}. \citetalias{McClure-Griffiths_2010} proposed the RM excess in the region as indication of magnetized LA II, but the Antlia SNR in the foreground complicates the determination of the dominant Faraday rotator in the region.
\end{enumerate}
Our inspection of the noticeable structures in the RM grid indicates that it is not uncommon to find multiple gaseous objects along the lines-of-sights. Especially, the situation is complicated in Region III and Region V due to the coincidence of the Antlia SNR and the Magellanic Leading Arms.
Additional pieces of information is required in order to determine where along the lines-of-sight the Faraday rotation occurs in this region.
\subsection{Radio continuum and diffuse polarised emission}\label{sec:diffuse}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{zoom-in.pdf}
\caption{
The zoom-in image of a $|\nabla \mathcal{P}|/|\mathcal{P}|$ double-jump profile filament in Region III (the north-western edge of the Antlia SNR) and its $H\alpha$ counterpart.
}
\label{fig:zoom-in}
\end{figure}
Figure \ref{fig:stokes_i} shows the S-PASS Stokes $I$ image of the field\footnote{We used colour maps provided by {\sc CMASHER} python package (\citealt{cmasher}) in Figure \ref{fig:stokes_i}, \ref{fig:p}, and \ref{fig:p_grad}.}.
Our field of interest covers a wide range of synchrotron emitting regions including the bright Galactic disc (e.g., Region I) as well as the lower-brightness diffuse structures.
There is weak radio synchrotron emission in the Antlia SNR region. However, its continuum emission is not as prominent as in $H\alpha$ except for the south western edge of the SNR where it interacts with the Gum Nebula (Region II).
In comparison to the Stokes $I$ image, the linear polarisation intensity map (Figure \ref{fig:p}) shows rich filamentary structures on top of the smooth polarised emission.
These depolarised filaments appear if there is a sharp change in electron and/or magnetic properties in the foreground.
\iffalse
The properties of linearly polarised radiation are often characterised with the complex Stokes vector $\mathcal{P}$, as defined in equation \ref{eq:Stokes_P}. However, observed Stokes $Q$ and $U$, and therefore, $\mathcal{P}$, can easily be affected by any offset and rotation in the $Q-U$ plane that depends on telescopes used for observations, e.g., interferometer or single-dish, as well as polarisation calibration during the imaging process.
As an alternative, \citet{Gaensler_2011} introduced a simple parameter that characterises the polarised emission that is invariant to the transformations in the $Q-U$ plane: ``the spatial gradient of $\mathcal{P}$'' defined as follows.
\fi
The properties of linearly polarised radiation are often characterised with the complex Stokes vector $\mathcal{P}$, as defined in equation \ref{eq:Stokes_P} (\citealt{Gaensler_2011}).
Assuming the spatial gradient in $\mathcal{P}$, i.e.,
\begin{equation}
|\nabla \mathcal{P}| = \sqrt{\left( \frac{\partial Q}{\partial x}\right)^{2} + \left( \frac{\partial U}{\partial x}\right)^{2} + \left( \frac{\partial Q}{\partial y}\right)^{2} + \left( \frac{\partial U}{\partial y}\right)^{2}},
\end{equation}
arises from fluctuations in the Faraday rotation in foreground Faraday screens, we adopt the normalized parameter, $|\nabla \mathcal{P}|/|\mathcal{P}|$ (Figure \ref{fig:p_grad}). Taking the spatial gradient makes it easier to trace edges of the filaments standing out in Figure \ref{fig:p}. Bright features in the $|\nabla \mathcal{P}|/|\mathcal{P}|$ map therefore highlight depolarisation arising from complex layers of Faraday screens in this region of the sky.
With the normalized polarisation gradient map in Figure \ref{fig:p_grad}, we revisit the regions discussed earlier in Section \ref{Sec:RM} where we identify noticeable distinction in the distributions of RMs.
\begin{enumerate}[start=1,label={Region \Roman*},wide = 0pt, leftmargin = 4em]
\item (bottom left): This area is covered with chaotic small angular scale depolarised filaments. These fuzzy structures are typical at low galactic latitude (e.g., \citealt{Uyaniker_2003}; \citealt{Iacobelli_2014}) where significant depolarisation is expected from the turbulent ISM in the Galactic disc.
\item (bottom right): The bright web of depolarised filaments spread along the north-eastern edge of the Gum Nebula indicates that the object effectively operates as a Faraday rotating screen.
This is not surprising given its (i) high emissivity in $H\alpha$ (see Figure \ref{fig:hi+halpha}) suggesting the presence of a substantial amount of free electrons and (ii) the morphological coherence in the RM grid indicating that the nebula is magnetized.
\item (right edge of the dotted circle): There are depolarised filaments and loops extended along the western edge of the Antlia SNR. The remarkable spatial coherence of $H\alpha$ emission and the $|\nabla \mathcal{P}|/|\mathcal{P}|$ distribution in this region was earlier reported by \citet{Iacobelli_2014}.
Furthermore, we discovered a narrow filament with ``double-jump'' profile at $(l,b) \approx (263, 27)$ which indicates a delta function-like distribution of $n_{\rm e}$ and/or magnetic field (e.g., a strong shock; \citealt{Burkhart_2012}). Indeed, \citet{Fesen_2021} studied the spectral line ratios of the very filament and supported the shock origin of the filament. We present a zoom-in image of this filament in Figure \ref{fig:zoom-in}.
\item (middle right): The striking feature standing out in this region is a narrow loop.
Its single-jump profile suggests a step-function-like change in magnetic field properties in and outside the loop (\citealt{Burkhart_2012}).
The negative RMs found in Figure \ref{fig:hi+rm} are well enclosed within the loop, indicating the enhanced magnetic field strength in this region.
\item (left edge of the dotted circle):
Similarly to Region III, there is a complex network of filaments in this region including a narrow filament at $(l,b) \approx (284,29)$ that clearly overlap with a narrow $H\alpha$-emitting filament.
\end{enumerate}
\section{Discussion}\label{sec:discussion}
\subsection{Is the Antlia SNR a Faraday rotator?}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{Illustration.pdf}
\caption{Illustration of the Antlia SNR/Magellanic Leading Arm field. The polarised radiation from the extra galactic point source propagate through both the Magellanic Leading Arm and the Antlia SNR, while the Galactic diffuse polarised emission knows only about the Antlia SNR. This image is for illustration purpose only. The size of and the distance to the objects does not correspond to the real size and the distance.
}
\label{fig:illust}
\end{figure*}
The observed RM of a distant source ($\phi_{\rm obs}$) is a superposition of the Faraday rotation occurring at every magneto-ionized medium between the source and the observer. This includes the Faraday rotation at the polarised emitting source ($\phi_{\rm intrinsic}$) and the Milky Way foreground ($\phi_{\rm MW}$):
\begin{equation}\label{eq:phi_obs}
\phi_{\rm obs} = \phi_{\rm intrinsic}+\sum_{i=1}^{N}\phi_{\rm obj,\, i}+\phi_{\rm MW},
\end{equation}
where $\phi_{\rm obj, i}$ represents the Faraday rotation taking place at different distance along the line-of-sight and $N$ is the number of such Faraday screens which is very likely unknown for any sight-lines.
RM catalogues from all-sky surveys revealed large Galactic-scale structures in the RM grids which indicates that $\phi_{\rm MW}$ is likely to dominate the observed RMs in most of the sky (\citealt{Taylor_2009}; \citealt{Schnitzeler_2019}).
In studies of objects with smaller angular scales like HVCs, the contribution of $\phi_{\rm MW}$ is often estimated using off-object RMs of the region and subtracted from $\phi_{\rm obs}$.
The variation of the intrinsic polarisation of sources ($\phi_{\rm intrinsic}$) is random and therefore negligible on the basis of the large number statistics.
In the region of the sky studied in this paper, there are several known localized objects that could possibly induce Faraday rotation ($\phi_{\rm obj, i}$), if magnetized: the Gum Nebula, the Antlia SNR, and the Magellanic Leading Arm.
The Gum Nebula leaves an imprint on the RM grid that closely follows the morphology of its $H\alpha$ emission (see Region II in Figure \ref{fig:hi+rm}), making it clear that it is the dominating Faraday rotator in the region.
However, the overlap of the Antlia SNR and the Magellanic Leading Arm on the sky makes it indeterminate whether the features in the RM grid arise due to either of or both the objects.
They are not physically associated given their distinct observed velocities and distances.
Therefore, if the Antlia SNR is magnetized and significantly affects the RM towards the sightlines, the RM signiture identified by \citetalias{McClure-Griffiths_2010} can no longer be clearly associated with the Leading Arm.
To test the possibility of whether the intriguing features appearing in the compact-source RM map are associated with the Magellanic Leading Arm in the Galactic halo or the Antlia SNR in the foreground, we bring extra information from the diffuse polarised radiation emitted from the large-scale Galactic interstellar medium.
The diffuse polarised emission traces large-scale smooth polarised emission from Galactic ISM.
On top of that, Faraday screens that alter the polarisation properties of radiation coming from behind produce depolarised structures.
When polarised radiation emitted from different patches of ISM are combined into a beam, their properties are inevitably averaged out due to the turbulent nature of ISM (i.e. large fluctuations in properties) and finite beam sizes in radio observations.
This depolarisation effect result in iconic filamentary structures in a polarisation emission, e.g., depolarised canals (e.g., \citealt{Haverkorn_2004}; \citealt{Fletcher_2006}).
Figure \ref{fig:illust} illustrates the relative location of the objects and roughly where the polarised radiation of point-sources and the Galactic diffuse emission come from with respect to the objects.
Unlike extragalactic radiation that propagate through both the Magellanic Leading Arm and the Antlia SNR, the Galactic diffuse polarisation does not experience Faraday rotation (if there is any) at the Magellanic Leading Arm since the object is beyond the Galactic disc where most of the emission is coming from.
Therefore, the morphological correspondence between the low-velocity $H\alpha$ filaments (Figure \ref{fig:hi+halpha}) and the depolarised canals (Figure \ref{fig:p} and Figure \ref{fig:p_grad}) is the smoking gun evidence that the Antlia SNR is a Faraday rotator and severely affects the observed RM towards the region where the excess of RM was reported by \citetalias{McClure-Griffiths_2010}.
Our findings lead to a conclusion that it is not feasible to draw any certain conclusions about the magnetic fields of the Magellanic Leading Arm using the RM grid technique. In other words, \emph{it is hard to interpret the RM excess in this region as evidence of the ``magnetized'' Leading Arm}
We attempt to perform a similar test on the Smith cloud, which is another candidate of magnetized HVCs, but it was impossible since the cloud is located at the celestial equator which is right at the border of the S-PASS sky coverage. To our knowledge, there is no comparable polarimetric survey at the same frequency that covers the northern sky.
Yet, the low-velocity H$\alpha$ emission in the region does not show any structures above $5\,\rm Rayleigh$ unlike the case of the Magellanic Leading Arm. We conclude that the Smith Cloud region has a less chance of being affected by the Galactic foreground.
\subsection{Magnetic field strength of the Antlia SNR}\label{sec:b_los}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{b_los.pdf}
\caption{
Panel (a): A Voronoi diagram coloured with $\phi_{\rm obs}$. The RM excess is enclosed within a dashed circle.
Panel (b): the Milky Way corrected RM ($\phi_{\rm obs}-\phi_{\rm MW}$) in the RM excess region and the surrounding.
Panel (c): the distribution of $B_{\parallel}$ estimated from sources in the RM excess region
}
\label{fig:b_los}
\end{figure}
In this Section, we calculate the line-of-sight magnetic field strength at the eastern edge of the Antlia SNR (i.e., Region V of Figure \ref{fig:hi+rm}, \ref{fig:stokes_i}, \ref{fig:p}, and \ref{fig:p_grad}) assuming the RM excess is entirely due to the magneto-ionized shell of the Antlia SNR.
This is to check if the estimated field strength is in a reasonable range expected from typical SNRs, in other words, whether the observed RM excess can be explained solely by Faraday rotation at the Antlia SNR.
From the definition of the RM (equation \ref{eq:RM}), the line-of-sight magnetic field strength can be expressed as follows:
\begin{equation}\label{eq:B}
B_{\parallel} = \frac{\phi_{\rm SNR}}{0.812 \left<n_{\rm e}\right> f L},
\end{equation}
where $\phi_{\rm SNR}$ is RM of the SNR, $f$ is a volume filling factor of the ionized gas set to 0.5 and $L$ is a path length .
Note that, here we work with the product of an average electron density and the path length through the SNR, $\left<n_{\rm e}\right> L$, since the distribution of the electron density along the sight-line, $n_{\rm e}(r)$, is not known.
Therefore, $B_{\parallel}$ is by assumption the electron-density weighted average magnetic field strength along the line-of-sight.
Under the assumption that the SNR is a dominant source of Faraday rotation other than the smoothly varying large-scale Milky Way ISM, equation \ref{eq:phi_obs} can be expressed as
\begin{equation}
\phi_{\rm SNR}=\phi_{\rm obs}-\phi_{\rm MW}
\end{equation}
in the region where the excess of RM is.
The Voronoi diagram in panel (a) of Figure \ref{fig:b_los} shows the overall distribution of $\phi_{\rm obs}$ near the eastern edge of the Antlia SNR.
The RM excess region is enclosed with a dashed circle.
On the other hand, in the surrounding domain where the RM distribution does not show any correspondence with the SNR, we can consider that the Milky Way is the only Faraday screen (i.e., $\phi_{\rm obs}\approx\phi_{\rm MW}$).
We estimate $\phi_{\rm MW}$ of the RM excess region based on $\phi_{\rm obs}$ measurements in the surrounding.
This approximation holds only at a immediate vicinity where the variance in the large-scale Milky Way field is small and there is no other apparent Faraday screens (e.g., the Gum nebula). Therefore, we restrict our sample to sources within a region bounded by the solid line in Panel (a).
The median $\phi_{\rm obs}$ in this region is adopted as $\phi_{\rm MW}=-56.55\,\rm rad\,m^{-2}$.
Panel (b) shows the distribution of $\phi_{\rm obs}-\phi_{\rm MW}$.
The corrected RMs are mostly positive within the RM excess region as previously pointed out by \citetalias{McClure-Griffiths_2010} and nearly zero in the surrounding region.
The path length through the SNR ($L$) is estimated from a simple geometric model of a 3D spherical shell.
We define $L$ as a function of the angular separation ($\theta$) from the center of the SNR at $(l, b) = (276.5^{\circ}, +19^{\circ})$:
\begin{equation}
L(\theta) =
\begin{cases}
2D\sqrt{\sin^{2}\theta_{\rm shell}-\sin^{2}{\theta}} &\text{if $\theta_{\rm shell} - d\theta<\theta<\theta_{\rm shell}$}\\
0 &\text{if $\theta>\theta_{\rm shell}$}\\
2 D d\theta &\text{if $\theta<\theta_{\rm shell} - d\theta$}
\end{cases},
\end{equation}
where $D=100\,\rm pc$ is the distance to the Antlia SNR, $\theta_{\rm shell}=18^{\circ}$ and $d\theta=8^{\circ}$ are the outer radius and the thickness of the shell, respectively, in angular scale.
The average electron density along the sight-lines ($\left<n_{\rm e}\right>$) is estimated from the emission measure ($EM$) of the SNR from its $H\alpha$ intensity:
\begin{equation}
EM = 2.75\left(\frac{T_{\rm e}}{10^{4}\rm K}\right)^{0.9} I_{H\alpha},
\end{equation}
where $T_{\rm e}=10^{4}\,\rm K$ is the electron temperature and $I_{H\alpha}$ is the $H\alpha$ emission in Rayleighs. From the definition of $EM$,
\begin{equation}
\left<n_{\rm e}\right> = \sqrt{\frac{EM}{f L}}.
\end{equation}
All parameters combined, equation \ref{eq:B} can be expressed as
\begin{equation}
B_{\parallel} = \frac{\phi_{\rm obs}-\phi_{\rm MW}}{0.673\sqrt{I_{H\alpha}} L }.
\end{equation}
The histogram in Panel (c) shows the distribution of $B_{\parallel}$ calculated using the individual polarised sources in the RM excess region.
The median of the distribution is at $B_{\parallel}\approx 5\,\rm\mu G$, which is similar to a typical magnetic field strength in the Galactic ISM.
\section{Summary}\label{sec:summary}
We hereby summarise three mutually related major points discussed throughout this paper.
First, we argue that the Faraday rotation towards the Magellanic Leading Arm is highly affected by the foreground SNR.
\citetalias{McClure-Griffiths_2010} identified the Magellanic Leading Arm as a magnetized HVC. However, we find that the Antlia SNR can equally well explain the observed RM excess.
Second, our work provides information about the structures and the magnitude of magnetic fields associated with the Antlia SNR.
The remnant is a Faraday rotator that severely depolarises the diffuse Galactic polarised emission in the background.
We also found double-jump profile filaments in the normalized polarisation gradient map which indicate a sharp enhancement along the shock regions that are bright in $H\alpha$ emission and studied by \citet{Fesen_2021}.
From the compact source RM grid, we estimated the line-of-sight magnetic field strength at the Eastern edge of the Antlia SNR to be $B_{\parallel}\approx5\,\rm\mu G$.
Finally, the lesson we learned from our study in the Antlia SNR/ Magellanic Leading Arm field raise caution for future studies using the RM grid technique.
Up-coming radio telescopes and polarisation all-sky surveys are expected to significantly increase the RM source density. This will allow us to study magnetism in much detail and even extend the RM grid technique to extra galactic objects (e.g. \citealt{Anderson_2021}).
However, such studies should always be aware that there are local Faraday rotators that can significantly affect the RM grid, like the Antlia SNR in our case. We suggest checking Galactic diffuse polarisation maps to identify local Faraday screens.
\section*{Acknowledgements}
We thank the anonymous referee for the positive and constructive review of the work.
The Australia Telescope Compact Array is part of the Australia Telescope National Facility which is funded by the Australian Government for operation as a National Facility managed by CSIRO. We acknowledge the Gomeroi people as the traditional owners of the Observatory site.
We also acknowledge the Ngunnawal and Ngambri people as the traditional owners and ongoing custodians of the land on which the Research School of Astronomy \& Astrophysics is sited at Mt Stromlo.
N.M.G. acknowledges the support of the ARC through Future Fellowship FT150100024.
A.S.H. is partly supported by a Discovery Grant from the Natural Sciences and Engineering Research Council of Canada.
Our analysis was performed using the Python programming language (Python Software Foundation, https://www.python.org). The following packages were used throughout the analysis: numpy (\citealt{Harris_2020}), SciPy (\citealt{Virtanen_2020}), and matplotlib (\citealt{Hunter_2007}). This research additionally made use of the publicly available tools: MIRIAD (\citealt{Sault_1995}), RM tools 1D (\citealt{Purcell_2020}), and {\sc CMASHER} python package (\citealt{cmasher}).
\section*{Data availability}
The data underlying this article were accessed from the CSIRO Australia Telescope National Facility online archive at https://atoa.atnf.csiro.au, under the project codes C2741. The derived data generated in this research are available in the article and in its online supplementary material.
\bibliographystyle{mnras}
|
\section{Introduction}
The Lennard-Jones (LJ) system is one of the most popular and extensively studied model systems in condensed matter physics. Many results on its transport properties have been published over the years. For recent reviews of simulation data available in the literature see e.g. Refs.~\cite{BellJPCB2019,HarrisJCP2020,AllersJCP2020}.
Numerical data can be particularly useful in verifying and testing various models and approaches to the transport properties of liquids. The latter research topic remains important and timely despite considerable progress achieved over many decades~\cite{HansenBook,GrootBook,MarchBook}.
Among the existing semi-quantitative models and scaling relationships perhaps the most familiar are the Stokes-Einstein relation between the self-diffusion and shear viscosity coefficients~\cite{FrenkelBook,ZwanzigJCP1983,Balucani1990,CostigliolaJCP2019,KhrapakMolPhys2019}, the excess entropy scaling of transport coefficients~\cite{RosenfeldPRA1977,DzugutovNature1996,
RosenfeldJPCM1999,DyreJCP2018}, density scaling~\cite{FragiadakisJCP2019,HarrisJCP2020} as well as the freezing temperature scaling~\cite{RosenfeldPRE2000,CostigliolaJCP2018,KhrapakAIPAdv2018}.
Some of the scaling relationships can be rationalized within the framework of the isomorph theory~\cite{DyreJPCB2014,VeldhorstPoP2015,DyreJCP2018}. This theory predicts that many liquids exhibit an approximate
``hidden'' scale invariance that implies the existence of lines in the thermodynamic phase diagram, so-called isomorphs, along which structure and dynamics in properly reduced units
are invariant to a good approximation. Excess entropy is constant along isomorphs and this can serve as an explanation of excess energy scaling~\cite{DyreJCP2018}. Additionally, the melting and freezing lines are approximate isomorphs~\cite{PedersenNatCommun2016,SaijaJCP2001} and therefore the lines that are parallel to them in some vicinity should be also isomorphs. This explains the observation that the transport coefficients of some simple fluids behave quasi-universally when plotted versus temperature scaled by the freezing temperature. But this also suggests that the transport coefficients should coincide when plotted versus density scaled by the freezing density. The purpose of this paper is to prove this conjecture.
For some systems the freezing density scaling is trivial. A special case is the hard-sphere (HS) system. Since all the structural and dynamical properties depend on the packing fraction (reduced density) alone, the freezing density scaling is obvious.
Another relevant example is related to strongly coupled Yukawa fluids, which are often used to model classical systems of charged particles immersed in a neutralizing medium, such as colloidal suspensions and complex plasmas~\cite{FortovUFN,FortovPR,IvlevBook}. Transport coefficients of strongly coupled Yukawa fluids turn out to be some quasi-universal functions of the reduced coupling parameter $\Gamma/\Gamma_{\rm fr}$
~\cite{RosenfeldPRE2000,OhtaPoP2000,SaigoPoP2002,VaulinaPRE2002,VaulinaPoP2002,
FortovPRL2003,FaussurierPRE2003,DonkoPRE2004,VaulinaPRE2010,
KhrapakPoP2012,KhrapakJPCO2018}. Here the coupling parameter $\Gamma=Q^2\rho^{1/3}/T$ is the ratio of the Coulomb interaction energy at a mean interparticle separation ($Q$ being the particle charge and $\rho$ particle density) to the kinetic energy, characterized by the system temperature $T$ (measured in energy units) and $\Gamma_{\rm fr}$ is the value of $\Gamma$ at the freezing point (the coupling parameter at the melting point, $\Gamma_{\rm m}$, can be used as well because the fluid-solid coexistence region is very narrow~\cite{HynninenPRE2003} and one normally does not discriminate between freezing and melting points in Yukawa systems). The success of this scaling for Yukawa systems was put in the context of the isomorph theory~\cite{VeldhorstPoP2015}. For the present discussion it is important that a universal scaling with $\Gamma/\Gamma_{\rm fr}$ implies not only scaling with $T/T_{\rm fr}$, but also a scaling with $\rho/\rho_{\rm fr}$, so that the freezing density scaling emerges trivially.
In this paper we demonstrate that the freezing density scaling is more general and is not limited to trivial examples given above. In particular, it works very well for sufficiently dense LJ fluids in a wide region of the LJ phase diagram. Simulation data on the transport coefficient of LJ fluids along isotherms fully comply with this scaling. Moreover, we demonstrate that the observed scaling is closely related to that of the hard-sphere fluid. It is shown that the Stokes-Einstein relation between the diffusion and viscosity coefficients is valid in the dense liquid regime, and that the boundary of this regime can be used to locate the crossover between gas-like and liquid-like regions on the phase diagram.
\section{Formulation}
The LJ potential, which is often used to approximate interactions in liquefied noble gases, reads
\begin{equation}
\phi(r)=4\epsilon\left[\left(\frac{\sigma}{r}\right)^{12}-\left(\frac{\sigma}{r}\right)^{6}\right],
\end{equation}
where $\epsilon$ and $\sigma$ are the energy and length scales (or LJ units), respectively. The density and temperature in LJ units are $\rho_*=\rho\sigma^3$ and $T_*=T/\epsilon$. The LJ potential exhibits repulsion at short interatomic distances and attraction at long distances. As a result, the phase diagram of a LJ system contains a liquid-vapor phase transition with liquid-vapor coexistence region and a liquid-vapor critical point, in addition to the fluid-solid phase transition (systems of purely repulsive particles, such as for instance hard spheres, do not show a liquid phase).
To demonstrate the validity and usefulness of the freezing density scaling we consider two sets of extensive numerical simulation data for the LJ transport coefficients. The viscosity and self-diffusion data have been published by Meier {\it et al.}~\cite{Meier2002,MeierJCP_1,MeierJCP_2}. The second data set for self-diffusion, viscosity, and thermal conductivity coefficients is due to Baidakov {\it et al.}~\cite{BaidakovFPE2011,BaidakovJCP2012,BaidakovJCP2014}.
The data have been obtained in a wide region of the LJ phase diagram. Though simulations protocols were different, the two datasets are in good agreement where they overlap~\cite{HarrisJCP2020}.
To be concrete, the viscosity and self-diffusion coefficients at $T_*=1, 1.5$ and 2 are taken from Refs.~\cite{BaidakovFPE2011,BaidakovJCP2012} and at $T_*=3$ and 4 from Ref.~\cite{Meier2002}. The thermal conductivity coefficients at $T_*=1, 1.5$ and 2 are taken from Ref.~\cite{BaidakovJCP2014}. We do not consider data points from the fluid-solid coexistence region.
We reduce the transport coefficients (originally expressed in LJ units) using the so-called system-independent Rosenfeld normalization~\cite{RosenfeldJPCM1999}:
\begin{equation}\label{Rosenfeld}
D_{\rm R} = D\frac{\rho^{1/3}}{v_{\rm T}} , \quad
\eta_{\rm R} = \eta \frac{\rho^{-2/3}}{m v_{\rm T}}, \quad \lambda_{\rm R} = \lambda \frac{\rho^{-2/3}}{v_{\rm T}},
\end{equation}
where $\rho^{-1/3}=\Delta$ defines the characteristic interatomic separation, $v_{\rm T}=\sqrt{T/m}$ is the thermal velocity, $m$ is the atomic mass, and the temperature is expressed in energy units ($k_{\rm B}=1$).
In the first approximation, the transport coefficients of a dilute gas consisting of hard spheres of diameter $\sigma$ (HS gas) are~\cite{LifshitzKinetics}
\begin{eqnarray}\label{HSD}
D &=& \frac{3}{8\rho\sigma^2}\left(\frac{T}{\pi m}\right)^{1/2} \Rightarrow D_{\rm R} = \frac{3}{8\sqrt{\pi}\rho_*^{2/3}}\simeq\frac{0.212}{\rho_*^{2/3}} , \\ \label{HSeta}
\eta &=& \frac{5}{16\sigma^2}\left(\frac{mT}{\pi}\right)^{1/2} \Rightarrow \eta_{\rm R} = \frac{5}{16\sqrt{\pi}\rho_*^{2/3}}\simeq\frac{0.176}{\rho_*^{2/3}}, \\ \label{HSlambda}
\lambda &=& \frac{75}{64\sigma^2}\left(\frac{T}{\pi m}\right)^{1/2} \Rightarrow \lambda_{\rm R} = \frac{75}{64\sqrt{\pi}\rho_*^{2/3}}\simeq\frac{0.661}{\rho_*^{2/3}}.
\end{eqnarray}
Remarkably, when expressed in Rosenfeld's units, the transport coefficients of dilute hard-sphere gas exhibit universal density scaling: All transport coefficients decay as $\propto \rho_*^{-2/3}$. This density scaling is of course only approximate for dilute LJ gases. The actual transport cross sections are different from the hard-sphere model and specifics of scattering in the LJ potential has to be properly accounted for (see e.g. Refs.~\cite{HirschfelderBook,HirschfelderJCP1948,SmithJCP1964,
KhrapakPRE2014_Scattering,
KhrapakEPJD2014,Kristiansen2020} and references therein for some related works).
Contrary to the situation with dilute gases no accurate general theory of transport process in liquids still exists, and that is why universal relationships and scalings are of considerable interest. The freezing density scaling along isotherms is illustrated in the next Section.
\section{Results and discussion}
Our main results are plotted in Figs.~\ref{Fig1} - \ref{Fig3}, where we plot the reduced transport coefficients versus the density normalized to its value at the freezing point, taken from Ref.~\cite{SousaJCP2012}. They demonstrate that the self-diffusion and viscosity data (Figs.~\ref{Fig1} and \ref{Fig2}) coincide excellently when plotted versus density reduced by its value at the freezing point. The coincidence is not so impressive for the thermal conductivity data shown in Fig.~\ref{Fig3}). It should be noted that the thermal conductivity coefficient is proportional to the specific heat $c_v$, which exhibits a critical enhancement near the critical point. Thus the behaviour of thermal conductivity on near-critical isotherms can be expected to be somewhat different from that on supercritical isotherms~\cite{KhrapakPRE2021}. Additionally, some scattering may reflect the accuracy of numerical data.
\begin{figure}
\includegraphics[width=8.5cm]{Figure1.pdf}
\caption{(Color online) Reduced self-diffusion coefficient of a LJ fluid $D_{\rm R}$ vs the reduced density $\rho_*/\rho_*^{\rm fr}$. The symbols correspond to numerical data tabulated in Refs.~\cite{Meier2002,BaidakovFPE2011}. The dashed and solid curves denote the HS gas and fluid asymptotes. The latter is based on the simulation data from Ref.~\cite{Pieprzyk2019}.}
\label{Fig1}
\end{figure}
\begin{figure}
\includegraphics[width=8.5cm]{Figure2.pdf}
\caption{(Color online) Reduced shear viscosity coefficient of a LJ fluid $\eta_{\rm R}$ vs the reduced density $\rho_*/\rho_*^{\rm fr}$. The symbols correspond to numerical data tabulated in Refs.~\cite{Meier2002,BaidakovJCP2012}. The dashed and solid curves denote the HS gas and fluid asymptotes. The latter is based on the simulation data from Ref.~\cite{Pieprzyk2019}.}
\label{Fig2}
\end{figure}
The quasi-universality of transport coefficients extend to very low density, down to $\rho_*/\rho_*^{\rm fr}\simeq 0.1$. For lower densities the freezing density scaling is not appropriate. Here the dilute gas regime is realised and the transport coefficients are determined by the corresponding transport cross sections~\cite{HirschfelderBook,KhrapakEPJD2014}.
\begin{figure}
\includegraphics[width=8.5cm]{Figure3.pdf}
\caption{(Color online) Reduced thermal conductivity coefficient of a LJ fluid $\lambda_{\rm R}$ vs the reduced density $\rho_*/\rho_*^{\rm fr}$. The symbols correspond to numerical data tabulated in Ref.~\cite{BaidakovJCP2014}. The dashed and solid curves denote the HS gas and fluid asymptotes. The latter is based on the simulation data from Ref.~\cite{Pieprzyk2020}.}
\label{Fig3}
\end{figure}
Two additional curves are plotted in Figs.~\ref{Fig1} - \ref{Fig3}. The dashed curves correspond to Eqs.~(\ref{HSD}) - (\ref{HSlambda}). As pointed out above, they do not represent accurate approximations for the LJ gas, but the main trends in the density dependence are very well reproduced. To plot these curves we assumed $\rho_{*}^{\rm fr}=1$, which is appropriate for the range of $T_*$ investigated due to relatively weak dependence of $\rho_*^{\rm fr}$ on $T_*$ (actually $\rho_{*}^{\rm fr}\simeq 1$ at $T_*\simeq 1.5$~\cite{SousaJCP2012}).
The solid curves correspond to the freezing density scaling of transport coefficients of a HS fluid. To plot these curves we have used extensive molecular simulation results by Pieprzyk {\it et al}.~\cite{Pieprzyk2019,Pieprzyk2020}. The use of large simulation systems and long simulation times allowed accurate prediction of transport coefficients in the thermodynamic limit. It is observed in Fig.~\ref{Fig1} that the self-diffusion coefficient of LJ fluids is in excellent agreement with HS asymptote down to the lowest density studied in Ref.~\cite{Pieprzyk2019}. For the shear viscosity coefficient the HS asymptote somewhat overestimates the LJ data, but multiplying the HS asymptote by a constant factor of $0.9$ brings LJ and HS data to a very good agreement (see Fig.~\ref{Fig2}). For the thermal conductivity coefficients, the agreement between the LJ and HS data is good in the moderate density regime, $\rho_*/\rho_{*}^{\rm fr}\lesssim 0.6$, but for higher relative density the thermal conductivity coefficient of the HS fluid is consistently larger and overestimates that of the LJ fluid by about $40\%$ at the freezing point.
It is not at all surprising that the HS model can reproduce the transport properties of softer particle systems. It is well recognized that properly reduced dynamical properties of fluids with strong repulsive forces can be mapped into those of HS fluids~\cite{Dymond1985,LopezFloresPRE2013,DyreJPCM2016}. The novelty of our approach is that the freezing density scaling seems to provide a simple and convenient mapping method, in particular for the self-diffusion coefficient.
The fact that the freezing density scaling of LJ fluid is relatively close to that of a HS fluid is also not very surprising. The effective repulsive (inverse power law) exponent of a LJ potential is quite high, $n_{\mathrm{eff}}\simeq 18$ (at moderate densities)~\cite{PedersenPRL2008}. This is because the attractive ($\propto r^{-6}$) term of the LJ potential makes its repulsive short-range branch considerably steeper than just the $\propto r^{-12}$ repulsive term. Only at high densities and temperatures $n_{\mathrm{eff}}$ approaches 12~\cite{BohlingJCP2014}.
In this regime the agreement between the LJ and HS scaling may somewhat decline.
\begin{figure}
\includegraphics[width=8.5cm]{Figure4.pdf}
\caption{(Color online) Stokes-Einstein relation without the hydrodynamic diameter in a LJ fluid. The parameter $\alpha_{\rm HS}$ is plotted as a function of the reduced density $\rho_*/\rho_*^{\rm fr}$. The symbols correspond to numerical data tabulated in Ref.~\cite{Meier2002,BaidakovFPE2011,BaidakovJCP2012}. The dashed and solid curves denote the HS gas and fluid asymptotes. The latter is based on the simulation data from Ref.~\cite{Pieprzyk2019}. The horizontal line at $\alpha_{\rm HS}= 0.15$, correspond to the simulation results from Ref.~\cite{CostigliolaJCP2019}.}
\label{Fig4}
\end{figure}
Let us briefly discuss the Stokes-Einstein (SE) relation between the self-diffusion and viscosity coefficient. According to Zwanzig's model~\cite{ZwanzigJCP1983} for simple atomic fluids this relation can be formulated without introducing a so-called ``hydrodynamic diameter'' (in fact, the interparticle separation plays the role of the effective hydrodynamic diameter):
\begin{equation}\label{SE}
\frac{D\eta}{\rho^{1/3}T}=\alpha_{\rm SE}.
\end{equation}
In the Rosenfeld's normalization Eq.~(\ref{SE}) reduces to $D_{\rm R}\eta_{\rm R}=\alpha_{\rm SE}$.
The parameter $\alpha_{\rm SE}$ can be related to the ratio of the transverse-to-longitudinal sound velocities~\cite{ZwanzigJCP1983,KhrapakMolPhys2019}:
\begin{displaymath}
\alpha_{\rm SE}\simeq 0.13\left(1+\frac{c_t^2}{2c_l^2}\right),
\end{displaymath}
where $c_t$ and $c_l$ are the transverse and longitudinal sound velocities. For LJ fluids near the freezing point we have $c_l\simeq 2 c_t$~\cite{KhrapakMolecules2020} and this should be relevant for a sufficiently dense regime not too far from the fluid-solid phase transition. Thus, we can expect $\alpha_{\rm SE}\simeq 0.15$ on approaching freezing. This is perfectly consistent with numerical data on the LJ fluids, see Fig.~\ref{Fig4}. The HS asymptote lies close, but slightly above the LJ data, because the HS viscosity is about $10\%$ higher than that of the LJ fluid. In the dilute regime we can expect an approach to the dilute HS gas asymptote, which is $D_{\rm R}\eta_{\rm R}\simeq 0.037/\rho_{*}^{4/3}$. Figure~\ref{Fig4} demonstrates that the data point lie quite close to this asymptote (as previously we used $\rho_*^{\rm fr}\simeq 1$ to plot this curve).
The final comment is related to the crossover between the gas-like and liquid-like regions on the phase diagram, a topic of considerable recent interest~\cite{Simeoni2010,BrazhkinJPCB2011,BrazhkinPRE2012,
BrazhkinUFN2012,BellJCP2020}. Stokes-Einstein relation provides us with a good pragmatic tool to estimate the location of this crossover. Figure~\ref{Fig4} demonstrates that numerical data follow very closely the dilute gas and the fluid asymptotes. The later intersect at $\rho_*/\rho_*^{\rm fr}\simeq 0.35$ and this can be chosen as the demarcation condition between the gas-like and liquid-like regions. Note that this is close to the location of the minimum of the reduced shear viscosity coefficient (Fig.~\ref{Fig2}). It would be interesting to verify to which extent this condition is applicable to other soft interacting systems.
\section{Conclusion}
To summarize, each of the properly reduced transport coefficients of self-diffusion, shear viscosity, and thermal conductivity of dense LJ fluids coincides along different isotherms when plotted as a function of density divided by its value at the freezing point. The quality of this coincidence is particularly high for the self-diffusion and viscosity coefficients and is somewhat lower for the thermal conductivity coefficient. The freezing density scaling is close to that of a HS fluid, in particular for the self-diffusion coefficient. This provides a simple useful procedure of mapping between soft (but strongly repulsive) and HS fluids. The Stokes-Einstein relation of the form $D_{\rm R}\eta_{\rm R}\simeq 0.15$ is satisfied to a very high accuracy in a dense fluid regime with $\rho_*/\rho_*^{\rm fr}\gtrsim 0.4$. At low densities the dilute HS asymptote is appropriate $D_{\rm R}\eta_{\rm R}\propto \rho_{*}^{-4/3}$. The two asymptotes are intersecting at about $\rho_*/\rho_*^{\rm fr}\simeq 0.35$ and this can serve as a practical demarcation condition between the gas-like and liquid-like regions on the LJ system phase diagram.
|
\section{Introduction}
Artificial Neural Networks (NN) have shown potential in enabling autonomy for tasks that are hard to formally define or specify.
For instance, NN can be used to learn and replicate multi-robot behaviors \cite{RiviereEtAl2020ral,ZhouEtAl2019iros,FanLLP2020ijrr}, or to model complex aircraft collision avoidance protocols \cite{JulianEtAl2016dasc}.
Driving is another example of such a task where not all driving situations can be predicted in advance and, therefore, it may be beneficial to extrapolate behaviors from previously learned scenarios.
Thus, learning for driving \cite{PanEtAl2018rss,KuuttiEtAl2021its,StricklandFBA2018icra} is gaining ground as a possible solution to the automated driving or driver assistance challenge problems.
Nevertheless, NN-based autonomous systems may produce unsafe behaviors when they encounter inputs in regions where an insufficient number of training samples was available.
Multiple methods exist that attempt to discover such unsafe behaviors (also known as adversarial samples or counter-examples) for autonomous systems \cite{TianEtAl2018icse,Dreossi2019,TuncaliEtAl2020tiv}.
Other methods investigate classes of systems and NN for which provable properties can be established \cite{SunKS2019hscc,IvanovEtAl2020,DuttaEtAl2018adhs,TjengXT2019iclr}.
In either case, an open challenge is how to repair or modify NN-based systems when they are found to not be safe.
In the case of testing based methods (and some verification methods), adversarial samples are returned that can be added to the training set to improve the accuracy of the NN \cite{YaghoubiF2019tecs,DreossiJS2018arxiv,DreossiGYKSS2018ijcai}.
However, adding multiple adversarial samples to the training set might produce over-fitting, or even ``oscillations", i.e., adding training samples from one region may result in reduced accuracy in another region.
On the other hand, many verification methods may not even return any information when an NN-based system fails to satisfy its safety requirements.
Hence, repairs may not be as straightforward as when counterexamples are available.
In this paper, we make progress toward the repair problem for NN-based systems.
We pose the following question: given a set of safety constraints on the output of the NN component, is it possible to modify the NN such that the safety constraints are satisfied while the NN accuracy over the original training set remains high?
We propose an optimization based solution by posing the repair problem as a Mixed-Integer Quadratic Program (MIQP).
Our framework is presented on a number of challenging problems relevant to robotics: forward kinematics of arm manipulators, aircraft collision avoidance, and mobile robot motion control.
The repair problem \cite{goldberger2020minimal,DongEtAl2021arxiv,CruzFS2021arxiv} has objectives that typically extend beyond just identifying adversarial samples to be added to the training set \cite{YaghoubiF2019tecs,DreossiJS2018arxiv,DreossiGYKSS2018ijcai}.
In \cite{CruzFS2021arxiv}, the authors address the repair problem for NN-based controllers with a Two-Level Lattice (TLL) architecture.
The main step in the framework is a local repair to the controller which is only activated in the states in the vicinity of any adversarial samples.
A condition for applying the repair is that any discovered unsafe regions do not overlap with previously trained regions.
A limitation of the method in \cite{CruzFS2021arxiv} is that it is only applicable to the NN architecture.
In contrast, our approach is applicable to general feed forward NN and there are no conditions on the adversarial samples.
The repair process in \cite{DongEtAl2021arxiv} is executed in the loop with a verification engine.
Given an adversarial sample, the neighborhood around it is partitioned until a region is identified where for every point in that region the safety constraints are violated.
Then, the neurons that cause the violation are identified and they are repaired.
Even though the method applies to general feedforward NN, the repair formulation does not capture the effect of the neuron modifications to the performance of the NN overall.
Finally, the authors in \cite{goldberger2020minimal} propose a verification-based method for minimally modifying the network weights so as to satisfy a given property.
Due to the use of verification, they guarantee the correctness of the repaired model.
However, this work restricts the repair to the last layer and only guarantees the minimality of the weight changes.
Their method is also limited to guaranteed modification of only one adversarial input.
The following are the main contributions of this paper.
First, we propose a layer-wise neural network repair framework using Mixed-integer Quadratic Programming (MIQP) to satisfy constraints on the output of the network for a given input domain (Sec. \ref{sec: MinRepair-MIP}).
Second, while our method repairs the network for the specified local input domain, it maintains the prediction/classification accuracy of the network over its original trained space.
Finally, we demonstrate the application of our method in addressing safety, model improvement, and verification problems in a variety of robot learning domains (Sec. \ref{sec: sim-sec}).
\section{Problem Statement}
\subsection{Notation}
We denote the set of variables $a_1,a_2,\cdots,a_N$ with $\{a_n\}^N_{n=1}$.
Let $N\!N$ be a neural network with $L$ hidden layers, as shown in Fig. \ref{fig:MLP}.
The nodes values at each layer $l \in \{l_i\}^L_{l=0}$ are represented by a vector $\mathbf{x}^l=[x^l_0,x^l_1,\cdots,x^l_{\lvert\mathbf{x}^l\rvert}]^T$, where $\lvert\mathbf{x}^l\rvert$ denotes the dimension of layer $l$. The network output is also denoted as $\mathbf{y}$.
Each two subsequent layers are fully connected with weighted edges and bias terms.
The notation
\[ \mathbf{x}^{l-1}\xrightarrow{\mathbf{W}^l,\mathbf{b}^l} \mathbf{x}^l \]
represents the weight matrix $\mathbf{W}^l$ and the bias vector $\mathbf{b}^l$ of layer $l$ connecting the vectors $\mathbf{x}^{l-1}$ to $\mathbf{x}^l$ with
$$
\mathbf{W}^l = \begin{bmatrix}
w^l_{00} & w^l_{10} &\cdots & w^l_{\lvert\mathbf{x}^{l-1}\rvert0}\\
w^l_{01} & w^l_{11} &\cdots & w^l_{\lvert\mathbf{x}^{l-1}\rvert1}\\
\vdots & \vdots & \ddots &\vdots\\
w^l_{0\lvert\mathbf{x}^l\rvert} & w^l_{1\lvert\mathbf{x}^l\rvert} &\cdots & w^l_{\lvert\mathbf{x}^{l-1}\rvert\lvert\mathbf{x}^l\rvert}
\end{bmatrix},
\mathbf{b}^l = \begin{bmatrix}
b^l_0\\
b^l_1\\
\vdots\\
b^l_{\lvert\mathbf{x}^l\rvert}
\end{bmatrix}.
$$
The training data set of $N$ inputs $\mathbf{x}^0_n$ and target outputs $\mathbf{t}_n$ is denoted by $\{(\mathbf{x}^0_n,\mathbf{t}_n)\}^N_{n=1} \subseteq \mathcal{X}_o\times\mathcal{T}\subseteq\mathbb{R}^{\lvert\mathbf{x}^0\rvert}\times\mathbb{R}^{\lvert\mathbf{t}\rvert}$.
We use $x^l_{j,n}$ to denote the $j$th node value of layer $l$ for sample $n$ and $\mathbf{x}^l_n$ to denote the vector of nodes at layer $l$ for sample $n$.
The value of each hidden node is calculated using the weighted sum of the nodes in its previous layer passed through a nonlinear activation function.
In this work, we solve the repair problem for networks that use the Rectified Linear Unit (ReLU) activation function $R(z) = \max\{0,z\}$.
Thus, given the $n$th sample, in the hidden layer $l:\mathbf{x}^{l-1}\xrightarrow{\mathbf{W}^l,\mathbf{b}^l}
\mathbf{x}^l$, we have $\mathbf{x}^l = R\left(\mathbf{W}^l\mathbf{x}^{l-1}+\mathbf{b}^l\right)$.
An activation function is not applied to the last layer, so $\mathbf{y} = \mathbf{W}^{L+1}\mathbf{x}^{L}+\mathbf{b}^{L+1}$.
\subsection{Problem Formulation}
\begin{definition}[Repair Problem] Let $N\!N_o$ be a trained neural network over the training input-output space $\mathcal{X}_o\times\mathcal{T}\subseteq\mathbb{R}^{\lvert\mathbf{x}^0\rvert}\times\mathbb{R}^{\lvert\mathbf{t}\rvert}$ and $\Psi(\mathbf{y},\mathbf{x}^0)$ be a constraint (predicate) on the output $\mathbf{y}$ of $N\!N_o$ for a set of inputs of interest $\mathbf{x}^0\in \mathcal{X}_r\subseteq\mathcal{X}_o$. The Repair Problem is to modify the weight and bias terms of $N\!N_o$ such that $\Psi(\mathbf{y},\mathbf{x}^0)$ holds for the new neural network $N\!N_r$.
\end{definition}
By repairing $N\!N_o$, we also aim to maintain the performance of the original network.
Thus, the repaired network $N\!N_r$ should both satisfy $\Psi(\mathbf{y},\mathbf{x}^0)$ for the inputs $\mathbf{x}^0\in\mathcal{X}^{r}$ and maintain the prediction/classification performance of $N\!N_o$ for the inputs in the original training space domain $\mathcal{X}_o\times\mathcal{T}$.
To satisfy the latter, the minimal network modification method proposed in \cite{goldberger2020minimal} minimizes the $d$-norm error between each two weight and bias vectors of $N\!N_o$ and $N\!N_r$ while satisfying $\Psi(\mathbf{y},\mathbf{x}^0)$.
Although the authors in \cite{goldberger2020minimal} showed that their method can successfully modify the network to correct the observed misclassified points in a classification problem, a minimal deviation from the original weights is not necessarily a sufficient guarantee for maintaining the prediction/classification performance of the original network over the whole input space $\mathcal{X}_r$.
As an example, assume $NN_o$ is trained as a binary classifier (the highest output determines the classifier label for a given input) with one hidden layer ($L=1$) and each layer of size two.
Assume the original weights and biases are
\begin{align*}
\mathbf{W}^1&= \begin{bmatrix}
0.2 & 1 \\
0.01 & 1
\end{bmatrix}~,
\mathbf{b}^1 = \begin{bmatrix}
0.2\\
-0.11
\end{bmatrix}, \\
\mathbf{W}^2&= \begin{bmatrix}
0.5 & 1 \\
2 & -1
\end{bmatrix},
\mathbf{b}^2 = \begin{bmatrix}
-0.2\\
-1
\end{bmatrix}.
\end{align*}
Assume $\mathbf{x}^0=[9,0]^T$ is found as a misclassified point in $N\!N_o$ for which the true label is $y_0>y_1$.
A minimal change of weight $w^1_{01} = 0.01$ to $0.02$ can repair this network to correctly classify this point.
However, this repair changes the label of the classifier for $\mathbf{x}^0=[100,20]^T$ and creates a new misclassified point that was correctly classified in the original network.
Hence, a subtle change in the weights may cause the network to significantly deviate from its original performance. Therefore, it is important for the repaired network $N\!N_r$ to also minimize the defined loss function $E(\mathbf{W},\mathbf{b})$ of the original network $N\!N_o$.
\begin{figure}[t]
\centering
\includegraphics[scale=0.27]{NN.pdf}
\caption{Multi-layer Perceptron}
\label{fig:MLP}
\end{figure}
\begin{definition}[Minimal Repair Problem] \label{MinRepair-formulation} Let $N\!N_o$ be a trained neural network over the training input-output space $\mathcal{X}_o\times\mathcal{T}\subseteq\mathbb{R}^{\lvert\mathbf{x}^0\rvert}\times\mathbb{R}^{\lvert\mathbf{t}\rvert}$ and $\Psi(\mathbf{y},\mathbf{x}^0)$ be a predicate that represents some constraints on the output $\mathbf{y}$ of $N\!N_o$ for the set of inputs of interest $\mathbf{x}^0\in \mathcal{X}_r\subseteq\mathcal{X}_o$.
The Minimal Repair Problem is to modify the weights of $N\!N_o$ such that the new network $N\!N_r$ satisfies $\Psi(\mathbf{y},\mathbf{x}^0)$ while preserving the topology of $N\!N_o$ and minimizing the original defined loss function $E(\mathbf{W},\mathbf{b})$ over the original domain $\mathcal{X}_o\times\mathcal{T}$.
\end{definition}
\section{Layer-wise Network Repair as a Mixed-Integer Quadratic Program (MIQP)} \label{sec: MinRepair-MIP}
To solve the Minimal Repair Problem, we can formulate it as an optimization to minimize the loss function $E(\mathbf{W},\mathbf{b})$ subject to $(\mathbf{x}^0,\mathbf{t})\in\mathcal{X}_o\times\mathcal{T}$ and $\Psi(\mathbf{y},\mathbf{x}^0)$ for $\mathbf{x}^0\in \mathcal{X}^{r}$.
However, the resulting optimization problem is non-convex and difficult to solve due to the nonlinear ReLU activation function and high-order nonlinear constraints resulted from the multiplication of terms involving the weight/bias variables.
An alternative approach is to obtain a sub-optimal solution by just focusing on a single layer repair.
Layer-wise repair just modifies the weights and biases of a single layer to adjust the predictions so as to minimize $E(\mathbf{W},\mathbf{b})$ and to satisfy $\Psi(\mathbf{y},\mathbf{x}^0)$.
The layer-wise Minimal Repair Problem is defined as follows
\begin{definition}[Layer-wise Minimal Repair Problem] \label{SingleLayMinRepair-formulation} Let $N\!N_o$ denote a trained neural network with $L$ hidden layers over the training input-output space $\mathcal{X}_o\times\mathcal{T}\subseteq\mathbb{R}^{\lvert\mathbf{x}^0\rvert}\times\mathbb{R}^{\lvert\mathbf{t}\rvert}$ and $\Psi(\mathbf{y},\mathbf{x}^0)$ denote a predicate that represents some constraints on the output $\mathbf{y}$ of $N\!N_o$ for the set of inputs of interest $\mathbf{x}^0\in \mathcal{X}_r\subseteq\mathcal{X}_o$.
The Layer-wise Minimal Repair Problem is to modify the weights of a layer $l\in\{1,\cdots,L+1\}$ in $N\!N_o$ such that the new network $N\!N_r$ satisfies $\Psi(\mathbf{y},\mathbf{x}^0)$ while preserving the topology of $N\!N_o$ and minimizing the original defined loss function $E(\mathbf{W}^l,\mathbf{b}^l)$ over the original domain $\mathcal{X}_o\times\mathcal{T}$.
\end{definition}
In our optimization problem, we modeled the ReLU activation function $R(z)=\max\{0,z\}$ with a disjunctive inequality constraint \cite{balas1979disjunctive}.
Assume $\phi$ is a Boolean variable and $z = x-s$, where $x,s\geq 0$.
In the disjunctive inequality (\ref{eq: disjunction}), we have $x=z$ when $\phi=1$ ($z\geq 0$) and $x=0$ when $\phi=0$ ($z=-s< 0$). Here $x$ determines the output of ReLU activation function $R(z)$,
\begin{align}\label{eq: disjunction}
\left[\begin{array}{c}
\phi = 0\\
x\leq 0
\end{array}\right]\bigvee\left[\begin{array}{c}
\phi = 1\\
s\leq 0
\end{array}\right].
\end{align}
Disjunctive Ineq. (\ref{eq: disjunction}) can be relaxed into mixed integer algebraic equations using the Big-M or Convex-Hull reformulation \cite{tsay2021partition,belotti2011disjunctive}.
We also assume that the predicate $\Psi(\mathbf{x}^0,\mathbf{y})$ is represented as a linear constraint.
\begin{Assumption}\label{assump: linear_predicate}
The predicates $\Psi(\mathbf{x}^0,\mathbf{y})$ are formulated as linear equality/inequality constraints.
\end{Assumption}
Moreover, since $\mathcal{X}_r$ and $\mathcal{X}_o$ are not necessarily convex, we formulate the layer-wise minimal optimization problem over a data set sampled from $\mathcal{X}_o\times\mathcal{T}$ and $\mathcal{X}_r\times \Tilde{{\mathcal{T}}}$, where $\Tilde{{\mathcal{T}}}$ is the set of original target values of inputs in $\mathcal{X}_r$.
\begin{remark}
The predicate $\Psi(\mathbf{x}^0,\mathbf{y})$ defined over $\mathbf{x}^0\in \mathcal{X}_r$ is not necessarily compatible with the target values in $\Tilde{\mathcal{T}}$.
It means that the predicate may bound the NN output for $\mathcal{X}_r$ input space such that not allowing an input $\mathbf{x}^0\in \mathcal{X}_r$ to reach its target value in $\Tilde{\mathcal{T}}$.
It is a natural constraint in many applications.
For instance, due to the safety constraints, we may not allow a NN controller to follow its original control reference for a given unsafe set of input states.
\end{remark}
For a given layer $l$, we also define $E(\mathbf{W}^l,\mathbf{b}^l)$ in the form of sum of square loss as follows
\begin{align}\label{eq: loss}
E(\mathbf{W}^l,\mathbf{b}^l) = \sum^{N}_{n=1}\lVert \mathbf{y}_n(\mathbf{x}_n^0,\mathbf{W}^l,\mathbf{b}^l)-\mathbf{t}_n\rVert^2_2,
\end{align}
where $\mathbf{W}^l$ and $\mathbf{b}^l$ are the weight and bias vectors of the layer $l$, respectively, and $\{(\mathbf{x}^0_n,\mathbf{t}_n)\}^N_{n=1}\sim \mathcal{X}_o\times\mathcal{T} \cup \mathcal{X}_r\times\Tilde{\mathcal{T}}$ ($\lVert\cdot\rVert_2$ denotes the Euclidean norm).
Here, since we only repair the weights and biases of the target layer $l$, the loss term $E$ is a function of the weight matrix and bias vector of $\mathbf{W}^l$ and $\mathbf{b}^l$, respectively.
Hence, the weight and bias terms of all layers except the target layer $l$ are fixed, as shown in Fig. \ref{fig:MLP-rpair}.
\begin{figure}[t]
\centering
\includegraphics[scale=0.29]{NN3.pdf}
\caption{Multi-layer Perceptron repair given a sample $n$}
\label{fig:MLP-rpair}
\end{figure}
Considering ReLU activation functions formulated by disjunctive Ineq. (\ref{eq: disjunction}), the loss function (\ref{eq: loss}) and Asmp. \ref{assump: linear_predicate}, we formulate the Layer-wise Minimal Repair Problem in Def. \ref{SingleLayMinRepair-formulation} as a Mixed Integer Quadratic Program (MIQP).
\begin{opt} \label{opt: minrepair1}
Let $N\!N_o$ be a neural network with $L$ hidden layers, $\Psi(\mathbf{y},\mathbf{x}^0)$ be a predicate, and $\{(\mathbf{x}^0_n,\mathbf{t}_n)\}^N_{n=1}$ be an input-output data set sampled from $(\mathcal{X}_o\times\mathcal{T}) \cup (\mathcal{X}_r\times\Tilde{\mathcal{T}})$ over the sets $\mathcal{X}_o$, $\mathcal{X}_r$, $\mathcal{T}$, and $\Tilde{\mathcal{T}}$ all as defined in Def. \ref{SingleLayMinRepair-formulation}. Also, let $E(\mathbf{W}^l,\mathbf{b}^l)$ be the loss function defined in (\ref{eq: loss}). Repair the weight and bias vectors $\mathbf{W}^l$ and $\mathbf{b}^l$ of the $l$\textsuperscript{th} layer in $N\!N_o$ (as shown in Fig. \ref{fig:MLP-rpair}) so as to minimize the cost function (\ref{eq:cost}) subject to the constraints (\ref{eq: repair weights-last-layer})-(\ref{eq: weights bound-mid-layer-delta}).
\begin{align}
& \label{eq:cost}
\underset{\substack{
\mathbf{W}^l,\mathbf{b}^l,\mathbf{y}_n,\delta,\\ \{\mathbf{x}_n^i\}_{i=l}^{L},
\{\mathbf{s}^i\}_{i=l}^{L},\{\bm{\phi}_n^i\}_{i=l}^{L}}
}
{\text{min}}~E(\mathbf{W}^l,\mathbf{b}^l)+\delta \\
& \text{s.t.}\nonumber\\
&~~~\label{eq: repair weights-last-layer}\mathbf{y}_n=\mathbf{W}^{L+1}\mathbf{x}^{L}_n+\mathbf{b}^{L+1},~~~~~~~~~~~~~~ \text{for }\{n\}_{n=1}^N,\\[3pt]
&~~~\label{eq: repair weights-mid-layer}\mathbf{x}^i_n-\mathbf{s}^i_n=\mathbf{W}^i\mathbf{x}^{i-1}_n+\mathbf{b}^i,\nonumber\\
&~\qquad\qquad\qquad\qquad\qquad~~~ \text{for } \{n\}_{n=1}^N \text{ and }\{i\}_{i=l}^{L},\refstepcounter{equation}\subeqn\\[3pt]
&~~~\mathbf{x}^i_n,\mathbf{s}^i_n\geq 0,\qquad\quad~~~~~~~ \text{for } \{n\}_{n=1}^N \text{ and }\{i\}_{i=l}^{L},\subeqn\\[3pt]
&~~~\bm{\phi}^i_n\in \{0,1\}^{\lvert\mathbf{x}^i\rvert},\quad~~~~~~~ \text{for } \{n\}_{n=1}^N \text{ and }\{i\}_{i=l}^{L},\subeqn\\[3pt]
&~~~\left[\begin{array}{c}
\phi^i_{j,n} = 0\\
x^i_{j,n}\leq 0
\end{array}\right]\bigvee\left[\begin{array}{c}
\phi^i_{j,n} = 1\\
s^i_{j,n}\leq 0
\end{array}\right],\nonumber\\
&~\qquad\qquad\qquad~~~~ \text{for } \{n\}_{n=1}^N,~\{i\}_{i=l}^{L}\text{ and }\{j\}_{j=0}^{\lvert\mathbf{x}^i\rvert},\subeqn\\[3pt]
&~~~\label{eq: requirement-mid-layer}\Psi(\mathbf{y}_n,\mathbf{x}^0_n), \quad~~~~~~~~~~ \text{for }\{n\}_{n=1}^N \text{ and }\mathbf{x}^0_n\in \mathcal{X}_r,\\[3pt]
&~~~\label{eq: weights bound-mid-layer}\lvert w^l_{kj}-w^{l,prev}_{kj}\rvert\leq\delta,~~~ \text{for } \{k\}_{k=0}^{\lvert\mathbf{x}^{l-1}\rvert}\text{ and }\{j\}_{j=0}^{\lvert\mathbf{x}^l\rvert},\refstepcounter{equation}\subeqn\\[3pt]
&~~~\lvert b^l_{j}-b^{l,prev}_{j}\rvert\leq\delta,\qquad\quad\quad~~~~~~~~~~~ \text{for } \{j\}_{j=0}^{\lvert\mathbf{x}^l\rvert},\subeqn\\[3pt]
&~~~\label{eq: weights bound-mid-layer-delta}\delta\geq 0.\subeqn
\end{align}
\end{opt}
The layer-wise repair scheme is shown in Fig. \ref{fig:MLP-rpair} for repairing $\mathbf{x}^{l-1}\xrightarrow{\mathbf{W}^l,\mathbf{b}^l} \mathbf{x}^l$. The sample values of $\mathbf{x}_n^{l-1}$ can be obtained by the weighted sum of the nodes in its previous layers starting from $\mathbf{x}_n^0$ for all $N$ samples $\{n\}^N_{n=1}$. In R-OPT, constraint (\ref{eq: repair weights-last-layer}) represents the forward pass of last layer $\mathbf{x}^{L}\xrightarrow{\mathbf{W}^{L+1},\mathbf{b}^{L+1}} \mathbf{y}$. Constraint
(5)
represents the forward pass of layers $\mathbf{x}^{i-1}\xrightarrow{\mathbf{W}^{i},\mathbf{b}^{i}} \mathbf{x}^{i}$ for $\{i\}_{i=l}^{L}$ with ReLU activation function formulated as disjunctive functions (\ref{eq: disjunction}). Constraint (\ref{eq: requirement-mid-layer}) is the given predicate on $\mathbf{y}$ defined over $\mathbf{x}^0\in\mathcal{X}_r$. Finally, Constraint
(7)
bounds the error between the weight and bias terms $\mathbf{W}$ and $\mathbf{b}$, and their corresponding original weight and bias terms $\mathbf{W}^{prev}$ and $\mathbf{b}^{prev}$ by $\delta$. By adding $\delta$ to (\ref{eq:cost}), we aim to minimize weights deviation as well as $E(\mathbf{W}^l,\mathbf{b}^l)$ in the repair process.
\begin{remark}
Since the output layer is not passed through a ReLU activation function, applying repair to the last layer $L+1$ does not create mixed integer constraints
(5). Therefore, Minimal Repair Problem for the last layer is a Quadratic Program (QP).
\end{remark}
\section{Empirical Experiments}\label{sec: sim-sec}
We show the application of the Layer-wise Minimal Repair framework in bounding an affine transformation, applying constraints on a learned forward kinematics model, correcting an erroneous NN in a classification problem, and applying safety constraints on a NN controller. In all experiments, except the classification problem, we first train a neural network with training data uniformly sampled from the input-output training domain $\mathcal{X}_o\times\mathcal{T}$. We use our proposed repair formulation R-OPT to repair the network given the target layer $l$ and the defined predicate $\Psi(\mathbf{y},\mathbf{x}^0)$ over the input space $\mathcal{X}_r\subseteq\mathcal{X}_o$. In network repair, we use equal samples from both $\mathcal{X}_o\times\mathcal{T}$ and $\mathcal{X}_r\times\Tilde{\mathcal{T}}$. We formulate the optimization using Pyomo open-source optimization language in Python \cite{hart2011pyomo,bynum2021pyomo}, and Pyomo.GDP for modeling the disjunctive constraints \cite{chen2018pyomo}. We use Gurobi 9.1 as solver \cite{gurobi}.
\subsection{In-place Rotation}
Our method is first tested on a network that learned in-place rotation.
We first train a two-hidden-layer network $N\!N_o$ ($L=2$) with $3$ nodes in each layer to learn an in-place 45-degrees-counterclockwise rotation function $\textsc{Rot}\!:\mathcal{X}_o\rightarrow\mathcal{T}$ in 2D space.
The original input space is $\mathcal{X}_o = \{\mathbf{x}^0=[x,y]\in \mathbb{R}^2\;|\; 1\leq x,y\leq4\}$. The output space $\mathcal{T}$ represents the set of counterclockwise rotated points in $\mathcal{X}_o$ by $45$ degrees defined as $\mathcal{T}=\{\mathbf{y}\in\mathbb{R}^{2}\;|\;\mathbf{y}=\textsc{Rot}(\mathbf{x}_0)\text{ for }\mathbf{x}_0\in\mathcal{X}_o\}$. We repair $N\!N_o$ to push all the outputs $\mathbf{y}$ inside a Quadrilateral defined as
\begin{align*}
\Psi(\mathbf{y},\mathbf{x}^0)=\{\mathbf{y}\in\mathcal{T}\;|\;\lVert[x,y] - [2.5,2.5]\rVert_1\leq \frac{5\sqrt{2}}{4}\},
\end{align*}
where $\lVert\cdot\rVert_1$ denotes the Taxicab norm. Here, $\mathcal{X}_r=\mathcal{X}_o$. We repair the network with $N=200$ samples.
Figure \ref{fig: affine-trans} (a) demonstrates network predictions after applying repair to the first layer $\mathbf{x}^0\xrightarrow{\mathbf{W}^1,\mathbf{b}^1}\mathbf{x}^1$.
As it is shown, the predicate $\Psi(\mathbf{y},\mathbf{x}^0)$ pushes the predicted outputs inside the green set.
The bar plot in Fig. \ref{fig: affine-trans} (b) illustrates the Mean Square Error (MSE) of the network predictions after applying repair to different layers (showed with colored bars) compared to the original network predictions MSE (showed with dashed line).
This plot separates the prediction errors of samples based on their targets whether located inside or outside of the green set.
Clearly, prediction error for the points with targets inside the green set is lower than the points with targets outside the green set.
The prediction performance improves by moving from the last layer to the first layer for both inside and outside cases. The large error reduction by hidden layers is due to the capture of non-linearities in their forward pass. However, the last layer only has the scaling factor, thereby, results smaller error reduction compared to the hidden layers.
Moreover, to achieve a significant error reduction, we can select any hidden layer since moving among hidden layers does not result a large error change.
In practice, we observed that repairing the latter hidden layers helps Gurobi solver to find a solution faster since the number of integer variables created by modeling ReLU activation function is increased by $2^l$ moving backward from each hidden layer $l$.
In this experiment, the optimal weight deviations for each layer are $\delta_1=0.391$, $\delta_2 = 0.4408$, and $\delta_3 = 0.7605$ for the first, second, and third layers, respectively. It implies that repairing by moving backward from the last layer needs smaller weight deviation to adjust the network's behavior.
\begin{figure}[t]
\centering
\includegraphics[scale=0.26]{temp.pdf}
\caption{In-place rotation: (a) predictions after applying repair to the first layer, (b) Mean Square Error (MSE) between testing and original targets.}
\label{fig: affine-trans}
\end{figure}
\subsection{Forward Kinematics}
In our second test, we bound the output of a three-hidden-layer network $N\!N_o$ ($L=3$) with $20$ nodes in each layer that learned the forward kinematics of a 6-DOF KUKA robot arm \cite{kazi2005research}.
The forward kinematics function $\textsc{Fk}\!:\mathcal{X}_0\rightarrow\mathcal{T}$ for this robot is a mapping from the $6$ joint angles to the location of end-effector in 3D space.
To train the original network, we generate a forward kinematics data set using RoboAnalyzer simulator \cite{gupta2017roboanalyzer}.
Here, the inputs to the $N\!N_o$ are the $\sin(\theta_q)$ and $\cos(\theta_q)$ of each joint angle $\theta_q$ for $\{q\}^6_{q=1}$ ($12$ inputs in total).
It helps to remove the discontinuity that occurs by mapping $\theta$ to $[-\pi,\pi]$.
Given the trigonometrically filtered input angles, we define the original input space as $\mathcal{X}_o = \big\{\mathbf{x}^0\in \mathbb{R}^{12}\;|\; -1\leq x^0_j\leq1 \text{ for }\{j\}^{12}_{j=1}\big\}$.
The output space $\mathcal{T}$ is then defined as $\mathcal{T}=\{\mathbf{y}\in\mathbb{R}^{3}\;|\;\mathbf{y}=\textsc{Fk}(\mathbf{x}_0)\text{ for }\mathbf{x}_0\in\mathcal{X}_o\}$, where the outputs $\mathbf{y}=[x,y,z]$ are the coordinates of end-effector in 3D space.
We repair the network to satisfy a predicate defined as $\Psi(\mathbf{y},\mathbf{x}^0)=\{\mathbf{y}\in\mathcal{T}\;|\; x\leq 0.5\}$ for $\mathcal{X}_r=\mathcal{X}_o$.
Here, $x>5$ can be the space of a human worker who is jointly interacting with the robot.
So, $\Psi(\mathbf{y},\mathbf{x}^0)$ bounds the location of the end-effector to not operate in this space. We repair the network with $N = 500$ samples.
Figure \ref{fig: forward-kin} illustrates the results of applying network repairs on the forth and third layers.
As shown in Fig. \ref{fig: forward-kin} (a), the repaired network successfully pushes the outputs inside the robot's operation region and has an accurate prediction performance for the points inside this bound (see Fig. \ref{fig: forward-kin} (b)).
Prediction errors also demonstrate that the repair of the third layer has lower error compared to the forth layer.
In this experiment, the optimal weight deviations for each layer are $\delta_3 = 0.5761$, and $\delta_4 = 0.8024$ for the third and forth layers, respectively.
\begin{figure}[t]
\centering
\includegraphics[scale=0.28]{temp2.pdf}
\caption{Forward kinematics: (a) output predictions, (b) Mean Square Error (MSE) between testing and original targets.}
\label{fig: forward-kin}
\end{figure}
\subsection{Unmanned Aircraft Collision Avoidance system X (ACAS Xu)}
ACAS Xu is a variant of Aircraft Collision Avoidance System X (ACAS X) that provides horizontal manoeuvre advisories for unmanned aircraft \cite{kochenderfer2011robust}. This system uses a look up table that produces 5 different horizontal advisory outputs of clear-of-conflict ($COC$), weak right/left heading rates of $\pm 1.5^{\circ} / s$ ($WR$ and $WL$), and strong right/left heading rates of $\pm 3^{\circ} / s$ degrees per second ($SR$ and $SL$), i.e. $\mathbf{y}=[COC,WR,WL,SR,SL]$. The inputs of system are 7 sensor measurements of distance from ownship to intruder $\rho$ ($m$), angle to intruder relative to ownership heading
direction $\theta$ ($rad$), heading angle of intruder relative to ownship
heading direction $\psi$ ($rad$), speed of ownship $v_{own}$ ($m/s$), speed of intruder $v_{int}$ ($m/s$), time until loss of vertical separation $\tau$ ($s$), and previous advisory $a_{prev}$ ($^{\circ} / s$). The advisory in $\mathbf{y}$ with the minimum value is the selected advisory for a given input. To improve storage efficiency, an array of 45 Deep Neural Networks (DNN) are trained in \cite{julian2016policy} for the discretized combinations of $\tau$ and $a_{prev}$ to learn the look up table. Inputs of each DNN are $\mathbf{x}^0=[\rho,\theta,\psi,v_{own},v_{int}]$. Each DNN has 6 hidden layers with 50 ReLU activation nodes in each layer. DNNs return a horizontal advisory given the input sensor measurements. Given the application of this system, it is very important for the system to return correct advisories. In this experiment, we focus on the DNN trained for $a_{prev} = WL$ and $\tau = 100$. Following the look up table, the desired output property of this network is to return $WL$ or $COC$ horizontal advisories for the input space \begin{align*}
\mathcal{X}_o=\{\mathbf{x}^0=&[\rho,\theta,\psi,v_{own},v_{int}]\in\mathbb{R}^5\;|\;-0.33\leq\rho\leq 0.68 \\
&\text{ and } -0.5\leq \theta,\psi,v_{own},v_{int}\leq 0.5\}.
\end{align*}
Thus, the predicate is defined as $\Psi(\mathbf{y},\mathbf{x}^0)=\{\mathbf{y}\in\mathbb{R}^5\;|\;COC,WL<WR,SR,SL\}.$ To find input samples in $\mathcal{X}_o$ that violate $\Psi(\mathbf{y},\mathbf{x}^0)$, we use Marabou verification tool proposed in \cite{katz2019marabou}.
Figure \ref{fig: classification} illustrates the projection of local misclassified set of inputs $\mathcal{X}^1_r$ and $\mathcal{X}^2_r$, shown with green and blue dots respectively, on $\theta$ and $\rho$ input spaces.
Our method successfully corrects the local misclassified set of points $\mathcal{X}^1_r\subseteq\mathcal{X}^1_o$ and $\mathcal{X}^2_r\subseteq\mathcal{X}^2_o$ by applying our NN repair technique on the last layer.
In this experiment, $\delta=9.8\times10^{-5}$ and $\delta=1.14\times10^{-3}$ for repairing $\mathcal{X}^1_r$ and $\mathcal{X}^2_r$, respectively, that demonstrates the satisfaction of property $\Psi(\mathbf{y},\mathbf{x}^0)$ with just a small deviation of last layer's original weights.
\begin{figure}[t]
\centering
\includegraphics[scale=0.34]{classify.pdf}
\caption{Collision Avoidance System: projection of repaired regions on $\theta$ and $\rho$ input spaces shown in green and blue. Misclassified points shown with dots. $\lVert \mathbf{x}^0_{adv1}-\mathbf{x}^0_{adv2}\rVert_2$ represents the average Euclidean distance between the two local repaired regions.}
\label{fig: classification}
\end{figure}
\subsection{Safe Controller}
In our last experiment, we repair a mobile robot point-to-goal controller to avoid a static obstacle.
Consider a robot that is following a unicycle model as
\begin{align}\label{eq: robot-model}
\dot{\mathbf{x}}=\mathbf{g}(\mathbf{x})\mathbf{u} ={\small\begin{bmatrix}
\cos(\theta) & 0\\
\sin(\theta) & 0\\
0 & 1
\end{bmatrix}}\mathbf{u}.
\end{align}
where states are $\mathbf{x}\!=\![x,y,\theta ]^T\in \mathcal{X}\!\in\! \mathbb{R}^2\!\times\![-\pi,\pi)$ and control inputs are $\mathbf{u}=[v,\omega]^T\in\mathcal{U}\in \mathbb{R}^2$.
The variables $x$, $y$, $\theta$ denote the longitudinal and lateral positions of the robot and heading angle, respectively.
The controls $v$ and $\omega$ represent the linear and angular velocities of robot, respectively.
We first use imitation learning to train a neural network controller $N\!N_o$ that imitates a QP-based Controller $\textsc{Ctr}:\mathcal{X}\rightarrow\mathcal{U}$ as an expert.
$N\!N_o$ receives the state of the robot as the input $\mathbf{x}^0=\mathbf{x}$ ($\mathcal{X}_o=\mathcal{X}$) and outputs the control action, i.e. $\mathbf{y}=\mathbf{u}$ ($\mathcal{T}=\mathcal{U}$).
The training samples are collected from $\mathcal{X}_o\times\mathcal{T}=\big\{\{(\mathbf{x}^0_n,\mathbf{t}_n)\}^N_{n=1}\in \mathcal{X}\!\times\!\mathcal{U}\;|\; \mathbf{x}^0_1\in\mathcal{X}_{init},\mathbf{x}^0_N\in\mathcal{X}_g,\mathbf{t}_n=\textsc{Ctr}(\mathbf{x}_n^0)\big\}$.
This controller steers the robot from an initial set of states $\mathcal{X}_{init}=\{\mathbf{x}^0\in \mathcal{X}_o ~\lvert ~x,y\in[-5,-3],\theta\in[0,\frac{\pi}{2}]\}$ to the goal set $\mathcal{X}_g=\{\mathbf{x}^0\in\mathcal{X}_o\;|\;\big\lVert V(\mathbf{x}^0) \leq 0 \}$.
Here $V:\mathcal{X}_o\rightarrow\mathbb{R}$ is a continuously differentiable function
\begin{align}\label{eq: lyap}
V(\mathbf{x}^0)=\big\lVert[x,y]^T-[0,0]\big\rVert^2_2 - 0.2^2.
\end{align}
We model $N\!N_o$ to have 2 hidden layers with 10 ReLU nodes per layer.
Figure \ref{fig:controller} (a) shows the trajectories of system under QP-based learned policy.
Readers are referred to \cite{yaghoubi2020training} for more details on using QP-based controllers in imitation learning.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.23\textwidth}
\centering
\includegraphics[scale=0.233]{original_prediction.pdf}
\end{subfigure}%
\hfill
\begin{subfigure}[t]{0.24\textwidth}
\centering
\includegraphics[scale=0.233]{repaired_prediction.pdf}
\end{subfigure}
\caption{Safe controller: (a) simulated trajectories of system under learned policy by the original network, (b) simulated trajectories of system under repaired policy.}
\label{fig:controller}
\end{figure}
Now assume for the same control task, the robot needs to avoid an unsafe region (the ellipsoidal violet region in Fig. \ref{fig:controller}) defined as
\begin{align}\label{eq: unsafe set}
\mathcal{X}_{unsafe} = \{ \mathbf{x}^0 \in \mathcal{X}_o ~\lvert ~h(\mathbf{x}^0)< 0\},
\end{align}
where $h:\mathcal{X}_o\rightarrow\mathbb{R}$ is a continuously differentiable safety measure that represents a static obstacle defined as
\begin{align}\label{eq: cbf}
h(\mathbf{x}^0) = \lVert[x,y]^T - [-1.5,-3]^T\rVert_2^2 - 0.2^2.
\end{align}
We treat $h(\mathbf{x}^0)$ as a Control Barrier Function (CBF) \cite{ames2019control} to derive a safety predicate for the output $\mathbf{y}=\mathbf{u}$ of $N\!N_o$ controller under which the states $\mathbf{x}$ of system ($\ref{eq: robot-model}$) that start in $\mathcal{X}\setminus \mathcal{X}_{unsafe}$ never enters the unsafe set $\mathcal{X}_{unsafe}$.
\begin{definition}[Control Barrier Function (CBF)]
Given the nonlinear control affine system (\ref{eq: robot-model}) and an unsafe set $\mathcal{X}_{unsafe}$, a continuously differentiable function $h:\mathcal{X}\rightarrow\mathbb{R}$ is a CBF for this system if there exists an extended $\mathcal{K}$ class function $\alpha:\mathbb{R}\rightarrow\mathbb{R}$ (strictly increasing and $\alpha(0)=0$) such that
\begin{align}
&h(\mathbf{x})\geq 0 &&\forall \mathbf{x}\in \mathcal{X}\setminus \mathcal{X}_{unsafe},\label{eq: CBF1}\\
&h(\mathbf{x})< 0 &&\forall \mathbf{x}\in \mathcal{X}_{unsafe},\label{eq: CBF2}\\
&L_g h(\mathbf{x}) \mathbf{u} +\alpha(h(\mathbf{x}))\!\geq\! 0 &&\forall \mathbf{x}\in \mathcal{X}\setminus \mathcal{X}_{unsafe}, \label{eq: CBF3}
\end{align}
where $L_g h(\mathbf{x})$ is the first order Lie derivative of $\frac{\partial h}{\partial \mathbf{x}}^T\mathbf{g}$.
\end{definition}
If the control $\mathbf{u}$ satisfies (\ref{eq: CBF3}), then the state of system (\ref{eq: robot-model}) never leaves $\mathcal{X}\setminus \mathcal{X}_{unsafe}$.
Our defined candidate function (\ref{eq: cbf}) satisfies the conditions (\ref{eq: CBF1}) and (\ref{eq: CBF2}) given the defined unsafe set (\ref{eq: unsafe set}).
Now to satisfy (\ref{eq: CBF3}), for all $\mathbf{x}^0\!\in\!\mathcal{X}_r=\mathcal{X}_o\setminus\mathcal{X}_{unsafe}$ we impose the predicate $\Psi_1(\mathbf{x}^0,\mathbf{y})=\{\mathbf{y}\!\in \!\mathcal{U}\;|\;L_g h(\mathbf{x}^0) \mathbf{y} +\gamma h(\mathbf{x}^0)\!\geq\! 0\}$ to the output $\mathbf{y}$ of $N\!N_o$, where $\gamma \geq 0$.
To ensure that the repaired controller steers the robot toward the goal set $\mathcal{X}_g$, we also treat $V(\mathbf{x}^0)$, defined in (\ref{eq: lyap}), as a Lyapunov-like function.
\begin{definition}[Lyapunov-like Function (CLF)]
Given the nonlinear control affine system (\ref{eq: robot-model}) and a goal set $\mathcal{X}_g$, a continuously differentiable function $V:\mathcal{X}\rightarrow\mathbb{R}$ is a CLF for this system if
\begin{align}
&V(\mathbf{x})> 0 &&\forall \mathbf{x}\in \mathcal{X}\setminus \mathcal{X}_g,\label{eq: CLF1}\\
&V(\mathbf{x})\leq 0 &&\forall \mathbf{x}\in \mathcal{X}_g,\label{eq: CLF2}\\
&L_g V(\mathbf{x}) \mathbf{u} \!\leq\! 0 &&\forall \mathbf{x}\in \mathcal{X}\setminus \mathcal{X}_g. \label{eq: CLF3}
\end{align}
\end{definition}
If there exists a control $\mathbf{u}$ that satisfies (\ref{eq: CLF3}), then the states of system (\ref{eq: robot-model}) reach the goal set $\mathcal{X}_g$ where $V(\mathbf{x})\leq 0$.
Similar to the safety condition, $V(\mathbf{x}^0)$, defined in (\ref{eq: lyap}), satisfies the conditions (\ref{eq: CLF1}) and (\ref{eq: CLF2}). So for all $\mathbf{x}^0\!\in\!\mathcal{X}_r$ we impose
$\Psi_2(\mathbf{x}^0,\mathbf{y})=\{\mathbf{y}\in \mathcal{U}\;|\;L_g V(\mathbf{x}^0) \mathbf{y} \leq \beta\}$ to the output $\mathbf{y}$ of $N\!N_o$ to satisfy condition (\ref{eq: CLF3}).
Here $\beta\geq 0$ is a slack variable that minimally relaxes the predicate $\Psi_2$ when enforcing the goal-reaching constraint to $\mathbf{u}$ contradicts the safety condition $\Psi_1$.
For further details on Lyapunov-like functions, CBFs, and their application in safe mobile robot navigation, readers are referred to \cite{ames2019control,yaghoubi2020risk,majd2021safe}.
Given the linear inequality constraints $\Psi_1$ and $\Psi_2$, we repair the last layer of $N\!N_o$ with 100 reference trajectories.
As illustrated in Fig. \ref{fig:controller} (b), the generated trajectories of system (\ref{eq: robot-model}) by the repaired policy successfully avoid the unsafe region showed in violet.
The repaired controller decreases the linear and angular velocities $v$ and $\omega$ for the states close to the unsafe region while the repaired policy converges to the original policy when states get farther from the unsafe region.
In this experiment, the optimal weight deviation is $\delta=2.45$.
\section{Conclusion and Future Work}\vspace{-3pt}
In this paper, we provided a framework for minimal layer-wise neural network repair.
For a repair, we adjust a trained neural network to satisfy a set of predicates that impose constraints on the output of NN for a given set of inputs of interest.
The experimental results demonstrated the success of our framework in repairing the trained networks to satisfy a set of predicates while maintaining the performance of the original network.
Inspired by our theoretical and experimental results, one is able to improve, change, and guarantee different aspects of performance of a trained neural network.
While we showed the capability in our framework in addressing safety, model improvement, and verification problems in different robot learning applications, MIQP formulation of network repair is not an easy problem to solve for high-dimensional robotic systems with large input spaces.
Our future work aims to address this problem by relaxing or decreasing the number of integer variables of our formulation.
Additionally, we aim to investigate an efficient sampling mechanism for the network repair that can help to decrease the number of integer variables as well.
In our experiments, we also observed that repairing the last layer did not necessarily satisfy a given property and, in some cases, it caused infeasibility.
While we did not observe this issue in repairing the hidden layers, we aim to explore more cases where infeasibility occurs.
Finally, we plan to create a public available neural network repair Python package integrated with TensorFlow model architectures \cite{tensorflow2015-whitepaper} in our future work.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
In an effort to create efficient electron structure methods, it has become commonplace to combine contrasting models of electron correlation ({\it e.g.} wave function, density functional, density matrix functional) in order to exploit their most effective attributes.\cite{Cizek1969, Jeziorski1981, Andersson1990, Mok1996, Handy2001, Cremer2001, Pulay2011, Becke2013, Crittenden2013, Tsuchimochi2014, Wallace2014, DNO1, RamosCordoba2016, BenavidesRiveros2017, Evangelista2018, ViaNadal2019, Fromager2007, Weimer2008, Toulouse2009, Janesko2009a, Chai2009, Stoyanova2013, Yamanaka2006, Rohr2010, NOFMP2a, NOFMP2b, Perez-Jimenez2007, Manni2014, Carlson2015, Gagliardi2017, Malcolm1998, Kohn2013, Hubert2016, DNO2, Hirao1992, Grimme1999, Grafenstein2000, Angeli2001, Angeli2002, Mostafanejad2020, Furche2008} Current methods that employ such a strategy, vary in computational cost and accuracy, and in most cases their remains room for improvement in both departments. The most common strategy is to decompose electron correlation into static (strong) and dynamic (weak) correlation, and approach each differently. The development of these methods involves the design of models that precisely capture each component, and also blend the approaches together in a seamless fashion. The former aspect can be thoroughly evaluated by applying the method to relatively simple systems, such as small hydrogen clusters.
Although they are not chemically complex, the electronic structure of \ce{H} atom clusters, particularly linear chains and regular lattices of \ce{H} atoms, can serve as a systematic testing ground for new treatments of electron correlation.\cite{Paldus1993, Kowalski1998PRL, Kowalski1998CPL, Schipper1999, Jankowski1999, VanVoorhis2000, Hachmann2006, AlSaidi2007, Rassolov2007, Sinitskiy2010, Stella2011, Bytautas2011, RamosCordoba2015, Bulik2015, Robinson2012a, Robinson2012b, Robinson2012c, Robinson2012d, Sand2013, Limacher2013, Kats2013, Henderson2014, Motta2017, Motta2020, Giner2020, Marie2021} Creating a model that accurately describes the surprisingly complex features of \ce{H} cluster potential energy surfaces can prove challenging. There have been several studies of such clusters with an array of methods, including density functional approximations, natural orbital functionals, and truncated post-HF methods. These systems have been shown to be effective for evaluating the ability of methods to describe both ground and excited state properties. While it is important for these methods to accurately describe the potential energy surface of these ``simple" systems, it is also important to accurately describe the underlying two-electron density. A method that correctly models the two-electron density, as well as the energy and one-electron density, is more likely to be universally successful.
The $\dno$ method combines a cumulant functional approach for the treatment of static correlation with an on-top density functional (or post-HF) to treat dynamic correlation.\cite{DNO1, DNO2} Amongst other approaches, it most resembles multiconfigurational-pair density funcitonal theory (MC-PDFT).\cite{Manni2014, Carlson2015, Gagliardi2017, Sand2017, Mostafanejad2020} In which, a multireference one-electron density matrix (1-RDM), and subsequent one-electron density, is used to determine the kinetic, nuclear-potential, and classical Coulomb energy. The on-top pair density, obtained from the multireference two-electron density matrix (2-RDM), is used to create a modified spin-density\cite{Manni2014} which is inserted into the usual exchange-correlation density functionals ({\it e.g.}~PBE\cite{Perdew1996}) to determine the exchange and correlation energy. The 2-RDM can be obtained from a traditional multireference self-consistent-field \cite{Manni2014, Carlson2015, Gagliardi2017, Sand2017}, a variational 2-RDM-driven complete active-space self-consistent field\cite{Mostafanejad2020}, or another economical alternative such as a pair-coupled cluster doubles calculation.\cite{Garza2015} The $\dno$ method uses an on-top density functional for only the dynamic correlation energy, where the functional depends directly on the on-top density. The remaining components of the energy are determined using the approximate $\dno$ 2-RDM.
This work begins by introducing a $\dno$ wave function (Subsection \ref{ssec:wf}) from which the $\dno$ two-electron density matrix and energy expression can be derived. The $\dno$ 2-RDM is then supplemented with an {\it ad hoc} term to correct for missing static correlation between electron pairs (Subsection \ref{ssec:hsc}). This is followed by a description of the determination of the dynamic correlation energy from the $\dno$ 2-RDM (Subsection \ref{ssec:odf}), and then an approach to reduce correlation double-counting (Subsection \ref{ssec:dc}). Then a new optimization algorithm for the $\dno$ 2-RDM is described (Subsection \ref{ssec:opt}), which involves the simultaneous optimization of the orbitals and the electron transfer variables. The performance of the optimization method is then evaluated relative to the previous $\dno$ 2-RDM optimization algorithm. The performance of $\dno$, in describing potential energy surfaces of small \ce{H} clusters, is then compared to established single- and multireference methods (Section \ref{sec:results}). The two-electron density is then used to analyze the static correlation provided by the $\dno$ model in the rectangular \ce{H4} cluster (Subsection \ref{ssec:rec2sq}). Finally, the results of the $\dno$ evaluation and analysis are summarized (Section \ref{sec:conc}).
\section{Theory}
\label{sec:theory}
\subsection{The $\dno$ wave function}
\label{ssec:wf}
The multideterminantal $\dno$ wave function,
\begin{equation}\label{eq:DNOwf}
\Psi_{\dno} = \sum_I c_I \Psi_I,
\end{equation}
is defined with respect to a reference configuration in which all open-shell (singly occupied) orbitals are considered frozen ({\it i.e.}~singly occupied in all determinants). Only doubly occupied orbitals are active and all excitations are $2n$-tuples (doubles, quadruples, ...) where no electrons are unpaired. In other words, with respect to the active orbitals, $\Psi_{\dno}$ is seniority-zero.\cite{Bytautas2011, Stein2014, Henderson2014, Poelmans2015, Chen2015, Limacher2015, VanMeer2018, Brzek2019, Vu2020} Also, each active pair of electrons has its own active space, consisting of the doubly occupied orbital in the reference wave function and some number of vacant (virtual) orbitals. Each determinant of the expansion is then defined by $2n$-tuple excitations of the reference configuration,
\begin{equation}
\Psi_I = \Psi_{i\bar{i}j\bar{j}k\bar{k}\dots}^{a\bar{a}b\bar{b}c\bar{c}\dots}
\end{equation}
where $i,j$ and $k$ refer to reference occupied orbitals, and $a,b$ and $c$ refer to all active orbitals. The absence of a bar, $i$, indicates an $\alpha$-spin orbital while the presence of a bar, $\bar{i}$, indicates a $\beta$-spin orbital. Considering this, it is useful to define the determinantal expansion coefficients, $c_I$, as a product of coefficients corresponding to specific double excitations of the active electron pairs,
\begin{equation}
c_I = d_i^a d_j^b d_k^c \dots
\end{equation}
In order to arrive at the $\dno$ formalism, the double excitation coefficients, $d_i^a$, are defined in terms of orbital occupancies, $n_i$, and electron transfer variables, $\Delta_{ia}$,
\begin{equation}
d_i^a = \delta_{ia} \sqrt{n_i} - \sqrt{\Delta_{ia}},
\end{equation}
where $\delta_{ia}$ is the Kroenecker delta and $\Delta_{ii}=0$. Orbital occupancies, occupied and vacant (with respect to the reference configuration), are defined as
\begin{align}
n_i & = 1 - \sum_a \Delta_{ia}, \\
n_a & = \sum_i \Delta_{ia}.
\end{align}
The $N$-representable components of the $\dno$ energy expression, $E^{\dno'}$, can then be obtained by expanding the expectation value of the $N$-electron hamiltonian in terms of $n_i$, $\Delta_{ia}$, and one and two-electron integrals over orbitals. Initially, the energy may be divided into diagonal and off-diagonal contributions,
\begin{align} \label{eq:EDNOp}
E^{\dno'} & = \langle \Psi_{\dno} | \hat{H} | \Psi_{\dno} \rangle \nonumber \\
& = \sum_I c_I^2 H_{II} + \sum_{I\ne J} c_I c_J H_{IJ} ,
\end{align}
where
\begin{equation}
H_{IJ} = \langle \Psi_I | \hat{H} | \Psi_J \rangle,
\end{equation}
and the electronic hamiltonian (in atomic units) is given by
\begin{equation}
\hat{H} = \sum_{i=1}^N -\tfrac{1}{2} \nabla^2_i - \sum_{i=1}^N \sum_{A=1}^M \frac{Z_A}{r_{iA}} + \sum_{i=2}^N \sum_{j<i} \frac{1}{r_{ij}},
\end{equation}
where $M$ is the number of nuclei, $Z_A$ is the nuclear charge, $r_{iA}$ is the electron-nucleus distance, and $r_{ij}$ is the electron-electron distance. The diagonal contributions can be expressed as,
\begin{equation}
\sum_I c_I^2 H_{II} = E^{\dno}_\text{0-1RDM} + E^{\dno}_\text{pair},
\end{equation}
where $E^{\dno}_\text{0-1RDM}$ is the energy associated with the zeroth-order term in the expansion of the two-electron density matrix (2-RDM) in terms of the one-electron density matrix (1-RDM),\cite{DNO2} which is the basis of cumulant functional theory.\cite{Kutzelnigg1999} The 0-1RDM energy is expressed as,
\begin{align}
E^{\dno}_\text{0-1RDM} & = \sum^\text{cl}_p 2n_p h_{pp} + \sum^\text{op}_p n_p h_{pp} \nonumber \\
& + \sum^\text{cl}_p \sum^\text{cl}_q n_p n_q \left( 2J_{pq} - K_{pq} \right) \nonumber \\
& + \sum^\text{cl}_p \sum^\text{op}_q n_p n_q \left( 2J_{pq} - K_{pq} \right) \nonumber \\
& + \sum^\text{op}_p \sum^\text{op}_q \frac{n_p n_q}{2} \left( J_{pq} - K_{pq} \right) ,
\end{align}
where $\text{cl}$ and $\text{op}$ denote the set of closed and open-shell orbitals, respectively. The one-electron integrals over spatial orbitals, $\phi_p$, are defined as,
\begin{equation}
h_{pp} = \int \phi^*_p(\mathbf{r}) \left[ -\frac{1}{2}\nabla^2 - \sum_{A} \frac{Z_A}{r_A} \right] \phi^*_p(\mathbf{r}) d\mathbf{r}
\end{equation}
where $\mathbf{r} = (x,y,z)$ and the two-electron Coulomb, $J_{pq}$, and exchange, $K_{pq}$, integrals are defined as,
\begin{equation}
J_{pq} = \int \frac{\phi^*_p(\mathbf{r}_1)\phi^*_q(\mathbf{r}_2)\phi_p(\mathbf{r}_1)\phi_q(\mathbf{r}_2)}{r_{12}} d\mathbf{r}_1 d\mathbf{r}_2
\end{equation}
and
\begin{equation}
K_{pq} = \int \frac{\phi^*_p(\mathbf{r}_1)\phi^*_q(\mathbf{r}_2)\phi_q(\mathbf{r}_1)\phi_p(\mathbf{r}_2)}{r_{12}} d\mathbf{r}_1 d\mathbf{r}_2 .
\end{equation}
After expanding the expectation values for the diagonal contributions to $E^{\dno'}$ [first term of Equation \eqref{eq:EDNOp}], simplifying, and removing the contribution of $E^{\dno}_\text{0-1RDM}$, the pair correction energy, $E^{\dno}_\text{pair}$, remains. The pair correction energy is the first term associated with the cumulant and is given by,
\begin{equation} \label{eq:epair}
E^{\dno}_\text{pair} = \sum_{pq} \eta_{pq} \left( 2J_{pq} - K_{pq} \right),
\end{equation}
where
\begin{align}
\eta_{pq} & = \delta_{pq} n_p (1 - n_p) \nonumber \\
& + \left(1 - \delta_{pq}\right) \Bigg[ O_p V_q \Delta_{pq} \left( n_q - n_p - \Delta_{pq} \right) \nonumber \\
& \hphantom{\left(1 - \delta_{pq}\right)\Bigg[} + V_p O_q \Delta_{qp} \left( n_p - n_q - \Delta_{qp} \right) \nonumber \\
& \hphantom{\left(1 - \delta_{pq}\right)\Bigg[} - V_p V_q \sum_r \Delta_{rp} \Delta_{rq} \Bigg] .
\end{align}
The $O_p$ and $V_p$ coefficients are elements of orbital-basis-sized vectors consisting of 1s and 0s,
\begin{equation}
O_p = \begin{cases} 1\;,\;p \in \mathcal{O} \\ 0\;,\; \text{otherwise} \end{cases}
\end{equation}
\begin{equation}
V_p = \begin{cases} 1\;,\;p \in \mathcal{V} \\ 0\;,\; \text{otherwise} \end{cases}
\end{equation}
where $\mathcal{O}$ denotes the set of active orbitals that are occupied in the reference wave function, and $\mathcal{V}$ denotes the set of active vacant orbitals.
The off-diagonal terms of the expectation value expansion are responsible for electron correlation. In the case of $\dno$, excitations are limited to low-lying near-degenerate orbitals which is aimed at capturing the so-called ``static" correlation energy,
\begin{equation}
\sum_{IJ} c_I c_J H_{IJ} = E^{\dno}_\text{stat}
\end{equation}
By employing Slater-Condon rules\cite{Szabo1996} for the evaluation of matrix elements between Slater determinants, and considering that the active spaces for the electron pairs are disjoint, the static correlation energy can be simplified to give,
\begin{equation} \label{eq:estat}
E^{\dno}_\text{stat} = \sum_{pq} \left( \zeta_{pq} - \xi_{pq} \right) L_{pq} ,
\end{equation}
where
\begin{equation}
\zeta_{pq} = V_p V_q \sum_r \sqrt{\Delta_{rp} \Delta_{rq}} ,
\end{equation}
and
\begin{equation}
\xi_{pq} = O_p V_q \sqrt{n_p \Delta_{pq}} + O_q V_p \sqrt{n_q \Delta_{qp}} .
\end{equation}
The time-inversion exchange integral is defined as
\begin{equation}
L_{pq} = \int \frac{\phi^*_p(\mathbf{r}_1)\phi^*_p(\mathbf{r}_2)\phi_q(\mathbf{r}_1)\phi_q(\mathbf{r}_2)}{r_{12}} d\mathbf{r}_1 d\mathbf{r}_2 .
\end{equation}
The $\dno$ wave function contains only $2n$-tuple excitations, in which the electrons remained paired, and each electron pair is excited to its own (restricted) active space. Therefore, $E^{\dno'}$ is an upper bound to the more general doubly-occupied configuration interaction (DOCI) energy.\cite{Bytautas2011,Bytautas2015}
\subsection{The high-spin correction}
\label{ssec:hsc}
As defined above, the $\dno$ wave function and the subsequent 2-RDM do not account for the correlation between active electron pairs (interpair correlation). The high-spin correction is added to $\dno$ to account for this correlation exactly in the strong correlation limit ({\it e.g.}~multiple bond dissociation). The term is {\it ad hoc}, and therefore does not guarantee $N$-representability below the strong correlation limit. Nevertheless, it proved effective previously\cite{DNO2} and is added to the $N$-representable $\dno$ energy, derived in Subsection \ref{ssec:wf}, to give the total $\dno$ energy,
\begin{equation}
E^{\dno} = E^{\dno'} + E^{\dno}_\text{HSC} .
\end{equation}
The high-spin correction energy is defined as
\begin{equation} \label{eq:ehsc}
E^{\dno}_\text{HSC} = -\sum_{pq} \kappa_{pq} K_{pq},
\end{equation}
where
\begin{multline} \label{eq:kappa}
\kappa_{pq} = (1-\delta_{pq}) \Bigg[ \sum_{r \ne s} \xi_{pr} \xi_{ps}\\
+ W_p \frac{n_p}{2\sqrt{2}} \sum_r \xi_{qr} + W_q \frac{n_q}{2\sqrt{2}} \sum_r \xi_{pr} \Bigg] ,
\end{multline}
where the last two terms have been added here for modelling open-shell systems. The coefficient $W_p$ indicates if the orbital is singly-occupied,
\begin{equation}
W_p = \begin{cases} 1\;,\;p \in \mathcal{S} \\ 0\;,\; \text{otherwise} \end{cases}
\end{equation}
where $\mathcal{S}$ is the set of singly-occupied orbitals. By following the same reasoning as the high-spin correction for the doubly-occupied active orbitals (first term of Equation \ref{eq:kappa}), a factor of $\frac{1}{2}$ for the last two terms would be predicted using the strong-correlation limit.\cite{DNO2} However, that results in too severe of a correction and therefore, for this first application of $\dno$ to open-shell systems, it has been reduced by a factor of $\sqrt{2}$.
\subsection{Dynamic correlation}
\label{ssec:odf}
As mentioned above, the $\dno$ wave function is designed to capture static correlation, which means dynamic correlation must be accounted for by some other means. In this study, the dynamic correlation energy is captured using an on-top density functional (ODF). The dynamic correlation energy, $E^{\dno}_\text{dyn}$, is added to the total $\dno$ energy to give the total energy including static and dynamic correlation,
\begin{equation}
E^{\dno\text{-dyn}} = E^{\dno} + E^{\dno}_\text{dyn} .
\end{equation}
Within the current formalism, it is necessary to remove extra correlation that is erroneously included, or ``double counted". Therefore, the total dynamic correlation energy is the sum of the dynamic correlation energy supplied by the ODF and a double-counting correction,
\begin{equation}
E^{\dno}_{\text{dyn}} = E_\text{ODF}[\Gamma^{\dno}(\mathbf{r},\mathbf{r})] + E_\text{DC}[\Gamma^{\dno}(\mathbf{r},\mathbf{r})] ,
\end{equation}
where both are a function of the two-electron on-top density, $\Gamma(\mathbf{r},\mathbf{r})$. The on-top density is the probability that two-electrons may be found at the same coordinate, $\mathbf{r}$. It may be calculated from the 2-RDM, which is given by
\begin{multline}
\Tilde{\Gamma}(\mathbf{x}_1,\mathbf{x}_2,\mathbf{x}_1',\mathbf{x}_2') = \frac{N(N-1)}{2} \int \Psi^*(\mathbf{x}_1',\mathbf{x}_2',\mathbf{x}_3,\dots,\mathbf{x}_N)
\\
\times \Psi(\mathbf{x}_1,\mathbf{x}_2,\mathbf{x}_3,\dots,\mathbf{x}_N) d\mathbf{x}_3 \dots d\mathbf{x}_N,
\end{multline}
where $\Psi$ is the $N$-electron wave function and $\mathbf{x} =(\mathbf{r},\omega)$ is a combination of spatial $\mathbf{r}$ and spin $\omega$ coordinates. The spinless 2-RDM may be obtained through integration over the spin-coordinates,
\begin{equation}
\Gamma(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1',\mathbf{r}_2')
= \int \left. \Tilde{\Gamma}(\mathbf{x}_1,\mathbf{x}_2,\mathbf{x}_1',\mathbf{x}_2') \right|_{\substack{\omega_1' = \omega_1\\ \omega_2' = \omega_2}} d\omega_1 d\omega_2,
\end{equation}
and the aforementioned two-electron density is the diagonal, $\Gamma(\mathbf{r}_1,\mathbf{r}_2) = \Gamma(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1,\mathbf{r}_2)$. The spinless 2-RDM can then be resolved into different spin-components corresponding to the relative spin of electrons 1 and 2,
\begin{equation}
\begin{split}
\Gamma(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1',\mathbf{r}_2')
& =
\Gamma^{\alpha\alpha}(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1',\mathbf{r}_2') + \Gamma^{\beta\beta}(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1',\mathbf{r}_2')
\\
& + \Gamma^{\alpha\beta}(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1',\mathbf{r}_2') + \Gamma^{\beta\alpha}(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1',\mathbf{r}_2'),
\end{split}
\end{equation}
where $\alpha$ and $\beta$ denote spin-up and spin-down electrons, respectively.
The 2-RDM, and its components, can also be expanded in an orbital basis,
\begin{equation} \label{eq:2RDMbasis}
\Gamma(\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_1',\mathbf{r}_2') = \sum_{pqrs} \Gamma_{pqrs} \phi^*_p(\mathbf{r}_1')\phi^*_q(\mathbf{r}_2')\phi_r(\mathbf{r}_1)\phi_s(\mathbf{r}_2).
\end{equation}
where, in this case, the $\phi_p$ are the $\dno$ orbitals. Following a procedure similar to the derivation of $E^{\dno'}$, but using second quantization, one can arrive at an expression for the elements of the spin-resolved $\dno$ 2-RDM over the $\dno$ basis,
\begin{equation}
\Gamma^{\dno,\sigma\sigma}_{pqrs} = \frac{f_{pq}^{\sigma\sigma} n_p n_q + \eta_{pq}}{2} \delta_{pr}^{qs}
\end{equation}
and
\begin{multline} \label{eq:2RDMos}
\Gamma^{\dno,\sigma\sigma'}_{pqrs} = \frac{f_{pq}^{\sigma\sigma'} n_p n_q + \eta_{pq}}{2} \delta_{pr}\delta_{qs} \\ - \frac{\kappa_{pq}}{2}\delta_{ps}\delta_{qr} + \frac{\zeta_{pr} - \xi_{pr}}{2}\delta_{pq}\delta_{rs},
\end{multline}
where $\sigma=\alpha$ or $\beta$ (with $\sigma \ne \sigma'$), $\delta_{pr}^{qs} = \delta_{pr}\delta_{qs} - \delta_{ps}\delta_{qr}$ and the $f^{\sigma\sigma'}_{pq}$ ensures there are only interactions with $\alpha$ electrons in singly-occupied orbitals,
\begin{subequations}
\begin{align}
f^{\alpha\alpha}_{pq} & = 1, \\
f^{\alpha\beta}_{pq} & = 1 - W_q, \\
f^{\beta\alpha}_{pq} & = 1 - W_p, \\
f^{\beta\beta}_{pq} & = \left( 1 - W_p \right)\left( 1 - W_q \right) .
\end{align}
\end{subequations}
Two ODFs are used in this study. The first is a combination of the Opposite-spin exponential cusp (OSEC) and Fermi-hole correction (FHC) functionals, referred to as the OF functional.\cite{OF2018} The second is the Colle-Salvetti (CS) functional.\cite{Colle1975,Colle1979} The OSEC functional depends on the total $\alpha\beta$ on-top density, whereas the FHC functional depends on the $\alpha\alpha$ and $\beta\beta$ on-top densities (more specifically, their Laplacians),
\begin{multline}
E_\text{OF}[\Gamma(\mathbf{r},\mathbf{r})] = E_\text{OSEC}[\Gamma^{\alpha\beta}(\mathbf{r},\mathbf{r}) + \Gamma^{\beta\alpha}(\mathbf{r},\mathbf{r})] \\ + E_\text{FHC}[\Gamma^{\alpha\alpha}(\mathbf{r},\mathbf{r})] + E_\text{FHC}[\Gamma^{\beta\beta}(\mathbf{r},\mathbf{r})]
\end{multline}
The FHC functional effectively widens the Fermi-hole, which exists in any 2-RDM that obeys antisymmetry, using an exponential-cusp ansatz for the second-order cusp. The functional is given by
\begin{widetext}
\begin{multline}
E_\text{FHC}[\Gamma^{\sigma\sigma}] = \int \Bigg\{ \pi\left(4\delta\left[ 3 \sqrt{\pi} \left(\delta + 16 \delta^3 \right) - 40 \delta^2 -1 \right]e^{-\frac{1}{16\delta^2}} \right.\\ \left. + \sqrt{\pi}\left[ 3\sqrt{\pi} \left(\delta + 24 \delta^3\right) - 48\left(\delta^2 + 4 \delta^4\right) - 1 \right]\left[ 1 + \text{erf}\left(\frac{1}{4\delta}\right) \right] \right) L^{\sigma\sigma}(\mathbf{r}) \Bigg\} \\
/ \Bigg\{ 3 \delta^4 \left(4\left(\delta + 40 \delta^3\right)e^{-\frac{1}{16\delta^2}} + \sqrt{\pi} \left[ 1 + 48 \left( \delta^2 + 4 \delta^4 \right) \right]\left[ 1 + \text{erf}\left(\frac{1}{4\delta}\right) \right] \right) \Bigg\} d\mathbf{r},
\end{multline}
\end{widetext}
where $L^{\sigma\sigma}(\mathbf{r})$ is the Laplacian of the two-electron density with respect to the interelectronic coordinate, $\bm{u} = \mathbf{r}_1 - \mathbf{r}_2$, at the electron-electron coalescence point ($u=0$),
\begin{equation}
L^{\sigma\sigma}(\mathbf{r}) = \nabla_{\bm{u}}^2 \left.\Gamma^{\sigma\sigma}(\mathbf{r}+\bm{u}/2,\mathbf{r}-\bm{u}/2) \right|_{u=0},
\end{equation}
where $\mathbf{r} = \frac{\mathbf{r}_1 + \mathbf{r}_2}{2}$. The $\delta$ determines the correlation length,
\begin{equation}
\delta \equiv \delta(\mathbf{r}) = \kappa L^{\sigma\sigma}(\mathbf{r})^{1/8}
\end{equation}
where $\kappa = 2.30$.\footnote{This value differs from the value determined in reference \citenum{OF2018}, $\kappa=1.73$, because it was found to overestimate the correlation energy between electrons of parallel-spin in these \ce{H} clusters.}
The OSEC functional approximates the input $\alpha\beta$ two-electron density as constant in the neighbourhood of the electron-electron coalescence point, and determines the energy associated with introducing an exponential cusp. The OSEC energy is given by the functional
\begin{widetext}
\begin{equation}
E_\text{OSEC}[\Gamma^{\alpha\beta}(\mathbf{r},\mathbf{r})] = \int \frac{2\pi\left(2\lambda \left(\sqrt{\pi} \lambda - 1 \right)e^{-\frac{1}{4\lambda^2}} + \sqrt{\pi} \left(\sqrt{\pi} \lambda - 2\lambda^2 -1 \right)\left[ 1 + \text{erf}\left(\frac{1}{2\lambda}\right) \right] \right) \Gamma^{\alpha\beta}(\mathbf{r},\mathbf{r})}{\lambda^2 \left( 2 \lambda e^{-\frac{1}{4\lambda^2}} + \sqrt{\pi}\left( 1 + 2\lambda^2 \right) \left[ 1 + \text{erf}\left(\frac{1}{2\lambda}\right) \right] \right)} d\mathbf{r},
\end{equation}
\end{widetext}
where the correlation length is determined by $\lambda$,
\begin{equation}
\lambda \equiv \lambda(\mathbf{r}) = q_\text{OSEC} \rho(\mathbf{r})^{1/3}
\end{equation}
with $q_\text{OSEC} = 2.54$ and $\rho(\mathbf{r})$ is the one-electron density.\cite{OF2018}
The Colle-Salvetti ODF depends on the total (sum of all spin components) on-top density,
\begin{multline}
E_\text{CS}\left[\Gamma\left(\bm{r},\bm{r}\right)\right] = -4 a \int \frac{\Gamma\left(\bm{r},\bm{r}\right)}{\rho\left(\bm{r}\right)}\\ \left( \frac{ 1 + b \rho\left(\bm{r}\right)^{-8/3}e^{-c\rho\left(\bm{r}\right)^{-1/3}} L(\mathbf{r}) }{1 + d \rho\left(\bm{r}\right)^{-1/3} } \right) d\bm{r}
\end{multline}
where $L(\mathbf{r})$ is the Laplacian of the total on-top density at $u=0$, $a=0.049$, $b=0.132$, $c=0.2533$ and $d=0.349$.\cite{Colle1975}
\subsection{Double-counting correction}
\label{ssec:dc}
The ODFs, OF and CS, are both derived from correcting the two-electron density in the neighbourhood of the electron-electron coalescence point, or in the short-range of interelectronic distance. Multideterminantal treatments of electron correlation, such as $\dno$, capture the full range of electron correlation. Therefore, there is overlap, or double-counting, that can occur when wave function methods are combined with density functional methods in the present manner. The main motivation for specifically using ODFs with $\dno$, is that as electrons become strongly correlated this information is revealed to the ODF via $\Gamma^{\dno}$. However, the ODF only depends on the OD, $\Gamma^{\dno}(\mathbf{r},\mathbf{r})$, which does not possess information regarding $\dno$ short-range correlation beyond the electron-electron coalescence point (in the neighbourhood of $u=0$). Because the ODF is not sensitive to the effects of correlation at small $u\ne0$, a correction must be applied.
The general scheme behind the derivation of the OF functional\cite{OF2018} (similar to that of the CS functional\cite{Colle1975}), is the approximation of the exact two-electron density $\Gamma\left(\mathbf{r}_1,\mathbf{r}_2\right)$ by the uncorrelated two-electron density $\Gamma^0\left(\mathbf{r}_1,\mathbf{r}_2\right)$ with a cusp, $c(u)$, inserted at $u=0$,
\begin{equation}
\Gamma\left(\mathbf{r}_1,\mathbf{r}_2\right) = \Gamma^0\left(\mathbf{r}_1,\mathbf{r}_2\right)\left( 1 - g(u)\left[ 1 - \Phi(\mathbf{r}) c(u) \right] \right)
\end{equation}
where $\Phi(\mathbf{r})$ is a function that maintains the normalization of $\Gamma\left(\mathbf{r}_1,\mathbf{r}_2\right)$. The correction is applied in the short-range ({\it i.e.} small $u$) and the range is controlled by the function $g(u)$,
\begin{equation}
g(u) = e^{-\lambda(\mathbf{r})u^2},
\end{equation}
which depends on the correlation length $\lambda(\mathbf{r})$. The derivation is completed by Taylor-expanding $\Gamma^0\left(\mathbf{r}_1,\mathbf{r}_2\right)$ about $u=0$. The expansion is zeroth-order for the OSEC functional,
\begin{equation}
\Gamma^0\left(\mathbf{r}_1,\mathbf{r}_2\right) \approx \Gamma^0\left(\mathbf{r},\mathbf{r}\right),
\end{equation}
and second-order for the FHC functional. The CS functional also uses a second-order expansion, but multiple approximations are made to ensure the integrand is well-behaved.
If the ODF is to exclusively account for the correlation energy in the short-range, then the short-range correlation already present in $\Gamma$ must be removed,
\begin{multline}
E^{\dno}_\text{dyn} = E^{\dno}_\text{ODF}[\Gamma] \\- \frac{1}{2}\int \frac{\left[ \Gamma(\mathbf{r}_1,\mathbf{r}_2) - \Gamma^0(\mathbf{r}_1,\mathbf{r}_2) \right] g(u)}{u} d\mathbf{r}_1 d\mathbf{r}_2 .
\end{multline}
To arrive at a practical expression for the double-counting correction, $\Gamma(\mathbf{r}_1,\mathbf{r}_2)$ and $\Gamma^0(\mathbf{r}_1,\mathbf{r}_2)$, are treated with a zeroth-order Taylor expansion about $u=0$ ({\it i.e.}~constant), identical to the OSEC functional. Also, the double-counting correction is not required in the strong correlation limit, when correlation causes electrons to completely avoid each other and $\Gamma(\mathbf{r},\mathbf{r})=0$, and must be turned off. Therefore, the correction is mutliplied by an electron-pair specific factor, $\frac{\Gamma_{pp}}{\Gamma^0_{pp}}$, and the resulting double-counting correction is given as
\begin{multline}
E_\text{DC}^{\dno}\left[\Gamma\left(\mathbf{r},\mathbf{r}\right)\right] = 2\pi c_\text{DC} \\
\int \sum_p O_p \left( \frac{\Gamma_{pp}}{\Gamma^0_{pp}} \right) \frac{\Gamma_{pp}^0\left(\mathbf{r},\mathbf{r}\right)
- \Gamma_{pp}\left(\mathbf{r},\mathbf{r}\right)}{\lambda(\mathbf{r})^2} d\mathbf{r}
\end{multline}
where $\Gamma^0$ is the uncorrelated two-electron density, which for $\dno$ is given by
\begin{equation}
\Gamma^{\dno,0} = \Gamma^{\dno}_\text{0-1RDM} + \Gamma^{\dno}_\text{pair} .
\end{equation}
The two-electron densities are decomposed into electron-pair components, which for $u=0$, are given by
\begin{equation}
\Gamma_{pp}^0\left(\bm{r},\bm{r}\right) = n_p \left| \phi_p(\mathbf{r}) \right|^4 + \sum_q \Delta_{pq} \left| \phi_q(\mathbf{r}) \right|^4,
\end{equation}
and
\begin{multline}
\Gamma_{pp}\left(\bm{r},\bm{r}\right) = \Gamma_{pp}^0\left(\bm{r},\bm{r}\right) + \sum_{qs} \Delta_{pq}\Delta_{ps} \left| \phi_q(\mathbf{r}) \right|^2 \left| \phi_s(\mathbf{r}) \right|^2\\ - 2\sum_q \xi_{pq} \left| \phi_p(\mathbf{r}) \right|^2 \left| \phi_q(\mathbf{r}) \right|^2 .
\end{multline}
The above correction is added to the OF functional energy to give the total dynamic correlation energy. In the case of the CS functional,
\begin{equation}
\lambda(\mathbf{r}) = q_\text{CS} \rho(\mathbf{r})^{1/3}
\end{equation}
where $q_\text{CS} = 2.29$ (original value from CS derivation).
\subsection{2-RDM optimization}
\label{ssec:opt}
The $\dno$ 2-RDM depends upon the $\dno$ orbitals, $\{ \phi_p \}$, and the electron transfer variables, $\{ \Delta_{ia} \}$. The optimal $\dno$ 2-RDM is found by minimizing $E^{\dno}$ with respect to both $\{\phi_p\}$ and $\{\Delta_{ia}\}$. The optimization can be performed for both $\{\phi_p\}$ and $\{\Delta_{ia}\}$ simultaneously, using the trust-region method.
In order to obtain the optimal $\dno$ orbitals, $\{\phi_p\}$, the initial set of orthonormalized orbitals, $\{ \Tilde{\phi}_r\}$, are rotated amongst themselves via a unitary transformation,
\begin{equation}
\phi_p = \sum_r U_{rp} \Tilde{\phi}_r .
\end{equation}
To determine such a transformation, the unitary matrix is parameterized with the exponential of a skew matrix,\cite{HelgakerBook} $\mathbf{Y}$,
\begin{equation}
\mathbf{U} = e^{\mathbf{Y}} ,
\end{equation}
where
\begin{equation}
Y_{pq} = y_{pq}, \quad Y_{qp} = -y_{pq} .
\end{equation}
Any energy, $E[\mathbf{y}]$, that is a functional of a set of orbitals, may be expanded about the optimal orbitals, in a Taylor series, in terms of the skew matrix parameters, $\mathbf{y}$,
\begin{equation}
E[\mathbf{y}] \approx E[\mathbf{0}] + \mathbf{y} \cdot \mathbf{g}^\phi + \frac{1}{2} \mathbf{y}^\dag \mathbf{H}^{\phi\phi} \mathbf{y} ,
\end{equation}
and approximated to second-order.\cite{Bozkaya2011, Bozkaya2013, Bozkaya2014, Bozkaya2016} The gradient, $\mathbf{g}^\phi$, and hessian, $\mathbf{H}^\phi$, consist of the first and second derivatives of the energy with respect to the skew-matrix parameters,
\begin{equation}
g^\phi_{pq} = \left. \frac{\partial E^{\dno}}{\partial y_{pq}} \right|_{\mathbf{y}=0} , \quad H^{\phi\phi}_{pq,rs} = \left. \frac{\partial^2 E^{\dno}}{\partial y_{pq} \partial y_{rs}} \right|_{\mathbf{y}=0} .
\end{equation}
General expressions for the above derivatives may be found elsewhere,\cite{Bozkaya2011} but for $\dno$ specifically see Appendix \ref{app:derivs}. In the case of $E^{\dno}$, the energy is invariant ($g^\phi_{pq} =0$) with respect to rotations between orbital pairs where both belong to the inactive (frozen) closed-shell, open-shell, or inactive vacant orbital subspaces. Consequently, optimization must be performed with respect to all other possible rotations.
In the case of $\dno$, the energy is also a function of the electron transfer variables, $\{ \Delta_{ia} \}$. To simultaneously optimize the energy with respect to both $\{ \phi_p \}$ and $\{ \Delta_{ia} \}$, the orbital gradient and hessian are combined with those for $\{ \Delta_{ia} \}$. However, in order to avoid violation of $N$-representability of the 1-RDM, the $\{ \Delta_{ia} \}$ are first parameterized such that they remain between 0 and 1/2,
\begin{equation} \label{eq:theta}
\Delta_{ia} = \frac{\cos^2 \theta_{ia}}{2} .
\end{equation}
The total gradient and hessian are then given by
\begin{equation}
\mathbf{g} = \begin{bmatrix} \mathbf{g}^\phi \\ \mathbf{g}^\theta \end{bmatrix} , \quad \mathbf{H} = \begin{bmatrix} \mathbf{H}^{\phi\phi} & \mathbf{H}^{\phi\theta} \\ \mathbf{H}^{\phi\theta} & \mathbf{H}^{\theta\theta} \end{bmatrix},
\end{equation}
where
\begin{equation}
g^\theta_{pq} = \frac{\partial E^{\dno}}{\partial \theta_{pq}} , \quad H^{\theta\theta}_{pq,rs} = \frac{\partial^2 E^{\dno}}{\partial \theta_{pq} \partial \theta_{rs}},
\end{equation}
and
\begin{equation}
H^{\phi\theta}_{pq,rs} = \frac{\partial^2 E^{\dno}}{\partial y_{pq} \partial \theta_{rs}} .
\end{equation}
Minimization can be performed using the trust-region Newton method, in which an augmented hessian is prepared and a step is taken within a dynamic trust-region radius. A standard trust-region approach is described in Nocedal and Wright (Algorithm 4.1).\cite{Nocedal2006} The procedure begins with the calculation of an augmented hessian, $\mathbf{A}$,
\begin{equation}
\mathbf{A} = \mathbf{H} + \epsilon \mathbf{I}
\end{equation}
where $\mathbf{I}$ is the identity matrix and $\epsilon$ is chosen such that $\mathbf{A}$ is positive definite. Then a Newton step, $\mathbf{p}$, is calculated by solving
\begin{equation}
\mathbf{A} \mathbf{p} = -\mathbf{g} .
\end{equation}
If the step is within the trust-region radius, $p = \sqrt{\mathbf{p}\cdot\mathbf{p}} < t$, then the step is taken, otherwise a subproblem must be solved (Equation 4.5 in reference \citenum{Nocedal2006}) in which a step within the trust-region radius is found iteratively. Each step, whether pure Newton-Raphson or a solution to the subproblem, is evaluated by comparing the actually change in the function ($E$) to the change predicted by a second-order Taylor expansion. If the reduction is less than a minimum threshold ({\it e.g.}~25\% of predicted) then no step is taken and $t$ is reduced. If the reduction is more than the minimum, a step is taken, and if it is more than a maximum threshold ({\it e.g.}~75\% of predicted) then $t$ is increased.
In order to avoid optimizations to saddle points in $\{ \phi_p \}$ and $\{ \Delta_{ia} \}$ space, a modification can be made before commencing the trust-region algorithm. The non-augmented hessian is diagonalized to evaluate the eigenvalues,
\begin{equation}
\mathbf{Z}^T \mathbf{H} \mathbf{Z} = \mathbf{H}'
\end{equation}
where $\mathbf{H}'$ is a diagonal matrix of hessian eigenvalues and $\mathbf{Z}$ is the matrix that transforms to normal coordinates. For any normal coordinate (mode) with a negative eigenvalue, $H_{ii}'<0$, and a gradient component with a magnitude less than a threshold, $\left|g_i'\right| < g_0$, the gradient is amplified to promote escape from the saddle point,
\begin{equation}
\mathbf{g}_\text{mod} = \mathbf{Z} \mathbf{g}_\text{mod}'
\end{equation}
where
\begin{equation}
g_{\text{mod},i}' = \begin{cases} \text{sign}(g_i')g_0, &\text{ If } H_{ii}'< 0 \text{ and } \left| g_i' \right| < g_0 \\ g_i', &\text{ otherwise} \end{cases}
\end{equation}
The trust-region method then continues as usual.
When an optimization step is calculated, the unitary matrix for the orbital rotation is prepared by first approximating $e^{\mathbf{Y}}$ to second order,
\begin{equation}
\mathbf{U} \approx \mathbf{I} + \mathbf{Y} + \frac{1}{2} \mathbf{Y} \mathbf{Y}^{\dag} ,
\end{equation}
and then ensuring it is unitary by applying Gram-Schmidt orthonormalization.
\subsection{Conditions of $N$-representability}
\label{ssec:Nrep}
Due to the addition of the high-spin correction, the $\dno$ 2-RDM is not strictly $N$-representable. Therefore, it is useful to have some means of evaluating the extent to which $N$-representability might be violated. The necessary, but not sufficient, $N$-representability conditions that depend only on the 2-RDM are referred to as the $PQG$ (or $DQG$) conditions.\cite{Garrod1964,Mazziotti2006, Mazziotti2012} They require that the following matrices, in a spin-orbital basis and defined in terms of the second quantization, be positive semi-definite ({\it i.e.}~all non-negative eigenvalues)
\begin{subequations}
\begin{align}
P_{ijkl} & = \langle \Psi | a_j^\dagger a_i^\dagger a_k a_l | \Psi \rangle, \\
Q_{ijkl} & = \langle \Psi | a_j a_i a_k^\dagger a_l^\dagger | \Psi \rangle, \\
G_{ijkl} & = \langle \Psi | a_j^\dagger a_i a_k^\dagger a_l | \Psi \rangle,
\end{align}
\end{subequations}
where $\Psi$ is the $N$-electron wave function, $a_i^\dagger$ and $a_i$ are creation and annihilation operators, respectively. Sufficient conditions for $N$-representability require the calculation of higher-order matrices and are not considered here.\cite{Mazziotti2012,Li2021} For a spin-resolved 2-RDM over spatial orbitals [\eqref{eq:2RDMps} and \eqref{eq:2RDMos}], and a diagonal 1-RDM, the elements of $\bm{P}$, $\bm{Q}$, and $\bm{G}$ can be expressed as
\begin{subequations}
\begin{align}
P^{\sigma\sigma'\sigma\sigma'}_{pqrs} = &\; \Gamma^{\sigma\sigma'}_{pqrs}, \\
Q^{\sigma\sigma'\sigma\sigma'}_{pqrs} = &\; \Gamma^{\sigma\sigma'}_{pqrs} + \frac{\delta_{pr}\delta_{qs}}{2} \left( 1 - n_p - n_q \right) \nonumber \\
& - \delta_{\sigma\sigma'} \frac{\delta_{ps}\delta_{qr}}{2} \left( 1 - n_p - n_q \right), \\
G^{\sigma\sigma'\sigma\sigma'}_{pqrs} = &\; \frac{\delta_{pr}\delta_{qs}}{2}n_q - \Gamma^{\sigma\sigma'}_{rqps}, \\
G^{\sigma\sigma\sigma'\sigma'}_{pqrs} = &\; \Gamma^{\sigma\sigma'}_{qrps} \quad (\sigma \ne \sigma'),
\end{align}
\end{subequations}
where $\sigma = \alpha$ or $\beta$, and $\sigma$ can be equivalent to $\sigma'$ (except for $G^{\sigma\sigma\sigma'\sigma'}_{pqrs}$). The expression of $\bm{P}$, $\bm{Q}$, and $\bm{G}$ in terms of spatial-orbital indices (subscript) requires specifying the spin-function (superscript) used to form the corresponding spin orbital. The matrices can be decomposed into spin-blocks for construction and analysis, which is evident from their structure,\cite{Mazziotti2002}
\begin{subequations}
\begin{align}
\bm{P} & = \begin{bmatrix} \bm{P}^{\alpha\alpha\alpha\alpha} & 0 & 0 & 0 \\
0 & \bm{P}^{\beta\beta\beta\beta} & 0 & 0 \\
0 & 0 & \bm{P}^{\alpha\beta\alpha\beta} & 0 \\
0 & 0 & 0 & \bm{P}^{\beta\alpha\beta\alpha} \end{bmatrix}, \\
\bm{Q} & = \begin{bmatrix} \bm{Q}^{\alpha\alpha\alpha\alpha} & 0 & 0 & 0 \\
0 & \bm{Q}^{\beta\beta\beta\beta} & 0 & 0 \\
0 & 0 & \bm{Q}^{\alpha\beta\alpha\beta} & 0 \\
0 & 0 & 0 & \bm{Q}^{\beta\alpha\beta\alpha} \end{bmatrix}, \\
\label{eq:Gmat}
\bm{G} & = \begin{bmatrix} \bm{G}^{\alpha\alpha\alpha\alpha} & \bm{G}^{\alpha\alpha\beta\beta} & 0 & 0 \\
\bm{G}^{\beta\beta\alpha\alpha} & \bm{G}^{\beta\beta\beta\beta} & 0 & 0 \\
0 & 0 & \bm{G}^{\alpha\beta\alpha\beta} & 0 \\
0 & 0 & 0 & \bm{G}^{\beta\alpha\beta\alpha} \end{bmatrix}.
\end{align}
\end{subequations}
The sum of the negative eigenvalues of these matrices can be used to quantify the extent to which they are not positive semi-definite and subsequently violate the $PQG$-conditions.\cite{RodriguezMayorga2017} These sums are denoted by
\begin{equation} \label{eq:Xneg}
X_\text{neg} = \sum_{\substack{i = 1\\ x_i < 0}}^{K^2} x_i,
\end{equation}
where $K$ is the size of the orbital basis, $\bm{X} = \bm{P}, \bm{Q}$ or $\bm{G}$ and $x_i$ is an eigenvalue of $\bm{X}$.
\section{Method}
\label{sec:method}
\subsection{Reference energies and properties}
The energies from single-reference methods: completely renormalized coupled-cluster (CR-CCL)\cite{Piecuch2002, Piecuch2005} and spin-unrestricted density functional approximation (B3LYP),\cite{Becke1993, Stephens1994, Hertwig1997} and multireference energies: multireference self-consistent field (MRSCF)\cite{Chaban1997} and multireference M{\"o}ller-Plesset second-order perturbation theory (MRMP2),\cite{Hirao1992a, Hirao1992b, Hirao1992c, Hirao1993} were calculated using GAMESS.\cite{GAMESS} The active space for the MRSCF and MRMP2 calculations was chosen to be all occupied orbitals of the reference configuration, and an equivalent number of the lowest lying vacant (virtual) orbitals. Full configuration interaction (FCI) energies, and 2-RDMs, were determined using Quantum Package 2.\cite{qp2} All calculations were performed using the cc-pVTZ basis set.\cite{BSE2019, Dunning1989}
\subsection{$\dno$ energies and properties}
The $\dno$, $\dnoof$, and $\dnocs$ calculations were performed with MUNgauss.\cite{mungauss} The active space for each electron pair consisted of the orbital they occupy in the reference configuration and one vacant (virtual) orbital. The $\dno$ 2-RDMs were optimized using the the trust-region method outlined in Subsection \ref{ssec:opt}. The iterative subproblem was solved using the dgqt routine from MINPACK-2,\cite{More1983} translated to Fortran 90. Restricted Hartree-Fock orbitals are used as the intial guess, and the intial trust-region radius was $t = 0.05$ with a maximum radius value of 2.00. The gradient threshold, required for saddle-point escape, was $g_0 = 10^{-4}$. The convergence criteria was $\bar{g} < 10^{-6}$ where $\bar{g} = \frac{|\mathbf{g}|}{n}$ is the normalized total gradient length, where $n$ is the total number of orbital optimization parameters, $\{ y_{pq} \}$ and $\{ \Delta_{ia} \}$.
The OF and CS energies are calculated post-$\dno$ using an SG-1 grid\cite{Gill1993} and Becke atomic weights.\cite{Becke1988}
The value of the double-counting correction parameter, $c_\text{DC}$, was chosen to minimize the error in the $\dnoof$, and $\dnocs$, \ce{H2} potential energy curves, relative to FCI. The values were 0.40 and 0.35 for $\dnoof$ and $\dnocs$, respectively.
\section{Results}
\label{sec:results}
\subsection{Cluster geometries}
The geometries of the \ce{H} clusters studied are presented in Figure \ref{fig:clusters}.
\begin{figure}
\begin{center}
\includegraphics[width=0.42\textwidth]{fig/clusters}
\caption{\label{fig:clusters} Geometries of \ce{H} clusters}
\end{center}
\end{figure}
The potential energy surfaces were calculated with established single-reference (FCI, CR-CCL and B3LYP) and multireference methods (MRSCF and MRMP2) as well as with $\dno$, $\dnoof$ and $\dnocs$. The potential energy surface is defined as,
\begin{equation}
U(\mathbf{X}) = E(\mathbf{X}) + V_{nn}(\mathbf{X})
\end{equation}
where $E(\mathbf{X})$ is the electronic energy provided by the electronic structure method, $V_{nn}(\mathbf{X})$, is the nuclear-repulsion energy, and both are a function of the nuclear coordinates, $\mathbf{X}$.
\subsection{2-RDM optimization}
The most time consuming part of the $\dno$ 2-RDM optimization is the transformation of the two-electron integrals from the atomic orbital basis to the $\dno$ orbital basis, which scales as $N_\text{act} K^4$, where $N_\text{act}$ is the number of active orbitals and $K$ is the number of atomic basis functions. Recently, this cost has been reduced for similar calculations in natural orbital functional methods by introducing the resolution-of-the-identity (DoNOF-RI).\cite{LewYee2021}
\begin{table}
\caption{\label{tab:2RDMopt} Comparison of the number of integral transformations required to optimize the $\dno$ 2-RDM}
\begin{ruledtabular}
\begin{tabular}{lccc}
& & \multicolumn{2}{c}{integral transformations} \\
\cline{3-4}
cluster & $R$ (\AA) & ID\footnote{ID = iterative diagonalization of pseudo-Fock matrix} & TR\footnote{TR = trust-region method} \\
\hline
\ce{H2} & 1.2 & 170 & 10 \\
\ce{H3} & 1.2 & 924 & 13 \\
\ce{H4} linear & 1.2 & 252 & 47 \\
\ce{H4} square & 1.2 & 320 & 20 \\
\end{tabular}
\end{ruledtabular}
\end{table}
For the systems studied, the use of the trust-region method, on average, reduced the number of optimization steps (and therefore the number of integral transformations) by an order of magnitude compared to the previous iterative diagonalization algorithm\cite{Piris2009,DNO1,DNO2} (Table \ref{tab:2RDMopt}). Furthermore, the introduction of an analytical hessian, and a saddle-point escape procedure, enabled the trust-region method to find global minima that the previous iterative-diagonalization algorithm could not locate.
In some cases, in order to determine whether optimized 2-RDMs were indeed global minima, multiple different initial guesses were used. This includes using orbitals different from RHF, including MRSCF orbitals and FCI natural orbitals, choosing different vacant orbitals for the active space, and also using 2-RDMs (both the orbitals and the electron-transfer variables) from other cluster geometries. The minima found by the 2-RDM optimization algorithm were invariant with respect to the choice of initial orbitals or which vacant orbitals to include in the active space. Local minima were only located when an alternative 2-RDM guess was supplied from another cluster geometry ({\it e.g.}~2-RDM from \ce{H4} at $\theta=90^\circ$ for \ce{H4} at $\theta=85^\circ$).
\subsection{\ce{H2}}
The error in the \ce{H2} potential energy curves relative to the FCI result is presented in Figure \ref{fig:H2error}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/H2error}
\caption{\label{fig:H2error} Error in \ce{H2} potential energy curves compared to FCI ($\Delta U = U - U_\text{FCI}$).}
\end{center}
\end{figure}
To illustrate the importance of the double-counting correction, the $\dnoof$ and $\dnocs$ result without including the double-counting correction ($c_\text{DC} = 0$) is shown. Without the double-counting correction, both $\dnoof$ and $\dnocs$ predict $\sim$ 30 kJ mol$^{-1}$ too much correlation energy near the equilibirum bond length ($R = 0.743$ \AA). However, as the amount a static correlation grows with increasing $R$, the error approaches zero. The static correlation included in the $\dno$ 2-RDM causes the on-top density to vanish, and therefore there is no superfluous correlation energy at dissociation. If the double-counting correction is included, the error in $\dnoof$ and $\dnocs$ is dramatically reduced with a maximum error of $\sim$ 3 kJ mol$^{-1}$ (a positive deviation near equilibrium, and negative at stretched bond lengths). The error in the B3LYP potential energy curve is also included, where the most notable deviation is the energy at the dissociation limit. The dissociation limit predicted by B3LYP is 5.6 kJ mol$^{-1}$ higher than FCI.
The equilibrium bond lengths and dissociation energies predicted by each method for the various \ce{H} clusters are presented in Tables \ref{tab:Re} and \ref{tab:DE}. In the case of \ce{H2}, $\dnoof$ and $\dnocs$ predict equilibrium bond lengths within 0.004 {\AA} of the FCI value, 0.0743 \AA, while the other methods are slightly more accurate. Also for \ce{H2}, $\dno$ and MRSCF are identical and therefore predict that same bond length (0.756 \AA) and dissociation energy (399 kJ mol$^{-1}$).
\begin{table}
\caption{\label{tab:Re} Equilibrium bond lengths (in \AA) of \ce{H} clusters}
\begin{ruledtabular}
\begin{tabular}{lcccc}
& \multicolumn{4}{c}{cluster} \\
\cline{2-5}
method & \ce{H2} & \ce{H3} &\ce{H4} linear&\ce{H4} square \\
\hline
$\dno$ & 0.756 & 0.942 & 0.908 & 0.910 \\
$\dnoof$ & 0.739 & 0.924 & 0.882 & 0.863 \\
$\dnocs$ & 0.739 & 0.918 & 0.879 & 0.857 \\
B3LYP & 0.744 & 0.931 & 0.885 & 0.841 \\
CR-CCL & 0.743 & 0.930 & 0.887 & 0.878 \\
MRSCF & 0.756 & 0.957 & 0.906 & 0.912 \\
MRMP2 & 0.745 & 0.934 & 0.889 & 0.868 \\
FCI & 0.743 & 0.930 & 0.887 & 0.864
\end{tabular}
\end{ruledtabular}
\end{table}
Other than $\dno$ and MRSCF, all methods predict fairly accurate bond dissociation energies for \ce{H2}. The high accuracy of $\dnoof$ (0 kJ mol$^{-1}$ error) and $\dnocs$ (1 kJ mol$^{-1}$ error) is due to the calibration of the double-counting correction coefficient, $c_\text{DC}$. The strategy is to use the bond of \ce{H2} as a prototype for the static and dynamic correlation energy balance between paired electrons.
\begin{table}
\caption{\label{tab:DE} Dissociation energies (in kJ mol$^{-1}$) of \ce{H} clusters}
\begin{ruledtabular}
\begin{tabular}{lcccc}
& \multicolumn{4}{c}{cluster} \\
\cline{2-5}
method & \ce{H2} & \ce{H3} &\ce{H4} linear&\ce{H4} square \\
\hline
$\dno$ & 399 & 317 & 588 & 168 \\
$\dnoof$ & 454 & 411 & 722 & 288 \\
$\dnocs$ & 455 & 413 & 718 & 288 \\
B3LYP & 461 & 444 & 765 & 349 \\
CR-CCL & 454 & 412 & 726 & 292 \\
MRSCF & 399 & 327 & 610 & 202 \\
MRMP2 & 444 & 400 & 713 & 278 \\
FCI & 454 & 412 & 733 & 288
\end{tabular}
\end{ruledtabular}
\end{table}
\subsection{\ce{H3}}
The potential energy curves (PECs) for \ce{H3}, in a linear arrangement, are presented in Figure \ref{fig:H3}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/H3}
\caption{\label{fig:H3} Potential energy of a linear \ce{H3} cluster with neighbouring \ce{H} atoms separated by a distance $R$.}
\end{center}
\end{figure}
The $\dno$ and MRSCF PECs are above the PECs of the other methods designed to also capture dynamic correlation. Unlike \ce{H2}, the $\dno$ and MRSCF PECs differ and it is apparent that, even with the high-spin correction applied to the unpaired electron (Equation \ref{eq:ehsc}), $\dno$ is missing correlation at moderately stretched bond lengths. Nevertheless, the correct dissociation limit is reached by $\dno$, $\dnoof$ and $\dnocs$.
The error in the PECs, for the methods that include both static and dynamic correlation ($\dnoof$, $\dnocs$, MRMP2 and B3LYP), compared to the FCI PEC is shown in Figure \ref{fig:H3error}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/H3error}
\caption{\label{fig:H3error} Error in \ce{H3} potential energy curves compared to FCI ($\Delta U = U - U_\text{FCI}$).}
\end{center}
\end{figure}
At small $R$, $\dnoof$ and $\dnocs$ overestimate the correlation energy by a maximum of 4 and 11 kJ mol$^{-1}$, respectively. At moderately stretched bond lengths, both $\dnoof$ and $\dnocs$ underestimate the correlation energy (23 and 29 kJ mol$^{-1}$, respectively), which is due to the missing static correlation in $\dno$ between the electron pair and the unpaired electron.
The largest deviation of the MRMP2 PEC from the FCI PEC is at small $R$ (16 kJ mol$^{-1}$), and the error decreases monotonically as $R$ increases. In the case of B3LYP, the correlation energy is significantly overestimated around equilibrium $R$ (24 kJ mol$^{-1}$) and like \ce{H2}, the energy at the dissociation limit is incorrect. These errors are reflected in the predicted equilibrium $R$ values and dissociation energies (Tables \ref{tab:Re} and \ref{tab:DE}).
Similar to \ce{H2}, both $\dno$ and MRSCF overestimate the equilibrium separation. Whereas, $\dnoof$ and $\dnocs$ slightly underestimate $R_e$, by 0.006 {\AA} and 0.012 \AA, respectively. The values of $R_e$ from both MRMP2 (0.934 \AA) and B3LYP (0.931 \AA) differ only slightly from the FCI value. Also, like \ce{H2}, the CR-CCL PEC is equivalent to FCI.
For those methods designed to capture all the correlation energy, it is only the dissociation energy predicted by B3LYP (444 kJ mol$^{-1}$) that differs substantially from the FCI value (412 kJ mol$^{-1}$).
\subsection{Linear \ce{H4}}
The PECs for the linear \ce{H4} cluster, with interatomic spacing $R$, are presented in Figure \ref{fig:H4lin}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/H4lin}
\caption{\label{fig:H4lin} Potential energy of a linear \ce{H4} cluster with neighbouring \ce{H} atoms separated by a distance $R$.}
\end{center}
\end{figure}
Similar to \ce{H3}, the $\dno$ and MRSCF PECs appear above the others (as expected), and the B3LYP PEC lies below the others at equilibrium and above at dissociation. Interestingly, while the CR-CCL PEC is essentially exact close to $R_e$, and eventually approaches the correct dissociation limit, there is a significant negative deviation from the FCI PEC at large $R$.
The error in the linear \ce{H4} PECs, compared to FCI, are explicitly shown in Figure \ref{fig:H4linerror}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/H4linerror}
\caption{\label{fig:H4linerror} Error in linear \ce{H4} cluster potential energy curves compared to FCI ($\Delta U = U - U_\text{FCI}$).}
\end{center}
\end{figure}
The behaviour of the error in the $\dnoof$ and $\dnocs$ PECs for linear \ce{H4} is the same as that seen for \ce{H2}, although substantially larger. Both $\dnoof$ and $\dnocs$ underestimate the correlation energy at small $R$, and overestimate at large $R$, while eventually reaching the correct dissociation limit. Like \ce{H3}, the error in the MRMP2 PEC is largest at small $R$ and decreases as $R$ is increased. In the case of CR-CCL, the PEC is essentially exact for small $R$ and then begins overestimating the correlation energy at $\sim$ 1.4 \AA. The error in the B3LYP PEC is similar that seen for both \ce{H2} and \ce{H3}.
With the exception of $\dno$ and MRSCF, all methods provide good estimates of $R_e$ compared to FCI, with $\dnoof$ and $\dnocs$ showing the most deviation (0.005 and 0.008 \AA, respectively). Both $\dno$ and MRSCF predict larger $R_e$ values, and are within close agreement with each other (differing by 0.002 \AA).
The dissociation energies predicted by each method are more varied. The dissociation energies predicted by $\dnoof$ and $\dnocs$, 722 kJ mol$^{-1}$ and 718 kJ mol$^{-1}$, respectively, are in decent agreement with the FCI value, 733 kJ mol$^{-1}$. Only CR-CCL provides a better estimate, 726 kJ mol$^{-1}$.
\subsection{Square \ce{H4}}
The PECs for the dissociation of \ce{H4} in a square arrangement ($\theta = 90^\circ$, see Figure \ref{fig:clusters}) are shown in Figure \ref{fig:H4sq}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/H4sq}
\caption{\label{fig:H4sq} Potential energy of a square \ce{H4} cluster where each at is a distance $R$ from the centre (see Figure \ref{fig:clusters}).}
\end{center}
\end{figure}
The square arrangement of \ce{H} atoms results in much more varied performance amongst the methods assessed compared to the linear arrangement. The B3LYP PEC is far below the others near $R_e$, but above again at the dissociation limit. There is substantial separation between the $\dno$ and MRSCF PECs near $R_e$. Furthermore, this system is a known issue for single-reference coupled-cluster methods, and while CR-CCL performs better here than CCSD(T), the error is significant. The explicit errors are shown in Figure \ref{fig:H4sqerror}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/H4sqerror}
\caption{\label{fig:H4sqerror} Error in square \ce{H4} cluster potential energy curves compared to FCI ($\Delta U = U - U_\text{FCI}$).}
\end{center}
\end{figure}
The error in the $\dnocs$ PEC is relatively small, and provides, arguably, the best agreement with the FCI PEC. The usual underestimation of the correlation energy at small $R$ by MRMP2 is relatively small for this system, and therefore MRMP2 also provides a good estimate of the FCI PES. The negative deviation of $\dnoof$ at moderate $R$ values reaches a maximum of $\sim$ 14 kJ mol$^{-1}$ before going to the correct dissociation limit. Whereas, the CR-CCL PES overestimates the correlation energy by more than 30 kJ mol$^{-1}$ before approaching the dissociation limit.
Despite the errors in the PECs, accurate predictions of the FCI $R_e$, 0.864 \AA, are provided by $\dnoof$, $\dnocs$, MRMP2 and CR-CCL, with values of 0.863, 0.857, 0.868 and 0.878, respectively. As usual, the dissociation energies vary more widely. Both $\dnoof$ and $\dnocs$ provide very accurate estimates of the dissociation energy, and MRMP2 and CR-CCL differ by only 10 and 4 kJ mol$^{-1}$, respectively. While B3LYP overestimates the dissociation energy by 61 kJ mol$^{-1}$.
\subsection{Rectangle to square \ce{H4}}
\label{ssec:rec2sq}
At fixed $R$, if $\theta$ of the, $D_{4h}$ symmetry, square arrangement of \ce{H4} (Figure \ref{fig:clusters}) is changed from 90$^\circ$, the result is a rectangular, $D_{2h}$ symmetry, arrangement. The PECs for $\theta=70^\circ$ to $\theta=110^\circ$ ($D_{2h}$ $\to$ $D_{4h}$ $\to$ $D_{2h}$) for fixed $R$ values of 0.8, 1.2 and 1.7 {\AA} are presented in Figure \ref{fig:H4theta}.
\begin{figure*}
\begin{center}
\includegraphics[width=0.315\textwidth]{fig/H4thetaRpt8}
\includegraphics[width=0.315\textwidth]{fig/H4thetaR1pt2}
\includegraphics[width=0.315\textwidth]{fig/H4thetaR1pt7}
\caption{\label{fig:H4theta} Potential energy of \ce{H4} cluster as a function of $\theta$ and $R$ (see Figure \ref{fig:clusters}).}
\end{center}
\end{figure*}
Like the PECs for the other \ce{H} clusters, the $\dno$ and MRSCF PECs lie above those of the methods designed to capture all of the correlation energy. As $R$ increases, and static correlation increases while dynamic correlation decreases, the $\dno$ and MRSCF PECs approach the others (note the scale of the $y$-axis in Figure \ref{fig:H4theta}).
In the case of CR-CCL, coupled-cluster methods are known to have a concave (downward pointing) derivative discontinuity,\cite{Robinson2012c, RamosCordoba2015, Marie2021} or cusp, at $\theta=90$, which becomes more prominent as $R$ increases. This is in contrast to the convex (upward pointing) cusp seen for RHF and spin-restricted density functional methods. The cusp in the CR-CCL PEC is barely noticable at $R=0.8$ \AA, but is evident at the larger $R$ values. The MRMP2 PEC closely resembles the FCI PEC at all $R$ values, and the absolute error decreases with increasing $R$. The height of the PEC is significantly underestimated by B3LYP at all $R$ values.
The $\dnoof$ and $\dnocs$ methods accurately model the PEC near $\theta=70^\circ$ and 90$^\circ$, however there is an obvious cusp at $\theta \sim 85^\circ$ and $\theta \sim 95^\circ$. This is due to an inadequate description of the 2-RDM upon transitioning from a $D_{4h}$ to $D_{2h}$ arrangement, which is analysed, and elaborated upon, later.
The barrier height, defined as the difference in energy at $\theta= 90^\circ$ and 70$^\circ$, predicted by each method at each $R$ value is presented in Table \ref{tab:BH}.
\begin{table}
\caption{\label{tab:BH} Barrier height\footnote{Barrier height is defined as $E[90^\circ] - E[70^\circ]$} (in kJ mol$^{-1}$) for $D_{2h}\to D_{4h} \to D_{2h}$ transition of \ce{H4} cluster at varying $R$.}
\begin{ruledtabular}
\begin{tabular}{lccc}
& \multicolumn{3}{c}{$R$ (\AA)} \\
\cline{2-4}
method & 0.8 & 1.2 & 1.7 \\
\hline
$\dno$ & 305 & 212 & 76 \\
$\dnoof$ & 287 & 200 & 76 \\
$\dnocs$ & 284 & 200 & 78 \\
B3LYP & 238 & 157 & 50 \\
CR-CCL & 292 & 186 & 51 \\
MRSCF & 281 & 190 & 72 \\
MRMP2 & 286 & 204 & 80 \\
FCI & 293 & 208 & 81
\end{tabular}
\end{ruledtabular}
\end{table}
At $R=0.8$ \AA, $\dnoof$ and $\dnocs$ underestimate the barrier by 6 and 9 kJ mol$^{-1}$, respectively. Compare that to $\dno$, which overestimates the barrier by 12 kJ mol$^{-1}$. MRMP2 provides accuracy similar to $\dnoof$ and $\dnocs$, while CR-CCL only differs from the FCI barrier by 1 kJ mol$^{-1}$. As $R$ is increased, the agreement of $\dno$ with FCI improves, which highlights the importance of static correlation to these PECs. At $R=1.2$, $\dnoof$ and $\dnocs$ both differ from FCI by 8 kJ mol$^{-1}$, and differ by 5 and 3 kJ mol$^{-1}$, respectively, at $R=1.7$ \AA. At the same time, the agreement of MRMP2 with FCI also improves, to within 4 kJ mol$^{-1}$ at $R=1.2$ {\AA} and within 1 kJ mol$^{-1}$ at $R=1.7$ \AA. In contrast, due to the cusp, the CR-CCL values worsen.
Although the barrier heights predicted by $\dnoof$ and $\dnocs$ are in good agreement with FCI values, the deviation of the PECs from FCI approximately 5$^\circ$ from $\theta = 90^\circ$ is concerning. An initial analysis of the correlation captured by $\dno$ is provided by the electron transfer variables, $\{ \Delta_{ia} \}$. The values of $\Delta_{14}$ and $\Delta_{23}$ from $\theta=70^\circ$ to $\theta=110^\circ$, at $R = 1.7$ \AA, are presented in Figure \ref{fig:Deltas}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/DeltaR1pt7}
\caption{\label{fig:Deltas} $\Delta$ values for rectangle-to-square \ce{H4} cluster as a function of $\theta$ at $R= 1.7$ \AA.}
\end{center}
\end{figure}
From $\theta=70^\circ$ to $\theta=77^\circ$, $\Delta_{14}$ and $\Delta_{23}$ are equivalent, indicating that the motion of both electron pairs is correlated in the same manner. This points to the rectangular cluster behaving as two \ce{H2} molecules, which is not surprising at particularly large separations like $R=1.7$ \AA. Beyond $\theta=77^\circ$, the values of $\Delta_{14}$ and $\Delta_{23}$ diverge, and there is an obvious transition that occurs from $\theta = 84^\circ$ to $\theta = 87^\circ$, and from $\theta = 93^\circ$ to $\theta = 96^\circ$ (grey rectangles). At the $\theta=90^\circ$ side of the transition, the electron pairs have significantly different amounts of static correlation. The $\Delta_{23}$ electron pair is very near the strong correlation limit, $\Delta \to 1/2$, while $\Delta_{14} \approx 0.28$. Essentially, there are two regimes of correlation. For $\theta$ near 70$^\circ$ (and 110$^\circ$) there is the \ce{2H2} regime, and for $\theta$ near 90$^\circ$ there is the \ce{H4} regime.
The presence of two different regimes, or models, for the correlation in \ce{H4} by $\dno$, and their transition, is supported further by an analysis of the $\dno$ orbitals (Figure \ref{fig:H4orb}).
\begin{figure}
\begin{center}
\begin{tabular}{c|c|c|c|c|}
\cline{2-5}
& \multicolumn{4}{c|}{ $\theta$ (degrees) } \\
\hline
\multicolumn{1}{|c|}{orbital} & 84 & 85 & 86 & 87 \\
\hline \hline
\multicolumn{1}{|c|}{\raisebox{0.035\textwidth}{1}}
& \includegraphics[width=0.08\textwidth]{fig/NO1_84}
& \includegraphics[width=0.08\textwidth]{fig/NO1_85}
& \includegraphics[width=0.08\textwidth]{fig/NO1_86}
& \includegraphics[width=0.08\textwidth]{fig/NO1_87} \\ \hline\hline
\multicolumn{1}{|c|}{\raisebox{0.035\textwidth}{2}}
& \includegraphics[width=0.08\textwidth]{fig/NO2_84}
& \includegraphics[width=0.08\textwidth]{fig/NO2_85}
& \includegraphics[width=0.08\textwidth]{fig/NO2_86}
& \includegraphics[width=0.08\textwidth]{fig/NO2_87} \\ \hline\hline
\multicolumn{1}{|c|}{\raisebox{0.035\textwidth}{3}}
& \includegraphics[width=0.08\textwidth]{fig/NO3_84}
& \includegraphics[width=0.08\textwidth]{fig/NO3_85}
& \includegraphics[width=0.08\textwidth]{fig/NO3_86}
& \includegraphics[width=0.08\textwidth]{fig/NO3_87} \\ \hline\hline
\multicolumn{1}{|c|}{\raisebox{0.035\textwidth}{4}}
& \includegraphics[width=0.08\textwidth]{fig/NO4_84}
& \includegraphics[width=0.08\textwidth]{fig/NO4_85}
& \includegraphics[width=0.08\textwidth]{fig/NO4_86}
& \includegraphics[width=0.08\textwidth]{fig/NO4_87} \\\hline
\end{tabular}
\end{center}
\caption{\label{fig:H4orb} Change in $\dno$ active orbitals of \ce{H4} with $\theta$.}
\end{figure}
From Figure \ref{fig:H4orb}, it is seen that from $\theta=84^\circ$ to $85^\circ$ the $\dno$ orbitals resemble those of two separate \ce{H2} molecules, albeit with small tails on the opposite pair of protons. At $\theta=86^\circ$, the orbitals begin to extend significantly to the neighbouring pair of protons, and by $\theta=87^\circ$ they are almost evenly distributed. This transition occurs quickly with respect to the PEC, in just 3$^\circ$.
Analysis of the $\Delta$ values and the orbitals of $\dno$ partially explains the behaviour of the model for the this challenging \ce{H4} PEC, but it is does not explain why it differs from the FCI result. A comparison of the $\dno$ and FCI two-electron densities, $\Gamma^{\dno}(\mathbf{r}_1,\mathbf{r}_2)$ and $\Gamma(\mathbf{r}_1,\mathbf{r}_2)$, can show conclusively how the $\dno$ model of electron correlation differs from how the electrons actually behave. The $\dno$ and FCI opposite-spin ($\alpha\beta$) and parallel-spin ($\alpha\alpha$) two-electron densities for \ce{H4} at $\theta= 70^\circ, 85^\circ$ and 90$^\circ$, with $R=1.7$ {\AA}, are shown in Figure \ref{fig:2eden}.
\begin{figure*}
\begin{center}
$$\Gamma^{\alpha\beta}(\mathbf{r},\mathbf{r}_0)$$
\includegraphics[width=0.325\textwidth]{fig/FCIab70}
\includegraphics[width=0.325\textwidth]{fig/FCIab85}
\includegraphics[width=0.325\textwidth]{fig/FCIab90}
$$\Gamma^{\dno,\alpha\beta}(\mathbf{r},\mathbf{r}_0)$$
\includegraphics[width=0.32\textwidth]{fig/DNOab70}
\includegraphics[width=0.325\textwidth]{fig/DNOab85}
\includegraphics[width=0.325\textwidth]{fig/DNOab90}\\
$$\Gamma^{\alpha\alpha}(\mathbf{r},\mathbf{r}_0)$$
\includegraphics[width=0.325\textwidth]{fig/FCIaa70}
\includegraphics[width=0.325\textwidth]{fig/FCIaa85}
\includegraphics[width=0.325\textwidth]{fig/FCIaa90}\\
$$\Gamma^{\dno,\alpha\alpha}(\mathbf{r},\mathbf{r}_0)$$
\includegraphics[width=0.325\textwidth]{fig/DNOaa70}
\includegraphics[width=0.325\textwidth]{fig/DNOaa85}
\includegraphics[width=0.325\textwidth]{fig/DNOaa90}
\caption{\label{fig:2eden} FCI and $\dno$ two-electron densities of \ce{H4} cluster when a test electron is placed at a \ce{H} nucleus, $\mathbf{r}_0$ (orange dot). Densities are shown for \ce{H4} at $R = 1.7$ {\AA} and $\theta =$ $70^\circ$ (left), $85^\circ$ (middle), and $90^\circ$ (right).}
\end{center}
\end{figure*}
The two-electron densities are plotted for a test-electron placed at the upper right \ce{H} atom (indicated by an orange dot). Therefore, the plots show the probability of finding the other electron(s), opposite or parallel-spin, within the \ce{H4} cluster.
At $\theta = 70^\circ$ (left column of Figure \ref{fig:2eden}), the $\dno$ and FCI two-electron densities closely resemble each other. For opposite-spin electrons (top half of Figure \ref{fig:2eden}), if an $\alpha$ electron is found at the upper-right \ce{H} atom, then it is most likely to find a $\beta$ electron on the closest \ce{H} atom, to which it is paired. It would then be equally likely to find the other $\beta$ electron on either of the two more distant \ce{H} atoms. In the case of parallel-spin electrons (bottom half of Figure \ref{fig:2eden}), if an $\alpha$ electron is found at the upper-right \ce{H} atom then the other $\alpha$ electron must be found in the ``other" \ce{H2} molecule, with an equal probability of being found at either proton.
At $\theta = 90^\circ$ (right column of Figure \ref{fig:2eden}), the behaviour is much different. For electrons of opposite-spin, when an $\alpha$ electron is found at the upper-right \ce{H} atom, it is equally likely to find the $\beta$ electron at either of the nearest protons, with a reduced probability of finding a $\beta$ electron at the furthest proton. Comparing the $\dno$ and FCI two-electron densities, the same qualitative behaviour is seen but the $\dno$ 2-RDM results in too high a probability of finding the $\beta$ electron on the furthest proton compared to the probability at the nearest protons. In the case of parallel-spin electrons, if an $\alpha$ electron is found at the upper-right \ce{H} atom then it is most likely that another $\alpha$ electron is on the furthest proton, with only a slight chance of an $\alpha$ electron on the nearest protons. Again, $\dno$ models this behaviour qualitatively, but compared to FCI, $\dno$ predicts the probability of finding an $\alpha$ electron on the nearest protons to be too high relative to the furthest proton. This suggests that the amount of exchange, or Fermi correlation, which occurs between parallel-spin electrons, in the $\dno$ 2-RDM is slightly inadequate at $\theta=90^\circ$.
Finally, at $\theta=85^\circ$ (middle column of Figure \ref{fig:2eden}), the difference between $\dno$ and FCI two-electron densities is the most pronounced. It is evident that the electrons of the $\dno$ 2-RDM at $\theta=85^\circ$ behave more like the electrons at $\theta=70^\circ$ (\ce{2H2} regime) than the actual behaviour given by the FCI two-electron density. In the case of the $\dno$ two-electron densities, both opposite and parallel-spin, the probabilities of finding electrons on the left-side protons (other \ce{H2}) is near equivalent. Whereas, the FCI two-electron density shows uneven probabilities for the left-side protons, which more resembles the behaviour at $\theta=90^\circ$ (\ce{H4} regime).
The inability of $\dno$ to accurately model the PEC from $\theta = 70^\circ$ to $90^\circ$, is due to the energetic preference of $\dno$ to model \ce{H4} as \ce{2H2}, at $\theta$ values close to $90^\circ$. The error observed in the parallel-spin two-electron density suggests that $\dno$ is incorrectly describing the correlation between different electron pairs (interpair correlation), which is similar to the conclusion reached in a study with PNOF6.\cite{RamosCordoba2015} For $\dno$, this could mean an error in the exchange component (Fermi correlation), or possibly something else, but it implies that the high-spin correction (HSC) is not adequate. The HSC provides the correct description in the strong correlation limit (both pairs), but it has deficiencies for moderate amounts of static correlation and must be improved upon.
The inadequacy of the HSC is highlighted further by examining the $N$-representabiltity of the $\dno$ 2-RDM; in particular, the violation of the $PQG$-conditions (Subsection \ref{ssec:Nrep}). The sums of the negative eigenvalues of the $\bm{P}$, $\bm{Q}$, and $\bm{G}$ matrices corresponding to the $\dno$ 2-RDM for rectangle-to-square \ce{H4} are shown in Figure \ref{fig:PQGneg}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{fig/PQGneg}
\caption{\label{fig:PQGneg} Sum of the negative eigenvalues of $\bm{P}$, $\bm{Q}$ and $\bm{G}$ for the rectangle-to-square \ce{H4} cluster as a function of $\theta$ at $R= 1.7$ \AA.}
\end{center}
\end{figure}
The sums are separated into contributions from different spin-blocks of the $\bm{P}$, $\bm{Q}$, and $\bm{G}$ matrices. The $\bm{P}$ and $\bm{Q}$ matrices are separated into spin-blocks for each specific spin-pair, whereas the $\bm{G}$ matrix is separated into only three blocks, $\bm{G}^{\alpha\beta\alpha\beta}$, $\bm{G}^{\beta\alpha\beta\alpha}$, and $\bm{G}^{\sigma\sigma\sigma\sigma}$, due to the coupling of the parallel-spin blocks [Equation \eqref{eq:Gmat}]. For all values of $\theta$, the parallel-spin blocks of the $\bm{P}$ and $\bm{Q}$ matrices have no negative eigenvalues. This is completely expected considering that all parallel-spin terms of the $\dno$ 2-RDM are derived from the $\dno$ wave function. The only terms not originating from the $\dno$ wave function are those corresponding to the HSC, which are applied to the opposite-spin block of the 2-RDM only [Equation \eqref{eq:2RDMos}]. The off-diagonal, $\bm{G}^{\alpha\alpha\beta\beta}$ and $\bm{G}^{\beta\beta\alpha\alpha}$, spin-blocks of $\bm{G}^{\sigma\sigma\sigma\sigma}$ contain 2-RDM elements corresponding to the HSC and they are responsible for the contribution to $G^{\sigma\sigma\sigma\sigma}_\text{neg}$. The most severe violation of the $PQG$-condition is seen for $\bm{G}^{\alpha\beta\alpha\beta}$, also known as the particle-hole density matrix. The sum of the negative eigenvalues of $\bm{G}^{\alpha\beta\alpha\beta}$ reach a minimum at $\theta=90^\circ$, where $G^{\alpha\beta\alpha\beta}_\text{neg} = -0.229$. Due to the fact that the HSC is applied to the exchange-like terms ({\it i.e.}~$\Gamma^{\alpha\beta}_{pqqp}$) of the $\dno$ 2-RDM, the sum of the negative eigenvalues of the $\bm{P}^{\alpha\beta\alpha\beta}$ and $\bm{Q}^{\alpha\beta\alpha\beta}$ matrices are equivalent for all values of $\theta$. In the case of $P_\text{neg}$ and $Q_\text{neg}$, the minimum value observed is -0.0818, which occurs at $\theta = 86^\circ$ and $\theta = 94^\circ$. As seen for the energy and orbital occupancies, the violation of the $PQG$-conditions exhibits different behaviour, with respect to $\theta$, in the neighbourhood of $\theta = 90 ^\circ$ (between the grey rectangles) compared to $\theta$ closer to $70^\circ$ or $110^\circ$ (outside the grey rectangles).
\FloatBarrier
\section{Conclusions}
\label{sec:conc}
The $\dno$ method is one of many that combines different models for different aspects of electron correlation in an effort to efficiently capture the total. The $\dno$-ODF approach is unique in the manner that it combines an approximate 2-RDM for static correlation with an ODF for dynamic correlation. With the exception of the high-spin correction, the $\dno$ 2-RDM and energy can be derived from a multideterminantal wave function with restricted $2n$-tuple excitations. Combining the $\dno$ 2-RDM for static correlation with an ODF for dynamic correlation requires the inclusion of a double-counting correction, which can be done by replacing the short-range correlation counted by the $\dno$ 2-RDM.
Previously, $\dno$, and other NOF methods, optimized the 2-RDM through alternating optimization of the occupation numbers and the orbitals until convergence was reached. The orbitals themselves were optimized by iterative diagonalization of a pseudo-Fock matrix. The trust-region Newton method introduced here, which allows for simultaneous optimization of the orbitals and occupancies, drastically reduces the number of optimization steps required and has the ability to identify and avoid saddle points.
Even simple hydrogen clusters can prove challenging for electronic structure methods, particularly single-reference methods. Overall, the $\dno$-ODF approach performs well when describing the potential energy surfaces of \ce{H2}, \ce{H3} and \ce{H4}, relative to the more accurate, but costly, multireference wave function method, MRMP2.
Analysis of the two-electron density of \ce{H4} reveals that $\dno$ incorrectly describes interorbital correlation. Despite being accurate in the strong-correlation limit, the {\it ad hoc} HSC does have deficiencies, particularly near the $D_{2h} \to D_{4h}$ transition. A more complete description of interorbital correlation within $\dno$ is a target of future development.
\begin{acknowledgments}
JWH thanks the Natural Sciences and Engineering Research Council of Canada (NSERC) for a Discovery Grant, Compute/Calcul Canada for computing resources and the Discovery Institute for Computation and Synthesis for useful consultations.
\end{acknowledgments}
\section*{Data Availability}
The data that support the findings of this study are available from the corresponding author upon request.
|
\section{Introduction}
With the rapid development of deep learning technology, various face forgery methods \cite{Thies_2016_CVPR,10.1145/3306346.3323035,karras2019stylebased,faceswap,deepfake} based on deep learning have been proposed. Nowadays, more and more forged faces without any visual artifacts are widely spread on social media. If these forgery images are abused, it probably leads to a series of moral, ethical, and safety issues. Nowadays, face forgery detection is facing severe challenges.
Up to now, many deepfake detection methods \cite{mccloskey2018detecting,yu2018attributing, Huaxiao2018,guarnera2020deepfake,
9072088,li2018ictu,nguyen2019multitask,capsule,li2020face,mesoNet,zhao2021multiattentional} have been proposed.
Most existing methods are mainly based on Convolutional Neural Network (CNN). For instance, in \cite{li2018ictu, 9072088}, the authors make full use of physiological signal, such as the inconsistency of human blinks, to detect deepfake;
in \cite{capsule}, the authors first introduce the capsule network for face forgery detection, and show that it can achieve very good detection performances compared with related methods;
in \cite{yu2018attributing}, the authors try to detect those artifacts generated by GAN (Generative adversarial network) for exposing face forgery; in \cite{zhao2021multiattentional}, the authors consider deepfake detection as a fine-grained classification problem, and propose a new multi-attentional detection network.
Typically, the existing detection methods can achieve good detection accuracy for high-quality databases. However, the detection accuracies would drop compared with their high-quality versions. Furthermore, for those face images with more diverse contents, such as images from WildDeepfake\cite{zi2021wilddeepfake}, the detection performances of existing methods are still far from satisfactory.
The previous results in \cite{rssler2019faceforensics} show that Xception \cite{chollet2017xception} has better sensitivity to manipulated face images created by deepfake, and it can outperform many related methods for face forgery detection. In this paper, therefore, we use the Xception as the backbone of the proposed method, and introduce dual attention mechanism and feature fusion in the middle flow and exit flow in the original Xception model for face forgery detection. In the middle flow, we firstly obtain different high-dimensional features using different levels of convolution, and refine the features via Convolutional Block Attention Module (CBAM) simultaneously. Finally, we fuse them to get a more comprehensive high-dimensional features for subsequent network analysis. In the exit flow, we employ the self-attention mechanism and depthwise separable convolution to learn the global information and local information of the fused features separately to improve the classification ability of the proposed model. Experimental results show that the proposed method outperforms the related methods both in effectiveness and generalization ability.
The rest of the paper is organized as follows. Section \ref{sec:Relative} describes related works. Section \ref{sec:Proposed} introduces the proposed method in detail. Section \ref{sec:experiments} shows comparative results and discussions. Finally, the concluding remarks of this paper and future works are given in Section \ref{sec:Conclusion}.
\section{Related work}
\label{sec:Relative}
In this section, we will describe three related works used in the proposed method, that is, Xception \cite{chollet2017xception}, Convolutional Block Attention Module (CBAM) \cite{woo2018cbam} and self-attention Mechanism \cite{aug_attention_conv}.
\subsection{Xception}
As illustrated in Fig. \ref{Fig:xception}, Xception \cite{chollet2017xception} consists of three flows: Entry flow, Middle flow, and Exit flow. Entry flow includes a classical convolution block and 3 residual separableConv blocks; Middle flow includes 8 residual separableConv blocks; Exit flow includes 1 residual separableConv block and 1 separableConv block. Expect for the first block and the last block, all blocks have linear residual connections around them, which aims to prevent the gradient from disappearing during the training process of the network. Xception is a network architecture based on depthwise separable convolution, which can not only significantly reduce the number of parameters, but also independently learn channel correlation and spatial correlation separately. Note that Xception is originally used for traditional image classification. In \cite{rssler2019faceforensics}, Xception is introduced to detect face forgery and achieves good detection results. Thus, Xception usually serves as a baseline network for comparative study in most recent related works.
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.45]{./img/xception.pdf}
\caption{The framework of Xception, which includes entry flow, middle flow and exit flow. Note that all Convolution and SeparableConvolution layers are followed by batch normalization \cite{BN} }
\label{Fig:xception}
\end{figure*}
\subsection{Convolutional Block Attention Module}
CBAM \cite{woo2018cbam} is a lightweight and general module, which aims to refine features via attention mechanism. As illustrated in Fig \ref{Fig:CBAM}, given an input feature map, CBAM infers attention maps along two separate dimensions, i.e., channel and spatial, then the attention maps are multiplied to the input feature map for adaptive feature refinement. In this way, CBAM can help the model refines the feature effectively by learning what and where to emphasize or suppress. CBAM is used in sign language recognition \cite{CBAM-resnet3d}, generative model \cite{CBAM-GAN} and target detection \cite{CBAM-target-detection}. To our best knowledge, however, there are no related works for deepfake detection.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{./img/CBAM.pdf}
\caption{The framework of the Convolutional Block Attention Module}
\label{Fig:CBAM}
\end{figure}
\subsection{Self-Attention Mechanism}
\label{subsec:Self-Attention}
The self-attention mechanism \cite{self_attention} tries to focus on key information and ignore irrelevant information via learning different weights corresponding to the feature maps. The self-attention mechanism can be described as mapping a query and a set of key-value pairs to an output, where the query (Q), keys (K) and values (V) can be obtained from the input feature map.
Let $H$, $W$, $F_{in}$ denote the height, width and number of channel of input feature map. $d_{k}$ denotes the keys of dimension.
Given an input feature map of shape $( H,W,F_{in})$, we firstly flatten it to a matirx $X\in$ $\mathbb{R}^{HW\times F_{in}}$. The output of the single-head attention mechanism can be calculated as:\\
\begin{equation}
\label{formula_self_attention}
\begin{aligned}
Q = XW_{q}; K = XW_{k}; V = XW_{v}\\
output = softmax( \frac{ Q K^{T}}{\sqrt{d_{k} } } ) V
\end{aligned}
\end{equation}
\noindent where $W_{k}, W_{q}$ and $W_{v}$ represent randomly initialized coefficient matrices, which can be updated during training stage. The softmax denotes the normalization operation. Through the self-attention mechanism, we can adaptively learn the global feature information. Note that the classical convolution operation has a significant weakness in that it only operates on a local neighbourhood, thus missing global information. In some applications, self-attention mechanism can be used to enhance the convolution operation.
\section{Proposed Method}
\label{sec:Proposed}
As illustrated in Fig. \ref{fig:model}, the proposed method is an improved version of Xception, and it also includes three flows, that is, entry flow, middle flow and exit flow. We will describe the three modules in the following subsections.
\begin{figure*}[t]
\centering
\includegraphics[scale=0.56]{./img/model.pdf}
\caption{The proposed framework based on Xception. The symbol $\otimes$ denotes the feature map concatenation, and $\oplus$ denotes feature map addition. }
\label{fig:model}
\end{figure*}
\subsection{Entry Flow}
The entry flow in the proposed framework is exactly the same as that in original Xception model (refer to Fig. \ref{Fig:xception}), which consists of a classical convolution block and 3 residual separabelconv blocks.
\subsection{Middle Flow Combined with Feature Fusion \& CBAM}
The middle flow in the original Xception is composed of 8 residual separableConv blocks within a single branch. In the proposed middle flow, however, we firstly reduce the number of residual separableConv blocks from 8 to 6, and assign them into three parallel branches. As illustrated in Fig. \ref{fig:model}, we assign 1, 2, and 3 residual separableConv blocks in the three branches respectively. In this way, the proposed method is able to learn different high-dimensional semantic features of image by using different levels of convolution. Furthermore, we introduce the CBAM after every residual separabelConv block to enhance the characterization ability of the features and refines intermediate features. Finally, we merge the reorganized high-dimensional features in the three branches, and employ a $1 \times 1$ convolution to learn the channel correlation to fully integrate these features. By doing so, we can also significantly reduce the dimension of feature map that feed to the exit flow.
\subsection{Exit Flow Combined with Self Attention}
The exit flow in Xception consists of a residual separableConv block and a separableConv block (including two depthwise separable convolutions). Thus, just local information of the fused features is considered here. Inspired by method \cite{aug_attention_conv}, we introduce a self-attention module after the first depthwise separable convolution, and this module is paralleled with another depthwise separable convolution in the proposed exit flow as illustrated in Fig.\ref{fig:model}. As described in section \ref{subsec:Self-Attention}, each attention map over the fused features are firstly computed from keys (K) and query (Q) in self-attention mechanism. These attention maps are used to compute weighted averages of the values (V). Then we concatenate the results and reshape them to match the original spatial dimensions via a pointwise convolution. In addition, we employ the depthwise separable convolution at the other branch to learn the local information of the fused features. Finally, we combine the global information (via self-attention) and local information (via depthwise separable convolution) of the fused features, and feed them to a fully connected layer for classification.
\section{Experiments}
\label{sec:experiments}
\begin{table*}[htb]
\renewcommand\arraystretch{1.5}
\caption{Detection accuracies of different detection networks. Some results are obtained from \cite{zi2021wilddeepfake}. In all following Tables, those values with an asterisk (*) denote the best performances in the corresponding cases. }
\centering
\setlength{\tabcolsep}{7mm}{
\begin{tabular}{cccc}
\hline
Network & TIMIT (LQ) & FF++deepfake (LQ) & WildDeepfake \\
\hline
AlexNet \cite{AlexNet} & 94.77\% & 90.58\% & 60.37\% \\
VGG16 \cite{vgg16} & 98.73\% & 90.19\% & 60.92\% \\
ResNetV2-50 \cite{resNet} & 94.88\% & 90.91\% & 63.99\% \\
ResNetV2-152 \cite{resNet} & 95.68\% & 88.00\% & 59.33\% \\
Inception-v2 \cite{inceptionV3} & 90.30\% & 89.44\% & 62.12\% \\
MesoNet-4 \cite{mesoNet} & 91.18\% & 87.75\% & 64.47\% \\
MesoNet-Inception\cite{mesoNet} & 97.85\% & 84.82\% & 68.48\% \\
Xception \cite{rssler2019faceforensics} & 99.65\% & 90.25\% & 74.32\% \\
ADDNet-2D \cite{zi2021wilddeepfake} & 99.54\% & 90.42\% & 76.25\% \\
Capsule \cite{capsule} & 99.85\% & 93.81\% & 77.60\% \\
Proposed Method & \bf{99.86\%*} & \bf{96.24\%*} & \bf{83.32\%*} \\
\hline
\end{tabular}}
\label{Table:Comparative}
\end{table*}
In our experiments, we employ three popular databases, including Deepfake-TIMIT ( denoted as TIMIT for short) \cite{korshunov2018deepfakes}, FaceForensics++ (denoted as FF++) \cite{rssler2019faceforensics}, and WildDeepfake \cite{zi2021wilddeepfake}. In training stage, we use MTCNN \cite{zhang2016joint} to locate and align the face regions in video frames, and then resize the face images to the size of $224\times224$. For a fair comparison, all detection results are evaluated on the same test dataset as existing methods.
We employ the cross-entropy loss and Adam optimizer with the batch size 64. We set the initial learning rate to be 0.0001. We train the networks in total of 20 epochs, and reduce the learning rate by a half every 5 epochs, and select the model with the best accuracy as the final model.
\subsection{Evaluation on FF++, TIMIT and WildDeepfake}
In this experiment, just the low-quality (LQ) versions of TIMIT and FF++(deepfake) databases are considered since their high-quality (HQ) versions are relatively detected by the current methods. The detection accuracies are shown in Table \ref{Table:Comparative}. From Table \ref{Table:Comparative}, we obtain two following observations:
\begin{itemize}
\item First of all, the proposed method always outperforms other test methods in all cases. Compared with the advanced method (i.e., Capsule \cite{capsule}), the proposed method achieves 0.01\%, 2.43\%, and 5.72\% improvements for TIMIT (LQ), FF++deepfake (LQ), and WildDeepfake respectively. Compared with the original Xception \cite{rssler2019faceforensics} model, the proposed method achieves 0.21\%, 5.99\%, and 9.00\% improvements for the three databases respectively, which means that the feature fusion and dual attention mechanism introduced in the proposed model can effectively enhance the ability of the original Xception for face forgery detection.
\item Compared with TIMIT and FF++, the detection accuracies for WildDeepfake are relatively lower for all detection networks. The main reason is due to the diversity of WildDeepfake. Note that those video clips of WildDeepfake are collected on the Internet, they are from different sources and may be compressed several times. Based on our experiments, the proposed method still obtain 83.32\% detection accuracy for WildDeepfake. While the detection accuracies are all lower than 77.70\% for other test networks.
\end{itemize}
\subsection{Cross-dataset Evaluation on Celeb-DF}
To show the generalization of the proposed method, cross-dataset evaluation on a new database - Celeb-DF \cite{Celeb_df} is considered. In this section, we first train a classifier based on the high-quality version of FF++ and WildDeeepfake respectively, and then use the resulting classifier to detect those face forgeries on Celeb-DF. The comparative results are shown in Table \ref{Table:Cross1} and Table \ref{Table:Cross2} respectively.
\begin{table}[t]
\renewcommand\arraystretch{1.5}
\center
\caption{Cross-dataset evaluation on Celeb-DF (AUC) via training on FF++(HQ). Some results are obtained from \cite{zhao2021multiattentional}.}
\begin{tabular}{ccc}
\hline
network & FF++ & Celeb-DF \\
\hline
MesoNet-Inception \cite{mesoNet} & 83.00\% & 53.60\% \\
Xception\cite{rssler2019faceforensics} & {99.70}\% & 65.30\% \\
Capsule \cite{capsule} & 90.61\% & 67.92\% \\
Proposed Method & 98.09\% & \bf{68.39\%*} \\
\hline
\end{tabular}
\label{Table:Cross1}
\end{table}
\begin{table}[t]
\renewcommand\arraystretch{1.5}
\center
\caption{Cross-dataset evaluation on Celeb-DF(ACC) via training on WildDeepfake. }
\begin{tabular}{ccc}
\hline
network & WildDeepfake & Celeb-DF \\
\hline
MesoNet-Inception \cite{mesoNet} & 68.48\% & 49.11\% \\
Xception\cite{rssler2019faceforensics} & 74.32\% & 51.87\% \\
Capsule \cite{capsule} & 77.60\% & 53.00\% \\
Proposed Method &{83.32}\% & \bf{72.62\%*} \\
\hline
\end{tabular}
\label{Table:Cross2}
\end{table}
The results in the two above tables show that the generalization of the proposed method is better than three other modern networks. Especially for the case of training on Wilddeepfake, the proposed method achieves around 20\% improvements.
\subsection{Ablation Study}
\label{ablation}
In the proposed method, we introduce dual attention mechanism (i.e., CBAM in middle flow and self-attention in exit flow) and feature fusion (in the middle flow) into original Xception model for face forgery detection. In order to verify the rationality of the proposed method, we conduct the several ablation experiments in this section. For simplicity, WildDeepfake is used for performance evaluation.
\begin{table}[t]
\caption{Ablation results on WildDeepfake(ACC)with different attention mechannisms. }
\center
\renewcommand\arraystretch{1.5}
\setlength{\tabcolsep}{2mm}{
\begin{tabular}{cc}
\hline
Setups & WildDeepfake \\
\hline
Proposed Method w/o CBAM\&Self-Attention & 78.28\% \\
Proposed Method w/o Self-Attention & 80.09\% \\
Proposed Method w/o CBAM & 82.78\% \\
Proposed Method & \textbf{83.32\%*} \\
\hline
\end{tabular}}
\label{ablation_study}
\end{table}
\begin{itemize}
\item Analysis on Dual Attention Mechanism: In this experiment, we evaluate the the proposed model after removing the CBAM and/or self-attention. The experimental results are shown in Table \ref{ablation_study}. From Table \ref{ablation_study}, we observe that when both CBAM and self-attention are removed from the proposed model, the detection accuracy is 78.28\%. When self-attention or CBAM is removed, the corresponding detection accuracies increase to 80.09\% and 82.78\% respectively. When the self-attention and CBAM are preserved, the proposed method works the best, and achieves 83.32\% detection accuracy. The above results show that the dual attention mechanism is useful for face forgery detection.
\vspace{0.5em}
\item Analysis on Different Setups in Middle Flow: In the proposed middle flow, we design three branches (i.e., 1, 2, and 3 residual SparavleConv Blocks and CBAM respectively) to extract different high-dimensional semantic features of face image, and then perform feature fusion on all branches for subsequent network analysis. In this experiment, we evaluate the performance on using different setups in the middle flow. Some comparative results are shown in Table \ref{tab:middle_flow}. From Table \ref{tab:middle_flow}, we observe that the proposed middle flow achieves the best detection performance. When adding the 4th branch (i.e., 4 residual SparavleConv Blockss and CBAM) or removing the 3rd branch in proposed middle flow, the corresponding detection accuracies will drop from 83.32\% to 79.20\% and 74.99\% respectively, which means that setting 3 branches is reasonable in the proposed model. In addition, we compared other four setups with a single branch, that is, just using the 1st, 2nd, 3rd branch respectively, and the original middle flow in Xception. Their detection accuracies all are lower than 80.05\%, which means that feature fusion is helpful to enhance the model performance.
\end{itemize}
\section{Conclusion}
\label{sec:Conclusion}
In this paper, we first introduce dual attention mechanism (i.e., CBAM and self-attention) and feature fusion into Xception for face forgery detection, and show that the proposed method can significantly improve the detection performance of the original Xception model, and achieve state-of-the-art results both in the effectiveness and generalization ability compared with related methods. In addition, we also provide some ablation experiments to verify the rationality of introducing dual attention mechanism and feature fusion in the proposed method. In future, there are several important issues in the proposed framework worthy of in-depth study. For instance, we just consider the spatial information as network input in the proposed method. We would combine some features in the frequency domain, and construst a two stream network for further enhancing the detection performance.
In addition to high-dimensional facial semantic features, we try to use multiple attention maps to explore discriminative local region (such as eyes and nose) on the face, and fuse these features in the middle flow for face forgery detection.
\begin{table}[t]
\caption{Ablation results on WildDeepfake (ACC) with different setups in the proposed middle flow. }
\center
\renewcommand\arraystretch{1.5}
\setlength{\tabcolsep}{1mm}{
\begin{tabular}{c c c}
\hline
Setups in Middle Flow & Branch Number & WildDeepfake \\
\hline
Proposed middle flow & 3 & \bf{83.32\%*} \\
Adding the 4th branch & 4 & 79.20\% \\
Removing the 3rd branch & 2 & 74.99\% \\
Just using the 1st branch & 1 & 79.60\% \\
Just using the 2nd branch & 1 & 79.57\% \\
Just using the 3rd branch & 1 & 80.04\% \\
Original middle flow in Xception & 1 & 78.53\% \\
\hline
\end{tabular}}
\label{tab:middle_flow}
\end{table}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
It is a well-known fact that trigonometric functions are periodic: if $f(x)$ is any trigonometric function, then $f(x + 2\pi) = f(x)$ for all values of $x$ in the domain of $f$. Therefore, it is natural to define trigonometric functions on the unit circle, where all multiples of $2\pi$ are identified when we wrap the real line onto the circle. Because of this definition, trigonometric functions are also called circular functions. In this setting, the trigonometric functions $\sin t$ and $\cos t$ are just the unit circle's parametrization with respect to arc length.
Recall that the unit circle is the locus of all points in the plane $\mathbb{R}^2$ that are at a distance of one unit from the origin, where distances are measured using the standard Euclidean norm: $\lVert \vec{x} \rVert = (x_1^2+x_2^2)^{1/2}$. What if we switch to an $L_p$-norm: $\lVert \vec{x} \rVert_p = (|x_1|^p+|x_2|^p)^{1/p}$, $(p \ge 1 )$? We then get a new family of curves defined by the equations $|x|^p+|y|^p =1$. These are called unit $p$-circles and are shown in the figure below. Because these curves are in between a square and a circle, they are also called squircles.
\begin{figure}[h]
\centering
\includegraphics[width=0.3\textwidth]{squircleThickFix.png}
\caption{$p$-circles for $p = 1, 2, 4$, and $10$ from inside to outside, respectively }
\end{figure}
Can we parametrize these $p$-circles to get $p$-trigonometric functions $x = \sin_p t$ and $y= \cos_p t$ such that, when $p=2$, we recover the standard trigonometric functions? What properties and identities do these generalized trigonometric functions have? Can we do calculus over these curves? What can be said about the periods of these functions? How does the curvature change along a $p$-circle? What is the area it encloses? What are the rational points on $p$-circles? Note that for any $p \ge 1 $, $L_p$, as defined above, gives a norm, but this norm is induced by an inner product only when $p=2$ (\cite{SK}).
Therefore, $p=2$ is a special case of interest; however, all of the aforementioned questions are well defined for any $p \ge 1$. The goal of this paper is to investigate these questions. Our primary reference for this research is \cite{WP}. While we follow the general outline given in \cite{WP}, we also do some independent investigation.
There are at least three ways to generalize trigonometric functions. These correspond to 3 different \break parametrizations of the unit $p$-circle: areal, arc length, and angular. It turns out that these three parametrizations are equivalent only when $p=2$! The parametrization we will be working with corresponds to the areal parametrization. Our investigation of these generalized trigonometric functions and their inverses led to several interesting connections involving double angle formulas, norms induced by inner products, Stirling numbers, Bell polynomials, Lagrange inversion, gamma functions, and generalized $\pi$ values.
These $p$-trigonometric functions have several applications, specifically in design. Rather than using rounded rectangles, Apple uses $p$-circles for their icons, as the curvature continuity leads to a more sleek look, unifying the design of their hardware and icons \cite{AS}. Another design application can be found in squircular dinner plates, designed to allow a greater surface area for food while taking up the same amount of cabinet space as their circular counterparts \cite{PL}.
The paper is organized as follows. In Section \ref{ptrig}, we define $p$-trigonometric functions using a differential equations approach and derive some basic properties of these functions. We show that for any positive integer $k$, the well-known double angle formula for $\sin(2x)$ holds for $\sin_k(2x)$ if and only if $k=2$. In Section \ref{taylor}, we focus on the successive derivatives of $\sin_p(x)$. This revealed a connection between the coefficients of the terms in the derivatives and Stirling numbers of the first kind. We derive the Taylor series of $\sin_p^{-1}x$ using Newton's binomial series and then find the Taylor series of its inverse using Lagrange inversion theorem. It is shown that both $\sin_p x$ and $\sin_p^{-1} x$ are analytic functions at $x = 0$. Our work gave rise to the concept of rigidity of functions, which deals with the simultaneous vanishing of the derivatives of a function and its inverse. A generalization of $\pi$ for $p$-circles, $\pi_p$, and its properties are examined in Section \ref{pi_p} using beta and gamma functions. Furthermore, we use a Monte Carlo method to compute $\pi_p$. In Section \ref{optimal}, we determine the value of $p$ for which the unit $p$-circle is halfway between the unit circle and the square that contains it from the lenses of area, perimeter, and curvature.
Rational points on $p$-circles are determined in Section \ref{rational}. We end the paper with some questions for future work in
Section \ref{conclusion}.
\vskip 3mm \noindent
\textbf{Acknowledgements:} This paper is the outcome of the MAT 268 (Introduction to Undergraduate Research in Mathematics) course taught by the first author to the remaining authors at Illinois State University in Spring 2021. We want to thank the department of mathematics for providing us with the necessary resources for this research. Discussions with Anindya Sen led to the notion of rigidity in Section \ref{taylor}. Pisheng Ding raised several interesting questions and comments after reading this paper. We thank both of them for their interest and input in this paper. Finally, we are grateful to an anonymous referee for many comments and suggestions.
\section{$p$-trigonometric functions} \label{ptrig}
Unless stated otherwise, $p$ will denote a positive real number that is at least 1.
\subsection{Coupled Initial Value Problem}\label{CIVP}
The standard trigonometric functions sine and cosine that parametrize the unit circle are famously coupled by the derivative relation
$\sin't = \cos t,\, \cos't = -\sin t$. If we take $x(t) = \cos t$ and $y(t) = \sin t$,
we see that the pair is one of many solutions to the system of differential equations
$$x'(t) = -y(t), \ \ \ y'(t) = x(t).$$ However, with the inclusion of the initial conditions
$$x(0) = 1,\ y(0) = 0,$$ differential equation theory guarantees that the sine and cosine functions are, in fact, the only solutions
to this system \cite{BP}, better known as the Coupled Initial Value Problem (CIVP).
For $p \geq 1$, a natural extension of the CIVP considers the functions $x(t),\, y(t)$ satisfying
$$x'(t) = -y(t)^{p-1}, \ y'(t) = x(t)^{p-1}, \ x(0) = 1, \ y(0) = 0.$$
The motivation for this extension comes from that fact that any functions $x(t)$ and $y(t)$ that satisfy the above CIVP parametrize the curve $x^p + y^p = 1$. This is seen by differentiating $h(t) := x(t)^p + y(t)^p$ with respect to $t$, to get $h'(t) = px(t)^{p-1}x'(t) + py(t)^{p-1}y'(t)$. Substituting $x'(t) = -y(t)^{p-1}, \ y'(t) = x(t)^{p-1}$, will show that $h'(t) = 0$. This means $h(t)$ is a constant function. Using the initial conditions, we can conclude that $h(t) = 1$, i.e., $x^p +y^p = 1$, as desired.
Again, from the general theory of differential equations, the above CIVP has a unique solution. We can define $\cos_pt = x(t)$
and $\sin_pt = y(t)$ as the unique solution to the generalized CIVP. But these functions do not parametrize $p$-circles in general. For instance, when $p$ is an odd positive integer, these functions parametrize $p$-circles only in the first quadrant where $x$ and $y$ are both positive. To circumvent this issue, we restrict the domain of the solutions of the CIVP and then extend them to functions on the real line using symmetry and periodicity. This is done in the next three subsections.
Once we have $\sin_p t$ and $\cos_p t$ in place, we may then define the other trigonometric functions $\tan_p t := \frac{\sin_p t}{\cos_p t},\, \csc_p t := \frac{1}{\sin_pt},\, \sec_pt := \frac{1}{\cos_pt}$, and $\cot_pt := \frac{1}{\tan_pt}$ such that the familiar inverse relations are maintained.
\subsection{Inverse $p$-trigonometric functions}\label{Inverse functions}
Starting with the equation $x=\sin_p{y}$, we use the CIVP to find $\sin_p^{-1}{x}$. Differentiating both sides with respect to $y$ and simplifying, we find:
\begin{eqnarray*}
\frac{dx}{dy} &=& \cos_p^{p-1}{y}\\
&=& (\cos_p^p{y})^{\frac{p-1}{p}}\\
&=& (1-\sin_p{y}^p)^{\frac{p-1}{p}} \\
&=& (1-x^p)^{\frac{p-1}{p}}.
\end{eqnarray*}
This is a separable differential equation. To solve it, we separate and integrate both sides. This gives:
\begin{eqnarray*}
\frac{dx}{dy} &=& (1-x^p)^{\frac{p-1}{p}}\\
\int{\frac{dx}{(1-x^p)^{\frac{p-1}{p}}}} &=& \int{dy} \\
\int_0^x{\frac{dt}{(1-t^p)^{\frac{p-1}{p}}}} &=& y = \sin_p^{-1}{x}.
\end{eqnarray*}
We can do the same for $x=\cos_p{y}$ to get
\begin{eqnarray*}
\cos_p^{-1}{x}= \int_x^1{\frac{dt}{(1-t^p)^{\frac{p-1}{p}}}}.
\end{eqnarray*}
\subsection{Areal parametrization of $p$-circles}
The unit circle has a useful property that a sector with angle measure $\theta$ in radians has an area of $\theta/2$. We can use this property to find sine and cosine in terms of area where $x=\cos(2a)$, $y=\sin(2a)$, and $a$ is the area of the sector made by the points $(1,0)$ and $(x,y)$. It is then natural to ask if this property extends to all $p$-circles.
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{areaOfSector.png}
\caption{Area of a $p$-sector}
\label{fig:areaOfSector}
\end{figure}
\begin{prop}\label{areapara}
Let $(x,y)$ be a point in the first quadrant of the unit $p$-circle, and $a$ be the area of the sector made by the points $(1,0)$ and $(x,y)$. It holds that $x=\cos_p(2a)$ and $y=\sin_p(2a)$.
\end{prop}
\begin{proof}
This argument is in the spirit of Levin \cite{LA}. Working in the first quadrant, the area of the sector in a $p$-circle can be given by the area of $A1+A2$ as denoted in Figure \ref{fig:areaOfSector}. This can be given by $a=\frac{1}{2}x(1-x^p)^{\frac{1}{p}}+\int_x^1{(1-t^p)^{\frac{1}{p}}dt}$. We can differentiate both sides with respect to $x$ and simplify to get the following:
\begin{eqnarray*}
\frac{da}{dx} &=& \frac{1}{2}\left((1-x^p)^{\frac{1}{p}}+x\frac{1}{p}(1-x^p)^{\frac{1}{p}-1}(-px^{p-1})\right)-(1-x^p)^{\frac{1}{p}}\\
&=&(1-x^p)^{\frac{1}{p}}\left(\frac{1}{2}-\frac{x^p}{2}(1-x^p)^{-1}-1 \right)\\
&=&(1-x^p)^{\frac{1}{p}}\left(\frac{(1-x^p)-x^p-2(1-x^p)}{2(1-x^p)} \right)\\
&=&-\frac{(1-x^p)^{\frac{1}{p}-1}}{2}.
\end{eqnarray*}
Using the fundamental theorem of calculus, we can write this as $a=\int_x^1{\frac{(1-t^p)^{\frac{1}{p}-1}}{2}dt}+c$. When $a=0$ and $x=1$, we get $c=0$. From here, we can conclude that $a=\frac{1}{2}\arccos_p{x}$. Solving for $x$ gives $x=\cos_p{(2a)}$.
We can do the same thing in terms of $y$ to get $a=\int_0^y{\frac{(1-t^p)^{\frac{1}{p}-1}}{2}dt}$+c. When $a=0$ and $y=0$, we get $c=0$. From here, this equation has been shown to be $a=\frac{1}{2}\arcsin_p{y}$ and thus $y=\sin_p{(2a)}$. As such, this shows that this property does extend to all unit $p$-circles.
\end{proof}
\subsection{Definition and Graphs of $\sin_px$ and $\cos_px$} \label{sec:extension}
To generalize the formula $\pi/2= \sin^{-1} (1)$, we first set $\pi_p/2 := \sin_p^{-1} (1)$.
Since we have shown that the $p$-trigonometric functions can be parametrized by area, we can now extend then to functions defined on the entire real line as follows. We first restrict them to $[0, \sin_p^{-1} (1)] =[0 , \pi_p/2]$ and then extend the domain to $[0, 2\pi_p]$ using symmetry:
\[
\sin_p t := \begin{cases}
\sin_p(\pi_p - t) & \pi_p/2 < t \le \pi_p, \\
-\sin_p(2\pi_p - t) & \pi_p < t < 2\pi_p.
\end{cases}
\]
We then periodically extend that it $(-\infty,\infty)$
by setting $\sin_p(t+2\pi_pk) = \sin_p(t)$ for any integer $k$. The definition of $\cos_p(t)$ is similar. The resulting graphs are shown below.
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{sinplabel.png}
\caption{Graph of $\sin_p{x}$ for $p=1,2,$ and $10$}
\label{fig:sinp}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=1\textwidth]{cosplabel.png}
\caption{Graph of $\cos_p{x}$ for $p=1,2,$ and $10$}
\label{fig:cosp}
\end{figure}
\subsection{Trigonometric identities}
While we may have defined the generalized CIVP in a manner similar to the original, there is no guarantee that $\sin_p t,\, \cos_p t$ thus defined satisfy familiar trigonometric properties and identities. In this section, we explore a few identities of the $p$-trigonometric functions.
\begin{lem}[$p$-Pythagorean Equation]\label{p-pythag} \cite[p. 268]{WP}
The functions $\sin_p t,\, \cos_p t$ satisfy $|\sin_pt|^p + |\cos_pt|^p = 1$ for all real $t$.
\end{lem}
\begin{proof}
This is clear from the definition of these functions using the CIVP and extension using symmetry and periodicity.
\end{proof}
\begin{comment}
\begin{proof}
We will work in the first quadrant and use the symmetry of the unit $p$-circle for the other quadrants. We restrict our domain and assume that $\sin_pt$ and $\cos_pt$ are non-negative. Now we define $H(t) := \sin_p^pt + \cos_p^pt$. Then $H'(t) = p\cos_p^{p-1}t\cos_p't + p\sin_p^{p-1}t\sin_p't$. Thus, by
the definition of $\sin_pt,\, \cos_pt$, we have $H'(t) = p\sin_p't\cos_p't + p(-\cos_p't)\sin_p't = 0$.
Therefore, $H(t) \equiv C$ for some constant $C$. If we consider $t = 0$, we see that
$\cos_p^p(0) + \sin_p^p(0) = C$; thus, by the initial conditions of the CIVP, we have $C = 1$.
\end{proof}
\end{comment}
It is clear from the $p$-Pythagorean Equation that the functions $\sin_p t,\, \cos_pt$ are bounded and
$|\sin_pt| \leq 1,\, |\cos_pt| \leq 1$. Dividing all terms of the $p$-Pythagorean equation by $|\sin_pt|^p$ and $|\cos_pt|^p$ gives the identities
$1 + |\cot_pt|^p = |\csc_pt|^p$ and $|\tan_pt|^p + 1 = |\sec_pt|^p$, respectively.
\begin{lem}
The functions $\sin_pt$ and $\cos_pt$ are odd and even, respectively.
\end{lem}
\begin{proof}
The functions $\alpha(t) := -\sin_p(-t)$ and $\beta(t) = \cos_p(-t)$
satisfy $\alpha'(t) = -(-\sin_p'(-t)) = \cos_p^{p-1}(-t) = \beta(t)^{p-1}$ and
$\beta'(t) = -\cos'_p(-t) = \sin_p^{p-1}(-t) = -\alpha(t)^{p-1}$. Note that
$\alpha(0) = -\sin_p(0) = 0$ and $\beta(0) = \cos_p(0) = 1$; thus, the functions
$\alpha, \beta$ satisfy the generalized CIVP. Then, by the uniqueness of solutions, we must have
$\sin_pt = -\sin_p(-t)$ and $\cos_pt = \cos_p(-t)$.
\end{proof}
However, not all standard $2$-trigonometric identities are satisfied. For instance, we show that for positive integer values of $p$, $\sin_p(2t) = 2\sin_pt\cos_pt$ is satisfied if and only if $p = 2$.
A double angle formula for generalized trigonometric functions is still sought after \cite{ED, SS}.
\begin{prop}\label{doubleangle}
Let $k \in \mathbb{Z}_+$. Then $\sin_k(2t) = 2\sin_k(t)\cos_k(t)$ if and only if $k = 2$.
\end{prop}
\begin{proof}
The desired identity is well known for $k = 2$. We suppose the identity holds for $k \ge 1$ and show that $k$ must be 2.
We consider the cases $k = 1$ and $k > 1$ separately. For $k = 1$, we note that the CIVP gives the unique solution
$\sin_1(t) = t$ and $\cos_1(t) = 1-t$. Then $\sin_1(2t) = 2t \ne 2t(1-t) = 2\sin_1 t \cos_1 t$. If
$k > 1$, then by Lemma \ref{p-pythag}, the functions $\sin_k$ and $\cos_k$ satisfy
$|\sin_k t|^k + |\cos_k t|^k = 1$. By the Intermediate Value Theorem, there exists some $t_0$ in $[0, \pi_p/2]$
such that $\sin_k(t_0) = \cos_k(t_0)$. As $t_0 \geq 0$, the substitution
$\sin_k(t_0) = \cos_k(t_0)$ into the $p$-Pythagorean identity gives
$2\sin_k^k (t_0) = 1$, therefore $\sin_k^k (t_0) = \frac{1}{2}$. Then, by the assumption that $\sin_k(2t) = 2\sin_k t \cos_k t$ is satisfied for all $t$, we may
raise all terms to the power $k$ and evaluate at the point $t_0$ to obtain
$\sin_k^k(2t_0) = 2^k (\frac{1}{2}) \frac{1}{2} = 2^{k-2}$. Since $\sin_k^k t$ is bounded above by $1$, we obtain
$2^{k-2} \leq 1$, which implies that $k \leq 2$. Together with the assumption that $k > 1$, we obtain
$k = 2$.
\end{proof}
It is known that the $L_p$ norm is induced by an inner product if and only if $p = 2$ \cite{SK}. Then together with
Proposition \ref{doubleangle}, we make the following remark.
\begin{remark}
The following are equivalent for $k \in \mathbb{Z}_+$:
\begin{itemize}
\item $L_k$ is a norm induced by an inner product,
\item $\sin_k(2t) = 2\sin_k t \cos_k t$, and
\item $k = 2$.
\end{itemize}
\end{remark}
\section{Taylor Series} \label{taylor}
Now that we have defined $p$-trigonometric functions and their derivatives by the CIVP, it is natural to study the higher derivatives of these functions. We begin by observing that for any $p > 1$, all the successive derivatives of $\sin_p x$ and $\cos_p x$ are defined for all values of $x$. In this section, we provide an algorithm for differentiating these functions, demonstrate some patterns and connections present in their successive derivatives, and formulate the Taylor series for $\sin_p^{-1}x$ and $\sin_p x$. The Taylor series representations of these functions provide a tool to express all of the derivatives of the $p$-trigonometric functions in one formula.
\subsection{Higher derivatives and the bracket notation}
Because of the simplicity and utility of the closed formulas for differentiation of $\sin_2x,\, \cos_2x$, it is natural to wonder about higher derivatives of $\sin_px,\, \cos_px$. We find these higher derivatives by utilizing the definition given by the CIVP in Section \ref{CIVP}. However, these derivatives become complex rather quickly. To help address this, we introduce a notation that will be used throughout this section in relation to higher derivatives of these $p$-trigonometric functions: $[m,n]_p:=\cos_p^m(x)\sin_p^n(x)$.
\begin{lem}\label{bracketNotaiton}
The derivative of $\cos_p^m(x)\sin_p^n(x)$ satisfies $ \frac{d}{dx}[m,n]_p=-m[m-1,n+p-1]_p+n[m+p-1,n-1]_p$.
\end{lem}
\begin{proof} Applying the standard rules of differentiation, we get the following.
\begin{align*}
\frac{d}{dx} [m,n]_p &=
\frac{d}{dx} (\cos_p^m(x)\sin_p^n(x)) \\
& = -m\cos_p^{m-1}(x)\sin_p^{p-1}(x)\sin_p^n(x)+\cos_p^m(x)\cdot n\sin_p^{n-1}(x)\cos_p^{p-1}(x) \\
&=-m\cos_p^{m-1}(x)\sin_p^{n+p-1}(x)+n\cos_p^{m+p-1}(x)\sin_p^{n-1}(x). \\
&=-m[m-1,n+p-1]_p+n[m+p-1,n-1]_p. \qedhere
\end{align*}
\end{proof}
Although we do not have a closed formula for finding derivatives of these functions, Lemma \ref{bracketNotaiton} serves as a recursive algorithm for computing successive derivatives, as demonstrated in the following example.
\begin{example}\label{derivativeExample}
Lemma \ref{bracketNotaiton} can be iteratively applied to $\sin_px$ to find the first few derivatives:
\begin{align*}
\sin_p x&=[0,1]_p \\
\frac{d}{dx} \sin_p x&=0+1[p-1,0]_p\\
\frac{d^2}{dx^2} \sin_p x&=0+(-p+1)[p-2,p-1]_p+0 \\
\frac{d^3}{dx^3} \sin_px&=0+-(p-1)(-(p-2)[p-3,2p-2]_p+(p-1)[2p-3,p-2]_p)+0 \\
&=0+(p^2-3p+2)[p-3,2p-2]_p+(-p^2+2p-1)[2p-3,p-2]_p+0.
\end{align*}
\end{example}
There seems to be no clear pattern that arises from these derivatives like there is for $\sin x$. However, in the next subsection, we will see one pattern in the coefficients of the first terms of these derivatives.
\subsection{Connection to Stirling numbers}
For any variable $x$ and a non-negative integer $n$, the falling factorial is defined as follows.
\[(x)_n:= \begin{cases}
1 & \text{ if } n =0, \\
x(x-1)(x-2) \cdots (x-n+1) & \text{ if } n \ge 1 .
\end{cases}\]
For $n \ge 1$, $(x)_n$ is a non-constant polynomial of degree $n$ whose coefficients are the Stirling numbers of the first kind. More precisely, we set:
\[ (x)_n = \sum_{k=1}^n s(n, k) x^k.\]
We will now show a connection between the successive derivatives of $\sin_p x$ and Stirling numbers.
Building a tower from the coefficients in Example \ref{derivativeExample}, we get:
\begin{center}
1\\
0 \ \ \verb"|" \ \ \underline{1}\\
0 \ \ \verb"|" \ \ \underline{$-p+1$} \ \ \verb"|" \ \ 0\\
0 \ \ \verb"|" \ \ \underline{$p^2-3p+2$} \ \ \verb"|" \ \ $-p^2+2p-1$ \ \ \verb"|" \ \ 0
\end{center}
We observed that the coefficients of the polynomials in the second column (underlined) can be expressed using Stirling numbers of the first kind $s(n,k)$. For instance, corresponding to the polynomial $p^2-3p+2$ (corresponding to the 3rd derivative of $\sin_p x$), we have $s(3, 3) = 1$, $s(3, 2) = -3$ and $s(3, 1) = 2$. To prove this, we need the following lemma.
\begin{lem}
For any $n \ge 1$, the first term of $\frac{d^n}{dx^n} (\sin_p(x))$ is given by
\[ (-1)^{n-1} (p-1)_{n-1} [p-n, (n-1)(p-1)]_p.\]
\end{lem}
\begin{proof}
We prove this using mathematical induction. For $n=1$, $\frac{d}{dx} (\sin_p(x)) = \cos_p^{p-1} (x) = 1 [p-1, 0]_p$, which agrees with the answer obtained with $n=1$ in the given expression. Having proved the base case, let us assume that the result is true for $n=k$. Differentiating the first term of $\frac{d^k}{dx^k} (\sin_p(x))$ using the chain rule, and only picking the first term of the resulting expression will give us
\[ (-1)^{k-1} (p-1)_{k-1} \left( (-1) (p-k) [p-(k+1), k (p-1)] \right) = (-1)^k (p-1)_{k-1} (p-1) [p-(k+1), k(p-1)].\]
The recursive nature of the falling factorial tells us that $(p-1)_{k-1} (p-1) = (p-1)_k$. This shows that the first term of $\frac{d^{k+1}}{dx^{k+1}} (\sin_p(x))$ is given by $(-1)^k(p-1)_k [p-(k+1), k(p-1)]$. By the principle of mathematical induction, the result is true for all $n \ge 1$.
\end{proof}
The connection to Stirling numbers and the successive derivatives of the $\sin_p(x)$ is now clear. Simplifying the coefficient of the first term of $\frac{d^n}{dx^n} (\sin_p(x))$ obtained from the above lemma gives:
\[ (-1)^{n-1} (p-1)_{n-1} = (-1)^{n-1} \frac{(p)_{n}}{p} = \frac{(-1)^{n-1}}{p} \sum_{k=1}^n s(n, k) p^k.\]
\begin{comment}
Building a tower from the coefficients in Example \ref{derivativeExample}, we get:
\begin{center}
1\\
0 \ \ \verb"|" \ \ \underline{1}\\
0 \ \ \verb"|" \ \ \underline{$-p+1$} \ \ \verb"|" \ \ 0\\
0 \ \ \verb"|" \ \ \underline{$p^2-3p+2$} \ \ \verb"|" \ \ $-p^2+2p-1$ \ \ \verb"|" \ \ 0
\end{center}
There is hardly a pattern here, but one that we did find was that the coefficients in the second column (underlined) can be expressed using Stirling numbers of the first kind $s(n,k)$. For instance, corresponding to the polynomial $p^2-3p+2$, we have $s(3, 3) = 1$, $s(3, 2) = -3$ and $s(3, 1) = 2$.
We can show this by using the recurrence relation for the second column. When thought of as a triangle, the bracket notation defined in Lemma \ref{bracketNotaiton} can be thought of as two operations in the triangle. The left term moves you to the left in the triangle, and the right term moves you to the right. Because of this, we use the right term once to get into the second column, then use the left term there on out to stay in the second column. Doing this gives us $1[p-1,0]$, then $(1-p)[p-2, p-1]$, then $(1-p)(2-p)[p-3, p-2]$ and so on. The coefficient is just a rising factorial that can be expressed in Pochhammer notation $(1-p)_{n-1}$ where $n$ is the number of derivatives. We can use a separate definition for rising factorials to get it in terms of Stirling numbers
\begin{align*}
(x)_n &= \sum_{k=0}^n{(-1)^{n-k}s(n,k)x^k}.
\end{align*}
Substituting in our rising factorial gives
\begin{align*}
\sum_{k=0}^{n-1}{(-1)^{n-k-1}s(n-1,k)(1-p)^k}.
\end{align*}
\end{comment}
\subsection{Newton's binomial series} Let $p$ be any integer that is greater than 1.
As the previous section demonstrates, finding a formula for the successive derivatives of $\sin_px$ to compute its Taylor series is complicated. Instead, we examine $\sin_p^{-1}x$, whose Taylor series at $x=0$ is more manageable, and use this to find the Taylor series of $\sin_px$ at $x= 0$ through the Lagrange inversion theorem. To do this, we apply Newton's binomial series to derive the Taylor series of $\sin_p^{-1}x$. Newton's binomial series tells us the following for any exponent $a$ and $|x| < 1$:
\begin{align*}
(1-x)^{-a}&=1+ax+\frac{a(a+1)}{2!}x^2+\frac{a(a+1)(a+2)}{3!}x^3+... \\
&=\sum_{k=0}^\infty\frac{a^{(k)}x^k}{k!},
\end{align*}
where $a^{(k)}=a(a+1)(a+2)\cdots(a+k-1)$ is the rising factorial \cite[p. 742]{JS}. Note that, by convention, $a^{(0)} = 1$.
\begin{prop} \label{taylor-arcsin}
We can express $\sin_p^{-1}x$ as the following Taylor series:
\begin{equation*}
\sin_p^{-1}x=\sum_{k=0}^\infty\left(\frac{p-1}{p}\right)^{(k)}\frac{x^{kp+1}}{k!(kp+1)}.
\end{equation*}
\end{prop}
\begin{proof}
Beginning with the integral form of $\sin_p^{-1}x$ derived in Section \ref{Inverse functions}, we apply Newton's binomial series:
\begin{align*}
\sin_p^{-1}x&=\int_0^x(1-t^p)^{-(\frac{p-1}{p})}dt \\
&=\int_0^x\left(\sum_{k=0}^\infty\left(\frac{p-1}{p}\right)^{(k)}\frac{t^{kp}}{k!} \right) dt.
\end{align*}
Power series have the property that they can be integrated term by term within the interval of convergence. Thus, when we integrate and apply the fundamental theorem of calculus, the result follows.
\end{proof}
\begin{example} \label{talorexample}
Applying Proposition \ref{taylor-arcsin} for $p=2$ gives the following well-known result:
\begin{equation*}
\sin_2^{-1}x=x+\frac{1}{6}x^3+\frac{3}{40}x^5+\frac{5}{112}x^7+ \cdots +{2n \choose n}\frac{x^{2n+1}}{2^{2n}(2n+1)}+ \cdots.
\end{equation*}
Similarly, when $p=4$, we get the first few terms as follows:
\begin{equation*}
\sin_4^{-1}x=x+\frac{3}{20}x^5+\frac{7}{96}x^9+\frac{77}{1664}x^{13}+ \cdots.
\end{equation*}
\end{example}
It would be helpful to have a closed-form solution for these higher derivatives. In the next section, we introduce some tools and discuss what this will look like.
\subsection{$\sin_p^{-1} x$ through the gamma function}
We now introduce a special function to shed light on $\sin_p^{-1} x$. The gamma function, $\Gamma(z)$, is defined as $\Gamma(z)=\int_{0}^{\infty} e^{-t}t^{z-1} \,dt$, for $z > 0$. This converges for any real number $z > 0$, and it
is an extension of the factorial function: $\Gamma(n) = (n-1)!$. It is well-known that $\Gamma(1/2) = \sqrt{\pi}$. Two important properties of the gamma function are
\[\Gamma(x+1) = x \Gamma(x) \ \ \text{ and } \ \ \Gamma(x) \Gamma(1-x) = \frac{\pi}{ \sin(\pi x)}.\]
\\
Using the gamma function, for any integer $p > 1$, we can further simplify the Taylor series for $\sin^{-1}_p(x)$ as follows.
We begin by the formula from Proposition \ref{taylor-arcsin} which states that
\begin{equation*}
\sin_p^{-1}x=\sum_{k=0}^\infty\left(\frac{p-1}{p}\right)^{(k)}\frac{x^{kp+1}}{k!(kp+1)}.
\end{equation*}
Then we have the following:
\begin{align*}
\sin_p^{-1}x & = \sum_{k=0}^\infty\left(\frac{p-1}{p}\right)^{(k)}\frac{x^{kp+1}}{k!(kp+1)} \\
& = \sum_{k=0}^\infty\left(1-\frac{1}{p}\right)^{(k)}\frac{x^{kp+1}}{k!(kp+1)} \\
& = \sum_{k=0}^\infty \left(1-\frac{1}{p}\right) \left(2-\frac{1}{p}\right)\left(3-\frac{1}{p}\right) \cdots \left(k-\frac{1}{p}\right)\frac{x^{kp+1}}{k!(kp+1)} \\
& = \sum_{k=0}^\infty\Gamma\left(1- \frac{1}{p}\right) \left(1-\frac{1}{p}\right) \left(2-\frac{1}{p}\right)\left(3-\frac{1}{p}\right) \cdots \left(k-\frac{1}{p}\right) \frac{1}{\Gamma(1- \frac{1}{p}) }\frac{x^{kp+1}}{k!(kp+1)} \\
& = \sum_{k=0}^\infty\Gamma\left(2- \frac{1}{p}\right) \left(2-\frac{1}{p}\right)\left(3-\frac{1}{p}\right) \cdots \left(k-\frac{1}{p}\right) \frac{1}{\Gamma(1- \frac{1}{p}) }\frac{x^{kp+1}}{k!(kp+1)} \\
& \ \ \vdots \\
& = \sum_{k=0}^\infty\Gamma\left(k- \frac{1}{p}\right) \frac{1}{\Gamma(1- \frac{1}{p}) }\frac{x^{kp+1}}{k!(kp+1)} \\
& = \sum_{k=0}^\infty \frac{\Gamma\left(k- \frac{1}{p}\right) }{\Gamma\left(1 - \frac{1}{p}\right)}\frac{x^{kp+1}}{k!(kp+1)}.
\end{align*}
\begin{thm} \label{thm1}
Let $n > 1$ be a positive integer. Then for any positive integer $l$, let $k$ and $r$ be the integers given by the division algorithm: $l = nk + r$ where $k \ge 0$ and $0 \le r \le n-1$. Then we have
\[ \left(\frac{d^l}{dx^l} \sin_n^{-1}(x) \,\right) \biggr \rvert_{x=0} = \begin{cases}
\frac{\Gamma\left(k- \frac{1}{n}\right) }{\Gamma\left(1 - \frac{1}{n}\right)}\frac{(kn)!}{k!}, & \text{ if } r = 1, \\
0, & \text{ if } r \ne 1.
\end{cases}
\]
\end{thm}
\begin{proof}
The Taylor series for $\sin^{-1}_n(x)$ at $x = 0$ has the form
\[\sin^{-1}_n(x) = h(x) = \sum_{m=0}^{\infty} \frac{h^{(m)}(0)}{m!} x^m. \]
On the other hand, from the above calculation, we know that
\[\sin^{-1}_n(x) = \sum_{k=0}^\infty \frac{\Gamma\left(k- \frac{1}{n}\right) }{\Gamma\left(1 - \frac{1}{n}\right)}\frac{x^{kn+1}}{k!(kn+1)} =
\sum_{k=0}^\infty \left(\frac{\Gamma\left(k- \frac{1}{n}\right) }{\Gamma\left(1 - \frac{1}{n}\right)}\frac{(kn)!}{k!} \right) \frac{1}{(kn+1)!} x^{kn+1}.\]
Equating the coefficients of like-powers of $x$ in both these series, we get the theorem.
\end{proof}
Now that we have derived the Taylor series of $\sin_p^{-1}x$, we can apply the Lagrange inversion theorem as outlined in the next section.
\subsection{Lagrange inversion}\label{lagrange} A function $z = f(w)$ is said to be analytic at $c$ if it is infinitely differentiable at $c$ and if the Taylor series for $f(w)$ at $w=c$ converges to $f(w)$ for all $w$ in a neighborhood of $c$.
For an equation $z=f(w)$, where $f$ is analytic at $c$ and $f'(c)\neq 0$, the Lagrange inversion theorem can be used to find the equation's inverse, $w=g(z)$, in a neighborhood of $0$. This inverse is given by the formula \cite[Chapter 3]{MA}:
\begin{align*}
g(z)&=c+\sum_{n=1}^\infty g_n\frac{(z-f(c))^n}{n!},\ \text{ where} \\
g_n&=\lim_{w\to c}\frac{d^{n-1}}{dw^{n-1}} \left[\left(\frac{w-c}{f(w)-f(c)}\right)^n \right].
\end{align*}
\begin{comment}
\begin{example}
Using this theorem, we can invert the Lambert $W$ function, $f(w)=we^w$. If we let $c=0$, we meet all of the conditions given above and can apply the theorem as follows:
\begin{align*}
g(z)&=\sum_{n=1}^\infty g_n\cdot\frac{z^n}{n!} \\
g_n&=\lim_{w\to 0}\frac{d^{n-1}}{dw^{n-1}}(e^{-nw})
\end{align*}
After taking the first few derivatives of $e^{-nw}$, we are able to create a formula for the $(n-1)$th derivative:
\begin{align*}
\frac{d^{n-1}}{dw^{n-1}}(e^{-nw})&=(-n)^{n-1}e^{-nw}
\end{align*}
Substituting this into the equation for $g_n$ above and returning to the equation for $g(z)$, we find
\begin{align*}
g_n&=\lim_{w\to 0}(-n)^{n-1}e^{-nw} \\
&=(-n)^{n-1} \\
g(z)&=\sum_{n=1}^\infty(-n)^{n-1}\frac{z^n}{n!}.
\end{align*}
\end{example}
\end{comment}
For power series, this theorem takes a slightly different form. Specifically, when $f$ and $g$ are formal power series expressed as
\begin{equation*}
f(w)=\sum_{k=0}^\infty f_k\frac{w^k}{k!}\quad \text{and}\quad g(z)=\sum_{k=0}^\infty g_k\frac{z^k}{k!},
\end{equation*}
with $f_0=0$ and $f_1\neq 0$, applying the Lagrange inversion theorem gives us the following \cite{MA}:
\begin{align*}
&g(z)=c+\sum_{n=1}^\infty g_n\frac{(z-f(c))^n}{n!},\quad \text{with} \\
&g_n=\frac{1}{f_1^n}\sum_{k=1}^{n-1}(-1)^kn^{(k)}B_{n-1,k}(\hat{f}_1,\hat{f}_2,...,\hat{f}_{n-k}),\quad n\geq2,\quad \text{where} \\
&\hat{f}_k=\frac{f_{k+1}}{(k+1)f_1}, \quad g_1=\frac{1}{f_1}, \quad n^{(k)}=n(n+1)\cdots(n+k-1)\text{, and} \\
B_{n,k}(x_1,&x_2,...,x_{n-k+1})= \sum_ {} \frac{n!}{j_1!j_2!...j_{n-k+1}!}\left ( \frac{x_1}{1!} \right )^{j_1} \left ( \frac{x_2}{2!} \right )^{j_2} ... \left ( \frac{x_{n-k+1}}{(n-k+1)!} \right )^{j_{n-k+1}},
\end{align*}
where this sum is taken over all sequences $j_1,j_2,j_3,..., j_{n-k+1}$ of non-negative integers that satisfy $j_1+j_2+...+j_{n-k+1}=k$ and
$j_1+2j_2+3j_3+...+(n-k+1)j_{n-k+1}=n$. These are the Bell polynomials.
The Taylor series expansion of $\sin_p x$ is obtained when the above theorem is applied to
\begin{equation*}
\sin_p^{-1}x =\sum_{k=0}^\infty \left(\frac{p-1}{p}\right)^{(k)} \frac{x^{kp+1}}{k!(kp+1)},
\end{equation*}
which was derived in the previous section. We are able to apply this theorem to $\sin_p^{-1}x$, as it meets the initial conditions given above: $f_0=0$ and $f_1\neq 0$.
\begin{example}
When $p=2$, we can apply Lagrange Inversion Theorem with $c=0$, as $f(c)=0$ and $f'(c)=1$.
To do so, we must calculate $f_k$ for the first few terms. Expanding $\sin_2^{-1}x$, we find
\begin{equation*}
f_0=0,\ f_1=1,\ f_2=0,\ f_3=1,\ f_4=0,\ f_5=9,\ f_6=0.
\end{equation*}
Using these values, we can find $\hat{f_k}$:
\begin{equation*}
\hat{f_1}=0, \ \hat{f}_2=\frac{1}{3},\ \hat{f}_3=0,\ \hat{f}_4=\frac{9}{5},\ \hat{f}_5=0.
\end{equation*}
We may now use these values to find the first few $g_n$ using the formulas given above. To this end, we record a couple of special Bell polynomials that will be used below: $B_{n, n}(x_1) = (x_1)^n$ and $B_{n, n-1} (x_1, x_2) = {n \choose 2} (x_1)^{n-2}x_2$. These are obtained by simplifying the general Bell polynomial given above.
When $n=1$, $g_1=\frac{1}{f_1} = \frac{1}{1}=1$. When $n=2$, we have $g_2 = (-1)^1\cdot2^{(1)}B_{1,1}(0)=0$. Similarly, when $n=3$, we have
\begin{eqnarray*}
g_3 & = & \frac{1}{f_1^3} \left( (-1)^1 3^{(1)} B_{2,1} (\hat{f_1}, \hat{f_2}) + (-1)^2 3^{(2)} B_{2, 2}(\hat{f_1})\right) \\
& = & \frac{1}{1^3} \left( -3 B_{2,1} (0, 1/3) + 12 B_{2, 2}(0)\right) \\
& = & -3 {2 \choose 2} \frac{1}{3} + 12 (0^2) = -1.
\end{eqnarray*}
In the same manner, applying this formula to the next few values of $n$, we find that $g_4=0$ and $g_5=1$.
Substituting these values into the formula for $g(z)$ given by Lagrange Inversion Theorem above, we have:
\begin{align*}
g(z)&=0+\sum_{n=1}^\infty g_n \frac{(z-0)^n}{n!}. \\
\sin_2(z) &=z-\frac{z^3}{3!}+\frac{z^5}{5!}+\cdots.
\end{align*}
When $p=4$, these computations get more tedious. Using SageMath, we find that
\begin{equation*}
\sin_4x=x-\frac{18}{5!}x^5+\frac{14364}{9!}x^9- \cdots.
\end{equation*}
\end{example}
The above ideas prove the following theorem.
\begin{thm} \label{thm:analytic}
For any integer $p > 1$, the functions $\sin_p^{-1} x$ and $\sin_p x$ are analytic at $x = 0$.
\end{thm}
It is well-known that $\sin x/x \rightarrow 1$ as $x \rightarrow 0$. We now generalize this result.
\begin{cor}
Let $p >1$ be an integer. Then we have
\[ \lim_{x \rightarrow 0} \frac{\sin_p x}{x} = 1.\]
\end{cor}
\begin{proof}
By Theorem \ref{thm:analytic}, we know that $\sin_p x$ is analytic at $x = 0$, and moreover, from the CIVP, $\sin_p 0 = 0$. Therefore, we can express $\sin_p x$ as a power series whose constant term is $0$:
\[\sin_p x = a_1x + a_2x^2 + a_3 x^3 + \cdots + a_nx^n+ \cdots. \]
Differentiating both sides and invoking the CIVP gives:
\[(\cos_p x)^{p-1} = a_1+2a_2x + 3a_3x^2 + \cdots + nx^{n-1} + \cdots. \]
Since $\cos_p(0) =1$, setting $x=0$ in the above equation tells us that $a_1 = 1$.
Finally, we have
\[\lim_{x \rightarrow 0} \, \frac{\sin_p x}{x} = \lim_{x \rightarrow 0} \, \frac{x+a_2x^2 + a_3x^3 + \cdots}{x} = \lim_{x \rightarrow 0} \, 1+a_2x +a_3x^2 + \cdots = 1.\]
\end{proof}
Note that from this result it also follows that $\tan_p x/x \rightarrow 1$ as $x \rightarrow 0$. One can also prove these limits using l'H\^{o}pital's rule. In the same vein, one can also show the following.
\begin{cor}
\[ \lim_{x \rightarrow 0} \frac{\sin_4 x - x}{x^5} = -\frac{18}{5!}.\]
\end{cor}
\subsection{Rigidity} Note that the missing terms of the Taylor series for $\sin_4(x)$ are exactly the ones that were also missing in $\sin_4^{-1}(x)$; see Example \ref{talorexample}. In fact, for both functions, the non-zero terms in the Taylor series correspond to powers of $x$ that form an arithmetic progression of the form $4m+1$. We proved this fact in Theorem \ref{thm1} for $\sin_n^{-1}(x)$. We now conjecture that this is also true for $\sin_n(x)$.
\begin{conjecture} \label{conj}
Let $n$ be a positive integer. Then
\[ \left(\frac{d^l}{dx^l} \sin_n (x) \right) \biggr \rvert_{x=0} \ne 0 \iff l \equiv 1 \mod n. \]
\end{conjecture}
This led to the following, more general question in analysis.
\vskip 3mm
\noindent
\textbf{Question:} Suppose $f(x)$ is a real-valued function that is infinitely differentiable at $x=a$ such that $f’(a) \ne 0$. Let $f(a) = b$ and let $g(x)$ be the local inverse of $f(x)$ (this exists by the inverse function theorem) at $x = a$. Is it true that for every positive integer $n$, the $n$th derivative of $f(x)$ at $x=a$ is non-zero if and only if the nth derivative of $g(x)$ at $x=b$ is non-zero?
\vskip 3mm
It turns out that, in general, the above answer is no. Take for example $f(x)=x^2$. We have $f(1) = 1$ and $f'(1) = 2 \ne 0$. At $x=1$, the local inverse of $f(x)$ is $g(x) = \sqrt{x}$. Note that for all $k \ge 3$, $f^{(k)}(1) = 0$ but $g^{(k)}(1) \ne 0$. On the other hand, for the function $f(x) = \sin(x)$, the above question has an affirmative answer because the Taylor series for $\sin x$ and $\sin^{-1} x$, have only odd terms. This leads naturally to the following definition.
\vskip 3mm
\begin{defn}
Let $y=f(x)$ be a function that is infinitely differentiable at $x=a$ such that $f’(a) \ne 0$. We say that $f(x)$ is rigid at $x=a$ if for any positive integer $k$, $f^{(k)}(a) \ne 0$ if and only if $g^{(k)} (b) \ne 0$, where $g(x)$ is the local inverse of $f(x)$ at $x = a$ and $b = f(a)$.
\end{defn}
In this terminology, $f(x) =x^2$ is not rigid at $x=1$ but $f(x) = \sin x$ is rigid at $x=0$. Conjecture \ref{conj} can now be restated as follows. For any positive integer $k$, $\sin_k x$ is rigid at $x=0$.
\vskip 3mm
\noindent
\textbf{Question:} What are necessary and sufficient conditions for a function $y= f(x)$ that is infinitely differentiable at $x=a$ to be rigid at $a$?
\begin{comment}
\subsection{Associahedra}
Lagrange inversion theorem can be formulated in geometric terms as follows.
The associahedron $K_n$ is a convex $(n-2)$-polytope where distinct
vertices correspond to distinct parenthetical groupings of $n$ symbols, and edges are drawn between vertices if they are obtainable through applying the associative law once; see figure \ref{associahedron} for $K_5$. Let $z = f(w)$ be a function that is analytic at $w = 0$ such that $f(0) = 0$ and $f'(0) = 1$. Then, given the formal power series $f(w) =
\sum_{k=1}^\infty f_k \frac{w^k}{k!}$ with $f_0 = 0$,
$f_1 = 1$, its local inverse at $0$ is given by $g(z) = \sum_{k=0}^\infty g_k \frac{z^k}{k!}$, where
\begin{equation*}
g_n = \sum_{F \ \text{face of } K_n} (-1)^{n-\dim F} f_F,
\end{equation*}
and $f_F = f_{i_1} \cdots f_{i_m}$ for any face
$F = K_{i_1} \times \cdots \times K_{i_m}$ of $K_n$.
\begin{figure}[H] \label{associahedron}
\centering
\includegraphics[width=0.35\textwidth]{Associahedron_K5_front.png}
\caption{Associahedron $K_5$ \cite{JN}}
\label{fig:associahedra}
\end{figure}
\end{comment}
\section{Generalized $\pi$ values} \label{pi_p}
\subsection{Organic definition}
As we generalize trigonometric functions in the $p$-norm, we must also take into consideration generalizing the value of $\pi$. Recall that $\pi =2\sin^{-1}(1)$. Using this as our inspiration, we can organically define $\pi_p$ as $\pi_p:=2\sin_p^{-1}(1)$. Using our $\sin_p^{-1}x$ formula we derived in Section \ref{Inverse functions} and letting $x=1$, we get
\begin{equation}\label{pi_peq}
\pi_p= 2\int_{0}^{1}{ \frac{1}{(1-t^p)^{\frac{p-1}{p}}}} \,dt.
\end{equation}
Note that, unless otherwise indicated, when we refer to $\pi$, we are referring to $\pi_2$.
When $p=2$, we find that the area of the unit circle is equal to $\pi$. It is then natural to wonder if $\pi_p$ has any relation to the area of a unit $p$-circle.
\begin{prop}
The area of a unit $p$-circle is $\pi_p$, when $p \ge 1$.
\end{prop}
\begin{proof}
In Proposition \ref{areapara}, we found that the area of the sector of the $p$-circle that connects the points $(x,y)$ and $(1,0)$ is given as a function of $y$ by $a(y)=\frac{1}{2}\int_{0}^{y}(1-t^p)^{\frac{1}{p}-1} \,dt$. If we let $(x,y)$ be the point $(0,1)$, we get the area of the unit $p$-circle in the first quadrant, given by $a(1)=\frac{1}{2}\int_{0}^{1}(1-t^p)^{\frac{1}{p}-1} \,dt$. Since the unit $p$-circle has 4-fold symmetry, we can multiply both sides of the equation by four to find the area of the entire $p$-circle:
\begin{equation*}
4a(1)=2\int_{0}^{1}(1-t^p)^{\frac{1}{p}-1} \,dt.
\end{equation*}
From Section \ref{Inverse functions}, we know that $2\int_{0}^{x}(1-t^p)^{\frac{1}{p}-1} \,dt=2\sin_p^{-1}x$. Therefore, we know that the right hand side of the equation is $2\sin_p^{-1}(1)$, which is equal to $\pi_p$ as shown in Equation (\ref{pi_peq}). We have also already established that $a(1)$ is the area of the quarter unit $p$-circle, so $4\cdot a(1)$ gives us the area of the entire unit $p$-circle. Therefore, we find that the area of the unit $p$-circle is $\pi_p$.
\end{proof}
\begin{cor}\label{0>pi_p>4}
For any $p \ge 1$, we have $2 \le \pi_p <4$.
\end{cor}
\begin{proof}
As shown above, $\pi_p$ is the area of a unit $p$-circle. When $p = 1$, we get the region bounded by the square $|x| + |y| = 1$, which has area $2$.
Similarly, since the $p$-circle is inscribed in a square of side length two, we know that the area of the $p$-circle is bounded by the square's area, which is 4. This shows that for any $p \ge 1$, we have $2 \le \pi_p<4$.
\end{proof}
\subsection{A formula for $\pi_p$}
We now show how we can compute $\pi_p$ in terms of the gamma function. To this end, we need another special function called
the beta function, $\beta(x,y)$, which is closely related to the gamma function and can be defined as $\beta(x,y)=\int_{0}^{1} t^{x-1}(1-t)^{y-1} \,dt$, for any two real numbers $x,y$ such that $x>0$ and $y>0$. We can put the beta function in terms of gamma using the property
\begin{equation}\label{BetaGamma}
\beta(x,y)=\frac{\Gamma(x)\Gamma(y)}{\Gamma(x+y)}.
\end{equation}
\begin{prop} For any $p \ge 1$, we have
\begin{equation*}
\pi_p =\frac{2\Gamma^2(\frac{1}{p})}{p\Gamma(\frac{2}{p})}.
\end{equation*}
In particular, $\pi_p$ is a differentiable function of $p$.
\end{prop}
\begin{proof}
Referring back to our definition for $\pi_p$, if we let $u=t^p$, we can express $\pi_p$ in terms of the beta function as follows:
\begin{equation*}
\pi_p= 2\int_{0}^{1}{ \frac{1}{(1-t^p)^{\frac{p-1}{p}}}} \,dt = \frac{2}{p} \int_{0}^{1} (1-u)^{\frac{1}{p}-1}\cdot u^{\frac{1}{p}-1}\,du = 2 \beta(1/p, 1/p).
\end{equation*}
We can then use Equation \eqref{BetaGamma} to put $\pi_p$ in terms of the gamma function, and that gives the formula stated in the proposition.
Since $\Gamma(x)$ is a differentiable function and compositions and quotients of differentiable functions are again differentiable, it follows that $\pi_p$ is differentiable.
\end{proof}
\begin{example} \label{pi_p:2,3,4}
Using the above equation, we can numerically approximate $\pi_p$ for any $p$. For $p=2, 3$ and $4$, we get:
\begin{equation*}
\pi_2=\frac{2\Gamma^2(\frac{1}{2})}{2\Gamma(1)}\approx 3.1415,
\quad
\pi_3=\frac{2\Gamma^2(\frac{1}{3})}{3\Gamma(\frac{2}{3})}\approx3.533
\quad \text{and}\quad
\pi_4=\frac{2\Gamma^2(\frac{1}{4})}{4\Gamma(\frac{2}{4})}\approx3.708.
\end{equation*}
\end{example}
\begin{comment}
Using Equation \ref{Gammapi_p}, we can derive a formula for $\Gamma \left (\frac{1}{2^n} \right )$ in terms of $\pi_p$. We need to use the gamma identity, $\Gamma(x) \Gamma(1-x) = \frac{\pi}{ \sin(\pi x)}$, twice, first letting $x=\frac{1}{p}$ and then letting $x=\frac{1}{2^n}$.
We are then able to solve for $\Gamma\left(\frac{1}{2^n}\right)$ after substituting.
First let $x=\frac{1}{p}$ and solve for $\Gamma \left (\frac{1}{p}\right )$ and then let $x=\frac{1}{2^n}$ and solve in terms of $\pi$.
\begin{equation*}
\Gamma \left (\frac{1}{p}\right )=\frac{\pi}{\sin(\frac{\pi}{p})\Gamma(1-\frac{1}{p})} \ \ \ \ \text{and} \ \ \ \ \pi=\Gamma \left (\frac{1}{2^n} \right)\Gamma \left(1-\frac{1}{2^n} \right)\sin \left(\frac{\pi}{2^n}\right)
\end{equation*}
Substituting our new $\pi$ equation into our $\Gamma \left (\frac{1}{p}\right)$ equation, we find $\Gamma \left (\frac{1}{p}\right)=\frac{\Gamma\left (\frac{1}{2^n} \right )\Gamma \left (1-\frac{1}{2^n}\right)\sin \left (\frac{\pi}{2^n}\right)}{\sin \left(\frac{\pi}{p}\right)\Gamma \left (1-\frac{1}{p} \right)}.$
Substituting this into Equation \ref{Gammapi_p} and simplifying, we find
\begin{equation*}
\Gamma \left (\frac{1}{2^n}\right)=\frac{\sin \left(\frac{\pi}{p}\right )\Gamma \left (1-\frac{1}{p}\right)\sqrt{p\cdot \pi_p\Gamma \left(\frac{2}{p}\right)}}{\Gamma \left (1-
\frac{1}{2^n}\right)\sin \left (\frac{\pi}{2^n}\right)\sqrt{2}}.
\end{equation*}
\begin{example}
Using the above equation, we can find the value of $\Gamma(\frac{1}{2})$ in terms of $\pi_4$ by letting $n=1$ and $p=4$.
\begin{equation*}
\Gamma \left (\frac{1}{2}\right)=\frac{\sin \left(\frac{\pi}{4}\right )\Gamma \left (1-\frac{1}{4}\right)\sqrt{4\cdot \pi_4\Gamma \left(\frac{2}{4}\right)}}{\Gamma \left (1-
\frac{1}{2}\right)\sin \left (\frac{\pi}{2}\right)\sqrt{2}}.
\end{equation*}
Solving for the right-hand side of the equation, we find that it is equivalent to $\sqrt{\pi}$.
\end{example}
\end{comment}
\subsection{Properties of $\pi_p$}
We have already seen that $\pi_p$ is a differentiable function of $p$ for all $p > 0$. Is it a monotonic function?
Example \ref{pi_p:2,3,4} suggests that $\pi_p$ increases with $p$. We now prove that fact.
\begin{prop}
$\pi_p$ is an increasing function on $(0, \infty)$.
\end{prop}
\begin{proof}
Recall that $\pi_p$ is the area of a unit $p$-circle. Since $p$-circles have a 4-fold symmetry, we get $\pi_p= 4\int_{0}^{1}{(1-x^p)^{\frac{1}{p}}} \,dt$. We will be done if we can show that, for any fixed value of $x$ in $(0, 1)$, $(1-x^p)^{\frac{1}{p}}$ is an increasing function in $p$. This is because if
$(1-x^{p_1})^{\frac{1}{p_1}} < (1-x^{p_2})^{\frac{1}{p_2}}$ for all $x$ in $(0, 1)$ and $p_1 < p_2$, then
\[ 4 \int_0^1 (1-x^{p_1})^{\frac{1}{p_1}} \, dx < 4 \int_0^1 (1-x^{p_2})^{\frac{1}{p_2}} \, dx,\]
showing that $\pi_{p_1} < \pi_{p_2}$ whenever
$0 < p_1 < p_2$.
To this end, let $\psi(p):=(1-x^p)^{\frac{1}{p}}$, where $x$ is a fixed number in $(0,1)$. Taking the natural logarithm and differentiating with respect to $p$ on both sides, we get
\begin{eqnarray*}
\ln(\psi(p)) &=& \frac{\ln(1-x^p)}{p}, \\
\frac{\psi'(p)}{\psi(p)} & =& \frac{-\ln(1-x^p)}{p^2}+\frac{-\ln(x)x^p}{p(1-x^p)}, \\
\psi'(p) & =& (1-x^p)^{\frac{1}{p}} \left( \frac{-\ln(1-x^p)}{p^2}+\frac{-\ln(x)x^p}{p(1-x^p)} \right).
\end{eqnarray*}
For $0 < x < 1$ and $p >0$, note that $0 < 1-x^p < 1$. Therefore, $\ln(x)$ and $\ln(1-x^p)$ are both negative. This shows that all parts of the derivative are positive. Therefore, $\psi'(p) > 0$, which means $\psi(p)$ is an increasing function.
\end{proof}
Having shown above that $\pi_p$ is an increasing function and that it has an upper bound of 4 in Corollary \ref{0>pi_p>4}, we know that a limit exists. It is then only natural to wonder what the limit of $\pi_p$ is.
\begin{prop}
$\lim_{p\to\infty} \pi_p=4$.
\end{prop}
\begin{proof}
Using $\pi_p=\frac{2\Gamma^2(\frac{1}{p})}{p\Gamma(\frac{2}{p})}$, we can take the limit of $\pi_p$ as $p$ approaches infinity. Note that we have not yet stated Legendre's duplication formula,
${\Gamma(2z)}=\frac{\Gamma(z)\Gamma(z+\frac{1}{2})}{2^{1-2z}\sqrt{\pi}}$.
\begin{align*}
\pi_p&=\frac{2}{p}\cdot\frac{\Gamma^2(\frac{1}{p})}{\Gamma(\frac{2}{p})} \\
&=\frac{2}{p}\cdot\frac{\Gamma^2(\frac{1}{p})\cdot2^{1-\frac{2}{p}}\sqrt{\pi}}{\Gamma(\frac{1}{p})\Gamma\left(\frac{1}{p}+\frac{1}{2}\right)} \\
&=\frac{2}{p}\cdot\frac{\Gamma(\frac{1}{p})\cdot2^{1-\frac{2}{p}}\sqrt{\pi}}{\Gamma\left(\frac{1}{p}+\frac{1}{2}\right)} \\
&=\frac{2\cdot \Gamma(\frac{1}{p}+1)\cdot2^{1-\frac{2}{p}}\sqrt{\pi}}{\Gamma\left(\frac{1}{p}+\frac{1}{2}\right)}.
\end{align*}
\[ \lim_{p\to\infty} \pi_p=\frac{2\cdot \Gamma(1)\cdot2\sqrt{\pi}}{\Gamma(\frac{1}{2})}
=\frac{ 1 \cdot 4\sqrt{\pi}}{\sqrt{\pi}}
=4. \]
\end{proof}
\begin{figure}[H]
\centering
\includegraphics[width=0.4\textwidth]{pip.png}
\caption{Graph of $\pi_p$ }
\label{fig:pi_p}
\end{figure}
Recall that it is well-known that $\pi_2 = \pi$ is an irrational number (a number that is not the ratio of two integers). On the other hand, $\pi_1 = 2$, a rational number. ($\pi_1 =2$ because it is the area enclosed by the square $|x| +|y| = 1$ of side length $\sqrt{2}$.) It is natural to ask for what values of $p$, is $\pi_p$ irrational? This is a hard question. Since $\pi_p$ is a continuous function, it takes rational and irrational values infinitely often; see Figure \ref{fig:pi_p}.
\begin{comment}
However, when $\pi_p$ is irrational, we can get rational approximations for it, similar to $\pi \approx 22/7$. This is done by taking the partial sums of the following series whose terms are rational numbers when $p$ is a positive integer.
\begin{prop}
For any $p > 1$, we have $$\pi_p=2\sum_{k=0}^{\infty} \frac{\left ( \frac{p-1}{p} \right)^{(k)}}{k!(kp+1)}.$$
\end{prop}
\begin{proof}
This is obtained by setting $x=1$ in the Taylor series for $\sin_p^{-1}(x)$ from Section \ref{lagrange}.
\end{proof}
\begin{example}
When evaluating this series at $p=4$, we find that $\pi_4 \approx 3.708$ \end{example}
\end{comment}
\begin{comment}
\begin{equation*}
\pi_4=2 \left (1+\frac{3}{20}+\frac{7}{96}+\frac{77}{1664}+... \right)
\end{equation*}
\end{comment}
\subsection{$\pi_p$ with the Monte Carlo method}
Since we have shown that $\pi_p$ can be described as the area of a $p$-circle, we can use a rather fun technique to approximate the value of $\pi_p$. Given a $p$-circle shaped dartboard inscribed inside a square, what is the probability that a uniformly random throw will land on the dartboard (assuming that the dart must land inside the square)? The probability is the ratio of the area of the board to the area of the box. Therefore, if we have $n$ throws where $t$ of them land on the dartboard, the probability would be $\frac{t}{n}=\frac{\pi_p}{4}$. We can solve for $\pi_p$ to get $\pi_p=\frac{4t}{n}$. Because of the law of large numbers, when $n\to\infty$, the ratio goes to the true ratio, and we find the true value of $\pi_p$. Writing a simple program to do this for us, at $n=1,000,000,000$, we get $\pi_3 \approx 3.53324$ and $\pi_4 \approx 3.7081$.
\section{Optimal unit $p$-circles} \label{optimal}
One question that naturally arises when examining $p$-circles is, ``At what value of $p$ is the corresponding squircle halfway between a unit circle ($p=2$) and a square ($p \to\infty$)?" This question was examined from three lenses: area, perimeter, and curvature.
\subsection{Area}
We sought to find the value of $p$ for which the area enclosed by the $p$-circle $|x|^p+|y|^p=1$ is $\frac{\pi+4}{2}$, which is the average of the areas of the unit circle and the square that the unit circle is inscribed in. Because the $p$-circle is symmetric, we can examine the first quadrant only, resulting in the following equation:
\begin{equation*}
\int_0^1\sqrt[p]{1-x^p}\ dx=\frac{\pi+4}{8}.
\end{equation*}
\begin{comment}
This equation can be solved by applying the $u$-substitution $u=x^p$ on the left side of the equation, putting the resulting integral in terms of the beta function, and applying Equation \ref{BetaGamma}. This results in the following:
\begin{equation*}
\frac{\Gamma\left(\frac{1}{p}+1\right)^2}{\Gamma\left(\frac{p+2}{p}\right)}=\frac{\pi+4}{8}
\end{equation*}
\end{comment}
Using SageMath, the root of this equation can be found, giving the approximation $p\approx 3.162038$. As such, we can conclude that the value of $p$ for which the area of the $p$-circle is exactly halfway between the areas of the unit circle ($p=2$) and the square in which it is inscribed is $p\approx 3.162038$.
\subsection{Perimeter}
\begin{comment}
With the changing shape of a $p$-circle, it is intuitive to wonder at what value of $p$ is a squircle exactly halfway between a circle and a square? We had already shown that when $p\approx3.162$, the area of the unit squircle is approximately halfway between the areas of a unit circle and a square with side length 2.
\end{comment}
Next, we want to find the value of $p$ for which the perimeter of a unit $p$-circle is halfway between those of a unit circle and the square that the unit circle is inscribed in. The circumference of a unit circle is $2\pi$, and the perimeter of a square that contains the unit circle is 8. Therefore, we have to find the value of $p$ for which the perimeter of a unit $p$-circle is $\pi+4$.
To find the perimeter of the unit $p$-cirle, we apply the Euclidean arc length formula to the defining equation of a $p$-circle. We equate the resulting integral to $\pi+4$ to obtain the equation:
\[
\pi+4 = 4\int_{0}^{1} {\sqrt{1+(1-x^p)^{2(1-p)/p} x^{2(p-1)}} \; dx.}
\]
We solved this equation numerically using SageMath to find that $p \approx 4.667489$.
\subsection{Curvature} Finally, we wish to find $p$ such that the curvature of the unit $p$-circle is halfway between that of a square (here said to have curvature 0) and the $2$-unit circle (which has curvature $1$).
For a given smooth curve $C$ in $\mathbb{R}^2$, the curvature is
a measure of how different our curve is from a circle at a given point. While there are many equivalent
formulations of the curvature of a given curve, the following gives the curvature for a curve defined implicitly by
$F(x,y) = 0$:
\begin{align*}
\kappa = \frac{|F_y^2F_{xx} - 2F_xF_yF_{xy} + F_x^2F_{yy}|}{(F_x^2+F_y^2)^{\frac{3}{2}}}.
\end{align*}
Using the relation $F(x, y) = x^p + y^p - 1 = 0$ for the unit $p$-circle, we obtain
$$\kappa = (p-1)\frac{x^py^{2p}+x^{2p}y^p}{(x^{2p}y^2+y^{2p}x^2)^{\frac{3}{2}}}(xy).$$
If we investigate the curvature of the unit $p$-circle at the point $x = y$, we find that
\begin{align*}
\kappa &= (p-1)x^2 \frac{2x^{3p}}{(2x^{2p+2})^{\frac{3}{2}}} = \frac{p-1}{\sqrt{2}x}.
\end{align*}
When $x = y$, we can write the relation for the unit circle as $2x^p = 1$, which gives $x = 2^{-\frac{1}{p}}$. Substituting for $x$ gives
\begin{align*}
\kappa &= \frac{p-1}{\sqrt{2} \cdot 2^{-\frac{1}{p}}} = \frac{p-1}{2^{\frac{1}{2}-\frac{1}{p}}} = (p-1)2^{\frac{1}{p}- \frac{1}{2}}.
\end{align*}
Therefore, if we solve for $p$ such that the unit $p$-circle has curvature 1/2, we find that $p \approx 1.43643264$.
\subsection{Resulting graphs}
Graphing these 3 $p$-circles gives Figure \ref{fig:opgraphs} where the unit circle and square are dashed, and the $p$-circle is solid. For the optimal curvature, we also have $p=1$ since both $p=1$ and $p\to\infty$ have the same curvature.
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{optimal3Dashed.png}
\caption{All 3 optimal $p$-circles}
\label{fig:opgraphs}
\end{figure}
\section{Rational Points on $p$-circles} \label{rational}
\subsection{2-circles and Pythagorean triples}
Right triangles (and as a result, Pythagorean triples) have long been objects of mathematical interest, studied intensely by the Babylonians even more than a thousand years before Pythagoras \cite{NO}.
Given any Pythagorean triple $(x,y,z)$ satisfying $x^2 + y^2 = z^2$, we may divide all parts by $z^2$ to obtain
$\frac{x^2}{z^2} + \frac{y^2}{z^2} = 1$. Then the point $(\frac{x}{z}, \frac{y}{z})$ is a rational point which lies on the $2$-unit circle defined by $x^2 + y^2 = 1$. On the other hand, given any rational number $\frac{v}{u}$, we obtain
the Pythagorean triple $(u^2-v^2, 2uv, u^2 + v^2)$ \cite{SJ}. In this manner, we may translate contexts between rational points on the unit circle and right triangles with integer side lengths.
\subsection{$p$-circles and Fermat's Last Theorem}
We can generalize the known results for rational points on $2$-circles and ask the same question for $p$-circles where $p$ is an integer greater than 2. There are certainly 4 trivial rational points along the axes of the graph, which are the points $(0,1), (1,0), (0,-1), (-1,0)$. To find the others, we may look at the rational solutions in the first quadrant and use symmetry to extend our answers to the entire unit $p$-circle.
Let $p$ be an integer greater than 2 and let $P = (\frac{p_1}{q_1},\frac{p_2}{q_2}) \in \mathbb{Q}^2$ be a rational point on the unit $p$-circle lying in the first quadrant. As $P$ lies on the unit $p$-circle and $P$ is in the first quadrant, we must have $(\frac{p_1}{q_1})^p + (\frac{p_2}{q_2})^p = 1$ and $\frac{p_1}{q_1}>0, \frac{p_2}{q_2}>0$. We then find
$(p_1q_2)^p + (p_2q_1)^p = (q_1q_2)^p$. However, by Fermat's Last Theorem, there are no positive integers $p_1, p_2, q_1, q_2$ that satisfy this relation. Thus, there exist no rational solutions in the first quadrant. Then, by symmetry, we see that the only rational points on the circle are exactly those along the axes.
\section{Future Research} \label{conclusion}
The results in this paper seem to indicate that $p$-trigonometric functions have interesting but complex behavior. For instance, even basic formulas such as the double-angle formulas for $\sin (2x)$ and $\cos (2x)$ seem to have no straightforward generalization. Similarly, understanding higher derivatives of $\sin_p(x)$ at $x=0$ looks very difficult; see Conjecture \ref{conj} and the questions following it. There are several other open questions. We list a few that we think merit further study.
\begin{enumerate}
\item We know the derivatives of $\sin_p x$ and $\cos_p x$. What about $ \int \sin_p x \, dx $ and $\int \cos_p x \, dx$? Using the Taylor series for $\sin_p x$ and $\cos_p x$, one can evaluate these integrals as series. But are there closed-form answers for these integrals?
\item The parametrization of $p$-circles we considered in this paper are with respect to area. We can also parametrize these curves with respect to arc length. These give yet another generalization of the $p$-trigonometric functions. What properties do these functions have?
\item Can we extend this work for $(p, q)$-trigonometric functions that come from looking at the curves $|x|^p + |y|^q = 1$? Parametrizing these curves will give us $\sin_{p, q} x$ and $\cos_{p, q} x$. What can be said about these functions?
\item So far, we have been working in $\mathbb{R}^2$. Can we extend this work to $\mathbb{R}^3$? To this end, we should look at the unit $p$-sphere $|x|^p + |y|^p +|z|^p= 1$. For $p=2$, this is the standard unit sphere, and as $p$ goes to infinity, we get a cube that encloses the unit sphere. These surfaces can be called sphubes ($p$-spheres), analogous to our squircles ($p$-circles). It opens gates to a whole new area of research. What are the parametric equations of these surfaces? Can we do sphubical trigonometry that is similar to spherical trigonometry? What are the volume and surface areas of the regions enclosed by these surfaces? What is the Gaussian curvature function of these surfaces?
\end{enumerate}
\begingroup
\raggedright
\bibliographystyle{plain}
|
\section{Introduction}
There are holographic representations of AdS that are so precisely
defined that in principle one could program a quantum computer
to simulate and test them.
By contrast there is no such completely concrete\footnote{By concrete I mean that it is sufficiently definite that it could be programmed into a quantum computer.} representation of de Sitter space. Not only do we not have an example; we don't even know the rules. My purpose in this paper is to briefly lay out some principles, and then to construct a well-defined example that realizes those principles.
The framework is static-patch (SP) holography \cite{Susskind:2021omt}\cite{Susskind:2021dfc}. The assumption is that there exists a unitary Hamiltonian quantum mechanics of a de Sitter static-patch in which the degrees of freedom are located on the stretched horizon. We begin by considering the roles of entanglement, chaos, and complexity, and derive necessary requirements---very different from those for AdS---for a quantum system to be dual to de Sitter space. We will see that these requirements are met by a non-standard, but perfectly definite limit of the SYK system. Quantum-simulating the system should be no harder than simulating SYK in the usual range of parameters.
\subsection{Four Conjectures}
The paper revolves around four conjectures:
\begin{enumerate}
\item There exists a de Sitter generalization (which I will describe) of the RT and HRT equations for entanglement entropy in which the AdS boundary is replaced by the boundary of the static patch; namely the horizon.
\item{The fundamental holographic horizon degrees of freedom are ``hyperfast" scramblers which scramble on a time scale of order the dS horizon scale $R,$ rather than the fast-scrambling time scale $R \log S.$}
If true this implies that the Hamiltonian is not of the usual k-local \ type but is more complex.
\item The hyperfast scrambling property implies that complexity growth is also hyperfast. I argue that hyperfast complexity growth is the holographic dual of the most essential feature of de Sitter space: its exponential growth.
\item The last section of the paper introduces a ``hyperfast" limit of the SYK model which has the required features for a holographic de Sitter dual. This leads to the unexpected conjecture that a limit of SYK describes de Sitter space.
\end{enumerate}
There are very few equations in this paper, but lots of figures. The figures efficiently summarize calculations that were done in earlier papers by many people.
\subsection{A Word About Time}
The metric of the static patch of de Sitter space is,
\begin{equation}
ds^2 =-f(r) dt^2 + \frac{1}{f(r)}dr^2 + r^2 d\Omega^2
\end{equation}
where,
\begin{equation}
f(r) = \left( 1- \frac{r^2}{R^2} \right).
\end{equation}
In what follows we will sometimes use a dimensionless time coordinate on the stretched horizon denoted by $\omega.$ The relation between $t$ and $\omega$ is,
\begin{equation}
\omega = t/R
\end{equation}
Near the horizon the geometry can be approximated by Rindler space. In figure \ref{time} the relation between $\omega$ and the light-like coordinate $x^+$ is shown. The stretched horizon is indicated by the hyperbola at a distance $\rho$ from the light-cone.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.6]{time.pdf}
\caption{The near horizon region of de Sitter space is approximately Rindler space. The Rindler time $\omega$ is a hyperbolic angle. The stretched horizon is the hyperbola at a distance $\rho$ from the bifurcate horizon.}
\label{time}
\end{center}
\end{figure}
The relation between $\omega$ and $x^+ $ is,
\begin{equation}
\omega = \log{x^+} - \log{\rho}
\end{equation}
The value of $\omega$ on the stretched horizon for a given $x^+$ is mildly dependent on the value of $\rho$ that we assign to the stretched horizon. More importantly, time differences along the stretched horizon are not dependent on $\rho.$
\bigskip \noindent
There is of course something uncomfortable about assuming a universal eternal time in de Sitter space. In AdS boundary time can have unlimited accuracy provided by clocks located on, or even beyond, the asymptotically frozen boundary. In the static de sitter patch the best we can do is to assume the horizon itself defines a clock (for example through the growth of complexity), but no clock built out of a system of entropy $S$ can keep time for times longer than $e^S.$ It is also likely that no de Sitter vacuum can be stable against decay for longer than $e^S.$ The things I will talk about here do not
require such long times.
\subsection{A Word About ``Tomperature"} \label{tom}
In a system with a finite Hilbert space the temperature can be infinite while the energy and all physical time scales remain finite. An example is a spin in a magnetic field. The limit of infinite temperature is just the limit in which spin-up and spin-down are equally probable, but both spin states have finite energy and the Larmor frequency is finite. In this type of system the idea of infinite temperature can be misleading.
The SYK system is an example of this kind in which the temperature can be infinite without the energy-per-degree-of-freedom being infinite or time scales going to zero. For example the energy of a fermion excitation and the decay time for the fermion two-point function are both finite in the infinite temperature limit.
It can be useful to define a quantity ${\cal{T}}$ whch Henry Lin and I called tomperature, which reflects the finiteness of energies and time scales for such systems. The precise definition of tomperature is not important but one property involves the decay of perturbations. We will assume that tomperature is defined so that the decay of typical two-point functions behaves like $$e^{-{\cal{T}} t.}$$
The important thing about ${\cal{T}}$ is that it remains finite as $T\to \infty,$ and at low temperatures it is equal to the usual temperature $T.$
\section{Entanglement in de Sitter Space}
It is all but certain that the holographic principle \cite{tHooft:1993dmi}\cite{Susskind:1994vu}, entanglement \cite{VanRaamsdonk:2010pw}\cite{Ryu:2006bv}\cite{Hubeny:2007xt}, and complexity growth \cite{Susskind:2014rva} are the essential quantum-mechanical mechanisms that lead, in appropriate limits,
to the emergence of classical space-time geometry. It would be very surprising if they did not play an outsize role in cosmology, but so far we know very little about how they work in de Sitter space, or in other cosmological space-times.
%
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.4]{stretch.pdf}
\caption{Penrose diagram for dS and stretched horizon}
\label{stretch}
\end{center}
\end{figure}
In this paper I will assume that there is a holographic description of a static (SP) patch in four-dimensional de Sitter space\footnote{The various mechanisms and calculations described in this paper apply to four dimensions except for in the last section where a possible relation between SYK and two-dimensional de Sitter space is conjectured. Generalization to other number of dimensions is non-trivial and I will not undertake the task here. }; but unlike AdS, the spatial slices of de Sitter space have no asymptotic cold boundaries where the holographic degrees of freedom are located. Instead, they are nominally located on the boundary of the SP (see for example \cite{Dyson:2002pf}\cite{Banks:2006rx}\cite{Susskind:2011ap}\cite{Banks:2016taq}\cite{Banks:2020zcr}\cite{Susskind:2021omt}\cite{Susskind:2021dfc}); that is to say, the stretched horizon.
Static patches come in opposing pairs. To account for the pair, two sets of degrees of freedom are required. The Penrose diagram of dS in figure \ref{stretch} shows such a pair of SPs along with their stretched horizons. Following \cite{Susskind:2021omt} the center of the SPs (sometimes thought of as the points where observers are located) will be called the ``pode" and the ``antipode." I'll also refer to the two SPs as the pode-patch and the antipode-patch.
\subsection{RT in dS}
Although it is clear from the Penrose diagram that the two SPs are entangled, it would be good to have a generalization of the RT and HRT equations
\cite{Ryu:2006bv}\cite{Hubeny:2007xt} to support this claim. Let's recall the original statement of the RT formula for AdS. It begins by dividing the boundary (on a fixed time-slice) into two subsets $A$ and $B.$
\bigskip \noindent
\it The entanglement entropy of $A$ and $B$ is $1/4G$ times the minimal area
of a surface homologous to either $A$ or $B$.
\rm
\bigskip \noindent
In dS there is no boundary of a global time-slice but there is a boundary to a SP, namely the horizon. Therefore we try the following formulation:
\bigskip \noindent
\it The entanglement entropy between the pode and antipode patches is $1/4G$ times the minimal area
of a surface homologous to the stretched horizon (of either side).
\rm
\bigskip \noindent
This however will not work. Figure \ref{homologous1} shows the spatial slice and the adjacent pair of stretched horizons. The red curve represents a surface homologous to the pode's stretched horizon. It is obvious that that curve can be shrunk to zero, which if the above formulation were correct would imply vanishing entanglement between the pode and antipode static patches.
\rm
\bigskip \noindent
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.3]{homologous1.pdf}
\caption{A $t=0$ slice of dS and the stretched horizons shown as purple great circles. The red surface is homologous to the light blue horizon. It can be shrunk to a point.}
\label{homologous1}
\end{center}
\end{figure}
To correct this problem begin by separating the two stretched horizons. This is a natural thing to do since they will separate after a short time, as is obvious from figure \ref{stretch}. The space-like surface shown in green in the left panel of figure \ref{bulge} passes through the separated stretched horizons. In the right panel the geometry of the space-like slice is illustrated.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.3]{bulge.pdf}
\caption{A spatial slice through dS. In the right panel the geometry of the slice is shown.}
\label{bulge}
\end{center}
\end{figure}
The reason for the bulge between the horizons is that in de Sitter space the local $2$-sphere exponentially grows as one moves behind the horizons.
Let us now reformulate a dS-improved version of the RT principle:
\bigskip \noindent
\it The entanglement entropy of the pode-antipode systems is $1/4G$ times the minimal area
of a surface homologous to the stretched horizon of the pode, and \underline{lying between the two sets of degrees} \underline{of freedom};
in this case between the two stretched horizons.
\rm
\bigskip \noindent
It is evident from the geometry of the space-like slice that the minimum-area two-dimensional surface lying between the two horizons is degenerate: there are two equal minimal-area surfaces, and they lie right at the horizons.
This version of the RT principle is illustrated in figure \ref{homologous3}
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.3]{homologous3.pdf}
\caption{The red curve represents the minimal surface lying between the two stretched horizons shown in purple. There are two such surfaces, only one of which is shown, lying right on top of the horizons. }
\label{homologous3}
\end{center}
\end{figure}
\bigskip \noindent
Clearly the area of this de Sitter version of the RT surface is the area of the horizon.
This gives the entanglement entropy that we expect \cite{Gibbons:1977mu}, namely,
\begin{equation}
S_{ent} =\frac{\rm Horizon \ Area}{4G}.
\end{equation}
One thing to note, is that in anti-de Sitter space the phrase ``lying between the two sets of degrees of freedom" is unnecessary. The degrees of freedom lie at the asymptotic boundary and any minimal surface will necessarily lie between them.
This version of the de Sitter RT formula is sufficient for time-independent geometries. A more general HRT ``maxmin" formulation goes as follows:
Pick a time on the stretched horizons and anchor a three-dimensional surface $\Sigma$ connecting the two \footnote{After version 1 of this paper appeared I became aware that Edgar Shaghoulian had proposed a similar replacement of the AdS boundary by the dS stretched horizon, especially in the context entanglement, in a seminar at MIT last March. As in this paper Shaghoulian suggested anchoring space-like surfaces on the stretched horizon. The details of the theory appear in a recent preprint \cite{Shaghoulian:2021cef} that followed V1 of this paper. There are some differences with the formulation in this paper but the results are similar if not identical}. This is shown in figure \ref{RTdS}.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.4]{RTdS.pdf}
\caption{ The black dots represent the anchoring points of a space-like surface $\Sigma$ connecting the two horizons at a particular time. The minimal two-sphere cutting $\Sigma$ lies at the anchoring points.}
\label{RTdS}
\end{center}
\end{figure}
Find the minimum-area two-dimensional sphere that cuts the three-dimensional surface $\Sigma$ and call its area $A_{min}(\Sigma).$
Now maximize $A_{min}(\Sigma)$ over all space-like $\Sigma.$ Call the resulting area $$A_{maxmin}.$$ The entanglement entropy between the pode and antipode static patches is,
\begin{equation}
S_{ent} = \frac{A_{maxmin}}{4G}.
\end{equation}
However, in the present case, because $A_{min}(\Sigma)$ occurs at the anchoring points the maximization of $A_{min}(\Sigma)$ is redundant: the minimum area is independent of $\Sigma.$ Later we will see an example in which this is not true.
All of this may be more intuitive in the bit-thread picture \cite{Freedman:2016zud} which is illustrated in fig \ref{threads}.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.6]{threads.pdf}
\caption{The green lines represent bit-theads.}
\label{threads}
\end{center}
\end{figure}
\bigskip \noindent
We introduce a set of bit-threads ending on the two horizons. The bit-threads have a thickness or area of Planckian size. The maximum number of bit-threads that can be squeezed through without overlapping is the entanglement entropy. In general the minimum area surface defines the bottleneck which controls the maximum number of bit-threads. Since in the present case the
bottleneck is at the horizon, it follows that the entanglement entropy is determined by the horizon area.
\subsection{Black Holes}
We come now to black holes in dS. In particular we will consider a pair of Schwarzschild-de Sitter black holes located near the pode and the antipode.
The relevant Penrose and embedding diagrams are shown in figure \ref{smallbh}.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.4]{smallbh.pdf}
\caption{Penrose and embedding diagrams for the Schwarzschild -de Sitter black hole. The diagrams should be periodically identified along the dashed vertical lines. The two black holes at the pode and antipode are connected by an Einstein-Rosen bridge.}
\label{smallbh}
\end{center}
\end{figure}
If we construct a space-like slice at a positive time it will look like figure \ref{smallbh2}, the bulge in the middle representing the exponential growth of the geometry behind the cosmic horizon. The two black holes are connected by an Einstein-Rosen bridge and are therefore entangled.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.4]{smallbh2.pdf}
\caption{RT surfaces and bit-thread diagram for the dS-Schwarzschild \ geometry.}
\label{smallbh2}
\end{center}
\end{figure}
The full horizon (generalized horizon) of the static patch now has two components: the large cosmic horizon and the small black hole horizon. The minimal surface homologous to the generalized horizon lying between the two generalized horizons consists of two components shown in red and the entanglement entropy is the sum of the two areas divided by $4G.$
The bit-thread picture is especially clear. Bit theads from the generalized horizon of one side---say the pode side---can end on the generalized horizon of the other side. The maximum number of non-overlapping bit-threads is obtained by the configuration in the lower panel of figure \ref{smallbh2}.
In this case the maxmin surface behind the smaller black hole horizon does not lie at the anchoring points but at the bifurcate horizon. That is because the black hole geometry shrinks as one moves behind the horizon.
A generalization to include bulk entanglement is possible. In fact
the black hole contribution to the entanglement entropy may also be thought of as bulk entanglement in which the entanglement of the cosmic horizons is supplemented by the entanglement of bulk matter; namely the two entangled black holes.
\section{Hyperfast Scrambling and Quasi-normal Modes}
Now we come to the second subject of this paper: chaos and scrambling in de Sitter space.
\subsection{Scrambling}
At first it seems obvious: if a particle is dropped from near the pode it falls to the horizon in a manner very similar to the way it would fall to the horizon of a black hole. The momentum grows exponentially which, according to the momentum-size correspondence \cite{Susskind:2018tei} \cite{Susskind:2019ddc}, means that the operator size also grows exponentially with the standard exponent,
\begin{equation}
size = \ e^{2\pi t / \beta}
\end{equation}
or in terms of
Rindler time,
\begin{equation}
size = e^{\omega}.
\end{equation}
This implies a scrambling time (time at which the size saturates at value $S$)
\begin{eqnarray}
t_* &=& \frac{\beta}{2\pi} \log{S} \cr \cr
\omega_* &=& \log{S}.
\end{eqnarray}
Thus it would naively appear that the static patch Hamiltonian should be a conventional fast-scrambler\footnote{This argument was put forth in \cite{Susskind:2011ap}. I now believe it is incorrect.}.
But what do we mean by size? Whose size, and how is it defined? Consider the operator that creates a particle at the pode. In the theory where size is best understood---the SYK model---size refers to the number of simple fermionic operators in the time-dependent operator that perturbs the boundary \cite{Roberts:2018mnp}. At the time the operator acts to create a fermion at the boundary its size is one. We might imagine that in dS the operator that creates a particle at the pode is also simple and of small size. This is wrong: in de Sitter holography the simple operators are located at the stretched horizon, not at the pode. The operator that creates a particle at the pode is likely to be highly complex. One sees this explicitly in the dS-Matrix theory \cite{Banks:2006rx}\cite{Susskind:2011ap}\cite{Banks:2016taq} where creating a single particle at the pode requires constraining all $N$ ``off-diagonal'' matrix degrees of freedom \ that link the cosmic horizon to the particle.
\subsection{Quasinormal Modes}
That raises the question: what is the bulk interpretation of the simple degrees of freedom---for example a single matrix degree of freedom---in dS holography? The answer is clear. By analogy with AdS where simple operators describe perturbations of the boundary, in dS holography the simple operators describe disturbances of the horizon. A simple operator excites a single quantum of energy in the low angular momentum quasi-normal modes of the cosmic horizon.
\subsection{Scrambling}
The $\log{S}$ factor in the the scrambling time in SYK
has a simple intuitive bulk meaning as the time it takes for a single fermion to fall from the boundary, where the holographic degrees of freedom are located, to the stretched horizon \cite{Sekino:2008he}. But in static patch holography the holographic degrees of freedom \ are already located at the stretched horizon.
Therfore there is no $\log{S}$ factor in the scrambling time. The scrambling time for an initially simple operator is of the same order as the decay time of the quasinormal modes. In Rindler units that means a time of order unity,
\begin{eqnarray}
\omega_* &\sim& 1, \cr
t_* &\sim& \beta = R.
\label{wstareq1}
\end{eqnarray}
One could say that scrambling of the fundamental degrees of freedom in de Sitter space is ``hyperfast." Any system that satisfies \eqref{wstareq1} I'll call a hyperfast system.
\bigskip \noindent
One might object that hyperfast scrambling violates the rigorous fast-scrambling bound of \cite{Maldacena:2015waa}. However that bound is based on an important assumption---k-locality of the Hamiltonian:
\bigskip \noindent
\it The meaning of k-locality: The Hamiltonian consists of a sum of monomials (in the fundamental degrees of freedom) each of which has degree less-than-or-equal-to $k,$ where $k$ is parametrically of order unity in the limit of large entropy. In otherwords $k$ is fixed as $R$ becomes large in Planck units.
\rm
\bigskip \noindent
Evidently hyperfast scrambling requires a violation of $k$ locality. One way to do this is by loosening the requirement that $k$ be independent of the overall size of the system. An example would be a system of $N$ qubits in which we retain the requirement that the Hamiltonian consists of a sum of monomials of degree less-than-or-equal-to $k,$ but allow $k$ to grow with $N$.
In general this will allow the system to evade the fast-scrambling bound. In section \ref{hype} we will analyze an example of this type.
In what follows I will reserve the term k-local \ strictly for the case in which $k$
is fixed and independent of $N.$
\subsection{Scrambling and Thermalization}
Scrambling and thermalization are two different phenomena. Scrambling is diagnosed by out-of-time-order four-point functions while thermalization is diagnosed by the conventional two-point function or more simply the decay rate of a simple perturbation. For example the decay time of quasinormal modes for black holes is order $\beta,$ the inverse temperature, while the scrambling time is of order $\beta \log{S}.$
Nevertheless I am arguing that de Sitter space is a hyperfast system meaning that the two time-scales---scrambling time and quasinormal mode decay time---are of the same order,
\begin{equation}
t_{thermal} \sim t_* \sim \beta.
\label{hyperfast}
\end{equation}
In section \ref{hype} I will give a quantitative argument for this behavior in a certain limit of qubit models including SYK.
\section{Hyperfast Complexity Growth}
Hyperfast scrambling has implications for complexity growth. For k-local \ Hamiltonians complexity growth is bounded and cannot exceed linear growth with a coefficient proportional to the product of temperature and entropy. Typically this is saturated for black holes. Equivalently, for black holes complexity grows with the time like,
\begin{equation}
{\cal{C}} \sim STt = S \omega
\label{CsimSw}
\end{equation}
But if the Hamiltonian itself is complex---for example if it contains terms of very high order in the elementary degrees of freedom---then one might expect complexity growth to be hyperfast. This, in turn, should manifest itself in the space-time geometry through the holographic complexity dualities. We will see that hyperfast complexity growth is closely related to the key property of de Sitter space: its exponential growth.
\subsection{Maximal Slices in dS} \label{maxslice}
Let us consider generalizations of holographic complexity to de Sitter space. First the CV proposal. Instead of anchoring spatial slices on the boundaries (which do not exist) we anchor them on the horizons. Then, subject to the anchoring conditions, we determine the maximal space-like surface that stretch between the two horizons. The holographic complexity conjecture is that the complexity is proportional to the volume of those slices.
For early times the maximal slices are smooth but relative to AdS they bulge up into the interior region as shown in figure \ref{CV}.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.5]{CV.pdf}
\caption{The top panel shows the early smooth maximal slices bending upward due to the expanding local two-spheres. At some point the maximal surfaces intersect future infinity. The bottom panel shows the maximum slices, assuming an appropriate cutoff, at later times. }
\label{CV}
\end{center}
\end{figure}
\bigskip \noindent
The reason is that the geometry of de Sitter space exponentially grows behind the horizon. Therefore, in order to capitalize on that growth, the maximal surfaces bend upward toward the future boundary. Let us consider the midpoint of the maximal surface and assume that it is a proper time $\tau$ from the bifurcate horizon. At the midpoint the area of the local two-sphere grows like $\exp{\frac{\tau}{R}}.$ This has the effect that the volume of the maximal slice also grows exponentially.
As figure \ref{CV} shows, a point quickly occurs at which the surfaces become light-like and reach the future boundary. This happens after a time of order $\beta$ or a Rindler time of order unity.
Beyond this point the classical rate of volume growth becomes infinite. Since the true complexity is limited to be exponential in the entropy, I take this to mean that the rate of complexity growth becomes exponentially large once the anchoring points have moved a short distance along the stretched horizons. A reasonable interpretation is that for a hyperfast system, after a Rindler time of order unity the complexity grows linearly with time, but with a rate that is non-polynomially large in $S,$
\begin{eqnarray}
{\cal{C}}(\omega) &\sim& \exp{\left(\sqrt[a]{S}\right)} \times \omega \cr
1<&a&<\infty,
\label{nonpoly}
\end{eqnarray}
instead of as in \eqref{CsimSw}.
Later in section \ref{sykds} we will see that this is exactly what is predicted by quantum circuits in the hyperfast regime.
\subsection{Inflation: The Dual of Hyperfast Complexity Growth}
The growth of maximal surfaces is a geometric feature whose dual is complexity growth.
Hyperfast growth of maximal surfaces, as in figure \ref{CV} reflects the inflationary growth of the interior of de Sitter space. (The same things are true for the complexity-action duality.)
This leads to the following conjectured duality:
\bigskip \noindent
\it
The property of hyperfast complexity growth is the holographic dual of the exponential growth of de Sitter space behind the horizon of the static patch.
\rm
\bigskip \noindent
What does it take for complexity to grow in a hyperfast way? A necessary condition is that the Hamiltonian not be k-local \ (unless $k$ is allowed to grow with $N$).
\section{SYK and dS}\label{sykds}
I now want to conjecture that a particular limit of a well-known system displays features of de Sitter space:
\subsection{A Radical Conjecture}
\it
\bigskip \noindent
The SYK system, and other similar qubit systems, in the limit of high temperature (but finite tomperature), with $q$ scaling as a power of $N,$
\begin{eqnarray}
T&\to& \infty \cr
q &\sim& N^{p} \cr
0&<&p<1,
\label{hyperlimit}
\end{eqnarray}
are hyperfast and have features that strongly resemble de Sitter space.
\rm
In particular systems defined by \eqref{hyperlimit} are hyperfast scramblers with the scrambling and thermalization times of the same order, as in equation \eqref{hyperfast}. Both time scales are set by the fundamental energy scale of SYK, namely the parameter ${\cal{J}}.$
\begin{equation}
t_{thermal} \sim t_* \sim {{\cal{J}}}^{-1},
\label{TtsimJ}
\end{equation}
where ${\cal{J}}$ is the usual SYK energy scale that controls the magnitude of the $q$-local terms in the Hamiltonian.
Most significantly, the SYK system in the limit \eqref{hyperlimit} exhibits hyperfast complexity growth, the importance of which I explained in section \ref{maxslice}. In section \ref{hype} an
explicit calculation will confirm this\footnote{I thank Adam Brown for sharing his calculations of complexity growth in SYK at large $q.$ }.
\bigskip \noindent
\subsection{The Boundary Becomes the Horizon}
SYK in this parameter regime has other similarities with de Sitter holography. Consider the Penrose diagram for the bulk dual of SYK; namely JT gravity. The geometry for low temperature is shown in the left panel of figure \ref{JTDS}.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.4]{JTDS.pdf}
\caption{Penrose diagram for the JT bulk dual of SYK. The dynamical Schwarzian boundary bends inward toward the horizon. At low temperature shown in the left panel the boundary bends in very little. At high $T$ it bends inward and almost hugs the horizon.}
\label{JTDS}
\end{center}
\end{figure}
As is well known, the boundary of the geometry bends inward toward the horizons by an amount that depends on the temperature \cite{Maldacena:2016upp}. For very low $T$ the boundaries bend in very little and the distance between them at $t=0$ is large as in the left panel. But as the temperature increases the boundaries move closer to the horizon and if one follows the formulas to $T \to \infty$ (or tomperature to ${\cal{J}}$) they get within a cutoff distance
${{\cal{J}}}^{-1}$ of the horizons as in the right panel. In the high temperature limit the boundary merges with the stretched horizon and the holographic degrees of freedom cannot be distinguished from horizon modes. To put it another way, the boundary degrees of freedom become the quasinormal horizon modes.
If we combine this picture with the conjecture that for $q\sim N^{p}$ complexity growth (see next section \ref{hype}) is hyperfast then it follows that the geometry between the horizons grows rapidly, just as one might expect for de Sitter space.
All in all, I think it is possible that in the limit $T\to \infty, \ {\cal{T}} \to {\cal{J}},$ and $ \ q\sim N^{p},$ the SYK system may have more in common with de Sitter space than with a black hole. But two things are missing: an argument
that there is an emergent space-time in the grey shaded region of the right panel of figure \ref{JTDS}, i.e., geometry near the pode and antipode; and evidence that the model has the symmetries of de Sitter space.
As I argued earlier, the degrees of freedom near the pode region should be constructed from somewhat complex combinations of the horizon degrees of freedom
and might not be easy to see. As for the symmetries I don't have anything to say right now.
\subsection{Hyperfast Scrambling in SYK}\label{hype}
This rest of this section is based on calculations that were explained to me by Adam Brown, Steve Shenker, Douglas Stanford, and Zhenbin Yang. The detailed calculation in the Brownian\footnote{As in Brownian motion.} ``epidemic model" was done by Adam Brown. It applies to the SYK model, but also more general qubit models in the limit of large temperature and large k-locality parameter which in SYK is called $q$.
At high temperature it is sufficient to consider so-called Brownian models. We begin with an epidemic model for operator growth. The rule is that in each step of a quantum circuit $\left( N/q \right)$ random q-local gates act on the $N$ qubits. Each qubit is represented in one gate and no qubit interacts in more than one gate.
At the start one qubit is ``infected." After $n$ steps the probability that any qubit is infected is $P(n).$ According to Brown, $P(n)$ satisfies the recursion relation,
\begin{equation}
P(n+1) = P(n) + \big[ 1-P(n) \big] \big[ 1 - \big( 1-P(n)\big)^{q-1} \big].
\label{discp}
\end{equation}
In the continuum Brownian limit this becomes,
\begin{equation}
\frac{dP}{dn} = \big( 1-P \big) \left(( 1 - (1-P)^{q-1} \right).
\label{cntinp}
\end{equation}
Integrating \eqref{cntinp} gives,
\begin{equation}
P(n) = 1 - \big( 1+e^{(q-1)n} \big)^{\frac{-1}{q-1}}
\label{solut}
\end{equation}
shown in figure \ref{graph}.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=.6]{graph.pdf}
\caption{The function in \eqref{solut} for $q=2$ (red), $q=6$ (blue), and
$q=100$ (green). By the time we get to $q=100$ the curve has become indistinguishable (for $n>0$) from the late-time behavior in \eqref{earlylate}.
The very early exponentially growing region collapses to zero as $q$ increases. }
\label{graph}
\end{center}
\end{figure}
The solution involves two exponential behaviors:
\begin{eqnarray}
P(n) &=& \frac{1}{N} e^{(q-1)n} \ \ \ \ \ \ \ \ {\rm early \ time} \cr \cr
P(n) &=& 1- e^{-n} \ \ \ \ \ \ \ {\rm late \ time}
\label{earlylate}
\end{eqnarray}
The early behavior is the usual limiting Lyapunov behavior associated with the scrambling bound \cite{Maldacena:2015waa}. This determines the relation between $n$ and the Rindler time for the usual SYK theory,
\begin{equation}
\omega = (q-1)n.
\end{equation}
The scrambling time is determined by $P(n)\to 1,$
\begin{equation}
n_* = \frac{1}{q-1} \log{N}.
\end{equation}
But analyzing the solution \eqref{solut} we see that as $q$ increases the range of time over which the Lyapunov behavior is relevant shrinks to zero with large $q.$ This is easily seen in figure \ref{graph}. When $q$ is extremely large the scrambling time $n_*$ is determined by the late-time exponent and we find,
\begin{equation}
n_* =1.
\end{equation}
This applies whenever $q\sim N^p.$
On the other hand the thermal time (the time-constant for the decay of the two-point function) in the limit $q\sim N^p$ can also be calculated from the Brownian model. One finds that the thermalization time is proportional to the scrambling time\footnote{I thank Zhenbin Yang for this observation.},
\begin{equation}
\omega_{therm} = \frac{\omega_*}{2}.
\end{equation}
We may interpret this to be dual to the statement that the scrambling time and the decay time for quasinormal modes are equal apart from a factor of order unity. This, as I explained, is what one expects if the holographic degrees of freedom are located on the stretched horizon of the de Sitter static patch.
The thermal time scale is also the decay time for quasinormal modes, and in Rindler units it is of order unity. This observation allows us to identify the circuit time $n$ with Rindler time up to a constant of order unity,
\begin{equation}
\frac{n}{\omega} \sim 1.
\end{equation}
\subsection{Hyperfast Complexity Growth in SYK}
In an attempt to place bounds on complexity growth in quantum circuits Nielsen
\cite{Nielsen} introduced a right-invariant ``complexity geometry" on the group $SU(2^N).$ The geometry is characterized by certain penalty factors\footnote{The notation in this section is that of Brown and Susskind \cite{Brown:2017jil}} ${\cal{I}}(w)$ which Nielsen took to be exponentially large or even infinite, in all directions other than those associated with operators involving $k$ or fewer qubits,
\begin{eqnarray}
{\cal{I}}(w) &=&1 \ \ \ \ \ \ \ \ \ w\leq k \cr \cr
{\cal{I}}(w) &\geq& 4^N \ \ \ \ \ \ \ \ \ w>k
\end{eqnarray}
The geometries with infinite penalty factors are called sub-Riemannian and have pathological properties such as infinite sectional curvatures. The Nielsen and sub-Riemannian geometries are difficult to work with. Technically the reason is that the cut-locus is very close to the origin. Nevertheless \cite{Nielsen} was able to derive some interesting bounds.
Nielsen in his work did not suggest that geodesic length (in the
complexity geometry) was an actual definition of complexity. That
suggestion was made by Brown and Susskind \cite{Brown:2017jil}, who argued for a
geometry with a much more moderate schedule of penalty factors,
\begin{equation}
{\cal{I}}(w) \sim 4^w.
\end{equation}
In this section I will assume results from a soon-to-be published
paper by Adam Brown, Henry Lin, Michael Freedman, and myself in which we argue that there is a universality of complexity geometries. The conjecture of universality is the following:
\begin{itemize}
\item The complexity growth for the sub-Riemannian, Nielsen, and
Brown-Susskind geometries are approximately the same. Moreover the same can be said for any schedule lying between subRiemannian and Brown-Susskind.
\item In generic directions the cut locus moves from being very close
to the origin for Nielsen geometry, to being exponentially far
from the origin for the Brown-Susskind geometry.
\end{itemize}
The details are not important for us in this paper. Two things
are important: first that the great distance of the cut locus from the
origin makes it easy to use the geometry in order to estimate how
complexity grows, for very long times. Second, the Brown-Susskind
schedule implies that at large $q$, after a short transient time the
complexity growth for the SYK system satisfies,
\begin{equation}
{\cal{C}}(t) = 2^q \epsilon Nt
\end{equation}
where $\epsilon$ is the variance in the energy per degree of freedom,
\begin{equation}
\langle H^2\rangle =N\epsilon^2.
\end{equation}
Ordinarily for a system at high temperature $\epsilon$ would grow with
temperature, but for reason discussed in section \ref{tom} we identify it
with the tomperature which remains finite as $T\to \infty.$
Thus, the complexity as a function of time is,
\begin{equation}
{\cal{C}}(t) = 2^q {\cal{T}} S t,
\end{equation}
where I’ve used the fact that the entropy is $\sim N.$ Finally the
dimensionless time $\omega$ is defined by,
$\omega ={\cal{T}} t.$
The complexity takes the form,
\begin{equation}
{\cal{C}}(\omega) =2^q S \omega.
\end{equation}
Next let us go to the hyperfast limit, $q = N^p$,
\begin{equation}
{\cal{C}}(\omega)=\left( 2^{N^p} \omega. \right)
\label{531}
\end{equation}
This is to be compared with the ordinary complexity growth for
an ordinary fast-scrambler,
\begin{equation}
{\cal{C}}(\omega) \sim N \omega.
\label{532}
\end{equation}
The rate of growth in \eqref{531} is exponentially larger than in \eqref{532}
in exactly the way described in \eqref{nonpoly} (with $a = 1/p$). In the large
$N$ limit it is infinitely faster than \eqref{532}, as required by the lower
panel of figure \ref{CV}. The only bulk interpretation that I know of is
that the geometry between the horizons grows exponentially, just
as it does in de Sitter space.
The behavior in \eqref{531} breaks down for short times $t \leq {\cal{T}}^{-1} $
and
for exponential times when the complexity has reached its maximum value and must stop growing.
One can use \eqref{531} to compute how long it takes to reach maximum complexity. The maximum complexity of a $N$-qubit system evolving under the action of a time-independent Hamiltonian is $2^N$.
The time that it takes to reach that value is,
\begin{equation}
\omega \sim 2^{N-N^p} \sim 2^N.
\end{equation}
Thus even with hyperfast growth the complexity still takes time
of order $2^N$ to reach its maximum, after which classical GR must
break down for the global space-time geometry.
This completes the argument that SYK (and other qubit models)
at high temperature and $q = N^p$
are hyperfast. I have argued that
the bulk dual of hyperfastness is:
\begin{enumerate}
\item A scrambling time equal (within a factor of $2$) to the decay
time of quasinormal modes.
\item Exceptionally rapid growth of volume between the horizons,
dual to the hyperfast growth of complexity.
\end{enumerate}
Both of these are \it inconsistent \rm with the behavior of black holes,
but are distinctive features of de Sitter space.
If correct, the conjectures of this paper open a new direction of
research in which the tools of quantum information theory can be
applied to cosmological space-times.
\section*{Acknowledgements}
I thank Adam Brown, Edgar Shaghoulian, Steve Shenker, Douglas Stanford, and Zhenbin Yang for crucial discussions on the material in section \ref{hype}.
\bigskip \noindent
LS was supported in part by NSF grant PHY-1720397.
\section*{Notes Added}
\begin{enumerate}
\item The special case of $q=N^{p}$ with $p=1/2$ is called the double-scaling limit of SYK and was studied more fully in \cite{Cotler:2016fpe}\cite{Berkooz:2020uly}. I thank Henry Lin for pointing this out.
\item
I can't recall who it was, but after this paper was circulated someone suggested a similarity with mechanisms discussed by Silverstein and collaborators \cite{Lewkowycz:2019xse}\cite{Shyam:2021ciy} in their discussion of the static patch. Consider going from
low temperature at fixed $q$ (the conformal limit) to the limit $T\to \infty, \ q\to N^p$ in two steps. First, keeping $q$ fixed we increase $T$ which has the effect of pulling in the boundaries as in figure \ref{JTDS}, but without driving the system into the hyperfast regime. This pushes the theory from a low temperature near-extremal black hole to a higher temperature non-extremal black hole, but not toward de Sitter.
Then we let $q$ increase. That is what brings the system from black hole behavior to a de Sitter-like phase.
Compare this with Silverstein and collaborators who first deform a holographic CFT by a $\bar{T} T$ deformation. If I understand correctly, this has the effect of pulling in the boundaries in much the same way as a cutoff; again illustrated by figure \ref{JTDS}. In a certain sense it is also increases the temperature, or more exactly the ratio of the temperature to the cutoff scale of the CFT. Of course at the end there is not much left of the CFT.
If the $\bar{T} T$ deformation is anything like a conventional cutoff this makes the theory spatially non-local but does not affect the k-locality of the couplings. (The couplings in an ordinary gauge theory are at most quartic which means the theory is 4-local.)
They then follow this operation by what they call a $\Lambda$ deformation, which, it is argued, drives the system to a de Sitter phase. If it does produce de Sitter space with a inflating interior, then by the arguments in this paper it should be possible to show that the $\Lambda$ deformation leads to a breakdown of k-locality, possibly by the introduction increasingly high powers of the fields, and to hyperfast scrambling.
\item Herman Verlinde has called to my attention a number of talks in which he also conjectured a relation between SYK in the double scaling region and de Sitter space \cite{Verlinde}. See also Akash Goel and Herman Verlinde, in preparation.
\end{enumerate}
|
\section{Introduction}
High-redshift ($z\geq 3$) quasars are important to study because they provide information about the growth of supermassive black holes and the evolution of active galactic nuclei (AGNs) in the early Universe. Most of high-redshift and radio-loud objects are expected to be blazars, a subclass of AGNs with a relativistic jet pointing toward the observer \citep{1995PASP..107..803U}. Relativistically beamed sources dominate the high-redshift radio source population due to the Doppler boosting effect. The number of radio-loud ($R_{1.4}\geq 100$) quasars decreases with increasing redshift. For example, at $z>6$ only several radio-loud quasars are found, and radio emission is not detected at $z>7$ in optically detected quasars \citep{2007AJ....134..617W,2014AJ....147....6M}.
An observed radio spectrum of a blazar is a combination of a flat spectrum by a compact core and a steep spectrum by a kiloparsec jet. The relationship between the core and the jet power determines the shape of the observed radio spectrum based on their contribution. Hence, the majority of radio sources at high redshifts should have a flat spectrum. However, we observe the well-known steep-spectrum \citep{2008AJ....136..344M,2008A&A...484L..39F,2010A&A...524A..83F} and the gigahertz-peaked spectrum (GPS) population high-redshift radio-loud quasars \citep{1990MNRAS.245P..20O,2010A&A...524A..83F}.
Currently, only a few high-redshift blazars are found at $z\geq 5$, for example SDSS J114657.79+403708.6 \citep{2010A&A...524A..83F,2014MNRAS.440L.111G}, SDSS J102623.61+254259.5 \citep{2012MNRAS.426L..91S,2015MNRAS.446.2921F}, SDSS J164854+460328 \citep{2019MNRAS.484..204C}, Q0906+6930 \citep{2004ApJ...610L...9R}, and PSO J047.4478+27.2992 \citep{2020A&A...635L...7B}. Their radio loudness is much less than that of quasars at redshifts $3 \leq z \leq 5$.
The most distant known blazar at $z=6.1$, PSO J047.4478+27.2992, was discovered by \cite{2020A&A...635L...7B} in 2020 and is considered a compact steep-spectrum or a megahertz-peaked spectrum (CSS/MPS) candidate based on the RATAN-600 and Very Large Array (VLA) measurements \citep{2020A&A...643L..12S,2021MNRAS.503.4662M}. It is a radio-loud blazar with a $\log R\sim 2.5$, which is higher by several orders of magnitude than for quasars at $z= 3$--5.
Several studies have compiled and studied statistically high-redshift blazars based on the optical (Sloan Digital Sky Survey (SDSS), e.g. \citealt{2013MNRAS.433.2182S}) or radio (Cosmic Lens All Sky Survey (CLASS)) observations (e.g. \citealt{2019MNRAS.484..204C}). A systematic study of 30 high-redshift ($z>4.5$) quasars was presented by \cite{2016MNRAS.463.3260C,2017MNRAS.467.2039C}, where multifrequency radio spectra were used to classify the spectral type of the sources and no preferable spectral type for high-redshift objects was found. \cite{1979A&A....80...13B}, \cite{1979A&AS...35..153T}, \cite{1980MNRAS.190..903L}, and \cite{2000A&AS..143..303D} demonstrated that sources with higher redshifts had steeper radio spectra. In some studies an ultra-steep spectrum is used either as an indicator of high redshift (e.g. \citealt{2001MNRAS.326.1563J, 2014A&A...569A..52S}) or as a reasonable tracer of intermediate-redshift ($z\geq 1$) galaxies \citep{2014MNRAS.443.2590S}.
There are only four radio-loud quasars known at $z > 6$ to date, PSO J172.3556+18.7734 at $z= 6.82$ is the most distant among them \citep{2021ApJ...909...80B}. It has a steep radio spectrum ($\alpha_{1.5-3}=-1.31$) and radio loudness about $R_{4400}\sim 70$.
Unfortunately, for most quasars at $z > 3$ observed data are available at very few radio frequencies.
In this study, new RATAN-600 observed data for 102 high-redshift quasars are presented. In 2017--2020 we obtained radio spectra measured quasi-simultaneously at 1.2, 2.3, 4.7, 8.2, 11.2, and 22 GHz and compiled a catalogue of six-frequency flux densities. Combining our new observations with those from the literature, we investigated the radio spectra of quasars at $z\geq 3$. We estimated the radio luminosity and radio loudness for these objects and made an analysis for different redshift ranges. Spectral classification was performed and new GPS/MPS candidates were found.
\section{The sample}
The complete sample consists of 102 quasars at $z\geq 3$ with a flux density level $S_{1.4}\geq100$ mJy in the declination range $-35\degr \leq {\rm Dec} \leq 49\degr$ (Fig.~\ref{fig1}).
The spectroscopic redshifts were taken from the National Aeronautics and Space Administration (NASA)/Infrared Processing and Analysis Center (IPAC) Extragalactic Database (NED),\footnote{https://ned.ipac.caltech.edu} and the redshift median value was found to be 3.26. There are 10 quasars at $z\geq 4$ in the sample and only one, B2 1023+25, at $z> 5$ ($z= 5.28$; \cite{2012MNRAS.426L..91S}). The number of objects decreases exponentially with increasing redshift (Fig.~\ref{fig2}). There are 22 bright radio quasars with flux densities 0.5--2.4 Jy in the sample, and the median value of the flux density at 1.4 GHz is 0.22 Jy (Fig.~\ref{fig3}).
Table~\ref{tab:param} presents the list of the sources with their characteristics, where Column~1~is the source name, Column~2 is the redshift $z$, Column~3 is the average flux density at 4.7 GHz (RATAN-600 measurements), Column~4 is the radio luminosity at 4.7 GHz, Column~5 is the radio loudness $\log~R$, Columns~6 and 7 are the spectral indices, Column~8 is the radio spectrum type, and Column~9 is the blazar type.
The list includes 48 quasars that are classified as blazars according to the 5th Edition of the Roma-BZCAT catalogue \citep{2015Ap&SS.357...75M}. Almost all blazars (46) are flat-spectrum radio quasars (FSRQs), and two of them are blazars of an uncertain type.
For high-energy bands, we found that only seven sources have their matches in the \textit{Fermi}-LAT 4th catalog \citep{2020arXiv200511208B}, and 30 sources have X-ray photon fluxes available according to the NED database.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig1}
\caption{Hammer-Aitoff projection in galactic coordinates of the sky distribution of the quasars. The size of the circles corresponds to the flux density level at 1.4 GHz, and the color indicates the redshift value.}
\label{fig1}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig2}}
\caption{Redshift distribution for the sample.}
\label{fig2}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig3}}
\caption{Flux density distribution for the sample at 1.4 GHz.}
\label{fig3}
\end{figure}
\section{Observations and data reduction}
We carried out the observations in transit mode with the RATAN-600 radio telescope \citep{1993IAPM...35....7P,2020gbar.conf...32S} at six frequencies between 1.2 and 22 GHz simultaneously. For a single transit observation, one can obtain an instantaneous spectrum of a source. RATAN-600 continuum radiometer parameters are presented in \citet{2019AstBu..74..348S}.
The sample was observed in 2017--2020. We observed the sources 3--15 times for each observing epoch to improve the signal-to-noise (S/N) ratio. The median number of the observing epochs is $N_{\rm obs}=9$ for the sample.
The observations were processed using the Flexible Astronomical Data Processing System (FADPS) standard data reduction software developed by \citet{1997ASPC..125...46V} for the broadband RATAN-600 continuum radiometers and the automated data reduction system \citep{2011AstBu..66..109T,2018AstBu..73..494T,2016AstBu..71..496U}.
We used the following eight flux density secondary calibrators: 3C48, 3C138, 3C147, 3C161, 3C286, 3C295, 3C309.1, and NGC 7027.
The flux-density scale is based on measurements by \cite{1977A&A....61...99B} and \cite{2013ApJS..204...19P,2017ApJS..230....7P}; these are in good agreement, and their differences are within the measurement errors. Additionally we used traditional RATAN-600 flux density calibrators: J0240-23, J1154-35, and J0521+16 \citep{2019AstBu..74..497S}.
The measurements of the calibrators were corrected for angular size and linear polarisation, according to the data from \citet{1994A&A...284..331O} and \citet{1980A&AS...39..379T}.
The measurement uncertainties and the calibration procedure are described in \cite{2001A&A...370...78M} and \cite{2021MNRAS.503.4662M}.
\section{Results}
\subsection{Flux densities}
\label{flux}
The total number of observations is more than 1400, and the detection rates are 46 per cent and 100 per cent at 22 and 4.7 GHz, respectively (Table~\ref{tab:detection}). The number of sources with RATAN data available at five to six frequencies is 75. The median value of the S/N ratio varies from 9 to 17 at 4.7--22 GHz. At 1.2 and 2.3 GHz, the median S/N values are 25 and 31 due to strong radio frequency interference (RFI). The strong RFI contamination at 11.2 GHz is also the reason for the non-detection of objects in the declination range $-10\degr \leq {\rm Dec.} \leq 0\degr$.
The flux density standard errors are 5--20 per cent for 11.2, 8.2, and 4.7 GHz, and 10--35 per cent for 2.3, 1.2, and 22 GHz. Table~\ref{tab:detection} presents the detection rate statistics, median flux densities, and their uncertainties for different frequencies. Fig.~\ref{fig4} presents the average flux density distribution at the six frequencies.
\begin{table}
\caption{\label{tab:detection} Observational statistics: the mean value of the RATAN flux densities $\tilde{S}$ and their uncertainties $\tilde\sigma_{S}$, number of observations N$_{\rm obs}$, and detection rates. The standard deviations are given in parentheses.}
\centering
\begin{tabular}{rcrrr}
\hline
Frequency & $\tilde{S}$ & $\tilde\sigma_{S}$ & N$_{\rm obs}$ & Detections \\
(GHz) & (Jy) & (\%) & & (\%) \\
\hline
22 & 0.56 (0.59) & 18 & 1418 & $46$ \\
11.2 & 0.41 (0.55) & 11 & 1418 & $89$ \\
8.2 & 0.32 (0.51) & 11 & 801 & $99$ \\
4.7 & 0.44 (0.52) & 9 & 1418 & $100$ \\
2.3 & 0.38 (0.37) & 31 & 801 & $82$ \\
1.2 & 0.77 (0.50) & 25 & 801 & $12$ \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\centerline{\includegraphics[width=.48\textwidth]{fig4}}
\caption{Averaged flux density distributions for the sample at RATAN frequencies.}
\label{fig4}
\end{figure}
\begin{table*}
\caption{\label{tab:ratan}RATAN-600 flux densities for 102 quasars, measured simultaneously at six frequencies (1.2--22 GHz). A sample is shown here; the full version is available as online supporting material.}
\begin{tabular}{cccccccc}
\hline
NVSS name & JD & $S_{22}$ & $S_{11.2}$ & $S_{8.2}$ & $S_{4.7}$ & $S_{2.3}$ & $S_{1.2}$ \\
\hline
000108$+$191434 & 2457793 & -- & $0.168\pm0.016$ & $0.166\pm0.017$ & $0.167\pm0.015$ & $0.151\pm0.055$ & -- \\
000108$+$191434 & 2457838 & $0.104\pm0.023$ & $0.134\pm0.014$ & $0.156\pm0.016$ & $0.190\pm0.017$ & $0.207\pm0.055$ & -- \\
000108$+$191434 & 2457853 & $0.147\pm0.029$ & $0.129\pm0.013$ & $0.155\pm0.016$ & $0.150\pm0.013$ & $0.136\pm0.055$ & -- \\
000108$+$191434 & 2458027 & $0.146\pm0.029$ & $0.135\pm0.014$ & $0.150\pm0.015$ & $0.164\pm0.014$ & $0.207\pm0.055$ & $0.339\pm0.096$\\
000108$+$191434 & 2458104 & $0.111\pm0.024$ & $0.126\pm0.013$ & $0.141\pm0.015$ & $0.170\pm0.015$ & $0.167\pm0.055$ & -- \\
000108$+$191434 & 2458192 & -- & $0.108\pm0.012$ & $0.133\pm0.014$ & $0.157\pm0.014$ & $0.144\pm0.055$ & -- \\
000108$+$191434 & 2458236 & -- & $0.129\pm0.013$ & $0.151\pm0.016$ & $0.153\pm0.013$ & $0.167\pm0.055$ & -- \\
000108$+$191434 & 2458309 & $0.112\pm0.024$ & $0.117\pm0.012$ & $0.152\pm0.016$ & $0.168\pm0.015$ & $0.194\pm0.055$ & -- \\
000108$+$191434 & 2458530 & -- & $0.094\pm0.010$ & $0.107\pm0.012$ & $0.134\pm0.012$ & $0.153\pm0.055$ & -- \\
000108$+$191434 & 2458717 & $0.103\pm0.022$ & $0.119\pm0.012$ & $0.118\pm0.013$ & $0.139\pm0.012$ & $0.125\pm0.056$ & $0.181\pm0.067$ \\
000108$+$191434 & 2458771 & -- & $0.088\pm0.010$ & $0.106\pm0.012$ & $0.120\pm0.011$ & $0.122\pm0.056$ & -- \\
\hline
\end{tabular}
\end{table*}
Table~\ref{tab:ratan} presents the flux densities, where Columns 1 and 2 are the source name and the averaged observing epoch (JD), Columns 3--14 are the flux densities and their uncertainties (Jy). The data is distributed in the VizieR Information System.\footnote{https://vizier.u-strasbg.fr/viz-bin/VizieR} The RATAN-600 flux densities and the instantaneous spectra for some quasars are published in the BL Lac database\footnote{www.sao.ru/blcat/} maintained by the Special Astrophysical Observatory. The database is constantly updated with new data and is freely available \citep{2014A&A...572A..59M}.
\subsection{Radio spectra and classification}
Figs.~\ref{fig:spectra}--\ref{fig:spectra4} show the broad-band radio spectra of the quasars. We compiled the spectra using both RATAN (red circles) and CATS (black circles) data \citep{2005BSAO...58..118V, 1997BaltA...6..275V}. The RATAN-600 measurements cover the time period 2017--2020, except for several objects which were observed from 2014. The CATS data cover a time period of several dozens years, and the main data are represented by the National Radio Astronomy Observatory (NRAO) VLA Sky Survey (NVSS: \citet{1998AJ....115.1693C}), Faint Images of the Radio Sky at Twenty cm (FIRST: \citet{1994ASPC...61..165B}), Westerbork Northern Sky Survey (WENSS: \citet{1997A&AS..124..259R}), Green Bank 6-cm survey (GB6: \citet{1996ApJS..103..427G}), Australia Telescope 20 GHz Survey (ATCA20: \citet{2010MNRAS.402.2403M}), and VLA measurements.
Low-frequency data are well presented by the GaLactic and Extragalactic All-sky Murchison Widefield Array (GLEAM) survey at 72--231 MHz (2013--2014; \cite{2017MNRAS.464.1146H}) and the Giant Metrewave Radio Telescope Sky Survey (TGSS) at 150 MHz (2015, \cite{2017A&A...598A..78I}); the data were obtained almost simultaneously with the RATAN-600 observations.
We analysed the averaged radio spectra of quasars compiled from the RATAN and CATS data. The spectral parameters were estimated by fitting with polynomials, which were calculated by weight averaging of measurements in the FADPS software package \citep{1997BaltA...6..275V}. The resulting approximations are shown by the green dotted lines in Figs.~\ref{fig:spectra}--\ref{fig:spectra4}. We defined the spectral index from the power-law $S_{\nu} \propto \nu^{\alpha}$ ($S_{\nu}$ is a flux density at the frequency $\nu$, and $\alpha$ is a spectral index).
Table~\ref{tab:type} presents the criteria based on which the spectral classification was made. We determined the spectral indices $\alpha_{\rm low}$ and $\alpha_{\rm high}$ for peaked and upturn spectra for the frequencies where the spectral slope changes its sign from positive to negative or vice versa. For the rest of the spectra, the low-frequency spectral index was calculated between the TGSS, GLEAM, WENSS, and other decimeter-wavelength measurements. The high-frequency spectral index was estimated between centimeter catalog data points. We considered a spectrum to be flat if the spectral index $-0.5 \leq \alpha \leq 0$, and inverted if $\alpha > 0$. The spectrum is assumed to be complex for indeterminate shapes with two or more maxima or minima. The spectrum is called steep or ultra-steep for $-1.1 < \alpha < -0.5$ and $\alpha \leq -1.1$, respectively.
Fig.~\ref{fig5} presents the spectral-type distribution in the convinient form of a two-color diagram in which $\alpha_{\rm low}$ is plotted against $\alpha_{\rm high}$. The point size is proportional to the radio luminosity at 4.7 GHz. The 'blazar box' corresponds to the spectral index range from $-$0.5 to 0.5. The median values of $\alpha_{\rm low}$ and $\alpha_{\rm high}$ are $-$0.01 (0.52) and $-$0.44 (0.33), respectively (Fig.~\ref{fig6}). Thus, the majority of the radio spectra steepens at higher frequencies.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig5}}
\caption{Two-color diagram. The dotted square indicates the 'blazar box', the dashed lines mark the zero spectral indices, and the ultra-steep spectrum sources area is presented with a pink color. The point size is proportional to the radio luminosity at 4.7 GHz.}
\label{fig5}
\end{figure}
We found 47 (46 per cent) of quasars had peaked radio spectra (PS) with a turnover in their average spectrum (Table~\ref{tab:PS}). There are 17 sources with classical \citep{1997A&A...321..105D} GPS spectral features and 13 quasars with a peak at MHz frequencies (MPS). We present 24 new GPS and MPS candidates in Table~\ref{tab:PS}.
A large part of the sample (24 per cent) has flat spectra and 15 per cent has steep ones. A small number of the sources have complex, upturn, and inverted spectra.
We have not found ultra-steep radio spectra. The spectrum of PKS 2313-330 has an ultra-steep part with $\alpha_{4.7{\text{--}}11.2}=-1.58\pm0.09$ and is classified as PS. The spectral index was determined based on the multi-epoch RATAN observations in 2015--2020 at four frequencies simultaneously. For the second quasar, 87GB 012817.1+434250, we found an ultra-steep spectral part at low frequencies with $\alpha_{1.4{\text{--}}4.7}=-1.16\pm0.02$. This object is classified as having a complex spectrum. Its ultra-steep spectral part was obtained from two non-simultaneous measurements (NVSS and RATAN) on a time-scale of 20 years.
For nine quasars, the radio spectra are well described by a sum of two spectral components at low (LFC) and high (HFC) frequencies \citep{2002PASA...19...83K,2020AdSpR..65..745K}, which are denoted by '*' in Table~\ref{tab:param}. LFC indicates an optically thin synchrotron emission of extended kpc-scale structures with a peak frequency much less than 1 GHz due to its characteristic size \citep{1963Natur.199..682S}. HFC dominates at frequencies of several GHz and more and is associated with the pc-scale jet.
\begin{table}
\begin{small}
\caption{\label{tab:type}Spectral types in the sample.}
\centering
\begin{tabular}{lccc}
\hline
Type & Criteria & N & {\%} \\
\hline
Peaked & $\alpha_{\rm low}>0$, $\alpha_{\rm high}<0$ & 47 & 46 \\
Flat & $-0.5 \leq \alpha \leq 0$ & 25 & 24 \\
Inverted & $\alpha>0$ & 8 & 8 \\
Upturn & $\alpha_{\rm low}<0$, $\alpha_{\rm high}>0$ & 2 & 2 \\
Steep & $-1.1 < \alpha < -0.5$ & 15 & 15 \\
Ultra-steep & $\alpha \leq -1.1$ & 0 & 0 \\
Complex & Two or more maxima/minima & 5 & 5 \\
\hline
\end{tabular}
\end{small}
\end{table}
\begin{figure}
\centerline{\includegraphics[width=.48\textwidth]{fig6}}
\caption{Distribution of the low-frequency (shaded) and high-frequency (colored) spectral indices for the averaged radio spectra.}
\label{fig6}
\end{figure}
\subsection{The average spectrum}
To study the distant radio sources population, we used continuum radio spectra averaged over the objects located within close redshift intervals $\Delta z = 0.1$. The spectrum of each object was preliminarily recalculated into the rest frame using the $(1+z)$ correction. The spectra of individual objects were then approximated by a polynomial of the first or second degree. Finally, the average spectrum of all objects in each bin was calculated using the formula \citep{2018AstBu..73..393V}
\begin{equation}
\log\,S(\nu_{j}) = \frac{1}{N}\sum\limits_{i=1}^N \log\,f_{i}(\nu_{j}),
\end{equation}
where $\nu_{j}$ is the set of frequencies in the approximating data set for the $i$the source, $f_{i} (\nu_{j})$ is the flux density calculated for the $i$th source approximation at a frequency $\nu_{j}$, and $N$ is the number of sources. The averaging procedure was carried out using the $spcalc$ utility of the FADPS data processing system. As a result, we have calculated eight average spectra in the range $z = 3.0$--3.8. The mean error of the averaged spectrum in each redshift bin is determined by the
quadratic sum of spectrum approximation uncertainties. The uncertainty of the spectrum approximation of an individual object was calculated using errors of polynomial fitting coefficients by a standard method. At higher redshift intervals, there are only a few objects, and the averaging procedure was not applied.
Fig.~\ref{fig7} presents the normalised average radio spectra for the redshift bins $\Delta z=0.1$. The cumulative radio spectrum is shown by the black line. A 99 per cent confidence intervals (3$\sigma$) are shown as strips with different colors.
The eight average spectra in the redshifts range 3.0--3.8 show a similar shape within their uncertainties. The values of low and high spectral indices for the cumulative radio spectrum are $\alpha_{\rm low}$=0.20$\pm$03 and $\alpha_{\rm high}$=-0.40$\pm$0.02, respectively. The peak frequency for the average spectrum is about 1.6 GHz in the rest frame.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig7}}
\caption{Average spectra of the sources normalised by the peak flux density. The coloured lines relate to the redshift bins $z = 3.0$--3.8 and corresponding strips represent 3$\sigma$ uncertainties. The black line is the average spectrum of all sources at $z=3.0-3.8$.}
\label{fig7}
\end{figure}
\subsection{$z$--$\alpha$ correlation}
We have not found any significant correlation between redshifts and spectral indices at low and high frequencies (Fig.~\ref{fig8}). The Pearson correlation coefficient is $r=0.07$ for $\alpha_{\rm low}-z$ and $r=0.01$ for $\alpha_{\rm high}-z$, with 46 and 95 per cent of probability that these weak correlations appear by chance and not weighty, although this dependence was reported for radio sources in several studies (e.g. \citealt{1979A&A....80...13B,1980MNRAS.190..903L,2000A&AS..143..303D}).
The result does not contradict the existing theories, explaining $z$--$\alpha$ correlation by the interaction of the radio lobes with the dense intergalactic medium at $z\sim~2$--3. Our sample contains distant ($z>~3$) and very compact radio sources, mostly with peaked radio spectra, for which the core emission dominates at the GHz frequencies and the influence of radio lobes is negligible \citep{2012MNRAS.420.2644K}.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig8}}
\caption{$\alpha$ versus $z$ plot for low and high frequencies. The filled symbols represent the median values of the spectral indices for redshift bins of 0.1, from 3.0 to 4.2.}
\label{fig8}
\end{figure}
\subsection{Radio luminosity}
We used the $\Lambda$CDM cosmology with $H_0 = 67.74$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_m$=0.3089,
and $\Omega_\Lambda$=0.6911 \citep{2016A&A...594A..13P} to estimate the radio luminosity at 4.7 GHz:
\begin{equation}
L_{\nu} = 4 \pi D_{\rm L}^2 \nu S_{\nu} (1+z)^{-\alpha -1},
\end{equation}
where $\nu$ is the frequency, $S_{\nu}$ the measured flux density,
$z$ the redshift, $\alpha$ the spectral index calculated at 4.7 GHz by linear approximation of the spectrum between 1 and 8 GHz, and $D_{\rm L}$ the luminosity distance.
The median value of the radio luminosity for our sample is $\sim 2 \times 10^{44}$ erg s$^{-1}$, and the source J0525$-$3343 at $z=4.41$ has the highest luminosity of $3.38 \times 10^{45}$ erg s$^{-1}$. Figs~\ref{fig9} and \ref{fig10} show the luminosity distribution and the redshift versus luminosity plot, respectively.
In Fig.~\ref{fig10} we compare the luminosities of our sample with the luminosities calculated for the 17 distant quasars at $z>5.5$ selected from the FIRST and Pan-STARRS1 surveys cross-match in \citet{2015ApJ...804..118B}. Also, we draw in the figure five individual sources: BZQ J0906+6930 at $z=5.47$, PSO J047.4478+27.2992 at $z=6.1$ and NVSS J164854+460328 at $z=5.38$ \citep{2021MNRAS.503.4662M}, VIK J2318-3113 at $z=6.44$ \citep{2021arXiv210111371I}, and PSO J172.3556+18.7734 at $z=6.82$ \citep{2021ApJ...909...80B}.
The sample of \citet{2015ApJ...804..118B} consists of the sources with the 1.4 GHz flux densities in the range 0.035--3.04 mJy, and accordingly their luminosity values are two orders of magnitude lower compared to the values in our sample.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig9}}
\caption{Radio luminosity distribution.}
\label{fig9}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig10}}
\caption{Redshift versus luminosity. The sample of distant blazars from \citet{2015ApJ...804..118B} is shown with the red circles and other five distant quasars from the literature are marked with the red stars.}
\label{fig10}
\end{figure}
\subsection{Radio loudness}
The radio loudness is defined as
\begin{equation}
R = \frac{S_{\nu, {\rm radio}}}{S_{\nu, {\rm opt}}}
\end{equation}
where $S_{\nu,{\rm radio}}$ is the estimated radio flux density at 4.7 GHz and $S_{\nu, {\rm opt}}$ is the optical flux density corresponding to the filter $B$ or $g$.
Optical data were taken from the SIMBAD database,\footnote{https://simbad.u-strasbg.fr/simbad/} where $B$ or $g$ magnitudes are presented for 71 of 102 objects. The magnitudes were transformed into flux densities using the Pogson law.
We adopted $S_{\nu,{\rm opt}} = 3631$ Jy for $g$ = 0,\footnote{\url{https://www.sdss.org/dr12/algorithms/fluxcal/\#SDSStoAB}}, $S_{\nu,{\rm opt}} = 4260$ Jy for $B = 0$ \citep{1979PASP...91..589B}, and the optical spectral index $\alpha=-0.3$ \citep{2019A&A...630A.110G}.
Fig.~\ref{fig11} presents the radio loudness distribution. The $\log~R$ value spans from 2.1 to 5.43, the median is 3.5, and the majority of the sources are highly radio loud with $\log~R > 2.5$ \citep{2019MNRAS.482.2016Z}. Only 5 of 71 quasars have $\log~R < 2.5$.
In Fig.~\ref{fig12} we draw 24 radio-loud blazars at $z>4$ selected from the Cosmic Lens All Sky Survey (CLASS: \citealt{2019MNRAS.489.2732I}) as well as five individual sources - BZQ J0906+6930 at $z=5.47$ \citep{2004ApJ...610L...9R}, PSO J352.4034-15.3373 at $z=5.84$ \citep{2018ApJ...861L..14B}, PSO J047.4478+27.2992 at $z=6.1$ \citep{2020A&A...635L...7B}, VIK J2318-3113 at $z=6.44$ \citep{2021arXiv210111371I}, and PSO J172.3556+18.7734 at $z=6.82$ \citep{2021ApJ...909...80B}, - to compare with our results.
The sample of \citet{2019MNRAS.489.2732I} was selected with a lower limit of 30 mJy at 1.4 GHz, therefore it contains fainter radio sources with lower radio-loudness values than our sample, which is reflected in Fig.~\ref{fig12}. We also discuss selection effects in the following section.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig11}}
\caption{Radio loudness distribution.}
\label{fig11}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{fig12}}
\caption{Redshift versus radio loudness. The sample of distant blazars from \citet{2019MNRAS.489.2732I} is shown with the red circles, and other five distant quasars from the literature are marked with the red stars.}
\label{fig12}
\end{figure}
\section{Discussion}
We have classified 46 per cent spectra of $z\geq 3$ quasars as PS type, which is quite a large fraction of bright distant quasars. Thus, we confirm the earlier result of \citet{1990A&AS...84..549O} who revealed that seven of the 14 quasars studied at $z\geq 3$ are GPS. Also, \citet{1998PASP..110..493O} expects about 40 per cent of all bright radio sources to be PS type (GPS and CSS).
If a PS source has a peak between 0.5 and 10 GHz, then it is classified as GPS; however, if the spectrum has a peak around 1 GHz, the spectral indices of the optically thin and thick parts are about 0.51 and $-$0.73, and the source is non-variable, then it is classified more precisely as a classical or canonical GPS \citep{1997A&A...321..105D}. In our sample, 17 sources (17 per cent) fit the criteria of classical GPS. 12 of them are already known as GPS or MPS in \citet{1997A&A...321..105D}, \citet{1998PASP..110..493O}, and \cite{2017ApJ...836..174C}. For the other five, further study of their variability (Sotnikova 2021, in preparation) will help classify them more precisely. The variability is a crucial parameter when distinguishing between PS quasars and classical GPS \citep{2020arXiv200902750O}.
Several studies \citep{2005A&A...435..839T,2012A&A...544A..25M,2013AstBu..68..262M,2019AstBu..74..348S} have shown that only 2--3 per cent of PS objects are classical GPS, whereas other GPS sources could be, in fact, the blazars that temporarily have broader peaked spectra during their flares \citep{2011A&A...536A..14P}. 50 per cent of PS objects in the sample (24 of 47) are FSRQs .
We compared the spectra of GPS quasars in the samples from \citet{2012A&A...544A..25M,2013AstBu..68..262M} and \citet{2019AstBu..74..348S}. There is no difference between averaged optically thick and thin spectral indices within the measurement uncertainties (Table~\ref{tab:comparing}). This can mean that the high redshift GPS have, on average, the same spectral indices as the low- and medium-redshift GPS, and there is no evidence for the evolution of the radio spectrum.
Among 47 PS objects, we determined that 13 (13 per cent) MPSs in the sample that are the sources with a peak below 1 GHz in the observer’s frame. They are also interesting as young compact (5--10 mas) radio sources \citep{2015MNRAS.450.1477C,2016MNRAS.459.2455C,2017ApJ...836..174C}, possibly evolved versions of GPS sources \citep{2010MNRAS.408.2261K}.
On the whole, MPS sources are a combination of PS sources, the peak frequency of which have been shifted to low frequencies due to cosmological redshift \citep{2021A&ARv..29....3O}. Two of our 13 MPSs had already been identified as MPSs in \citet{2017ApJ...836..174C}. We note that 15 objects of the sample with the steep spectrum could also be MPS candidates, and additional low-frequency ($<$1 GHz) data are needed to obtain their radio spectra at megahertz frequencies and define a possible peak.
\citet{1957MNRAS.117..680W}, \citet{1979A&A....80...13B}, \citet{1979A&AS...35..153T}, \citet{1980MNRAS.190..903L}, \citet{2000A&AS..143..303D}, \citet{2001MNRAS.326.1563J}, \citet{2010AstL...36....7V}, and \citet{2014A&A...569A..52S} reported a correlation between the observed spectral index and redshift and also proposed to use the presence of an ultra-steep spectral index as a method for selecting high-redshift radio sources. However, this correlation was detected mostly for radio galaxies, while more and more distant quasars with flat radio spectra have been revealed later. The steeper spectra for distant quasars could be explained by selection effects and higher inverse Compton losses due to the increased photon energy density of the cosmic microwave background \citep{2018MNRAS.480.2726M}. We have no quasars with the ultra-steep spectrum in the sample.
The number of objects in our sample decreases with increasing redshift, whereas the radio loudness increases. This means that our initial criterion of $S_{1.4}\geq100$ mJy selects only the most radio-loud quasars at high redshifts. There are other radio-loud quasars underrepresented in our study because of their low radio flux density at $\nu_{\rm obs} = 1.4$ GHz. The quasar radio-loud fraction was investigated in several studies, and it was shown that it remains about $\sim$ 10 per cent in optically selected quasar samples up to redshift $z \sim 6$ \citep{1989AJ.....98.1195K,2000AJ....119.1526S,2021ApJ...908..124L}.
When higher redshifts are considered, the lower-frequency observations are needed to detect the radio emission because of the steep spectrum at MHz--GHz frequencies. We think that the artificial trend we obtained in the $\log R$--$z$ plane will be eliminated if the selection criterion is changed toward a lower flux density limit at 1.4 GHz, or if a lower frequency survey is applied to the high-redshift sample construction.
\begin{table*}
\caption{\label{tab:comparing} Comparison of spectral indices of quasars with the classical GPS features in different samples: $\alpha_{\rm below}$ and $\alpha_{\rm above}$ are the median spectral indices below and above the turnover, respectively.}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
$z$ & N & $\alpha_{\rm below}$ & $\alpha_{\rm above}$ & Reference \\[3pt]
\hline
0.14--4.56 & 45 & 0.7 (0.1) & $-$0.7 (0.1) & {\cite{2012A&A...544A..25M}} \\
0--5 & 43 & 0.9 (0.1) & $-$0.6 (0.1) & \cite{2013AstBu..68..262M} \\
& & & & $S_{5}\geq 0.2$ Jy \\
0--5 & 71 & 0.7 (0.2) & $-$0.7 (0.2) & \cite{2019AstBu..74..348S} \\
& & & & $S_{5}\geq 0.2$ Jy \\
3--5 & 12 & 0.7 (0.2) & $-$0.7 (0.3) & This study \\
& & & & $S_{1.4}\geq 0.1$ Jy \\
\hline
\end{tabular}
\end{table*}
\section{Summary}
We present the results of multifrequency RATAN observations in 2017--2020 of a complete sample of bright quasars at $z \gid 3$. The flux densities were measured at frequencies of 1.1, 2.3, 4.7, 8.2, 11.2, and 22 GHz quasi-simultaneously with uncertainties of 9--31 per cent. The detection rate is 100, 89, and 46 per cent at 4.7, 11.2, and 22 GHz, respectively. The main conclusions are the following:
\begin{enumerate}
\item The sample has peaked and flat radio spectra (46 per cent and 24 per cent, respectively). There are 15 per cent quasars with a steep spectrum ($-1.1<\alpha<-0.5$), indicating the dominance of bright compact core emission and insignificant contribution of extended optically thin kpc-scale components in observed radio spectra. There are no quasars with ultra-steep radio spectra ($\alpha\leq-1.1$) in the sample. We have not found any significant correlation between the redshifts and spectral indices.
\item Eight new MPSs (J0214$+$0157, J0232$+$2317, J0525$-$3343, J0624$+$3856, J0905$+$0410, J1045$+$3142, J1418$+$4250, J2019$+$1127) and 16 new GPS (Table~\ref{tab:PS}) candidates are suggested. Further study of their variability and additional low-frequency observations are needed to classify them precisely.
\item The radio luminosity $L_{4.7}$ has been calculated with a median value of $\sim 2 \times 10^{44}$ erg s$^{-1}$ for the sample.
\item Using new radio data, we have estimated the radio loudness for 71 objects with a mean value of 3.5. The $\log~R$ values span from 2.1 to 5.43. The majority of the quasars are highly radio-loud with $\log~R >~2.5$, and only five quasars have $\log~R <~2.5$.
\end{enumerate}
We are preparing a follow-up work with a further study of the sample, including assessing the variability and notes on individual objects (Sotnikova, in preparation).
\section*{Acknowledgements}
We thank the referee for providing useful suggestions and comments that significantly improved the article. This work was supported in the framework of the national project ``Science'' by the Ministry of Science and Higher Education of the Russian Federation under the contract 075-15-2020-778. The observations were carried out with the RATAN-600 scientific facility. This research has made use of the CATS database, operated at SAO RAS, Russia.
The research has made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. We used of the SIMBAD database, operated at CDS, Strasbourg, France.
\section*{Data Availability}
The data underlying this article are available in the article and in its online supplementary material. The measured flux densities are distributed in the VizieR Information System and partly published in the BL Lac database available at the Special Astrophysical Observatory website.
\bibliographystyle{mnras}
|
\section{Introduction}
In recent times, the competing effects of electronic correlations and strong spin-orbit
coupling have generated considerable interest in the condensed matter
community \cite{doi:10.1146/annurev-conmatphys-031115-011319,PhysRevLett.109.066401,Schaffer_2016}.
Topological materials \cite{doi:10.1146/annurev-conmatphys-031214-014501,doi:10.1146/annurev-conmatphys-062910-140432,doi:10.7566/JPSJ.82.102001,Hasan_2015,RevModPhys.82.3045,RevModPhys.83.1057,Yan_2012}
provide an excellent platform for the potential realization of exotic
electronic phases and novel phenomena resulting from such
an interplay.
A particular example is provided by Weyl semimetals (WSMs) \cite{doi:10.1146/annurev-conmatphys-031016-025458,doi:10.1146/annurev-conmatphys-031113-133841,RevModPhys.90.015001},
which are topologically nontrivial gapless systems featuring pairs of non-degenerate
bands, touching each other at isolated points in the band structure,
with a low-energy description in terms of massless, linearly dispersing
Weyl fermions. Electronic interaction effects in these systems have been extensively explored, using various approaches \cite{PhysRevB.95.201102,PhysRevLett.113.136402,PhysRevLett.109.196403,PhysRevB.98.241102,PhysRevB.96.195160,PhysRevB.94.241102,PhysRevB.94.125135,PhysRevB.90.035126,PhysRevB.89.235109,PhysRevB.89.014506,PhysRevB.96.165203,PhysRevB.100.245137,Yi2017possible}.
More generally, the behavior of three-dimensional
semimetals under a large applied magnetic field has presented
a problem of recurring interest for the community\cite{Arnold2017,Yakovenko93,Kumar_2010,LeBeouf2017,PhysRevB.79.241101,PhysRevLett.110.266601,PhysRevLett.54.1182,TAKAHASHI1994384,Yaguchi_2009}.
For such systems, in the so-called quantum limit, where the Fermi level
intersects only the lowest Landau band, the nesting of Fermi surfaces
is greatly enhanced, stabilizing field-induced symmetry-breaking phases,
such as density wave orders \cite{Arnold2017,PhysRevB.79.241101,Yaguchi_2009,TAKAHASHI1994384}.
The quantum limit can be achieved relatively easily in Dirac/Weyl semimetals,
owing to their low carrier density, and the extra valley or orbital
degrees of freedom may further enrich the types of phases that can
be realized.
On the experimental side, high-field measurements of Dirac/Weyl semimetals have also given rise to anomalous features \cite{Zhang2016,Wang2020,Tang2019,Ramshaw2018,Liu2016} that hint at the possibility of field-induced phase transitions in the quantum limit, thus contributing to the interest in this area. Such anomalies
are often attributed to interaction-induced instabilities occurring
at high fields\cite{Zhang2016,Ramshaw2018,Tang2019}. Besides, the high-field regime provides a setting that can be useful for potentially
isolating decisive experimental signatures of Weyl fermions\cite{Hui2019,Kim2017,Ramshaw2018,rinkel2019,PhysRevB.86.115133,PhysRevX.4.031035,Zhang2019}.
From a purely theoretical point of view, much of the existing effort has
been directed specifically towards examining the effect of electronic interactions
in three-dimensional semimetals at high fields\cite{Roy2015, Miransky2015, Shindou2019,Song2017,Trescher2017,Xiao2016,Yang2011,TAKAHASHI1994384}.
In this context, the appearance of a chiral symmetry-breaking, fully
gapped charge-density wave (CDW) order, even for sufficiently weak
repulsive electron-electron interactions, for topological Dirac and Weyl semimetals has been discussed in the literature \cite{Roy2015,Song2017,Trescher2017,Xiao2016,Yang2011}.
Such studies
have employed both analytical and numerical approaches, but have generally
focused their attention solely on the quantum limit.
On the other hand, while phonons in Weyl semimetals have gained considerable attention\cite{Song2016, Hui2019,rinkel2019,Moller2017,nguyen2020,PhysRevB.101.085202,PhysRevB.95.165114,PhysRevB.97.195106,PhysRevLett.110.046402,PhysRevLett.119.107401,PhysRevLett.125.146402, Wang2020} due to their interplay with the nontrivial electronic band topology,
the role of phonons in the magnetic field-induced {\em instabilities} of Weyl semimetals
has barely been explored.
The need for including electron-phonon interactions in theoretical
treatments of such systems has been raised in part by recent high-field measurements
in the Weyl semimetal TaAs \cite{Ramshaw2018}, where a strong increase
in the ultrasonic attenuation at low temperatures has been observed.
Likewise, it has been recently suggested that electron-phonon interactions can be important for the field-induced ordered states reported in graphite\cite{LeBeouf2017} and in the Dirac semimetal ZrTe$_{5}$\cite{Zhao2020, Qin2020}, although the discussion on the presence or absence of order in ZrTe$_5$ is still ongoing\cite{Okada1982, Tian2021, Galeski2021, Ehmcke2021}.
In this paper, we aim to contribute along the preceding line of research by studying possible instabilities of a Weyl semimetal at high magnetic fields,
through the interplay of electron-electron and electron-phonon interactions.
The primary objective is to theoretically predict the high-field lattice instabilities
for a minimal (two-node) Weyl semimetal model, as a function of the magnitude
and direction of the magnetic field, in the presence and absence of improper symmetries relating the two Weyl nodes, and for both the adiabatic and nonadiabatic phonon regimes.
To that end, we adopt the Kadanoff-Wilson renormalization group (RG) approach\cite{Bourbon89,Bourbon91},
which has the advantage of treating different electronic and lattice instabilities on an equal footing and allows to
include multiple energy scales with ease.
It follows from our analysis that electron-phonon interactions tend to augment the charge-density wave fluctuations in the system and that, under certain conditions, lattice instabilities preempt those of purely electronic origin.
It is, therefore, important to take into account the effect of electron-phonon interactions while studying potential instabilities in Weyl semimetals at high magnetic fields.
Our analysis also evidences that particle-particle scattering (a.k.a. the "Cooper channel"), which has been often neglected before, plays an important role when the two Weyl nodes are related by a mirror plane. Indeed, the Cooper channel interferes destructively with the particle-hole scattering (a.k.a. the "Peierls channel"), thereby preventing purely electronic instabilities while enabling lattice instabilities.
Another important finding from our study is that breaking the mirror symmetry (which can be done for example by rotating the external magnetic field) suppresses the Cooper channel, thereby increasing the critical temperature for the lattice instability.
The paper is organized as follows. In Sec. \ref{sec:prel}, we begin by recalling the minimal
model of two Weyl nodes in a quantizing magnetic field and in the
presence of electron-phonon and electron-electron interactions. Then, we briefly review the Kadanoff-Wilson renormalization group
method.
In Sec. \ref{sec:cll}, we write down the RG equations when the Fermi level intersects
only the chiral Landau levels.
We derive the condition for phonon softening from the renormalization
of the phonon part of the action, and use it to obtain the Peierls
transition temperature, in hitherto unexplored scenarios and as a function of different tuning parameters.
For instance, we take into account the effect of an asymmetry in
the position of the Weyl nodes of opposite chiralities, as well as
in their velocities.
We also extend our analysis to the nonadiabatic regime, where the bare phonon energy
exceeds the energy scale of the Peierls transition.
We consider competing logarithmic divergences in the Cooper and Peierls
channels whenever relevant, and find that the results are sensitively
dependent on the competition between the two contributions.
In Sec. \ref{sec:noncll}, we describe the
corresponding results for the case when the Fermi level also intersects
the first nonchiral Landau level. Here, we distinguish between the
behavior of the scalar and pseudoscalar phonon modes, and we find that the latter are qualitatively more prone to undergo an instability.
In Sec. \ref{sec:real}, we briefly extrapolate our results to real Weyl semimetals and comment on their applicability to Dirac semimetals.
In Sec. \ref{sec:conc}, we summarize our results and state the main conclusions.
We also compare our work with recent theories\cite{Zhao2020,Qin2020} of lattice instabilities in ZrTe$_5$, as well as with theories of magnetic catalysis\cite{Miransky2015}.
Finally, the appendices contain details of the RG calculation.
\section{Model}
\label{sec:prel}
For most of this work, we limit ourselves to a minimal model of two untilted Weyl nodes of opposite chirality.
In Sec. \ref{sec:real}, we discuss briefly how the results from our toy model might extrapolate to models with various pairs of Weyl nodes, with a tilt.
\subsection{Free Weyl fermions in a magnetic field}
We consider a Weyl semimetal of spatial dimensions $L_x\times L_y\times L_z$ under a magnetic field ${\bf B}=B\hat{\bf z}$, with periodic boundary conditions in the $z$ direction.
The Landau levels originating from a Weyl fermion of a given chirality $\tau$ (with $\tau=\pm 1$) are characterized by the following good quantum numbers: the Landau level index $n$ ( $n\in\mathbb{Z}$, with $n>0$ for conduction-band Landau levels, $n<0$ for valence-band Landau levels, $n=0$ for the chiral Landau level), the guiding center $X$, and the wave vector $k$ along the direction of the external magnetic field, measured from the position of the Weyl node.
Note that $k$ is assumed to be bound by a cutoff, within which the Weyl fermion approximation is valid.
This cutoff is smaller than the internodal distance.
Below, we will often denote the collection of good quantum numbers with the letter $\alpha$, i.e. $\alpha=(n,X,k,\tau)$.
In addition, we disregard the Zeeman effect, whose effect in our model is limited to a shift in the Weyl nodes' location.
The eigenenergies are independent of $X$ and given by
\begin{align}
\label{eq:LL_energies}
\epsilon_{n X k \tau} &=\left\{\begin{array}{cc} \tau b_0 + \epsilon_{n k\tau} & \text{ $(n\neq 0)$}\\
\tau b_0 +\hbar v_\tau \tau k & \text{ $(n=0)$}\end{array}\right.,
\end{align}
where $\epsilon_{n k\tau} = \hbar v_\tau \tau {\rm sign}(n)\sqrt{k^2+ 2 |n|/l_B^2}$,
$v_\tau$ is the node-dependent magnitude of the Dirac velocity (not to be confused with the $v_{n k \tau}$ coefficient below), and $2 b_0$ is the energy shift between the two nodes of opposite chirality.
In the presence of an improper symmetry that relates the two nodes (e.g. a mirror plane), $b_0=0$ and $v_+ = v_-$.
Below, we will study the general case in which both $b_0$ and $v_+-v_-$ maybe nonzero.
The eigenspinors, written in the pseudospin basis $\{\sigma\} = \{\uparrow,\downarrow\}$ that describes the two degenerate bands at the Weyl node in the absence of magnetic fields, are
\begin{equation}
\label{eq:spinor_cll}
\langle{\bf r}|\Psi_{0 X k \tau}\rangle= \frac{e^{i (k+k_\tau) z}}{\sqrt{L_z}}
\left(\begin{array}{c} 0\\
h_{0 X}(x,y)
\end{array}\right)
\end{equation}
for the chiral ($n=0$) Landau level, and
\begin{equation}
\label{eq:spinor_ncll}
\langle{\bf r}|\Psi_{n X k \tau}\rangle= \frac{e^{i (k+k_\tau) z}}{\sqrt{L_z}}
\left(\begin{array}{c} u_{n k \tau} h_{|n|-1, X}(x,y)\\
v_{n k\tau} h_{|n|, X}(x,y)
\end{array}\right)
\end{equation}
for the nonchiral ($n\neq 0$) Landau levels,
where $k_\tau$ is the projection of the momentum of the Weyl node of chirality $\tau$ along the magnetic field,
\begin{align}
u_{n k\tau} &= \frac{1}{\sqrt{2}} \tau {\rm sign}(n) \sqrt{1+\frac{\hbar v_\tau k}{\epsilon_{n k \tau}}}\\
v_{n k\tau} &= \frac{1}{\sqrt{2}} \sqrt{1-\frac{\hbar v_\tau k}{\epsilon_{n k \tau}}}
\end{align}
are position-independent coefficients of the Landau-level spinors,
\begin{align}
h_{n X}(x,y) &= \frac{(-1)^{n}}{\sqrt{L_y}} e^{-i X y/l_B^2} \varphi_{n}(x-X)\\
\varphi_n(x) &=\left(\frac{1}{\pi l_B^2}\right)^{1/4}\frac{1}{\sqrt{2^n n!}} H_n\left(\frac{x}{l_B}\right) e^{-x^2/2 l_B^2}
\end{align}
are Landau-level wave functions and $H_n(x)$ are the Hermite polynomials.
It is useful to note that
\begin{align}
&\langle h_{|n|, X} | h_{|n'|, X'}\rangle\equiv \int dx dy\, h^*_{|n| X} h_{|n'| X'}= \delta_{X X'} \delta_{|n| |n'|},\nonumber\\
& u_{n k\tau}^2 + v_{n k \tau}^2 =1.
\end{align}
Using the shorthand notation, the free fermion Hamiltonian can be written as
\begin{equation}
{\cal H}_e^{(0)} = \sum_\alpha \epsilon_\alpha c^\dagger_\alpha c_\alpha,
\end{equation}
where $c^\dagger_\alpha\equiv c^\dagger_{n X k \tau}$ is a (dimensionless) operator that creates an electron in state $|\Psi_{n X k \tau}\rangle$.
For the moment, we neglect the single-particle hybridization gap due to magnetic tunneling between chiral Landau-levels of opposite chirality.
This gap can become measurable at high magnetic fields\cite{Zhang2017, Ramshaw2018}, provided that (i) the magnetic field is perpendicular to the wave vector that separates Weyl nodes in momentum space ; (ii) the magnetic length is comparable to or shorter than the inverse of the distance in momentum space that separates two Weyl nodes at zero field.
The hybridization gap depends strongly on the orientation of the magnetic field, and it is believed to be relatively negligible when the magnetic field is parallel to the wave vector connecting the two Weyl nodes\cite{Kim2017, Chan2017, Saykin2018}.
In Sec.~\ref{sec:real}, we will briefly comment on the implications of the hybridization gap in our theory.
\subsection{Electron-phonon interaction}
The electron-phonon interaction in the band eigenstate basis is given by
\begin{widetext}
\begin{equation}
{\cal H}_{\rm ep} =\sum_{\alpha \alpha'} \sum_{{\bf q}, j} \tilde{g}^{ep}_{\alpha\alpha', j}({\bf q}) c^\dagger_{\alpha} c_{\alpha'} (a^\dagger_{{\bf q},j}+a_{-{\bf q},j}),
\end{equation}
where $a^\dagger_{{\bf q},j}$ is a (dimensionless) operator that creates a phonon mode $j$ with momentum ${\bf q}$, and
\begin{align}
\label{eq:e_ph_me}
&\tilde{g}_{\alpha\alpha', j}^{ep}({\bf q})\equiv \tilde{g}_{\tau\tau',j}({\bf q})\langle \Psi_{n X, k, \tau} |e^{i {\bf q}\cdot{\bf r}}| \Psi_{n' X' k' \tau'}\rangle\nonumber\\
&=\tilde{g}_{\tau\tau',j}({\bf q})\delta_{k'+k_{\tau'}, k+k_\tau-q}\delta_{X',X+q_y l_B^2}\left[u_{n k\tau} u_{n' k' \tau'} \langle h_{|n|-1, X}|e^{i {\bf q}_\perp\cdot{\bf r}_\perp} | h_{|n'|-1, X'}\rangle
+ v_{n k\tau} v_{n' k' \tau'} \langle h_{|n|, X} |e^{i {\bf q}_\perp\cdot{\bf r}_\perp}| h_{|n'|, X'}\rangle\right]\nonumber\\
&=\tilde{g}_{\tau\tau',j}({\bf q})\delta_{k'+k_{\tau'}, k+k_\tau-q}\delta_{X',X+q_y l_B^2} e^{i q_x(X+X')/2}\left[u_{n k\tau} u_{n' k' \tau'} F_{|n|-1, |n'|-1}({\bf q}_\perp)+ v_{n k \tau} v_{n' k' \tau'} F_{|n|,|n'|}({\bf q}_\perp)\right]
\end{align}
is the electron-phonon matrix element (with dimensions of energy) in the low-energy Hilbert space spanned by Weyl fermions.
For brevity, Eq. (\ref{eq:e_ph_me}) has been written using the eigenstates of the nonchiral Landau levels (Eq.~(\ref{eq:spinor_ncll})).
Yet, it can be easily adapted to include chiral Landau levels via $u_{0,k,\tau}\to 0$ and $v_{0,k,\tau}\to 1$.
In the numerical estimates performed below, we will use\cite{Kaasbjerg2012, Mahan2000b}
\begin{equation}
\label{eq:ep_approx}
\tilde{g}_{\tau\tau',j}({\bf q})^2 \sim\frac{\hbar d_{{\rm op},j}^2}{2 \rho {\cal V} \omega_{0,j}({\bf q})}
\end{equation}
for long-wavelength optical phonons, where ${\cal V}$ is the crystal volume, $\rho$ is the atomic mass density, $d_{{\rm op},j}$ is the optical deformation potential (dimensions of energy divided by length) and $\omega_{0,j}({\bf q})$ is the unperturbed phonon frequency for mode $j$. For the case of long-wavelength acoustic phonons, we replace $d_{{\rm op},j} \to |{\bf q}| d_{{\rm ac},j}$, where $d_{{\rm ac},j}$ is the acoustic deformation potential (dimensions of energy).
In Eq.~(\ref{eq:e_ph_me}), we have defined the form factors\cite{MacDonald1986,Goerbig2011}
\begin{equation}
\label{eq:form_factor}
F_{n m}({\bf q}_\perp) \equiv \sqrt{\frac{{\rm min}(n,m)!}{{\rm max}(n,m)!}} \left(\frac{(\mp q_y - i q_x) l_B}{\sqrt{2}}\right)^{|n-m|} L_{{\rm min}(n,m)}^{|n-m|}\left(\frac{q_\perp^2 l_B^2}{2}\right) e^{-q_\perp^2 l_B^2/4},
\end{equation}
where $n,m$ are nonnegative integers, the $+$ and $-$ signs correspond to the cases $n>m$ and $n<m$, respectively, and the functions $L_n^\alpha$ are the generalized Laguerre polynomials.
Also, we have used ${\bf q}=q \hat{\bf z} + {\bf q}_\perp$, ${\bf r}= z \hat{\bf z} + {\bf r}_\perp$.
We note that, for long-wavelength phonons ($q$ smaller than the internodal distance), Eq.~(\ref{eq:e_ph_me}) vanishes unless $\tau=\tau'$.
In addition, in the limit ${\bf q}_\perp\to 0$, Eq.~(\ref{eq:e_ph_me}) becomes
\begin{equation}
\langle \Psi_{n X, k, \tau} |e^{i q z}| \Psi_{n' X' k' \tau'}\rangle = \delta_{k'+k_{\tau'}, k+k_\tau-q}\delta_{|n|, |n'|} \delta_{X, X'}
\left(u_{n,k,\tau} u_{n', k', \tau'} + v_{n,k,\tau} v_{n', k', \tau'}\right).
\end{equation}
In Eq. (\ref{eq:e_ph_me}), we have assumed (for simplicity) that the electron-phonon coupling is diagonal in the pseudospin basis.
Admittedly, for a generic nesting wave vector, there is no symmetry-based reason that would preclude the electron-phonon coupling from having pseudospin-dependent terms. However, the pseudospin structure of the electron-phonon coupling is not qualitatively crucial for our theory; what is crucial is that the coupling between Fermi points connected by the nesting wave vector be nonzero. Any pseudospin-dependence of such electron-phonon interaction will simply modify (by a numerical factor) the effective electron-phonon coupling appearing in our expressions for the critical temperature in Secs \ref{sec:cll} and \ref{sec:noncll}.
\subsection{Electron-electron interaction}
The electron-electron interaction in the band eigenstate basis is given by
\begin{align}
{\cal H}_{ee} \simeq \frac{1}{2 {\cal V}} \sum_{\bf q} \sum_{\{\alpha\}}
\tilde{g}_{\alpha_1\alpha_2}^{\alpha_3\alpha_4} ({\bf q})c^\dagger_{\alpha_1} c_{\alpha_3} c^\dagger_{\alpha_2} c_{\alpha_4},
\end{align}
where we have neglected umklapp processes,
\begin{equation}
\label{eq:e_e_me}
\tilde{g}_{\alpha_1\alpha_2}^{\alpha_3\alpha_4} ({\bf q}) \equiv V({\bf q}) \langle\Psi_{n_1 X_1 k_1 \tau_1}|e^{i {\bf q}\cdot{\bf r}} |\Psi_{n_4 X_4 k_4 \tau_4}\rangle \langle\Psi_{n_2 X_2 k_2 \tau_2}|e^{-i {\bf q}\cdot{\bf r}} |\Psi_{n_3 X_3 k_3 \tau_3} \rangle
\end{equation}
\end{widetext}
is the Coulomb matrix element, $V({\bf q})=e^2/(\epsilon_0\epsilon_\infty {\bf q}^2)$ is the Coulomb potential and $\epsilon_\infty$ is the cutoff-dependent contribution from "high-energy" electrons (of energy exceeding the ultraviolet RG cutoff) to dielectric screening.
Screening effects originating from low-energy fermions will be discussed below.
\subsection{Partition function}
In what follows, we will apply the renormalization group approach\cite{Bourbon89} to the Peierls instability of the above model.
The starting point is the functional integral representation of the partition function in the presence of a magnetic field along $z$,
\begin{equation}
\label{eq:Z}
Z=\int \mathfrak{D}\psi^\dagger \mathfrak{D}\psi\int D\phi
\,e^{S[\psi^\dagger,\psi,\phi,h]/\hbar},
\end{equation}
which is expressed in terms of a trace over the fermion ($\psi^{(\dagger)})$ and phonon ($\phi$) fields. Here, $S$ stands for the action
\begin{align}
S[\psi^\dagger,\psi,\phi,h] &= \ S^{(0)}_e[\psi^\dagger,\psi]+ S_{ee}[\psi^\dagger,\psi]\nonumber\\
&+ S^{(0)}_p[\phi]+S_{ep}[\psi^\dagger,\psi,\phi] + S_{h}[\psi^\dagger,\psi,h],
\end{align}
which comprises five parts.
First, the free electron action is
\begin{equation}
\label{eq:freefer}
S_e^{(0)}[\psi^\dagger,\psi] =\sum_\alpha \sum_{\omega_n}\left(i\hbar\omega_n+\mu-\epsilon_\alpha\right)\psi^\dagger_\alpha (\omega_n)\psi_\alpha (\omega_n),
\end{equation}
where
$\psi_\alpha$ and $\psi_{\alpha'}^\dagger$ are Grassmann fields (with dimensions $\sqrt{\rm time}$), $\omega_n=(2n+1)\pi k_BT/\hbar$ ($n\in\mathbb{Z}$) are fermionic Matsubara frequencies at temperature $T$ and $\mu$ is the chemical potential.
In Eq.~(\ref{eq:freefer}), we have used the convention
\begin{equation}
\psi_\alpha(t)=\sqrt{\frac{1}{\hbar\beta}} \sum_{\omega_n} e^{-i\omega_n t} \psi_\alpha(\omega_n),
\end{equation}
where $\beta=1/k_B T$.
Second, the free phonon action is
\begin{equation}
S_p^{(0)} [\phi]= -\sum_{{\bf q},j} \sum_{\omega_m} [\mathscr{D}^0_j({\bf q},\omega_m)]^{-1}|\phi_j({\bf q},\omega_m)|^2,
\end{equation}
where
\begin{equation}
\label{D}
\mathscr{D}_j^0({\bf q},\omega_m) = [\omega_m^2+\omega_{0,j}({\bf q})^2]^{-1}
\end{equation}
is the free phonon propagator and
$\omega_m=2\pi k_BT m/\hbar$ ($m\in\mathbb{Z}$) are bosonic Matsubara frequencies, $\omega_{0,j}(\bf q)$ is the bare phonon dispersion, and
$\phi_j({\bf q},\omega_m)$ is the phonon displacement field for mode $j$ with wave vector ${\bf q}$ (with dimensions of $\sqrt{\rm action} \times {\rm time}$).
In second quantized form,
\begin{equation}
\phi_j ({\bf q},t) \to \frac{1}{2}\sqrt{\frac{\hbar}{\omega_{0,j}({\bf q})}} \left (a_{{\bf q},j}(t)^\dagger + a_{-{\bf q}, j}(t)\right).
\end{equation}
Third, the action for the electron-phonon interaction is
\begin{widetext}
\begin{equation}
S_{ep}[\psi^\dagger,\psi,\phi]=-\int_0^{\hbar\beta} dt {\cal H}_{\rm ep}(t)=-\sqrt{\frac{\pi v_F}{\beta \cal V}}\sum_{\omega_n,\omega_m} \sum_{{\bf q},j} \sum_{\alpha,\alpha'} z_{\alpha\alpha', j} g^{ep}_{\alpha\alpha', j}({\bf q})\psi^\dagger_{\alpha}(\omega_n+\omega_m) \psi_{\alpha'}(\omega_n) \phi_j({\bf q},\omega_m),
\end{equation}
where $z_{\alpha\alpha', j}$ stands as a renormalization factor of the electron-phonon coupling ($z_{\alpha\alpha', j}=1$ at the bare level) and
\begin{equation}
\label{eq:ep_conv}
g^{ep}_{\alpha \alpha', j}({\bf q})= \frac{2}{\hbar} \sqrt{\frac{\omega_{0,j}({\bf q}) {\cal V}}{\pi v_F}}\tilde{g}^{ep}_{\alpha\alpha',j}({\bf q})
\end{equation}
is the electron-phonon coupling with dimensions of velocity ($\tilde{g}_{\alpha\alpha', j}^{ep}$ was defined in Eqs.~(\ref{eq:e_ph_me}) and~(\ref{eq:ep_approx})).
The action for electron-electron interactions is
\begin{equation}
S_{ee}[\psi^\dagger,\psi] =-\int_0^{\hbar\beta} dt {\cal H}_{\rm ee}(t)=- \frac{\pi v_F}{\beta \cal V} \sum_{\{\alpha\},\{\omega_n\}} \sum_{\bf q}g_{\alpha_1\alpha_2}^{\alpha_3\alpha_4}({\bf q}) \psi^\dagger_{\alpha_1}(\omega_{n_1}) \psi^\dagger_{\alpha_2}(\omega_{n_2})\psi_{\alpha_3}(\omega_{n_3})\psi_{\alpha_4}(\omega_{n_1}+\omega_{n_2}-\omega_{n_3}),
\end{equation}
where
\begin{equation}
\label{eq:ee_conv}
g_{\alpha_1\alpha_2}^{\alpha_3\alpha_4}({\bf q})=\frac{\tilde{g}_{\alpha_1\alpha_2}^{\alpha_3\alpha_4}({\bf q})}{2\pi \hbar v_F}
\end{equation}
is the electron-electron coupling parameter with dimensions of length squared ($\tilde{g}_{\alpha_1\alpha_2}^{\alpha_3\alpha_4}$ was defined in Eq.~(\ref{eq:e_e_me})).
Finally, the source field contribution for susceptibilities is given by
\begin{equation}
\label{sh0}
S_{h}[\psi^\dagger,\psi,h] = \sqrt{\frac{\pi v_F}{\beta \cal V}} \sum_{\nu}\sum_{{\bf q},\omega_m} \sum_{\sigma,\sigma'} z_{\nu;\sigma,\sigma'} \Big\{O^\dagger_{\nu; \sigma,\sigma'} (\omega_m) h_{\nu; \sigma,\sigma'}({\bf q},\omega_m) + O_{\nu; \sigma,\sigma'} (\omega_m) h^*_{\nu; \sigma,\sigma'}({\bf q},\omega_m)\Big\},
\end{equation}
where $\sigma$ and $\sigma'$ are pseudospin labels, $h_{\nu;\sigma,\sigma'}$ are the source fields, and the composite fields for charge density-wave ($\nu={\rm CDW}$) and Cooper pairing ($\nu={\rm SC}$) take their respective forms:
\begin{equation}
\label{Ocdw}
O_{{\rm CDW};\sigma,\sigma'} (\omega_m) = \sum_{\alpha,\alpha'}\sum_{\omega_n} f_{\alpha,\alpha'; \sigma,\sigma'}^{\rm CDW}\psi^\dagger_{\alpha}(\omega_n+\omega_m) \psi_{\alpha'}(\omega_n)
\end{equation}
and
\begin{equation}
\label{Osc}
O_{{\rm SC};\sigma,\sigma'} (\omega_m) = \sum_{\alpha,\alpha'}\sum_{\omega_n} f_{\alpha,\alpha'; \sigma,\sigma'}^{\rm SC}\psi_{\alpha}(\omega_n+\omega_m) \psi_{\alpha'}(\omega_n).
\end{equation}
Here, $z_{\nu;\sigma,\sigma'}$ stands for the renormalization factor for vertex part of the susceptibility, and $f_{\alpha,\alpha';\sigma,\sigma'}^\nu$ is the related form factor (more on this below).
For a CDW susceptibility that is pseudospin-independent, we take $h_{{\rm CDW};\sigma,\sigma'} \propto \delta_{\sigma,\sigma'}$; for a pseudospin-singlet Cooper pairing susceptibility, we take $h_{{\rm SC};\sigma,\sigma'} \propto \delta_{\sigma,-\sigma'}$, and so on.
\subsection{Method}
\label{subsec:method}
Starting from Eq.~(\ref{eq:Z}), we perform successive partial integrations of fermion fields located in the outer energy shell of width $\Lambda_n(l) dl$ on both sides of the Fermi level. For each outer shell integration a complete trace over Matsubara frequencies is carried out from which the temperature dependence of the flow is obtained.
Here, $\Lambda_n(l)=\Lambda_n e^{-l}$ is the distance in energy from the Fermi level to the outer shell at step $l$ of the integration ($l>0$), $2\Lambda_n$ is the initial bandwidth cutoff of the Landau band $n$, and $dl\ll 1$ is the integration step.
The fermion fields in the outer shell are denoted as $\{\bar{\psi}, \bar{\psi}^\dagger\}$.
The recursive procedure is carried out perturbatively, keeping fixed the inner ($<$) shell variables and using $S^{(0)}_e[\bar{\psi}^\dagger,\bar{\psi}]$ as the outer shell free fermion part. This yields
\begin{align}
\label{RGTransf}
Z &\sim \int D\phi\int_< \mathfrak{D}\psi^\dagger\, \mathfrak{D}\psi \,e^{S[\psi^\dagger,\psi,\phi,h]_l/\hbar} \int \mathfrak{D}\bar{\psi}^\dagger \mathfrak{D}\bar{\psi} \,e^{S[\psi^\dagger,\psi,\bar{\psi}^\dagger,\bar{\psi},\phi,h]_{dl}/\hbar}\nonumber\\
& \propto \int D\phi\int_< \mathfrak{D}\psi^\dagger \mathfrak{D}\psi \,e^{S[\psi^\dagger,\psi,\phi,h]_{l+dl}/\hbar}.
\end{align}
Using the linked cluster theorem, the outer shell integration at the one-loop level yields the recursion relations
\begin{align}
\label{Sp0}
S_p^{(0)}[\phi]_{l+dl} &= S_p^{(0)}[\phi]_l+\frac{1}{2}\langle S_{e p}^2\rangle_{_{dl}} + {\cal O}(\phi^4) \\
\label{zeph0}
S_{e p}[\psi^\dagger,\psi,\phi]_{l+dl} &= S_{e p}[\psi^\dagger,\psi,\phi]_l+\langle S_{e p} S_{ee}\rangle_{_{dl}}+\ldots \\
\label{g1g2}
S_{ee}[\psi^\dagger,\psi]_{l+dl} &= S_{ee}[\psi^\dagger,\psi]_l + \frac{1}{2} \langle S_{ee}^2\rangle_{_{dl}}+ \ldots\\
\label{sh}
S_{h}[\psi^\dagger,\psi,h]_{l+dl} &= S_{h}[\psi^\dagger,\psi,h]_l+\langle S_{h} S_{ee}\rangle_{_{dl}}+\ldots
\end{align}
which are expressed in terms of free fermion averages $\langle \ldots \rangle_{_{dl}}$ in the outer energy shell.
Equation (\ref{Sp0}) gives the recursion relation for the renormalization of the purely phonon part of the action. At the harmonic $\phi^2$ level it leads to the phonon the self-energy correction $\delta \scalebox{1.55}{$\pi$} $ to the phonon propagator via $\mathscr{D}_{j,l}^{-1} = \mathscr{D}_{j,0}^{-1} - \scalebox{1.55}{$\pi$}_{j,l} $, as shown in Fig.~\ref{fig_RG1}(a). For simplicity, we neglect the hybridization between different phonon modes induced by the electron-phonon coupling; this is a reasonable approximation for phonon modes that are well-separated in energy from the rest.
Accordingly, $\scalebox{1.55}{$\pi$}$ and $\mathscr{D}$ have a single phonon mode index. At the one-loop level, we do not consider electronic self-energy corrections to $S_e^{(0)}$ since for effective momentum independent couplings these only affect the chemical potential which can rescaled back to its original value at each step of the RG at fixed band filling.
Equation (\ref{zeph0}), depicted by Fig.~\ref{fig_RG1}(b), refers to the renormalization of the electron-phonon vertex part $z_{\alpha,\alpha',j}$ due to electron-electron interactions and Eq. (\ref{g1g2}), depicted in Fig.~\ref{fig_RG1}(c), gives the recursion relation for a combination of electron-electron couplings $g_{\alpha_1,\alpha_2}^{\alpha_3,\alpha_4} $ involved in the renormalization of $z_{\alpha\alpha',j}$ and source field vertex $z_\nu$ of Fig.~\ref{fig_RG1}(d).
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.48]{fig1.pdf}
\caption{Renormalization group flow equations for (a) the phonon self-energy; (b) electron-phonon vertex part; (c) electron-electron scattering amplitudes (full square); (d) and source field vertex for CDW (SC) susceptibility. The full and dotted thick (thin) loop fermion lines refer to outer (inner) energy shell fermions near the right and left Fermi points of the Landau band $n$ (we consider $|n|=0$ in Sec. \ref{sec:cll} and $|n|=0, 1$ in Sec. \ref{sec:noncll}). }
\label{fig_RG1}
\end{center}
\end{figure}
\end{widetext}
\section{Peierls instability from chiral Landau levels}
\label{sec:cll}
We begin by considering the theory of the Peierls instability in the quantum limit, where the Fermi energy intersects only the chiral Landau levels.
\subsection{Electron-phonon and electron-electron couplings}
Figure~\ref{fig:ql} shows the low-energy electronic structure in the quantum limit.
The distance between the two Fermi points along the direction of the magnetic field is denoted as $2 k_F= k_{F+} - k_{F-} $ (see Fig.~\ref{fig:ql}).
According to this definition, the Fermi wave vector depends on the distance between the Weyl nodes (note that a different definition for $k_F$ will be used when we treat nonchiral Landau levels in Sec. \ref{sec:noncll}).
We take the magnitudes of the Dirac velocities to be $v_\tau = v_F$ for $\tau=1$ and $v_\tau=v_F + \delta v$ for $\tau=-1$, where $\delta v \geq -v_F$ (the latter condition is due to the fact that the two nodes have opposite chirality).
Assuming that the total carrier density per unit volume $n_e$ is independent of the magnetic field, the $B-$dependence of $k_F$ in the quantum limit is obtained from
\begin{equation}
2 k_F = |4\pi^2 l_B^2 n_e+b|,
\end{equation}
where $b=|k_+-k_-|$ is the momentum separation between the two Weyl nodes along the direction of the magnetic field.
Here, $n_e$ is defined by counting the number of electronic states between $k_\tau$ and the Fermi energy for each node $\tau$, and then summing over the two nodes.
For an overall electron-doped system ($n_e>0$), $2 k_F > b$ and $k_F$ decreases with $B$.
For an overall hole-doped system ($n_e<0$), ${\rm min}({k_F})=0$.
Regardless of the doping, $2 k_F \to b$ as $B\to\infty$.
Thus, at very high magnetic field, $k_F$ can be made small by orienting the magnetic field perpendicular to the direction of separation between the two nodes (because $b=0$ in that case).
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\columnwidth]{fig2.pdf}
\caption{Landau level spectrum of two Weyl nodes of opposite chirality (Eq.~(\ref{eq:LL_energies})) in the quantum limit, where the Fermi energy $\epsilon_F$ intersects only chiral Landau levels. The parameter values in Eq. (\ref{eq:LL_energies}) are $v_+=v_F$, $v_-=1.5 v_F$, $b_0=0.1 \hbar v_F/l_B$. The distance between the two Fermi points is denoted as $2 k_F=k_{F+} - k_{F-}$, where $k_{F+}$ and $k_{F-}$ are the Fermi wave vectors measured from the center of the Brillouin zone (along the direction of the magnetic field).
The locations of the Weyl nodes, $k_+$ and $k_-$, are also indicated and measured from the center of the Brillouin zone (along the direction of the magnetic field).
}
\label{fig:ql}
\end{center}
\end{figure}
We will be interested in phonons whose wave vectors connect the two Fermi points, as these are subject to a Peierls instability.
We believe such internodal electron-phonon scattering is generically allowed, despite assertions to the contrary in the literature\cite{nguyen2020}.
Indeed, from a group theory point of view, there is no selection rule that forbids electron-phonon scattering between nodes of opposite chirality.
In other words, because the little group of Weyl nodes located at arbitrary points in the Brillouin zone comprises only the identity operation, generic phonon modes of the crystal can scatter electrons between two nodes of opposite chirality.
From Eq.~(\ref{eq:e_ph_me}), the bare electron-phonon matrix element for the chiral Landau level can be written as
\begin{align}
\label{eq:cll_me}
g_{\alpha\alpha',j}^{ep} = & g_{\tau\tau',j} \langle\Psi_{0 X, k, \tau} |e^{i {\bf q}\cdot{\bf r}}| \Psi_{0 X' k' \tau'}\rangle\nonumber\\
=& g_{\tau\tau',j} \delta_{k'+k_{\tau'}, k+k_\tau-q}\delta_{X',X+q_y l_B^2}\nonumber\\
& \times e^{i q_x(X+X')/2} e^{-q_\perp^2 l_B^2/4},
\end{align}
where $g_{\tau\tau',j}$ is related to $\tilde{g}_{\tau\tau',j}$ in Eq.~(\ref{eq:ep_approx}) via the relation (\ref{eq:ep_conv}).
When $q_\perp l_B\gtrsim 1$, the electron-phonon matrix element is exponentially suppressed.
For a phonon wave vector connecting the two Fermi points, we set $\tau=-\tau'$ in Eq.~(\ref{eq:cll_me}).
Hereafter, we label $g_{\tau,-\tau,j}\equiv g_{x,j}$ and we neglect the momentum-dependence of $g_{x,j}$.
The latter approximation is justified because we will be interested in phonon momenta close to $2 k_F \hat{\bf z}$ and, in such vicinity, the ${\bf q}$-dependence of $g_{x,j}$ will be smooth and qualitatively unimportant for our theory.
The matrix elements for the vertex part of electronic CDW and SC susceptibilities lead to the form factors entering in Eqs. (\ref{Ocdw}) and (\ref{Osc}),
\begin{align}
\label{fcdw}
f_{\alpha\alpha';\sigma,\sigma'}^{\rm CDW} & = \delta_{\sigma,\downarrow}\delta_{\sigma',\downarrow}\delta_{k'+k_{\tau'}, k+k_\tau-q}\delta_{X',X+q_y l_B^2}\nonumber\\
&\times e^{i q_x (X+X')/2} e^{-q_\perp^2 l_B^2/4} \\
\label{fsc}
f_{\alpha\alpha';\sigma,\sigma'}^{\rm SC} & = \delta_{\sigma,\downarrow}\delta_{\sigma',\downarrow}\delta_{k'+k_{\tau'}, -k-k_\tau-q}\delta_{X',-X+q_y l_B^2}\nonumber\\
& \times e^{i q_x (X+X')/2} e^{-\left[q_x^2+\left(q_y-2X/l_B^2\right)^2\right]l_B^2/4},
\end{align}
where we have used the fact that the $n=0$ Landau levels are pseudospin-polarized (recall Eq.~(\ref{eq:spinor_cll})). This allows us to omit the $\sigma$ and $\sigma'$ labels from the CDW and SC renormalization factors, denoting them simply as $z_{\rm CDW}$ and $z_{\rm SC}$.
Similarly, the Coulomb matrix elements for the electron-electron interactions (see Eqs. (\ref{eq:e_e_me}) and (\ref{eq:ee_conv})) read
\begin{align}
\label{eq:cll_me2}
g_{\alpha_1\alpha_2}^{\alpha_3\alpha_4} ({\bf q}) =& \frac{V({\bf q})}{2\pi\hbar v_F} e^{i q_x(X_1+X_4)/2} e^{-i q_x(X_2+X_3)/2} e^{-q_\perp^2 l_B^2/2}\nonumber\\
&\times \delta_{X_4,X_1+q_y l_B^2} \delta_{X_3,X_2-q_y l_B^2} \nonumber\\
&\times \delta_{k_4+k_{\tau_4}, k_1+k_{\tau_1}-q} \delta_{k_3+k_{\tau_3}, k_2+k_{\tau_2}+q}.
\end{align}
In what follows, we separate $g_{\alpha_1\alpha_2}^{\alpha_3\alpha_4}$ in two parts, which are the most important marginal couplings involved in the lattice instabilities.
The first part is the long-wavelength part, known as the forward Coulomb scattering.
For this scattering, the momentum transfer ${\bf q}$ is small compared to the internodal distance.
Thus, we take $\tau_1=\tau_4$ and $\tau_2=\tau_3$ in Eq.~(\ref{eq:cll_me2}).
Moreover, the divergence of the Coulomb potential at ${\bf q}\to 0$ is removed by replacing
\begin{equation}
\label{eq:Vsc}
V({\bf q}) \to \lim_{q\to 0}V({\bf q})/\epsilon({\bf q},0)\equiv g_2 \pi \hbar v_F,
\end{equation}
where
\begin{equation}
\epsilon({\bf q},\omega) = 1-V({\bf q}) \Pi({\bf q},\omega)
\end{equation}
is the contribution of Weyl fermions to the dielectric function,
\begin{equation}
\Pi({\bf q},\omega) = \frac{1}{{\cal V}}\sum_{\alpha,\alpha'}|\langle \Psi_{\alpha}|e^{i{\bf q}\cdot{\bf r}}|\Psi_{\alpha'}\rangle|^2 \frac{f_\alpha-f_{\alpha'}}{E_\alpha-E_{\alpha'}+\hbar\omega+i 0}
\end{equation}
is the electronic polarization function in the random phase approximation (RPA) and $f_\alpha$ is the Fermi distribution.
The reason for using static screening in Eq.~(\ref{eq:Vsc}) is that frequency-dependence is irrelevant in the RG sense.
Then, keeping only the chiral Landau level contribution to the zero-temperature polarization function \cite{rinkel2019}, which is justified for a UV cutoff $\Lambda_0$ that does not exceed the distance in energy between the $n=0$ and $n=1$ Landau levels, we get
\begin{equation}
\label{g20}
g_2= 2 \pi l_B^2 \frac{v_F+\delta v}{v_F + \delta v/2}.
\end{equation}
We emphasize that this RPA expression for $g_2$ is only the bare or initial value of the coupling in the RG flow, i.e. $g_2\equiv g_2(l=0)$.
We can rewrite Eq.~(\ref{g20}) as $g_2 \pi \hbar v_F = e^2/(\epsilon_\infty q_{TF}^2)$,
where
\begin{equation}
q_{TF}^2 = \frac{e^2}{\epsilon_\infty} \nu(\epsilon_F)
\end{equation}
is the square of the Thomas-Fermi screening wave vector and
\begin{equation}
\nu(\epsilon_F) = \frac{1}{4\pi^2 l_B^2 \hbar v_F}\frac{2+\delta v/v_F}{1+\delta v/v_F}
\end{equation}
is the density of states at the Fermi level.
For $|\delta v|/v_F\ll 1$, $(q_{TF} l_B)^2$ is equal to an effective "fine structure constant" $e^2/(2\pi^2 \hbar v_F \epsilon_\infty)$, which is smaller than unity in typical Weyl semimetals.
Thus, we expect $q_{TF} l_B \lesssim 1$ in common Weyl semimetals in the quantum limit.
The second part of $g_{\alpha_1\alpha_2}^{\alpha_3\alpha_4}$ is the "short"-wavelength part, known as the backward Coulomb scattering.
Here, the momentum transfer ${\bf q}$ connects the two nodes of opposite chirality.
Thus, we take $\tau_1=-\tau_4$, $\tau_2=-\tau_3$, $\tau_4=-\tau_3$ in Eq.~(\ref{eq:cll_me2}) and we replace
\begin{equation}
V({\bf q}) \to V (2 k_F \hat{\bf z}+{\bf q}_\perp).
\end{equation}
Then, we define
\begin{align}
\label{g10}
g_1&=\frac{1}{\pi \hbar v_F}\frac{\int_0^{1/l_B} dq_\perp\,q_\perp V(2 k_F\hat{\bf z}+{\bf q}_\perp)}{\int_0^{1/l_B} dq_\perp\, q_\perp} \nonumber\\
&= \frac{e^2 l_B^2}{\pi\hbar v_F \epsilon_\infty} \ln\left(1+\frac{1}{4 k_F^2 l_B^2}\right).
\end{align}
Like for $g_2$ in Eq.~(\ref{g20}), the value of $g_1$ in Eq.~(\ref{g10}) is only the initial (bare) value of the coupling in the RG flow.
Two remarks are in order here. First, our expression for the bare $g_1$ is unscreened. This is well justified when $k_F l_B\gtrsim 1$. In such regime, $(2 k_F)^2 \gg q_{TF}^2$ and the non-logarithmic part of the RPA screening (coming from the "fast" electrons) is qualitatively unimportant.
Regarding the logarithmically divergent part of the RPA screening at $q=2 k_F$, it is not to be included in the bare $g_1$, but rather it is treated below in the course of the RG flow.
Second, the averaging over the transverse momentum ${\bf q}_\perp$ with a cutoff $1/l_B$ in Eq.~(\ref{g10}) is motivated by simplicity (so that the effective electron-electron coupling is independent of wave vector) and by the fact that the Coulomb matrix element in Eq.~(\ref{eq:cll_me2}) is exponentially suppressed for $q_\perp l_B\gtrsim 1$.
This averaging procedure is justified provided that $k_F l_B$ is not small. When $k_F l_B\to 0$, the backward Coulomb scattering becomes singular and this singularity has to be treated on an equal footing as the one in the forward Coulomb scattering. The study of the regime $k_F l_B \ll 1$ goes beyond the scope of our paper.
The relative strength between short- and long-range Coulomb interactions plays an important role in the Peierls instability.
As we show below, a larger value of $g_1/g_2$ results in a lower critical temperature for the Peierls instability.
From Eqs.~(\ref{g20}) and (\ref{g10}), the ratio between the bare couplings is
\begin{equation}
\frac{g_1}{g_2} = \frac{e^2}{2\pi^2 \hbar v_F \epsilon_\infty} \frac{1+ \delta v/2 v_F}{1 +\delta v/v_F}\ln\left(1+\frac{1}{4 k_F^2 l_B^2}\right),
\end{equation}
which depends on the magnetic field explicitly via $l_B$ and implicitly via $k_F$.
When $k_F l_B\gg 1$, we have $g_1/g_2 \sim q_{TF}^2/k_F^2$ (note that $k_F l_B\gg 1$ is {\em not} incompatible with being in the quantum limit, provided that the Weyl nodes are well-separated in momentum space).
When $k_F l_B\lesssim 1$, $g_1/g_2 \sim (q_{TF} l_B)^2 \ln[1+1/(2 k_F l_B)^2]$
increases with $B$.
In summary, for a weakly interacting Weyl semimetal (small fine structure constant),
one expects $g_1/g_2\ll 1$.
However, for more strongly interacting systems with $k_F l_B\lesssim 1$, it is possible that $g_1 \simeq g_2$.
\subsection{RG equations for the adiabatic Peierls instability}
The outer shell corrections entering the RG recursive relations (Eqs.~(\ref{Sp0}), (\ref{zeph0}) and (\ref{g1g2})) and calculated in Appendix A are all evaluated in the limit of small $q_\perp $ up to the cutoff $1/l_B$, above which all the couplings are
exponentially suppressed (see Eqs. (\ref{eq:cll_me}) and (\ref{eq:cll_me2})). Within the magnetic length scale $l_B$, the couplings can then be considered as essentially local in the transverse directions, which yield one-dimensional-like loop corrections for the flow equations.
From Eq. (\ref{zeph0}) and the results of Appendix A, the RG transformation for the internode electron-phonon vertex part $z_{\tau,-\tau}$ (which renormalizes the bare coupling $g_{x,j}$ introduced below Eq.~(\ref{eq:cll_me}))
is purely electronic and independent of the phonon mode $j$; it leads to the scaling equation
\begin{equation}
\label{zpm}
\frac{d}{dl} \ln z_{\tau,-\tau}= {\alpha_0\over 2} { (g_2-g_1)}\lambda_P,
\end{equation}
where
\begin{equation}
\label{alpha0}
\alpha_0=\frac{1}{2\pi l_B^2 (1+ \delta v/2v_F)}
\end{equation}
is a factor that originates from the three-dimensionality of the system, and $\lambda_P(l,T,\delta v)$ is the loop cutoff function of the Peierls scattering channel resulting from an exact summation over intermediate Matsubara fermion frequencies ($0\le\lambda_P\le 1$, see Eq. (\ref{IPS})).
The flow of the electron-phonon vertex then depends on the combination of electron-electron couplings $g_2-g_1$ governed by Eq. (\ref{g1g2}).
The derivations in Appendix A at the one-loop level lead to the scaling equations
\begin{align}
\label{g1g2ChiralA}
{dg_1\over dl} & = -{1\over 2}g_1^2 \alpha_0\lambda_P + g_2g_1\alpha_0(\lambda_P + \lambda_C) \\
{dg_2\over dl} & = {1\over 2}g_1^2\alpha_0 \lambda_C + {1\over 2}g_2^2\alpha_0(\lambda_P + \lambda_C),
\label{g1g2ChiralB}
\end{align}
which can be merged to give
\begin{equation}
\label{g2mg1}
{d(g_2-g_1)\over dl} = {1\over 2} \alpha_0 (g_2-g_1)^2 (\lambda_P+ \lambda_C),
\end{equation}
as a relevant combination of couplings in presence of effectively spinless fermions shown in Fig. \ref{fig_RG1} (c). In our local transverse scheme for the couplings, this equation is similar to the one obtained by Yakovenko for the problem of 3D electron gas in the presence of strong magnetic field\cite{Yakovenko93}. Here $\lambda_C(l,T,\Delta,\delta v)$ is the loop cutoff function of the Cooper scattering channel, as obtained from the summation over Matsubara fermion frequencies ($-1\leq\lambda_C\leq 0$, see Eq. (\ref{Cooper})). This $\lambda_C$ is opposite in sign to $\lambda_P$ and it is also cut off by the energy scale
\begin{equation}
\label{eq:Delta_main}
\Delta\equiv \hbar v_F(k_{F+}+k_{F-})
\end{equation}
associated to mirror symmetry-breaking.
In the RG equations (\ref{zpm}), (\ref{g1g2ChiralA}) and (\ref{g2mg1}), the variables $g_1$, $g_2$, $\lambda_C$ and $\lambda_P$ vary with $l$ (they are not to be confused with their bare values). In what follows we shall proceed to the integration of the flow equations as a function of $l$. At sufficiently large $l $, the functions $\lambda_{P,C}$ will cut-off the flow providing in turn either a $T$ or a $\Delta$ dependence to various quantities depending whether $k_BT< \Delta$ or $k_BT> \Delta$. From the $T$ dependence, for instance, one can then extract the temperature scale of various instabilities of the model.
When the two Weyl nodes are related by an improper symmetry operation that remains unbroken in the presence of the external magnetic field, we get $\Delta=0$.
Such is the case, for instance, if at zero magnetic field the two nodes were related to one another by a mirror plane and the applied magnetic field is oriented perpendicular to that plane.
If the magnetic field is tilted away from the normal to the mirror plane, one generically has $\Delta\neq 0$.
Thus, the energy scale $\Delta$ is partially tunable with the magnitude and orientation of the magnetic field.
Of course, if the two nodes are unrelated by an improper symmetry at zero magnetic field, then $\Delta\neq 0$ irrespective of the magnetic field.
When ${\bf q} = 2 k_F \hat{\bf z}$, the RG transformation for the phonon propagator of mode $j$
leads to the phonon self-energy corrections $\delta \scalebox{1.55}{$\pi$}_j $ shown in Fig.~\ref{fig_RG1}~(a) and in Eq. (\ref{Sp0}). In the adiabatic case, these corrections are purely electronic in character. At the one-loop level at step $\ell$, one has
\begin{align}
\label{Kohn_chiral}
\mathscr{D}_{j,l}^{-1} = &\, \mathscr{D}_{j,0}^{-1} - \scalebox{1.55}{$\pi$}_{j,l}\cr
= & \, \omega_m^2 + \omega_{0,j}^2(\mathbf{q})\left[1- \frac{\alpha_0}{2} g_{x,j}^{\prime2}\chi(l)\right],
\end{align}
where $g_{x,j}'\equiv g_{x,j}/\omega_{0,j}$ and
\begin{equation}
\label{Ki0}
\chi(l) = \int_0^l z_{\tau,-\tau}^2(l')\lambda_P dl'
\end{equation}
is the charge density-wave susceptibility of the electron system along $\hat{\bf z}$ and near $2k_F$.
Equation~(\ref{Kohn_chiral}) describes the softening of phonons due to their coupling to Weyl fermions.
The vanishing of the term inside the square bracket sets the temperature scale for the Peierls instability.
For pedagogical reasons, we compare Eq.~(\ref{Kohn_chiral}) with the standard RPA expression for the renormalized phonon frequency (see e.g. Refs. [\onlinecite{mahan2000, rinkel2019}]).
The connection emerges if we set the bare $g_2$ to zero and if we neglect $\lambda_C$ in our RG equations.
Under those conditions, Eq.~(\ref{Ki0}) becomes
\begin{equation}
\label{eq:chi_RPA}
\chi(l)=\frac{\int_0^l \lambda_P dl'}{1+ \frac{\alpha_0 g_1}{2} \int_0^l \lambda_P dl'},
\end{equation}
where $g_1$ here is the bare coupling for the backward Coulomb scattering.
If we substitute Eq.~(\ref{eq:chi_RPA}) in Eq.~(\ref{Kohn_chiral}), we recover the usual RPA result for the renormalized phonon frequency, with $\alpha_0\int_0^l \lambda_P dl'/2$ playing the role of the electronic polarization bubble and $1+ g_1\alpha_0 \int_0^l \lambda_P dl'/2$ describing the contribution from low-energy electrons to the dielectric function.
Thus, we learn that the widely used RPA expressions are valid only when the forward Coulomb scattering and the scattering amplitude in the Cooper channel are negligible.
Yet, as we see below, both $g_2$ and $\lambda_C$ play an important role in the Peierls instabilities of Weyl semimetals at high magnetic fields.
\subsubsection{Adiabatic regime with mirror symmetry}
\label{sec:cll_mirror}
Let us suppose that the two Weyl nodes of opposite chirality are related by a mirror plane at zero magnetic field.
This is a common circumstance in a variety of real Weyl semimetals.
In the presence of a magnetic field, mirror symmetry is preserved provided that the field is oriented perpendicular to the mirror plane. Then, one has $\Delta=0$, $\delta v=0$ and $\lambda_P(l,T) =-\lambda_C(l,T)$, the latter of which become cut off only by the temperature. It follows that interference between both (Peierls and Cooper) scattering channels is maximum, so that
\begin{align}
\label{g1g2Chiral}
\frac{d g_1}{dl}&=-{\alpha_0\over 2} g_1^2\lambda_P\cr
\frac{d g_2}{dl}&= {\alpha_0\over 2} g_1^2\lambda_C =\frac{d g_1}{dl}.
\end{align}
The solution for $g_1(l)$ at low temperature (where $\lambda_{P}\approx1$) is given by
\begin{equation}
\label{g1l}
g_1(l) = {g_1 \over 1+ {1\over 2} \alpha_0 g_1 l},
\end{equation}
indicating that repulsive backward scattering decays with $l$ and is marginally irrelevant. From Eqs. (\ref{g1g2Chiral}) or (\ref{g2mg1}), we see that for the mirror symmetric situation,
$g_1-g_2$ is an invariant of the RG flow. This is a consequence of the fact that, since fermions in chiral Landau levels are effectively spinless, the $g_1$ term can be transformed by exchange to a $g_2$ process, which conserves particles on each node and satisfies electron-hole symmetry at $\Delta=0$ \cite{Fowler76}.
The invariant combination $g_2(l)-g_1(l)=g_2-g_1$ remains marginal and nonuniversal. One can then write down the solution for the electron-phonon vertex part (Eq. (\ref{zpm})) as
\begin{equation}
\label{zeph}
z_{\tau,-\tau} = e^{{1\over 2}\gamma \int_0^l \lambda_P(l') dl'},
\end{equation}
where
\begin{equation}
\label{eq:gamma}
\gamma= {\alpha_0(g_2-g_1)}
\end{equation}
is a nonuniversal power law exponent dependent on Coulomb interactions and the magnetic field.
At ${\bf q}=2k_F \hat{\bf z}$, $\lambda_P=1$ up to the temperature cutoff value
\begin{equation}
l_T\equiv \ln (\Lambda_0/k_B T),
\end{equation}
and $\lambda_P\simeq 0$ for $l>l_T$.
Then, from Eq.~(\ref{Kohn_chiral}), the vanishing of the renormalized phonon frequency
in the static ($\omega_m=0$) limit leads to the Peierls critical temperature scale for the $j$ mode,
\begin{equation}
\label{Tc0}
T_{c,0}^j={\Lambda_0\over k_B}\left( \alpha_0 g_{x,j}^{\prime2}\over 2\gamma + \alpha_0 g_{x,j}^{\prime2}\right)^{1/\gamma}.
\end{equation}
In practice, the Peierls instability is set by the mode $j$ for which $T_{c,0}^j$ is maximum.
The power law dependence for the critical temperature originates from the competition between the Peierls and the Cooper scattering channels, and differs qualitatively from the exponential, BCS-like forms proposed in earlier mean-field treatments \cite{Yang2011, Zhang2016, Qin2020}.
The latter studies have implicitly ignored the possibility of a Cooper channel contribution. We emphasize that the Cooper channel is important even though the two Weyl nodes are not time-reversed partners of one another;
the reason is that the mirror plane relating the two Weyl nodes passes from the $\Gamma$ point, so that the two nodes have opposite momenta along the direction of the magnetic field.
It is as though the presence of the magnetic field rendered an effectively one-dimensional problem with the $k\to -k$~ symmetry that is required for a full Cooper channel contribution.
Let us comment on some qualitative aspects of Eq.~(\ref{Tc0}). When $g_2> g_1$, Peierls instability occurs at any nonzero value of $g_{x}^{\prime2}$ and charge-density wave correlations of the electron system enhance the Peierls temperature with respect to the BCS-like limit $k_BT^j_{c,0} = \Lambda_0 \exp[-2/ (\alpha_0 g_{x,j}^{\prime2})]$, which we recover when electron-electron interactions are vanishingly small ($\gamma\to 0^+$).
When $g_2< g_1$ (i.e. $\gamma <0$),
a Peierls instability can still occur but only if the electron-phonon coupling strength exceeds a threshold value ($\alpha_0 g_{x,j}^{\prime2} > -2\gamma$).
In this sense, the phonon softening is qualitatively weaker if $g_1>g_2$ than if $g_1<g_2$.
As mentioned above, $g_1>g_2$ takes place if $q_{TF}> {\rm max}(k_F, 1/l_B)$.
The fact that the screening due to the short-range ($2 k_F$) part of the Coulomb interaction tends to suppress the Peierls instability when $q_{TF} > k_F$ is consistent with the literature in quasi one-dimensional organic conductors at zero magnetic field \cite{barisic83}.
Next, we carry out a numerical estimate of Eq.~(\ref{Tc0}) (see Fig. \ref{fig:Tc} for the variation of $T^j_{c,0}$ as a function of the field value $B$).
Combining Eqs.~(\ref{eq:ep_approx}) and (\ref{eq:ep_conv}), we have
\begin{equation}
\alpha_0 g^{\prime2}_{x,j}\simeq \frac{e B d_j^2}{\pi^2 v \rho \hbar^2 \omega_{0,j}({\bf q})^2}.
\end{equation}
We consider optical phonons with $d_j= 5\, {\rm eV}/\AA$ and $\hbar\omega_{0,j}({\bf q}) = 10\, {\rm meV}$.
We also take $\rho=10^4\, {\rm kg/m}^3$, $v_\tau=10^5\, {\rm m/s}$, $b=0.01 \AA^{-1}$ and $n_e=5\times 10^{17}\, {\rm cm}^{-3}$.
We take two possible values for $\epsilon_\infty$, $30$ and $80$.
In Fig. \ref{fig:Tc}, the dependence of $T^j_{c,0}$ on $B$ is nonmonotonic.
For low-to-moderate magnetic fields, $g_1/g_2\ll 1$ and hence $\gamma\simeq 1$ (far from the BCS-like regime).
In this regime, $\gamma\gg \alpha_0 g^{\prime2}_{x,j}$ and $k_B T^j_{c,0}/\Lambda_0 \simeq \alpha_0g^{\prime2}_{x,j}/2$ grows linearly with $B$.
As the magnetic field is made stronger, $g_1/g_2$ increases and as a result $\gamma$ decreases.
At high enough magnetic field, this eventually leads to a suppression of $T_c$ due to $g_1$.
The magnitude of the maximum critical temperature
depends sensitively on the parameters of the model.
For instance, if $d_j=2 {\rm eV}/\AA$, everything else being equal, the maximum of $T^j_{c,0}$ is suppressed by almost an order of magnitude.
For a given material, higher carrier density implies a higher value of the maximal $T^j_{c,0}$.
In Fig. \ref{fig:Tc}, the maximum of $T^j_{c,0}$ attains $\sim 1\%$ of the cutoff energy scale $\Lambda_0$.
The value for $\Lambda_0$ is roughly set by the distance in energy between the Fermi level and the bottom of the $n=1$ nonchiral Landau level band. We estimate $\Lambda_0\sim \hbar v_F/l_B \gtrsim 20\, {\rm meV}$ for the quantum limit, which is inferior to the values used in the recent literature \cite{Zhang2016}.
Based on these numbers, we conclude that the emergence of a measurable Peierls instability in a mirror-symmetric Weyl semimetal at high magnetic field is far from obvious. We must emphasize, however, that while these numbers are reasonable across a wide family of materials, reliable numerical estimates for $T_{c}$ in individual Weyl semimetal materials are beyond the scope of our work. For one thing, the value of the electron-phonon coupling is not known reliably in these materials. Likewise, the presence of multiple pairs of nodes in several real Weyl semimetals introduces additional complications in making quantitative predictions about field-induced instabilities (see Sec.\ref{sec:real} for further comments in this regard).
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig3.pdf}
\caption{(a) Estimate of the Peierls transition temperature in the adiabatic phonon approximation as a function of the magnetic field, with mirror symmetry (solid lines) and without mirror symmetry (dashed lines) for $\Lambda_0=50\, {\rm meV}$ (considering a field-dependent cutoff does not introduce qualitative changes to our results). For the curves with the dashed lines, we take $\Delta=5\, {\rm meV}$ as the mirror symmetry-breaking energy scale. Only the magnetic field values for which the system is in the quantum limit are displayed.
Two different values of $\epsilon_\infty$ are considered: $30$ (blue and green curves), $80$ (orange and red curves).
The transition temperature $T_{c}$ is enhanced by mirror-symmetry breaking. Also, the transition temperature is higher for larger $\epsilon_\infty$.
(b) The electron-electron interaction parameter $\gamma$ (Eq.~(\ref{eq:gamma})) as a function of the magnetic field, for two different values of $\epsilon_\infty$.
(c) $k_F l_B$ as a function of the magnetic field. The parameter values are indicated in the main text.
}
\label{fig:Tc}
\end{center}
\end{figure}
\subsubsection{Adiabatic regime with broken mirror symmetry}
\label{sec:cll_no_mirror}
We now consider the more general case with broken mirror symmetry.
Nonzero $\delta v$ and $\Delta=\hbar v_F (k_{F+}+k_{F-})$
are inevitable if the two Weyl nodes of opposite chirality are unrelated by an improper symmetry.
In the instance in which the two nodes are related by a mirror plane at zero magnetic field, such mirror symmetry can be effectively broken by orienting the magnetic field away from the normal to the mirror plane.
In such case, $\Delta$ can be large (of the order of $\hbar v_F (k_++k_-))$ when the magnetic field is parallel to the mirror plane.
From Eq.~(\ref{g2mg1}), the solution for $g_2(l)-g_1(l)$ can be obtained,
\begin{equation}
\label{C1}
g_2(l)-g_1(l) = {g_2-g_1\over 1- {1\over 2}\alpha_0(g_2-g_1) \int_0^l (\lambda_P + \lambda_C)dl'}.
\end{equation}
A finite $\delta v$ alone does not affect the logarithmic singularity of the scattering channels, but leads to a density of states renormalization by the factor $(1+ \delta v/2v_F)^{-1}$, which we have included in the definition for $\alpha_0$ (Eq.~\ref{alpha0}).
In contrast, $\Delta$ acts as a pair breaking perturbation for the logarithmic singularity of the Cooper scattering channel, which is thereby suppressed at sufficiently low energy. Electron-hole symmetry at $2k_F$ (nesting), however, is preserved and the related logarithmic singularity in the Peierls channel therefore remains essentially unaffected at all $l< l_T$.
From Eq. (\ref{Cooper}), one can distinguish three regimes for the cutoff functions $\lambda_C$ and $\lambda_P$ as a function of $l$.
First, at $l <l_\Delta= \ln (\Lambda_0/\Delta)$, the influence of $\Delta$ is small and one still has $\lambda_C\approx -\lambda_P=-1$. Therefore, the flow equations (\ref{g1g2Chiral}-\ref{zeph}) remain valid together with the expression for the Peierls critical temperature (Eq. (\ref{Tc0})), with the proviso that $k_BT^j_{c,0}> \Delta$.
The second regime, where $l> l_\Delta$, becomes applicable if $\Delta$ exceeds the Peierls transition temperature for the mirror-symmetric case (Eq. (\ref{Tc0})).
In this regime, the Cooper loop is cut off by $\Delta$ and becomes vanishingly small ($\lambda_C\approx 0$). Accordingly, the expression for $T^j_{c,0}$ is modified. The flow of interactions (Eq.~(\ref{C1})) for $\lambda_P\approx 1$ up to $l_T$ will develop a simple pole singularity at the characteristic temperature
\begin{equation}
\label{ }
T^*_0 = {\Delta\over k_B } e^{-2/\alpha_0(g_2-g_1)},
\end{equation}
which corresponds to the scale of an intrinsic instability of the electron gas against a $(0,0,2k_F)$ charge-density-wave formation. This can be checked by looking at the vertex part $z_{\rm CDW}$ for the electronic CDW susceptibility. From (\ref{sh0}), (\ref{sh}) and (\ref{fcdw}), its flow equation reads
\begin{equation}
\label{zcdw}
\frac{d}{dl} \ln z_{\rm CDW}= {\alpha_0\over 2} { (g_2-g_1)}\lambda_P,
\end{equation}
which unsurprisingly coincides with (\ref{zpm}) for $z_{\tau,-\tau}$. Using (\ref{C1}), one has
\begin{equation}
\label{zdelta}
z_{\rm CDW}(l) = {z_{\rm CDW}(\Delta)\over [1-{1\over 2}\alpha_0(g_2-g_1) (l-l_\Delta)]},
\end{equation}
where $z_{\rm CDW}(\Delta)=z_{\tau,-\tau}(\Delta)$, as given by Eq.~(\ref{zeph}) when evaluated at $l=l_\Delta$. The expression for $z_{\rm CDW}(l_T)$ and thus the CDW susceptibility (\ref{Ki0}) develops a simple pole singularity at $T^*_0$.
However, in the presence of a nonzero electron-phonon coupling, this instability is preempted by the Peierls transition.
According to Eq.~(\ref{C1}), the phonon softening condition (Eq.~(\ref{Kohn_chiral})) then gives a Peierls transition temperature
\begin{equation}
\label{Tc0Delta}
T^j_{c,0}= {\Delta \over k_B} \exp\left( -{2- \alpha_0g_{x,j}^{\prime 2}\chi(\Delta)\over \gamma [1 -{1\over2}\alpha_0\chi(\Delta) g_{x,j}^{\prime 2}] +\alpha_0z_{\tau,-\tau}^2(\Delta)g_{x,j}^{\prime 2}}\right),
\end{equation}
where $\chi(\Delta)$ is the contribution to the charge density-wave electronic susceptibility (\ref{Ki0}) up to $l_\Delta$. These correlations enhance the transition temperature with respect to the case without electron-phonon interactions.
We remind the reader that Eq.~(\ref{Tc0Delta}) holds only if $k_B T^j_{c,0}\leq \Delta$; in the opposite regime, one can ignore mirror symmetry-breaking and use Eq.~(\ref{Tc0}).
We also note that $T_{c,0}^j$ exceeds the temperature scale $T_0^*$ of the purely electronic instability, with $T_{c,0}^j\to T_0^*$ when $g_{x,j}\to 0$.
Finally, the third regime takes place when $\Delta>\Lambda_0$, i.e. when mirror symmetry is strongly broken. In this case, $l_\Delta\to 0$ and the Peierls transition temperature is still given by Eq.~(\ref{Tc0Delta}), but with $\chi(\Delta)\to 0$ in the exponent and $\Delta\to\Lambda_0$ in front of the exponential, i.e.
\begin{equation}
\label{Tc0Delta2}
T^j_{c,0}= {\Lambda_0\over k_B} \exp\left( -\frac{2}{\gamma +\alpha_0 g_{x,j}^{\prime 2}}\right).
\end{equation}
Figure \ref{fig:Tc}(a) shows the evolution of the transition temperature as a function of the applied magnetic field for $\Lambda_0=50~{\rm meV}$ (considering a field-dependent cutoff does not introduce qualitative changes to our results) and ${\Delta=5~{\rm meV}}$ (assumed for simplicity to be field-independent), for the same parameters as the in the mirror-symmetric case.
Interestingly, the transition temperature is significantly enhanced when mirror symmetry is broken.
As shown by Eq. (\ref{zdelta}), the presence at finite $l$ of a simple pole singularity magnifies the charge density-wave response of the electron system with respect to the power law (\ref{zeph}) for $\Delta=0$. This is responsible for such an enhancement in $T^j_{c,0}$.
In sum, there is a qualitative difference for the Peierls instabilities of the chiral Landau levels depending on whether or not they are related by an improper symmetry.
In the presence of mirror symmetry, the instability vanishes in the absence of electron-phonon interaction. This is because of the destructive interference between the Cooper and the Peierls channels.
When the mirror symmetry is broken, the Cooper channel is suppressed and this opens the possibility for having a purely electronic instability even when the electron-phonon coupling is negligible.
However, the presence of electron-phonon coupling makes the instability more pronounced, and thus a lattice instability preempts the purely electronic instability.
\subsection{Nonadiabatic Peierls instability}
\label{sec:nonad}
Thus far, our results were obtained in the so-called adiabatic regime, where $\hbar\omega_{0,j}(2 k_F \hat{\bf z}) \ll k_BT^j_{c,0}$ and the electronic degrees of freedom are `faster' than the phonons.
In this case, the effective electron-electron interaction induced by the exchange of phonons is strongly retarded and its contribution to higher order vertex corrections in the phonon self-energy (Eq. (\ref{Kohn_chiral})), which would result from the contraction of the generated anharmonic terms in the phonon action, is negligible. This leaves us with the diagram in Fig. \ref{fig_RG1}(a) of the adiabatic approximation, in which vertex corrections are purely electronic in character.
However, it is apparent from the numerical estimates of Fig.\ref{fig:Tc} that the scenario $\hbar\omega_{0,j}(2 k_F \hat{\bf z}) \gg k_BT^j_{c,0}$
can be commonly realized in practice in Weyl semimetals.
This defines the nonadiabatic regime where phonon-mediated interactions between electrons tend to be nonretarded. Such interactions make a significant contribution to the electronic vertex in Fig. \ref{fig_RG1}(b-c) and
contribute to the phonon self-energy depicted in Fig. \ref{fig_RG1}(a). It follows that the criterion of instability against the formation of a superstructure, as derived in Eq. (\ref{Kohn_chiral}) for phonons, no longer holds and must be modified. Therefore, at energy scales below $\hbar\omega_{0,j}(2 k_F \hat{\bf z})$, we focus on the electronic degrees of freedom and derive
the condition for the instability of the electron system against the formation of a $2k_F \hat{\bf z}$ charge-density wave, using Eqs. (\ref{g1g2ChiralA}) and (\ref{g1g2ChiralB}).
We shall tackle this problem using a two-cutoff RG approach \cite{Caron84}. First, we integrate the flow equations of the previous section in the adiabatic sector, where $\Lambda_0(l) \ge \hbar\bar{\omega}_{0}$, here defined for simplicity as the minimum energy of the set of phonon modes $\{\hbar\omega_{0,j}\}$.
The scaled cutoff energy reaches the characteristic phonon energy at $l^*$, i.e. $\Lambda_0(l^*)\equiv \hbar\bar{\omega}_{0}$.
For $l>l^*$, all phonon modes enter the nonadiabatic regime and we carry out the partial trace integration $\int D\phi\ e^{(S^{(0)}_{p,l^*}+ S^{}_{ep,l^*})/\hbar}$ over phonon degrees of freedom in the partition function (Eq. (\ref{RGTransf})). By dropping anharmonic terms generated by the RG transformation, the integration is gaussian and yields attractive, phonon-mediated interactions between electrons. The backscattering amplitude mediated by phonons takes the form
\begin{equation}
\label{g1ph}
\mathfrak{g}_1 = - {1\over 2}z_{\tau,-\tau}^2(l^*)\sum_j\mathscr{D}_{j,l^*}(0,2k_F \hat{\bf z})g_{x,j}^{2}.
\end{equation}
This coupling can be considered as essentially nonretarded, i.e. independent of frequency.
As such, its amplitude will simply add to the purely repulsive $g_1(l^*)$ part obtained from Eq. (\ref{g1l}), so that the initial condition at $l^*$ for (\ref{g1g2Chiral}) becomes $g_1^*= g_1(l^*) + \mathfrak{g}_1$. With sizable electronic vertex corrections $z_{\tau,-\tau}(l^*)$ and Peierls fluctuations, which imply a large $\mathscr{D}_{j,l^*}(0,2k_F \hat{\bf z})$, the likelihood for a total attractive backscattering amplitude $g_1^*$ is high.
The integration of phonon modes at small ${\bf q}$ will also give rise to an effective attractive forward scattering amplitude
\begin{equation}
\label{g2ph}
\mathfrak{g}_2 = - {1\over 2}\sum_j\mathscr{D}_{j,0}(0,q\hat{\bf z}\sim0)g_{\tau,\tau,j}g_{-\tau,-\tau,j},
\end{equation}
which will add to $g_2(l^*)$ to give $g_2^*= g_2(l^*) + \mathfrak{g}_2$ as the boundary condition for the forward scattering at $l^*$. However, in the absence of electronic enhancement ($z_{\tau,\tau}=1$) and a lack of ${\bf q} \sim 0$ phonon softening, this attractive contribution is expected to be small, so that $g_2^*$ remains repulsive.
Given these corrections to the electronic interactions, we will investigate the possibility for an electronic instability in the Weyl semimetal model, in the nonadiabatic regime.
\subsubsection{Nonadiabatic regime with mirror symmetry}
\label{sec:nonad_mir}
We first consider the possibility for instabilities when the two Weyl nodes are related by perfect mirror symmetry ($\Delta=0,\delta v=0)$. In these conditions, $g_1$ and $g_2$ at $l> l^*$ are still governed by the flow of Eq.~({\ref{g2mg1}), for which the interference between the Cooper and the Peierls scattering channels is maximum ($\lambda_C=-\lambda_P=-1$). The invariant $g_2(l) -g_1(l)=g_2^*-g_1^*$ is now fixed by the boundary conditions at $l^*$. Therefore, the vertex part (\ref{zcdw}) of the $2k_F$ charge-density-wave susceptibility at $l>l^*$ takes the scaling form
\begin{equation}
\label{zna}
z_{\rm CDW}(l) = z_{\rm CDW}(l^*) e^{\gamma^*(l-l^*)/2},
\end{equation}
where $\gamma^*= \alpha_0 (g_2^*-g_1^*)$ and $z_{\rm CDW}(l^*)$ corresponds to the same expression given in Eq. (\ref{zeph}).
The $2k_F\hat{\bf z}$ charge-density-wave susceptibility, Eq. (\ref{Ki0}), when evaluated at the temperature cutoff value $l_T$, has the power law form
\begin{equation}
\label{ }
\chi(T) = \chi(l^*) + {{z^{2}_{\rm CDW}(l^*)}\over \gamma^{*}}\left( \left({ \hbar\bar{\omega}_{0}\over k_BT}\right)^{\gamma^*} -1\right),
\end{equation}
where $\chi(l^*)$ is the contribution to the susceptibility (\ref{Ki0}) up to $l^*$. In the nonadiabatic regime, the chiral mirror symmetric Landau level thus evolves towards a Luttinger liquid behavior with no instability at finite temperature.
As a function of phonon frequency, it follows that the system will crossover from a classical $2k_F\hat{\bf z}$ Peierls superstructure at small $ \hbar\bar{\omega}_{0}\ll k_B T_{c,0}^j$ to a Luttinger liquid at high phonon frequency $ \hbar\bar{\omega}_{0}\gg k_B T_{c,0}^j$. Such a quantum-classical crossover is reminiscent to the one encountered in the model of spinless electrons coupled to optical phonons in one dimension. The zero temperature Peierls state is known to be suppressed by quantum lattice fluctuations at large phonon frequency \cite{Hirsch83,Caron84}. Regarding the strength of correlations, attractive corrections coming from phonons yield $\gamma^*>\gamma $ at any field values, so that charge density wave correlations are enhanced by phonons in the nonadiabatic domain.
One may also wonder whether other types of correlations, such as the superconducting ones, may become singular in the nonadiabatic domain. It turns out that the phonon renormalization of $g_1^*$ to the attractive sector will make this possibility unlikely. This is especially the case for the equal-pseudospin Cooper pairing susceptibility. From Eqs.~(\ref{sh0}), (\ref{Osc}) and (\ref{sh}), its vertex part obeys the flow equation of Fig.~\ref{fig_RG1} (e), namely
\begin{equation}
\label{ }
{d\ln z_{\rm SC}\over dl} = -{\alpha_0\over 2} { (g_1-g_2)}\lambda_C.
\end{equation}
Its solution takes the scaling form $z_{\rm SC} \sim e^{\gamma_{\rm SC} l^*/2}e^{\gamma_{\rm SC}^*(l-l^*)/2}$, but with a negative power law exponent $\gamma_{\rm SC}^*=-\gamma^*$ that strongly suppresses superconducting correlations in the nonadiabatic domain.
\subsubsection{Nonadiabatic regime with broken mirror symmetry}
\label{sec:nonad_nomir}
In the absence of mirror symmetry ($\Delta\neq 0$ and $\delta v\neq 0$), one can distinguish two regimes of nonadiabaticity. The first one is characterized by $l^*<l_\Delta$, i.e. $ \hbar \bar{\omega}_{0} > \Delta$.
In this regime, the nonadiabatic conditions described in Sec. \ref{sec:nonad_mir} come first, where the Peierls scale $T_{c,0}^j$ becomes irrelevant and Luttinger liquid conditions for electrons prevail in the interval $l_\Delta> l>l^*$. Within our sharp cutoff boundary approach, the logarithmic singularity of the Cooper scattering channel is suppressed at $l > l_\Delta$, where $\lambda_C\approx0$, and the couplings in Eq. (\ref{C1}) develop a simple pole singularity:
\begin{equation}
\label{ }
g_2(l)-g_1(l) = {g^*_2-g^*_1\over 1- {1\over 2}\alpha_0(g^*_2-g^*_1) (l-l_\Delta)},
\end{equation}
where $g_{1,2}^*=g_{1,2}(l^*) + \mathfrak{g}_{1,2}$ are evaluated at $l^*$.
The singularity influences in turn the vertex part of the charge density-wave susceptibility,
\begin{equation}
z_{\tau,-\tau}(l) = {z_{\tau,-\tau}(l_\Delta)\over 1- {1\over 2}\alpha_0(g^*_2-g^*_1) (l-l_\Delta)},
\end{equation}
where $z_{\tau,-\tau}(l_\Delta)$ is given by Eq. (\ref{zna}). The Luttinger liquid is therefore unstable against the formation of a charge-density-wave state. Using $l_T=\ln (\Lambda_0/k_BT)$, the temperature scale of the instability is given by
\begin{equation}
\label{Tc0nonadia}
T^*_0 = {\Delta\over k_B } e^{-2/\alpha_0(g^*_2-g^*_1)}.
\end{equation}
Next, we examine the second regime, where $l^*>l_\Delta$, i.e. $ \hbar \bar{\omega}_{0} < \Delta$. In this regime, symmetry breaking effects described in Sec.~\ref{sec:cll_no_mirror}~first set in, before the onset of nonadiabatic regime at $l^*$. In this case the combination of couplings at $l>l^*$ shows a simple pole structure
\begin{equation}
\label{ }
g_2(l)-g_1(l) = {g^*_2-g^*_1\over 1- {1\over 2}\alpha_0(g^*_2-g^*_1) (l-l^*)}.
\end{equation}
Here $g^*_2-g^*_1$ is given by the solution of Eq. (\ref{C1}), first up to $l_\Delta$ and then to $l^*$ where the phonon induced contributions $\mathfrak{g}_{1,2}$ are added to $g_{1,2}(l^*)$. According to Eq. (\ref{zpm}) the vertex part of the charge-density-wave susceptibility will have the form
\begin{equation}
\label{ }
z_{\tau,-\tau}(l) = {z_{\tau,-\tau}(l^*) \over 1- {1\over 2}\alpha_0(g^*_2-g^*_1) (l-l^*)},
\end{equation}
where $z_{\tau,-\tau}(l^*)$ is given by (\ref{zdelta}). The temperature scale of the instability against charge-density-wave superstructure becomes
\begin{equation}
\label{Tc0nonadiadelta}
T^*_0 = { \hbar \bar{\omega}_{0} \over k_B } e^{-2/\alpha_0(g^*_2-g^*_1)}.
\end{equation}
One can conclude that mirror symmetry-breaking with finite $\Delta$ invariably makes the Luttinger liquid of the nonadiabatic regime unstable against the formation of a charge-density-wave state.
The nonadiabatic regime has also been addressed in Ref. [\onlinecite{Shindou2019}], albeit for a different model of semimetal with spin-polarized electron and hole bands in the quantum limit. In that paper, an attractive electron-electron interaction mediated by a screened electron-phonon coupling have been considered, and phases such as an excitonic insulator, charge Wigner crystal and charge-density wave have been predicted using a parquet renormalization group analysis.
\section{Peierls instability from nonchiral Landau levels}
\label{sec:noncll}
In this section, we consider the case in which the Fermi level crosses not only the chiral Landau levels, but also the $|n|=1$ nonchiral Landau levels.
For now, we will be interested in the Peierls instability that results from electronic transitions taking place inside the $|n|=1$ Landau level.
In the end of the section, we will compare this instability to the one taking place in the $n=0$ Landau level.
We will neglect the possibility of instabilities arising from transitions connecting chiral and nonchiral Landau levels.
For such transitions, the form factors in Eq.~(\ref{eq:form_factor}) vanish when ${\bf q}$ is parallel to the magnetic field.
\subsection{Electron-phonon and electron-electron couplings}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\columnwidth]{fig4.pdf}
\caption{Landau level spectrum of two Weyl nodes of opposite chirality, near the quantum limit, where $\rm{k}_{||}$ refers to the component of the momentum along the direction parallel to the magnetic field. The Fermi energy $\epsilon_F$ intersects the first nonchiral Landau level on each node ($n=1$ for the positive chirality node, and $n=-1$ for the negative chirality node).
The wave vectors ${\bf q}_1$, ${\bf q}_2$, ${\bf q}_3={\bf q}_1+{\bf q}_2$ and ${\bf q}_4=2 {\bf q}_1 + {\bf q}_2$ connect different Fermi points within the $|n|=1$ set of Landau levels.
The energy $\Lambda_1$ corresponds to the distance from the bottom of the $|n|=1$ Landau levels to $\epsilon_F$; it plays the role of an ultraviolet cutoff.
In this figure, it is assumed that the two nodes of opposite chirality are related by a mirror plane.
Otherwise, the vector ${\bf q}_1$ would have a different length for the two nodes.}
\label{fig:nonc}
\end{center}
\end{figure}
We begin by assuming that the two Weyl nodes of opposite chirality are related by a mirror plane.
Then, we linearize the dispersion for the $|n|=1$ Landau levels in the vicinity of the Fermi points.
Within the subspace of $|n|=1$ Landau levels, several different phonon wave vectors $\pm {\bf q}_i$ ($i=1,2,3,4$) connect different Fermi points (see Fig. \ref{fig:nonc}).
The wave vectors at which instabilities are more likely are $\pm {\bf q}_1$, $\pm {\bf q}_2$ and $\pm {\bf q}_4=\pm (2 {\bf q}_1+{\bf q}_2)$, because they connect electronic branches that are perfectly nested.
On the one hand, the theory of phonon instabilities at $\pm {\bf q}_2$ and $\pm {\bf q}_4$, which connect nodes of opposite chirality, is very similar to the one discussed in the preceding subsection for chiral Landau levels.
The only significant difference is that the ultraviolet cutoff appearing in the expression for the Peierls transition temperature is now given by $\Lambda_1$, instead of $\Lambda_0$, and that the matrix elements of the different couplings are multiplied by a factor $\lambda$ defined below in Eq.~(\ref{eq:lambda}).
On the other hand, the instability at wave vector $\pm {\bf q}_1$, which connects two $|n|=1$ branches at the same Weyl node, presents some conceptual novelties.
Thus, hereafter we concentrate on this wave vector, whose magnitude we denote as $2 k_F$.
The maximum value of $k_F$ we consider is $\sqrt{2}/l_B$ (beyond this value, the Fermi level intersects the $|n|=2$ Landau levels).
Let us label $L$ (left) and $R$ (right) the two Fermi points separated by $q_1$ in a node of a given chirality $\tau$.
The electron-phonon matrix element connecting $L$ to $R$ reads
\begin{align}
\label{eq:gLR}
&g_{LR,\tau}^{ph}(\mathbf{q}) =g^{ep}_{\tau}\langle\Psi_{n,X,L,\tau}|e^{i {\bf q}\cdot{\bf r}} |\Psi_{n,X^{\prime},R,\tau}\rangle\nonumber\\
& \simeq \lambda\, g^{ep}_{\tau} \delta_{q, -2 k_F}\delta_{X^{\prime},X+q_{y}l_{B}^{2}}
e^{iq_{x} (X+X^\prime)/2} e^{-q_{\bot}^{2}l_{B}^{2}/4},
\end{align}
where we have used $n=1$ ($n=-1$) for $\tau=1$ ($\tau=-1$), and we have defined
\begin{equation}
\label{eq:lambda}
\lambda=\frac{1}{\sqrt{(k_{F}l_{B})^{2}/2+1}}.
\end{equation}
One can similarly obtain an expression for $g_{RL,\tau}^{ph}(\mathbf{q})$; the only difference with respect to Eq.~(\ref{eq:gLR}) is that $\delta_{q, -2 k_F}$ is replaced by $\delta_{q, 2 k_F}$.
In the derivation of Eq.~(\ref{eq:gLR}), we have approximated $\exp(-q_\perp^2 l_B^2/4) \left[1 + O(q_\perp^2 l_B^2)\right]\simeq \exp(-q_\perp^2 l_B^2/4) $.
This approximation is motivated by the fact that, for $q_{\bot}l_{B}\geq1$, the electron-phonon matrix element is exponentially suppressed.
Comparing Eq.~(\ref{eq:gLR}) with Eq.~(\ref{eq:cll_me}), we recognize that the electron-phonon matrix elements for chiral and nonchiral Landau levels differ by a factor $\lambda$.
This difference emerges from the pseudospin textures of the respective Landau levels.
For chiral Landau levels, the expectation value of the pseudospin is either perfectly aligned or antialigned with the $z$ axis (depending on the sign on $B$), irrespective of the wave vector and the chirality (cf. Eq.~(\ref{eq:spinor_cll})).
Hence, the overlap factor for the spinors of the two chiral Landau levels of opposite chirality is unity.
In contrast, for the $|n|=1$ level, the expectation value of the pseudospin rotates in momentum space: it points along $x$ at the bottom of the $|n|=1$ band, and gradually tilts towards $z$ as $k$ departs from the energy minimum (cf. Eq.~(\ref{eq:spinor_ncll})).
Moreover, the left- and right-moving branches in a nonchiral Landau level have opposite orientations of the pseudospin along $z$.
This leads to an overlap factor $\lambda<1$ between the branches connected by the wave vector ${\bf q}_1$.
The overlap factor is gradually suppressed for larger $q_1 l_B$.
In practice, $q_1=2 k_F$ is a very small fraction of the Brillouin zone. Therefore, the ${\bf q}$-dependence of $g^{ep}_\tau$ in Eq. (\ref{eq:gLR}) can be neglected and its transformation properties under symmetry operations can be determined from the irreducible representations at zero wave vector.
In the presence of mirror symmetry, one may have $q=0$ phonons whose deformation potentials are even or odd under the mirror operation.
We denote these phonons as scalar and pseudoscalar, respectively.
In reality, scalar and pseudoscalar phonons are invariant under all proper symmetry operations; however, for our model of two Weyl nodes, only the mirror operation is relevant.
Consequently, we have
\begin{align}
\label{geph}
g^{ep}_{1} &=g^{ep}_{-1}\equiv g_{s} \text{ (scalar phonons)}\nonumber\\
g^{ep}_{1} &=-g^{ep}_{-1}\equiv g_{ps} \text{ (pseudoscalar phonons)}.
\end{align}
The Coulomb matrix elements can be calculated in the same way as for the chiral Landau levels.
The long-wavelength (${\bf q}\simeq 0$) part of the Coulomb interaction (counterpart of Eq.~(\ref{g20})) is now given by
\begin{equation}
\label{eq:g2_nc}
g_{2}=\frac{2\pi l_{B}^{2}}{1+2\sqrt{1+\frac{2}{k_F^2 l_B^2}}},
\end{equation}
which incorporates the static screening from electrons in the $|n|=0$ and $|n|=1$ levels, at zero temperature (we assume $\Lambda_1\gg k_B T$).
Likewise, the short-wavelength ($q\simeq q_1 =2 k_F$) part corresponding to interbranch momentum transfer
within a node reads
\begin{equation}
g_{1} =\frac{e^{2}\lambda^{2} l_{B}^{2}}{\pi \hbar v_{F}\epsilon_{\infty}}\ln\left(1+\frac{1}{4k_{F}^{2}l_{B}^{2}}\right),
\end{equation}
where we have averaged over $q_\perp$ like in Eq. (\ref{g10}), assuming that $k_F l_B$ is not small.
Note that this expression is identical to that of the previous section, except for the $\lambda^2$ factor coming from the momentum-space pseudospin texture of the nonchiral Landau levels.
Then,
\begin{align}
\frac{g_1}{g_2} &= \frac{e^2}{\pi^2 \hbar v_F \epsilon_\infty} \lambda' \ln\left(1+\frac{1}{4 k_F^2 l_B^2}\right)\nonumber\\
\lambda' &\equiv \frac{1}{2 + k_F^2 l_B^2} \left(1+ 2 \sqrt{1+ \frac{2}{k_F^2 l_B^2}}\right).
\end{align}
Once again, compared to the discussion of Sec. \ref{sec:cll} on chiral Landau levels, a dimensionless factor $\lambda'$ appears in the ratio $g_1/g_2$ for nonchiral Landau levels.
\subsection{RG equations and phonon spectrum}
Like in Sec.~\ref{sec:cll}, the outer shell corrections entering the RG recursive relations (Eqs.~(\ref{Sp0}), (\ref{zeph0}) and (\ref{g1g2})) are all evaluated in the limit of small $q_\perp $ up to the cutoff $1/l_B$, above which all the couplings are
exponentially suppressed. Within the magnetic length scale $l_B$, the couplings can then be considered as essentially local in the transverse directions, which yield one-dimensional-like loop corrections for the flow equations.
From the results of Appendix B, the RG transformations for the electron-phonon vertices $z_j$ of Fig.~\ref{fig_RG1} (b), corresponding to scalar ($j=s$) and pseudoscalar ($j={ps}$) modes, are given by
\begin{align}
\label{eq:rg_ep_nc}
\frac{d}{dl} \ln z_j &=\frac{\gamma_j(l)}{2}\lambda_P,
\end{align}
where
\begin{align}
\gamma_s(l) &= \alpha_1(g_{2}(l)-2g_{1}(l))\nonumber\\
\gamma_{ps}(l) &=\alpha_1 g_{2}(l)
\end{align}
and $\alpha_{1}=1/(2\pi l_{B}^{2})$. Here the cutoff function $\lambda_P$ of the Peierls loop for the non chiral Landau is given by Eq.~(\ref{eq:ipnc}) of Appendix B.
Remarkably, the RG flow and the corresponding phonon softening for the {\em pseudoscalar} phonons is independent of the coupling $g_{1}$. As explained in Appendix B, this is only true for the case where the two Weyl nodes are related by mirror symmetry.
The physical meaning behind the difference in scalar and pseudoscalar phonon modes originates from the fact that the latter do not create any net charge fluctuations and hence they are not screened by $g_1$.
In other words, $g_1$ couples to total charge fluctuations, which are absent in the case of pseudoscalar phonons.
In contrast, $g_2$ (which is associated to vertex corrections) couples to all phonons.
As for the electron-electron interactions, the results of Appendix B lead to the following RG equations:
\begin{align}
\label{eq:rg_ee_nc}
\frac{dg_{1}}{dl}&=(-g_{1}^{2}+g_{1}g_{2})\alpha_{1}\lambda_P\nonumber\\
\frac{dg_{2}}{dl}&=\frac{g_{2}^{2}}{2}\alpha_{1}\lambda_P.
\end{align}
Unlike in Sec. \ref{sec:cll}, the one-loop corrections to both the couplings $g_{1}$ and $g_{2}$ arise from the Peierls channel alone. The contribution from the Cooper channel, neglected here, is no longer logarithmically divergent because the two electronic branches connected by $q_1$ are not symmetric with respect to center of the Brillouin zone.
The last diagram of Fig.~\ref{fig_RG1} (c) is therefore (approximately) absent.
The solutions of Eq.~(\ref{eq:rg_ee_nc}) can be written as
\begin{align}
\label{eq:ee_nc_sol}
\gamma_{j}(l) &=\frac{\gamma_{j}}{1-\frac{\gamma_{j}}{2}\int_0^l\lambda_P dl'},
\end{align}
where $\gamma_j$ is the bare ($l=0$) value of $\gamma_j(l)$.
In the absence of a logarithmically singular Cooper channel, there is no invariant combination of couplings in the course of the RG flow. While $\gamma_{ps}(l)$ is marginally relevant, $\gamma_{s}(l)$ may be marginally irrelevant or relevant depending upon the relative values of $g_{2}(l)$ and $2g_{1}(l)$.
Using Eq.~(\ref{eq:ee_nc_sol}), Eq.~(\ref{eq:rg_ep_nc}) can be solved to yield
\begin{align}
\label{eq:ep_nc_sol}
z_{j} &=\frac{1}{1-\frac{\gamma_{j}}{2}\int_0^l\lambda_P dl'}.
\end{align}
Thus, the simple poles in the electron-electron couplings translate into corresponding poles in the electron-phonon vertices $z_{s}$ and $z_{ps}$, which essentially demonstrates the coupling of different charge-density wave correlations to the scalar and pseudoscalar phonon modes.
The charge density wave emerging at the instability of pseudoscalar phonons is special in that the total charge density is uniform, while the {\em difference} in the charge density between the two Weyl nodes of opposite chirality oscillates in space with a period of $\pi/k_F$.
The self-energy corrections for the scalar and pseudoscalar phonon propagators shown in Fig.~\ref{fig_RG1} (a) are given by
\begin{align}
\label{eq:phprop_nc}
\mathscr{D}_l^{-1} = &\, \mathscr{D}_0^{-1} - \scalebox{1.55}{$\pi$}_l\cr
= & \, \omega_m^2 + \omega_{0,j}^2(\mathbf{q})\left[1- {\alpha_1} g_{j}^{\prime2}\chi_j(l)\right],
\end{align}
where $\omega_{0,j}$ is the bare frequency for the scalar {($ j =s$)} and pseudoscalar {($ j =ps$)} phonons, $g_{j}^{\prime2}= g_{j}^{2}\lambda^2/\omega_{0,j}^2$, and
\begin{equation}
\label{Ki}
\chi_{j}(l) = \int_0^l z_{j}^2(l') \lambda_P dl'
\end{equation}
refers to the corresponding electronic susceptibility involved in the phonon softening.
Since we are interested in ${\bf q} = 2 k_F \hat{\bf z}$, we take $\lambda_P \approx 1$ for $l< l_T= \ln(\Lambda_1/k_B T)$ and $\lambda_P\simeq 0$ for $l>l_T$ in Eq.~(\ref{Ki}).
Equation (\ref{eq:phprop_nc}) allows us to study the adiabatic phonon softening caused by electrons.
When $\alpha_1 g_{j}^{\prime2}\chi_{j}(l_T)=1$ at $\omega_m=0$, the system undergoes a Peierls instability.
The transition temperature for the instability at $2 k_F\hat{\bf z}$ is given by
\begin{align}
\label{eq:Tc_nc}
T_{c,1}^j &=\Lambda_{1}\exp\left(-\frac{2}{2\alpha_{1} g_{j}^{\prime2}+\gamma_{j}}\right) \text{ ($j=s,ps$)},
\end{align}
where the subscript $1$ is to remind that the instability considered here emerges from within the $|n|=1$ nonchiral Landau levels and their coupling to phonons.
The Peierls transition temperature is of the BCS (exponential) form, in agreement to the phenomenological expression used in Ref. [\onlinecite{Liu2016}], although renormalized by electron-electron interactions.
This is unlike in the chiral Landau levels related by a mirror plane (or by another improper symmetry), where the Cooper scattering channel is present, and consequently, a power law behavior of the transition temperature is realized.
We note that Eq.~(\ref{eq:Tc_nc}) is valid only when $2\alpha_{1} g_{j}^{\prime2}+\gamma_{j}>0$.
If $2\alpha_{1} g_{j}^{\prime2}+\gamma_{j}<0$, $T_{c,1}^j=0$ and there is no Peierls instability.
For scalar phonons, $\gamma_s$ becomes negative at high magnetic fields, which leads to $2\alpha_{1} g_{s}^{\prime2}+\gamma_{s}<0$.
In contrast, for pseudoscalar phonons, $\gamma_{ps}$ is positive regardless of the magnetic field (because $g_2>0$) and therefore $2\alpha_{1} g_{ps}^{\prime2}+\gamma_{ps}$ is always positive.
Thus, pseudoscalar phonons remain susceptible to a Peierls instability even when the scalar phonons are not.
The qualitative difference in the softening between scalar and pseudoscalar phonons was first noticed in Ref.~[\onlinecite{rinkel2019}], in the context of the random phase approximation.
This approximation amounts to neglecting the bare value of $g_2$ in Eqs. (\ref{eq:rg_ep_nc}) and (\ref{eq:rg_ee_nc}).
Under such condition, our theory would predict
\begin{align}
\label{eq:chi_RPA_nc}
\chi_s(l) &= \frac{\int_0^l \lambda_P dl'}{1+g_1 \alpha_1 \int_0^l \lambda_P dl'}\nonumber\\
\chi_{ps}(l) &= \int_0^l \lambda_P dl'.
\end{align}
Hence, the Peierls instability (the divergence of $\chi(l)$ at $l\to\infty$, which takes place when $\lambda_P\simeq 1$) would be suppressed for scalar phonons due to RPA screening, but not for pseudoscalar phonons.
According to Eqs.~(\ref{eq:chi_RPA_nc}) and (\ref{eq:phprop_nc}), pseudoscalar phonons would undergo a Peierls instability for an infinitesimally weak electron-phonon coupling, while it would take a minimum strength of electron-phonon coupling to have a Peierls instability of scalar phonons.
This statement is in agreement with the results of Ref.~[\onlinecite{rinkel2019}].
However, our more complete theory, which goes beyond RPA, adds nuance to such results.
For instance, the $g_2$ contribution ignored in the RPA helps the Peierls instability of scalar phonons, which can now take place for infinitesimally weak electron-phonon interactions, provided that $g_2> 2 g_1$ (as shown in Eq.~(\ref{eq:Tc_nc})).
In Eq.~(\ref{eq:Tc_nc}), the cutoff energy scale $\Lambda_{1}$ decreases as the strength of the applied magnetic field increases (unlike $\Lambda_0$ in the preceding section), ultimately vanishing at the threshold of the quantum limit.
The exact functional dependence can be calculated analytically.
When $\Lambda_{1}$ is vanishingly small, there is no Peierls transition in the nonchiral Landau level.
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig5.pdf}
\caption{(a) Peierls transition temperature for scalar and pseudoscalar phonons, at the nesting wave vector ${\bf q}_1$ (see Fig.~\ref{fig:nonc}).
The parameter values are the same as in Fig.~\ref{fig:Tc}.
Only the magnetic field range in which the Fermi level intersects the $n=0$ and $|n|=1$ Landau levels is shown. Also, we limit the data to $q_1 l_B > 1$.
For a given strength of electron-phonon interactions, the transition temperature for pseudoscalar phonons is enhanced with respect to that of scalar phonons; the two converge to one another when $\epsilon_\infty$ is large.
(b) The bare value of $\gamma_j$ as a function of the magnetic field. For the scalar phonon, $\gamma_s=g_2 - 2g_1$ changes sign at sufficiently high magnetic fields, eventually eliminating the Peierls instability. For the pseudoscalar phonon, $\gamma_{ps}= g_2$ remains always positive. We note that the numerical values for $\gamma_j$ are much smaller than in the quantum limit (Fig.~\ref{fig:Tc}) because $g_2$ is relatively reduced by the additional screening coming from the $|n|=1$ Landau levels. As a a result, the critical temperatures in this figure are much lower than the ones expected in Fig.~\ref{fig:Tc}.}
\label{fig:Tc_nonchiral}
\end{center}
\end{figure}
Figure \ref{fig:Tc_nonchiral} provides a numerical estimate of $T_{c, 1}^s$ and $T_{c,1}^{ps}$.
For equal magnitude of electron-phonon coupling, the transition temperature for pseudoscalar phonons is higher than that of scalar phonons, because $\gamma_{ps}>\gamma_s$.
This difference can be very significant when $g_1/g_2$ is not small, i.e. at stronger magnetic fields, for reasons mentioned above.
Therefore, we conclude that pseudoscalar phonons are more prone than scalar phonons to Peierls instabilities.
For the same parameter values as in Sec.~\ref{sec:cll}, the transition temperature for the Peierls instability at wave vector ${\bf q}_1$ is very low (sub Kelvin).
This is so for three reasons: (i) when the Fermi level intersects only $n=0$ and $|n|=1$ Landau levels, the cutoff $\Lambda_1$ can be no larger than $(2-\sqrt{2}) \hbar v/l_B$; (ii) $g_2$ is smaller than in the quantum limit, due to the additional screening coming from the partially occupied $|n|=1$ Landau levels (compare Eqs.~(\ref{eq:g2_nc}) and (\ref{g20}));
(iii) as the magnetic field is made stronger, $\Lambda_1$ and $g_2$ both decrease, while $g_1/g_2$ increases.
In view of the preceding paragraph, a Peierls transition that originates exclusively from $|n|=1$ Landau levels at wave vector $q_1$ appears difficult to observe.
For a given Fermi energy, the instability will be often dominated by the nesting wave vector connecting nodes of opposite chirality, mainly because
$g_1$ is smaller for the larger (internode) nesting wave vectors.
An exception may take place if $g_1>g_2$ for both $n=0$ and $|n|=1$ bands. In that case, for a weak enough electron-phonon interaction strength, there would be no instability originating from the chiral Landau levels (recall Eq.~(\ref{Tc0})). There could be, however, an instability originating from the $|n|=1$ levels for (and only for) pseudoscalar phonons, although at very low temperature.
We conclude this section by discussing the cases of broken mirror symmetry and nonadiabaticity.
Let us suppose that the mirror symmetry is broken (which can be done by rotating the direction of the magnetic field).
In this case, the wave vector ${\bf q}_1$ is different (by $\delta k_F$) in the two nodes of opposite chirality.
If the energy scale $\hbar v_F \delta k_F$ is small compared to the critical temperature obtained from Eq.~(\ref{eq:Tc_nc}), then the effect of the mirror-breaking perturbation on the Peierls instability is unimportant.
If, to the contrary, $\hbar v_F \delta k_F$ is large, only one node will contribute to the Peierls instability.
In this case, the Peierls transition temperature will still be of exponential form, like in Eq.~(\ref{eq:Tc_nc}).
However, there is no longer a qualitative distinction between the softening of scalar and pseudoscalar phonons.
In fact, a single phonon mode will couple to electrons via a combination of scalar and pseudoscalar coupling.
The corresponding transition temperature is given by
\begin{align}
\label{eq:Tc_nc_bm}
T_{c} &=\Lambda_{1}\exp\left(-\frac{2}{\alpha_{1} g_{1}^{\prime2}+\gamma_{1}}\right),
\end{align}
where $\gamma_{1}=\alpha_{1}(g_{2}-g_{1})$, and ($g_{2}$, $g_{1}$, $g_{1}^{\prime}$) refer to the electron-electron and electron-phonon couplings defined for a single nonchiral Landau level.
While Eq.~(\ref{eq:Tc_nc_bm}) has the same form as Eq.~(\ref{Tc0Delta2}), it is quantitatively much lower, mainly because $g_2$ is suppressed when the $|n|=1$ bands are populated and hence contribute to screening.
Next, we briefly discuss the nonadiabatic regime of instabilities for nonchiral Landau levels. In the two-cutoff scaling scheme, when for some given magnetic field interval, $\Lambda_1(l) $ is scaled down to $\Lambda_1(l^*)=\hbar\omega_{0,j}< \Lambda_1$, the adiabatic softening condition (Eq.~(\ref{eq:phprop_nc})) no longer holds and must be replaced by the flow equations shown in Eq. (\ref{eq:rg_ee_nc}). In these, the conditions at $l^*$ for non retarded couplings become $g_{1,2}^* = g_{1,2}(l^*) + \mathfrak{g}_{1,2} $, where $ \mathfrak{g}_{1,2} $ are the attractive phonon-induced backward and forward scattering amplitudes. These can be obtained from Eqs.~(\ref{g1ph}-\ref{g2ph}) after the substitution of the appropriate electron-phonon coupling matrix element (Eqs.~(\ref{geph}) and (\ref{eq:gLR})), of the renormalized phonon propagator (Eq.~(\ref{eq:phprop_nc})) and of vertex parts $z_{j}(l^*)$ (Eq.~(\ref{eq:ep_nc_sol})) at $l^*$. Note that when the applied magnetic field is sufficiently large and $\Lambda_1 < \hbar \omega_{0,j}$, the electron-phonon system is entirely nonadiabatic so that $l^*=0$.
It follows from Eqs.~(\ref{eq:rg_ee_nc}) and (\ref{eq:rg_ep_nc}) that, for $\lambda_P\approx 1$, the combinations of couplings
\begin{align}
\gamma_j(l) = {\gamma_j^*\over 1 -{{\gamma}^*_j \over 2}(l-l^*)}
\end{align}
and the electron-phonon vertex parts
\begin{align}
z_j(l) = {z_j(l^*)\over 1 -{{\gamma}^*_j \over 2}(l-l^*)}
\end{align}
are both developing a simple pole structure. Here $ \gamma_{ps}^*=\alpha_{1} g_2^*$ and $\gamma_{s}^*= \alpha_{1} (g_2^* -2g_1^*)$, whereas the expressions for $z_j(l^*)$ are given by Eq.~(\ref{eq:ep_nc_sol}) at $l^*$. As a function of temperature, we obtain the characteristic scale
\begin{equation}
\label{Tc1nonadia}
T_{1,j}^{*} = {\Lambda_{1,j}\over k_B} e^{-2/\gamma_j^*}
\end{equation}
for an instability of the electron system against a $2k_F\hat{\bf z}$ charge density wave, which is distinct for the scalar ($j=s$) and pseudoscalar ($j=ps$) cases. Here the cutoff energy is $\Lambda_{1,j}= \ {\rm min}\{\hbar\omega_{0,j},\Lambda_1\}$ where $\Lambda_1$ is field-dependent. Therefore at a given applied magnetic field nonadiabatic corrections, which predominantly screen $g_1$ to the attractive sector, will enhance (reduce) the critical temperature (Eq. (\ref{eq:Tc_nc})) for scalar (pseudoscalar) phonons.
The corresponding transition temperature for broken mirror symmetry between the two Weyl nodes, where we effectively consider a single nonchiral Landau level for our analysis, is given by \begin{equation}
\label{Tc1nonadia_bm}
T_{1}^{*} = \frac{\Lambda_{1}}{k_B} e^{-2/\gamma_{1}^*},
\end{equation}
where $\gamma_{1}^*=\alpha_{1} (g_2^* -g_1^*)$.
\section{Application to real materials}
\label{sec:real}
Thus far, we have considered a toy model for two Weyl nodes of opposite chirality, with untilted Weyl cones.
In this section, we extrapolate our results to more realistic situations.
This extrapolation is nontrivial, due to at least three aspects.
First, in the absence of magnetic fields, the Fermi surfaces of several real WSM are non spherical around the Weyl nodes, and moreover host topologically trivial hole bands away from the nodes. This goes well beyond the simple model that we have considered at zero magnetic field. Fortunately, in the strong magnetic field regime that we are interested in, the topologically trivial hole pockets are pushed away from the Fermi energy and the bands at the Fermi level become simpler and rather universal: one-dimensional (though highly degenerate) chiral Landau and nonchiral Landau levels.
Admittedly, our model neglects the curvature of the chiral Landau level. Nevertheless, the main effect of such curvature in our theory (namely a magnetic-field-dependence of the Fermi velocity in the chiral LL) is qualitatively unimportant.
Second, one common element of real Weyl semimetals, which we have ignored throughout the paper, is the tilt of the Weyl cones. It turns out that such a tilt does not change the main results of our theory because the Peierls scattering amplitude (and also the Cooper scattering amplitude, in the presence of mirror symmetry) retains its logarithmic divergence; this can be checked by an explicit calculation.
Third, real WSM most often host multiple pairs of Weyl nodes. This has various implications for our theory, which are worth elaborating on for the remainder of this section.
For example, TaAs and related materials (NbAs, TaP, NbP) display two symmetry-inequivalent multiplets of Weyl nodes (8 W1 nodes and 16 W2 nodes).
The eight symmetry-equivalent W1 nodes are located on the $k_z=0$ plane of the Brillouin zone (see Fig. \ref{fig:taas}).
The sixteen symmetry-equivalent W2 nodes are distributed into two planes ($k_z=\pm k_0$), which are related to one another by time-reversal.
Because the Fermi energy of the W1 nodes is significantly greater than that of the W2 nodes, the quantum limit is first attained in the latter.
\begin{figure}[h!]
\centering
\includegraphics[width=0.3\textwidth]{fig6.png}
\caption{$k_z=0$ plane of the Brillouin zone in TaAs. Eight Weyl nodes are shown (blue and red correspond to opposite chiralities). The grey dashed lines denote the mirror planes.}
\label{fig:taas}
\end{figure}
In the absence of a magnetic field, TaAs belongs to point group $C_{4 v}$.
Weyl nodes of opposite chirality in Fig. \ref{fig:taas} are related to one another by a mirror plane.
At zero field, none of the long-wavelength phonons in TaAs are pseudoscalar\cite{Aroyo2006a, Aroyo2006b}.
Let us suppose that the magnetic field is applied along $z$. Then, all mirror planes are broken but the rotational symmetry is preserved.
The eight W1 nodes of Fig. \ref{fig:taas} can then be subdivided into four symmetry-related nodes of positive chirality, and another four symmetry-equivalent nodes of negative chirality.
The sixteen W2 nodes can be similarly subdivided.
In the quantum limit, the possible phonon instabilities occur for wave vectors connecting Weyl nodes of opposite chirality.
Indeed, since chiral Landau levels of the same chirality have the same dispersion along the direction of $B$, it follows that the nesting condition necessary for the logarithmic singularity of the Peierls and Cooper channels is not satisfied for phonon wave vectors connecting two nodes of the same chirality.
For a wave vector in the $xy$ plane connecting two nodes of opposite chirality, two pairs of chiral Landau levels (in the case of W1 nodes) or four pairs of chiral Landau levels (in the case of W2 nodes) contribute to the instability. For the remaining pairs, the phonon wave vector does not connect the two nodes of opposite chirality.
If we neglect the Coulomb scattering between different pairs of nodes, then our theory of Sec. \ref{sec:cll_no_mirror} for the Peierls transition temperature can be applied, with some modifications, to the case of multiple pairs of nodes.
Specifically, we note the following points: (i) since ${\bf B}||\hat{\bf z}$ breaks all mirrors, the Peierls transition temperature will be of the BCS form;
(ii) the coupling $g_2$ will be reduced by a factor of $N$, where $N$ is the number of pairs of Weyl nodes contributing to the screening of the Coulomb interaction;
(iii) the effective electron-phonon parameter $g_{x,j}^2$ will be enhanced by a factor $N'$, where $N'$ is the number of pairs of Weyl nodes that are connected by the same nesting wave vector.
While (ii) leads to a depletion of the transition temperature, (iii) implies an increase. Hence, the transition temperature for multiple pairs of Weyl nodes can be either higher or lower than the transition temperature for a single pair of Weyl nodes, depending on whether the dominant contribution comes from electron-electron or electron-phonon interactions.
It must also be mentioned that, in the case of multiple pairs of nodes, there could be additional types of instabilities that are not present in the single-pair model.
One aspect of having the magnetic field along the $z$ axis is that the field is perpendicular to the separation between all the Weyl nodes in a constant $k_z$ plane.
Therefore, significant single-particle hybridization gaps between chiral Landau levels of opposite chirality are expected at high fields.
When the size of these gaps exceeds the Peierls transition temperature computed for gapless chiral Landau levels, we expect that the instability will be suppressed.
Next, let us suppose that the magnetic field is applied along the $x$ direction and let us consider a nesting wave vector ${\bf q}$ along $x$ as well.
The single-particle hybridization between chiral Landau levels is relatively unimportant in this configuration, because the magnetic field is parallel to the vector separating the two Weyl nodes that contribute to the Peierls instability.
Such field reduces the symmetry of the crystal from $C_{4v}$ to $C_{1h}$.
As a result,\cite{dresselhaus2008} the $E$ phonons of TaAs, doubly degenerate in the absence of a magnetic field, split into nondegenerate scalar ($A'$) and pseudoscalar ($A''$) phonons.
According to Eq.~(\ref{eq:Tc_nc}), we would expect to see a more significant softening for the $A''$ phonon than for the $A'$ phonon, in the regime in which both $n=0$ and $|n|=1$ Landau levels intersect the Fermi energy.
One caveat is that the hybridization between $A'$ and $A''$ phonons, neglected in our theory, may be significant due to the near degeneracy of these modes.
In the quantum limit with ${\bf B}=B\hat{\bf x}$, the results from Sec.~\ref{sec:cll} indicate that the $B$-dependence of the Peierls transition temperature would be qualitatively different for ${\bf q}||\hat{\bf x}$ and ${\bf q}||\hat{\bf y}$ phonons (compare e.g. Eqs.~(\ref{Tc0}) and ~(\ref{Tc0Delta})), because the former phonons would couple two Weyl nodes related by a mirror plane and the latter phonons would couple Weyl nodes not related by a symmetry.
To close this section, we note that our theory is also applicable to a Dirac semimetal, under the condition that the Zeeman splitting can transform it into a Weyl semimetal.
\section{Summary and conclusions}
\label{sec:conc}
\begin{table*}
\label{table}
\begin{centering}
\begin{tabular}{|c||c|c|c|c|}
\hline
Magnetic field regime & Phonon regime & Mirror symmetry & Field-dependence of transition temperature\tabularnewline
\hline
\hline
\multirow{4}{*}{Quantum limit} & \multirow{2}{*}{Adiabatic} & Yes & Power-law [Eq. (\ref{Tc0})].\tabularnewline
\cline{3-4} \cline{4-4}
& & No & Exponential [Eq. (\ref{Tc0Delta})]. \tabularnewline
\cline{2-4} \cline{3-4} \cline{4-4}
& \multirow{3}{*}{Nonadiabatic} & Yes & No instability (Luttinger liquid).\tabularnewline
\cline{3-4} \cline{4-4}
& & \multirow{2}{*}{No} & \multirow{2}{*}{Exponential [Eq. (\ref{Tc0nonadia}) for $\Delta < \hbar\omega_0$ and Eq. (\ref{Tc0nonadiadelta}) for $\Delta>\hbar\omega_0$].
}\tabularnewline
& & & \tabularnewline
\hline
\multirow{4}{*}{Near-quantum limit} & \multirow{2}{*}{Adiabatic} & Yes & Exponential [Eq. (\ref{eq:Tc_nc})]. Pseudoscalar $>$ scalar.\tabularnewline
\cline{3-4} \cline{4-4}
& & No & Exponential [Eq. (\ref{eq:Tc_nc_bm})].\tabularnewline
\cline{2-4} \cline{3-4} \cline{4-4}
& \multirow{2}{*}{Nonadiabatic} & Yes & Exponential [Eq. (\ref{Tc1nonadia})]. Pseudoscalar $>$ scalar.\tabularnewline
\cline{3-4} \cline{4-4}
& & No & Exponential [Eq. (\ref{Tc1nonadia_bm})].\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\caption{Summary of the main results. {\em First column}: in the quantum limit, only the chiral Landau levels intersect the Fermi energy. In the "near-quantum" limit, the chiral and the first nonchiral Landau levels intersect the Fermi energy. {\em Second column}: the adiabatic (nonadiabatic) phonon regime is realized when $k_B T_c \gg \hbar \omega_0$ ($k_B T_c \ll \hbar\omega_0$), where $T_c$ is the critical temperature for the lattice instability and $\omega_0$ is the bare phonon frequency.
{\em Third column}: when present, mirror symmetry relates the two nodes of opposite chirality. Mirror symmetry may be broken intrinsically (due to lattice structure) or extrinsically (due to the application of a magnetic field that is not perpendicular to the preexisting mirror plane). The mirror symmetry-breaking energy scale $\Delta$ is defined in Eq. (\ref{eq:Delta_main}).
{\em Fourth column}: the magnetic-field-dependence of $T_c$ for the lattice instability can be exponential or power-law. For the near-quantum regime, we consider instabilities at nesting wave vectors that connect electronic states of the same chirality. The power-law behavior in $T_c$ originates from the interference between the Peierls and Cooper channels. In the near-quantum regime with mirror symmetry, "pseudoscalar $>$ scalar" stands for the fact that the instability is qualitatively more likely to occur for pseudoscalar phonons than for scalar phonons, because the former do not generate net charge fluctuations and are thus not screened by Coulomb interactions.}
\label{table}
\end{table*}
To summarize, we have developed a theory of phonon instabilities for a simple model of Weyl semimetal
with two Weyl nodes of opposite chiralities, in the presence of a large applied
magnetic field and electron-electron interactions,
using a Kadanoff-Wilson renormalization group (RG) approach.
In this section, we begin by reviewing the main results collected in Table \ref{table}.
We have focused on two different parameter regimes for the applied
magnetic field $B$: the quantum limit, where the Fermi energy intersects
only the chiral ($n=0$) Landau levels, and the "near-quantum" limit,
where the Fermi level also crosses one nonchiral ($n=1$) Landau level on each node.
We have assumed that the hybridization gap coming from the single-particle magnetic tunneling is smaller than the Fermi energy, so that the system is a metal in the noninteracting limit.
In the quantum limit, the instabilities take place for nesting wave vectors connecting $n=0$ Landau levels of opposite chirality.
We have first considered the situation where the two
Weyl nodes of opposite chirality are related by a mirror plane, and
the field is oriented perpendicular to this plane. In this case, the Peierls
transition temperature $T_c$ obeys a power-law dependence on $B$ (Eq. (\ref{Tc0})), which
originates from the destructive interference between the Peierls and Cooper scattering channels.
The Cooper channel is important even though the two Weyl nodes are not time-reversed partners of one another, provided that the two nodes have opposite momenta along the direction of the magnetic field.
The Peierls transition in this situation depends crucially on the electron-phonon coupling, that is to say we find no instabilities in the absence of electron-phonon interactions.
Moreover, the dependence of $T_c$ on $B$ can be nonmonotonic (Fig.~\ref{fig:Tc}).
Again in the quantum limit, when the two Weyl nodes of opposite chirality
are not related by symmetry, the
Cooper channel is suppressed due to the asymmetry in the positions of the Weyl nodes in momentum space.
On the other hand, the Peierls channel remains unaffected because of the perfect nesting.
This opens the possibility of having a purely electronic instability.
However, in the presence of a nonzero electron-phonon coupling, this instability is preempted by an instability of the lattice, with a BCS-like expression for $T_c$
(Eq. (\ref{Tc0Delta})).
In a material that is mirror-symmetric in the absence of a magnetic field, such symmetry can be broken in a controllable fashion by rotating the magnetic field away from the normal to the mirror plane.
One can therefore gradually turn from Eq. (\ref{Tc0}) to Eq. (\ref{Tc0Delta}); the crossover takes place when the energy scale associated to mirror symmetry-breaking exceeds the $T_c$ in Eq. (\ref{Tc0}).
Still in the quantum limit, we have investigated the nonadiabatic regime, where the bare phonon energy
exceeds the value of $T_c$ obtained in the adiabatic regime.
In this regime, if the two Weyl nodes are related by mirror symmetry,
Luttinger liquid behavior is predicted, with no instability at finite
temperature.
On the other hand, if mirror symmetry is broken,
the aforementioned Luttinger liquid state becomes
unstable against the formation of a charge-density wave order (with $T_c$ given by Eq. (\ref{Tc0nonadia}) or (\ref{Tc0nonadiadelta}), depending on whether the energy scale for mirror symmetry-breaking exceeds or not the bare phonon energy).
Next, we have considered lower magnetic fields.
In contrast to the quantum limit, here lattice instabilities can also emerge at nesting wave vectors that connect two Fermi points of $n=1$ Landau levels within the same chirality. In the next two paragraphs, we focus on such instabilities emerging from $n=1$ Landau levels.
For the adiabatic phonon regime, we obtain a Peierls instability for $n=1$ Landau levels with a $T_c$ of the BCS form; $T_c$ decreases as $B$ increases (Fig. ~\ref{fig:Tc_nonchiral}).
Here, the Cooper channel is suppressed because the Fermi points connected by the nesting wave vector are not symmetric with respect to the $\Gamma$ point.
When mirror symmetry is present, a qualitative difference appears between the behaviors of pseudoscalar phonons (whose deformation potentials have equal magnitudes and opposite signs on Weyl nodes of opposite chirality) and scalar phonons (whose deformation potentials have equal magnitudes and equal signs on Weyl nodes of opposite chirality).
Because pseudoscalar phonons do not generate net charge fluctuations, they are less screened than scalar phonons by Coulomb interactions and are accordingly more prone to undergo a Peierls instability.
In the nonadiabatic regime with mirror symmetry, we obtain a characteristic scale for an instability of the electron system against a charge-density wave order, which is distinct for scalar and pseudoscalar phonons (Eq. (\ref{Tc1nonadia})).
When the mirror symmetry relating the Weyl nodes is broken, there is no longer a qualitative difference between the softening of scalar and pseudoscalar phonons (Eq.~(\ref{eq:Tc_nc_bm})).
If we extrapolate our results to lower fields, where the Fermi level intersects multiple nonchiral Landau levels, we anticipate that the leading instability will occur at the nesting wave vector connecting chiral Landau levels of opposite chirality. Such instability will however take place at significantly lower temperatures than the corresponding one in the quantum limit, mainly because of the contribution of the nonchiral Landau levels to the screening of the forward Coulomb scattering.
This completes the summary of our work.
The main messages are the following: (1) it is important to take electron-phonon interactions into account, along with electronic interaction effects, for a complete understanding of interacting phases of matter of Weyl semimetals at high fields; (2) electron-phonon interactions tend to augment the charge-density wave fluctuations in the system and, whenever a Peierls transition is possible, lead to an enhancement of the corresponding transition temperature;
(3) the Cooper channel plays an important role when the the two nodes are related by a mirror plane, and can prevent purely electronic instabilities while enabling non-BCS-like lattice instabilities; (4) breaking mirror symmetry increases $T_c$ for the Peierls transition by suppressing the Cooper channel; (5) $T_c$ is expected to be higher for instabilities emerging from chiral (as opposed to nonchiral) Landau levels.
For the remainder of this section, we compare our theory to the existing literature.
First, the importance of electron-phonon interactions in the putative charge-density wave ordered state in the Dirac semimetal ZrTe$_5$ at high magnetic field
has been suggested in a recent theoretical study\cite{Qin2020}.
The authors of this paper use a mean-field approach to compare the threshold magnetic field for the order parameter, calculated using electron-phonon and electron-electron interactions alone, whereas we perform a detailed analysis of the Peierls transition resulting from the interplay of these two effects, to compare their relative importance for the transition. Moreover, Ref.~[\onlinecite{Qin2020}] neglects the Cooper channel contribution, which, as we have shown, can qualitatively change the behavior of the transition through its interference with the Peierls channel.
Second, Ref.~[\onlinecite{Zhao2020}] has developed an RG theory for a metal-insulator transition in the quantum limit of a Dirac material, treating electron-electron interactions, electron-phonon interactions and disorder on an equal footing. The authors of Ref.~[\onlinecite{Zhao2020}] do not consider forward scattering and vertex corrections, and do not take into account the Cooper channel contributions, all of which play a role in our final expression for the Peierls transition temperature. Moreover, Ref.~[\onlinecite{Zhao2020}] considers phonon-induced electronic interactions as the starting point, which would correspond to the nonadiabatic regime in our analysis. We, in turn, do not consider the effect of disorder. These differences make it difficult to directly compare the results of Ref.~[\onlinecite{Zhao2020}] with ours.
Third, the lattice instabilities studied in our work differ in a number of ways from the ones that are well-known in quasi one-dimensional conductors\cite{Pouget2015}.
The nonadiabatic regime is less likely to be relevant in the latter, because they are soft materials with low-frequency phonons and a relatively high value for the tight-binding electron-phonon matrix element. Also, the cutoff band energy entering the Peierls transition temperature is significantly larger compared to the values found for WSM.
Moreover, in quasi-1D conductors, the low-energy electron bands are symmetric with respect to the $\Gamma$ point. This is not the case for Weyl fermion bands in Weyl semimetals, which can affect the importance of the Cooper channel depending on the orientation of the magnetic field.
Finally, there is no spin degeneracy in WSM, contrary to quasi-1D conductors. In these, the possibility of a density-wave instability for electrons induced by an applied magnetic field
is known to exist, but is restricted to the spin sector in the form of a field-induced spin-density wave state\cite{Lebed08}.
Fourth, we compare our work to the literature on magnetic catalysis, which has been long-studied in high-energy physics and has recently received attention in the context of Weyl and Dirac semimetals\cite{Miransky2015}.
Magnetic catalysis consists of a dynamical generation of a Dirac mass in an interacting system of initially massless Dirac (or Weyl) fermions subjected to a strong magnetic field.
Our work can be regarded as a theory of magnetic catalysis in two-node Weyl semimetals, in the presence of both electron-electron and electron-phonon interactions.
If we turn off electron-phonon interactions in our theory, we can compare our results to earlier works in the literature of magnetic catalysis. While Ref. [\onlinecite{Miransky2015}] reviews mainly the mean-field approach, consistent results have been reported in RG approaches that are closer in spirit to our work (see e.g. Refs. [\onlinecite{Hong1996}] and [\onlinecite{Hattori2017}]).
In the latter RG approaches, an effective electron-electron interaction parameter $g$ flows to strong coupling provided that the bare value of $g$ is positive, thereby producing the dynamical generation of a Dirac mass. The dynamical mass is BCS-like. The flow equation for $g$ in those theories coincides with our flow equation (\ref{g2mg1}) for $g_2-g_1$ (i.e. difference between forward and backward Coulomb scattering), but only provided that the Cooper channel contribution is negligible.
While the neglect of the Cooper channel may be justified in models with multiple fermion flavors,
we have shown that the Cooper channel contribution is important in the quantum limit of two-node Weyl semimetals, if the nodes are related by a crystal symmetry. In that situation, the Cooper channel interferes destructively with the Peierls channel, thereby leading to an effective electron-electron coupling $g_2-g_1$ that no longer flows to strong coupling. In other words, the Cooper channel contribution precludes the emergence of magnetic catalysis in the absence of electron-phonon interactions.
When electron-phonon interactions are included, we find the magnetic catalysis re-emerges in the form of a Peierls instability, but with a critical temperature that is no longer BCS-like.
In our theory, the contribution from the Cooper channel becomes suppressed when the two Weyl nodes are not related by a crystal symmetry. In this case, only the Peierls channel remains intact and $g_2-g_1$ flows to strong coupling provided that the bare value of $g_2-g_1$ is positive, like predicted in the RG treatments of magnetic catalysis. Yet, even in this case, our theory offers a new result: the purely electronic instability resulting from the divergence of $g_2-g_1$ is preempted by a lattice instability, because the electron-phonon vertex diverges before $g_2-g_1$ does.
To conclude, we mention that possible avenues for future work include a generalization of our theory to semimetals with multiple pairs of nodes, and its application to study electronic and lattice instabilities in other interesting semimetals such as graphite, bismuth and chiral multifold fermion systems.
\acknowledgements
This paper has been financially supported by the Canada First Research Excellence Fund (CFREF) and by the Natural Sciences and Engineering Research Council of Canada (NSERC). C.B. thanks NSERC under Grant No. RGPIN- 2016-06017. I.G. thanks NSERC under Grant No. RGPIN- 2018-05385. S.K. acknowledges financial support from the Postdoctoral Fellowship from Institut Quantique, and from UF Project No. P0224175 - Dirac postdoc fellowship, sponsored by the Florida State University National High Magnetic Field Laboratory (NHMFL). I.G. thanks R. C\^ot\'e for a useful discussion. S.K. gratefully acknowledges useful discussions with Vikram Tripathi.
|
\section*{Abstract}
{\bf
I present results for soft anomalous dimensions through three loops for many QCD processes. In particular, I give detailed expressions for soft anomalous dimensions in various processes with electroweak and Higgs bosons as well as single top quarks and top-antitop pairs.
}
\vspace{10pt}
\noindent\rule{\textwidth}{1pt}
\tableofcontents\thispagestyle{fancy}
\noindent\rule{\textwidth}{1pt}
\vspace{10pt}
\section{Introduction}
The calculation of higher-order soft-gluon corrections in perturbative QCD requires calculations of soft anomalous dimensions, $\Gamma_S$, for the corresponding processes \cite{NKGS}. The current state-of-the-art for $\Gamma_S$ for many processes is three loops. In this paper, I present results for $\Gamma_S$ for various processes at hadron colliders. These include processes with $W$, $Z$, $\gamma$, and $H$ bosons, as well as single-top and top-pair production, and $2 \to 3$ processes involving top quarks produced in association with electroweak or Higgs bosons.
Soft-gluon corrections are very important because they are typically large and they dominate the perturbative corrections for a multitude of processes, especially those involving top quarks. We consider partonic processes $p_a +p_b \to p_1 +p_2+ \cdots$
and define $s=(p_a+p_b)^2$, $t=(p_a-p_1)^2$, $u=(p_b-p_1)^2$ and $s_4=s+t+u-\sum m_i^2$. At partonic threshold $s_4 \rightarrow 0$,
and the soft corrections at order $\alpha_s^n$ involve logarithmic terms of the form $\ln^k(s_4/M^2)/s_4$, with $M$ a hard scale and $k \le 2n-1$. In order to resum these soft corrections in the (differential) cross section at NLL, NNLL, and N$^3$LL accuracy, we need
to calculate soft anomalous dimensions at, correspondingly, one loop, two loops, and three loops.
If we take transforms of the cross section, with transform variable $N$, then we can write a factorized expression as
\begin{equation}
\sigma^{ab \to 12\cdots}(N)
= {\rm tr} \left\{H^{ab \to 12\cdots} \,
S^{ab \to 12\cdots}\left(\frac{\sqrt{s}}{N \mu_F}\right) \right\} \,
\psi_a \left(N_a,\mu_F\right) \,
\psi_b \left(N_b,\mu_F\right) \,
\prod J_i \left(N,\mu_F\right)
\nonumber
\end{equation}
where the $\psi$ and $J$ functions describe collinear emission from incoming and outgoing partons, $H^{ab \to 12\cdots}$ is a short-distance hard function, and $S^{ab \to 12\cdots}$ is a soft function which describes soft-gluon emission \cite{NKGS} and which satisfies the renormalization group equation
\begin{equation}
\left(\mu_R \frac{\partial}{\partial \mu_R}
+\beta(g_s)\frac{\partial}{\partial g_s}\right)\,S^{ab \to 12\cdots}
=-\Gamma_S^{\dagger \, ab \to 12\cdots} \, \, S^{ab \to 12\cdots}
-S^{ab \to 12\cdots} \, \, \Gamma_S^{ab \to 12\cdots} \, .
\nonumber
\end{equation}
The soft anomalous dimension $\Gamma_S^{ab \to 12\cdots}$ controls the evolution of the soft function which gives the exponentiation of logarithms of $N$ in the resummed cross section. For a recent review of soft anomalous dimensions for many QCD processes, see Ref. \cite{NKuni}.
\section{Cusp anomalous dimension}
The cusp anomalous dimension \cite{AMP,BNS,IKR,GKAR,NK2loop,GHKM,NK3loopcusp} is the simplest type of $\Gamma_S$ and a basic ingredient of calculations for QCD processes.
For eikonal lines with momenta $p_i$ and $p_j$ we define the cusp angle $\theta=\cosh^{-1}(p_i\cdot p_j/\sqrt{p_i^2 p_j^2})$.
The perturbative series is $\Gamma_{\rm cusp}=\sum_{n=1}^{\infty} (\alpha_s/\pi)^n \, \Gamma^{(n)}_{\rm cusp}$
where at one loop $\Gamma_{\rm cusp}^{(1)}=C_F (\theta \coth\theta-1)$, at two loops
\begin{eqnarray}
\Gamma_{\rm cusp}^{(2)}&=& K_2 \, \Gamma_{\rm cusp}^{(1)}
+\frac{1}{2}C_F C_A \left\{1+\zeta_2+\theta^2
-\coth\theta\left[\zeta_2\theta+\theta^2
+\frac{\theta^3}{3}+{\rm Li}_2\left(1-e^{-2\theta}\right)\right] \right.
\nonumber \\ && \hspace{30mm} \left.
{}+\coth^2\theta\left[-\zeta_3+\zeta_2\theta+\frac{\theta^3}{3}
+\theta \, {\rm Li}_2\left(e^{-2\theta}\right)
+{\rm Li}_3\left(e^{-2\theta}\right)\right] \right\} \, ,
\nonumber
\end{eqnarray}
and at three loops
$\Gamma_{\rm cusp}^{(3)}= K_3 \Gamma_{\rm cusp}^{(1)}
+2 \, K_2 (\Gamma_{\rm cusp}^{(2)}-K_2 \, \Gamma_{\rm cusp}^{(1)})+C^{(3)}$,
where $K_3$ and $C^{(3)}$ have long expressions (see Refs. \cite{NKuni,NK3loopcusp} for explicit expressions)
and $K_2=C_A (67/36-\zeta_2/2)-(5/18)n_f$.
In the case of the production of heavy-quark pairs, with mass $m$, we can also write the above expressions in terms of $\beta=\tanh(\theta/2)=\sqrt{1-(4m^2/s)}$, and denote them by
$\Gamma_{\rm cusp}^{(n) \, \beta}$.
If eikonal line $i$ represents a massive quark and eikonal
line $j$ a massless quark, then we have simpler expressions.
At one loop $\Gamma_{\rm cusp}^{(1) \, m_i}=C_F [\ln(2 p_i \cdot p_j/(m_i \sqrt{s}))-1/2]$,
at two loops $\Gamma^{(2) \, m_i}_{\rm cusp}=K_2 \, \Gamma_{\rm cusp}^{(1) \, m_i}+(1/4) C_F C_A (1-\zeta_3)$, and at three loops
\begin{equation}
\Gamma^{(3) \, m_i}_{\rm cusp}=K_3 \, \Gamma_{\rm cusp}^{(1) \, m_i}
+ \frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right) \, .
\nonumber
\end{equation}
If both eikonal lines are massless, then
$\Gamma_{\rm cusp}^{\rm massless}=C_F \ln(2 p_i \cdot p_j/s) \, \sum_{n=1}^{\infty} (\alpha_s/\pi)^n \, K_n$.
\section{$\Gamma_S$ for some simple processes}
For processes with trivial color structure, the soft anomalous dimension is very simple. In fact
$\Gamma_S$ vanishes for the following:
Drell-Yan processes $q{\bar q} \to \gamma^*$, $q{\bar q} \to Z$;
$W$-boson production via $q{\bar q'} \to W^{\pm}$;
Higgs production via $b{\bar b} \to H$ and $gg \to H$;
electroweak-boson pair production $q{\bar q} \to \gamma \gamma$, $q{\bar q} \to Z Z$, $q{\bar q} \to W^+ W^-$;
production of two different electroweak bosons $q{\bar q} \to \gamma Z$, $q{\bar q'} \to W^{\pm} \gamma $, $q{\bar q'} \to W^{\pm} Z$; charged Higgs production via $b{\bar b} \to H^- W^+$, $b{\bar b} \to H^+ H^-$, $gg \to H^+ H^-$.
Also, for Deep Inelastic Scattering (DIS), $lq \to lq$ with subprocess $q \gamma^* \to q$, we have
at one loop: $\Gamma_S^{(1) \, q \gamma^* \rightarrow q}=C_F \ln(-t/s)$;
at two loops: $\Gamma_S^{(2) \, q \gamma^* \rightarrow q}=K_2 \, C_F \ln(-t/s)$;
and at three loops: $\Gamma_S^{(3) \, q \gamma^* \rightarrow q}=K_3 \, C_F \ln(-t/s)$.
More generally, when all external lines in a process are massless, then $\Gamma_S^{(2)}$ is proportional to
$\Gamma_S^{(1)}$ \cite{ADS}, but this is not true for processes with massive lines. Furthermore, at three loops for multi-leg scattering there are contributions from four-parton correlations \cite{ADG}.
\section{$\Gamma_S$ for large-$p_T$ $W$, $Z$, $\gamma$, $H$ production}
Let $V$ denote a $W$ or $Z$ boson or a photon or a Higgs boson. The soft anomalous dimension for these processes is a simple function (not a matrix) \cite{LOS,NKsv,NKRG} (see also \cite{NKuni}).
For the processes $qg \to W^{\pm}q'$, $qg \to Zq$, $qg \to \gamma q$, and $bg \to Hb$, we have
at one loop: $\Gamma_S^{(1) \, qg \to Vq'}=C_F \ln(-u/s)
+(C_A/2) \ln(t/u)$;
at two loops: $\Gamma_S^{(2) \, qg \to Vq'}=K_2 \, \Gamma_S^{(1) \, qg \to Vq'}$; and
at three loops: $\Gamma_S^{(3) \, qg \to Vq'}=K_3 \, \Gamma_S^{(1) \, qg \to Vq'}$.
The same $\Gamma_S$ also describes the reverse processes such as $\gamma q \rightarrow qg$.
For the processes $q {\bar q'} \to W^{\pm}g$, $q {\bar q} \to Zg$, $q{\bar q} \to \gamma g$, and $b{\bar b} \to Hg$,
we have at one loop: $\Gamma_S^{(1) \, q{\bar q'} \to Vg}=(C_A/2) \ln(tu/s^2)$;
at two loops: $\Gamma_S^{(2) \, q{\bar q'} \to Vg}=K_2 \, \Gamma_S^{(1) \, q{\bar q'} \to Vg}$;
and at three loops: $\Gamma_S^{(3) \, q{\bar q'} \to Vg}=K_3 \, \Gamma_S^{(1) \, q{\bar q'} \to Vg}$. The same $\Gamma_S$ also describes the reverse processes such as $\gamma g \rightarrow q {\bar q}$.
\section{$\Gamma_S$ for single-top production}
We continue with results for single-top production \cite{NKsingletop,NKsch,NKtWH,NKtch,NK3loop} (see also \cite{NKuni,NKdis21}.
For single-top $t$-channel production,
${\Gamma}_S^{bq \to tq'}$ is a $2 \times 2$ matrix \cite{NKsingletop,NKtch,NK3loop}.
Using a $t$-channel singlet-octet color basis, the matrix elements are
at one loop
\begin{eqnarray}
&& \hspace{-7mm}
{\Gamma}_{S \, 11}^{(1) bq \to tq'}=
C_F \left[\ln\left(\frac{t(t-m_t^2)}{m_t s^{3/2}}\right)-\frac{1}{2}\right] \, , \quad \quad \quad
{\Gamma}_{S \, 12}^{(1) bq \to tq'}=\frac{C_F}{2N_c} \ln\left(\frac{u(u-m_t^2)}{s(s-m_t^2)}\right) \, ,
\nonumber \\ && \hspace{-7mm}
{\Gamma}_{S \, 21}^{(1) bq \to tq'}= \ln\left(\frac{u(u-m_t^2)}{s(s-m_t^2)}\right) \, , \quad \quad
{\Gamma}_{S \, 22}^{(1) bq \to tq'}= \frac{C_A}{2}\left[\ln\left(\frac{u(u-m_t^2)}{m_t \, s^{3/2}}\right)-\frac{1}{2}\right]
\nonumber \\ && \hspace{50mm}
{}+\left(C_F-\frac{C_A}{2}\right) \left[\ln\left(\frac{t(t-m_t^2)}{m_t \, s^{3/2}}\right)-\frac{1}{2}
+2\ln\left(\frac{u(u-m_t^2)}{s(s-m_t^2)}\right)\right] \, ,
\nonumber
\end{eqnarray}
at two loops
\begin{eqnarray}
&& \Gamma_{S \, 11}^{(2) bq \to tq'}= K_2 \, \Gamma_{S \, 11}^{(1) bq \to tq'}+\frac{1}{4} C_F C_A (1-\zeta_3)\, ,
\quad \quad
\Gamma_{S \, 12}^{(2) bq \to tq'}= K_2 \, \Gamma_{S \, 12}^{(1) bq \to tq'} \, ,
\nonumber \\ &&
\Gamma_{S \, 21}^{(2) bq \to tq'}= K_2 \, \Gamma_{S \, 21}^{(1) bq \to tq'} \, , \quad \quad
\Gamma_{S \, 22}^{(2) bq \to tq'}= K_2 \, \Gamma_{S \, 22}^{(1) bq \to tq'}+\frac{1}{4} C_F C_A (1-\zeta_3) \, ,
\nonumber
\end{eqnarray}
and at three loops
\begin{eqnarray}
&& \hspace{-5mm} \Gamma_{S \, 11}^{(3) bq \to tq'}= K_3 \, \Gamma_{S \, 11}^{(1) bq \to tq'}
+ \frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right) \, ,
\nonumber \\ && \hspace{-5mm}
\Gamma_{S \, 12}^{(3) bq \to tq'}= K_3 \, \Gamma_{S \, 12}^{(1) bq \to tq'} + X_{12}^{(3) bq \to tq'} \, , \quad \quad
\Gamma_{S \, 21}^{(3) bq \to tq'}= K_3 \, \Gamma_{S \, 21}^{(1) bq \to tq'} + X_{21}^{(3) bq \to tq'} \, ,
\nonumber \\ && \hspace{-5mm}
\Gamma_{S \, 22}^{(3) bq \to tq'}= K_3 \, \Gamma_{S \, 22}^{(1) bq \to tq'}
+ \frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right)
\nonumber \\ && \hspace{15mm}
{}+ X_{22}^{(3) bq \to tq'} \, ,
\nonumber
\end{eqnarray}
where the $X_{ij}^{(3) bq \to tq'}$
denote unknown terms from four-parton correlations in the last three matrix elements at three loops. It is important to note that due to the color structure of this process, only the first three-loop matrix element, $\Gamma_{S \, 11}^{(3) bq \to tq'}$, contributes to the N$^3$LO soft-gluon corrections; therefore, the unknown terms in the other three-loop matrix elements do not pose a problem in deriving N$^3$LO results.
For single-top $s$-channel production,
${\Gamma}_S^{q{\bar q}' \to t{\bar b}}$ is also a $2 \times 2$ matrix \cite{NKsingletop,NKsch,NK3loop}. Using an $s$-channel singlet-octet color basis, we have at one loop
\begin{eqnarray}
&& \hspace{-8mm}
\Gamma_{S \, 11}^{(1) q{\bar q}' \to t{\bar b}}=C_F \left[\ln\left(\frac{s-m_t^2}{m_t\sqrt{s}}\right)
-\frac{1}{2}\right] \, ,
\quad \quad \quad
\Gamma_{S \, 12}^{(1) q{\bar q}' \to t{\bar b}}=\frac{C_F}{2N_c} \ln\left(\frac{t(t-m_t^2)}{u(u-m_t^2)}\right) \, ,
\nonumber \\ && \hspace{-8mm}
\Gamma_{S \, 21}^{(1) q{\bar q}' \to t{\bar b}}= \ln\left(\frac{t(t-m_t^2)}{u(u-m_t^2)}\right) \, , \quad \quad \quad
\Gamma_{S \, 22}^{(1) q{\bar q}' \to t{\bar b}}=\frac{C_A}{2} \left[\ln\left(\frac{t(t-m_t^2)}{m_t \, s^{3/2}}\right)-\frac{1}{2}\right]
\nonumber \\ && \hspace{53mm}
{}+\left(C_F-\frac{C_A}{2}\right) \left[\ln\left(\frac{s-m_t^2}{m_t \sqrt{s}}\right)-\frac{1}{2}+2\ln\left(\frac{t(t-m_t^2)}{u(u-m_t^2)}\right)\right] \, ,
\nonumber
\end{eqnarray}
at two loops
\begin{eqnarray}
&&
\Gamma_{S \, 11}^{(2) q{\bar q}' \to t{\bar b}}=K_2 \, \Gamma_{S \, 11}^{(1) q{\bar q}' \to t{\bar b}}+\frac{1}{4} C_F C_A (1-\zeta_3) \, , \hspace{8mm}
\Gamma_{S \, 12}^{(2) q{\bar q}' \to t{\bar b}}=K_2 \, \Gamma_{S \, 12}^{(1) q{\bar q}' \to t{\bar b}} \, ,
\nonumber \\ &&
\Gamma_{S \, 21}^{(2) q{\bar q}' \to t{\bar b}}= K_2 \, \Gamma_{S \, 21}^{(1) q{\bar q}' \to t{\bar b}} \, , \hspace{8mm}
\Gamma_{S \, 22}^{(2) q{\bar q}' \to t{\bar b}}= K_2 \, \Gamma_{S \, 22}^{(1) q{\bar q}' \to t{\bar b}}+\frac{1}{4} C_F C_A (1-\zeta_3) \, ,
\nonumber
\end{eqnarray}
and at three loops
\begin{eqnarray}
&& \hspace{-2mm}
\Gamma_{S \, 11}^{(3) q{\bar q}' \to t{\bar b}}= K_3 \, \Gamma_{S \, 11}^{(1) q{\bar q}' \to t{\bar b}}
+\frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right) \, ,
\nonumber \\ && \hspace{-2mm}
\Gamma_{S \, 12}^{(3) q{\bar q}' \to t{\bar b}}=K_3 \, \Gamma_{S \, 12}^{(1) q{\bar q}' \to t{\bar b}} + X_{12}^{(3) q{\bar q}' \to t{\bar b}} \, , \hspace{8mm}
\Gamma_{S \, 21}^{(3) q{\bar q}' \to t{\bar b}}= K_3 \, \Gamma_{S \, 21}^{(1) q{\bar q}' \to t{\bar b}} + X_{21}^{(3) q{\bar q}' \to t{\bar b}} \, ,
\nonumber \\ && \hspace{-2mm}
\Gamma_{S \, 22}^{(3) q{\bar q}' \to t{\bar b}}= K_3 \, \Gamma_{S \, 22}^{(1) q{\bar q}' \to t{\bar b}}
+\frac{1}{2} K_2 C_F C_A (1-\zeta_3)+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right)
\nonumber \\ && \hspace{18mm}
{}+ X_{22}^{(3) q{\bar q}' \to t{\bar b}} \, ,
\nonumber
\end{eqnarray}
where the $X_{ij}^{(3) q{\bar q}' \to t{\bar b}}$
denote unknown terms in the last three matrix elements. Again, we note that only the first three-loop matrix element, $\Gamma_{S \, 11}^{(3) q{\bar q}' \to t{\bar b}}$, contributes to the N$^3$LO soft-gluon corrections.
For associated $tW$ production the soft anomalous dimension is a simple function \cite{NKsingletop,NKtWH,NK3loop}.
At one loop
\begin{equation}
\Gamma_S^{(1) bg \to tW}=C_F \left[\ln\left(\frac{m_t^2-t}{m_t\sqrt{s}}\right)
-\frac{1}{2}\right] +\frac{C_A}{2} \ln\left(\frac{u-m_t^2}{t-m_t^2}\right) \, ,
\nonumber
\end{equation}
at two loops
\begin{equation}
\Gamma_S^{(2) bg \to tW}=K_2 \, \Gamma_S^{(1) bg \to tW}
+\frac{1}{4}C_F C_A (1-\zeta_3) \, ,
\nonumber
\end{equation}
and at three loops
\begin{equation}
\Gamma_S^{(3) bg \to tW}=K_3 \, \Gamma_S^{(1) bg \to tW}+\frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right) \, .
\nonumber
\end{equation}
The same soft anomalous dimension applies for the process $bg \to tH^-$, and for the FCNC processes, via anomalous top-quark couplings, $qg \to tZ$, $qg \to tZ'$, and $qg \to t\gamma$.
\section{$\Gamma_S$ for top-antitop pair production}
We continue with soft anomalous dimension matrices for $t{\bar t}$ production \cite{NKGS,NK2loop,FNPY,NKttbar} (see also \cite{NKuni,NKdis21}).
For top-antitop pair production via the $q{\bar q} \to t{\bar t}$ channel,
${\Gamma}_S^{q{\bar q} \to t{\bar t}}$ is a $2 \times 2$ matrix and we use an $s$-channel singlet-octet color basis. At one loop for $q{\bar q} \to t{\bar t}$
\begin{eqnarray}
\Gamma^{(1) q{\bar q} \to t{\bar t}}_{S \, 11}&=&\Gamma_{\rm cusp}^{(1) \, \beta} \, ,
\quad
\Gamma^{(1) q{\bar q} \to t{\bar t}}_{12}=
\frac{C_F}{C_A} \ln\left(\frac{t-m_t^2}{u-m_t^2}\right) \, ,
\quad
\Gamma^{(1) q{\bar q} \to t{\bar t}}_{21}=
2\ln\left(\frac{t-m_t^2}{u-m_t^2}\right) \, ,
\nonumber \\
\Gamma^{(1) q{\bar q} \to t{\bar t}}_{22}&=&\left(1-\frac{C_A}{2C_F}\right)
\Gamma_{\rm cusp}^{(1)}
+4C_F \ln\left(\frac{t-m_t^2}{u-m_t^2}\right)
-\frac{C_A}{2}\left[1+\ln\left(\frac{s m_t^2 (t-m_t^2)^2}{(u-m_t^2)^4}\right)\right] \, ,
\nonumber
\end{eqnarray}
and at two loops
\begin{eqnarray}
&& \hspace{-8mm} \Gamma^{(2) q{\bar q} \to t{\bar t}}_{S\, 11}=\Gamma_{\rm cusp}^{(2) \, \beta}, \, \hspace{2mm}
\Gamma^{(2) q{\bar q} \to t{\bar t}}_{12}=
\left(K_2-C_A \, N_2^{\beta}\right) \Gamma^{(1) q{\bar q} \to t{\bar t}}_{12}, \, \hspace{2mm}
\Gamma^{(2) q{\bar q} \to t{\bar t}}_{21}=
\left(K_2+C_A \, N_2^{\beta}\right) \Gamma^{(1) q{\bar q} \to t{\bar t}}_{21} \, ,
\nonumber \\ && \hspace{-8mm}
\Gamma^{(2) q{\bar q} \to t{\bar t}}_{22}=
K_2 \Gamma^{(1) q{\bar q} \to t{\bar t}}_{22}
+\left(1-\frac{C_A}{2C_F}\right)
\left(\Gamma_{\rm cusp}^{(2) \, \beta}-K_2 \Gamma_{\rm cusp}^{(1) \, \beta}\right)
+\frac{1}{4} C_A^2(1-\zeta_3) \, ,
\nonumber
\end{eqnarray}
where
\begin{equation}
N_2^{\beta}=\frac{1}{4}\ln^2\left(\frac{1-\beta}{1+\beta}\right)
+\frac{(1+\beta^2)}{8 \beta} \left[\zeta_2
-\ln^2\left(\frac{1-\beta}{1+\beta}\right)
-{\rm Li}_2\left(\frac{4\beta}{(1+\beta)^2}\right)\right] \, .
\nonumber
\end{equation}
At three loops for $q{\bar q} \to t{\bar t}$ we can write the last matrix element as
\begin{eqnarray}
\Gamma_{S \, 22}^{(3) q{\bar q} \to t{\bar t}}&=&
K_3 \, \Gamma_{S \, 22}^{(1) q{\bar q} \to t{\bar t}}
+\left(1-\frac{C_A}{2C_F}\right) \left(\Gamma_{\rm cusp}^{(3) \, \beta}-K_3 \Gamma_{\rm cusp}^{(1) \, \beta}\right)+\frac{1}{2}K_2 C_A^2(1-\zeta_3)
\nonumber \\ &&
{}+C_A^3\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2\zeta_3+\frac{9}{16}\zeta_5\right)+X_{22}^{(3) q{\bar q} \to t{\bar t}} \, ,
\nonumber
\end{eqnarray}
where $X_{22}^{(3) q{\bar q} \to t{\bar t}}$ denotes unknown three-loop contributions from four-parton correlations.
The other matrix elements are also not fully known at three loops, but they have an analogous structure to that at two loops (essentially, replace (2)'s by (3)'s in the superscripts as well as replace $K_2$'s by $K_3$'s, and add $X$ terms for unknown contributions).
For top-antitop pair production via the $gg \to t{\bar t}$ channel, $\Gamma_S^{gg \to t{\bar t}}$ is a $3\times3$ matrix, and we use a color basis
$c_1=\delta^{ab} \delta_{12}$, $c_2= d^{abc} T^c_{12}$, $c_3= if^{abc} T^c_{12}$.
We have
\begin{eqnarray}
\Gamma_S^{gg \to t{\bar t}}=\left[\begin{array}{ccc}
\Gamma_{S \, 11}^{gg \to t{\bar t}} & 0 & \Gamma_{S \, 13}^{gg \to t{\bar t}} \vspace{2mm} \\
0 & \Gamma_{S \, 22}^{gg \to t{\bar t}} & \Gamma_{S \, 23}^{gg \to t{\bar t}} \vspace{2mm} \\
\Gamma_{S \, 31}^{gg \to t{\bar t}} & \Gamma_{S \, 32}^{gg \to t{\bar t}} & \Gamma_{S \, 22}^{gg \to t{\bar t}}
\end{array}
\right] \, .
\nonumber
\end{eqnarray}
At one loop for $gg \to t{\bar t}$
\begin{eqnarray}
\Gamma_{S \, 11}^{(1) gg \to t{\bar t}}&=& \Gamma_{\rm cusp}^{(1) \, \beta} \, ,
\quad
\Gamma_{S \, 13}^{(1) gg \to t{\bar t}}= \ln\left(\frac{t-m_t^2}{u-m_t^2}\right) \, , \quad
\Gamma_{S \, 31}^{(1) gg \to t{\bar t}}= 2 \ln\left(\frac{t-m_t^2}{u-m_t^2}\right),
\nonumber \\
\Gamma_{S \, 22}^{(1) gg \to t{\bar t}}&=&\left(1-\frac{C_A}{2C_F}\right) \Gamma_{\rm cusp}^{(1) \, \beta}
+\frac{C_A}{2}\left[\ln\left(\frac{(t-m_t^2)(u-m_t^2)}{s\, m_t^2}\right)-1\right] \, , \quad
\nonumber \\
\Gamma_{S \, 23}^{(1) gg \to t{\bar t}}&=&\frac{C_A}{2} \ln\left(\frac{t-m_t^2}{u-m_t^2}\right) \, , \quad
\Gamma_{S \, 32}^{(1) gg \to t{\bar t}}=\frac{(N_c^2-4)}{2N_c} \ln\left(\frac{t-m_t^2}{u-m_t^2}\right) \, ,
\nonumber
\end{eqnarray}
and at two loops
\begin{eqnarray}
\Gamma_{S \, 11}^{(2) gg \to t{\bar t}}&=& \Gamma_{\rm cusp}^{(2) \, \beta}, \,
\Gamma_{S \, 13}^{(2) gg \to t{\bar t}}=\left(K_2-C_A N_2^{\beta}\right)
\Gamma_{S \, 13}^{(1) gg \to t{\bar t}}, \,
\Gamma_{S \, 31}^{(2) gg \to t{\bar t}}=\left(K_2+C_A N_2^{\beta}\right)
\Gamma_{S \, 31}^{(1) gg \to t{\bar t}},
\nonumber \\
\Gamma_{S \, 22}^{(2) gg \to t{\bar t}}&=& K_2 \, \Gamma_{S \, 22}^{(1) gg \to t{\bar t}}
+\left(1-\frac{C_A}{2C_F}\right) \left(\Gamma_{\rm cusp}^{(2) \, \beta}-K_2 \Gamma_{\rm cusp}^{(1) \, \beta}\right)+\frac{1}{4} C_A^2(1-\zeta_3) \, ,
\nonumber \\
\Gamma_{S \, 23}^{(2) gg \to t{\bar t}}&=& K_2 \, \Gamma_{S \, 23}^{(1) gg \to t{\bar t}} \, , \quad
\Gamma_{S \, 32}^{(2) gg \to t{\bar t}}= K_2 \, \Gamma_{S \, 32}^{(1) gg \to t{\bar t}} \, .
\nonumber
\end{eqnarray}
At three loops for $gg \rightarrow t{\bar t}$, we can write the $22$ matrix element as
\begin{eqnarray}
\Gamma_{S \, 22}^{(3) gg \to t{\bar t}}&=&
K_3 \, \Gamma_{S \, 22}^{(1) gg \to t{\bar t}}
+\left(1-\frac{C_A}{2C_F}\right) \left(\Gamma_{\rm cusp}^{(3) \, \beta}-K_3 \Gamma_{\rm cusp}^{(1) \, \beta}\right)+\frac{1}{2}K_2 C_A^2(1-\zeta_3)
\nonumber \\ &&
{}+C_A^3\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2\zeta_3+\frac{9}{16}\zeta_5\right)+X_{22}^{(3) gg \to t{\bar t}} \, ,
\nonumber
\end{eqnarray}
where $X_{22}^{(3) gg \to t{\bar t}}$ denotes unknown three-loop contributions from four-parton correlations. The other matrix elements are, again, also not fully known at three loops, but they have an analogous structure to that at two loops.
\section{$\Gamma_S$ for $tqH$, $tqZ$, $tq\gamma$, $tqW$ production}
We consider processes $bq \to tq'H$ as well as $bq \to tq'Z$, $bq \to tq'\gamma$, $bq \to tqW^-$, $qq \to tq'W^+$. We use a $t$-channel singlet-octet color basis, and we further define
$s'=(p_1+p_2)^2$, $t'=(p_b-p_2)^2$, $u'=(p_a-p_2)^2$. All these processes have the same soft anomalous dimension matrix \cite{MFNK}. We have at one loop
\begin{eqnarray}
&& {\Gamma}_{S \, 11}^{(1) \, bq \to tq'H}=
C_F \left[\ln\left(\frac{t'(t-m_t^2)}{m_t s^{3/2}}\right)-\frac{1}{2}\right],
\nonumber \\ &&
{\Gamma}_{S \, 12}^{(1) \, bq \to tq'H}=\frac{C_F}{2N_c} \ln\left(\frac{u'(u-m_t^2)}{s(s'-m_t^2)}\right),
\quad
{\Gamma}_{S \, 21}^{(1) \, bq \to tq'H}= \ln\!\left(\frac{u'(u-m_t^2)}{s(s'-m_t^2)}\right) \, ,
\nonumber \\ &&
{\Gamma}_{S \, 22}^{(1) \, bq \to tq'H}= C_F \left[\ln\left(\frac{t'(t-m_t^2)}{m_t s^{3/2}}\right)-\frac{1}{2}\right]
-\frac{1}{N_c}\ln\left(\frac{u'(u-m_t^2)}{s(s'-m_t^2)}\right)
+\frac{N_c}{2}\ln\left(\frac{u'(u-m_t^2)}{t'(t-m_t^2)}\right) \, ,
\nonumber
\end{eqnarray}
at two loops
\begin{eqnarray}
&& \Gamma_{S\, 11}^{(2) \, bq \to tq'H}= K_2 \, \Gamma_{S\, 11}^{(1) \, bq \to tq'H}
+\frac{1}{4} C_F C_A (1-\zeta_3) \, ,
\quad
\Gamma_{S\, 12}^{(2) \, bq \to tq'H}= K_2 \, \Gamma_{S\, 12}^{(1) \, bq \to tq'H} \, ,
\nonumber \\ &&
\Gamma_{S\, 21}^{(2) \, bq \to tq'H}= K_2 \, \Gamma_{S\, 21}^{(1) \, bq \to tq'H} \, ,
\quad
\Gamma_{S\, 22}^{(2) \, bq \to tq'H}= K_2 \, \Gamma_{S\, 22}^{(1) \, bq \to tq'H}
+\frac{1}{4} C_F C_A (1-\zeta_3) \, ,
\nonumber
\end{eqnarray}
and at three loops
\begin{eqnarray}
&& \hspace{-6mm} \Gamma_{S \, 11}^{(3) bq \to tq'H}= K_3 \, \Gamma_{S \, 11}^{(1) bq \to tq'H}
+ \frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right) ,
\nonumber \\ && \hspace{-6mm}
\Gamma_{S \, 12}^{(3) bq \to tq'H}= K_3 \, \Gamma_{S \, 12}^{(1) bq \to tq'H} + X_{12}^{(3) bq \to tq'H} \, , \quad \quad
\Gamma_{S \, 21}^{(3) bq \to tq'H}= K_3 \, \Gamma_{S \, 21}^{(1) bq \to tq'H} + X_{21}^{(3) bq \to tq'H} \, ,
\nonumber \\ && \hspace{-6mm}
\Gamma_{S \, 22}^{(3) bq \to tq'H}= K_3 \, \Gamma_{S \, 22}^{(1) bq \to tq'H}
+ \frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right)
\nonumber \\ && \hspace{15mm}
{}+ X_{22}^{(3) bq \to tq'H} \, ,
\nonumber
\end{eqnarray}
where the $X_{ij}^{(3) bq \to tq'H}$ denote unknown terms in the last three matrix elements which, however, do not contribute to the soft-gluon corrections at N$^3$LO.
We next consider the processes $q{\bar q'} \to t {\bar b} H$ as well as $q{\bar q'} \to t{\bar b}Z$, $q{\bar q'} \to t{\bar b}\gamma$, $q{\bar q} \to t{\bar b}W^-$, $q{\bar q'} \to t{\bar q''}W^+$, which all have the same soft anomalous dimension matrix \cite{MFNK}, and we use an $s$-channel singlet-octet color basis. We have at one loop
\begin{eqnarray}
&& \Gamma_{S \, 11}^{(1) \, q{\bar q'} \to t{\bar b}H}=C_F \left[\ln\!\left(\frac{s'-m_t^2}{m_t\sqrt{s}}\right)-\frac{1}{2}\right],
\nonumber \\ &&
\Gamma_{S \, 12}^{(1) \, q{\bar q'} \to t{\bar b}H}=\frac{C_F}{2N_c} \ln\left(\frac{t'(t-m_t^2)}{u'(u-m_t^2)}\right), \quad \quad
\Gamma_{S \, 21}^{(1) \, q{\bar q'} \to t{\bar b}H }= \ln\!\left(\frac{t'(t-m_t^2)}{u'(u-m_t^2)}\right) \, ,
\nonumber \\ &&
\Gamma_{S \, 22}^{(1) \, q{\bar q'} \to t{\bar b}H}=C_F \left[\ln\left(\frac{s'-m_t^2}{m_t \sqrt{s}}\right)-\frac{1}{2}\right]
-\frac{1}{N_c}\ln\left(\frac{t'(t-m_t^2)}{u'(u-m_t^2)}\right)
+\frac{N_c}{2} \ln\left(\frac{t'(t-m_t^2)}{s(s'-m_t^2)}\right) \, ,
\nonumber
\end{eqnarray}
at two loops
\begin{eqnarray}
&& \Gamma_{S\, 11}^{(2) \, q{\bar q'} \to t{\bar b}H}= K_2 \, \Gamma_{S\, 11}^{(1) \, q{\bar q'} \to t{\bar b}H}+\frac{1}{4} C_F C_A (1-\zeta_3) \, ,
\quad
\Gamma_{S\, 12}^{(2) \, q{\bar q'} \to t{\bar b}H}= K_2 \, \Gamma_{S\, 12}^{(1) \, q{\bar q'} \to t{\bar b}H} \, ,
\nonumber \\ &&
\Gamma_{S\, 21}^{(2) \, q{\bar q'} \to t{\bar b}H}= K_2 \, \Gamma_{S\, 21}^{(1) \, q{\bar q'} \to t{\bar b}H} \, ,
\quad
\Gamma_{S\, 22}^{(2) \, q{\bar q'} \to t{\bar b}H}= K_2 \, \Gamma_{S\, 22}^{(1) \, q{\bar q'} \to t{\bar b}H}+\frac{1}{4} C_F C_A (1-\zeta_3) \, ,
\nonumber
\end{eqnarray}
and at three loops
\begin{eqnarray}
&& \hspace{-6mm} \Gamma_{S \, 11}^{(3) q{\bar q}' \to t{\bar b}H}= K_3 \, \Gamma_{S \, 11}^{(1) q{\bar q}' \to t{\bar b}H}
+\frac{1}{2} K_2 C_F C_A (1-\zeta_3)
+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right) ,
\nonumber \\ && \hspace{-6mm}
\Gamma_{S \, 12}^{(3) q{\bar q}' \to t{\bar b}H}=K_3 \, \Gamma_{S \, 12}^{(1) q{\bar q}' \to t{\bar b}H} + X_{12}^{(3) q{\bar q}' \to t{\bar b}H} \, , \hspace{8mm}
\Gamma_{S \, 21}^{(3) q{\bar q}' \to t{\bar b}H}= K_3 \, \Gamma_{S \, 21}^{(1) q{\bar q}' \to t{\bar b}H} + X_{21}^{(3) q{\bar q}' \to t{\bar b}H} \, ,
\nonumber \\ && \hspace{-6mm}
\Gamma_{S \, 22}^{(3) q{\bar q}' \to t{\bar b}H}= K_3 \, \Gamma_{S \, 22}^{(1) q{\bar q}' \to t{\bar b}H}
+\frac{1}{2} K_2 C_F C_A (1-\zeta_3)+C_F C_A^2\left(-\frac{1}{4}+\frac{3}{8}\zeta_2-\frac{\zeta_3}{8}-\frac{3}{8}\zeta_2 \zeta_3+\frac{9}{16} \zeta_5\right)
\nonumber \\ && \hspace{15mm}
{}+ X_{22}^{(3) q{\bar q}' \to t{\bar b}H} \, ,
\nonumber
\end{eqnarray}
where the $X_{ij}^{(3) q{\bar q}' \to t{\bar b}H}$ denote unknown terms in the last three matrix elements which, however, do not contribute to the soft-gluon corrections at N$^3$LO.
\section{Conclusion}
Soft anomalous dimensions are fundamental in describing soft-gluon emission in QCD processes. In this contribution, I presented results for soft anomalous dimensions for many processes through three loops.
These results are needed in calculations of high-order corrections.
\paragraph{Funding information}
This material is based upon work supported by the National Science Foundation under Grant No. PHY 2112025.
|
\section{Introduction}
\label{sec:intro}
\IEEEPARstart{W}ith the increasingly powerful intelligence capability, future wireless networks will transform from ``connected things'' to ``connected intelligence'' \cite{Saad2020AVision,Letaief2019The,Xiao2020Toward}. To adapt this new trend, the intelligent devices in future wireless networks should not only be able to accomplish complex smart applications, but also be utilized for more efficient, robust communications.
Fortunately, this communication paradigm shift coincides with the Shanno and Weaver's envisions for future communications. It is known that conventional wireless networks are based on Shannon's classical information theory. Reviewing Shanno and Weaver's seminal work \cite{Weaver1949Recent}, we can find that communications can be categorized into three levels 1) How accurately can the symbols of communication be transmitted? (The technical problem.); 2) How precisely do the transmitted symbols convey the desired meaning? (The semantic problem.); 3 How effectively does the received meaning affect conduct in the desired way? (The effectiveness problem.).
The existing wireless networks limit to the technical level, which aim at transmitting signals as massively and accurately as possible.
In contrast, communications at the semantic level intend to transmit source semantic information.
Here, semantic information can be deemed as the desired ``meaning'' of the source.
For instance, word ``four'' and ``4'' have the same the meaning at the semantic level. However, their representations in bit sequences at the technical level can be totally different. Therefore, semantic communications need to characterize the meanings behind the bits.
Beyond text, the images also contain semantic information.
Fig. \ref{fig:example} shows an example of semantic communications of images, and the target is to detect the dog in the image. Communications at the semantic level only transmit the semantic information related to the dog.
This is because the information is sufficient to represent the meaning of the source, i.e. whether there is a dog in the image.
All the other information related to the background or the cat do not affect the decision-making process, and thus it is not transmitted. However, communications at the technical level treat all information equally important, even though most parts of it do not affect the accomplishment of the artificial intelligence (AI) task. Therefore, communications at the semantic level can significantly improve the communication efficiency \footnote{As stated in Shanno and Weaver's seminal work \cite{Weaver1949Recent}, the effectiveness problem is closely interrelated with the semantic problem, and overlaps it in a rather vague way.
Following this cognition, we generally term the communication method considered in this work as semantic communications. However, it definitely involves the problem in effectiveness level.}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.9\linewidth]{Semantic_communication_example.png}
\end{center}
\caption{An illustration example of the communications at the semantic level and the technical level.}
\label{fig:example}
\end{figure}
Even though having promising advantages, there are still some key challenges in semantic communications including the representation and mathematical quantification of semantic information, and the implementation of the semantic communications. The authors in \cite{Floridi2004Outline} proposed a quantitative theory of semantic information based on the truth-values and semantic discrepancy with respect to a given situation.
The authors in \cite{baoj2011towards} proposed a general model of semantic communications, and the lossless semantic data compression was discussed.
Though the initial theoretical studies such as \cite{Floridi2004Outline,baoj2011towards} shed some light on semantic communications, it is still challenging to implement semantic communications in practice.
In recent years, with the development of machine learning, deep neural networks based semantic communications have been attracting increasing attentions.
The authors in \cite{Xie2021Deep} proposed a semantic communication system for text transmission based on deep learning.
Based on \cite{Xie2021Deep}, the authors in \cite{Xie2021ALite} proposed a lite distributed semantic communication system for internet of thing (IoT) applications.
Even though the existing works are interesting, none of them considered the relations between the AI tasks and the semantic information. Actually, the semantic information is closely related to the target AI task.
For instance, from Fig. \ref{fig:example}, if the target of the transmission changes to detect the cat, the information related to the dog is no longer necessary. Therefore, semantic communication should take AI tasks into considerations.
The main contribution of this paper is a semantic communication paradigm with AI tasks (SC-AIT).
First, we propose the semantic communication architecture for SC-AIT, in which the functions of the three communication levels are clearly identified.
In particular, the effectiveness level mainly concerns with the source information, the desired conduct of the AI tasks, and forms a knowledge base (KB). The semantic level conducts semantic information extraction and coding based on KB and the neural networks. Then, following the design principles of the architecture, we accomplish SC-AIT for a classification task to show its feasibility. Next, an experimental prototype is established to implement SC-AIT for a surface defect detection task, and experimental results verify the efficiency of SC-AIT. We also summarize some open issues and challenges for future semantic communications. Finally, key points of this paper are concluded.
\vspace{-0.3cm}
\section{Architecture and Implementations}
This section elaborates the architecture and implementations of SC-AIT. There are some priori art on the architecture design of semantic communications \cite{Weaver1949Recent,baoj2011towards}. Different from these studies, the proposed architecture introduces AI task block, and investigates the underlaying relations between AI tasks and semantic information to instruct semantic communications.
This is necessary since not all semantic information is required by AI tasks. However, this is also challenging due to the complex relations between the AI tasks and the source information. To circumvent the challenge, this work employs neural networks to achieve the design target. In particular, a study case of SC-AIT is implemented for classification tasks based on an end-to-end neural network. The large number of model parameters in neural networks are utilized to characterize the complex relations between the AI tasks and the source information so as to instruct semantic communications.
\subsection{Architecture Design}
In the first subsection, we introduce the proposed architecture and mainly analyze the necessity and the function of AI task block for the overall semantic communications. The implementation details of the architecture will be elaborated in the next subsection.
As shown in Fig.~\ref{fig:architecture_en}, the semantic communication architecture consists of the effectiveness level, the semantic level and the technical level \cite{Weaver1949Recent}, and each layer is closely interrelated to each other.
At the effectiveness level, there is a source block, which contains the information to be transmitted.
In this work, we assume that the target of the communications is to accomplish a certain AI task that is known by both transmitter and receiver \cite{Letaief2019The}.
The contained information, the AI tasks, and their inherent relations constitute a KB, which can be used to store the above information and instruct the following semantic communications.
Since for different AI tasks, information can have different degrees of importance, it is necessary to introduce AI task in KB.
For instance, as shown in Fig. 1, once the AI task becomes detecting the cat, the information related to dog is no longer necessary.
Moreover, no matter whether the target is the dog or cat, the background information is not needed.
Therefore, the effectiveness level needs to quantify the degrees of importance of semantic information for the accomplishment of AI tasks, so as to decide the parts of semantic information that should be transmitted. At the receiver, there is also a KB to store the knowledge including the related semantic information and AI tasks.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=1\linewidth]{architecture_en_0617.png}
\end{center}
\caption{The proposed semantic communications architecture.}
\label{fig:architecture_en}
\end{figure}
At the semantic level, the semantic information of the source node first needs to be extracted and expressed in a certain form.
After the semantic extraction and expression block, all the semantic information will be delivered to the semantic encoder.
Similar to the conventional source coding, the purpose of the semantic encoder is to reduce source information redundancy.
In particular, SC-AIT proposes to reduce semantic information redundancy according to the contribution to the accomplishment of AI tasks.
The information that is irrelevant to the success of AI tasks can be discarded, and only the useful semantic information is encoded at this block.
Note that the KB has already quantify the relations between the AI tasks and the semantic information, and thus this process can be conducted according to knowledge stored in KB.
At the receiver of the semantic level, a semantic decoder is employed to recover the semantic information.
In addition, there is also a KB at the receiver, which can interact with the source KB via a shared KB at the semantic level.
When the KB at one side alters, the KB at the other side can synchronize the knowledge.
In particular, it is recognized that the destination should be able to increase its level of knowledge by the received message. After the redundancy reduction at the semantic level, the information will be transmitted at the technical level via physical channel.
Note that the local KB can be stored at the transmitter and the receiver, while the shared KB can either be stored in an authoritative third party or just be a virtual KB to synchronize the KBs at the two sides.
At the technical level, the structure is mainly the same as the existing communication systems. One exception is that the source coding is accomplished at the semantic level. This is because the source coding is finished at the semantic encoder block, where the most relevant semantic information is extracted. Compared with the conventional source coding at the technical level, which aims at compressing source information in terms of the amount of the bits, the semantic encoder aims at compressing source information in terms of the content, thus transmitting the ``meaning'' of the source information. Then, channel encoder is employed to enhance the transmission reliability, and the bit stream is modulated for transmission.
\subsection{Implementations}
In this subsection, following the design principles of the proposed architecture, a study case of SC-AIT is implemented for image classification tasks based on convolutional neural networks (CNN). In particular, this subsection elaborates how to establish KB and how to investigate the underlaying relations between AI tasks and semantic information to instruct semantic communications.
Image classification is a fundamental AI task that has extensive applications such as surface defect detection in industrial internet, MNIST.
In the considered classification tasks, the source typically contains a number of images, and they will be transmitted to the receiver for analysis.
This is particularly suitable for IoT applications, where the transmitter sensors may have limited computation capability for accurate inferring.
Following the architecture shown in Fig. \ref{fig:architecture_en}, the source image will first be delivered to semantic extraction and expression block to represent the information of the source image. This process can be achieved by a CNN due to its powerful representation learning capability. In particular, a CNN network can extract the key features of the source image and express them into a series of feature maps.
The extracted feature maps indicate different aspects of features of the source images, such as the color, the texture.
Since different AI tasks require different features of the images, it is necessary to further refine the most important features from the extracted features.
This is finished at the semantic encoder block, which further refine the semantic information that is closely related to the AI task.
As shown in Fig. \ref{fig:architecture_en}, the semantic encoder block is instructed by source KB. Therefore, before communications, KBs need to be established in advance.
Generally, it is particularly challenging to establish a general KB due to the diversity and the complex relations between the semantic information and the AI tasks.
Fortunately, CNN contains a large number of model parameters. These model parameters are trained not only to minimize a certain loss function for the best AI task performance, but also to determine the best forms of features representing the original image, i.e. semantic information of the source.
Therefore, the model parameters in CNN can naturally characterize the inherent relations between the semantic information and the AI tasks.
In particular, SC-AIT uses the gradients of the output of the CNN with respect to feature maps as the importance weights of the feature map to different classes. Note that the output of the CNN directly decides the class of the image, while the features maps can be deemed as the extracted semantic information in CNN. The gradient actually implies the degrees of importance of semantic information with respect to different classes.
In this way, KB is established, which stores the importance weights of all feature maps for each class. Thus, we quantify the relations between the feature maps (i.e. semantic information) and each class in KBs. Next, the semantic encoder can be conducted based on the established source KB. In particular, we can always transmit the important feature maps by jointly considering the bandwidth and the performance requirements.
Since SC-AIT is achieved in an end-to-end fashion, the receiver KB also requires the knowledge to accomplish the task, and the system needs to synchronize the knowledge once AI tasks or the semantic information changes.
Finally, the compressed semantic information is delivered to channel encoder and modulation for robust transmission.
At the receiver, the destination first demodulates and channel decodes the received message. Fully-connected layers at the receiver act as the semantic decoder. By selecting the class corresponding to the maximum value in the output vector, the destination is able to determine which class the transmitted image belongs to. In this way, SC-AIT is completed for classification tasks.
\vspace{-0.3cm}
\section{Prototype Establishment and Performance Evaluation}
In this section, we establish a semantic communication prototype to implement SC-AIT. In particular, the surface defect detection of hot-rolled steel strip is considered, which is an image classification task that has been extensively used in industrial IoT to replace subjective and repetitive process of manual inspection.
Then, the performance of SC-AIT is evaluated and compared with baseline schemes based on conventional JPEG transmission and normal semantic communications (SC) without considering AI tasks.
\vspace{-0.3cm}
\subsection{Proof-Of-Concept Prototype}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.9\linewidth,height=0.5\textheight]{system.png}
\end{center}
\caption{Semantic communication prototype for surface defect detection tasks.}
\label{fig:system}
\end{figure}
Fig.~\ref{fig:system} illustrates the established prototype. As shown in Fig.~\ref{fig:system}, we assume the transmitter is a sensor used to inspect surface defects. In particular, the transmitter is composed of a camera, a universal software radio peripheral (USRP), transmitter antennas.
The camera acts as a source to capture the images of the hot-rolled steel strip for surface defects. Then, the images from the camera are delivered to the sensor. Here, sensor acts as an intelligent node used to inspect surface defects, and its function is achieved by a computer in our prototype. In particular, the sensor first extracts and codes the semantic information of the images according to the established source KB, and then modulates the signals at the technical level. The USRP at the transmitter side (USRP A) reads the byte streams from the User Datagram Protocol (UDP) port of the sensor via Peripheral Component Interconnect express (PCIe) bus. USRP is software-defined radios, which conducts digital to analogue conversions and adapts signals to suitable forms for transmission. The LabVIEW software is used to control the transmission parameters, such as the frequency of the carrier and the transmission gains, as shown in Fig.~\ref{fig:system}. Finally, the antennas will transmit the signals.
The receiver is composed of receiver antennas, a USRP, and an edge server acted by a computer at the receiver side. In particular, the receiver antennas first deliver the received signal to the USRP at the receiver side (USRP B). In USRP B, the signals representing the semantic information of the image are demodulated and decoded into byte streams. Then, the byte streams are delivered to edge server's UDP port via PCIe bus. The edge node then decodes the semantic information and outputs the class that semantic information indicates.
The key system parameters are summarized in Fig. \ref{fig:system}. The dataset is NEU surface defect dataset\footnote{http://faculty.neu.edu.cn/yunhyan/NEU\_surface\_defect\_database.html}, which is suitable for industrial Internet scenarios. As shown in Fig.~\ref{fig:system}, six kinds of typical surface defects of the hot-rolled steel strip are collected in the dataset, i.e., rolled-in scale, patches, crazing, pitted surface, inclusion and scratches. The dataset includes 1800 grayscale images (300 samples for each kind), and the original resolution of each image is $200 \times 200$ pixels.
\subsection{Performance Evaluations}
This subsection evaluates the performance of the proposed SC-AIT. A conventional scheme that directly transmits images in JPEG forms for classifications is employed as the baseline scheme. In addition, the proposed communication method under different semantic compression ratios (CR) are considered. For instance, $CR=0\%$ indicates that all the feature maps are transmitted without compression, while $CR=30\%$ indicates 30\% of the feature maps are discarded, while the remaining 70\% feature maps are transmitted.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.8\linewidth]{ACvsSNR.png}
\end{center}
\caption{Simulation results of classification accuracy under different SNRs.}
\label{fig:simulation2}
\end{figure}
Fig.~\ref{fig:simulation2} first compares the classification accuracy of SC-AIT and the conventional JPEG transmission scheme under different signal to noise ratios (SNRs). As shown in Fig.~\ref{fig:simulation2}, compared with JPEG compressed scheme, the semantic communication can always obtain better performance.
In particular, SC-AIT with $CR=98\%$ can still harvest more than 40\% accuracy gains even at an aggressive compression ratio and 10 dB SNR when compared with the JPEG transmission scheme, since the top 2\% most relevant semantic information is transmitted.
We can also observe that the performance gain is more significant at low SNR regions. This is because the semantic communication system conducts in an end-to-end fashion and takes channel state information into the training. In contrast, the conventional JPEG transmission scheme has much worse bit error rate at low SNRs, leading to worse classification performance. Moreover, when $CR = 87\%$, there is almost no loss of classification accuracy when SNR varies. This indicates that SC-AIT can significantly reduce the redundant information without affecting the task performance.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.8\linewidth]{ACvsbpp.png}
\end{center}
\caption{Experimental results of classification accuracy with different available bandwidths.}
\label{fig:simulation1}
\end{figure}
Fig.~\ref{fig:simulation1} compares the performance of the proposed SC-AIT with a conventional JPEG transmission scheme and a normal semantic communication method without considering AI tasks (SC) under different bits per pixel (bpp). Here, different bpps indicates different bandwidth requirements of the communication methods. The lower the bpp is, the more extreme the source has to be compressed, and thus less bandwidth is required. In addition, since normal SC does not consider AI task, it tries to extract and transmit all the features. When the bpp constraint does not support transmitting all the feature maps, SC will randomly transmits parts of feature maps, as it does not consider the relations between the AI tasks and feature maps. In contrast, the proposed SC-AIT always transmits the most important parts of feature maps to satisfy different bpp constraints.
As we can observe from Fig.~\ref{fig:simulation1}, both SC-AIT and SC can achieve much lower bpps when compared with conventional JPEG transmission scheme. In particular, both the schemes can achieve convergent performance at $10^{-2}$ level of bpp.
Moreover, we can also observe that SC-AIT can achieve more significant performance gains when bpp is extremely low (lower than $10^{-3}$). This is because SC-AIT chooses the most important feature maps for transmission, which are the semantic information closely related to the success of AI tasks.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.8\linewidth]{Delayvsschemes.png}
\end{center}
\caption{Experimental results of complexity in terms of runtime.}
\label{fig:delay}
\end{figure}
Fig.~\ref{fig:delay} compares the complexity of SC-AIT and the baseline schemes in terms of the system runtime. In particular, the runtime includes transmission delay and process delay. The computers used for experiments are equipped with Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz, 100G RAM, and Tesla V100 32G graphics card.
As shown in Fig.~\ref{fig:delay}, the proposed SC-AIT have both lower process delay and transmission delay compared with the conventional JPEG transmissions.
The total delay of SC-AIT is only 70\% of the conventional JPEG transmission scheme.
This is because the SC-AIT uses the feature extraction network to extract the semantic information of the image, which has lower complexity than that of the conventional, complex source coding. The total complexity is thus reduced. When the feature maps are further compressed by 87\%, the transmission delay is further reduced by 21\%.
This is because the transmitted information is further compressed. Therefore, we can conclude that the semantic communication system has low complexity and is particularly suitable for delay-sensitive applications and systems with limited computation capability.
\vspace{-0.2cm}
\section{Challenges and Opportunities}
The above work can serve as basis for semantic communications that takes AI tasks into consideration. This section further summarizes some key challenges remaining in the research field, and sheds some light on future research opportunities.
\subsection{Mathematical Fundamentals of Semantic Communications}
It is known that conventional wireless networks are based on Shannon's classical information theory, in which basic definitions such the amount of information and channel capacity are rigorously derived. Similarly, semantic communications also require such basic definitions to support. There are a series of key issues to be determined: How to quantify semantic information? What is the lossless semantic compression? What is the best representation of semantic information?
These basic questions are essential to derive the theoretical semantic communication capacity bound. There are several initial works that discussed this topic. For instance, in \cite{baoj2011towards}, the authors analyzed the lossless semantic compression and establish theoretical bounds in lossless semantic data compression. However, the work has non-trivial assumptions such as adopting the logic-based approaches to capture semantic in human communication, which can significantly limit the application of the bound and utility communication in the future. In addition, general theories such as the information bottleneck theory may be applied. The information bottleneck theory extends the rate-distortion theory of data compression, and is able to explain the deep learning process. In particular, the information bottleneck theory can be used to express the source information into a briefest form, which may used to instruct the semantic compression and communications. In conclusion, the mathematical fundamentals of semantic communications is the basis of semantic communications. It is vital to quantify, express and derive the lossless bound of semantic communications.
\subsection{Neural Network Structure Optimization}
In our implementations of SC-AIT, we extract the semantic information based on a given network. Actually, the neural network structure is also closely related to the extracted semantic information.
For instance, in \cite{net2vec2018} , the Net2Vec framework is proposed, which maps semantic information to the corresponding feature filter space, and analyzes how semantic information is encoded by CNN filters.
In \cite{gradcam2017}, Grad-CAM, a class-discriminative localization technique is proposed, which analyzes the importance of feature maps to concepts, and generates visual explanations for the CNN network architecture.
Both works verify that the semantic features extracted by CNN is closely related to the structure of the CNN.
Therefore, the neural network structure needs to be optimized according to the information in the KB, so that CNN can represent semantic information more accurately.
Furthermore, the loss function of a neural network can also affect the extracted semantic information.
A appropriate loss function can make feature extracted have better task performance \cite{perceptualloss2016}.
Therefore, the neural network structure and the loss function needs to be optimized to represent semantic information more well while at moderate cost.
\subsection{C4 Resource Allocation}
Semantic communications require seamless integration of communication and intelligence. Therefore, besides communication resources, computation resource, caching resource, and control strategies (C4) are also deeply involved in semantic communications. Moreover, the C4 resources are closely interacted\cite{3C2017}, and all of them can significantly affect the final AI task performance. Therefore, the C4 resource allocation plays a key role in semantic allocation. It is challenging to consider the joint allocation of 4C resources according to the mutual constraint and restriction relationship of 4C resources. Besides, the resource allocation rule may also alter in semantic communications. In conventional wireless networks, the resources are typically allocated according to the QoS or QoE of users. In contrast, in semantic communications, the C4 resources need to be allocated according to the importance of the semantic information. Therefore, the resource allocation in semantic communications also requires new rules or principles.
\section{Conclusion}
This work has proposed a semantic communication method that takes AI tasks into consideration. The architecture of the proposed SC-AIT has been introduced, which is compatible with the Shanno and Weaver's envision for semantic communications. The function of each block in the proposed architecture has also been introduced in detail. Then, we have elaborated the proposed SC-AIT for classification tasks. A prototype is established for SC-AIT with surface defect detection tasks. Experimental results have verified the efficiency of the proposed semantic communications. We have also shed some light for the future researchers.
|
\section{Introduction}
The NASA {\it Kepler} main mission has shown the power of using asteroseismology to precisely characterize the properties of solar-like stars \citep[e.g.][]{2017ApJ...835..173S,2017ApJS..233...23S}. Based on the latest catalogs of \kep\ stellar properties \citep{2014ApJS..211....2H,2017ApJS..229...30M,2018ApJ...866...99B,2020AJ....159..280B}, more than 125,000 stars are supposed to be main-sequence solar-like stars according to their effective temperatures, surface gravities as well as the {\it Gaia} DR2 observations \citep[e.g.][]{2018A&A...616A...8A,2018A&A...616A...1G}, which were included in the latest \kep\ catalog. The turbulence of the outer layers of those stars excite oscillations known as solar-like oscillations \citep[e.g.][]{2019LRSP...16....4G}. Given the radius, mass, and gravity of these stars, the modes are expected to be above 300\,$\mu$Hz, which means that we need to observe them in short cadence (sampling of $\sim$\,58.85\,s) in order to be able to detect the acoustic modes.
\noindent During the survey phase of the \kep\ mission, around 2,600 stars were observed between May 2009 and March 2010 for a duration of approximately one month each and in a short-cadence mode. \citet{2011Sci...332..213C} performed a seismic analysis of the data available at that time and were able to detect solar-like oscillations in more than 500 stars. {\color {black} This led to the seismic characterization of that sample, providing their masses, radii, and ages \citep{2017ApJS..233...23S}. More recently, \citet[][]{2020FrASS...7...85B} analyzed hot main-sequence stars with $T_{\rm eff} >$\,6,000\,K in order to study the location of the $\delta$ Scuti/$\gamma$ Doradus instability strip and detected solar-like oscillations in 70 new \kep\, targets.}
\noindent Later, for the close-out of the nominal \kep\ mission in 2016, a new data release \citep[DR25,][]{Thompson2016} \footnote{\url{http://archive.stsci.edu/kepler/release_notes/release_notes25/KSCI-19065-002DRN25.pdf}} was done by the \kep\ Science Office. The calibration of the data was improved, in particular how some instrumental effects were corrected, which had an impact on about 50\% of the short-cadence data.
\noindent While the effect of the calibration should be negligible, the \kep\ Science Office could not assess the impact on the asteroseismic studies. Using the newly calibrated data of DR25, \citet{2017EPJWC.16001007S} re-analyzed the new lightcurves for 18 solar-analogs. In particular, the authors re-computed their global seismic parameters such as the mean large frequency separation that scales with the root-mean-square of the mean stellar density \citep{1995A&A...293...87K} and the frequency of maximum oscillation power that is related to the surface gravity of the star \citep{1991ApJ...371..396B}. Comparing the new results with those obtained with the previous data release ({ DR23}), \citet{2017EPJWC.16001007S} found that the impact on the measurement of the mean large frequency separation, the frequency of maximum oscillation power, and the background parameters was negligible for the stars with high signal-to-noise ratio (SNR). { Nevertheless,} the improvement of the lightcurves could still have an impact on the detection of the modes. With this in mind, we conduct the current work to determine the impact of the improved DR25 on the mode detection for \kep\ solar-like stars
{ \noindent One would hope that the newly calibrated DR25 data would lead to a higher number of detections of solar-like oscillations. However, that is without taking into account the surface magnetism of stars. Indeed, for the Sun and several solar-like stars, it has been shown that the amplitude of the modes is anti-correlated with magnetic activity \citep[e.g.][]{2010Sci...329.1032G,2015MNRAS.454.4120H,2017A&A...598A..77K,2018ApJS..237...17S}. As a consequence, the non detection of modes in the remaining stars can partially be explained by the high magnetic activity level of the stars \citep{2011ApJ...732L...5C} but other factors, such as metallicity or binarity, can also be at play \citep{2019FrASS...6...46M}.}
\noindent { In order to learn about the magnetic activity of stars with photometric data, we can use rotational modulation that results from the presence of active regions on the stellar surface. Additionally, the study of stellar rotation can provide key information for the understanding of angular momentum transport { \citep[e.g.][]{2019ARA&A..57...35A,2019ApJ...872..128V,2020AJ....160...90A,2020ApJ...904..140C,2021ApJ...912..127S}}. Several studies have been led with \kep\ and K2 data to measure surface rotation periods in a large sample of stars \citep[e.g.][]{2014ApJS..211...24M,2019ApJS..244...21S,2020A&A...635A..43R,2021ApJ...913...70G,2021ApJS..255...17S}. Furthermore, combined with precise ages of stars, such as those obtained with asteroseismology or for clusters, it is possible to improve rotation-age relationships \citep[e.g.][]{2003ApJ...586..464B,2008ApJ...687.1264M,2015MNRAS.450.1787A,2016Natur.529..181V,2021NatAs.tmp...71H,2021arXiv210101183G}.}
\noindent In Section 2, we describe the data that are used here and the procedure to calibrate the lightcurves to optimize them for asteroseismic studies. Then we explain how the search for the acoustic modes was done in Section 3 and consolidate the sample along with the stellar atmospheric parameters. Section 4 presents the study of the convective background parameters as well as the maximum amplitude of the modes and their correlation with the global seismic parameters. In Section 5, we compute the masses and radii based on the seismic scaling relations and look at the rotation and magnetism of this new sample compared to the previously known sample of solar-like stars with detections of acoustic modes. Finally, we finish with the discussion on the stellar parameters of the new seismic detections and provide a summary of this work (Sections 6 and 7).
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=15cm]{Figures/comp_PSD_DR24_DR25.pdf}
\caption{Comparison of the Power Spectrum Density of KIC~4484238 using the DR24 data (black curve) and the DR25 data (red curve).}
\label{PSD_DR24_25}
\end{center}
\end{figure*}
\section{Data description}\label{sec:Data}
Here we processed both the short-cadence (SC) data (sampling of 58.85s) and the long-cadence (LC) data (sampling of $\sim$\,29.42min). The SC data were used for the asteroseismic analysis whereas the LC data were used for the study of the rotation and magnetic activity as for those phenomena a high cadence is not necessary.
\noindent For the SC, the lightcurves used in this analysis are based on the last \kep\ data release DR25, where the \kep\ Science Office improved the calibration of the data. Indeed, as mentioned earlier, since the beginning of the mission, the short-cadence lightcurves were produced by correcting for the effects of {\it Kepler} shutterless readout. This scrambled the collateral smear and affected half of the short-cadence data. Hence, Presearch Data Conditioning - Simple Aperture Photometry \citep[PDC-SAP;][]{2010ApJ...713L..87J} lightcurves were obtained from the Mikulski Archive for Space Telescopes (MAST)\footnote{\url {https://mast.stsci.edu/portal/Mashup/Clients/Mast/Portal.html}}.
\noindent The LC data were also obtained from the MAST but the calibrated lightcurves produced by the \kep\ Science Office are not necessarily optimized for rotation studies as some filters are applied. { Therefore, from the target pixel files, we first selected a larger aperture. Starting by the pixel of the center of the star, adjacent pixels are added to the aperture if the average flux of the new pixel is decreasing (i.e., new added pixels should have less flux than the previous one in each direction). The addition of pixels to the aperture is stopped if: a) the mean flux of the new pixel increases, which suggests the presence of a nearby star; or b) the mean flux of the new pixel is less than a given threshold, which is established at 100 e-/s. This threshold allows us to maximize the flux from the target and yields larger apertures than those employed for the standard PDC-SAP ones.}
\noindent For both cadences, the obtained lightcurves were then calibrated for asteroseismology with the \kep\ Asteroseismic Data Analysis Calibration Software \citep[KADACS;][]{2011MNRAS.414L...6G}. We first removed outliers, corrected jumps, and removed instrumental trends. We also filled the gaps using the inpainting technique (with a multiscale discrete cosine transform) as described in \citet{2014A&A...568A..10G} and \citet{2015A&A...574A..18P}. This gap-filling technique allows us to reduce the noise at all frequencies, especially for lightcurves with rotation modulation. For the long-cadence lighcurves, we concatenated the different quarters and applied three high-pass filters with a cut-off period of 20, 55, and 80 days. Both short-cadence and long-cadence corrected lightcurves, called KEPSEISMIC, are available on the MAST\footnote{\url {https://archive.stsci.edu/prepds/kepseismic/}}.
\noindent In Figure~\ref{PSD_DR24_25}, we show the example of KIC~4484238, where the DR25 lightcurve leads to a significant improvement in the SNR. Indeed, while with the previous data release we could not detected any excess of power due to the oscillation modes, the DR25 power spectrum density (PSD) presents a bump around 2000\, $\mu$Hz.
\noindent Finally, for the long-cadence data, we also used the PDC-MAP \citep[Maximum A Posteriori,][]{Thompson_2013} data to which we applied the same KADACS correction software and gap filling technique to check the rotation period retrieved from the lightcurves and to make sure that there is no nearby star polluting the aperture.\\
\section{Looking for new detection of solar-like oscillations}\label{sec:detection}
\noindent Using the short-cadence lightcurves calibrated in Section~\ref{sec:Data}, we first performed a blind analysis of 2,572 lighcurves with the A2Z pipeline \citep{2010A&A...511A..46M}. This analysis provided global seismic parameters, in particular the mean large frequency separation, $\Delta \nu$, and the frequency of maximum power, $\nu_{\rm max}$. We then visually checked the stars for which the relation between $\Delta \nu$ and $\nu_{\rm max}$ agreed { within 20\% of the empirical relation derived from the {\it Kepler} data \citep{2011ApJ...743..143H}}. This led to a sample of 20 stars where the detection seemed to be clear. { In addition, following the procedure of \citet{2010A&A...511A..46M}, we found that the measurement of $\Delta \nu$ had} more than 95\% probability of being due to stellar signal.
\noindent We then visually checked the outputs of the A2Z pipeline as well as the power spectrum densities of the remaining stars, flagging some clear detections as well as some candidates. We also computed the FliPer metric \citep[Flicker in Power,][]{2018A&A...620A..38B}. { We used the sklearn.ensemble.RandomForest Regressor function \citep{2012arXiv1201.0490P} to train and combine predictions from a set of $200$ independent random trees. By combining the effective temperature of the star (\citet{2014ApJS..211....2H} and \citet[][hereafter M17]{2017ApJS..229...30M}) and FliPer granulation parameters \citep[see][for more details about the intrinsic methodology and the training of the supervised machine learning algorithm]{2018A&A...620A..38B}, the trained Random Forest \citep{2001MachL..45....5B} provides rough estimates of the frequency of maximum power. Based on the power contained in the power spectrum, this procedure flags stars that have a power spectrum consistent with the ones of solar-like stars.} This allowed us to select 20 additional stars where the blind run of A2Z agreed within 20\% of the FliPer predicted $\nu_{\rm max}$.
{\noindent For the remaining stars, we ran again the A2Z pipeline forcing $\nu_{\rm max}$ to the predicted value from the FliPer metric. For that subsample, we selected the stars where $\Delta \nu$ agreed within 20\% with the value expected from $\nu_{\rm max}$ based on the seismic scaling relations.
\noindent These steps led to a total sample of {\color {black} 105} stars with possible or confirmed detections. {\color {black} After cross-checking with the sample of new detections by \citet[][hereafter Ba20]{2020FrASS...7...85B}, we added 43 stars to the analysis that were not selected and were part of our original sample. Our sample of 148 targets includes 60 stars that are part of the new detections of Ba20.}
\noindent All the stars were also analyzed independently by the COR pipeline \citep{2009A&A...508..877M}, in a blind way first, and then in an iterative way consisting in using the seismic information provided by the $\nu_{\rm max}$ values derived by A2Z, in order to check for a possible agreement with this pipeline. We furthermore analyzed the sample with pySYD\footnote{https://github.com/ashleychontos/pySYD} { \citep{2021arXiv210800582C}}, an open-source python-based adaptation of the SYD pipeline \citep{2009CoAst.160...74H}.
\subsection{Consolidating surface gravity, effective temperature, and metallicity from literature}\label{sec:atmosph}
In order to better characterize the stellar parameters, we consolidated the atmospheric parameters for the full sample of {\color {black}148} stars by collecting values available in the literature. While all targets had $\log g$, $T_{\rm eff}$, and [Fe/H] available from the M17, many of those values were obtained from photometry, which could have large uncertainties. However, since then a few spectroscopic surveys have observed many of the targets in the \kep\ field.
\noindent { This is the case of the APOGEE survey \citep[Apache Point Observatory for Galactic Evolution Experiment,][]{gunn+2006,2019PASP..131e5001W}, which is part of the SDSS-IV \citep[Sloan Digital Sky Survey-IV][]{2017AJ....154...28B} and that provided high-resolution spectroscopy for more than 15,000 \kep\, targets. We used the DR16 \citep{2020ApJS..249....3A} for which the data were calibrated following \citet{2015AJ....150..148H}, \citet{2015AJ....150..173N}, and \citet{2016AJ....151..144G}. The majority of those stars} were red giants { and a few hundreds were dwarfs and subgiants as part of APOKASC \citep{2017ApJS..233...23S,2018ApJS..239...32P}}, a collaboration between APOGEE and the \kep\, Asteroseismic Science Consortium, but they also included several thousands of dwarfs without seismic detections for other science objectives.
\noindent Another large spectroscopic survey, the Large Sky Area Multi-ObjectFiber Spectroscopic Telescope \citep[LAMOST,][]{2012RAA....12..723Z,2015ApJS..220...19D,2018ApJS..238...30Z}, also observed the \kep\, field with low-resolution spectroscopy.
\noindent Finally, the \emph{Gaia} mission is also providing precise parallaxes for a large sample of the \kep\, field stars. \citet[][hereafter B20]{2020AJ....159..280B} derived new stellar parameters for the \kep\,-{\it Gaia} targets by fitting isochrones using the \emph{Gaia} DR2 { \citep{2018A&A...616A...1G}.}
\noindent We used the following prioritization to build our list of spectroscopic parameters for the {\color {black} 148} targets previously selected. When available we first selected the stellar parameters obtained with high-resolution spectroscopy from APOGEE. Then we completed the parameters with low-resolution spectroscopy from LAMOST DR5 \citep{2018MNRAS.477.4641R}. Finally we took the stellar parameters from the latest \kep\,-{\emph{Gaia} stellar parameters catalog. {\color {black} We found 102 stars with APOGEE spectra, 33 stars with LAMOST stellar parameters, and 13 stars with \emph{Gaia} parameters.}
\subsection{Finalizing the sample of confirmed detections}\label{sec:final}
The comparison of the results obtained by the three seismic pipelines, A2Z, COR, and pySYD allowed us to build a list of confirmed detections constituted of {\color {black} 99} stars where at least two pipelines agreed within 10\% in $\nu_{\rm max}$ as it has already been demonstrated that such scatter between pipelines is reasonable in particular for low SNR cases \citep[e.g.][]{2020ApJS..251...23Z}. In the remainder of the paper, we will refer to those 99 stars as {\it our sample}.
\noindent The final step was to consolidate the values of the mean large frequency separation, $\Delta \nu$, which was more difficult to determine for the cases with very low SNR and low resolution. We visually checked the \'echelle diagrams of the {\color {black} 99} targets and found 19 stars where the value of $\Delta \nu$ could be improved to have straighter ridges for the radial modes. So we refined $\Delta \nu$ by applying an additional analysis with the A2Z+ pipeline where we created a template to mimic the modes with five orders where the initial value for $\Delta \nu$ was taken as the value expected from scaling relations. We cross-correlated that template with the region of the PSD around $\nu_{\rm max}$ and swept the value of the mean large separation by +/-20\%. The value for which we obtained the maximum correlation is taken as the new $\Delta \nu$. This allowed us to converge for 7 stars. For the remaining 12 stars (KIC~3124465, 5818478, 6289367, 7009852, 7255919, 7598321, 7708535, 8349736, 8652398, 9892947, 9894195, 9912680), we selected $\Delta \nu$ that straightened the ridges. The associated uncertainties for those stars were obtained by computing the difference between the selected value and the value obtained by the A2Z pipeline.
\noindent For the 60 stars in common with the sample of Ba20, we confirmed the seismic detections for { 53} stars. The values of $\nu_{\rm max}$ of these stars are in agreement with the values reported by Ba20 within 10\%. { More specifically, there is an average offset of 4.3\% in $\nu_{\rm max}$ with a dispersion of 9.2\% and an average offset of 1.4\% in $\Delta \nu$ with a dispersion of 4.1\%.} The mean large frequency spacings agree within 5\% except for KIC~7669332, which appeared as an outlier in Figure 3 of Ba20. Our visual check of the \'echelle diagram confirms the reliability of our estimation. We also note that two targets of our sample (KIC~7215603 and 9715099) were included in \citet{2013ApJ...767..127H} and \citet[][hereafter C14]{2014ApJS..210....1C} but no $\nu_{\rm max}$ values were reported due to the aforementioned issue in the DR24.
\noindent We tested whether the apparent large fraction of stars more massive than usually observed in previous seismic data sets could be related to any artifact due to, for instance, the low signal-to-noise of the detection. Any effect translating in an overestimation of $\nu_{\rm max}$ or underestimation of $\Delta \nu$ will translate in an overestimated stellar mass derived from the seismic scaling relation. A general bias toward large $\nu_{\rm max}$ value is not expected for seismic data dominated by noise \citep[Eq. 23 of][]{2019A&A...622A..76M}. In order to identify any bias in $\Delta \nu$, we used the formalism developed by \citet{2013A&A...550A.126M}, which expresses the oscillation pattern in a parametric form. Therefore, we slightly smoothed the oscillation spectrum, with a smoothing function with a typical width of $\Delta \nu / 50$, and correlated it with the generic oscillation pattern defined by \citet{2013A&A...550A.126M}. This allowed us to do a more precise measurement of the large separation, so that we can consider that the fraction of stars more massive than usually observed is real.
\noindent In Table~\ref{tab1} we provide the global seismic parameters as well as the atmospheric parameters of the {\color {black} 99 stars with confirmed detection of solar-like oscillations}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/Diag_HR_seismic_Kepler_newdet_DR25_Balona_confirmed.pdf}
\caption{Seismic Hertzsprung-Russell Diagram where the mean large frequency separation is used instead of the luminosity. The C14 solar-like stars are represented with grey circles where the effective temperature is taken from the { M17}. The {\color {black} 99} stars with confirmed seismic detections are shown with red squares where the effective temperature is coming from APOGEE, LAMOST or B20 (see Section~\ref{sec:atmosph}). The position of the Sun is indicated by the $\odot$ symbol and the grey solid lines represent evolution tracks from ASTEC { \citep{2008Ap&SS.316...13C}} for a range of masses at solar composition (Z$_\odot$ = 0.0246). Typical uncertainties are represented in the bottom left corner.}
\label{HRD}
\end{center}
\end{figure}
\noindent Figure~\ref{HRD} represents the seismic Hertzsprung-Russell Diagram of the \kep\ solar-like stars for which C14 had detected oscillations (grey symbols) along with the new confirmed detections (red symbols). We can see that many of the { stars in our sample} are hotter than the Sun, hence more on the massive side, as well as more evolved solar-like stars, with a large fraction of subgiants. We also populate more the early red giant region where C14 had reported around tens of such evolved stars.
\noindent { We note that six of our targets have a flag 'BinDet\_NoCorr' in the B20 catalog, suggesting that they are binary candidates, which will have an impact on the derived effective temperature towards the red. We then looked for the {\it Gaia} DR2 effective temperatures of those stars. For four of them the values agreed with the ones reported by the spectroscopic surveys within the uncertainties. For one star, there is no {\it Gaia} DR2 $T_{\rm eff}$. For the target KIC~3633538, the {\it Gaia} effective temperature is of 5676\,K compared to 5100\,K in B20. The lower temperature in B20 can be explained by the binarity so we use the {\it Gaia} effective temperature for that star. This higher value is also more compatible with the location of the solar-like oscillation modes for that star.}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/New_det_Dnu_numax_Chaplin_confirmed.pdf}
\caption{Mean large frequency spacing, $\Delta \nu$, as a function of the frequency of maximum power, $\nu_{\rm max}$ for the C14 sample (grey circles) and { our sample} (red squares). Typical uncertainties are shown in the upper left corner.
\label{dnu_numax}
\end{center}
\end{figure}
\noindent The results of the global seismic analysis are shown in a $\Delta \nu$-$\nu_{\rm max}$ diagram (Figure~\ref{dnu_numax}) for the { stars in our sample} (red squares) to be compared to the C14 sample (grey circles). We can see that the relations are very similar. Nevertheless, we can also note some of the { stars in our sample} below the general trends, { in particular near $\nu_{\rm max}$ of 500\,$\mu$Hz}, suggesting that those stars are slightly more massive than the rest of the sample, which was shown for instance by \cite{2010A&A...517A..22M}. The average uncertainties are of 2.4\% in $\nu_{\rm max}$ and 5.8\% on $\Delta \nu$. The latter is slightly high compared to the usually 5\% reported in C14 but representative of the low SNR.
\noindent Using the surface gravity and effective temperature in the seismic scaling relations, we can estimate a predicted value for the frequency of the maximum power as follows:
\begin{equation}
\nu_{\rm max} = \nu_{\rm max, \odot} \frac{g}{g_{\odot} \sqrt{T_{\rm eff}/T_{\rm eff, \odot}}},
\end{equation}
\noindent where $\nu_{\rm max, \odot}$ is the frequency of maximum power for the Sun, taken as 3100\,$\mu$Hz, $T_{\rm eff, \odot}$= 5777\,K, and $\log g_{\odot}$=4.4377\,dex.
\noindent In Figure~\ref{numax_comp}, the measured $\nu_{\rm max}$ for the confirmed sample is compared to the predicted value, $\nu_{\rm max, pred}$ from Eq. 1. Here we used the $\log g$ from B20 as they have been shown to be more reliable. The B20 catalog contains 186,301 \kep\ stars and a surface gravity value was available for {\color {black} 97} of our stars. The agreement between the observed and the predicted $\nu_{\rm max}$ is in general within 20\% with an average discrepancy of {\color {black} 1.2\%}, with the predicted $\nu_{\rm max}$ overestimating the observed one. However the disagreement can reach up to 50\%.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=9cm]{Figures/numax_pred_numax_obs_confirmed_errors_allnew_B20.pdf}
\caption{Ratio of the observed $\nu_{\rm max, obs}$ and the predicted $\nu_{\rm max, pred}$ using the $\log g$ from B20. The { dashed} line shows the equality of both values.}
\label{numax_comp}
\end{center}
\end{figure}
\noindent We did a similar comparison by using the spectroscopic surface gravities (see Appendix A) and the disagreement is larger with an average discrepancy of {\color {black} 30\%}, again with an overestimation of the prediction.
\noindent We note that around 10 stars have high enough SNR to fit the individual modes. The determination of the frequencies of the individual modes will allow us to do boutique modeling for that subsample of stars, which will provide more precise stellar parameters as well as ages. However this is out of the scope of this paper and is part of a subsequent paper (Mathur et al. in prep.).
\noindent We also visually checked the \'echelle diagrams of the remaining 49 stars of the 148 stars without confirmation of acoustic-mode detection. A list of candidates of 26 stars was kept where either the \'echelle diagram seemed to show some ridges, though with a very poor SNR, or when there was some agreement with the predicted $\nu_{\rm max}$. Note that some of these candidates did not necessarily have an agreement between pipelines and three of them are part of the Ba20 sample. We show the properties of those candidates in the Appendix B.
\noindent {Finally, we re-analyzed the C14 sample using the DR25 providing a homogeneous catalog with the global seismic properties of the main-sequence and sub-giant stars with detection of solar-like oscillations observed in short-cadence during the survey phase of the {\it Kepler} mission (see Table~\ref{tab2}). To that sample we also added a few tens of stars from \citet{2011A&A...534A...6C,2011ApJ...733...95M,2012A&A...543A..54A,2015A&A...582A..25A,2017A&A...601A..82W}. In 19 stars we have realized that the SAP aperture of the new DR25 is not appropriate for asteroseismic studies because they were too small (2 or 3 pixels). By applying the same aperture extraction methodology described in this paper for the LC dataset we were able to reduce the noise level of the resultant light curve and the modes were detectable. In total, we detected the modes for 526 stars with 514 stars in common with the C14 sample. As shown by \citet{2017EPJWC.16001007S} who analyzed a subsample of solar analogs with DR23 and DR25, for the high SNR targets, the global seismic parameters remain the same. The values reported by C14 and the A2Z results with the DR25 have a median offset of 0.5\% in $\nu_{\rm max}$ and of 0.2\% in $\Delta \nu$. {We note that we also report $\nu_{\rm max}$ values for 34 stars that were in C14 but for which only $\Delta \nu$ was provided and no $\nu_{\rm max}$ was given.} We do not report the seismic parameters for 4 stars from C14 as we could not detect solar-like oscillations. }
\section{Granulation and modes amplitudes}
\noindent The determination of the global seismic parameters provides invaluable information on the convective parameters of the solar-like stars as well as the maximum amplitude of the modes. It has been shown that there are tight relations between these parameters and the frequency of maximum power of the acoustic modes \citep[e.g.][]{2011ApJ...743..143H,2011ApJ...741..119M,2012A&A...537A..30M,2014A&A...570A..41K}. Here we study these parameters and compare them with the known sample of solar-like stars with detected oscillations. This comparison allows us to check any deviation from those known relations.
\subsection{Convective background parameters}
\noindent The study of a large number of stars with solar-like oscillations from the main sequence to the red-giant branch showed that the convective background parameters are correlated with the location of the modes and follow scaling relations \citep{2011A&A...529L...8K,2011ApJ...741..119M,2014A&A...570A..41K}.
\noindent {\color {black} The background fit was performed with a Monte Carlo Markov Chains (MCMC) strategy using the \texttt{apollinaire} module\footnote{The source code is available at \url{https://gitlab.com/sybreton/apollinaire}} (Breton et al. in prep) which uses the \texttt{emcee} package \citep{2013PASP..125..306F}.
For the fit, we considered the PSD above 5 $\mu$Hz. The model is composed of four components: a power law for the magnetic activity, two Harvey models \citep{1985ESASP.235..199H} for different scales of granulation, and the photon noise at high frequency. All the results have been visually inspected and some PSD were fitted again with a frequency cut at 50 $\mu$Hz. For that higher frequency cut, the fitted model was only constituted of two Harvey models and a flat noise profile. } { The Harvey law fitted has the form:
\begin{equation}
H(\nu) = \frac{P_{\rm gran}}{1+(\nu/\nu_{\rm gran})^\alpha},
\end{equation}
\noindent { where $P_{\rm gran}$ is the granulation power, $\nu_{\rm gran}$ is the characteristic frequency}, and the slope $\alpha$ is fixed to 4 following \citet{2014A&A...570A..41K}. { $P_{\rm gran}$ is related to the granulation amplitude as the $A_{\rm gran}^2$ per frequency bin.}
}
\noindent Figure~\ref{taugran_numax} shows the granulation frequency for the second Harvey model of convection (the one below the modes), { $\nu_{\rm gran}$ (top panel)}, and the granulation { power, $P_{\rm gran}$} (bottom panel), as a function of $\nu_{\rm max}$. The grey points represent the values for a subsample of 163 stars from the C14 sample. The analysis of that subsample was also done with the \texttt{apollinaire} pipeline on data calibrated with the KADACS software that generated the KEPSEISMIC lightcurves. This allows us to directly compare the parameters of the { stars in our} sample with { those} of C14.
We can see in both plots that the new confirmed detections have the convective parameters in the same range as the known oscillating solar-like stars and have similar trends following the relations derived for stars from the main sequence to the red-giant branch \citep{2014A&A...570A..41K} shown with blue dashed lines. For $P_{\rm gran}$, we used the power law relation with a slope of 2.1 { as given in Section 5.4 of \citet{2014A&A...570A..41K}.}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/comp_nugran_gold_newdet_Balona_Apo_err_K14.pdf}\\
\includegraphics[width=9cm]{Figures/comp_Pgran_gold_newdet_Balona_Apo_err_K14.pdf}
\caption{Granulation frequency (top panel) and { power} (bottom panel) as a function of the frequency of maximum power. The { stars in our sample} are represented with red squares while the subsample of 163 stars from C14 are represented with grey circles. { The blue dashed lines are the relations derived by \citet{2014A&A...570A..41K}.}}
\label{taugran_numax}
\end{center}
\end{figure}
\subsection{Maximum amplitude of the modes}
From the Gaussian fit around the region of the modes, we derived the bolometric maximum amplitude of the modes following conversion of \citet{2008ApJ...683L.175K} and \citet{2011A&A...531A.124B}. For three stars, namely KIC~3633538, 9529969, 10340511, the Gaussian fit could not converge due to the low SNR and did not provide the maximum amplitude of the modes. { However, a value for the frequency of maximum power was obtained from the maximum around the region of the modes in the power spectrum density that was smoothed with a boxcar average of width 2\,$\times \Delta\nu$.}
\noindent In Figure~\ref{Amax_numax}, we can see here again that the { stars in our sample} have a similar behavior compared to the C14 sample. { We note that for the most evolved stars of our sample -- evolved subgiants with $\nu_{\rm max}$ down to 500\,$\mu$Hz and early red giants with lower $\nu_{\rm max}$ \citep[see Fig.~1 of][]{2014A&A...572L...5M}--, $A_{\rm max}$ is closer to the lower edge. The blue dashed line corresponds to the fit of the form of a power law between $A_{\rm max}$ and $\nu_{\rm max}$ for the C14 sample. We find $A_{\rm max} \propto \nu_{\rm max}^{-0.95}$.} These lower amplitudes combined { with} the noisier DR24 data, could explain the difficulty to detect the modes in the early analysis of those targets by \citet{2011Sci...332..213C}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/Amax_Chaplin_newdet_Balona_conf_error_bars_linfit.pdf}
\caption{Bolometric maximum amplitude of the modes as a function of $\nu_{\rm max}$. The legend is the same as in Figure~\ref{taugran_numax}. { The blue dashed line corresponds to the fit for the stars from C14.}}
\label{Amax_numax}
\end{center}
\end{figure}
\section{Global stellar parameters}\label{sec:MR}
\noindent Using the seismic scaling relations \citep{1991ApJ...371..396B,2011A&A...529L...8K}, we computed the masses and radii of { our sample} where we combined the global seismic parameters and the atmospheric parameters from Section~\ref{sec:atmosph} { as follows:}
\begin{equation}
\frac{R}{R_\odot}= \left( \frac{\Delta\nu_{\odot}}{ \Delta \nu}\right)^{-2} \left(\frac{\nu_{\rm max}}{\nu_{\rm max, \odot}} \right) \left(\frac{T_{\rm eff}}{5777} \right)^{1/2}\\
\end{equation}
\begin{equation}
\frac{M}{M_\odot}=\left(\frac{\Delta\nu_{\odot}}{\Delta \nu} \right)^{-4} \left(\frac{\nu_{\rm max}}{\nu_{\rm max, \odot}} \right)^3 \left(\frac{T_{\rm eff}}{5777} \right)^{3/2}. \\
\end{equation}
\noindent { where $\nu_{\rm max, \odot}$\,=\,3100\,$\mu$Hz and $\Delta\nu_{\odot}$\,=\,135.2\,$\mu$Hz.}
\noindent Among the confirmed detections, {\color {black} 79 stars} have APOGEE spectroscopic parameters and {\color {black} 16} have LAMOST values. The remaining {\color {black} 4} stars have $T_{\rm eff}$ and [Fe/H] from the {\it Gaia-Kepler} catalog.
\noindent In Figure~\ref{MR}, we represent the mass-radius diagram with { our sample} with red squares and the previously known sample of solar-like stars with oscillation detections in \kep\ with grey circles. For the latter sample, masses and radii were computed by \citet[][hereafter S17]{2017ApJS..233...23S} who used grid-based modeling where the global seismic parameters ($\Delta \nu$ and $\nu_{\rm max}$) were combined { with} spectroscopic observables ($T_{\rm eff}$, $\log$\,g, and [Fe/H]). We can see that our new sample includes a larger fraction of massive stars as suggested above.
\noindent Given that the observed $\Delta \nu$ might not be obtained in the asymptotic regime as it should be measured in a frequency range above $\nu_{\rm max}$, several approaches have been developed to apply a correction on the observed mean large frequency separation in order to reduce the impact on the derivation of the stellar masses and radii { \citep[e.g.][]{2011ApJ...743..161W,2013A&A...550A.126M,2016MNRAS.460.4277G,2016ApJ...822...15S,2018A&A...616A.104K,2021A&A...648A.113B}}. For instance, \citet{2013A&A...550A.126M} derived an empirical correction by taking into account the second-order term in the Tassoul relation \citep{1980ApJS...43..469T} and by using seismic observations for main-sequence to red-giant stars. The approach by \citet{2016ApJ...822...15S} used a grid of models of red giants where the mean large frequency from the the sound-speed profile was compared to the one obtained by using the computed frequencies, leading to a grid of $\Delta \nu$ corrections for each model of the grid. Given that our targets have a rather small SNR and that they are main-sequence stars and subgiants, we decided to adopt the \citet{2013A&A...550A.126M} corrections. They led to smaller radii by 1.5\% { on} average and smaller masses by 2.9\% { on} average compared to the radii and masses obtained without $\Delta \nu$ corrections.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/M_R_diag_newdet_Balona_confirmed_error_bars_black.pdf}
\caption{Radius vs Mass diagram for the C14 stars using S17 parameters (grey circles) and the { stars in our sample} described in this work (red diamonds). For the { latter}, corrections on $\Delta \nu$ were applied following Mosser et al. (2013).
\label{MR}
\end{center}
\end{figure}
\noindent In Figure~\ref{MR}, we can see that 2 stars have masses above 2\,$M_\odot$, which are not expected in our sample { as they are in the instability strip and do not have an external convective envelope needed for the excitation of solar-like oscillations \citep{2010aste.book.....A}}. Given the large uncertainties on the seismic parameters that are reflected by the larger error bars for those stars they could be attributed to the low SNR of the modes.
\section{Rotation and magnetic activity}\label{sec:rot}
For the {\color {black} 99} stars with a confirmed detection of solar-like oscillations, we analyzed the long-cadence time series to look for the surface rotation periods and measure the level of magnetic activity. These measurements are based on the presence of active regions/spots that come in and out of view and lead to a modulation in the lightcurves that is related to the surface rotation. For this analysis we applied three different techniques: a time-frequency analysis based on wavelets \citep{1998BAMS...79...61T, liu2007,2010A&A...511A..46M}, the auto-correlation function \citep{2014A&A...572A..34G,2014ApJS..211...24M}, and the composite spectrum that combines the two previous methods \citep{2016MNRAS.456..119C,2017A&A...605A.111C,2019ApJS..244...21S}. The reliable rotation periods, $P_{\rm rot}$, were selected following criteria described in \citet{2019ApJS..244...21S}, \citet{2021A&A...647A.125B} and \citet{2021ApJS..255...17S}. { Briefly, we use the three different filtered KEPSEISMIC lightcurves to select the rotation period. We then use the PDC-MAP light curves to confirm that the rotational signal does not result from pollution (instrumental or stellar) as the KEPSEISMIC lightcurves have a larger aperture compared to the PDC-MAP and are calibrated without taking into account the information on the instrumental drifts embedded in the co-trending basic vectors used in PDC-MAP. We should note that PDC-MAP light curves are often filtered at 20 days and, in \citet{2019ApJS..244...21S}, we showed that solely by using PDC-MAP light curves the distribution of retrieved rotation periods is shifted towards shorter values. When the rotation periods inferred from both calibration systems agree inside the combined errors, { they are} less likely to come from some instrumental pollution { or pollution by nearby stars}. Some examples of lightcurves are shown in Appendix D.} We obtained surface rotation periods for {\color {black} 63} stars among which {\color {black} one} was classified as a close binary candidate { based on the fast rotation and the shape of the lightcurve \citep[See][for more details]{2019ApJS..244...21S,2021ApJS..255...17S}}.
\noindent For stars with rotation period measurements, we computed the photometric magnetic index, $S_{\rm ph}$, following \citet{2014JSWSC...4A..15M}. Briefly, we first calculate the standard deviation of subseries of length 5\,$\times P_{\rm rot}$ and compute the mean value. This index has been shown to provide a relevant proxy for magnetic activity based on the Sun and a few solar-like stars and comparison with classical magnetic activity indexes \citep{2016A&A...596A..31S,2017A&A...608A..87S}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/Prot_Teff_newdet_Balona_error_bars.pdf}
\caption{Surface rotation periods, $P_{\rm rot}$, as a function of effective temperature, $T_{\rm eff}$, for the G14 sample (grey circles) and { our sample} (red squares). Typical error bars are shown in the bottom right corner.}
\label{Teff_Prot}
\end{center}
\end{figure}
\noindent In Figure~\ref{Teff_Prot} we represent $P_{\rm rot}$ as a function of the effective temperature of the stars for the new seismic detections and the sample of \citet[][hereafter G14]{2014A&A...572A..34G}. The stars with new seismic detection seem to be quite homogeneously distributed. While we had noticed that a significant fraction of those stars are in their subgiant phase, very few stars have long rotation periods. Because stars spin-down as they evolve { \citep[e.g.][]{1972ApJ...171..565S,1988ApJ...333..236K,2013A&A...556A..36G}}, a larger number of slow-rotating ($P_\text{rot}>40$ days) subgiants were expected in comparison with what we found in this work. However many of the subgiants of our sample are in general more massive than 1.2\,M$_\odot$, which corresponds to the Kraft break \citep{1967ApJ...150..551K}. Stars above the Kraft break do not undergo magnetic braking due to stellar winds in contrast with the lower mass stars. This difference between the high-mass and low-mass solar-like stars could explain the rotation periods retrieved for our sample.
\section{Discussion}
We look here at the different properties of the stars with { detections} of solar-like oscillations { presented in this paper} and compare them to the C14 sample, in particular to investigate a possible gain in the parameter space that we are probing.
\subsection{Magnitude distribution}
We first look at the magnitude of the targets. Indeed, in addition to the calibration issue of the data, the previous non detection of the modes for those stars could also be affected by higher photon noise that depends on the stellar magnitude. The higher noise in the data could lead to a non detection. Figure~\ref{Kp_histo} shows the distribution of the \kep\, magnitude, $K_p$, for { our sample} and the C14 sample. While they probe a similar range of magnitudes, we can see that the { stars in our sample} have a larger fraction of stars fainter than 11 compared to the C14 sample. We find that {\color {black} 55.5\%} of the stars in our sample have $K_p\,>$\,11, compared to 39\% for the C14 sample.
\noindent We perform a Kolmogorov-Smirnov test to quantify the comparison of the distribution of the magnitudes between the C14 sample and our sample of new seismic detections. We find a deviation value of {\color {black} 0.22} as well as a probability on the distributions differences of {\color {black} 0.05\%}, which means that the magnitudes distributions are indeed very different. This confirms that our sample with { detections} of solar-like oscillations { presented in this paper} are fainter.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/histo_Kp_newdet_Balona_confirmed_normalized.pdf}
\caption{Normalized histogram of the \kep\, magnitude for the C14 sample (black { dashed} line) and { our sample} (red solid line). }
\label{Kp_histo}
\end{center}
\end{figure}
\subsection{Blends explaining the previous non detections}
With these { detections} of solar-like oscillations { in our sample} using the new DR25, we can wonder whether the previous non detection was related to the presence of nearby stars. To check for that possibility, we cross-matched the list of { stars in our sample} with Gaia EDR3 \citep{2021A&A...649A...1G} and checked two possible sources of amplitude dilution:
\begin{enumerate}
\item Companions that are unresolved by {\it Gaia}, as indicated by high {\it Gaia} RUWE (Renormalized Unit Weight Error) values (RUWE$>$1.4 typically indicates binaries). While the contrast between the main star and the companions is not known, it is possible that a pollution can arise.
\item All resolved stars in {\it Gaia} within 20\arcsec (5 {\it Kepler} pixels) of the target. For stars with multiple companions, we selected the brightest star within the search radius.
\end{enumerate}
\noindent From that analysis, { we find 15 stars where RUWE\,$>$\,1.4 and 9 stars for which the magnitude difference between our target and the brightest star within 20\arcsec\, is smaller than 3}. So 24 stars are flagged to have a nearby star that may dilute the amplitude of the flux and could have prevented the detection of the modes. This number represents an upper limit of the number of stars for which oscillations could not be detected due to a blend. We also checked the amplitude of the modes for those stars that we flagged and found that 7 of them have a lower amplitude of the modes, which can thus be explained by the possible blends.
\noindent { For the remaining 22 stars without a nearby polluting star it is very likely that the better quality of the data allows us to detect the modes presented in this paper.}
\subsection{Metallicity distribution}
Metallicity can play a role on the detectability of solar-like oscillations as was shown by \citet{2010A&A...509A..15S}. Indeed, they suggested that for metal-poor stars, the amplitude of the modes was decreased. Figure~\ref{FeH_histo} shows the distribution of the metallicity for the {\color {black} 95} stars { of our sample} (red solid line), where [Fe/H] comes from either APOGEE or LAMOST as metallicity from B20 is not as reliable. We can see that it peaks at {\color {black} solar metallicity}. For comparison the metallicity distribution of the C14 sample is represented on the same figure with the { dashed} line, also peaking at solar metallicity. We note that among the { stars in our} sample, {\color {black} 51\%} of { them} are metal poor compared to the Sun while 54\% of the C14 sample are metal poor but given the typical uncertainties of 0.1dex the difference does not seem significant. We quantify the differences between the distributions of the two sample using a Kolmogorov-Smirnov test. We obtain a deviation of {\color {black} 0.08} with a probability that one sample is different from the other of {\color {black} 62.4\%}, confirming that the distributions are not significantly different.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/histo_FeH_confirmed_Balona_spectro_normalized.pdf}
\caption{Normalized histogram of the \kep\ metallicity for the C14 sample (black { dashed} line) and { for our sample} (red solid line) for the 96 stars with new seismic detections and with APOGEE or LAMOST spectroscopic observations. }
\label{FeH_histo}
\end{center}
\end{figure}
\subsection{Surface magnetic activity}
{ As explained in Section~5.2,} from the photometric activity proxy, $S_{\rm ph}$, we can also investigate whether the { stars with} { of our sample} have different levels of magnetic activity compared to G14. In Figure~\ref{Sph_Prot}, we show the magnetic activity index $S_{\rm ph}$ as a function of the rotation period $P_{\rm rot}$ for the stars from G14 with reliable rotation periods and for which asteroseismic detection had been obtained (grey circles). The results for the { stars in our sample} are shown with the red squares and the close binary { candidate with a blue square}. We can see that most stars with detected oscillations have similar $S_{\rm ph}$ values that are in the same range as the solar ones between the minimum and maximum of its magnetic cycle { as computed in \citet{2019FrASS...6...46M} } (delimited by the dashed lines). There are some stars with magnetic activity levels above the range of the solar cycle for both the previously known sample and { our sample}. However the new sample of this work adds four stars that are slower rotators ($P_{\rm rot} \ge$ 30 days) and with high $S_{\rm ph}$ values. We also note that one star (KIC~8165738) has a very low $S_{\rm ph}$ value of $\sim$\,10ppm. The modulation seems to be real. We recall that since we do not have the inclination angles of the rotation axis of the stars, $S_{\rm ph}$ values that we measure are a lower limit of the real magnetic activity level.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=9cm]{Figures/Prot_Sph_newdet_Balona_confirmed.pdf}
\caption{Magnetic activity proxy, $S_{\rm ph}$ vs surface rotation period, $P_{\rm rot}$ for the { G14} sample (grey circles) and { our sample} (squares). The blue square is the close binary candidate, KIC 4255487. The dashed lines correspond to the $S_{\rm ph}$ values between minimum and maximum magnetic activity { from \citet{2019FrASS...6...46M}}, The typical error bars are represented in the bottom right-hand side.}
\label{Sph_Prot}
\end{center}
\end{figure}
\noindent { Among those stars, three are flagged as potential binary systems as discussed in Section 3.2 (KIC~3633538, KIC~777146, KIC~10969935).}
\begin{figure*}[h!]
\begin{center}
\includegraphics[width=9cm]{Figures/comp_RGaia_seis_newdet_Balona_med.pdf}
\includegraphics[width=9cm]{Figures/comp_RGaia_seis_Serenelli17_mosser_corr_med.pdf}
\caption{Left panel: Comparison between the seismic radii after applying the \citet{2013A&A...550A.126M} corrections on $\Delta \nu$ and the {\it Gaia} radii (top panel) and ratio of the radii for { our sample} (bottom panel). { The blue squares represent the median binned data.} Only 97 stars are shown as two targets do not have a {\it Gaia} radius. Right panel: Same for the S17 sample. }
\label{RGaia}
\end{center}
\end{figure*}
\subsection{Radius comparison with Gaia DR2}
\noindent As shown in Section~\ref{sec:atmosph}, the \kep\,-\emph{Gaia} stellar properties catalog made use of the \emph{Gaia} DR2 in order to improve the stellar parameters of the \kep\ targets by using the new precise parallaxes measurements. B20 fitted isochrones using spectroscopic and photometric information available from previous catalogs and combined them with the \emph{Gaia} data. As said before, 97 of our solar-like stars with new seismic detections are in the B20 catalog
\noindent We find a general agreement between the seismic radii, $R_{\rm seis}$ and the \emph{Gaia} radii, $R_{Gaia}$, for both { our sample} (left panel of Figure~\ref{RGaia}) and the S17 sample (right panel of Figure~\ref{RGaia}) for comparison. For more clarity in the figure, we represent the S17 sample without the uncertainties. In { our} sample, we note that for larger stars (radii above $\sim$\,2.5$R_{\odot}$), hence more evolved stars, the seismic radii uncertainties become larger, due to larger uncertainties on the seismic parameters in comparison with smaller stars. We find that in average the {\it Gaia} radii are overestimated compared to the seismic ones by {\color {black} 4.4\%, with a scatter of 12.3\%. A general trend can be seen, where with increasing seismic radii $R_{Gaia}/R_{\rm seis}$ decreases. For main-sequence stars, the average disagreement is of 10.2\%, where the seismic radii are underestimated compared to the {\it Gaia} radii. For subgiants with radii between 1.5 and 3\,$R_{\odot}$, we find that the seismic radii are underestimated by 3.3\% { on} average. Finally, for early red giants with radii above 3\,$R_\odot$, the seismic radii are overestimated by 1.2\% { on} average. }
\noindent These results are slightly different from what was found by \citet{2017ApJ...844..102H} where a similar comparison was done on a larger sample of stars from the main sequence to the red giants but using the Tycho-{\it Gaia} astrometric solution \citep{2015A&A...583A..68M,2016A&A...595A...2G}. Their sample included the previously known solar-like stars with detection of acoustic modes. They found a disagreement of 5\% for radii between 0.8 and 8\,$R_\odot$, where seismic radii for dwarfs and subgiants were underestimated compared to {\it Gaia} radii. Later \citet{2018MNRAS.481L.125S}, \citet{2019A&A...628A..35K}, and \citet{2019ApJ...885..166Z} looked at differences between {\it Gaia} radii obtained with DR2 and the seismic ones, for dwarfs and/or red giants. For red giants, \citet{2019A&A...628A..35K} reported a 2\% discrepancy between the seismic and astrometric radius and showed how stellar radii can be calibrated with \kep\ and {\it Gaia} (see their Figs. 17 and 19). For dwarfs, \citet{2018MNRAS.481L.125S} and \citet{2019ApJ...885..166Z} found that the seismic radii were underestimated compared to the {\it Gaia} ones by 2\%. In particular, \citet{2019ApJ...885..166Z} investigated random and systematic uncertainties related to luminosity, effective temperatures, bolometric corrections and estimated the systematic uncertainties to be around 2\% as well. Our findings for the new seismic detections show larger discrepancies. Taking the systematics errors, the results obtained for the new seismic detections are thus consistent with the previously known solar-like stars when using {\it Gaia} DR2.
\noindent To check wether any of the bias or trends between seismic radii and radii inferred with {\it Gaia} observations are due to any effect of the effective temperature scale, we compute the seismic radii using the effective temperatures from B20. We find that { on} average the {\it Gaia} radii are overestimated by {\color {black} 3.3\%} with a scatter of {\color {black}12.4\%}. Using the effective temperatures from B20 should bring the discrepancy closer to 0. Thus the difference that remains here { is} likely to be related to the input from the global seismic parameters { (see Appendix E)}.
\noindent Finally, in order to better assess the agreement between the two ways of obtaining the stellar radii, we compare the radii differences as a function of the combined uncertainties from both seismology and {\it Gaia} (see Figure~\ref{Diff_R_histo}). For {\color {black} 50.5\%} of the stars, the differences between the two methodologies are within 1\,$\sigma$.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=14cm, height=9cm]{Figures/comp_RGaia_seis_newdet_Balona_sigma_histo_gauss.pdf}
\caption{Left panel: Differences between the {\it Gaia} radius and the seismic radius in units of the statistical uncertainty ($\sigma$), computed as the quadratic sum of the uncertainties from the B20 { radii} and the scaling relations { radii}. Right panel: histogram of the same differences. The { dashed} lines show the equality of both radii and the dot-dash lines represent the $\pm$\,1\,$\sigma$ limits. Only 97 stars are shown as two targets do not have a {\it Gaia} radius. { The blue dashed line shows a Gaussian fit of the histogram.}}
\label{Diff_R_histo}
\end{center}
\end{figure*}
\section{Conclusions}
We analyzed the lightcurves from the latest data release of the \kep\ mission for $\sim$\,2,600 stars that were observed in short cadence for one month during the survey phase of the mission. We obtained reliable seismic detections for { 99} solar-like stars, among which { 46} stars with newly reported detections. Our analysis increased by more than 15\% the number of \kep\, dwarfs with a full seismic study. { We also re-analyzed the DR25 of the sample of stars with previously detected modes by C14 and provide a homogeneous catalog of global seismic parameters for additional 526 main-sequence and sub-giant stars observed by {\it Kepler} during the survey mission, yielding a homogeneous seismic properties catalog for { 625 stars for which a complete table of the global seismic parameters are given in Appendix E}.}
\noindent For the full sample, we consolidated atmospheric parameters from APOGEE, LAMOST, and the \kep\,-{\it Gaia} catalogs. Using the surface gravity and effective temperature from the {\it Kepler-Gaia} catalog, we predicted frequency of maximum oscillation power of the modes based on seismic scaling relations. We found that { on} average the predicted $\nu_{\rm max}$ is slightly higher than the observed one.
\noindent The computation of the seismic masses and radii from the scaling relations and with the \citet{2013A&A...550A.126M} corrections on the mean large frequency spacing suggested that { our target stars on average more massive and off the main sequence compared to the existing \kep\ dwarfs and subgiants in the literature}.
\noindent While the background parameters have similar trends with $\nu_{\rm max}$ { as previously found in the literature, the maximum amplitude of the modes appears to be on the lower end of the known trends.} We also found that the sample of new seismic detections is constituted of fainter stars, which can lead to lower amplitude of the modes. This could partly explain the non detection of the acoustic modes for those stars when using the DR24 that was affected by some calibration issues concerning the short-cadence data of \kep. Both our sample and the sample of \cite{2014ApJS..210....1C} have { on} average a solar metallicity
\noindent { By using {\it Gaia} data to look for stars close to our targets, we found that 7 of the stars with { detections} of solar-like oscillations { in our sample} have a nearby companion that may pollute the observations. This could explain the non detections of acoustic modes with the previous {\it Kepler} data releases.}
\noindent From the analysis of the surface rotation of the lightcurves, we obtained reliable rotation periods for 64 stars. These stars have magnetic activity levels in the same range as the Sun along its activity cycle. Even though the sample with new seismic detections has a larger number of subgiants we did not find many slow rotators. This is probably because these are massive subgiants that are above the Kraft break on the main sequence and did not go through strong magnetic braking.
\noindent Finally, we found that seismic radii are { on} average 4.4\% underestimated compared to the {\it Gaia} DR2 radii with a scatter of 12.3\% and a decreasing trend with evolutionary stage, which is in agreement with previous comparisons done for main-sequence to red-giant stars. Using a different scale of temperature from the {\it Kepler-Gaia} catalog, the discrepancy slightly decreased
\noindent Many of the stars with new seismic detections have a very low SNR, however around ten of them could have their individual modes characterized, which will be part of a subsequent paper. Even though \kep\ has stopped operating more than five years ago, the high precision and continuous data collected by the mission still represent a goldmine for asteroseismology.
\begin{table*}[h]
\caption{Seismic and stellar parameters of the 99 stars with seismic detections. { The surface gravity is derived from Eq.\,1.} The Flags provide the provenances of the atmospheric parameters ($T_{\rm eff}$ and [Fe/H]): 0 =APOGEE, 1= LAMOST, 2= {\it Kepler-Gaia} catalog (B20), { and 3= {\it Gaia} DR2}.}
\begin{center}
\begin{tabular}{ccccccccc}
\hline
\hline
KIC & $T_{\rm eff}$ (K)& $\log g$ (dex)& [Fe/H] (dex)& $\nu_{\rm max}$ ($\mu$Hz)& $\Delta \nu$ ($\mu$Hz)& M (M$_\odot$)& R (R$_\odot$) & Flag \\
\hline
2010835 & 5896\,$\pm$\, 126& 4.07\,$\pm$\, 0.01& 0.12\,$\pm$\, 0.15&
1312\,$\pm$\, 72& 72.73\,$\pm$\, 4.52& 0.91\,$\pm$\, 0.11&
1.46\,$\pm$\, 0.13&0\\
2578869 & 5395\,$\pm$\, 101& 3.84\,$\pm$\, 0.01& 0.01\,$\pm$\, 0.16&
809\,$\pm$\, 48& 48.87\,$\pm$\, 1.89& 0.91\,$\pm$\, 0.09&
1.89\,$\pm$\, 0.12&0\\
3102595 & 5770\,$\pm$\, 122& 3.98\,$\pm$\, 0.01& -0.11\,$\pm$\, 0.02&
1071\,$\pm$\, 52& 52.20\,$\pm$\, 1.66& 1.83\,$\pm$\, 0.13&
2.30\,$\pm$\, 0.11&0\\
3124465 & 5796\,$\pm$\, 120& 3.85\,$\pm$\, 0.01& -0.33\,$\pm$\, 0.01&
795\,$\pm$\, 46& 46.00\,$\pm$\, 1.56& 1.23\,$\pm$\, 0.09&
2.18\,$\pm$\, 0.11&0\\
3219634 & 6145\,$\pm$\, 133& 3.98\,$\pm$\, 0.02& 0.20\,$\pm$\, 0.01&
1054\,$\pm$\, 53& 53.86\,$\pm$\, 1.60& 1.69\,$\pm$\, 0.15&
2.19\,$\pm$\, 0.12&0\\
3238211 & 6092\,$\pm$\, 129& 3.82\,$\pm$\, 0.01& -0.13\,$\pm$\, 0.01&
720\,$\pm$\, 42& 42.44\,$\pm$\, 1.45& 1.35\,$\pm$\, 0.10&
2.38\,$\pm$\, 0.12&0\\
3241299 & 6501\,$\pm$\, 130& 4.07\,$\pm$\, 0.01& 0.28\,$\pm$\, 0.01&
1239\,$\pm$\, 59& 59.70\,$\pm$\, 2.30& 1.99\,$\pm$\, 0.18&
2.16\,$\pm$\, 0.13&0\\
3425564 & 6239\,$\pm$\, 136& 3.82\,$\pm$\, 0.01& -0.23\,$\pm$\, 0.02&
714\,$\pm$\, 38& 38.50\,$\pm$\, 5.24& 2.04\,$\pm$\, 0.54&
2.92\,$\pm$\, 0.55&0\\
3633538 & 5676\,$\pm$\, 85& 4.34\,$\pm$\, 0.03& -0.07\,$\pm$\, 0.01&
2500\,$\pm$\, 119&119.05\,$\pm$\, 3.91& 0.84\,$\pm$\, 0.13&
1.03\,$\pm$\, 0.09&3\\
3750375 & 6110\,$\pm$\, 129& 3.69\,$\pm$\, 0.01& 0.04\,$\pm$\, 0.01&
544\,$\pm$\, 31& 31.16\,$\pm$\, 4.62& 2.02\,$\pm$\, 0.59&
3.35\,$\pm$\, 0.69&0\\
3761010 & 6403\,$\pm$\, 128& 4.02\,$\pm$\, 0.01& 0.05\,$\pm$\, 0.15&
1111\,$\pm$\, 59& 59.67\,$\pm$\, 1.89& 1.38\,$\pm$\, 0.09&
1.92\,$\pm$\, 0.09&1\\
3936658 & 6106\,$\pm$\, 130& 3.70\,$\pm$\, 0.01& 0.10\,$\pm$\, 0.01&
554\,$\pm$\, 36& 36.66\,$\pm$\, 5.72& 1.09\,$\pm$\, 0.33&
2.44\,$\pm$\, 0.52&0\\
...\\
\hline
\end{tabular}
\end{center}
\label{tab1}
\end{table*}%
\begin{table*}[h]
\caption{{ Global seismic parameters from A2Z for the { 526} stars with previous detection of solar-like oscillations.}}
\begin{center}
\begin{tabular}{ccc}
\hline
\hline
KIC & $\nu_{\rm max}$ ($\mu$Hz)& $\Delta \nu$ ($\mu$Hz) \\
\hline
1430163 & 1807\,$\pm$\, 43& 85.71\,$\pm$\, 1.79\\
1435467 & 1369\,$\pm$\, 56& 70.80\,$\pm$\, 1.47\\
1725815 & 1040\,$\pm$\, 28& 55.97\,$\pm$\, 1.43\\
2010607 & 675\,$\pm$\, 7& 42.44\,$\pm$\, 1.49\\
2309595 & 646\,$\pm$\, 16& 38.97\,$\pm$\, 1.38\\
2450729 & 1078\,$\pm$\, 36& 61.05\,$\pm$\, 1.95\\
2837475 & 1638\,$\pm$\, 72& 75.71\,$\pm$\, 1.22\\
2849125 & 729\,$\pm$\, 31& 40.44\,$\pm$\, 1.29\\
2852862 & 988\,$\pm$\, 61& 54.68\,$\pm$\, 1.07\\
2865774 & 1260\,$\pm$\, 37& 64.20\,$\pm$\, 2.16\\
2991448 & 1127\,$\pm$\, 34& 61.22\,$\pm$\, 2.91\\
2998253 & 2034\,$\pm$\, 9& 89.00\,$\pm$\, 2.13\\
3112152 & 1263\,$\pm$\, 43& 65.00\,$\pm$\, 1.77\\
3112889 & 817\,$\pm$\, 30& 53.09\,$\pm$\, 1.86\\
3115178 & 431\,$\pm$\, 16& 28.72\,$\pm$\, 0.97\\
3123191 & 1704\,$\pm$\, 51& 88.00\,$\pm$\, 2.13\\
3236382 & 1692\,$\pm$\, 15& 73.74\,$\pm$\, 1.80\\
...\\
\hline
\end{tabular}
\end{center}
\label{tab2}
\end{table*}
\begin{table*}[h]
\caption{Rotation periods and magnetic activity levels of the 63 stars with seismic detections. }
\begin{center}
\begin{tabular}{ccc}
\hline
\hline
KIC & $P_{\rm rot}$ (days) & $S_{\rm ph}$ (ppm)\\%& Flag \\
\hline
3633538 & 32.10\,$\pm$\, 4.22& 172.9\,$\pm$\, 4.1\\
6881330 & 11.96\,$\pm$\, 0.70& 112.5\,$\pm$\, 5.6\\
3102595 & 14.80\,$\pm$\, 1.94& 330.2\,$\pm$\, 12.5\\
3124465 & 25.91\,$\pm$\, 1.72& 110.6\,$\pm$\, 3.7\\
3936993 & 10.66\,$\pm$\, 0.85& 115.4\,$\pm$\, 5.8\\
4255487 & 8.04\,$\pm$\, 0.66& 1484.4\,$\pm$\, 67.7\\
4270687 & 13.42\,$\pm$\, 0.81& 80.5\,$\pm$\, 4.8\\
4859338 & 5.83\,$\pm$\, 0.50& 313.6\,$\pm$\, 17.5\\
5112169 & 5.03\,$\pm$\, 0.98& 270.0\,$\pm$\, 17.9\\
5183581 & 6.10\,$\pm$\, 1.23& 193.9\,$\pm$\, 11.1\\
5394680 & 43.94\,$\pm$\, 4.78& 523.7\,$\pm$\, 10.6\\
5597743 & 13.75\,$\pm$\, 1.07& 65.0\,$\pm$\, 3.9\\
5696625 & 10.56\,$\pm$\, 1.33& 105.1\,$\pm$\, 5.9\\
5771915 & 9.27\,$\pm$\, 0.69& 92.9\,$\pm$\, 5.6\\
5791521 & 5.10\,$\pm$\, 0.64& 91.4\,$\pm$\, 6.5\\
5814512 & 5.65\,$\pm$\, 0.67& 82.5\,$\pm$\, 6.5\\
5856836 & 7.06\,$\pm$\, 0.68& 60.5\,$\pm$\, 4.7\\
6062024 & 12.97\,$\pm$\, 2.52& 24.7\,$\pm$\, 2.3\\
...\\
\hline
\end{tabular}
\end{center}
\label{tab1_1}
\end{table*}%
\begin{acknowledgements}
This paper includes data collected by the \emph{Kepler} mission. Funding for the \emph{Kepler} mission is provided by the NASA Science Mission directorate. Some of the data presented in this paper were obtained from the Mikulski Archive for Space Telescopes (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. S.~M. acknowledges support by the Spanish Ministry of Science and Innovation with the Ramon y Cajal fellowship number RYC-2015-17697 and the grant number PID2019-107187GB-I00.
R.~A.~G. and S.~N.~B acknowledge the support from PLATO and GOLF CNES grants. A.~R.~G.~S. acknowledges the support from National Aeronautics and Space Administration under Grant NNX17AF27G and STFC consolidated grant ST/T000252/1. D.H. acknowledges support from the Alfred P. Sloan Foundation, the National Aeronautics and Space Administration (80NSSC19K0597), and the National Science Foundation (AST-1717000). M.S. is supported by the Research Corporation for Science Advancement through Scialog award $\#$26080. Guoshoujing Telescope (the Large Sky Area Multi-Object Fiber Spectroscopic Telescope LAMOST) is a National Major Scientific Project built by the Chinese Academy of Sciences. Funding for the project has been provided by the National Development and Reform Commission. LAMOST is operated and managed by the National Astronomical Observatories, Chinese Academy of Sciences.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{ Open Issues}
Although blockchain has been applied in many fields, it still faces huge challenges requiring effective solutions when adopted by smart grids. In this section, we highlight some of these challenges and point out future directions to stimulate further research.
\noindent{\bf Reliability and safety}. Since a power grid is a critical infrastructure,
its reliability and safety are of paramount importance for all energy providers and consumers.
The smart grid has used digital information and control mechanisms such as load adjustment/balancing, demand-side management, and self-healing technologies to achieve reliability and safety.
Blockchain can be used to enhance reliability and safety of a smart grid,
but this demands a deep integration of blockchain
with the smart grid. Such an integration requires interdisciplinary collaborations among
experts in computer science, information security, and electric systems.
The inherent fault tolerance of blockchain can help to enhance reliability of smart grids.
Blockchain can be applied to components of a smart grid
like the intelligent control system and the operational system to make them more robust against
faults or attacks. Meanwhile, because of the immutability of blockchain, it can be
an ideal media to share critical control information for grid management.
\noindent{\bf Performance}. A smart grid is featured by real-time control, monitoring,
and information exchange, and it has strict performance requirements on
the communication latency. Unfortunately, blockchain is not designed
for real-time applications, since its consensus latency may be too long for control,
optimization or trading in smart grids. Even fast consensus algorithms such as DPoS
still require seconds to reach a consensus. How to design a blockchain system
that can process data in real time is both interesting and challenging.
It is also important to improve computation and communication efficiency of blockchain
to support real-time operations in smart grids. The blockchain validators/miners
need to synchronize and verify each transaction and block, which may introduce long delays for smart grids.
Thus efficient algorithms should be specifically designed for real-time operations in smart grids.
\noindent{\bf Scalability}. As a smart grid consists of many actuators, measurement units, sensors
and smart meters that scatter over a wide area, the blockchain system must be highly scalable to efficiently and speedily manage them. This calls for a scalable blockchain architecture design to support
large-scale smart grids. Such a consideration has not received enough attention from the research community.
Fortunately, many potential solutions have been proposed recently to improve blockchain
throughput. Blockchain sharding has been a very effective mechanism
to improve scalability of a blockchain system. The multi-chain mechanism divides
the entire system into independent components, with each running its own blockchain.
Sidechain and the corresponding pegged mechanism can also contribute to improve scalability.
For both multi-chain and sidechain, appropriate cross-chain protocols should be applied to ensure security for transactions.
The off-chain or layer-2 solutions such as Lightning Network may be used
to shorten the confirmation delay of transactions in blockchain.
\noindent{\bf Security issues}.
Blockchain can help to establish trust and it is deemed trustworthy, but inappropriate use of
the blockchain technology may lead to severe consequences, especially for smart grids.
Like most security-critical information systems, the most important security problem in a smart grid is how to securely manage cryptographic keys. However, many existent blockchain systems
do not adopt effective key management strategies, and they just allow users to generate random keys
without certification. This can result in severe security problems, e.g. impersonation attacks and sybil attacks.
Entity and data authentication are basic requirements for security of smart grids, and this is only possible when
effective key management mechanisms are applied to blockchain.
A number of security threats should be carefully studied in designing a blockchain system for smart grids.
Illegal access to critical data of smart grids should be effectively prevented with access control mechanisms,
which are important for data sharing in smart grids.
Hardware security should be considered in case of device compromise in smart grids.
For example, devices like smart meters are easy to be compromised and tamper-proof hardware may be used to guarantee
data security. For electricity trading, fairness should be guaranteed to thwart malicious consumers or providers.
\noindent{\bf Privacy issues}.
Privacy issues are even more challenging than security for both blockchain and smart grids.
Blockchain is well-known for its transparency and public verifiability, which means {\em every}
validator/miner can access the content of all transactions. Different types of privacy need to be
protected, including identity privacy, location privacy, amount privacy, relationship privacy,
and data privacy.
From the perspective of privacy protection,
the single point of failure in a centralized system is turned into {\em multiple} single points
of failures in a blockchain system. An attacker only needs to compromise one validator/miner to
obtain sensitive information of the smart grid. Therefore, the privacy problem is both challenging
and critical when applying blockchain to smart grids.
In addition, traffic analysis is also a potential threat to secure operations of smart grids.
Simple encryption cannot solve privacy issues as validators/miners still need to verify the blockchain data. Advanced cryptographic techniques can be applied to solve privacy issues. Ring signature,
zero knowledge proof, zk-SNARKs, homomorphic encryption have been proposed recently for privacy preservation in blockchains.
However, these techniques are normally inefficient in computation and/or communication, so
they must be carefully adapted to meet the requirements of smart grids.
\noindent{\bf Compatibility for low-end devices}. The actuators, sensors, smart meters and measurement units
in a smart grid are usually resource-constrained, and it is difficult for them to carry out
certain public key cryptographic computations. The security mechanisms should accommodate
these low-end devices, and the computation burden on these devices should be minimized without
losing security and privacy. For example, low-end devices should only carry out
symmetric encryption and hashing operations, and should avoid frequent signature generation/verification.
\section{Conclusion}
In this survey, we have reviewed the latest research progress on applying the blockchain technology to smart grids.
We first summarize characteristics of blockchain systems, the smart contract paradigm, and
the advantages of the blockchain technology. Then we categorize the relevant research works into
5 groups: intelligent energy management, energy trading, security and privacy, microgrid management, and electric vehicles.
For each category, we briefly discuss issues with traditional solutions without using blockchain.
Then we analyze and compare how blockchain is utilized to tackle the issues. A table is provided to give a clear comparison among these solutions, followed by a brief summary for each category.
For completeness of this survey we also list and summarize related projects and initiatives.
Finally, we elaborate on open problems and research challenges that have not been
sufficiently investigated by current research.
Apparently, the blockchain technology has many desirable properties that can be utilized to
construct a better smart grid. To this end, collaborations among professionals from different areas
are very important. We expect that this survey can provide principle reference and direction guidance for the future development of blockchain-enabled smart grids.
\section{Introduction}\label{intro}
The groundbreaking blockchain technology has gained tremendous attention from governments and
companies to research institutions all over the world in recent years.
This is due to the outstanding advantages of blockchain in
decentralization, transparency, immutability, fault tolerance, and strong security.
Essentially, blockchain is a decentralized consensus ledger managed by multiple
maintainers (called miners or validators), without relying on a centralized server.
For instance, the underlying blockchain of Bitcoin or Ethereum is maintained
by thousands of miners using a specific consensus mechanism (e.g. proof-of-work or proof-of-stake).
Each transaction sent to the blockchain is verified and executed by all the validators,
and it cannot be changed anymore after it is confirmed on the blockchain.
Besides transactions, blockchains like Ethereum and Hyperledger fabric
also support the so-called smart contracts, which are self-executable programs running over blockchains.
Consequently, blockchain functions as a highly secure and trustworthy ``consensus computer"
that can serve as a trusted intermediary.
After the surprising success in the cryptocurrency field, the blockchain technology has been rapidly
introduced to many other fields ranging from financing, banking, insurance, manufacturing,
supply chain management, provenance, healthcare, Internet-of-Things to cloud computing.
More than 70 international banks have joined R3 to develop the open blockchain platform Corda,
aiming to lower cost of financing and banking. The Internet giant Facebook has announced its own cryptocurrency Libra, which has drawn the attention of the whole world. IBM, Amazon, Samsung, and other companies have also shown their strong
interests in blockchain, e.g. IBM's supply chain and Walmart's blockchain. In addition, a number of countries have already started to apply blockchain
to improve transparency in government management. In conclusion, blockchain
has been employed to empower various applications in many fields. In this article, we focus on the application of blockchain in smart grids.
\subsection{Why Smart Grids Need Blockchain}
A smart grid is a critical infrastructure that can be significantly improved with the blockchain
technology. Current smart grids incorporate communication and control technologies into
the power grids, such that significant improvements on energy efficiency and grid safety can be achieved. To this end, numerous smart devices are distributed throughout a smart grid to efficiently manage power generation, transmission, distribution and consumption.
It is critical to effectively manage these smart devices and maintain security and
stability of the smart grid. However, traditional centralized approaches to manage smart grids
face severe challenges in various aspects, ranging from energy management, electricity trading, security \& privacy, microgrid management, to electric vehicle management, which are further detailed in the following descriptions.
\begin{itemize}
\item \textbf{Energy Management.} The centralized grid management mode leads to the single point of failure and severe security issues.
A potential attacker would launch various attacks specifically targeting at a control center to maximize damage.
For instance, denial-of-service attacks against the control center may result in serious damages to the whole power grid.
Moreover, the centralized grid management mode cannot make the best of the distributed renewable energy sources, energy storage units, and electric vehicles to achieve efficient energy management.
It is challenging for the control center to make good and timely scheduling decisions for the highly dynamic and distributed renewable energy sources and storage units.
At the same time, with the increase number of prosumers, the burden and cost of centralized energy management have been increased, which makes it impossible to achieve effective demand response in many cases.
Especially for demand-side management, some consumers are more inclined to purchase electricity from prosumers according to their own wishes, rendering energy management even more challenging.
\item \textbf{Electricity Trading.} The widespread use of distributed energy indicates that the transformation of the centralized smart grid transaction model to a decentralized one is inevitable.
The centralized mode cannot efficiently deal with energy trading among distributed energy providers, consumers, and prosumers throughout a smart grid.
Moreover, the price of centralized grid energy transactions cannot meet the expectations of the market, and it is easy to cause unreasonable monopoly of electricity prices.
Therefore, it is important to improve energy trading efficiency and reliability of the smart grid.
\item \textbf{Security \& Privacy.} There exist growing security concerns on centralized smart grids.
External malicious attacks, third-party reliance, and privacy leakage have caused huge economic losses of the power grids.
The traditional centralized power grid has limited solutions.
Most situations rely on human supervision. However, it is a challenging task to manage the large amount of smart devices for status monitoring, grid control and metering in a centralized manner.
On the other hand, numerous smart sensors generate huge amount of data, which is hard to be stored and processed by a centralized server.
Moreover, it is also inefficient for the control center to maintain security for so many smart devices in case of malicious intrusions.
\item \textbf{Microgrid Management.} As a specific application scenario of the power grid, the microgrid has the characteristics of regional management, solving the problems caused by long-distance energy transmission in a large power grid.
However, as a sub-system of the large power grid, it still has huge hidden dangers in energy distribution, security and privacy protection, and load balancing.
\item \textbf{Electric Vehicle Management.} As another application scenario of the power grid, the combination of energy vehicles and the power grid adds a great flexibility to long-distance power transmission.
Being an independent power storage unit, each vehicle has certain limitations on its computing power and is extremely vulnerable to external attacks.
Especially in the process of charging/discharging and power transmission, secure payment and demand response are greatly challenged.
Moreover, in V2G (vehicle-to-grid), because of the need to exchange road conditions, vehicles have higher requirements for real-time performance.
\end{itemize}
This survey dedicates to the application of the blockchain technology in smart grids for various purposes.
We divide our study into 5 broad areas: energy management, energy trading, security and privacy protection,
microgrid management, and electric vehicles. This corresponds to the five challenges mentioned above. The road map is shown in Fig. \ref{fig:structure}. For each area, we discuss issues in conventional approaches and highlight existing problems and issues therein, describe
the blockchain-based solutions that utilize the blockchain technology to address the corresponding issues, emphasize the applicability of blockchain in these blockchain-based solutions,
and discuss the lessons learned from these studies. We also summarize existing projects and initiatives that integrate blockchain with smart grids from a practical side. Finally we analyze the remaining research challenges and put forward a number of potential research problems.
\begin{figure}[htbp]
\includegraphics[width=0.7\linewidth]{structure_new.pdf}
\caption{Roadmap of blockchain-based smart grids.}\label{fig:structure}
\end{figure}
\subsection{Related works}
A number of surveys have been conducted on related topics recently.
Meisel et al.~\cite{8217069} focus on research works that apply blockchain to microgrids. They study 14 cases of integrating microgrids and blockchain, and analyze the advantages in each case.
Xie et al.~\cite{xie2019survey} investigate the research works on applying blockchain to smart cities, treating a smart grid as a part of a smart city. This survey analyzes about 20 related papers on combination of blockchain and smart grids.
Andoni et al.~\cite{andoni2019blockchain} focus on the development of the energy sector, particularly the emerging projects in the energy sector. They examine 140 emerging blockchain energy enterprise projects, while only summarizing a small number of research papers.
Ref.~\cite{8730307} studies 40 papers on combining blockchain and smart grids, and discusses several important aspects of smart grids, including energy trading, security, electric vehicles, and microgrids. The authors also describe 3 new frameworks and 10 practical projects that integrate blockchain with smart grids.
Alladi et al.~\cite{alladi2019blockchain} focus on the integration of blockchain and smart grids too. This survey considers energy transaction, energy trading based vehicles, security \& privacy, and business cases, provides a number of valuable views, and puts forward the corresponding challenges and prospects.
More recently, Sengupta et al.~\cite{sengupta2020comprehensive} summarize many papers on blockchain and IoT/Industrial IoT, among which 20 of them are on smart grid security. Both \cite{ajomand2020review} and \cite{mollah2020blockchain} analyze the related works
from the aspects of energy trading, security, microgrids, and electric vehicles, while ignorning energy management. Ajomand et al.~\cite{ajomand2020review} provide a review on smart grids but they only survey about 20 papers. Ref.~\cite{mollah2020blockchain} offers a survey on about 40 papers and the related projects.
\subsection{Uniqueness of This Survey Article}
The survey articles mentioned in the previous subsection have made outstanding contributions to the understanding of the state-of-the-art in applying the blockchain technology to smart grids. Although smart grid is not the focused research subject in some articles \cite{xie2019survey, andoni2019blockchain, sengupta2020comprehensive}, the results of their collation still have high reference values. However, the shortcomings are also obvious. Particularly, there lacks enough articles to comprehensively summarize the unique characteristics of integrating blockchain with smart grids. Furthermore, smart grid management, which is particularly important, is only slightly mentioned in \cite{andoni2019blockchain}. In addition, some articles only provide a certain concept expression, lacking the description and summary of specific programs \cite{alladi2019blockchain, ajomand2020review}. Finally, the rapid development of blockchain and smart grids requires a more trendy, comprehensive, specific, and innovative article that can summarize the major recent academic achievements.
These factors motivate us to write this article. We intend to provide a comprehensive survey covering as many research works as possible, including the latest papers accepted or published in 2021. We divide the subject of this survey into 5 areas, involving more than 150 specific papers. For each area, we first outline the current challenges, then survey blockchain-based solutions, and finally provide an analytical summary and comparison study. Particularly, we make a lot of effort to outline the main idea of each paper, exemplify representative solutions and key technical details in the forms of figures and tables. One can see that the scope of research and the number of papers surveyed by this article are both unprecedented. To further improve the completeness and practicality of this survey, we sort out related sources, including important field experiments and current projects. Moreover, our survey covers the management of smart grids that have been overlooked by most related surveys, and divides it into supply side energy distribution management and demand side management. One can see that this survey article provides a clear logic for readers to comprehensively understand the subject areas, from headlines to subtitles then to the theme of each paragraph.
\subsection{Organization}\label{Organization}
This article is organized as follows. In the next section, we provide a brief overview on the blockchain and smart contract technologies,
emphasizing the opportunities and advantages of the combination of blockchain and smart grids.
In Section \ref{Management}, we present the blockchain-based schemes on energy management, covering energy distribution and demand side management.
Section \ref{trading} discusses the blockchain-based energy trading schemes for smart grids, emphasizing how the pricing of a trading process is ensured.
Then we describe the efforts on applying the blockchain technology to protect security and privacy in Section~\ref{security}.
Section \ref{Microgrid} summarizes the blockchain-based approaches to microgrid management, demonstrating how safety, automation, and power distribution in microgrids are realized.
Section \ref{Vehicle} presents the blockchain-based solutions on management of electric vehicles, including vehicle charges/discharges and energy transmission.
In Section \ref{resource}, we collect the resources on projects and initiatives that integrate blockchain and power grids, focusing on related practical applications and public experiments.
Finally, we discuss open research issues to stimulate further research.
\section{Background}
\subsection{Blockchain and Smart Contract}
The blockchain technology is originated from Bitcoin proposed by Satoshi Nakamoto in 2008. With the deepening of research, Ethereum, Tendermint, IOTA, ZCash, and other blockchain platforms have been gradually developed. They have their own characteristics in terms of delay and functionality, but their essence is a distributed database. In a blockchain system, miners play an important role, both as participants and as maintainers of the blockchain, upholding the system through actively accounting and packaging verified transactions to obtain certain rewards.
The emergence of consensus algorithms improves the stability of blockchain systems and solves the problem of data asynchrony. More importantly, blockchain systems support various consensus algorithms such as PoW, PoS, and PBFT, which possess their own advantages and are suitable for different scenarios.
Blockchain systems can be classified as public blockchains, consortium blockchains, and private blockchains. These three categories differ in the restrictions on the members who can participate. Public blockchains are open to all the people and thus everyone can participate; consortium blockchains only allow alliance members to participate; while private blockchains can only be maintained by a person or an organization. The specific differences are shown in Table \ref{Comparison}. Due to the unique design concepts, blockchain systems have many characteristics such as decentralization, transparency, immutability, and trustworthiness.
The idea of smart contracts is originally proposed by Nick Szabo. A smart contract is essentially a computer protocol to guarantee trusted transactions without the intervention of third parties. The advantages of smart contracts include reducing manpower, lowering costs, and realizing the automation of the system operation. Smart contracts exist in digital forms, which allow them to integrate perfectly with computers. They are generally formulated and executed together with users to better reflect their wishes. So far, not all blockchain platforms can support smart contracts. The most famous combination of smart contracts and blockchain is the implementation of Ethereum. There also exist many blockchain platforms that do not support smart contracts, including Bitcoin, Factom, etc.
\begin{table*}[htbp]
\begin{threeparttable} \footnotesize
\caption{Comparison Of Public, Private And Consortium Blockchain}
\vspace{-0.3cm}
\label{Comparison}
\begin{tabular}{c c c c c c c}
\toprule
Blockchain & Participants & Access control & Degree of decentralization & Security level & Transaction speed & Transaction cost \\
\specialrule{0em}{1pt}{1pt}
\midrule
Public blockchain & All users & No & Completely & High & Slow & High \\
\specialrule{0em}{1pt}{1pt}
Private blockchain & Partial users & Yes & Partially & Low & Fast & Low \\
\specialrule{0em}{1pt}{1pt}
Consortium blockchain & Partial users & Yes & Partially & Medium & Fast & Medium \\
\bottomrule
\end{tabular}
\end{threeparttable} \normalsize
\end{table*}
\subsection{Integrating Blockchain with Smart Grids: Opportunities and Advantages}
Smart grid is an intelligent transformation of traditional power grid. It solves a series of problems existing in traditional power grids such as poor elasticity, information islands, and single information flow. In a smart grid, the physical structure and the division of labor are clear. The information of the entire smart grid can be obtained by specialized agencies to achieve optimal distribution of energy. However, a smart grid still has a lot of problems and face a number of challenges (shown in Fig. \ref{intro}). This provides opportunities for blockchain to enter the smart grid field. More specifically, blockchain, as a new type of distributed structure, can be integrated with a smart grid (as illustrated in Fig. \ref{WechatIMG85}) and has a positive impact on the smart grid.
\begin{itemize}
\item Blockchain can transform and upgrade centralized grid management to distributed intelligent management. Firstly, a blockchain can realize effective management on the supply side of a grid. For a power grid with a large number of equipment and distributed energy units, the blockchain can provide decentralized supervision, accurate demand response, supply-demand balance, and optimized distribution of electrical energy. This enables the smart grid to cater to the growing trend of prosumers. Secondly, a blockchain can provide more democratic management for the demand side. Users can join the blockchain as nodes for energy data inspection, supervision, and even formulating corresponding rules according to their wishes through smart contracts. Finally, the decentralized nature of a blockchain can transform a smart grid from centralized storage management to distributed multi-point management.
The smart grid enabled by the blockchain technology can realize effective scheduling of electricity storage to minimize energy loss.
\item The financial nature of blockchain makes it an ideal platform for electricity trading in smart grids. A blockchain enabled smart grid can realize optimized data flow, energy flow and cash flow in the process of energy trading.
On one hand, energy trading in a traditional grid is usually centralized and thus is easy to cause monopoly. The decentralized nature of blockchain can allow prosumers to join the grid and trade electricity in a P2P manner. The immutability of blockchain can record the process of energy transactions and the interactions of related data.
Meanwhile, the transparency of blockchain allows users to verify grid data, which makes grid energy transactions and data more open and reliable.
On the other hand, blockchain can provide a secure cash flow for energy trading. Cryptocurrency has proven its security, credibility and convenience in payment processing.
The incentive mechanism of blockchain and smart contracts can realize dynamic pricing and flexible auctions between prosumers. This makes electricity trading in a smart grid more flexible and convenient.
\item Blockchain can greatly improve the security and privacy of the power grid for its decentralization and fault tolerance. A blockchain itself is highly tolerant to faults, which can enhance the underlying security of the power grid. Specifically, the consensus algorithm can effectively enhance the robustness of the system and resist malicious attacks. It is worth mentioning that the blockchain does not rely on third parties, which effectively prevents leaking private information to third parties. Considering the large number of users, the anonymity of blockchain can protect identity information.
\item Blockchain can also extend its advantages to microgrids. Due to the regional characteristics of a microgrid, strict access control mechanisms are required. Blockchains have their own identity management mechanisms, which can be adapted to suite the needs of microgrids. Moreover, a microgrid powered by the blockchain technology can facilitate electricity transmission within a short distance, thereby significantly reducing electricity loss. The combination of smart contracts and microgrids can also make the automated management of microgrids more rational and transparent.
\item Blockchain is well-suited for managing the V2G network in smart grids. The financial nature and security mechanisms of blockchain can ensure information security during the charging and discharging process. Each vehicle is responsible for energy storage and transmission, and can be used as a blockchain node for mining and reaching consensus. Lightweight consensus algorithms require low computing power and are suitable for energy vehicles.
\end{itemize}
\begin{figure*}[ht]
\includegraphics[width=0.9\linewidth]{sg_application.pdf}
\caption{Application scenarios of blockchain and smart grids.}\label{WechatIMG85}
\end{figure*}
\section{Blockchain for Intelligent Energy Management}\label{Management}
\subsection{Current Issues in Intelligent Energy Management}
Energy management is the general term for scientific planning, organization, inspection, control and supervision of the entire process of energy production, distribution, conversion and consumption. It can be roughly divided into supplier energy management and demand side energy management.
In order to achieve intelligent smart grid management, a number of schemes have been proposed to analyze consumption data for building models. However, existing methods have poor scalability and require expensive computation for real-time and dynamic smart grids. In addition, traditional energy management has problems with centralized storage, and hence is prone to single-point failures, which is easy to cause excessive waste of energy resources.
It turns out that solutions based on blockchain can implement effective management of electrical energy. They not only improve security of smart grids, reduce energy consumption and manpower, but also achieve higher scalability and decentralized design. These solutions can be further divided into two groups: effective electricity distribution and demand side management. For these two specific groups, we elaborate on the problems and technical characteristics of each solution. Finally, we conduct a comparative analysis on these solutions.
\subsection{Effective Electricity Distribution}
The goal of energy distribution is to meet electricity demand, so as to achieve reasonable energy management.
Tsolakis et al.~\cite {TsolakisA} establish a secure demand response platform based on OpenADR and blockchain. In this platform, OpenADR can make the equipment respond to peaks and troughs of electricity consumption, thereby reducing power consumption and realizing a reasonable energy distribution. In order to enhance robustness and realize supervision of the system, an information collection center is established in this system.
Like \cite {TsolakisA}, Van et al.~\cite{van2020cooperative} design an energy management platform for smart buildings, aiming to meet the user's demand for renewable energy. At the same time, the authors conduct an experimental test based on Ethereum and employ smart contracts for automatic management and monitoring. The results show that this scheme is suitable in a distributed environment with high power consumption. Edmonds et al.~ \cite{edmonds2020blockchain} propose an energy management platform for smart homes to address demand response. They make use of smart contracts and propose a privacy protection mechanism based on a private blockchain.
In addition to meeting demand response, effective energy distribution can optimize the entire energy system.
In \cite{alskaif2019decentralized}, AlSkaif et al. apply a distributed optimal power flow model to manage the local distribution network based on a private blockchain.
On this basis, the authors propose an optimization algorithm based on ADMM (alternating
direction method of multipliers), achieving more reasonable and efficient energy distribution. Similarly, Yang et.al~\cite{yang2021blockchain} use ADMM to achieve cost minimization modeling, which is especially suitable for flexible appliance scheduling.
Marius et al.~\cite{BusinessSmart2020} utilize reasonable energy distribution to achieve load optimization in virtual power plants. Moreover, research works~\cite{nakayama2019transactive, yahaya2020blockchain} aim to realize optimal management of energy costs. Ref.~\cite{nakayama2019transactive} uses smart contracts to implement economic scheduling among the prosumers and protect the interests of each prosumer. Yahaya et al.~\cite{yahaya2020blockchain} propose a model based on smart contracts and the demurrage technology to optimize consumers' energy consumption as well as achieve a reasonable distribution of energy.
Blockchain also shows a great potential in managing renewable resources.
Ref. \cite{LazaroiuBlockchain} uses smart contracts to achieve automation and transparent design. Lazaroiu et al. propose a price prediction algorithm to improve the accuracy of energy allocation. Experiments show that this scheme is particularly suitable for smart cities and prosumers. In addition, the main purpose of the scheme \cite{KimSecurity} is to ensure the security and privacy of renewable energy in the energy management process. Kim et al. summarize the potential security risks when renewable energy sources join the traditional power grid distribution management, and propose to use blockchain to improve the management of the smart grid.
Skowronski \cite{SkowronskiOn} proposes an energy management solution for green-field energy and realizes a smart grid based on the GRIDNET protocol. Smart devices in the smart grid are divided into light nodes and full nodes, and the maintenance of the entire blockchain is managed by the full nodes. Each node has a certificate when registering, and all full nodes run PoW consensus algorithm to manage the blockchain. The blockchain is utilized to prevent node misbehaviors and motivate benign behaviors.
Shao et al.~\cite {shao2019grid} propose a DER (Distributed Energy Resources) grid-connected VPP (virtual power plant) model. This model combines with the incentive measures of DER and blockchain to realize a reasonable energy distribution. In~\cite{wang2018blockchainassisted},
a secure crowdsourcing energy platform is proposed to adjust energy distribution. In this scheme, crowdsourcing sources come from two types: regular submissions of crowdsourcing tasks and real-time submissions. In order to motivate users, Wang et al.~\cite{wang2018blockchainassisted} design a monetary incentive mechanism, which is based on market demand and game theory.
Furthermore, Utz et al.~\cite{utz2018blockchain} implement an energy management platform based on smart contracts to meet the distributed problems caused by prosumers. This platform also implements related incentive mechanisms to facilitate rational use of distributed energy.
CBSG~\cite{FanConsortium} is designed to monitor and manage the data generated during energy distributions. It combines the data regulation, signcryption, and aggregation algorithms, ensuring the safety and integrity of energy data (shown in Figure \ref{fig:FanConsortium}).
Suciu et al.~\cite{suciu2019blockchain} propose SealedGrid to provide effective distributed energy management. On this basis, SealedGrid provides real-time monitoring functions to create a secure environment for participants.
\begin{figure}[htbp]
\includegraphics[width=0.95\linewidth]{FanConsortium.pdf}
\caption{ The workflow description of CBSG \cite{FanConsortium}.
\label{fig:FanConsortium}
\end{figure}
\subsection{Demand Side Management}
Blockchain based demand side management must first solve the problems of price and cost optimization to achieve economic savings.
Afzal et al.~\cite{afzal2020blockchain} apply smart contracts to realize automated billing functions in the community, and game theory is used to set a reasonable price. In this scheme, each user in the community can perform optimal energy management according to its own situation and protect its own privacy. Li et al.~\cite{li2019design} conduct in-depth research on Singapore's demand-side energy management. They develop a scheme based on smart contract and game theory for price optimization. However, it is worth noting that this scheme takes into account the uncertainties of renewable resources, and uses the receding horizon optimization technique to do demand side management.
Ref. \cite{wu2017m2m} proposes a process optimization scheme based on M2M (Machine-to-Machine) technology. This solution uses smart contracts to publish price records, which are analyzed to get a model of demand response. Similarly, Pop et al.~\cite{pop2018blockchain} adopt smart contracts and incentive mechanism in Ethereum, thereby achieving an accurate demand response mechanism.
Effective demand side management can reduce the negative impact of peak load.
In \cite{wang2019blockchain}, Wang et al.~discuss in detail the possibility of blockchain for demand-side management, and apply smart contracts to achieve load optimization and price optimization. Dang et al.~\cite{dang2019demand} analyze the demand-side management of large industries, and use individual load optimization to achieve optimization for the entire industry.
There exist a number of schemes that focus on the problem of demand response. The work in \cite{erenouglu2020blockchain} explores the feasibility of blockchain to achieve demand response and conducts field experiments. In \cite{Noor2018Energy} and \cite{hajizadeh2020blockchain}, the authors propose demand-side energy management for microgrids.
The model \cite{Noor2018Energy} reduce the peak-to-average ratio based on game theory.
Hajizadeh et al.~\cite{hajizadeh2020blockchain} use a smart contract to provide an automated processing mechanism. The smart contract is negotiated between both parties, so it meets the expectations of consumers and improves the flexibility of the system.
\subsection{Summary}
We summarize the design goals, targets, and other blockchain-related attributes of existing works discussed in this section, and report the results in Table \ref{management22}.
\begin{table*}[htbp]
\begin{threeparttable} \footnotesize
\caption{Blockchain For Intelligent Energy Management}
\vspace{-0.3cm}
\label{management22}
\begin{tabular}{c c c c c c c c c}
\toprule
Ref. & Design goal & Target & Consensus & Blockchain \\
\specialrule{0em}{1pt}{1pt}
\midrule
Wang\textit{ et~al.}\cite{wang2018blockchainassisted} & Market balance & Crowdsourcing & - & Hyperledger \\
\specialrule{0em}{1pt}{1pt}
Cutsem\textit{ et~al.}\cite {van2020cooperative} & Demand response & Smart building & PoW & Ethereum \\
\specialrule{0em}{1pt}{1pt}
Pop\textit{ et~al.} \cite {pop2018blockchain} & Demand response & Prosumer & PoS & Ethereum \\
\specialrule{0em}{1pt}{1pt}
Ioannis\textit{ et~al.}\cite{TsolakisA} & Demand response & Prosumer & - & Ethereum \\
\specialrule{0em}{1pt}{1pt}
Edmonds\textit{ et~al.}\cite {edmonds2020blockchain} & Demand response & Smart homes & PoW & Ethereum \\
\specialrule{0em}{1pt}{1pt}
Shao\textit{ et~al.}\cite {shao2019grid} & Connection freely & Any & iPoW & Semi-center \\
\specialrule{0em}{1pt}{1pt}
Tarek\textit{ et~al.}\cite {alskaif2019decentralized} & Flow optimization & Prosumer & - & Private \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite {li2019design} & Flow optimization & Prosumer & dPoS & Bitcoin \\
\specialrule{0em}{1pt}{1pt}
Dang\textit{ et~al.}\cite {dang2019demand} & Load optimization & Big industrial & PoW & - \\
\specialrule{0em}{1pt}{1pt}
George\textit{ et~al.}\cite{LazaroiuBlockchain} & Distribution & Prosumer & - & - \\
\specialrule{0em}{1pt}{1pt}
Rafal\textit{ et~al.}\cite{SkowronskiOn} & Distribution & Any & Hybrid & Ethereum \\
\specialrule{0em}{1pt}{1pt}
Fan\textit{ et~al.}\cite{FanConsortium} & Flexible regulation & Any & dPoS & Consortium \\
\specialrule{0em}{1pt}{1pt}
Wu\textit{ et~al.}\cite {wu2017m2m} & Cost optimization & Any & PoC & Bitcoin \\
\specialrule{0em}{1pt}{1pt}
Afzal\textit{ et~al.}\cite {afzal2020blockchain} & Cost optimization & Smart homes & - & Ethereum \\
\specialrule{0em}{1pt}{1pt}
Yahaya\textit{ et~al.}\cite {yahaya2020blockchain} & Cost optimization & Smart homes & PoW & Ethereum \\
\specialrule{0em}{1pt}{1pt}
\bottomrule
\end{tabular}
\end{threeparttable} \normalsize
\end{table*}
After reviewing the aforementioned blockchain-based solutions for smart grids,
we obtain the following findings:
\begin{itemize}
\item The decentralized nature of blockchain facilitates efficient energy management in smart grids, including energy distribution and demand side management. Many solutions use blockchain to carry out distributed transformation of the smart grid so that it can adapt to the management of new distributed energy sources. Distributed management can realize the decentralization of management rights and solve the centralized management problems of traditional power grids, which can facilitate the conversion of consumers to prosumers.
\item Most management solutions that combine blockchain and smart grids can realize effective demand response balance, entire system optimization, distributed storage management of resources, and overall market balance control \cite{patil2021study}.
\item Some solutions are suitable for various application scenarios and have good universality. There are also solutions that target specific scenarios, such as smart home, prosumer, and outsourcing fields. Blockchain has played an important role in all these schemes.
\item The multiple combinations of blockchain platforms and consensus algorithms provide a huge number of choices for various application scenarios. At the same time, some solutions even innovatively propose new blockchain platforms and consensus algorithms to perfectly adapt to their needs.
\item Smart contracts and incentive mechanisms closely integrated with blockchain can effectively realize automated operations of a system, eliminate the intervention of third parties, and attract more users to join. Especially in the management of smart grids, smart contracts can give full play to their greatest advantages, eliminate hidden human hazards, increase system security, and reduce required costs.
\item However, the biggest problem with most solutions is that they all ignore user privacy protection. They simply mention that their solutions can protect privacy, but do not provide technical details on privacy protection.
\end{itemize}
\section{Blockchain For Energy Trading}\label{trading}
\subsection{Current Issues in Energy Trading}
Trading energy in a competitive energy market can significantly improve energy efficiency,
reduce cost of energy consumers, and increase income of energy providers.
However, because the energy industry is intensive in capital, labor, and resource,
only large-scale energy giants can survive and thus monopolize the energy market.
As the emergence of distributed energy sources such as solar panels, wind turbines,
and electric vehicles, it is important for smart grids to support prosumers, which
may simultaneously be energy consumers and providers. Unfortunately,
the current wholesale energy market only accommodates large-scale electricity plants,
while the prosumers are normally small in capability. Although the retail energy market
may accommodate small-scale prosumers, it is hard to efficiently support so many prosumers.
Hence blockchain is introduced to smart grids
for improving energy efficiency and usage of renewable energy.
In this section, we categorize energy trading from two aspects: peer-to-peer energy trading
and transaction pricing.
We also provide a summary on existing schemes.
\subsection{Blockchain-Enabled Peer-to-Peer Energy Trading}\label{num3.1}
The first challenge in energy trading comes from the prosumers distributed in a smart grid, because prosumers in traditional power grids do not have sufficient rights, which makes traditional power grids hard to fulfill the current power market requirements \cite{pop2020trading}. On the other hand, as more prosumers use the rooftop solar photovoltaic technology to generate electricity, the surplus electricity of individuals and organizations become more common.
The decentralized nature of blockchain makes it well-suited to solve these problems. Ref. \cite{kumari2020blockchain} develops a blockchain-enabled energy trading platform to realize effective management of renewable resources, meeting the needs of residents, factories and other nodes accordingly.
The diversity of blockchain platforms also provides space for different design options.
Sabounchi and Wei \cite {dang2019demand} propose a transformation scheme based on Ethereum, which makes each prosumer a participant and manager of the system. At the same time, in order to incentivize participation of prosumers, the authors combine smart contracts to realize fair auction. Plaza et al.~\cite{plaza2018distributed} introduce a blockchain-based solution to serve the energy community sharing solar energy. Like the work in \cite{plaza2018distributed}, Yang \cite{yang2020blockchain} proves the feasibility of implementing distributed trading of renewable energy based on the Ethereum platform.
Moreover, the authors in~\cite{pipattanasomporn2018blockchain} propose to use Hyperledger
for trading surplus electricity and compensating the sellers accordingly. Similarly, Jamil et al. \cite{jamil2021peer} use machine learning and Hyperledger to implement an energy
trading platform that realizes electricity usage forecast and
on-demand distribution. In addition, Kavousi-Fard et al.~\cite{kavousi2021effective} propose a new consensus algorithm named Relaxed Consensus Innovation (RCI), which is suitable for the p2p energy trading market, especially microgrids.
Pee et al.~\cite{pee2019blockchain} make use of blockchain and smart contracts to implement automatic transactions, eliminating trusted third parties and ensuring transparency, immutability and traceability of energy trading.
Similarly, Yu et al.~\cite{yu2018distributed} incorporate smart contracts to realize transparent and independent energy transactions for prosumers.
Okoye et al.~\cite{okoye2020blockchain} improve the access rules of blockchain and greatly increase the transaction speed.
In addition to the above solutions for prosumers, there also exist proposals for traditional consumers to achieve supervision and optimization of the transaction system \cite{9121447, liu2020application}. ElectroBlocks \cite{tanwar2020electroblocks} solves the problems of insufficient demand and supply in traditional power grids through cost aware and store aware algorithms. Gao et al.~\cite{GaoGridMonitoring} use a sovereign blockchain to implement the supervision function of a trading system. This solution employs smart contracts to control consumer access, detect malicious nodes, and make transactions transparent. Danzi et al.~\cite {danzi2018blockchain} make use of a layered mechanism based on Ethereum to realize a transparent energy transaction market. They first implement a distributed architecture using blockchain. On this basis, the second blockchain is combined with smart contracts to achieve real-time information acquisition, implement fair pricing, and achieve demand response. The work~\cite{wang2019energy} proposes an optimization model based Hyperledger for operating a crowdsourcing energy system. In order to improve the rationality in energy trading, Wang et al.~present a two stage arithmetic algorithm. Furthermore, \cite{khorasany2020lightweight} develops an anonymous location proof algorithm, which employs distance and reputation to determine the most suitable trading partner. Umoren et al.~\cite{umoren2020blockchain} exploits the latest 5G technology and graphical interface to enhance the efficiency and supervision of the trading system. Like \cite{umoren2020blockchain}, ET-DeaL\cite{kumari2020deal} proposed by Aparna Kumari also uses the 5G technology. Additionally, it combines IPFS (InterPlanetary File System) with Ethereum to improve the security and privacy of energy transactions.
\subsection{Fair Energy Pricing}
The auction platform is of great significance in the pricing of energy transactions. Ref.~\cite{HahnSmart} proposes an auction mechanism based on smart contracts, which achieves reasonable energy pricing and eliminates third parties. More importantly, the mechanism uses Vickrey auctions to further ensure transaction fairness. The scheme presented in \cite{muzumdar2021trustworthy} also employs the Vickrey auction. It develops an iterative Vickrey–Clarke–Grove method and combines it with the proof-of-stake consensus blockchain to realize an energy trading mechanism with incentive functionality. Wang et al.~\cite{wang2018decentralized, wang2017novel} design a continuous double auction scheme to address the problem of excessive price fluctuations. In order to ensure security of the auction process, the authors use the multi-signature technique and the adaptive attack strategy to ensure the accuracy and stability of the pricing process. In \cite{stubs2020blockchain}, Marius et al. propose a mechanism similar to double auction, which optimizes distribution while effectively reducing the computational load of blockchain nodes. The work in \cite{hassan2021optimizing} analyzes the technical work of developing blockchain energy auctions from a green perspective.
Besides auction pricing, there also exist schemes that determine accurate and effective prices based on energy demand and supply. Ref. \cite{yue2019dynamic} proposes a dynamic pricing scheme for a blockchain-based ecosystem architecture, in which price is determined by an optimization problem between the microgrid and distributed system operators. This scheme incorporates smart contracts to realize automatic transaction execution. Li et al.~\cite{li2019blockchain} develop a pricing scheme to build a distributed non-cooperative market, which can achieve Nash equilibrium to ultimately maximize the welfare of the entire system. Blockchain and smart contracts are employed in this scheme to establish a transactive energy market. Similar to \cite{li2019blockchain}, the solution in~\cite{thomas2017automation} utilizes smart contracts to realize demand pricing and automated arrangements. However, this scheme provides certain rewards to users who actively adjust the price. Moreover, the solution proposed in~\cite{meng2019decentralized} is based on the information exchanged between prosumers. Prosumers conduct real-time multilateral pricing bids based on their respective power needs and actual conditions. This solution uses smart contracts to reflect the interests of the prosumers and encourage them to participate. In \cite{knirsch2018privacy}, Knirsch et al.~make use of load distribution information of power plants and customers for training and modeling, so as to predict electricity consumption and realize reasonable pricing.
The combination of blockchain and game theory can achieve fair and effective pricing in the energy trading market. Ref. \cite{LiConsortium} combines consortium blockchain to ensure transaction security, and uses the Stackelberg game to achieve reasonable pricing (shown in Figure. \ref{fig:trading}). Wu et al.~\cite{wu2018game} implement a secure energy trading platform to solve the problem of wind energy heating. The platform is a three-tier structure and has a side-payment mechanism. It is worth noting that the authors employ sequential games to achieve Nash equilibrium, ensuring reasonable and accurate energy pricing. In \cite{guo2020combined, guo2020architecture}, the authors make use of the Stackelberg game to achieve fair pricing between a grid and a communit, thyereby obtaining the maximum benefit. In these works, the blockchain platforms mainly serve to guarantee security and privacy. Smart contracts can effectively improve transaction efficiency. Guo et al.~\cite{guo2020architecture} design a new Byzantine consensus to improve the success rate. Doan et.al~\cite{doan2021peer} exploit a double auction-based Stackelberg game to encourage individual users to participate in energy trading and achieve maximum social welfare.
\begin{figure}[htbp]
\includegraphics[width=0.95\linewidth]{pic-trading.pdf}
\caption{Energy exchange process based on the scheme in \cite{LiConsortium}.}
\label{fig:trading}
\end{figure}
In addition to the above-mentioned methods, pricing in the energy trading process is also applied to certain special scenarios. Chen et al.~\cite{chen2019framework} propose a pricing algorithm for end devices in edge computing, and blockchain is used to realize distributed transactions. In \cite{wang2020research}, Wang et al.~apply smart contracts to implement a bidding system for V2G. This system employs a particle swarm to analyze and model the smart grid, so as to realize rational bidding. The authors also develop genetic algorithms to further optimize the bidding results and maximize the benefits of the entire system.
\subsection{Summary}
We summarize the key technologies and blockchain-related attributes of the works mentioned above in this subsection. The details are shown in Table \ref{trading22}.
\begin{table*}[htb]
\begin{threeparttable} \footnotesize
\caption{Blockchain For Energy Trading}
\vspace{-0.3cm}
\label{trading22}
\begin{tabular}{c c c c c c c c}
\toprule
Ref. & Target & Consensus & Demand res.
& Contract & Blockchain \\
\specialrule{0em}{1pt}{1pt}
\midrule
Gao\textit{ et~al.}\cite {GaoGridMonitoring} & Any & - & No
& Used & Sovereign \\
\specialrule{0em}{1pt}{1pt}
Okoye\textit{ et~al.}\cite{okoye2020blockchain} & Microgrid & PoS & No
& Used & Consortium \\
\specialrule{0em}{1pt}{1pt}
Yu\textit{ et~al.}\cite{yu2018distributed} & Any & - & Yes
& Used & Ethereum \\
\specialrule{0em}{1pt}{1pt}
Manisa\textit{ et~al.}\cite {pipattanasomporn2018blockchain} & Any & PoC & No
& Used & Hyperledger \\% & $\times$ $\times$ $\surd$ $\times$ \\
\specialrule{0em}{1pt}{1pt}
Wang\textit{ et~al.}\cite {wang2019energy} & Any & RBFT & Yes
& Used & Hyperledger \\% & $\times$ $\surd$ $\times$ $\times$ \\
\specialrule{0em}{1pt}{1pt}
Umoren\textit{ et~al.}\cite{umoren2020blockchain} & V2G & - & Yes
& Unused & Private \\% & $\times$ $\times$ $\times$ $\surd$ \\
\specialrule{0em}{1pt}{1pt}
Khorasany\textit{ et~al.}\cite{khorasany2020lightweight} & Any & A-PoL & Yes
& Used & Lightweight \\% & $\surd$ $\times$ $\times$ $\surd$ \\
\specialrule{0em}{1pt}{1pt}
Meng\textit{ et~al.}\cite{meng2019decentralized} & Prosumer & - & Yes
& Used & Ethereum \\% & - \\
\specialrule{0em}{1pt}{1pt}
Wang\textit{ et~al.}\cite{wang2018decentralized} & Microgrid & - & No
& Used & Bitcoin \\% & - \\
\specialrule{0em}{1pt}{1pt}
Hahn\textit{ et~al.}\cite {HahnSmart} & Any & PoC & Yes
& Used & Ethereum \\% & $\times$ $\times$ $\surd$ $\times$ \\
\specialrule{0em}{1pt}{1pt}
Wu\textit{ et~al.}\cite {wu2018game} & North China & - & No
& Used & Ethereum \\% & - \\
\specialrule{0em}{1pt}{1pt}
Danzi\textit{ et~al.}\cite {danzi2018blockchain} & Any & PoW & Yes
& Used & Ethereum \\%& $\times$ $\times$ $\surd$ $\times$ \\
\specialrule{0em}{1pt}{1pt}
Li \textit{ et~al.}\cite {LiConsortium} & IIoT & PoW & No
& Unused & Consortium \\
\specialrule{0em}{1pt}{1pt}
Yue\textit{ et~al.}\cite {yue2019dynamic} & Prosumer & - & Yes
& Used & - \\% & $\surd$ $\times$ $\times$ $\times$ \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite {li2019blockchain} & Any & - & No
& Used & Ethereum \\% & - \\
\specialrule{0em}{1pt}{1pt}
Guo\textit{ et~al.}\cite {guo2020combined} & Any & PoW & No
& Unused & Bitcoin \\% & $\times$ $\surd$ $\times$ $\surd$ \\
\specialrule{0em}{1pt}{1pt}
Wang\textit{ et~al.}\cite {wang2020research} & Prosumer & - & No
& Used & Ethereum \\% & $\times$ $\surd$ $\times$ $\times$ \\
\specialrule{0em}{1pt}{1pt}
Fabian\textit{ et~al.}\cite {knirsch2018privacy} & Any & PoW
& No & Used & - \\% & $\surd$ $\times$ $\times$ $\surd$ \\
\bottomrule
\end{tabular}
\end{threeparttable} \normalsize
\end{table*}
Having reviewed the aforementioned blockchain-based solutions for smart grids,
we obtain the following findings:
\begin{itemize}
%
\item Most of the schemes that integrate blockchain with energy trading can be applied to any transaction scenario. A few schemes are specifically developed for V2G, microgrids, prosumers, and industrial IoTs.
\item Blockchain as a carrier of the cryptocurrency or token has the desirable financial property. This makes blockchain well-suited for payment processes. More importantly, the decentralized nature of blockchain makes it both fault tolerant and highly secure. These features make it an ideal choice as a platform for energy trading in smart grids \cite{lotfi2020transition}.
\item In the application scenario of energy trading, the incentive mechanism of blockchain is particularly important. It has a positive effect on promoting energy trading and reasonable pricing. Due to the high requirements for reasonable pricing of energy transactions, current schemes also combine game theory, dynamic pricing, and auction technologies on the basis of incentive mechanism design.
\item The consensus mechanism can further ensure security and efficiency of the transaction process. Smart contracts can be employed to realize complex payment processing logics, such that payment can be made automatically and securely according to predefined rules
\item In terms of security, we find that some schemes cannot avoid the security and privacy risks that could occur during the energy trading process. Additionally, due
to the delay of blockchain itself, energy demand requests may not be responded in real time, which significantly affects the effective distribution of energy.
\end{itemize}
In addition, security and privacy issues are very important in energy trading, which are to be discussed in the next section, while energy trading in microgrids is covered in Sec.~\ref{Microgrid}.
\section{Security And Privacy Protection}\label{security}
\subsection{Current Issues in Smart Grid Security and Privacy}
A smart grid is an important infrastructure, so its security and privacy protection are the prerequisites for high-confidence system operations.
In recent years, power grids face serious security risks, threats and various attacks, including but not limited to:smart meter attacks, time synchronization attacks, physical network attacks, denial of service (DOS) attacks, false data injection attacks, replay attacks, double spend attacks, network attacks and GPS spoofing attacks.
Traditional schemes combine Kalman filter, stochastic game, sparse optimization and other technologies to prevent and detect attacks. However, most of these schemes can only be targeted at specific situations thus cannot resist all attacks. Furthermore, most of the traditional schemes are based on the centralized power grids, which can not adapt to the new distributed ones.
Blockchain, as a transparent distributed ledger, can adapt to the new smart grids. In addition, the mining mechanism, consensus mechanism, and the avoid of third parties of blockchain make it an effective way to resist most of the attacks and provide a safer environment.
In this section, we summarize the related works from the perspectives of security and privacy enabled by blockchain. We analyze each category comprehensively and present our insights obtained from this study in the end.
\subsection{Blockchain for Smart Grid Security}\label{num1}
In the process of transition to a distributed structure, the energy market is prone to potential security risks and crisis of confidence.
In \cite {hua2019blockchain}, Hua and Sun propose a transaction program for energy and carbon markets. In order to further improve security and robustness of the system, the authors adopt implement the so-called ``payment-to-public-key-hash'' with multi-signatures. Abdella et al.~\cite{abdella2019architecture} present a three-layer model to address the distributed energy management problem. The three-layer structure works in conjunction with each other to effectively improve security and efficiency. Ref. \cite{tanaka2017blockchain} highlights the important role of digital grid routers in smart grids, and proposes to integrate blockchain with digital grid routers to build a safer decentralized energy exchange platform.
The work in~\cite {MannaroCrypto} describes the method to change Sardinia's power grid in response to security risks caused by distributed transformation. This project implements a trading system in which blockchain acts as a middleman for the transaction. When a producer produces energy units, the smart meter sends the message to blockchain. Smart contracts automatically realize the interaction between information and tokens. There is also an agent in the system to help the sellers formulate strategies to maximize the seller's interests and facilitate rational use of resources.
Traditional power grids often rely on reliable third parties, which generally brings unknown security risks to the system. EnergyChain \cite{AggarwalEnergyChain} provides a solution specifically for smart homes without third parties.
External attacks are one of the common problems that cause security risks.
In \cite {ferrag2019deepcoin}, Ferrag and Maglaras propose an energy framework called DeepCoin based on blockchain and deep learning. The authors adopt a practical Byzantine fault-tolerant algorithm to improve efficiency. In order to counter against the cyber-attacks, different solutions have been proposed in \cite{mylrea2017blockchain, LiangDistributed}. The program in \cite{mylrea2017blockchain} simulates a power grid system with limited resources. Liang et al.~\cite{LiangDistributed} conduct experiments based on IEEE-118 and smart contracts to prove the effectiveness of the schemes. In order to enhance privacy protection and power
security, Kaur et al. \cite{kaur2021blockchain} propose a framework based on
software-defined networking and blockchain. The framework
combines a secure and efficient mutual authentication protocol
based on the elliptic curve cryptosystem and a smart contract
for demand response. For fake data injection attacks, Samy
et al. \cite{samy2021towards} develop an abstract model based on blockchain.
On the other hand,
many solutions have been optimized for secure payment and access for the energy market.
Researchers make effort to design secure payment platforms for energy exchange~\cite{luo2018distributed, shuaib2018using}.
The work in \cite{luo2018distributed} proposes a multi-layer grid architecture for consumers. The authors implement a trusted settlement mechanism at the second layer. Shuaib at al.~\cite{shuaib2018using} use smart contracts to verify the possibility of blockchain as a secure payment for distributed energy (shown in Figure. \ref{fig:security}).
Except the payment mechanism, Garg et al.~\cite{GargAn} develop a secure access mechanism for V2G by using elliptic curve cryptography. This scheme can ensure strong anonymity and low cost. Ouyang et al.~\cite{ouyang2017preliminary} focus on direct transactions for large scale power consumers. The framework adopts smart contracts to implement access and settlement mechanisms to ensure the security of transactions.
Bera et al. \cite{bera2021designing} propose
a new access control protocol named DBACP-IoTSG, which
is implemented in a private chain and has the characteristics
of low cost and high security.
Alcaraz et al.~\cite{alcaraz2020blockchain} consider the security problems caused by the high coupling degree of the entities, and propose a three-tier architecture to reduce the coupling degree. Tang et al. \cite{tang2021multiauthority} propose an
anonymous authentication scheme based on ring signatures,
which can realize tracking without exposing user information.
\begin{figure}[htbp]
\includegraphics[width=.95\linewidth]{security.pdf}
\caption{The workflow of the scheme in \cite{shuaib2018using}. }
\label{fig:security}
\end{figure}
In \cite{jiang2019blockchain} and \cite{huang2019development}, two solutions are mainly proposed for smart grids as an IoT system. While ensuring security, Jiang et al.~\cite{jiang2019blockchain} use smart contracts and DPoS consensus to improve efficiency. Huang at al.~\cite{huang2019development} employ Sigfox and Ethereum to confirm the feasibility of the solution. The work in \cite{sikeridis2020blockchain} mainly describes the security risks in the energy transmission system and the corresponding solutions making use of the decentralization, transparency and immutability properties of blockchain. Moreover, Ref. \cite{inayat2018load} discusses possible security issues in the AMI (Advanced Metering Infrastructure) and proposes a framework that can achieve load balancing. This framework integrates the blockchain incentive mechanism and the encryption/signature algorithm to ensure the security of the energy market and improve the efficiency of demand response. It is worth mentioning that BBARS \cite{wang2019bbars} employs a variety of public key encryption technologies to achieve the security of V2G's anonymous rewards.
Unlike the previous schemes, He et al.~\cite{he2020joint} discuss the methods to combine the advantages of the photovoltaic power and carbon market.
This scheme provides a dual-blockchain structure, in which the two chains cooperate and supervise each other to achieve data sharing. The authors also prove the usability on the Ethereum platform.
\subsection{Blockchain for Power Grid Privacy}\label{num1.2}
One of the reasons for privacy leaks lies in outside attacks or the introduction of dishonest third parties.
It is pointed out that many solutions use blockchain to transform a smart grid, but the public ledger may lead to user privacy leakage due to malicious attacks \cite{GaiPrivacy}.
This work adopts a consortium blockchain, where each energy transaction intermediary acts as a miner node, and the transaction intermediary is composed of energy reserve equipment and token banks.
It also uses smart contracts to design a black box and create a virtual account for each user.
Chen et al.~\cite{chen2020blockchainPrivacy-Preserving} employ group signatures based blockchain to build a platform, which can specifically solve the privacy leakage problems of WSNs (Wireless Sensor Networks).
In \cite {li2020blockchain}, Li et al.~propose FeneChain that makes use of anonymous authentication to strengthen privacy protection. The authors design a mechanism based on time commitment to ensure fairness and accuracy of energy transactions. Lu et al.~\cite {lu2019secure} incorporates a reputation-based fairness proof mechanism, reducing the user's computational burden and lowering the system's entry barrier. The scheme divides the electricity market transaction model into two levels that can cooperate with each other and provide a secure environment.
Ref. \cite{gai2019permissioned} points out that it is easy to launch attacks at the edge layer. Therefore, this work intends to use edge computing to enhance the security of edge nodes. At the same time, encrypted channels are introduced to ensure secure communications between edge devices. Blockchain can achieve decentralization to ensure the authenticity of the data, and users use pseudonyms to participate in transactions and preserve their privacy. Smart contracts are adopted to achieve the most efficient energy distribution strategy. In response to the privacy issues of edge devices, Guan et.al~\cite{guan2021blockchain} propose the BPM4SG scheme, which combines techniques such as multi-party secure computation, smart contracts, and ring signatures. The experimental results prove that this scheme has good performance advantages.
Sestrem et al.~\cite{sestrem2020cost} mainly tackle the potential privacy risks in the optimization process with blockchain. They protect the privacy of the entire system by establishing sidechains. Samuel et al.~\cite{OmajiA} propose an admission control mechanism based on Ethereum to solve the dilemma of service provision and privacy disclosure. The authors also combine Pagerank to encourage consumers to join. The work in \cite{guan2018privacy} applies the method of user group management to address the privacy leakage problem for smart meters. It establishes private blockchains to record the data of each group, and uses multiple pseudonyms to further protect the privacy of users (shown in Figure. \ref{fig:guan2018privacy}). Li et al. \cite{li2021lightweight} develop a privacy protection scheme in the process of price optimization,
which is based on the secure signature authentication mechanism
and identity-based proxy re-encryption strategies. In response to the privacy challenge of data aggregation, Singh et.al~\cite{singh2021blockchain} present a scheme called BHDA, which combines deep learning and homomorphic encryption to improve the performance of data aggregation and privacy protection with minimal computational overhead. Lu et.al~\cite{lu2021edge} propose a solution named EBDA to address the privacy issues caused by data aggregation , which integrates not only blockchain, but also edge computing technologies.
\begin{figure}[htbp]
\includegraphics[width=0.95\linewidth]{8419184.pdf}
\caption{ The description of the key steps of the scheme in \cite{guan2018privacy}.}
\label{fig:guan2018privacy}
\end{figure}
In the process of decentralization of smart grids, privacy issues are also extremely easy to arise.
In \cite{deng2019real} and \cite{son2020privacy}, the authors aim to achieve privacy protection for prosumers. Deng et al.~\cite{deng2019real} use a central operator to summarize the data, preventing prosumers from obtaining data not belonging to them. The proposed scheme also presents a new primal-dual gradient method for rational energy scheduling. Son et al.~\cite{son2020privacy} make use of functional encryption and smart contracts to improve system security and protect privacy. Lu et al.~\cite{lu2019blockchain} propose a solution based on SDN (Software Defined Network) technology. Experiments prove that this solution has good accuracy. Keshk et al.~\cite{keshk2019privacy} develop a two-level model. The privacy module is mainly an enhanced blockchain based on proof-of-work. It is used to ensure data integrity and mitigate data poisoning attacks. The anomaly detection module employs deep learning to integrate data. The two modules cooperate with each other, which greatly increases the security and efficiency. Zhang et al.~\cite{zhang2020privacy} employ blockchain and ring signature to enhance privacy protection. The innovation of this work lies in its adoption of the point multiplication of elliptic curve, which greatly reduces the system complexity. The work in \cite{liu2021privacy}
demonstrates a privacy-preserving network protocol based
on multi-level encryption. Guan et al. \cite{guan2021achieving} provide a general
model based on ciphertext-policy attribute-based encryption and
credibility-based equity proof consensus, which can achieve
fine-grained access control through transaction arbitration in
the form of ciphertext.
\subsection{Summary}
We summarize the design goals, key technologies, and other blockchain-related attributes of the existing solutions mentioned above. The results are reported in Table \ref{security22}.
\begin{table*}[htb]
\begin{threeparttable} \footnotesize
\caption{Blockchain For Security And Privacy Protection}
\vspace{-0.3cm}
\label{security22}
\begin{tabular}{c c c c c c c c c}
\toprule
Ref. & Design goal & Consensus & Contract & Blockchain \\
\specialrule{0em}{1pt}{1pt}
\midrule
Luo\textit{ et~al.}\cite{luo2018distributed} & Secure payment & -
& Used & Double-chain \\
\specialrule{0em}{1pt}{1pt}
Garg\textit{ et~al.}\cite{GargAn} & Security access & -
& Used & - \\
\specialrule{0em}{1pt}{1pt}
Huang\textit{ et~al.}\cite{huang2019development} & Wireless security & -
& Unused & Ethereum \\
\specialrule{0em}{1pt}{1pt}
He\textit{ et~al.}\cite{he2020joint} & Market integration & Hybrid
& Unused & Double-chain \\
\specialrule{0em}{1pt}{1pt}
EnergyChain~\cite{AggarwalEnergyChain} & Third party & PoW
& Unused & EnergyChain \\
\specialrule{0em}{1pt}{1pt}
Gai\textit{ et~al.}\cite{GaiPrivacy} & Third party & PBFT
& Unused & Consortium \\
\specialrule{0em}{1pt}{1pt}
Gai\textit{ et~al.}\cite {gai2019permissioned} & Service optimization & PoW
& Used & Permissioned \\
\specialrule{0em}{1pt}{1pt}
Guan\textit{ et~al.}\cite {guan2018privacy} & Service optimization & PoW
& Unused & Private
\\
\specialrule{0em}{1pt}{1pt}
Lu\textit{ et~al.}\cite{lu2019blockchain} & Distributed design & -
& Unused & - \\
\specialrule{0em}{1pt}{1pt}
Hua and Sun~\cite{hua2019blockchain} & Distributed design & -
& Used & - \\
\specialrule{0em}{1pt}{1pt}
Juhar\textit{ et~al.}\cite {abdella2019architecture} & Distributed design & PoS
& Used & Hybrid \\
\specialrule{0em}{1pt}{1pt}
Deng\textit{ et~al.}\cite{deng2019real} & Distributed design & -
& Unused & - \\
\specialrule{0em}{1pt}{1pt}
Keshk\textit{ et~al.}\cite {keshk2019privacy} & Distributed design & ePoW
& Unused & - \\
\specialrule{0em}{1pt}{1pt}
Son et al.~\cite{son2020privacy} & Distributed design & -
& Unused & - \\
\specialrule{0em}{1pt}{1pt}
DeepCoin~\cite{ferrag2019deepcoin} & Anti-attack & PBFT
& Unused & - \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite{li2020blockchain} & Anti-attack & PBFT
& Used & Consortium \\
\bottomrule
\end{tabular}
\end{threeparttable} \normalsize
\end{table*}
Having reviewed the aforementioned blockchain-based solutions for smart grid security and privacy,
we obtain the following findings:
\begin{itemize}
\item The research works discussed in this section have greatly improved security and privacy of smart grids. The transparency, immutability, and anonymity of blockchain properties can help to effectively protect the supervision, monitoring and management of smart grids. It is worth mentioning that security, as a basic attribute, is reflected in all fields such as wireless networks, smart homes, smart communities, V2G, and IIoT.
\item Blockchain-based solutions can effectively address various security problems caused by distributed transformation, third parties, malicious attacks, and other reasons.
\item Different encryption technologies such as asymmetric encryption, attribute-based encryption, and functional encryption, are employed to enhance security and privacy for smart grids and their users. Blockchain provides the basis for the implementation of these encryption algorithms.
\item It can be seen that the multi-level and multi-chain information transfer model can be realized on the basis of blockchain, and the protection of information privacy can be enhanced accordingly. We believe that it is indeed feasible to design solutions based on blockchain to protect user privacy.
\end{itemize}
\section{Blockchain for Microgrid Management}\label{Microgrid}
\subsection{Current Issues in Microgrid}
The concept of microgrid was first proposed by Lasseter. Although there is no formal standards for a microgrid system, this concept has been accepted by the communities and considerable research works have been dedicated to microgrids. We define a microgrid to be a small- to medium-sized and low-voltage power grid that combines various distributed power generations to provide electrical energy for local loads. A microgrid can improve the reliability of power supply on the demand side.
From the scale perspective, it is a small-scale power grid structure. But from the functionality perspective, a microgrid is an autonomous system that can realize self control, protection, and management, similar to a large-scale power grid.
However, a microgrid suffers from
high management cost and low revenue, so it can neither meet the requirements of investors nor guarantee trust for consumers.
As distributed and storage systems, various types of blockchains have been integrated with large-scale power grids. It is also possible to apply blockchain to microgrids. Recent years, many proposals based on the combination of microgrid and blockchain have been proposed, tackling issues such as security, power loss issues, automation and energy distribution in microgrids.
\subsection{Security Design of Microgrids}\label{num4.1}
A microgrid must first ensure that it can resist malicious attacks and threats from third-party organizations.
The solution in \cite{di2018energy} is based on Tendermint, which eliminates the dependence on third-party organizations. Mbarek et al.~\cite{mbarek2020enhanced} employ removable software to detect FDI (False Data Injection) attacks.
Different from the above schemes, \cite{BanksBlockchain} intends to solve the potential MIM (man-in-the-middle) attacks and DS (Data Spoofing) attacks in microgrids.
Based on Hyperledger, this work develops a blockchain-based system for sharing status and control information between microgrids. It can maintain dynamic stability for electricity generation, transmission, and distribution. Blockchain provides consensus nodes in the system to achieve trustworthy data communications between microgrids.
To enhance security and prevent cyber attacks for microgrids, Wang et al.~\cite{wang2019cybersecurity} use DAG (Directed Acyclic Graph) to improve traditional microgrids by combining the data recovery technology and the UT (Unscented Transform) technique for better efficiency and accuracy.
Like \cite{wang2019cybersecurity}, \cite{dabbaghjamanesh2019networked} also uses the UT technique to model uncertainties in demands and electricity supplies. The blockchain enabled IoT scheme can reduce operation cost and enhance security.
In order to further improve privacy protection of a microgrid, Zhang et al.~\cite {zhang2019privacy} make use of a consortium blockchain and the CDA (Continuous Double Auction) mechanism to reduce cost and improve transaction efficiency. They combine FBS (Fair Blind Signature) and the SSS (Secret Sharing Scheme) based $(t, n)$-threshold scheme to improve security and traceability of the system.
For the purpose of detecting faults and preventing damages due to fault damage propagations in a microgrid, Bayati et al.~\cite{bayati2020blockchain} propose a protection technique based on blockchain. This scheme employs encryption to protect differential relays from cyber attacks.
The security protection of a microgrid is also reflected in the field of finance. In \cite {jeon2019study}, Jeon et al.~use a hybrid blockchain to solve the hidden danger caused by the problem of double spending. And the work in \cite{warutai2019optimal} takes advantage of transparency and immutability of blockchain to implement an accounting model suitable for renewable energy. Experimental results show that it can promote the development of distributed energy resources in the future.
In \cite{AlamTowards}, security problems caused by one-way communications are addressed, where consumers can only accept passively and cannot respond their demands. Alam et al.~propose a double-chain structure to optimize the microgrid along with a negotiation algorithm. The algorithm generates smart contracts and accomplishes negotiations, thereby promoting the optimal transactions of credit.
\subsection{Power Losses of Microgrids}\label{num4.2}
One problem of the traditional power grid is power losses. Because a power station is usually far away from its users, it leads to a large amount of electricity loss during the transmission process. In a microgrid, power loss problems can be effectively alleviated because the microgrid is an energy ecosystem within a small area with a short transmission distance. However, the increasing number of microgrids brings other power loss problems \cite {sanseverino2017blockchain, hernandez2005fuel}. Power losses may also be resulted from issues such as old machines, coal consumption for power supply, and the superposition of energy transactions \cite{di2018technical}. We investigate the research works based on blockchain and find that it is indeed feasible to use blockchain-based solutions for solving the power loss problem in a microgrid.
With the deep promotion of distributed microgrids and the massive use of renewable energy sources, the power loss problem of renewable energy sources have become more serious.
Based on blockchain and smart contracts, Danzi et al.~\cite{danzi2017distributed} tackle the power loss problem caused by the high penetration rate of distributed energy resources. For renewable energy, the work in \cite {khalid2020blockchain} explores the challenges facing Pakistan's power imbalance, and proposes an anti-power solution based on Ethereum. Shishkov et al.~\cite{shishkov2019microgrid} employ blockchain to address the power system operation problem of Russian Federation.
Di Silvestre et al.~\cite {di2019ancillary} propose a comprehensive framework of microgrids. This work considers the issue of providing auxiliary services. In the proposed scheme, each photovoltaic generator plays a role in voltage regulation, and blockchain can increase the security of the system. Experimental results show that the scheme can effectively improve the voltage regulation, thereby enhancing the robustness of the system.
\subsection{Microgrid Automation}\label{num4.3}
Kang et al.~\cite {KangA} use PoW and smart contracts to implement a smart home system based Ethereum. It aims to automate the processing of renewable energy and eliminates the dependence of third parties (shown in Figure. \ref{fig:mic}).
Similarly, Afzal et al.~\cite{afzal2019blockchain} propose a microgrid solution for smart communities, where smart contracts can realize automation for the system.
\begin{figure}[htbp]
\includegraphics[width=.95\linewidth]{mic.pdf}
\caption{ Microgrid energy interaction based on the scheme in \cite {KangA}. }
\label{fig:mic}
\end{figure}
In \cite{xue2017blockchain}, the authors mainly address the complexity of the traditional power grid nanagement process and the high labor cost. The scheme proposed in Kounelis et al.~\cite{kounelis2017fostering} is committed to solving the problems of high cost of microgrid systems and low user profits. Both solutions \cite{xue2017blockchain, kounelis2017fostering} use smart contracts to achieve automated operations, while greatly reducing operating costs. Furthermore, Gazafroudi et al.~\cite{gazafroudi2020islanded} propose an automatic payment scheme that avoid labor and third parties, thus greatly reducing cost. However, it is not suitable for large-scale application scenarios.
\subsection{Power Distribution of Microgrid}\label{num4.4}
Similar to the traditional large-scale power grids, the energy distribution of a microgrid is optimized to meet demand response, improve service quality, and optimize price.
Alessandra et al.~\cite{AlessandraSmarter} intend to optimize the microgrid process to provide better services (shown in Figure. \ref{fig:mic-dis}).
\begin{figure}[htbp]
\includegraphics[width=.95\linewidth]{mic-dis.pdf}
\caption{Illustration of the scheme in \cite{AlessandraSmarter} in the field of microgrid.}
\label{fig:mic-dis}
\end{figure}
Like \cite{AlessandraSmarter}, \cite{li2019decentralized} meets the needs of users, and proposes a demand response mechanism for IoT (Internet of Things). This work employs game theory to achieve a more reasonable energy distribution and price.
Li and Nair \cite {li2018blockchain} propose a blockchain based system to generate and trade power, which can optimize price and cost in microgrids. It employs LSB (Linear Supply Bidding) to prove the equilibrium of the electricity market competition. The authors also introduce a DSB (Distributed Supply Bidding) algorithm to achieve reasonable pricing.
Dimobi et al.~\cite {dimobi2020transactive} analyze several energy market coordination schemes and propose an energy exchange scheme based on Hyperledger Fabric, which can increase the enthusiasm of users for energy investment and optimize cost.
How to achieve optimal power flow in microgrids has also attracted interests of researchers. The work in~\cite{munsing2017blockchains} uses the ADMM algorithm to ensure operational constraints and fair payment without relying on third-party agencies. Blockchain and smart contracts are employed to verify the effectiveness of the solution. With the smart contract technology, a power grid can run automatically and realize automatic payment of transactions.
Van Leeuwen et al.~\cite {van2020integrated} also incorporate the ADMM algorithm to optimize the system's power consumption. The solution utilizes smart contracts and eliminates third parties. In this scheme, the smart contract helps to distribute data to all participants and execute the ADMM algorithm.
\subsection{ Summary }
We summarize the key technologies and blockchain-related attributes of the existing works in this subsection. The specific results are presented in Table \ref{micogrid22}.
\begin{table*}[htb]
\begin{threeparttable} \footnotesize
\caption{Blockchain For Micogrid Management}
\vspace{-0.3cm}
\label{micogrid22}
\begin{tabular}{c c c c c c}
\toprule
Ref.
& Consensus & Incentive
& Blockchain & Key technique \\
\specialrule{0em}{1pt}{1pt}
\midrule
Alam\textit{ et~al.}\cite{AlamTowards}
& - & No
& Double-chain & - \\
\specialrule{0em}{1pt}{1pt}
Jeon\textit{ et~al.}\cite {jeon2019study
& - & No
& Hybrid & Plasma trading \\
\specialrule{0em}{1pt}{1pt}
Wang\textit{ et~al.}\cite {wang2019cybersecurity
& - & Yes
& Private & Unscented trans., DAG \\
\specialrule{0em}{1pt}{1pt}
Zhang \textit{ et~al.}\cite{zhang2019privacy}
& - & No
& Consortium & Blind signature \\
\specialrule{0em}{1pt}{1pt}
Danzi\textit{ et~al.}\cite {danzi2017distributed
& PoW & Yes
& Ethereum & Fairness Control \\
\specialrule{0em}{1pt}{1pt}
Silvestre\textit{ et~al.}\cite {di2019ancillary
& - & Yes
& Tendermint & Glow-worm swarm \\
\specialrule{0em}{1pt}{1pt}
Khalid\textit{ et~al.}\cite {khalid2020blockchain}
& PoA &
& Ethereum & Asymmetric crypto. \\
\specialrule{0em}{1pt}{1pt}
Kang\textit{ et~al.}\cite{KangA
& PoW & No
& Ethereum & - \\
\specialrule{0em}{1pt}{1pt}
Kounelis\textit{ et~al.}\cite {kounelis2017fostering}
& PoW & Yes
& Ethereum & Helios Coin \\
\specialrule{0em}{1pt}{1pt}
Xue\textit{ et~al.}\cite {xue2017blockchain}
& PoW & No
& Ethereum & - & - \\
\specialrule{0em}{1pt}{1pt}
Afzal\textit{ et~al.}\cite {afzal2019blockchain
& - & Yes
& Ethereum & Branch and Bound \\
\specialrule{0em}{1pt}{1pt}
Alessandra\textit{ et~al.}\cite {AlessandraSmarter}
& PoW & No
& Private & Machine learning \\
\specialrule{0em}{1pt}{1pt}
Dimobi\textit{ et~al.}\cite {dimobi2020transactive}
& PoW & Yes
& Hyperledger & - \\
\specialrule{0em}{1pt}{1pt}
Van\textit{ et~al.}\cite {van2020integrated}
& - & No
& Ethereum & ADMM \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite {li2019blockchain
& PoW & Yes
& Permissioned & Asymmetric crypto. \\
\specialrule{0em}{1pt}{1pt}
Eric\textit{ et~al.}\cite {munsing2017blockchains
& - & No
& Ethereum & ADMM \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite {li2019decentralized
& PoW & No
& - & Stackelberg game \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite{li2018blockchain}
& - & Yes
& Ethereum & Bidding algo. \\
\bottomrule
\end{tabular}
\end{threeparttable} \normalsize
\end{table*}
Having reviewed the aforementioned blockchain-based solutions for microgrids,
we obtain the following findings:
\begin{itemize}
\item Microgrids are subsystems of a large power grid, and its users are often communities or islands with relatively close daily connections, which conforms to the characteristics of the consortium blockchain. Therefore, we believe that the integration of blockchain with microgrids is more suitable than with a large power grid.
\item In terms of security, most solutions are based on the trust platform provided by blockchain, and incorporate cryptographic technologies such as blind signature and asymmetric encryption.
\item Power loss is mostly caused by long-distance transmissions and the excessive use of new renewable energy. The regional nature of a microgrid and the distributed structure of a blockchain can solve these problems. Some solutions combine machine learning and ADMM technologies to aggregate and analyze the energy usage data of a microgrid to achieve effective monitoring and forecasting of electricity.
\item Blockchain can provide transparent and unchangeable records for grid users. Its consensus and incentive measures can be exploited to reduce energy consumption. Smart contracts can provide automated management and eliminate trusted third parties. Based on the above characteristics, many schemes combine game theory and the financial characteristics of blockchain to establish auctions and dynamic pricing markets, making energy distribution more reasonable.
\item However, We find that some solutions for security enhancement still have certain flaws and the encryption algorithm used is relatively simple. Advanced techniques such as identity-based encryption, attribute-based encryption, and RSA encryption are rarely adopted in current research. Privacy protection is also mainly guaranteed by signature-based access control, which lacks in-depth research.
\end{itemize}
All in all, the integration of blockchain and microgrid is a complementary combination of two technical fields and there is still much room for further development.
\section{Blockchain For Electric Vehicles}\label{Vehicle}
\subsection{Current Issues in Electric Vehicles}
The scope of smart grids continues to expand, and electric vehicles are gradually becoming popular, especially in countries with policy supports such as China and Japan. Due to high mobility, electric vehicles can be seen as mobile terminals of the power grid, thereby realizing important functions. This is the so-called V2G technology, which can improve the performance of the power grid in terms of reliability, efficiency, and stability. However, electric vehicles are not perfectly integrated with smart grids, and there exist a series of problems such as energy shortage, security risks, and privacy leak.
As a distributed consensus ledger, blockchain can solve the problems of trust and privacy and demonstrate a great potential in the mobile field. We survey solutions on the combination of electric vehicles and blockchain from two perspectives: charging/discharging of electric vehicles and energy transmissions of electric vehicles. We elaborate on each category and summarize and compare related technologies in the end.
\subsection{Charging and Discharging of Electric Vehicles}\label{num2.1}
The blockchain technology can help to address the problems of excessive charging load and unstable voltage of electric vehicles. The works \cite {li2019iterative, li2020consortium} propose two models based on the consortium blockchain technology. It is worth mentioning that \cite{li2019iterative} presents a two-layer model, which can realize optimization of charging/discharging transactions. This scheme also considers a mixed integer programming problem, and proposes a new KHA (Krill Herd Algorithm) to solve the problem.
Moreover, the purpose of the scheme \cite{chen2020blockchain} is to forecast the power consumption in future to efficiently distribute power. It is built on an Ethereum platform, and its core is an auction quotation system and the corresponding power allocation algorithm.
In addition to load issues, the combination of blockchain and V2G can realize cost optimization.
Liu et al.~\cite{liu2018adaptive} optimize cost through the combination of IOEA (Iceberg Order Execution Algorithm) and smart contracts. Similarly, the authors \cite{liu2018blockchain} put forward their own design to optimize cost and price. The technical core of this solution lies in a JD (jump-diffusion) process for energy exchange. The innovation in \cite{ping2020coordinating} is the combination of Ethereum and Lagrange relaxation. This solution can not only maximize charging profit but also prevent malicious collusion attacks.
Blockchain's consensus mechanism and smart contracts can guarantee security for electric vehicle charging. The design goal of EVChain \cite{firoozjaei2019evchain} is to solve the privacy leakage and hidden security risks in the charging process. It uses the $k$-anonymity technique combined with blockchain, and the electric vehicles in this scheme can share credits based on reputation. Liu et al.~\cite{liu2019proof} propose a novel consensus algorithmic mechanism to resist attacks. This consensus is based on the ONPoB and is specially suitable for V2G environments.
Huang et al.~\cite{HuangLNSC} develop the LNSC system, a perfect security access mechanism without the dependence on third parties. In the aspect of computation costs, this system has a great advantage over traditional ones.
In \cite{zhou2019design}, Zhou et al.~conduct experiments based on an Ethereum platform, which greatly improves security and increases system efficiency.
As an emerging special application field, smart communities face certain problems in private charging piles and demand response \cite{wang2018contract}.
Like \cite{wang2018contract}, the community often determines the effective range for vehicle charging and discharging. In addition to security, the decentralization of blockchain can facilitate electric pile sharing, thereby improving the utilization efficiency of electric piles~\cite{hou2017resolution}. Su et al.~\cite{su2018secure} investigate the community residents' demand response to electric vehicle charging. They apply the PBFT consensus and smart contracts to utilize renewable energy, making the charging process more rational.
\subsection{Energy Transmissions Based on Electric Vehicles}\label{num2.2}
A smart grid can provide transit points for energy transactions through electric vehicles, which greatly improves the distance and flexibility of energy transmissions \cite{GuoDistributed} (as illustrated in Figure \ref{fig:car}). It can provide energy for edge base stations or nodes to solve the problem of insufficient energy at edges when necessary. Blockchain can provide trust, security, and privacy for the energy transmissions of V2G \cite{Kim2018Blockchainfor}.
\begin{figure}[htbp]
\includegraphics[width=.95\linewidth]{car-charge.pdf}
\caption{The model in \cite{GuoDistributed} involves three components: vehicle, blockchain network and power grid.}
\label{fig:car}
\end{figure}
Electric vehicles, as movable energy transmission nodes, combined with blockchain technology can reduce the cost of long-distance transmissions.
Ref. \cite {kang2017enabling} proposes a power trading model based on a consortium blockchain. The model mainly implements local power transactions between PHEVs in a smart grid. In addition, it employs an auction mechanism to solve the transaction pricing problem. Hassija et al.~\cite{hassija2020blockchain} apply DAG to improve the efficiency of system operation, and meanwhile combine it with game theory to achieve rational pricing between vehicles and the power grid. Liu et al.~\cite{liu2019electric} make use of the transparency and immutability properties of Ethereum to implement a price auction mechanism. Smart contracts in this scheme are used to realize dynamic pricing. Ref. \cite {xue2020design} employs game theory to balance the needs and interests of energy trading platforms.
The works \cite{zhou2019secure, zhou2018blockchain} combine blockchain, edge computing, and game theory to achieve fair pricing. On this basis, an ICC (iterative convex-concave) algorithm is designed to maximize the welfare of the entire system.
The combination of electric vehicles and blockchain can inherit their respective advantages and achieve demand response during energy transmissions.
The work \cite{islam2020blockchain} discusses the situation where renewable energy resources cannot meet vehicle needs at peak times; and the authors put forward their own opinions at the end of the paper. In \cite{liu2019peer}, the main contribution is to propose a new PoB consensus algorithm to stimulate demand response. The new consensus algorithm can be implemented on Ethereum and combined with smart contracts to eliminate third parties.
Moreover, blockchain can provide security for V2G.
Sheikh et al.~\cite{sheikh2019secured} mainly address the false data attack problem. The proposed scheme improves the PBFT consensus and confirms the feasibility of the scheme based on the standard IEEE33 bus system. Wang et al.~\cite{wang2019bsis} propose an incentive mechanism called BSIS for smart cities. This mechanism combines a credit-based consensus and the cryptocurrency technology, which can effectively prevent external attacks.
In addition to resist external attacks, the security risks brought by third parties cannot be ignored.
Bansal et al.~\cite{BansalSmartChain} point out that the blockchain technology can provide privacy protection, contribute an effective consensus mechanism and eliminate dependence on trusted third parties. As the computing resources of electric vehicles are limited, the authors design SmartChain, a smart and scalable ledger framework inspired by blockchain.
Sun et al.~\cite{sun2020blockchain} employ a consortium blockchain to eliminate third parties. Compared with other schemes, this approach improves the Byzantine algorithm based on DPoS consensus. The authors also design a new auction mechanism to optimize prices.
Both \cite{jindal2019survivor} and \cite{chaudhary2019best} propose security solutions based on SDN networks. The solution reported by Jindal et al.~\cite{jindal2019survivor} is named SURVIVOR, which randomly selects validators to perform security verification on all transactions. Chaudhary et al.~\cite{chaudhary2019best} present a scheme named BEST. The security of this scheme is also verified based on special nodes. This scheme is mainly used in the field of autonomous driving.
\subsection{Summary}
We summarize the design goals, key technologies and other blockchain-related attributes of the solutions outlined in this section and report the results in Table \ref{vehicle22}.
\begin{table*}[htbp]
\begin{threeparttable} \footnotesize
\caption{Blockchain For Electric Vehicle}
\vspace{-0.3cm}
\label{vehicle22}
\begin{tabular}{c c c c c c c c c}
\toprule
Ref. & Design goal & Consensus & Blockchain & Contract \\
\specialrule{0em}{1pt}{1pt}
\midrule
Su\textit{ et~al.}\cite {su2018secure} & Security and privac
& RDBFT
& Permissioned & Used \\
\specialrule{0em}{1pt}{1pt}
Zhou\textit{ et~al.}\cite {zhou2019design} & Security and privacy
& Hybrid
& Ethereum & Used \\
\specialrule{0em}{1pt}{1pt}
Wang\textit{ et~al.}\cite {wang2018contract} & Security and privacy
& RDBFT
& Permissioned & Used \\
Liu\textit{ et~al.}\cite {liu2019proof} & Security and privacy
& PoB
& Ethereum & Used \\
\specialrule{0em}{1pt}{1pt}
BEST\cite {chaudhary2019best} & Security and privac
& PoW
& Consortium & Unused \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite {li2019iterative} & Load optimization
& Kafka
& Hyperledger & Unused \\
\specialrule{0em}{1pt}{1pt}
Li\textit{ et~al.}\cite {li2020consortium} & Load optimization
& Kafka
& Hyperledger & Unused \\
\specialrule{0em}{1pt}{1pt}
Kang\textit{ et~al.}\cite {kang2017enabling} & Cost optimization
& PoW
& Consortium & Unused \\
\specialrule{0em}{1pt}{1pt}
Liu\textit{ et~al.}\cite {liu2019electric} & Cost optimization
& PoA
& Ethereum & Used \\
\specialrule{0em}{1pt}{1pt}
SmartChain\cite {BansalSmartChain} & Third party
& Smartchain & Smarrchain & Unused \\
\specialrule{0em}{1pt}{1pt}
Sun\textit{ et~al.}\cite {sun2020blockchain} & Third part
& DPoSP
& Consortium & Unused \\
\specialrule{0em}{1pt}{1pt}
Zhou\textit{ et~al.}\cite {zhou2019secure} & Edge computing
& PoW
& Consortium & Used \\
\specialrule{0em}{1pt}{1pt}
Zhou\textit{ et~al.}\cite{zhou2018blockchain} & Edge computing
& PoW
& Consortium & Used \\
\specialrule{0em}{1pt}{1pt}
Sheikh\textit{ et~al.}\cite {sheikh2019secured} & Anti-attac
& Byzantine
& Ethereum & Used \\
\specialrule{0em}{1pt}{1pt}
BSIS\cite {wang2019bsis} & Anti-attac
& PoR
& Permissioned & Unused \\
\specialrule{0em}{1pt}{1pt}
Liu\textit{ et~al.}\cite {liu2019peer} & Demand response
& PoB
& Ethereum & Used \\
\bottomrule
\end{tabular}
\end{threeparttable} \normalsize
\end{table*}
Having reviewed the aforementioned blockchain-based solutions for V2Gs,
we obtain the following findings:
\begin{itemize}
\item We conclude that using blockchain-based solutions to augment electric vehicles can provide safe charging/discharging and power transmission services. In the process of charging/discharging, the use of blockchain can make the private electric piles public and improve the resource utilization of the entire system.
In the process of energy transmission, vehicles can transmit electrical energy to remote areas during peak power periods, ensuring the stability of the entire system. Blockchain can provide vehicles with a transparent energy management platform.
\item More specifically, the combination of blockchain and V2G is mainly to deal with security risks, demand response, and process optimization. Security risks are mainly caused by external attacks, privacy leaks and third parties. Many solutions based on blockchain combine ECDSA, asymmetric encryption, and other cryptographic algorithms to greatly improve the security level of the system. Process optimization can be considered from the aspects of cost optimization, load optimization, service optimization, etc. Blockchain's own trust and financial attributes can attract a number of vehicles while reducing price fluctuations. Many schemes combine game theory and various improved algorithms on this basis to further realize optimization.
\item With the rapid increase in electric vehicles and the high complexity of traffic management, traditional V2G management consumes a lot of manpower and costs. The use of smart contracts can replace labor and realize automated operations of the system.
\item The diversity of blockchain platforms can provide application scenarios for different solutions. Due to the real-time requirements of V2G and the limitation of vehicle computing power, many solutions have abandoned the traditional PoW consensus and proposed various consensus algorithms such as Kafka, DPoSP and PoB, which effectively improve the efficiency of the system.
\item However, the problem of blockchain transmission delay in long-distance networks still exists. Whether the vision of blockchain joining V2G can be supported by most people also requires deeper thinking and further research.
\end{itemize}
\section{Current Projects and Initiatives}\label{resource}
For completeness of this survey, we summarize current projects and initiatives that apply blockchain to smart grids.
\textbf{Carbon Asset Development Platform (C.P) \cite{EnergyChainTechnology}}. This platform \cite{EnergyChainTechnology} is proposed by Energy Chain Technology Corporation and IBM, which is mainly used in the carbon market to achieve energy saving, emission reduction and green energy. The blockchain is intended to realize the traceability and transparent management of data, so as to further realize the validity and authenticity of information.
\textbf{Russian Carbon Market (R.C.M) \cite{IPCI}}. In response to the call for green energy, Russian company DAO IPCI employs blockchain technology to facilitate the business of local carbon industry companies. This platform is safe and open, and all nodes can monitor the data on the blockchain. In the implementation process, Hinprom takes the lead in experimental studies.
\textbf{Israel Microgrid (I.M) \cite{Israelcleanenergy}}. The Israel microgrid makes use of blockchain to realize the issuance of green certificates. The platform proposed by Greeneum can guarantee the correctness of energy production and then issue corresponding green certificates. On this basis, the project combines artificial intelligence to achieve electricity prediction and balance, making P2P transactions more reasonable
\textbf{Energy Web Foundation (E.W.F) \cite{Energyweb}}. Energy Web Foundation takes advantage of the underlying structure of Ethereum to integrate more than 100 power companies. Recently, it proposes EW-DOS to create a decentralized operating system for accelerating the construction of low-carbon power systems
\textbf{Tanzania X-Solar \cite{xsolarsystem}}. The X-Solar project is provided by M-PAYG with the corresponding technical support. The purpose of X-Solar is to solve the power shortage problem in developing countries of Africa such as Tanzania. It can provide solar power monitoring, as well as transaction and payment functions based on the blockchain
\textbf{Iberdrola Energy Tracking (I.E.T) \cite{Iberdrola}}. This project is established by Iberdrola and Kutxa investment, and has been successfully tested. It employs the blockchain technology to realize the management and monitoring of energy trading processes. In addition, blockchain is used to provide trusted energy certificates
\textbf{Poseidon Foundation (P.F) \cite{Poseidon}}. The Poseidon foundation focuses on reducing carbon emissions. It cooperates with Ben \& Jerry's to combine carbon trading with ice cream, effectively reducing carbon emissions. Stellar blockchain can improve the transparency of transactions. At present, it is actively working with the LCC (City Council of Liverpool) to build a carbon credit mechanism for reducing carbon emissions and protecting the environment of London
\textbf{TenneT and IBM Pilot (T.I.P) \cite{TenneT}}. In Germany, TenneT, IBM and Sonnen have cooperated to realize distributed storage of electricity. This effectively solves the high-cost management caused by excess energy supply. This project utilizes Hyperledger to record and trace the corresponding data information
An application of this project in electric vehicle charging in the Netherlands has also achieved certain results.
\textbf{Brooklyn Microgrid (B.M)\cite{Brooklynmicrogrid}}. In Brooklyn, the local power grid company (LO3) proposes to combine Ethereum with the smart grid and implement it locally. It allows residents to use solar energy to generate electricity and trade to obtain token rewards
\textbf{IDEO CoLab Solar Panel (I.C.S.P) \cite{IDEOCoLab}}. This is a joint project between IDEO CoLab and Filament. They aim to combine solar panels with the Internet of Things, and the bottom layer uses blockchain and smart contracts for monitoring, energy production, and automatic transactions
\textbf{Global Grid (G.G) \cite{globalgrid}}. Global Grid cooperates with a local Mexican developer to build a solar blockchain platform and a large photovoltaic grid. It is understood that this project also intends to expand to other renewable resource areas such as geothermal
\textbf{Electron \cite{Electron}}. Electron conducts community experiments in London based on Ethereum and IPFS. Residents can trade electricity with their neighbors through Electron
\textbf{Scanergy \cite{SCANERGY}}. This project is launched by the European Union, which aims to achieve green energy trading. Scanergy is based on Ethereum and smart contracts, and uses NRGcoin as the digital currency. It also makes use of the cloud market to derive a transparent energy auction platform
\textbf{Fremantle Project (F.P) \cite{Fremantlepowerledger}}. The Fremantle Project aims to optimize the use of distributed energy and water resources in Australia. Blockchain is employed to collect digital information of various energy sources to ensure its integrity.
\textbf{De La Salle University Microgrid (D.L.S.U.M) \cite{Philippinesmiero}}. This is a microgrid system constructed inside the De La Salle University. This system uses photovoltaic power generation, combined with the Qtum blockchain system for P2P transactions. The actual effect shows that the system can effectively realize the optimization of cost, environment, and safety
\textbf{PONTON Enerchain \cite{Ponton}}. PONTON is the underlying blockchain platform of Enerchain, which supports the electricity trading and transmission of more than 30 energy companies in Europe
\textbf{Solar Bankers (S.B) \cite{Solarbankers}}. Solar Bankers is a distributed solar trading platform proposed by Singapore. In Solar Bankers, blockchain is adopted to guarantee the security and transparency of solar energy production and transaction processes. Users can obtain the digital currency SunCoin by trading energy
\textbf{Rotterdam Heat Network (R.H.N) \cite{Eneco}}. CGI \& Eneco designs a heat energy trading platform based on Tendermint and conducts a field test in Rotterdam
\textbf{Austria Grid Project (A.G.P) \cite{GridSingularity}}. Grid Singularity is an Austrian energy trading platform. It has conducted experiments on the smart grid locally. It is implementing an open and secure distributed smart grid system
\textbf{Interbit \cite{BTLgroup}}. Interbit is a blockchain system created by BTL Group. It has conducted field experiments in the United Kingdom, Canada, and other places, and achieves good results in energy trading
\textbf{TEPCO \& Blockchain (T.B) \cite{electronTEPCO}}. TEPCO is a Tokyo-based power company that has worked with multiple companies such as Grid+ to create a new type of distributed smart grid. It creates a P2P solar trading system together with Conjoule.
Later, TEPCO joins the EWF (Energy Web Foundation) to accelerate the development of the global energy blockchain and collaborates with Electron on the decentralized transformation of smart grids
\textbf{Hoog Dalem LEF (H.D.L) \cite{energy21}}. It is a pilot project jointly cooperated by Stedin, Energy21, ABB, and iLeco. In this project, residents conduct P2P transactions and settle through blockchain
\textbf{SunContract \cite{SunContract}}. SunContract is a P2P energy trading platform supported by the European Union. The test in Slovenia shows that it can effectively reduce energy cost
\textbf{Share \& Charge (S.C) \cite{sharecharge}}. The main application area of the Share \& Charge platform is V2G. It makes the free private charging piles public. Each mobile app acts as an OCN node. Blockchain makes transactions transparent and open, and enhances the credibility of the system
\textbf{Oxygen Initiative (O.I) \cite{oxygeninitiative}}. Similar to Share \& Charge, Oxygen Initiative is also used in the charging field of vehicles. The predecessor of the Oxygen Initiative is the vehicle e-wallet project of RWE and ZF company. It is based on the Ethereum platform, which can provide secure payment services for vehicle expenses, and has been simulated in the US electricity market
\textbf{JuiceNet \cite{JuiceNet}}. JuiceNet is a marketing platform created by eMotorWerks. The platform realizes the charging function of P2P based on blockchain. Electric vehicles can be charged via other people's private charging piles according to demand, and pay the corresponding fees. This not only provides charging convenience for vehicles, but also encourages more residents to purchase charging piles to earn fees
We summarize the projects and initiatives mentioned above in Table~\ref{Resource}.
\begin{table*}[htbp]
\begin{threeparttable} \footnotesize
\caption{Comparison: Practical Applications and Public Experiments}
\vspace{-0.3cm}
\label{Resource}
\begin{tabular}{c c c c c c}
\toprule
Design Goal & Ref. & Organization & Energy type & Blockchain \\
\specialrule{0em}{1pt}{1pt}
\midrule
& C. P \cite{EnergyChainTechnology} & - & ECTC/ IBM & Carbon & - \\
\specialrule{0em}{1pt}{1pt}
& R.C.M \cite{IPCI} & Russian & DAO IPCI & Carbon & - \\
\specialrule{0em}{1pt}{1pt}
& I.M \cite{Israelcleanenergy} & Israel & Greeneum & Green energy & Ethereum \\
\specialrule{0em}{1pt}{1pt}
& E.W.F \cite{Energyweb} & Switzerland & Energy Web Foundation & Electricity & Ethereum \\
Management & X-Solar \cite{xsolarsystem} & Tanzania & M-PAYG & Solar & - \\
\specialrule{0em}{1pt}{1pt}
& I.E.T \cite{Iberdrola} & Spain & Iberdrola/ Kutxa & Renewable & - \\
\specialrule{0em}{1pt}{1pt}
& P.F \cite{Poseidon} & Liverpool & LCC & Carbon & Stellar \\
\specialrule{0em}{1pt}{1pt}
& T.I.P \cite{TenneT} & Germany & TenneT/ IBM/ Sonnen & Electricity & Hyperledger \\
\specialrule{0em}{1pt}{1pt}
\midrule
& B.M \cite{Brooklynmicrogrid} & Brooklyn & LO3 & Solar & Ethereum \\
\specialrule{0em}{1pt}{1pt}
& I.C.S.P \cite{IDEOCoLab} & - & IDEO CoLab/ Filament & Solar & - \\
\specialrule{0em}{1pt}{1pt}
& G.G \cite{globalgrid} & Mexico & SAPI de CV & Solar & - \\
\specialrule{0em}{1pt}{1pt}
& Electron \cite{Electron} & London & Electron & Electricity & Ethereum \\
\specialrule{0em}{1pt}{1pt}
& Scanergy \cite{SCANERGY} & Europe & - & Green energy & Ethereum \\
\specialrule{0em}{1pt}{1pt}
& F.P \cite{Fremantlepowerledger} & Australia & Australian government & Distributed & - \\
\specialrule{0em}{1pt}{1pt}
& D.L.S.U.M \cite{Philippinesmiero} & Philippines & Energo Labs & Solar & Qtum \\
\specialrule{0em}{1pt}{1pt}
Trading & Enerchain \cite{Ponton} & Germany & PONTON & Electricity & Tendermint \\
\specialrule{0em}{1pt}{1pt}
& S.B \cite{Solarbankers} & Singapore & Solar Bankers & Solar & Skyledger \\
\specialrule{0em}{1pt}{1pt}
& R.H.N \cite{Eneco} & Rotterdam & CGI \& Eneco & Heat & Tendermint \\
\specialrule{0em}{1pt}{1pt}
& A.G.P \cite{GridSingularity} & Austria & Grid Singularity & Electricity & - \\
\specialrule{0em}{1pt}{1pt}
& Interbit \cite{BTLgroup} & UK & BTL Group & Electricity & Interbit \\
\specialrule{0em}{1pt}{1pt}
& T.B \cite{electronTEPCO} & Tokyo & Conjoule/ Electron & Electricity & Ethereum \\
\specialrule{0em}{1pt}{1pt}
& H.D.L \cite{energy21} & Netherlands & Energy21/ Stedin & Electricity & Quasar \\
\specialrule{0em}{1pt}{1pt}
& SunContract \cite{SunContract} & Slovenia & SunContract & Electricity & Ethereum \\
\midrule
& S.C \cite{sharecharge} & Germany & Innogy/ Slock.it & Electricity & Ethereum \\
\specialrule{0em}{1pt}{1pt}
V2G & O.I \cite{oxygeninitiative} & US & Oxygen Initiative & Electricity & Ethereum \\
\specialrule{0em}{1pt}{1pt}
& JuiceNet \cite{JuiceNet} & North America & eMotorWerks & Electricity & - \\
\specialrule{0em}{1pt}{1pt}
\bottomrule
\end{tabular}
\end{threeparttable} \normalsize
\end{table*}
|
\section{Introduction}
The understanding of solid-liquid interfaces is of major importance for a sustainable energy future.\cite{Chu2012, Larcher2015} In particular, electrode-electrolyte interfaces are central for many processes, from the electrocatalytic water splitting for the production of green hydrogen to energy storage in lithium ion batteries supplying, e.g., portable electronic devices and electric vehicles.\cite{Bruce2008, Goodenough2010, Suntivich2011, Seh2017, Suen2017} For this purpose, a prominent material is the lithium manganese oxide spinel Li$_x$Mn$_2$O$_4$, with $0\leq x\leq2$, which is a frequently used positive electrode material in lithium ion batteries but can also be employed as electrocatalyst for the oxygen evolution reaction (OER) representing the limiting step of water splitting.\cite{Thackeray1983, Thackeray1997, Cady2015, Koehler2017} Beyond conventional lithium ion batteries containing organic electrolytes, the Li$_x$Mn$_2$O$_4$-water interface has recently received increasing attention regarding the development of environment-friendly aqueous rechargeable lithium ion batteries offering improved safety combined with higher ionic conductivity and lower-cost production.\cite{Kim2014, Alias2015}
Stoichiometric LiMn$_2$O$_4$ contains a one-to-one ratio of Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ ions.\cite{Takahashi2003, Akimoto2004, Piszora2004} The electrochemical incorporation or removal of Li ions during battery discharging and charging changes this ratio by the reduction or oxidation of Mn ions, respectively, ensuring overall charge neutrality\cite{Hunter1981, Thackeray1983} and allowing to control the electronic structure of the bulk material. This control is particularly important for the OER, as the Mn oxidation states are considered to be central for this process.\cite{Stamenkovic2006} Apart from the overall composition of the material, the OER activity is also determined by the details of the solid-electrolyte interface whose geometric and electronic structure as well as the atomic composition can be substantially different from the bulk.\cite{Gauthier2015, Schoenewald2020}
The active sites of electrocatalytic reactions are often embedded in a complex environment consisting of the -- possibly reconstructed -- solid surface and the electrical double layer. Small particles and porous materials have large surface-to-volume ratios, which are beneficial for a high activity, but the exposed surfaces can exhibit very different reactivities. The identification of active sites is therefore essential in a bottom-up approach for the design of improved catalysts.\cite{Hammer2000, Jaramillo2007, Norskov2009, Behrens2012, Jiao2015}
A practical challenge when using Li$_x$Mn$_2$O$_4$ as battery material is capacity fading, which is related to the disproportionation of Mn$^\mathrm{III}$ ions at the interface and the dissolution of the resulting Mn$^\mathrm{II}$ ions.\cite{Benedek2012, Bhandari2017, Leung2017} Consequently, the identification of tactics for controlling the Mn oxidation states at the interface is important for the construction of batteries with improved charge/discharge cycles and enhanced lifetime. Hence, to unravel the relationship between composition and reactivity, a comprehensive understanding of the geometric and electronic structure as well as of the dynamics and reactions at the interface is required.\cite{Hirayama2010, Leung2013, Benedek2017}
These insights can be gained in principle in computer simulations, but complex interface systems still pose a significant challenge as they require a first principles-quality description.\cite{Schaub2001, Rossmeisl2007, Bajdich2013, Tocci2014, Friebel2015} The electronic structure of Li$_x$Mn$_2$O$_4$ with coexisting Mn$^\mathrm{IV}$ and Jahn-Teller distorted high-spin Mn$^\mathrm{III}$ ions in the bulk as well as high-spin Mn$^\mathrm{II}$ ions at the interface is, however, difficult to describe by established methods like density functional theory (DFT). For a correct representation at least the level of the generalized gradient approximation including an additional Hubbard-like term for on-site Coulomb interactions (GGA$+U$) or a hybrid functional containing a fraction of exact Hartree-Fock exchange is needed.\cite{Karim2013, Kumar2014, Lee2016, Warburton2016, Eckhoff2020} A recent hybrid DFT benchmark of lithium manganese oxides showed that on-site Hartree-Fock exchange terms yield a correct description of partially filled shells of localized d electrons.\cite{Eckhoff2020} However, up to now ab initio molecular dynamics simulations using GGA$+U$ or hybrid DFT functionals could only be performed for rather small Li$_x$Mn$_2$O$_4$-water model systems containing a few hundred atoms on picosecond time scales due to the large computational effort.\cite{Leung2012, Choi2018, Intan2019, Okuno2019, Zhou2020} To consider the interplay of a variety of different structural motifs with a liquid solvent, picosecond time scales are not sufficient. For instance, for the equilibration of Li$_x$Mn$_2$O$_4$-water interfaces including the formation of hydroxide layers, electrical double layers, and/or strongly bound water at the interface as well as to obtain reliable statistics for elementary steps of proton transfer (PT) reactions and hydrogen bond networks significantly larger length and time scales are required.
Machine learning potentials (MLP) combine the efficiency of simple empirical potentials with the accuracy of quantum mechanics allowing to meet these requirements.\cite{Behler2016, Bartok2017, Noe2020, Behler2021a} Consequently, various MLPs have been developed, e.g., for water\cite{Morawietz2016, Cheng2016, Zhang2018, Cheng2019} and different solid-liquid interface systems.\cite{Wen2019, Artrith2019, Pun2020, Schran2021} In previous works we could further show that a high-dimensional neural network potential (HDNNP),\cite{Behler2007, Behler2014, Behler2015, Behler2017, Behler2021} a frequently used type of MLPs, is applicable to bulk materials containing transition metal ions in different oxidation states\cite{Eckhoff2020a, Eckhoff2020b, Eckhoff2021} and in different magnetic orders.\cite{Eckhoff2021} The ability to represent different oxidation states is mandatory for studying the Li$_x$Mn$_2$O$_4$ system. As the HDNNP's underlying functional form is unbiased with respect to different interaction types, an equally reliable description for the interactions in the bulk and at the interface is obtained as has been demonstrated, e.g., for the Cu-water\cite{Natarajan2016, Natarajan2017} and ZnO-water interfaces.\cite{Quaranta2017, Hellstroem2019, Quaranta2019} Moreover, like most MLPs, HDNNPs are reactive, i.e., they are able to describe the formation and cleavage of bonds making them applicable to PT reactions omnipresent in electrochemical systems.
Apart from the simulation of the atomistic structure, machine learning algorithms can also be applied to obtain information about the electronic structure, for example, atomic charges,\cite{Artrith2011, Ghasemi2015, Unke2019, Xie2020, Ko2020} electrostatic multipole moments,\cite{Houlding2007, Bereau2015} polarizabilities,\cite{Montavon2013} and even quantum mechanical wavefunctions.\cite{Schuett2019} Moreover, our recently developed high-dimensional neural network spin (HDNNS) prediction method\cite{Eckhoff2020b} can be used to identify the oxidation and spin states of the Mn ions. This method is based on local geometric changes associated to the details of the electronic structure. Thus, in combination with an HDNNP providing the energies and forces, nowadays nanosecond time scale investigations of the geometric and electronic structure are possible for systems containing about $10^4$-$10^5$ atoms.
In this work we investigate the Li$_x$Mn$_2$O$_4$-water interface employing an HDNNP in combination with an HDNNS. Several \{100\} and \{110\} surfaces with different terminations in contact with water are investigated to unravel the spatial distribution of Mn oxidation states and oxygen species, such as oxide ions O$^{2-}$, hydroxide ions OH$^-$, neutral water molecules H$_2$O, and hydronium ions H$_3$O$^+$. Starting from atomically smooth solid surfaces in contact with a liquid water phase, the formation of hydroxide and strongly bound water layers at the interface is studied to understand the fundamental properties of the interface structure. A detailed analysis of Jahn-Teller distortions and the hydrogen bond network is provided in the Supplementary Material. The analysis of electron hopping rates between Mn ions, various PT reactions, and water species residence lifetimes under equilibrium conditions yields a detailed understanding of the kinetics and dynamics. Finally, we compare the activity of the different surfaces and sites and show that initial steps of the OER occur already spontaneously under equilibrium conditions.
\section{Methods}
In HDNNPs\cite{Behler2007, Behler2021}, which we use to compute the energies and forces driving the molecular dynamics (MD) simulations, the potential energy is constructed as a sum of atomic energy contributions $E_n^\alpha$,
\begin{align}
E(\{\mathbf{R}\})=\sum_{\alpha=1}^{N_\mathrm{elem}}\sum_{n=1}^{N_\mathrm{atoms}^{\alpha}}E_n^\alpha\ .
\end{align}
Here, $\{\mathbf{R}\}$ are the nuclear coordinates for a system containing $N_\mathrm{elem}$ elements with $N_\mathrm{atoms}^{\alpha}$ for element $\alpha$.
The individual atomic energy contributions are represented by atomic feed-forward neural networks of the form
\begin{align}
\begin{split}
E_n=&\ b_1^4+\sum_{l=1}^{n_3}a_{l1}^{34}\cdot\tanh\Bigg\{b_l^3+\sum_{k=1}^{n_2}a_{kl}^{23}\cdot\tanh\Bigg[b_k^2\\
&\,+\sum_{j=1}^{n_1}a_{jk}^{12}\cdot\tanh\Bigg(b_j^1+\sum_{i=1}^{n_G}a_{ij}^{01}\cdot G_{n,i}\Bigg)\Bigg]\Bigg\}\ .
\end{split}\label{eq:energy_contributions}
\end{align}
The architecture $n_G$-$n_1$-$n_2$-$n_3$-1 of the atomic neural networks contains an input layer with $n_G$ neurons providing a description of the atomic environment. Moreover, in this work three hidden layers with $n_1$, $n_2$, and $n_3$ neurons, respectively, and an output layer with one neuron, which yields the atomic energy contribution, are used. The activation functions are hyperbolic tangents except for the output layer, for which a linear function is employed. The weight parameters $\{a_{\mu\nu}^{\rho\sigma}\}$ and $\{b_\nu^\sigma\}$ of the atomic neural networks are optimized to accurately reproduce a training data set consisting of energies $E^\mathrm{ref}$ and atomic force components $F^\mathrm{ref}$ of reference structures obtained, for example, in DFT calculations. For each element an individual atomic neural network is constructed making the neural network parameters element-specific. For clarity, the index $\alpha$ representing this element-dependence has been omitted in the quantities in Equation \ref{eq:energy_contributions}.
The atomic neural networks are able to describe the complex relation between the atomic energies and the local chemical environments of the atoms. These environments are described by vectors of many-body atom-centered symmetry functions (ACSF)\cite{Behler2011} $\mathbf{G}_n^\alpha$ serving as structural fingerprints of the local geometry inside a cutoff sphere of radius $R_\mathrm{c}$. ACSFs represent a general transformation from the Cartesian coordinates $\{\mathbf{R}\}$ to a translationally, rotationally, and permutationally invariant structural description based on interatomic distances and angles. Moreover, for all atoms of a given element, the ACSF vectors have the same dimensionality to ensure the applicability of the trained atomic neural networks to large-scale simulations of systems containing different numbers of atoms. As the ACSFs depend only on the elements and positions of the atoms, HDNNPs are able to describe the making and breaking of bonds. The parameters defining the spatial shapes of the radial and angular ACSFs can be adjusted to optimize the performance as described in the Supplementary Material. More detailed information about HDNNPs, ACSFs, their properties, and their construction are provided in several reviews.\cite{Behler2014, Behler2015, Behler2017, Behler2021}
The HDNNS method\cite{Eckhoff2020b} is closely related to HDNNPs and employs the same atomic neural network topology. However, instead of atomic energy contributions $E_n^{\alpha}$, the atomic neural networks (Equation \ref{eq:energy_contributions}) yield the atomic spins $S_n^{\alpha}$, i.e., they provide the $S_n^\alpha(\{\mathbf{R}\})$ relation. A HDNNS exploits the observation that different oxidation states as well as high- and low-spin states of transition metal ions typically lead to structurally different local environments. Thus, like for the energies and forces in HDNNPs, the method is based on the assumption that the atomic spins and oxidation states are uniquely defined by the structure. Consequently, consistent reference data corresponding to the ground state electronic structure are mandatory for a successful construction of the HDNNS. The absolute values of atomic reference spins obtained from DFT are used for training to circumvent the issue that the electronic ground state is twofold degenerate with respect to the absolute sign of all spins. Since the present work is restricted to the magnetic ground state, we do not explicitly include the degrees of freedom related to the relative orientations of the atomic spins, but we note that magnetic HDNNPs taking these degrees of freedom into account have been proposed.\cite{Eckhoff2021}
Using both, an HDNNP and an HDNNS, in MD simulations enables a simultaneous first principles-quality representation of the geometric and the qualitative electronic structure dynamics on nanosecond time scales for systems containing thousands of atoms.\cite{Eckhoff2020a, Eckhoff2020b, Eckhoff2021}
\section{Computational Details}
For the generation of the reference data the local hybrid exchange-correlation functional PBE0r\cite{Sotoudeh2017, Eckhoff2020} including D3 dispersion corrections\cite{Grimme2010, Grimme2011} was used in collinar spin-polarized DFT calculations. PBE0r considers only on-site Hartree-Fock exchange terms yielding an accurate description of the partially filled Mn d shell with a computational effort comparable to generalized gradient approximation functionals. The Car-Parrinello Projector Augmented-Wave (CP-PAW) code (version from September 28, 2016)\cite{Bloechl1994, CP-PAW} and the DFT-D3 code (version from June 14, 2016)\cite{Grimme2010, Grimme2011} were employed using the same setup as in our previous studies.\cite{Eckhoff2020, Eckhoff2020a, Eckhoff2020b}
The HDNNP and HDNNS were constructed using the RuNNer code (versions from October 19, 2020 and December 4, 2018, respectively).\cite{Behler2015, Behler2017, RuNNer} The architecture of the atomic neural networks is 180-25-20-15-1 for all elements in the HDNNP and 180-20-15-10-1 for Mn in the HDNNS, which is the only spin-polarized atom in the system. The parameters of the 180 radial and angular ACSFs per element with the cutoff radius $R_\mathrm{c}=10.5\,a_0$ are compiled in the Supplementary Material along with the description of a generally applicable scheme to adjust the parameter $\eta$ of the ACSFs to the element-specific nearest-neighbor distances. $a_0$ is the Bohr radius.
Instead of total energies, the DFT formation energies were used for training, which were obtained from the total energies minus the sum of the atomic energies calculated for the elements in their reference states, i.e., gaseous H$_2$, body centered cubic Li, gaseous O$_2$, and $\upalpha$-Mn. In addition to the formation energies also the DFT Cartesian atomic force components were used for training the HDNNP. Of all available energies and force components 90{\%} were used in the training set to determine the neural network parameters, while the remaining 10{\%} were employed to test the predictive power and reliability for structures not included in the training process.
The DFT reference atomic spins used to train the HDNNS are the absolute values of projections of the spin density onto the one-center expansions of the partial waves using atom-centered spheres with a radius of 1.2 times the atomic covalent radius. The atomic spin is therefore equal to the absolute difference in the number of spin-up and spin-down electrons at an atom in units of the electron spin $\tfrac{1}{2}\,\hbar$. The setup of the HDNNP and HDNNS construction is described in detail in the Supplementary Material.
\begin{figure*}[htb!]
\centering
\includegraphics[width=\textwidth]{Figure_01.pdf}
\caption{Side views ($zy$ plane) of equilibrated Li$_x$Mn$_2$O$_4$-water interface simulation cells with a volume of about $25\cdot25\cdot100\,\text{\AA}^3$ including the interfaces (a) $\{100\}_\mathrm{Li}$, (b) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, (c) $\{110\}_{\mathrm{LiMnO}_2}$, and (d) $\{110\}_{\mathrm{MnO}_2}$ are shown on the left. Top views ($xy$ plane) of non-equilibrated vacuum solid surfaces are provided on the right. The scale bar applies to all panels. The atoms are shown as balls whereby Li is colored green, Mn violet, and H white. The color/size of the oxygen atoms has been chosen according to the H connectivity: O$^{2-}$ is red/large, OH$^-$ turquoise/small, H$_2$O blue/small, and H$_3$O$^+$ orange/small. The O-H bonds are shown as sticks. The surrounding black line corresponds to the boundaries of the periodic simulation cell. A side view on the $zx$ plane is provided in the Supplementary Material Figures S1 (a) to (d). This figure was created with OVITO Pro (version 3.5.0).\cite{Ovito}}\label{fig:structure}
\end{figure*}
HDNNP-driven MD simulations were performed for Li$_x$Mn$_2$O$_4$-water interface systems in the isothermal-isobaric ($NpT$) ensemble at a temperature of $T=298$\,K and a pressure of $p=1$\,bar. The simulation cells include about $N\approx6\cdot10^3$ atoms in a volume of about $25\cdot25\cdot100\,\text{\AA}^3$. The volume ratio of Li$_x$Mn$_2$O$_4$ and water is approximately 1:1 with the phase boundaries parallel to the $xy$ plane. Four different cuts of bulk Li$_x$Mn$_2$O$_4$ in contact with water have been investigated, which are the \{100\} Li-terminated surface with Mn$_2$O$_4$ in the first subsurface layer ($\{100\}_\mathrm{Li}$), the \{100\} Mn$_2$O$_4$-terminated surface with Li in the first subsurface layer ($\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$), the \{110\} LiMnO$_2$-terminated surface ($\{110\}_{\mathrm{LiMnO}_2}$), and the \{110\} MnO$_2$-terminated surface ($\{110\}_{\mathrm{MnO}_2}$). Top views of the clean surfaces as well as side views of the employed interface slab models are shown in Figures \ref{fig:structure} (a) to (d), respectively. Both surfaces of each slab are structurally identical. Therefore, the solid slab is built from bulk stoichiometric LiMn$_2$O$_4$ with non-stoichiometric Li$_x$Mn$_2$O$_4$ surfaces. For each system, three simulations starting from different H$_2$O configurations, which initially did not include OH$^-$ and H$_3$O$^+$ ions, on top of the atomically flat solid surfaces were performed for an equilibration time of 1\,ns and a subsequent acquisition time of 5\,ns. In the same way, three bulk water simulations with different initial structures were carried out using approximately cubic cells containing around $1.5\cdot10^4$ atoms and fluctuating lattice parameters between about 50 and 60\,{\AA} in the $NpT$ ensemble. To identify the oxygen and water species, each H atom has been assigned to its closest O atom.
The HDNNP-driven simulations were performed using the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS)\cite{Plimpton1995, LAMMPS} and the neural network potential package (n2p2).\cite{Singraber2019, n2p2} They were run with a timestep of 0.5\,fs applying the Nos\'{e}-Hoover thermostat and barostat\cite{Nose1984, Hoover1985} with coupling constants of 0.05\,ps and 0.5\,ps, respectively, allowing for anisotropic changes of the simulation cell. The trajectory was stored in intervals of 0.1\,ps.
\section{Results and Discussion}
\subsection{High-dimensional neural networks}
The HDNNP and HDNNS are based on a reference data set consisting of 15228 Li$_x$Mn$_2$O$_4$ bulk structures,\cite{Eckhoff2020a, Eckhoff2020b} 5143 water bulk structures, and 17597 Li$_x$Mn$_2$O$_4$-water interface structures and their PBE0r-D3 DFT energies, atomic force components, and atomic spins. The structures include 32 to 255 atoms, with the interface structures containing between 122 and 194 atoms. A detailed description of the reference data set construction and composition is provided in the Supplementary Material.
\begin{figure*}[htb!]
\centering
\includegraphics[width=\textwidth]{Figure_02.pdf}
\caption{Energy errors $\Delta E=E^\mathrm{HDNNP}-E^\mathrm{DFT}$ as a function of the reference formation energy $E^\mathrm{DFT}$ of the (a) training set and (b) test set, force component errors $\Delta F=F^\mathrm{HDNNP}-F^\mathrm{DFT}$ as a function of the reference force components $F^\mathrm{DFT}$ of the (c) training set and (d) test set, and errors of the atomic spins $\Delta S=S^\mathrm{HDNNP}-S^\mathrm{DFT}$ as a function of the reference atomic spins $S^\mathrm{DFT}$ of the (e) training set and (f) test set. The color in the heatmaps represents the density of data points based on discretizing the plotting areas into grids of $200\times125$ points.}\label{fig:HDNN_quality}
\end{figure*}
The PBE0r-D3 DFT formation energies range from $-2.15$\,eV\,atom$^{-1}$ to $-0.83$\,eV\,atom$^{-1}$ therefore spanning a fitting interval of $1.32$\,eV\,atom$^{-1}$. In Figures \ref{fig:HDNN_quality} (a) and (b) the three structure types can be identified by their formation energies between about $-2.2$ and $-1.6$, $-1.6$ and $-1.3$, as well as $-1.0$ and $-0.8$\,eV\,atom$^{-1}$ corresponding to bulk Li$_x$Mn$_2$O$_4$, interface structures, and bulk water, respectively.
The HDNNP reproduces the energies of the training set with a root mean squared error (RMSE) of 1.9\,meV\,atom$^{-1}$, while it is able to predict the energies of the test set with an RMSE of 2.4\,meV\,atom$^{-1}$. The RMSE of only the bulk water training structures is about 1.0\,meV\,atom$^{-1}$ (test set 1.1\,meV\,atom$^{-1}$) and thus about half of the values of the RMSEs of the Li$_x$Mn$_2$O$_4$ and interface training structures, which are 2.0 and 2.1\,meV\,atom$^{-1}$, respectively (test sets 2.3 and 2.8\,meV\,atom$^{-1}$). We ascribe this difference to the less complex geometric and electronic structure of bulk liquid water that is typically very well represented by machine learning potentials.\cite{Morawietz2016, Cheng2019} We note that in spite of the increased complexity of the potential energy surface of the interface system, in particular the bulk Li$_x$Mn$_2$O$_4$ test set RMSE is very similar to the RMSE of 2.2\,meV\,atom$^{-1}$ of a previous HDNNP fitted to bulk Li$_x$Mn$_2$O$_4$ data only.\cite{Eckhoff2020a} The maximum energy error of all data in the training/test set is 13.2/15.6\,meV\,atom$^{-1}$. Only 2.3/5.5{\%} of the energy predictions in the training/test set have errors larger than 5\,meV\,atom$^{-1}$ (Figures \ref{fig:HDNN_quality} (a) and (b)).
The PBE0r-D3 reference data contains force components up to $|F^\mathrm{DFT}|\leq 5.06$\,eV\,$a_0^{-1}$. The force components RMSE of all data is 0.127\,eV\,$a_0^{-1}$ for both, training and test set. Again, the RMSE of bulk water (0.059\,eV\,$a_0^{-1}$) is lower than the RMSEs of bulk Li$_x$Mn$_2$O$_4$ (0.114\,eV\,$a_0^{-1}$) and the interface structures (0.143\,eV\,$a_0^{-1}$). The force components RMSE of bulk Li$_x$Mn$_2$O$_4$ is similar to the value of 0.107\,eV\,$a_0^{-1}$ of the aforementioned bulk only Li$_x$Mn$_2$O$_4$ HDNNP.\cite{Eckhoff2020a} As highlighted by the heatmaps in Figures \ref{fig:HDNN_quality} (c) and (d) most of the force components have an error smaller than 0.5\,eV\,$a_0^{-1}$ (99.33/99.32{\%}). The maximum errors are 3.13 and 3.12\,eV\,$a_0^{-1}$ for the training and test set.
The atomic spins of Mn are in the range $1.45\,\hbar\leq S^\mathrm{DFT}\leq 2.55\,\hbar$ (Figures \ref{fig:HDNN_quality} (e) and (f)). The RMSE of the HDNNS is $0.04\,\hbar$ for both training and test set. Only 0.75{\%} of the training data and 0.76{\%} of the test data show errors larger than $0.2\,\hbar$ possibly resulting in the assignment of a different spin and oxidation state, while the maximum errors are 0.57 and $0.54\,\hbar$, respectively. Consequently, the vast majority of the Mn oxidation states is accurately predicted. The assignment to the oxidation states Mn$^\mathrm{IV}$ (d electron configuration t$_\mathrm{2g}^3$e$_\mathrm{g}^0$), high-spin Mn$^\mathrm{III}$ (t$_\mathrm{2g}^3$e$_\mathrm{g}^1$), and high-spin Mn$^\mathrm{II}$ (t$_\mathrm{2g}^3$e$_\mathrm{g}^2$) has been set to the intervals $1.4\,\hbar\leq S<1.8\,\hbar$, $1.8\,\hbar\leq S<2.2\,\hbar$, and $2.2\,\hbar\leq S<2.6\,\hbar$, respectively, based on the distribution of spins shown in Figures \ref{fig:HDNN_quality} (e) and (f).
\subsection{Manganese oxidation state distribution}\label{sec:Mn_oxidation_state}
Using the obtained HDNNP and HDNNS, we investigate four solid-liquid interface systems with different geometric and electronic structure. In this section we first investigate the interface from the perspective of the solid phase, while we focus on the liquid phase in the next section and finally discuss the reactivity of interfacial water species in the last section.
In general, HDNNP-driven MD simulations allow to gain insights into the atomic structure and dynamics of the Li$_x$Mn$_2$O$_4$-water interface in equilibrium. Further, using the HDNNS we can in addition investigate the Mn oxidation state distribution in the trajectories. This distribution is of major importance for understanding capacity fading of Li$_x$Mn$_2$O$_4$ during battery usage because previous studies propose the origin to be disproportionation of Mn$^\mathrm{III}$ ions at the interface and subsequent dissolution of the emerging Mn$^\mathrm{II}$ ions. Mn$^\mathrm{II}$ is the most stable oxidation state in aqueous solution while Mn$^\mathrm{IV}$ is not soluble.\cite{Davies1969} Structural features exposing only small amounts of Mn$^\mathrm{II}$ and Mn$^\mathrm{III}$ ions at the interface could thus support the development of more durable electrode materials.
\begin{figure*}[htb!]
\centering
\includegraphics[width=\textwidth]{Figure_03.pdf}
\vspace*{-0.5cm}
\caption{Time averaged Mn oxidation state distributions and equilibrium charge transfer rates $\nu_\mathrm{CT}$ in the different layers of the (a) $\{100\}_\mathrm{Li}$, (b) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, (c) $\{110\}_{\mathrm{LiMnO}_2}$, and (d) $\{110\}_{\mathrm{MnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems. The oxidation state distribution is represented by the number of each species per Mn site $n$ in a layer. The lines are only shown to guide the eyes. The zero point of $z$ has been set to the center of the Li$_x$Mn$_2$O$_4$ slabs.}\label{fig:spin}
\end{figure*}
We start with the $\{100\}_\mathrm{Li}$ surface (Figure \ref{fig:structure} (a)), which is positively polarized due to Li termination, while the overall slab is neutral because of the charge compensation by the subsequent Mn$_2$O$_4$ layer containing formally a one-to-one ratio of Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ ions. We note that due to the structurally identical surfaces at both sides of the slab, the system in total contains 23 Li layers and only 22 Mn$_2$O$_4$ layers resulting in a slight overall excess of Mn$^\mathrm{III}$ compared to Mn$^\mathrm{IV}$ ions in the system. Such a situation is not unphysical but typical for Li$_x$Mn$_2$O$_4$, e.g., for different loads of Li ions compensated by different oxidation states of Mn.
In our simulations we observe that in the outermost Mn$_2$O$_4$ layer on each side almost only Mn$^\mathrm{III}$ ions are present (Figure \ref{fig:spin} (a)). The reason for the preference of Mn$^\mathrm{III}$ ions in the topmost layer is the undercoordination of the Mn ions at interface sites by only five O$^{2-}$ ions compared to the octahedral coordination in the bulk material. We have not observed any long-living OH$^-$ ions formed by dissociation of water molecules, Consequently, significant protonation of interfacial O$^{2-}$ ions and adsorption of hydroxide ions at the undercoordinated Mn ions does not take place. Due to this lack of negative ions, the formation of Mn$^\mathrm{III}$ ions is favored at the interface and Mn$^\mathrm{IV}$ ions are predominantly found in the second Mn$_2$O$_4$ layer (Figure \ref{fig:spin} (a)). Deeper layers contain a decreasingly pronounced alternating excess of either Mn$^\mathrm{III}$ or Mn$^\mathrm{IV}$ ions, respectively, which becomes small after several layers. Mn$^\mathrm{II}$ ions are very rarely observed and, if found, they only emerge in the topmost layer with about $10^{-6}$ Mn$^\mathrm{II}$ ions per Mn site.
\begin{figure}[htb!]
\centering
\includegraphics[width=\columnwidth]{Figure_04.pdf}
\caption{Equilibrated example structure of the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ Li$_x$Mn$_2$O$_4$-water interface including the solid surface and the first layer of the liquid projected on the $xy$ plane. The colors are according to the definition in Figure \ref{fig:structure} except that Mn$^\mathrm{III}$ ions are highlighted in pink, while Mn$^\mathrm{IV}$ ions are shown in violet.}\label{fig:structure_oxidation_states}
\end{figure}
In contrast to the $\{100\}_\mathrm{Li}$ system, both sides of the slab representing the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface (Figure \ref{fig:structure} (b)) are terminated by Mn$_2$O$_4$ resulting in total in 22 Li layers and 23 Mn$_2$O$_4$ layers. The reduced amount of Li$^+$ ions in the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ system compared to the $\{100\}_\mathrm{Li}$ system leads to a slight excess of Mn$^\mathrm{IV}$ ions instead of Mn$^\mathrm{III}$ ions. In the topmost layer the fraction is about three Mn$^\mathrm{III}$ to two Mn$^\mathrm{IV}$ ions (Figure \ref{fig:spin} (b)). The increased stability of Mn$^\mathrm{IV}$ ions in the topmost layer compared to the $\{100\}_\mathrm{Li}$ system can be explained by the coordination by OH$^-$ ions. For this system we find that OH$^-$ ions are adsorbed on top of about 43{\%} of the interface Mn ions. The majority of the corresponding protons formed in the dissociation of water molecules is attached to interface O$^{2-}$ ions, which are covered to about 21{\%} forming OH$^{-}$. This value is about half of the Mn coverage by OH$^-$ ions because there are twice as many interface O$^{2-}$ ions as Mn ions per layer. The OH$^-$ ions of the first liquid layer are preferably but not exclusively located at Mn$^\mathrm{IV}$ sites while Mn$^\mathrm{III}$ sites are typically coordinated by H$_2$O molecules (Figure \ref{fig:structure_oxidation_states}). The proposed intermediate state of the OER, in which two OH$^-$ ions are found on top of an [Mn$^\mathrm{III}_2$Mn$^\mathrm{IV}_2$O$_4$]$^{6+}$ unit,\cite{Cady2015} is therefore a rare configuration in our simulations explaining the low OER activity of the stoichiometric LiMn$_2$O$_4$ spinel.
The second layer on each side adapts to the Mn$^\mathrm{IV}$ excess in the system (Figure \ref{fig:spin} (b)). The alternating excess of either Mn$^\mathrm{III}$ or Mn$^\mathrm{IV}$ ions decays after a few layers converging to equal fractions of Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ ions in deeper layers. In contrast to the $\{100\}_\mathrm{Li}$ system in which basically no Mn$^\mathrm{II}$ ions form, in the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ system we find about $5\cdot10^{-4}$ Mn$^\mathrm{II}$ ions per Mn site in the topmost Mn$_2$O$_4$ layer with some statistical fluctuations depending on the surface and simulation. Substantially longer simulations are expected to be required to obtain fully converged values for such small fractions. As both sides of the Li$_x$Mn$_2$O$_4$ slab in every simulation are equal, converged results in Figure \ref{fig:spin} are symmetric with respect to $z=0$. Deviations from this behavior, which are small for most of our results, can consequently be employed to estimate the uncertainty caused by the finite simulation time.
The $\{110\}_{\mathrm{LiMnO}_2}$ (Figure \ref{fig:structure} (c)) and $\{110\}_{\mathrm{MnO}_2}$ (Figure \ref{fig:structure} (d)) interface systems show similar amounts of about $3\cdot10^{-4}$ and $4\cdot10^{-4}$ Mn$^\mathrm{II}$ ions per Mn sites in the topmost surface layers. While the $\{110\}_{\mathrm{LiMnO}_2}$ system contains a slight excess of Mn$^\mathrm{III}$ over Mn$^\mathrm{IV}$ ions, the opposite is the case for the $\{110\}_{\mathrm{MnO}_2}$ system. The topmost layer is still dominated by Mn$^\mathrm{III}$ ions in both systems with a ratio of nine Mn$^\mathrm{III}$ to one Mn$^\mathrm{IV}$ at the $\{110\}_{\mathrm{LiMnO}_2}$ interface (Figure \ref{fig:spin} (c)) and almost only Mn$^\mathrm{III}$ ions at the $\{110\}_{\mathrm{MnO}_2}$ interface (Figure \ref{fig:spin} (d)). The reason for this preference of Mn$^\mathrm{III}$ ions is the low coordination of the topmost Mn ions by only four O$^{2-}$ ions. As a consequence, the second layer in the $\{110\}_{\mathrm{LiMnO}_2}$ system and even the second and third layers in the $\{110\}_{\mathrm{MnO}_2}$ interface are predominantly occupied by Mn$^\mathrm{IV}$ ions. The influence of the interface on the oxidation state distribution vanishes already after about four layers. Regular oscillations of the alternating excess of either Mn$^\mathrm{III}$ or Mn$^\mathrm{IV}$, respectively, can be observed in the bulk (Figure \ref{fig:spin} (c) and (d)). As the amplitude of these oscillations varies for the $\{110\}_{\mathrm{LiMnO}_2}$ and $\{110\}_{\mathrm{MnO}_2}$ systems different arrangements of Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ ions can lead to local minimum configurations. Further, due to electron hopping processes, the Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ arrangements are dynamical at 298\,K and different local motifs are observed during the simulations even for the same interface systems. We note that the Mn oxidation state distribution is averaged over (001) planes in Figures \ref{fig:spin} (a) and (b) and over (110) planes in Figures \ref{fig:spin} (c) and (d) yielding different slices of the structure.
To compensate for the undercoordination of Mn ions by O$^{2-}$, about 94{\%} of the $\{110\}_{\mathrm{LiMnO}_2}$ interface Mn ions are in addition coordinated by approximately two OH$^-$ ions. These OH$^-$ ions are placed on bridge sites thus resembling the octahedral coordination in bulk Li$_x$Mn$_2$O$_4$ and are shared by two Mn ions (see Section \ref{sec:interface_structure}). About 46{\%} of the interface O$^{2-}$ ions contain adsorbed H$^+$ ions forming OH$^-$. For the $\{110\}_{\mathrm{MnO}_2}$ interface this value further increases slightly to about 48{\%}, while there is about one adsorbed OH$^-$ ion per interfacial Mn atom. In this case, the OH$^-$ ions originating from the liquid phase are not placed at Mn bridge sites but adopt empty octahedral coordination sites (see section \ref{sec:interface_structure}). To compare the OH$^-$ coverage between the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface and the $\{110\}_{\mathrm{LiMnO}_2}$ and $\{110\}_{\mathrm{MnO}_2}$ interfaces, the number of Mn ions per interface area has to be taken into account. This number is about 1.5 times higher at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface compared to the other two interfaces. Still, the OH$^-$ coverage is higher at the $\{110\}_{\mathrm{LiMnO}_2}$ and $\{110\}_{\mathrm{MnO}_2}$ interfaces. Hence, the dissociation degree of H$_2$O molecules is higher at these two interfaces, which provide more empty octahedral coordination sites of the interfacial Mn ions than the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface.
In addition to the Mn oxidation state distribution, we estimated the charge transfer rates related to electron hopping between the Mn$^\mathrm{II}$, Mn$^\mathrm{III}$, and Mn$^\mathrm{IV}$ ions. These rates were determined from the number of oxidation state changes per time. Because two oxidation states are changed by one electron hop, this number is divided by two. The trajectory data was collected every 0.1\,ps. The data in Figures \ref{fig:spin} (a) to (d) show that even the fastest processes are more than an order of magnitude slower than this sampling interval. This time scale difference ensures that the major fraction of processes is counted. To exclude counting of unsuccessful attempts to change the oxidation state, a transition is only considered in case the spin value of a Mn$^\mathrm{IV}$ ion increases above $1.9\,\hbar$, the spin value of a Mn$^\mathrm{III}$ ion increases above $2.3\,\hbar$ or decreases below $1.7\,\hbar$, or the spin value of a Mn$^\mathrm{II}$ ion decreases below $2.1\,\hbar$.
For all systems, the charge transfer rates are found to be largest close to the interface, with a maximum typically in the second to fourth Mn containing layer (Figures \ref{fig:spin} (a) to (d)). In the center of the solid slab the corresponding values are between about 0.02 to 0.03 charge transfers per ps and Mn site and hence much smaller than at the interface. These values are close to the value of 0.02 charge transfers per ps and Mn site obtained in our previous study of bulk LiMn$_2$O$_4$.\cite{Eckhoff2020b} The rates in the topmost layer of the $\{110\}_{\mathrm{LiMnO}_2}$ interface are higher than those in the topmost layers of the other interfaces, and in general lower charge transfer rates are found in layers with predominant single Mn oxidation states. For the $\{110\}_{\mathrm{LiMnO}_2}$ interface the rates are smaller in the second layer, which corresponds to the topmost layer at the $\{110\}_{\mathrm{MnO}_2}$ interface. Therefore, the charge transfer rates in the LiMnO$_2$ layers seem to be higher than those in the MnO$_2$ layers.
The electrons are not explicitly included in the HDNNP and the HDNNS identifies different oxidation states based on the local structural environment like, for example, the presence or absence of Jahn-Teller distortions. Consequently, the inhomogeneous distribution of Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ ions in the systems as well as the electron hopping processes raise the question if the overall numbers of these ions are conserved during the simulations. In principle, in the solid phase there has to be a one-to-one ratio between the number of Li$^+$ ions and the number of Mn e$_\mathrm{g}$ electrons. Since the average number of Mn$^\mathrm{II}$ ions contributing two e$_\mathrm{g}$ electrons is rather small or even negligible at all interfaces, we expect to find about the same number of Li$^+$ ions and Mn$^\mathrm{III}$ ions, each containing one e$_\mathrm{g}$ electron, in the system. Indeed we observe that the number of e$_\mathrm{g}$ electrons stays approximately constant during all interface simulations (see Supplementary Material Figures S2 (a) to (d)). This observation provides evidence of the conservation of total charge and number of electrons and confirms the consistent description of the systems by the HDNNP. Only small fluctuations in the predictions are observed, since electron hopping processes can give rise to intermediate structures in which the geometry-based assignment of the oxidation state is unavoidably physically ambiguous.\cite{Eckhoff2020b} In addition, remaining prediction errors of the HDNNS may contribute to these fluctuations as well.
Due to the slight excess of Li$^+$ ions in the $\{100\}_\mathrm{Li}$ system related to the surface geometry, there are more Mn$^\mathrm{III}$ than Mn$^\mathrm{IV}$ ions present in this system. About 409 Mn$^\mathrm{III}$ and 383 Mn$^\mathrm{IV}$ ions are predicted by the HDNNS on average over the full simulation time of all three $\{100\}_\mathrm{Li}$ interface simulations, which contain all the same number of Mn ions. Since 414 Li$^+$ ions are present, the error in the number of e$_\mathrm{g}$ electrons obtained from the HDNNS prediction is only about 1.2{\%}. In contrast to the $\{100\}_\mathrm{Li}$ system, the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ system contains more Mn$^\mathrm{IV}$ than Mn$^\mathrm{III}$ ions. Also in this case the HDNNS prediction is very accurate yielding about 400 Mn$^\mathrm{III}$ and 428 Mn$^\mathrm{IV}$ ions on average. Compared to the number of 396 Li$^+$ ions, the prediction error in the number of e$_\mathrm{g}$ electrons is again small (1.1{\%}). For the $\{110\}_{\mathrm{LiMnO}_2}$ system 401 e$_\mathrm{g}$ electrons are predicted compared to 408 Li$^+$ ions resulting in an underestimation of 1.8{\%}. Again, the excess of Mn$^\mathrm{III}$ over Mn$^\mathrm{IV}$ ions is predicted correctly (401:391). For the $\{110\}_{\mathrm{MnO}_2}$ system, for which 386 Mn$^\mathrm{III}$ and 406 Mn$^\mathrm{IV}$ ions are predicted by the HDNNS, an even better agreement with a deviation of only 0.4{\%} is reached, as this slab contains 384 Li$^+$ ions. In conclusion, the Mn oxidation states of all systems identified by the HDNNS are very accurately described via the geometric atomic environments of the Mn ions determined by the HDNNP energy surface.
The Mn$^\mathrm{II}$ ions in all systems have an above-average distance from the solid and are slightly displaced towards the liquid, which can be reasoned by the larger size of Mn$^\mathrm{II}$ ions. The Mn$^\mathrm{II}$ ions are preferably coordinated by H$_2$O instead of OH$^-$ of the water contact layer and the Mn-O distances are on average larger than for Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ ions. Dissolution of Mn$^\mathrm{II}$ ions was not observed in the 5\,ns MD simulations at 298\,K and 1\,bar employing atomically flat solid surfaces without defects and in the absence of external electric fields. Therefore, dissolution seems to be rare in equilibrium under standard conditions. Especially, the inclusion of steps and defects at the solid surface is expected to increase the dissolution rate.\cite{Dove2005} Moreover, possible surface reconstructions as proposed for the \{110\} and \{111\} surfaces\cite{Hirayama2010, Benedek2011, Karim2013, Kim2015} and the formation of surface layers of different stoichiometry such as Mn$_3$O$_4$\cite{Tang2014, Amos2016, Gao2019} might also be relevant for Mn$^\mathrm{II}$ dissolution. In particular, the Mn$_3$O$_4$ tetragonal spinel structure, in which Mn$^\mathrm{II}$ ions substitute the Li$^+$ ions at the tetrahedral sites of Li$_x$Mn$_2$O$_4$, is an interesting candidate for the formation of dissolved Mn$^\mathrm{II}$ ions. These Mn$^\mathrm{II}$ ions in addition block the Li$^+$ channels in the spinel structure and thus need to be removed during charging of the battery.
In summary, the weak interaction between water and the $\{100\}_\mathrm{Li}$ interface seems to be responsible for only very little formation of Mn$^\mathrm{II}$ ions. The outermost Li$^+$ layer separates the water molecules from the Mn and O$^{2-}$ ions, which are important for the dissociation of water and the formation of long-living OH$^-$ ions. A high coordination by O$^{2-}$ ions, i.e., a more bulk-like environment, favors the formation of higher Mn oxidation states and leads to a weaker interaction with water. On the one hand, electron hopping and hence electrical conductivity is increased in the vicinity of the interface leading to higher battery performance when using smaller particles sizes. On the other hand, the formation of Mn$^\mathrm{II}$ ions is only observed close to the surface suggesting more durable battery materials when using larger particles.
\subsection{Structural characterization of the interfaces}\label{sec:interface_structure}
The atomic structure as well as the reactivity of the interface are determined by the termination of the solid surface. For instance, we have seen that the Mn coordination can strongly affect the formation of OH$^-$ ions. The resulting degree of hydroxylation at the interface can be expected to be relevant for reactions at the surface such as the OER. Moreover, the structure and dynamics of the liquid in the vicinity of the interface and deviations of its properties from the bulk liquid are of high interest.
\begin{figure*}[htb!]
\centering
\includegraphics[width=\textwidth]{Figure_05.pdf}
\caption{Decadic logarithm of the time averaged concentration $c$ of different oxygen species as a function of the $z$ coordinate for the (a) $\{100\}_\mathrm{Li}$, (b) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, (c) $\{110\}_{\mathrm{LiMnO}_2}$, and (d) $\{110\}_{\mathrm{MnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems. The region of the Li$_x$Mn$_2$O$_4$ slab is highlighted by the gray background. The zero point of $z$ has been set to the center of the water slabs.}\label{fig:oxygen_species}
\end{figure*}
To assess the impact of the interface on the properties of the liquid, the system has to be sufficiently large to ensure the presence of a bulk-like region in the center of the liquid phase. This bulk-like region is not only important for comparing interfacial properties to those of the bulk, but also to obtain converged data for the interfacial properties. Figures \ref{fig:oxygen_species} (a) to (d) show the averaged atomic distributions in all four systems. The central region of the liquid phase is very similar for all different surfaces and shows only small fluctuations, which are much larger in the vicinity of the surfaces. The density of H$_2$O in the central 5\,{\AA} slice of the liquid phase is 0.946\,kg\,l$^{-1}$ ($c_{\mathrm{H}_2\mathrm{O}}=52.5\,\mathrm{mol\,l}^{-1}$). This density agrees very well with the value of 0.947\,kg\,l$^{-1}$ obtained in HDNNP-driven simulations of bulk water. Further, the properties of the hydrogen bond network are very similar in the centers of the liquid in all simulations (Supplementary Material). Consequently, the simulation cells, which all have a water region with a diameter of at least 50\,{\AA}, are large enough to yield bulk properties in the center, which is in excellent agreement with previous studies on other solid-water interfaces.\cite{Natarajan2016, Quaranta2017} An underestimation of the density compared to the experimental value of 0.997\,kg\,l$^{-1}$ at 298\,K and 1\,bar\cite{Wagner2002} is common in DFT calculations and was also observed in a previous study of water yielding 0.94\,kg\,l$^{-1}$\cite{Cheng2019} based on the revPBE0-D3\cite{Zhang1998, Adamo1999} DFT functional -- we note that our results are based on the PBE0r-D3\cite{Sotoudeh2017, Eckhoff2019} DFT functional.
The concentration profile as a function of the distance from the surfaces, which is proportional to the density profile with the molar mass as proportionality constant, shows two distinct OH$^-$ peaks of about the same size for each phase boundary (Figure \ref{fig:oxygen_species} (a) to (d)). These peaks correspond to protonated O$^{2-}$ ions of the solid and OH$^-$ ions adsorbed to Mn sites, respectively. Since these two peaks dominate the OH$^-$ concentration profile and are of about equal height, essentially all protons and hydroxide ions formed in the dissociation of water molecules are bound at the solid surface. The OH$^-$ concentration is about three orders of magnitude smaller at the $\{100\}_\mathrm{Li}$ interface compared to the other systems. The reason is that the OH$^-$ ions are not long-living at the $\{100\}_\mathrm{Li}$ interface.
Oscillations of the H$_2$O concentration in the vicinity of the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface are more pronounced than in the vicinity of the $\{100\}_\mathrm{Li}$ interface (Figure \ref{fig:oxygen_species} (a) and (b)). The depletion layer beyond the first water peak can be observed in the spatial atomic distributions in Figure \ref{fig:layer_distribution} (b) as well. The relatively high concentration of H$_3$O$^+$ ions at the same distance as the second OH$^-$ peak will be discussed in Section \ref{sec:proton_transfer}.
The structural deviation of the contact layer from the bulk liquid for the $\{110\}_{\mathrm{LiMnO}_2}$ and $\{110\}_{\mathrm{MnO}_2}$ systems is even more pronounced. In the case of the $\{110\}_{\mathrm{LiMnO}_2}$ system, the OH$^-$ concentration profile shows the presence of OH$^{-}$ ions rather far from the surface. The formation of a small additional O$^{2-}$ peak on top of the surface with maximum concentrations between 1 and 10\,mol\,l$^{-1}$ is particularly interesting. This peak implies that some H$_2$O molecules can be deprotonated twice to partially complete the octahedral coordination of the Mn ions as shown in Figure \ref{fig:structure_interface} (a). This formation of surface exposed O$^{2-}$ ions is potentially of interest for catalytic reactions due to their low coordination. Furthermore, a Mn$^\mathrm{IV}$ ion is often found close to a surface exposed O$^{2-}$ ion although the first solid layer is typically dominated by Mn$^\mathrm{III}$ ions at this surface. Thus, the surface exposed O$^{2-}$ ions have an impact on the distribution of Mn$^\mathrm{III}$ and Mn$^\mathrm{IV}$ ions.
At the $\{110\}_{\mathrm{MnO}_2}$ interface an H$_2$O concentration of 0.1 to 1\,mol\,l$^{-1}$ is found in the topmost layer of the solid surface, which implies the opposite process. Here, O$^{2-}$ ions are protonated twice thus forming water as shown in Figure \ref{fig:structure_interface} (b). A correlation of Mn$^\mathrm{II}$ formation to the formation of H$_2$O molecules in the topmost solid layer is not observed. Mn$^\mathrm{II}$ ions are most often observed in environments in which the Mn ion is coordinated by two H$_2$O molecules from above instead of typically one OH$^-$ ion and one H$_2$O molecule. The double deprotonation and protonation processes can be viewed as surface reconstructions of the $\{110\}$ Li$_x$Mn$_2$O$_4$ surfaces.
\begin{figure*}[htb!]
\centering
\includegraphics[width=\textwidth]{Figure_06.pdf}
\caption{Spatial atomic distributions projected onto the $zy$ plane of the (a) $\{100\}_\mathrm{Li}$, (b) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, (c) $\{110\}_{\mathrm{LiMnO}_2}$, and (d) $\{110\}_{\mathrm{MnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems are shown on the left. On the right the spatial atomic distributions up to 2.5\,{\AA} from the interface projected onto the $xy$ plane are represented with a ball model of the non-equilibrated vacuum solid surface in the background for reference. The time averaged concentration of each atomic species is represented by a linearly increasing opacity. In the case of oxygen, each of the species differing in hydrogen content is plotted individually. The spatial distributions are stacked on top of each other in the order from bottom to top H (yellow), Li (green), Mn (violet), O$^{2-}$ (red), O in H$_2$O (blue), and O in OH$^-$ (turquoise). The spatial distributions of Li, Mn, and O$^{2-}$ are not shown in the $xy$ projections.}\label{fig:layer_distribution}
\end{figure*}
\begin{figure}[htb!]
\centering
\includegraphics[width=\columnwidth]{Figure_07.pdf}
\caption{Equilibrated example structures of (a) an O$^{2-}$ ion (red) formed in an additional layer at the $\{110\}_{\mathrm{LiMnO}_2}$ interface and (b) an H$_2$O molecule (blue) formed in the topmost solid layer of the $\{110\}_{\mathrm{MnO}_2}$ interface. Both species are highlighted by a yellow circle.}\label{fig:structure_interface}
\end{figure}
On the left side of Figures \ref{fig:layer_distribution} (a) to (d) the time averaged spatial atomic distributions projected onto the $zy$ plane are shown for each interface. As expected the bulk solid has a regular pattern reflecting the crystal structure while the bulk liquid has a diffuse distribution in all simulations. However, at the different solid-liquid interfaces the liquid phase shows various interesting strongly bound water structural features, which are less mobile due to the strong interaction with the surface. An adaption of the interfacial water layers to optimize the interaction to the solid as well as the hydrogen bond network to the bulk liquid has been observed for different metal surfaces as well and can yield very specific water structures depending on the solid surface.\cite{Liriano2017, Gerrard2019} The thickness of the strongly bound water layer depends on the underlying solid surface. For the $\{100\}_\mathrm{Li}$ interface the strongly bound water layer has a diameter of about 1.5 to 2\,{\AA} only, while it is 3 to 4\,{\AA} thick for the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ and $\{110\}_{\mathrm{LiMnO}_2}$ interfaces. However, the structure at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface is dominated by a two-dimensional dense water layer on top of the solid, while the structure at the $\{110\}_{\mathrm{LiMnO}_2}$ interface is three-dimensional. A pattern similar to the latter one is also observed at the $\{110\}_{\mathrm{MnO}_2}$ interface. Here the strongly bound water layer is even 3.5 to 4.5\,{\AA}.
\begin{figure}[htb!]
\centering
\includegraphics[width=\columnwidth]{Figure_08.pdf}
\caption{Equilibrated example structures of the strongly bound water layers in the (a) $\{110\}_{\mathrm{LiMnO}_2}$ and (b) $\{110\}_{\mathrm{MnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems.}\label{fig:structure_crystalline_water}
\end{figure}
To investigate the strongly bound water layers in more detail the right panels of Figures \ref{fig:layer_distribution} (a) to (d) show the spatial atomic distributions in water films of 2.5\,{\AA} diameter above the surface starting from the H atom closest to the surface projected onto the $xy$ plane. The small strongly bound water layer on top of the $\{100\}_\mathrm{Li}$ interface forms due to the attractive interactions between the oxygen of H$_2$O and Li$^+$ ions as well as due to hydrogen bond formation between the hydrogen of H$_2$O and O$^{2-}$ ions at the surface (Figure \ref{fig:layer_distribution} (a)). A pattern can be observed at the interface but the concentrations at the strongly bound water sites are lower than those for the other interfaces (lower opacity of the first liquid layer in the right panels of Figure \ref{fig:layer_distribution} (a) than in (b) to (d)). Consequently, the H$_2$O molecules at the $\{100\}_\mathrm{Li}$ interface are more mobile than the various water species at the other interfaces, corresponding to the weaker interaction between water and the $\{100\}_\mathrm{Li}$ surface discussed above.
As mentioned in Section \ref{sec:Mn_oxidation_state} OH$^-$ ions are formed at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, $\{110\}_{\mathrm{LiMnO}_2}$, and $\{110\}_{\mathrm{MnO}_2}$ interfaces and cover the solid surface to a large fraction (Figures \ref{fig:layer_distribution} (b) to (d)). These OH$^-$ ions are rather strongly bound to specific sites and thus less mobile than the water molecules. Moreover, they are able to form strong hydrogen bonds and can order the water molecules. Therefore, they have a large impact on the formation of the strongly bound water layer. The H$_2$O/OH$^-$ distribution in the first layer of the liquid phase does not follow a regular pattern at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface (Figure \ref{fig:layer_distribution} (b) as well as Figure \ref{fig:structure_oxidation_states}). Hence, the Mn$^\mathrm{III}$/Mn$^\mathrm{IV}$ distribution in the first layer of the solid phase is also disordered (Figure \ref{fig:structure_oxidation_states}).
The order in the H$_2$O/OH$^-$ distribution increases for the $\{110\}_{\mathrm{LiMnO}_2}$ and $\{110\}_{\mathrm{MnO}_2}$ interfaces. At the $\{110\}_{\mathrm{LiMnO}_2}$ interface most OH$^-$ ions of the first layer of the liquid bridge the Mn sites, which are arranged in rows in the first solid layer (e.g., $y\approx5\,\text{\AA}$ in Figure \ref{fig:layer_distribution} (c) as well as Figures \ref{fig:structure_interface} (a) and \ref{fig:structure_crystalline_water} (a)). Most H$_2$O molecules in the first layer are aligned in rows close to the Li$^{+}$ ions with H atoms pointing to O$^{2-}$ ions of the solid (e.g., $y\approx1\,\text{\AA}$ in Figure \ref{fig:layer_distribution} (c)). In the first layer of the solid rows of alternating O$^{2-}$ and OH$^-$ ions are formed (e.g., $y\approx-1\,\text{\AA}$ in Figure \ref{fig:layer_distribution} (c)). O$^{2-}$ excess can lead to Mn$^\mathrm{IV}$ ions in the topmost solid layer.
In the first solid layer of the $\{110\}_{\mathrm{MnO}_2}$ interface an alternating pattern of O$^{2-}$ and OH$^-$ rows is formed, whereby O$^{2-}$ ions bridge the underlying Mn sites and the OH$^-$ ions are above the Li sites (e.g., $y\approx3\,\text{\AA}$ and $y\approx7\,\text{\AA}$ in Figure \ref{fig:layer_distribution} (d) as well as Figures \ref{fig:structure_interface} (b) and \ref{fig:structure_crystalline_water} (b)). In some simulations the H$_2$O molecules and OH$^-$ ions in the first liquid layer are arranged in alternating rows (e.g., $x\approx2\mathrm{\ to\ }3\,\text{\AA}$ and $x\approx5\mathrm{\ to\ }6\,\text{\AA}$ in Figure \ref{fig:layer_distribution} (d)).
The oxygen atoms of the water species in the first liquid layer generally tend to continue the oxygen face centered cubic (fcc) lattice of the solid yielding an energetically favored coordination of the Li and Mn ions (see Supplementary Material Figures S7 (a) and (b) for representations of the oxygen lattice only). The second liquid layer is most pronounced in the case of the $\{110\}_{\mathrm{MnO}_2}$ interface. The view on the $yz$ plane still seems to agree with the fcc lattice (Figure \ref{fig:structure_crystalline_water} (b)) but a view on the $xz$ plane shows that the second layer does not match the fcc lattice (Supplementary Material Figure S8). The same observation is obtained for the $\{110\}_{\mathrm{LiMnO}_2}$ interface. However, the strongly bound water layer cannot be assigned to a specific lattice structure or ice polymorph, since in most cases the OH$^{-}$ ions at the surface do not form a sufficiently regular pattern.
In conclusion, with increasing OH$^-$ concentration complex structural patterns can form which also affect the oxidation states of the underlying Mn ions. Interactions between the solid and the liquid lead to favored positions of the atoms. The oxygen atoms of OH$^-$ ions and H$_2$O molecules in the contact layer of the liquid prefer to be at sites which complete a bulk-like octahedral coordination of Mn or tetrahedral coordination of Li. The respective OH$^-$ ions thereby favor to bind to interfacial Mn ions. Therefore, strongly bound water structures form at the interface.
To compare the residence lifetimes of the O atoms in the strongly bound water layers at the different interfaces as well as in the bulk quantitatively, the correlation function,
\begin{align}
C_\mathrm{res}(t)=\dfrac{1}{N_\mathrm{tot}-N(t)}\sum_{t_0=0}^{t_\mathrm{tot}-t}\dfrac{\mathbf{r}_\mathrm{init}(t_0)\cdot\mathbf{r}(t_0+t)}{|\mathbf{r}_\mathrm{init}(t_0)|}\ ,
\end{align}
can be used with the total number of simulation steps $N_\mathrm{tot}$ and total simulation time $t_\mathrm{tot}$. $N(t)$ is the number of steps until time $t$. The quantity,
\begin{align}
r_i(t)=\begin{cases}1&\mathrm{for}\ z_\mathrm{min}\leq z_i(t)\leq z_\mathrm{max}\\0&\mathrm{otherwise}\end{cases}\ ,
\end{align}
keeps track if the oxygen $z$ coordinate of atom $i$ at time $t$ is outside the interval $[z_\mathrm{min},z_\mathrm{max}]$. O atoms which are close to the boundary at the time $t_0$ can lead to a large impact of short-term boundary crossing. To reduce this effect, the vector $\mathbf{r}_\mathrm{init}$ can employ a slightly smaller interval $[z_\mathrm{min},z_\mathrm{max}]$ than used for $\mathbf{r}$.
The residence correlation function is 1 if all initially resident O atoms are in the interval and becomes 0 if none of these O atoms is left in the interval. This function enables to measure the residence lifetime of the O atoms in given slices of the simulation cell. The residence lifetime can be extracted by a biexponential fit using
\begin{align}
C_\mathrm{res}^\mathrm{fit}(t)=C_1\exp\left(-\dfrac{t}{\tau_1}\right)+(1-C_1)\exp\left(-\dfrac{t}{\tau_2}\right)\ .
\end{align}
$\tau_1$ is the residence lifetime due to long-range reorganization caused by diffusion and $\tau_2$ accounts for short-term crossing of the boundaries. The weights of both processes $C_1$ and $1-C_1$ add up to one. For verification we show in the Supplementary Material Figure S9 monoexponential fits excluding the initial drop of the correlation functions which yield similar residence lifetimes $\tau_1$.
\begin{figure}[htb!]
\centering
\includegraphics[width=\columnwidth]{Figure_09.pdf}
\caption{Correlation functions $C_\mathrm{res}$ of the oxygen residence in the center of the liquid phase and in the interfacial liquid phase close to the solid surfaces. For the bulk case O atoms are considered which are located in the central 5\,{\AA} slice of the liquid phase at $t_0=0$. For the interface cases, O atoms in the two 2.5\,{\AA} slices at both surfaces of the solid at $t_0=0$ contribute. The contributions to the correlation function decrease from 1 to 0 if the O atoms leave the slices by more than 0.3\,{\AA} and they increase from 0 to 1 if the O atoms enter this region again. The results of all twelve simulations are averaged for the bulk water data. For the interface data the results of the three simulations for each interface are averaged. Dashed lines correspond to biexponential fits.}\label{fig:correlation_residence}
\end{figure}
The residence correlation functions of the central bulk water region and the different interface regions show that the long-range diffusion is faster in the bulk than at the interfaces (Figure \ref{fig:correlation_residence}). The residence lifetime of the O atoms in the central region is about 9\,ns. Since the $\{100\}_\mathrm{Li}$ interface has only a very thin strongly bound layer of H$_2$O molecules, the residence lifetime of about 14\,ns is still similar to bulk water, which is expected due to the weak interaction of this surface with the liquid. However, the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ and $\{110\}_{\mathrm{LiMnO}_2}$ interfaces containing large amounts of OH$^-$ ions bind the hydroxide/strongly bound water layer leading to higher residence lifetimes of about 37 and 31\,ns, respectively. For the $\{110\}_{\mathrm{MnO}_2}$ interface only very few oxygen species in the strongly bound layer can escape from the interface during the 5\,ns simulation. We note that the relative comparison of these numbers is more valuable than the absolute numbers because the underlying DFT method may generally tend to under- or overstructure water leading to faster or slower diffusion than obtained experimentally.\cite{Lin2012, Gillan2016, Chen2017}
\subsection{Proton transfer}\label{sec:proton_transfer}
An atomistic understanding of site-specific reaction mechanisms is crucial in the bottom-up development of advanced heterogeneous catalysts. The reactivity of solid surfaces can be highly dependent on the interface structure. The identification of active solid surface sites can therefore target the design of particle shapes and sizes yielding a higher activity. For instance, we have seen that the dissociation of water molecules is much more frequent at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, $\{110\}_{\mathrm{LiMnO}_2}$, and $\{110\}_{\mathrm{MnO}_2}$ interfaces compared to the $\{100\}_\mathrm{Li}$ interface. Since the initial steps of the proposed OER mechanism at Li$_x$Mn$_2$O$_4$ include the formation of two OH$^-$ ions on top of [Mn$^\mathrm{III}_2$Mn$^\mathrm{IV}_2$O$_4$]$^{6+}$,\cite{Cady2015} PT reactions at the interface are likely to play an important role.
\begin{figure*}[htb!]
\centering
\includegraphics[width=\textwidth]{Figure_10.pdf}
\caption{Decadic logarithm of the PT reaction rates $\nu_\mathrm{PT}$ between different water species as a function of the mean $z$ coordinate of the transferred H atom for the (a) $\{100\}_\mathrm{Li}$, (b) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, (c) $\{110\}_{\mathrm{LiMnO}_2}$, and (d) $\{110\}_{\mathrm{MnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems. The solid phase is highlighted by a gray background.}\label{fig:PT_species}
\end{figure*}
\begin{figure}[htb!]
\centering
\includegraphics[width=\columnwidth]{Figure_11.pdf}
\caption{Vectors connecting the oxygen atoms participating in PT events up to 7.5\,{\AA} from the interface projected onto the $xy$ plane of the (a) $\{100\}_\mathrm{Li}$, (b) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, (c) $\{110\}_{\mathrm{LiMnO}_2}$, and (d) $\{110\}_{\mathrm{MnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems. Missing translational symmetry is caused by limited sampling in the finite simulation time. A ball model of the ideal vacuum solid surface is shown in the background for reference. The colors represent the different water species participating in the PT and are defined in Figure \ref{fig:PT_species}.}\label{fig:PT_directions}
\end{figure}
\begin{figure}[htb!]
\centering
\includegraphics[width=\columnwidth]{Figure_12.pdf}
\caption{Top views of the dominant PT reactions in the (a) $\{100\}_\mathrm{Li}$, (b) and (c) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$, (d) $\{110\}_{\mathrm{LiMnO}_2}$, and (e) $\{110\}_{\mathrm{MnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems projected onto the $xy$ plane. In this figure all Mn oxidation states are colored violet.}\label{fig:PT_reactions}
\end{figure}
The PT reaction rates in Figures \ref{fig:PT_species} (a) to (d) are obtained from changes in the H atom assignments to the closest O atom, i.e., it is the number of assignment changes per unit time. The projected networks of the PT reactions near the interfaces are shown in Figures \ref{fig:PT_directions} (a) to (d). In the ideal case, all PT processes from one local minimum configuration to another one are counted, while proton rattling events, in which H atoms quickly jump back and forth between O atoms without settling on one of them, are excluded. However, an unambiguous distinction of both types is not possible. We decided to use a sampling interval of 0.1\,ps to reduce counting of proton rattling events because in this time interval a new local minimum can often be adopted. This approach yields a lower boundary of the absolute reaction rates since some PTs can be missed if forward and backward reactions happen rapidly. Also, nuclear quantum effects, which are not included in our simulations, are expected to increase the PT reaction rates.\cite{Litman2018} In the Supplementary Material we calculate the PT reaction rates of shorter trajectories using a reduced sampling interval of 1\,fs, while requiring that the assignment of the H atom to the closest O atom has to be maintained for 50\,fs after a transition to be counted as PT reaction. This approach yields mostly similar reaction rates of the frequent processes, while rare processes are sampled worse due to the shorter total simulation time (Supplementary Material Figures S10 (a) to (d)). Since the same approximations have been applied in the analysis of all PT processes, errors are expected to cancel to some extent resulting in comparable relative reaction rates. The PT reactions are classified by the water species involved in the reaction. In case a certain ion or molecule is participating in multiple PTs during a single sampling interval, the overall process is decomposed into the respective elementary steps. Since the PT rates of forward and backward reactions are equal in equilibrium according to the principle of detailed balance, their average has been calculated.
At the $\{100\}_\mathrm{Li}$ interface the dominant PT mechanism is the hydroxylation of the solid surface (Figure \ref{fig:PT_species} (a)), which has been called ``surface PT'' in earlier work.\cite{Quaranta2017, Hellstroem2019, Quaranta2019} An O$^{2-}$ ion from the solid accepts one proton from an H$_2$O molecule such that two OH$^-$ ions are formed (Figure \ref{fig:PT_reactions} (a)). The backward reactions most often happen between the same atoms (Figure \ref{fig:PT_directions} (a)) since the OH$^-$ ions are not long-living at the $\{100\}_\mathrm{Li}$ interface. PTs between H$_2$O molecules and OH$^-$ ions in the first liquid layer, i.e., ``adlayer PT'',\cite{Quaranta2017} happen rarely. Due to the effectively zero-dimensional proton transport by these PT reactions, the long-range transport is only given by diffusion and not by PT reactions following a Grotthus-like mechanism. In the liquid phase we can observe rare events of the autoionization of water, i.e., the formation of OH$^-$ and H$_3$O$^+$ ions from two H$_2$O molecules (very small green peaks between $-15\,\text{\AA}<z<15\,\text{\AA}$ in Figure \ref{fig:PT_species} (a)). These PT processes in the bulk can also be observed in some of the other simulations which are not shown in Figure \ref{fig:PT_species}. The OH$^-$ and H$_3$O$^+$ ions always recombine very quickly. This behavior is expected as a pH value of 7 in bulk water would mean that only a concentration of $10^{-7}\,\mathrm{mol\,l}^{-1}$ H$_3$O$^+$ ions are present. Consequently, there would be around 0.1 H$_3$O$^+$ ions in one of the $5\cdot10^4$ structures of a simulation trajectory including about $10^3$ H$_2$O molecules each.
The PT paths are completely different at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface. Here, the dominant mechanism is the transfer of one proton from a surface OH$^-$ ion to an H$_2$O molecule (Figure \ref{fig:PT_species} (b)). The H$_3$O$^+$ ion can be seen as transition state of the PT which rapidly transfers another proton to an O$^{2-}$ ion of the surface (Figure \ref{fig:PT_reactions} (b)). Such solvent-assisted PTs have also been observed at ZnO-water surfaces in previous work.\cite{Hellstroem2019} Surface PTs (violet) can be observed at the same locations as the solvent-assisted PTs (blue) in Figure \ref{fig:PT_directions} (b). Hence, the hydroxylation is in dynamic equilibrium at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface. The difference to the $\{100\}_\mathrm{Li}$ interface is that the equilibrium of surface PT reactions is shifted towards the occurrence of OH$^-$ ions at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface. Another frequent PT shown in Figure \ref{fig:PT_reactions} (c) happens in the perpendicular direction to the process in Figure \ref{fig:PT_reactions} (b) making the proton transport two-dimensional (Figure \ref{fig:PT_directions} (b)). Different dimensionalities of PT networks at different solid surfaces have been observed for ZnO-water interfaces before.\cite{Hellstroem2019} In the PT shown in Figure \ref{fig:PT_reactions} (c) the proton is transferred within the first liquid layer from an H$_2$O molecule to an OH$^-$ ion, i.e., adlayer PT (red).\cite{Quaranta2017} Further, the aforementioned intermediate H$_3$O$^+$ ions can alternatively transfer another proton to an adjacent H$_2$O molecule in the first liquid layer in the same direction as the PT in Figure \ref{fig:PT_reactions} (c) (yellow lines in Figure \ref{fig:PT_directions} (b)). In this way, long-range proton transport by PT reactions at the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ interface is possible. The intermediate H$_3$O$^+$ ions also explain the relatively high concentration of H$_3$O$^+$ in Figure \ref{fig:oxygen_species} (b).
As for the $\{100\}$ interfaces, surface PTs are a large fraction of the PT reactions at the $\{110\}_{\mathrm{LiMnO}_2}$ interface (Figure \ref{fig:PT_species} (c)). The dominant PT reaction, where the proton of an H$_2$O molecule is transferred to the O$^{2-}$ ion of the first solid layer, can be followed by a PT from an OH$^-$ ion of the first solid layer to the just created OH$^-$ ion (Figure \ref{fig:PT_reactions} (d)). The OH$^-$ ion in the first liquid layer is therefore an intermediate species in this two-step transfer (violet zigzag lines in Figure \ref{fig:PT_directions} (c)). In contrast to the other interfaces, OH$^-$ ions in the first liquid layer can also transfer their proton to another OH$^-$ ion forming surface exposed O$^{2-}$ ions at the $\{110\}_{\mathrm{LiMnO}_2}$ solid surface, i.e., H$_2$O molecules are deprotonated twice (second closest violet peak to the interface in Figure \ref{fig:PT_species} (c) and vertical violet lines in Figure \ref{fig:PT_directions} (c)). In addition, adlayer PTs can happen as well yielding in total a two-dimensional PT network.
At the $\{110\}_{\mathrm{MnO}_2}$ interface PTs between H$_2$O molecules and OH$^-$ ions are dominant (Figure \ref{fig:PT_species} (d)). In contrast to the other interfaces, these reactions can also form H$_2$O molecules in the first solid layer. Hence, the O$^{2-}$ ions of the solid can be protonated twice. The protonation of an OH$^-$ ion in the first solid layer by an H$_2$O molecule from the first liquid layer is shown in Figure \ref{fig:PT_reactions} (e). This reaction can be an indication of solid surface reconstruction and dissolution at longer time scales as the boundary between solid and liquid changes. Surface PTs have been observed at the $\{110\}_{\mathrm{MnO}_2}$ interface as well. The PT pathways form again a two-dimensional zigzag pattern at the $\{110\}_{\mathrm{MnO}_2}$ interface (Figure \ref{fig:PT_directions} (d)). The PT networks at the $\{110\}$ interfaces are thus very different to those at the $\{100\}$ interfaces (Figures \ref{fig:PT_directions} (a) to (d)).
\begin{figure*}[htb!]
\centering
\includegraphics[width=\textwidth]{Figure_13.pdf}
\caption{Decadic logarithm of the concentration $c$ for different oxygen species as a function of the oxygen $z$ coordinate of the (a) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ and (b) $\{110\}_{\mathrm{LiMnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems and decadic logarithm of the PT reaction rates $\nu_\mathrm{PT}$ between different water species as a function of the mean $z$ coordinate of the transferred H atom for the (c) $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ and (d) $\{110\}_{\mathrm{LiMnO}_2}$ Li$_x$Mn$_2$O$_4$-water interface systems. The solid phase is highlighted by a gray background.}\label{fig:Grotthuss}
\end{figure*}
In conclusion, PT processes are much more frequent at the interface compared to the bulk liquid. Many different processes can happen which can be very specific to the interface structure. Larger interface areas in principle increase the PT reaction rates, but some reactions, as observed, e.g., for solvent-assisted PTs and the formation of surface exposed O$^{2-}$ ions, only happen at particular sites. Only an increase of these surface sites would lead to a higher activity highlighting the importance of the site-specific atomistic understanding. The reaction mechanisms at the different interfaces determine how fast long-range transport of reactive species by PT reactions can be. The strongly bound water structure of the interface promotes alternating forward and backward reactions. The rate itself does not seem to be an sufficient indicator of the transport. An increased number of possible paths can increase the effective transport of the protons. The highest activity is found between the first solid and first liquid layer followed by processes taking place within the first liquid layer. Consequently, the structure of the first liquid layer is also of major importance for the reactivity.
As shown already in Figure \ref{fig:oxygen_species} a very few OH$^-$ and H$_3$O$^+$ ions can escape from the interface. In two of the simulations this process was successful as shown in Figures \ref{fig:Grotthuss} (a) and (b). The OH$^-$ ion in Figure \ref{fig:Grotthuss} (c) and the H$_3$O$^+$ ion in Figure \ref{fig:Grotthuss} (d) diffuse via the Grotthuss mechanism through the bulk. The H$_3$O$^+$ thereby escaped from the left interface and diffused for about 265\,ps in the bulk until it recombined at the initial interface. The diffusion via the Grotthuss mechanism is very fast as all PT events in the bulk originate from a single H$_3$O$^+$ species present only for 5.3{\%} of the simulation time. The PT rate from H$_2$O to OH$^-$ divided by the OH$^-$ concentration in the case of the $\{100\}_{\mathrm{Mn}_2\mathrm{O}_4}$ system, i.e., the difference between the red graph in Figure \ref{fig:Grotthuss} (c) and the blue graph in Figure \ref{fig:Grotthuss} (a), shows that the OH$^-$ ion in the bulk undergoes more than one PT per picosecond. In the liquid layer closest to the interface the rate is below one per picosecond. Liquid layers further away from the interface can also reach as high reaction rates per particle as in the bulk liquid. Thus, the high concentration of reactive species at the interface is the reason for the high PT reaction rates. The very specific reaction paths present at the interface cannot take place in the bulk.
\section{Conclusion}
A high-dimensional neural network potential has been used to perform first principles-quality simulations of several Li$_x$Mn$_2$O$_4$ spinel-water interfaces. These simulations reveal spontaneous water dissociation and the formation of a hydroxide layer at several solid surface cuts. The different water species can arrange in strongly bound water layers with significantly reduced mobility at the interface, whose structures are highly dependent on the underlying solid surface. The oxygen atoms of the different water species in the first liquid layer tend to mimic the bulk oxygen coordination of Li and Mn ions because these binding sites are preferred by the solid. A second strongly bound water layer is formed at interfaces with a large fraction of hydroxide ions. Its structure deviates from the face centered cubic lattice of bulk oxygen to optimize the interaction with the bulk liquid.
A high-dimensional neural network for spin prediction allows to determine the influence of the solid-liquid interface on the Mn oxidation state distribution. Weak interaction with water seems to be an indicator for only little formation of Mn$^\mathrm{II}$ ions at the interface. An increasing number of coordinating oxide ions of interfacial Mn ions stabilizes higher Mn oxidation states. The mobility of e$_\mathrm{g}$ electrons is increased in the vicinity of the interface compared to the bulk. Consequently, higher battery performance can be achieved with smaller particles sizes while smaller particles are also more prone to degradation since soluble Mn$^\mathrm{II}$ ions are formed at the interface.
Proton transfer occurs between water molecules and oxide ions at the solid surface. The formed hydroxide ions originating from water molecules are stabilized in the vicinity of Mn ions. Active sites at the interface increase the amount of water dissociation significantly compared to the bulk liquid. Beyond the solid surface, the reactivity is also highly dependent on the structure of the water contact layer because a large part of the proton transfer reactions takes place within this layer. The mobility of the water species at the interface is reduced. However, the protons can be transported by PT reactions depending on the effective dimensionality of the surface-specific proton transfer network. The number of possible paths as well as the underlying mechanisms are highly dependent on the interface structure. Therefore, a larger interface area increases the reactivity only if this interface includes the respective active sites.
In summary, in the present work a first step has been taken towards an understanding of site-specific reactions at the atomic scale to obtain control tactics for higher performance and durability of battery materials as well as higher reactivity of catalysts.
\section*{Supplementary Material}
See Supplementary Material for (I.A.) construction of the reference data set,\cite{Artrith2012, Eckhoff2019, Eckhoff2020a, Eckhoff2020b, Eckhoff2021} (I.B.) determination of the atom-centered symmetry function parameters,\cite{Behler2011} (I.C.) RuNNer settings,\cite{Eckhoff2020a, Eckhoff2020b} (II.A.) Jahn-Teller distortions,\cite{Eckhoff2020a} (II.B.) charge conservation,\cite{Zurbenko1986, Eckhoff2020b} (II.C.) hydrogen bonds, (II.D.) structure of the interface, (II.E.) residence correlation function, and (II.F.) proton transfer reaction rates.
\begin{acknowledgments}
This project was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 217133147/SFB 1073, project C03. We gratefully acknowledge computing time provided by the Paderborn Center for Parallel Computing (PC$^2$) and by the DFG project INST186/1294-1 FUGG (Project No.\ 405832858). Discussions with Peter E. Bl\"ochl are gratefully acknowledged.
\end{acknowledgments}
\section*{Data Availability}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
\section*{Conflicts of Interest}
The authors declare no conflicts of interest.
\section*{References}
|
\section{Introduction} \label{sec:Introduction}
Autonomous underwater vehicles (AUVs) with a capability of manipulation have significant potential for industry, including environmental conservation, aquaculture, oil and gas, and infrastructure maintenance \cite{Debruyn2020medusa}. Existing working-class vehicles with manipulators (typically weighted several tons) have been successfully utilised for deep, off-shore applications, particularly in the oil and gas industry. Recently, small-scale underwater vehicles (drones) with manipulators have emerged as a new robotic technology, enabling the shallow- and cluttered-water applications, such as pylon inspection and cleaning tasks \cite{Le2020the} and underwater panel operations \cite{Cieslak2015auto}.
\begin{figure}[t]\centering \vspace{12pt}
\epsfig{figure=./figs/under-ice-robot.jpg,width=0.9\linewidth}
\caption{An initial (non-pantograp based) prototype of an AUV developed at the UTS for the under-ice measurement in a test water-tank environment. The AUV is equipped with foldable arms with actuators that enable deployment through a small ice-hole. A white plastic sheet is used to simulate an iceplate and a simple spring-loaded probe is used in this initial design. The tether line is used to collect optical measurements in a host system.} \label{fig:rov}
\end{figure}
Recently, it has drawn attention to using the AUVs for polar missions, such as sampling and measuring micro-algae bioactivity under thin ice plates. Comprehensive monitoring of the algae requires direct measurements, including ice thickness and photosynthesis rate from optical sensors that contact the under-ice microbial community \cite{Manes2009small}. This sensor requires sufficient contact time (e.g. one second) to acquire a reliable measurement to function correctly. To date, this sensor has been deployed by human divers or heavy articulated arms since the requirements cannot be met by existing AUV platforms. Due to fatigue, safety, cost, reachability, and time windows of operations, current approaches usually cover only a small portion of areas of interest.
\begin{figure*}[t]\centering
\epsfig{figure=./figs/panto_diagram1.jpg,width=0.9\linewidth}
\caption{(left) The kinematic diagram of a pantograph showing two linkage bars and two joints. (right) If the pantograph constraint is enforced at the elbow joint, the end position is constrained to the vertical line.} \label{fig:1}
\end{figure*}
We provide a novel AUV system with a folding mechanism to address the challenges, which results in a compact geometry suitable for deployment through a small hole in the ice sheet. With the thrusters extended, the interventional AUV can provide sufficient control capabilities to stabilise the pose (position and orientation) against the water flow. More importantly, the AUV offers the ability of obtaining measurements by contacting the sea ice. Figure \ref{fig:rov} shows the AUV with the arm extended in a mock-up ice measurement in a water-tank experiment. A white plastic sheet was used to simulated an iceplate, and the AUV was deployed through a small radius hole ($<30$cm) simulating an ice-hole or a moonpool in an operating ship.
This approach will enable the measurements of undersea ice photophysiology across a large enough area to extrapolate an area estimate of the photophysiological stress of the microalgal communities, such as non-photochemical quenching of photosynthesis or a decline in chlorophyll content. These phenomena are known to be early indicators of environmental changes, disease infection, and nutrient deprivation in photoautotrophs \cite{Manes2009small}. Pollutants, invasive organisms, and sudden changes in the light or temperature all can negatively impact the microalgae under the sea ice, with consequences for the rest of the ecosystem as the net primary productivity declines. This work addresses this environmental changes by developing a lightweight prototype the AUV platform suitable to collect measurements with reliable contacts to the ice surface. The initial design used a simple spring-loaded probe which, however, introduced spring deflection and nonlinear contact forces. In this work, we investigate the use of pantograph mechanism to generate a reliabe contact force.
The remainder of the article is outlined as follows. Section~\ref{sec:related} provides the related work in AUVs with manipulators for interventional applications, and Section \ref{sec:panto} details the kinematic model of the pantograph. Section \ref{sec:design} provides the prototype design of the AUV and pantograph, and Preliminary experimental results are provided in Section~\ref{sec:pre}, analysing the performance of the pantograph and visual navigation system. Section~\ref{sec:conclusions} will conclude with future direction.
\begin{figure*}[t]\centering
\epsfig{figure=./figs/antarctic_foldable_arm.png,width=0.8\linewidth}
\caption{A CAD model of the pantograph-based AUV with arms folded (left), and with the arms stretched out (right). It is also equipped with a stereo-inertial system for visual localisation.} \label{fig:cad}
\end{figure*}
\section{Related Work} \label{sec:related}
AUVs have been used in many underwater applications, such as seabed survey, shipwreck search, and oil and gas industry. Such systems have been developed and studied through several projects, such as MEDUSA \cite{Debruyn2020medusa}, and PANDORA projects \cite{Maurelli2016the}. For the ice-covered ocean and polar applications, an AUV called PUMA was developed for the exploration of the Arctic seafloor \cite{Kunz2008deep}. Several underwater robots have been designed for the polar scenarios, most of them are cubic or torpedo-shaped, and some are even spherical. However, none of them can provide satisfying contact force control to obtain measurements from sensors (possibly fragile) contacting objects of interest under various uncertainties. In addition, above mentioned AUVs are not designed to operate under strong intertidal disturbances. Deployment through a small-radius ice hole or a moonpool from a ship imposes further restrictions on the design of AUVs.
\cite{Dayoub2015robotic} have demonstrated an interventional AUV in culling crown-of-thorns starfishes (COTS) in the Great Barrier Reef environment. The developed platform, COTSbot, is equipped with an injection-purpose 3-DOF manipulator underneath the vehicle, effectively shooting an end-effector to the target starfishes. \cite{Le2020the} have successfully demonstrated an underwater blast-cleaning robot, called a submersible pylon inspection robot (SPIR), at the Windang Bridge NSW, which utilises a 3-DOF manipulator for the jet-blasting. \cite{Debruyn2020medusa} demonstrated dual-robots with manipulators for underwater grasping and sampling applications. Although the underwater manipulator technology is gaining much maturity, the existing robot arms are still expensive due to the water-tight design, heavy weight due to the motors in the joints, computational complexity for Cartesian position control, and any impedence/addittance control requires additional torque/force sensors.
A practical yet effective interventional mechanism is a pantograph which has been widely applied in transportation systems. A pantograph is a redundant linkage arm that can be used to make a reliable contact between the arm and the environment. In electric trams and trains, a 'Z'-shaped pantograph has been used to provide the power, and there has been active research in maintaining a reliable contact and enough lift-forces between the pantograph and the power lines \cite{Yao2021analysis}. One of the key benefits is its simple mechanism, yet allowing the passive or active impedance/admittance control, providing a constant compliant force across a range of height variations. A manipulator with a force/torque sensor can achieve the constant force but typically requires additional cost. Another benefit is the pantograph can be driven by a passive spring or an active motor at the base, which reduces the requirements of the water-tight operation (the active servo can be housed within a vehicle body). It has also a low inertial for fast movement and no complex dynamics and control are need.
This mechanism can also be applied for ground and aerial applications, enabling a practical solution for the compliant interaction with the controlled reaction from the contact. Unfortunately, this mechanism has not been much investigated for robotics applications, and this work study the feasibility of the pantograph system.
\begin{figure*}[th!]\centering
\epsfig{figure=./figs/pantograpth_01.jpg,height=0.35\linewidth} \;\;\;\qquad
\epsfig{figure=./figs/stereo_rig.jpg,height=0.35\linewidth}
\caption{(left) A pantograph prototype (a demo link is available) and (right) a stereo-inertial system mounted in an enclosure for the test in an underwater environment (an outdoor rock pool).} \label{fig:panto}
\end{figure*}
\section{Pantograph Kinematics and Dynamics}\label{sec:panto}
Figure 2 illustrates the kinematics and geometry of a pantograph system, showing two joint angles ($\theta_1, \theta_2$) with the dimensions of linkages ($l_1, l_2$). The end position $p(x, y)$ of the pantograph then becomes,
\begin{align}\label{eq:1}
\left [ \begin{array}{l} x \\ y \end{array} \right ] =
\left [ \begin{array}{l}
l_1\cos\theta_1 + l_2\cos(\theta_1+\theta_2) \\
l_1\sin\theta_1 + l_2\sin(\theta_1+\theta_2)
\end{array} \right ],
\end{align}
The two linkage bars needs to satisfy the pantograph constraint between two bars,
\begin{align}\label{eq:constraint}
2\theta_1+\theta_2 = \pi,
\end{align}
which stems from the fact that the upper link bar must move twice as fast as the lower bar. A pair of gears with a $2:1$ ratio is used to implement this constraint. By applying this constraint, Equation (\ref{eq:1}) becomes
\begin{align}\label{eq:2}
\left [ \begin{array}{l} x \\ y \end{array} \right ] &=
\left [ \begin{array}{l}
l_1\cos\theta_1 + l_2\cos(\pi - \theta_1) \\
l_1\sin\theta_1 + l_2\sin(\pi -\theta_1)
\end{array} \right ] \\
&=\left [ \begin{array}{l}
(l_1 - l_2)\cos\theta_1 \\
(l_1 + l_2)\sin\theta_1
\end{array} \right ] \\
&=\left [ \begin{array}{c}
0 \\
2l_1\sin\theta_1
\end{array} \right ], \;\;\; \text{if}\;\; l_1 = l_2.
\end{align}
That is, by choosing the lengths of two bars equal, the motion of the endpoint can be constrained to the vertical line.
From the pantograph constraint Equation \ref{eq:constraint}, the angular velocity of two joints satisfies
\begin{align}\label{eq:3}
2\omega_1+\omega_2 = 0.
\end{align}
At the base, the acutator utilise a lever arm with the length of $r$ in the figure (perpendicular to the lower linkage bar) to which a driving input force $f_\text{i}$ applied. The resulting torques become
\begin{align}\label{eq:decoupsys}
\tau_2 &= f_\text{o} l_2\cos(\pi-\theta_1) \\
\tau_1 &= f_\text{i} r\sin(\alpha).
\end{align}
Using the fact that $\tau_1 = 2\tau_2$ and $\theta_1-\alpha=90^\circ$, the input and output force ratio becomes
\begin{align}
\frac{f_{\text{o}}}{f_{\text{i}}} = \frac{1}{2}\frac{r\sin(\alpha)}{l_2\cos(\pi-\theta_1)} = \frac{r}{2l_2} = \text{const.}
\end{align}
showing that the output force $f_{\text{o}}$ is not affected by the configuration of the pantograph but only by the input force which can be maintained as constant from a servor motor.
\section{Design of AUV with Foldable-Arms and Pantograph}\label{sec:design}
\begin{figure*}[t]\centering
\epsfig{figure=./figs/panto_test2.jpg,width=0.49\linewidth}
\epsfig{figure=./figs/panto_test1.jpg,width=0.49\linewidth}
\caption{Pantograph results with the measured force at the end point with respect to the height ($y$), showing nearly-constant force measurements along the height range of $100$mm - $300$mm. The right image is obtained by using a better-quality weight sensor.} \label{fig:meas}
\end{figure*}
\begin{figure*}[th!]\centering
\epsfig{figure=./figs/stereo_matching.jpg,width=1\linewidth}
\caption{Feature matching results between two consecutive images from the right camera, showing the features detected (left in blue) and matched/tracked (right, in green). Most features were detected in a narrow field-of-view of the camera due to the poor visibility. } \label{fig:feature}
\end{figure*}
A custom-built underwater AUV is constructed aiming to perform the under-ice algae monitoring in an Antarctic environment. One of the critical requirements is the easy deployment of the robot from a small-radius ice hole or the moonpool of a surveyor ship. Once deployed, it should maintain a stable position with variable ocean currents and uneven ice surface, as well as contacting the surface to measure the reflectance and wavelength-dependent fluorescence emission using an optical sensing instrument on an probing arm. Figure \ref{fig:cad} shows the computer-aided design (CAD) of the robot with the folderble actuation arm mechanism. This folding configuration enables easy deployment and reliable torque control of the vehicle. The initial prototype of the robot was also shown in Figure \ref{fig:rov} in a mock-up environment, which uses a non-pantograph-based probe (a simple spring-loaded rod) and a short-base line stereo-camera which is not suitable for the localisation.
To address these limitations, a pantograph and a visual-inertial module are developed separately, aiming for a full integration at later stage. A prototype of the pantograph is constructed using aluminium frames, which is lightweight and durable for the demonstration and experiments. To achieve the constant interaction force with adjustable force, we used a modular design: aluminium based frames for the upper and lower arms, 3D printed base frame for easy modification, a constant force spring mechanism, and a pantograph constraint coupler. The aluminium arms are accompanied with 3D printed components, which allow for more adjustments to be made to the pantograph simply by unbolting parts and moving them around the different hole positions.
Figure \ref{fig:panto}(left) shows the constructed pantograph\footnote{A demo video (by Mr Jacob Prideaux-Remin):\href{https://youtu.be/LtjLEpsexYA}{https://youtu.be/LtjLEpsexYA}}. We chose a constant-force spring to give the pantograph mechanism a constant force along the lever arm. This is due to the fact that the constant-force spring has a minimal radius change when being pulled and uncoiled. This design is very similar to what we can find in a tape measure. The spring is coupled to a lever arm which is directly connected to the lower pantograph arms. The lever arm can be easily changed with different lengths and different arm shapes depending on the required force output.
The kinematic constraint between the joints (Equation \ref{eq:constraint}) is imposed via a chain and sprocket, inspired by the train-style constant-force pantograph and its ability to not stretch under the load that the pantograph would be experiencing. It also supports laterally, so it does not deflect away from the pantograph arms during operation. A threaded adjustment arm has been used to adjust the pantograph with precision to achieve a linear movement by adjusting the chain length along the arm. Figure \ref{fig:panto}(right) shows the enclosure of the visual-inertial localisation system for the purpose of testing in an outdoor underwater environment. The localisation system utilises a stereo-inertial camera and an embedded computer in a water-tight enclosure. A VINS-fusion ROS package is used to test the performance in an outdoor rock pool environment.
\section{Preliminary Results}\label{sec:pre}
Due to the limited access to the water-tank facility at the UTS, we only presents the individual test results of the pantograph and localisation system in this work, although we are working towards the full integration and testing.
\begin{figure*}[th]\centering
\epsfig{figure=./figs/vins_result.jpg,width=0.8\linewidth}
\caption{(left) Preliminary results from the VINS fusion using the ZED2 camera. (right) Odometry output from the ZED2 camera showing significant drift in the position. The ground truth was not available for accurate evaluation, but the estimated trajectory properly captured the sweeping lawn-mower pattern during the experiments.} \label{fig:vins}
\end{figure*}
The designed pantograph is tested in the laboratory to analyse the force-height relationship. The pantograph mechanism is set at $100$mm between the lower axis and top of the pantograph arm, and the force is measured using a simple scale. It is then extended vertically by $25$mm, and then the force is measured, which process is repeated until $400$mm height is reached. Figure \ref{fig:meas} shows two test results of the force-height measurements. The first test (left) shows a relatively constant force between $100$ mm and $300$ mm while showing stong nonlinear effects when the height exceeds $300$mm. The second test (right) with a different measuring device is able to achieve a relatively constant force output between $100$mm and $350$mm. The variations of results stem from several factors, such as the friction in the system, which can cause resistance against the spring mechanism at certain points. Additionally, there is friction between the rope used and the pulley, which could cause the pantograph to exert all its force as the string is holding it back. The measuring device used is also a low-cost one causing some inaccuracy. However, the experiments clearly demonstrate the feasibility of the pantograph achieving a nearly constant force at the endpoint over a range of height variations.
A stereo-inertial navigation system is assembled and tested in a realistic underwater environment, an outdoor rock pool, as shown in Figure \ref{fig:panto} (right), which consists of a ZED2 camera and an NVIDIA NX embedded computer in a water-tight enclosure. The water was relatively clear with natural features in the rock wall and floor. The camera system was pointed forward and followed a lawn-mower trajectory pattern. Figure \ref{fig:feature} (left) shows the features detected from the right camera, and (right) the tracked feature (right) in the subsequent image. It can be observed that the detection range is quite narrow and short due to the low illumination and blurry features. Figure \ref{fig:vins} shows the preliminary localisation results using the VINS fusion package (left) and the direct ZED2 odometry output (right). It can be clearly seen that the ZED2 odometry output (IMU and stereo) is drifting rapidly, while the VINS fusion method shows more consistent accuracy. No ground truth trajectory is recorded in this preliminary experiment, but the sensing system followed a lawn-mower sweeping pattern which can be seen in the estimated trajectory. From the post-analysis, the narrow field-of-detection caused very sparse loop-closures between the tracks, distorting the straight trajectory. We are planning further experiments with a trajectory that has enough overlapping between the neighbouring routes and testing different viewing angles of the camera to maximise the chance of loop-closures.
\section{Conclusions}\label{sec:conclusions}
We presented a novel interventional AUV system that incorporates a pantograph-based interventional mechanism and a folding actuation. We reviewed the kinematics and dynamics of the pantograph system and provided the prototypes of the AUV platform and the pantograph system. Preliminary experimental results confirms that the system can maintain a constant contact force of around $190$ gram-force over the height variations of $100$mm - $300$mm, sufficient for the purpose of the optical measurements. The visual-inertial localisation system also showed consistent localisation performance in the realistic underwater condition. The future direction is to integrate the pantograph mechanism on the AUV and the autonomous operation of the robot using the visual-inertial localisation outputs.
\section*{Acknowledgments}
The prototype of the pantograph was constructed as part of a student design project at the UTS: Mr Jacob Prideaux-Remin, Mr Van-Troung Lam, Mr Darren Hayes, Mr Abdullah Azam, and the navigation system was tested by Mr Aman Singh. The initial AUV control system was designed by Dr Wenjie Lu at the UTS.
|
\section{Introduction}
\label{sec:intro}
The HPC community has been slow to adopt high-level, managed
languages---increasingly popular in industry---due to their apparent
performance overheads and due to the fact that rarely are these languages
designed with the parallelism or behavior of large-scale scientific workloads
in mind. However, such languages hold promise because they can relinquish the
programmer from the burden of low-level details of the system like pointers,
memory management, and type specification. This frees domain scientists from
having to be systems programmers and allows them to instead focus on their
application logic, thus reducing the ``time to science.''
The {\em de facto} standard for programming large-scale machines involves some
combination of MPI paired with C/C++/Fortran for inter-node parallelism, and
runtime support such as OpenMP~\cite{DAGUM:1998:OPENMP},
pthreads~\cite{BUTTLAR:1996:PTHREADS}, Cilk~\cite{BLUMOFE:1996:CILK}, or
Qthreads~\cite{WHEELER:2008:QTHREADS} for handling intra-node parallelism.
More recently, applications leverage heterogeneous hardware by using
accelerator frameworks like OpenCL~\cite{STONE:2010:OPENCL},
OpenACC~\cite{OPENACC} or CUDA~\cite{CUDA_PROG_GUIDE}. For domain experts,
this makes the problem worse, since not only must they worry about the ``how''
of parallelism across nodes, they must also deal with accelerator offloading,
synchronization, virtual memory, and other low-level details.
Ideally, the programmer would describe their problem in a high-level language
and the underlying machinery of the language would take care of mapping the problem
onto parallel and heterogeneous hardware. While we are not there yet, the
community {\em has} made strides in language design, compiler technology, and
parallel programming frameworks. Recent languages like
Regent~\cite{SLAUGHTER:2015:REGENT} and more established languages like
Chapel~\cite{CHAMBERLAIN:2007:CHAPEL}, HPF~\cite{MERLIN:1995:HPF},
UPC~\cite{CCS:1999:UPC}, X10~\cite{CHARLES:2005:X10}, and
Erlang~\cite{JOHANSSON:2000:HPE} all have made significant contributions in
bringing productive programming to the HPC world, yet they have not seen
widespread adoption. For a language to flourish in the HPC community, it must
be sophisticated (sporting a rich set of built-in functionality for scientific
programming), intuitive, and mature. Most importantly, however, {\em it must
perform well} for the increasingly broad set of large-scale applications that
the scientific community employs.
The Julia language\footnote{\url{https://julialang.org}} is an interesting
point on this spectrum not only because it takes inspiration from other popular
managed languages like Python and Ruby, but also because it was designed by
domain experts with scientific and technical computing in mind. It thus has
significant appeal for programmers brought up on Fortran and MATLAB. Unlike
those languages however, Julia has targeted scalable parallelism from the
start. Built atop the LLVM compiler framework, and sporting sophisticated
features like dynamic dispatch, high-performance math libraries, and JIT
compilation, Julia's single-threaded performance for scientific problems
approaches (or even surpasses) its low-level counterparts, C and Fortran. It
also features intuitive parallel programming abstractions, and others have
recently demonstrated its potential for petascale parallelism on a top
supercomputer~\cite{Regier2015CelesteVI}.
However, a large class of HPC applications involve significant inter-node
communication to achieve parallelism. A classic example is stencil computations
following the bulk-synchronous parallel (BSP) model, where in order to solve,
e.g. a partial differential equation (PDE) describing some dynamical system,
the domain is discretized and evolved in time using finite differences. While
the programmer must carefully balance the communication/computation ratio with
these methods, communication latencies due to language, library, system
software, and network interfaces can make or break performance.
As far as we are aware, the Julia language has not yet been evaluated in this
respect. Thus, in this paper, we seek to shed light on Julia's performance
for communication-intensive applications, both within and without the node. We
investigate Julia using a range of single-node and multi-node microbenchmarks,
as well as a port of a standard HPC benchmarking code encapsulating the above
behavior, high-performance conjugate gradient (HPCG).
We make the following contributions:
\begin{itemize}[nosep]
\item We provide a series of single-node and multi-node microbenchmarks for
Julia.
\item We evaluate the performance of Julia's parallel programming abstractions,
both on a single node and in a distributed environment.
\item We present the first port of a standard, large-scale, communication-intensive code (HPCG) to Julia.
\item We show through benchmarking that with a reasonable computation-to-communication
ratio, Julia can actually outperform C in distributed message passing scenarios,
demonstrating its potential for communication-intensive applications.
\end{itemize}
All code for our benchmarks and our experiments, as well as our experimental
data will be made freely available upon acceptance.
The remainder of this paper is organized as follows:
Section~\ref{sec:background} gives an overview of the Julia language and its
focus on scientific computing and parallelism. Section~\ref{sec:eval} presents the evaluation of Julia
for high-performance parallelism within the node. In Section~\ref{sec:distributed},
we analyze Julia's performance in a distributed environment. We discuss
our results in Section~\ref{sec:discuss}, outline related prior art
in Section~\ref{sec:related} and conclude in Section~\ref{sec:conc}.
\section{Background}
\label{sec:background}
In general, to be effective for efficient parallelism, a high-level language
will have high-performance primitives or constructs which support:
\begin{itemize}
\item Mapping execution contexts to hardware precisely (e.g. NUMA and processor affinity)
\item Hardware heterogeneity
\item Low-latency communication between execution contexts (threads/tasks/processes)
\item Low-latency creation and scheduling of tasks.
\item Efficient exploitation of parallelism (e.g. parallel \verb.map().s or parallel loop constructs)
\end{itemize}
Much of the underlying machinery for the above will be hidden from the
programmer.
Ideally the performance of the language will be close to the hardware's
capabilities, as is the case with lower-level languages. Some (including
Julia's designers) have made the argument that even if this is not true for the
HLL, the benefit to programmer productivity and application development time
should be considered as a counterbalance to the performance of the language,
especially since development time can take several years for large HPC codes.
This perspective may hold even more weight as new generations of domain experts
and programmers are being taught with high-level languages. Unfortunately, this
aspect of a language is more challenging to quantitatively measure in practice.
If the HLL's performance {\em can} match that of its lower-level counterpart,
all the better, since we get the best of both worlds. We seek to find out if
that is indeed the case for Julia.
Below we will outline the parallel programming primitives that Julia provides
and discuss some salient choices by the language designers.
\subsection{Overview of Julia}
\label{sec:julia}
Julia~\cite{BEZANSON:2017:JULIA} is a high-level programming language designed
for numerical computing. Our main interest in Julia stems from its promising
single-threaded performance (which the Julia developers have shown can approach
that of C and Fortran), its native parallel constructs, its extensibility, and
its rich set of user-developed packages. Particularly its performance relative
to other high-level languages, which is in part due to its usage of LLVM's JIT
compilation framework, indicates that it might be a suitable candidate for
communication-intensive HPC applications as well as for building parallel
runtime systems which support even higher-level (possibly domain-specific)
parallel languages. Some other noteworthy features of the Julia language include
multiple dispatch, dynamic types, meta-programming features, light-weight user
threads, easy support for native code invocation, and packages for distributed,
parallel computing.
Julia is already being used to solve a wide array of problems in technical
computing. For example, JuliaFEM~\cite{Rapo_Aho_Frondelius_2017} provides
a concise programming framework for solving large-scale problems using finite
element methods. Others have explored using high-level frameworks for GPU
computing with Julia~\cite{2016arXiv160400341Y}. Chen et al. investigated
using polymorphic code features in Julia applied to parallel and distributed
computing~\cite{Chen2014PrefixPolymorphism}. It has also been shown to work
well for classic scientific problems, such as convex
optimization~\cite{Udell2014ConvexJulia}, PDE
solving~\cite{Heitzinger2014JuliaPDEs}, and discrete event-driven
simulation~\cite{Kourzanov2018JuliaLiveSim}. Data scientists and AI experts are
increasingly employing Julia with success as well~\cite{Dinari2019JuliaMCMC,
Regier2015CelesteVI}.
\subsection{Parallelism in Julia}
\label{sec:parallelism}
Julia has an extensive array of features and packages that support parallel and
distributed computing. Julia natively provides an abstraction for task
parallelism called {\em Tasks}, which are essentially lightweight user threads.
Once tasks are defined they are invoked using the \verb.@schedule. macro.
{\em Workers} are an abstraction of processes and \verb.workers(). is
a function that returns an array of all workers deployed locally or
remotely. Processes do not share memory by default but can communicate
via \verb.Channels. or by using \verb.SharedArrays., discussed later
in this section. Multiple processes can be created locally and on other
nodes by calling \verb$@addprocs$.
{\em Threads} are Julia's version of kernel threads. Others have evaluated the
native threading support on multi-core
machines~\cite{Knopp2014JuliaMultiThreading}. With these threads, Julia allows
the user to wrap a value in \verb|Threads.Atomic| to avoid race conditions
while accessing or modifying shared state. Similar to OpenMP, one can indicate
that a code region is to be run on multiple threads using the \verb.@threads.
macro. A barrier is implicitly placed at the end of a \verb.@threads. block and
the code in that block will run on the number of threads specified by the
\verb.JULIA_NUM_THREADS. environment variable, again similar to standard OpenMP
programming practice. Additionally, Julia's threading package includes a set
of native atomic operations (usually provided by compiler built-ins in
lower-level languages) via LLVM, in addition to higher-level synchronization
primitives like locks, mutexes, and semaphores.
Communication between Julia processes is, by default, one-sided. It is
accomplished using the \verb.RemoteReference. and \verb.RemoteCall.
abstractions. A \verb.RemoteReference. is made for exchanging information
while \verb.RemoteCall. is used to request the remote execution of a function.
\verb.RemoteCall. returns a \verb.Future. object asynchronously---essentially
a \verb.RemoteReference.---and the result of the function can be obtained by
calling \verb.fetch. on the \verb.Future. object. The \verb.@spawn. and
\verb.@spawnat. macros are used to send work to an arbitrary and particular
process, respectively. Julia also provides an \verb.@everywhere. macro which
specifies that its wrapped expression shall run on all available workers.
Fine-grained parallelism is provided by \verb.@distributed. (for indicating
that loop iterations run in parallel) and \verb.pmap. (parallel mapping of
a function over a collection). The Julia documentation points out that if the
number of operations are large in number but not too complex (e.g. using synchronization),
\verb.@distributed. can be
used to perform them. On the contrary, if the operations are complex and not
numerous then use of \verb.pmap. is recommended.
\verb|@async| is a feature that can be used in scenarios where one wants to define
work but is not ready to have them executed right away. \verb.@async. will
launch tasks locally and can be sent to other processes or workers using
\verb.remotecall_fetch()..
Julia has macros for specifying operations as \verb.@sync. and \verb.@async. as
per need. \verb.@sync. can be employed on an otherwise \verb.@async. block and
vice versa. This, however, does not mean that Julia will block a task that uses another \verb.@async.
task. \verb.@async. can only block a synchronized function.
\verb.Channels. and \verb.RemoteChannels. allow the programmer to construct
producer/consumer queues between local and remote processes, respectively.
This is useful for transfer of primitive data across tasks. For more complex
data sharing like an array, \verb.SharedArray. and \verb.DistributedArray. are
used. A \verb.SharedArray. can be accessed by all processes in parallel, with
synchronization provided by the runtime. A \verb.DistributedArray. is similar
in concept but has the advantage of decomposing the array by distributing chunks
across local and remote processes.
Julia supports this set of parallel and distributed operations and
structures via the external packages \verb$Distributed.jl$ and \verb$Threads.jl$.
Importantly, Julia also sports a package to provide MPI bindings, which we will
leverage in Section~\ref{sec:distributed}.
While other features of Julia are beyond the scope of this paper, we refer
interested readers to other texts on the
language~\cite{DBLP:journals/corr/abs-1209-5145, Cabutto2018Julia,
Bezanson2018Julia, Zappa2018JuliaSubtyping} and its design choices, as well as
the official Julia documentation\footnote{\url{https://docs.julialang.org}}.
\section{Parallelism within the node}
\label{sec:eval}
In this section we present a series of experiments evaluating the performance
of Julia's runtime native parallelism primitives within a node compared to
low-level implementations in C. Our goal is to get a sense of what kinds of barriers to
parallel performance an HPC programmer might encounter on a compute node when
using Julia.
\subsection{Experimental Setup}
\label{sec:exp-setup}
All experiments were carried out on our in-house, 14-node cluster called {\em
mystic}. Each node comprises a dual-socket (2 NUMA nodes), 16-core (2 hardware
threads per physical core) Intel Xeon Silver SP 4108 (Skylake) CPU clocked at
1.80GHz and equipped with 64GB of 2.66 GHz DDR4 SDRAM. This machine runs
Ubuntu 18.04 LTS with a stock Linux kernel version 4.15. We configured the
BIOS to disable DVFS features.
For the Julia experiments, we use Julia version 1.3.1. All C/C++ benchmarks
were compiled with the highest optimization level (\verb.O3.) using the stock
Ubuntu version of gcc (7.5). Unless otherwise noted, all experiments were
conducted using 100 trials, with an additional 10 trials initially thrown out
to mitigate cold start effects. Timing was conducted using Julia's
native \verb.time_ns(). function and \verb.clock_gettime(). in C using the
\verb.CLOCK_REALTIME. time source. The file system configuration has no influence
on our experiments.
Many high-level languages support the notion of a {\em task}, a schedulable
unit of work which may or may not have the ability to modify state. Tasks can
essentially be thought of as glorified functions, but they can be moved around
the system. Tasks may be mapped to threads or processes, but ultimately they
should be very light-weight. The more overhead that the language introduces
when creating and managing tasks, the coarser the units of work must be to
amortize these costs, and the amount of parallelism available to the programmer
is reduced. Julia has support for such tasks (with the \verb.Task.
abstraction), which they describe alternatively as light-weight coroutines or
{\em green threads}. These are essentially user-level threads of which the
underlying OS is unaware. Our first experiment measures the latency of creating
these tasks. The baseline we compare this against is the time taken to create a pthread
(using \verb.pthread_create().). Since pthreads map directly to kernel threads
(visible to the OS scheduler), we expect that the Julia tasks should be much
cheaper to create. A comparison to an equivalent low-level user-threading package
(e.g. Qthreads~\cite{WHEELER:2008:QTHREADS} or OpenMP tasks) would be ideal
here, but we note that this is just to sanity check the task performance of
Julia. We leave such a comparison for future work.
To make the comparison as fair as possible, in Julia we measure the time to
create a task using the \verb.Task(). function followed by a call to
\verb.schedule(). which places the created task on the task scheduler's run
queue (which is local to the current Julia process on a single core). Note that
unlike most \verb.schedule(). functions, this Julia version does not
immediately yield to the scheduler.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=0.35\columnwidth]{figs/task_create_tput}
\label{fig:task-create}
} \subfigure[]{
\includegraphics[width=0.35\columnwidth]{figs/context_switches}
\label{fig:ctx-switch}
}
\caption{(a) Comparison of task creation latencies for Julia's native tasks and
kernel threads. (b) Context
switching latencies for Julia tasks and pthreads.}
\label{fig:tasks}
\end{figure}
Figure~\ref{fig:task-create} shows the results (note the log-scale on the y axis). Creation of a pthread (kernel thread)
costs roughly 20$\mu$s while Julia task creation and the subsequent schedule call
costs only about 400ns. This is consistent with the creation throughput shown in the figure, and is in line with our expectations.
However, tasks
on a remote core must be scheduled on a remote worker process, which must be
explicitly created with the \verb.addprocs(). routine. While this routine takes
roughly 400ms, one need only call it once, at runtime startup.
Often in HPC scenarios (i.e. many-task computing), there are many more tasks
than there are hardware threads available in the system. This allows a high
degree of concurrency when tasks perform blocking operations. It is important
that tasks be light-weight to maximize this concurrency. Here we measure the
context switching overhead of Julia tasks by creating two tasks in the same
worker process, having them cooperatively \verb.yield(). to each other a fixed
number of times, and approximating the latency from the number of yields in the
recorded time period. The same caveat as above applies to comparing kernel
threads with user threads. For our baseline we use the same strategy with two
pthreads pinned to the same core.
Figure~\ref{fig:ctx-switch} shows that Julia task context switching is roughly
one order of magnitude faster than for pthreads. Again, this is to be expected
since the kernel is not involved in a Julia context switch unless a task yields
and there is nothing else to run.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{figs/spawn-fetch}
\caption{Mean latencies for Julia's one-sided future-based
parallelism (spawn and fetch).}
\label{fig:spawn_fetch}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{figs/condvars}
\caption{Event notification latency for Julia and pthread condition variables.}
\label{fig:condvars}
\end{figure}
For the next experiment, we attempt to get an estimate of the overheads of
task operations between Julia workers.
Julia's native parallel processing features are
based on one-sided task-launch operations (futures).
When a future is created (using \verb.spawn().), its corresponding object is
immediately returned to the caller, without blocking. The work associated with the future may
complete immediately, or at some later point in time. The result is retrieved by
an explicit call to (\verb.fetch().) the future. This call
{\em will}
block if the execution has yet to take place.
Here we aim to investigate how the overheads of task creation relate to
the amount of work performed for fine-grained tasks.
For this experiment we created two benchmarks:
a tunable benchmark which returns the first $n$ Fibonacci numbers sequentially
and a baseline benchmark which creates a null task that performs no actual
work at all. The null task thus approximates the overheads of remote
task creation.
Figure~\ref{fig:spawn_fetch} depicts the results of spawning a remote task (on
any available worker) for various task granularities (where again, the
granularity is determined by the \verb.fib(). argument). These are compared
against the null task as we vary the number of Julia workers present in the
system. We cap our experiment at
16 worker processes since the system has 16 physical cores. Note that here we
are only spawning the task on {\em a single} remote worker, so the latency
should remain the same, independent of the number of workers. However, looking
at the figure this is not what we see. Instead, it seems that Julia's task
scheduling across workers scales roughly linearly with the number of
workers. While we make sure to create workers (\verb.@addprocs.) {\em
before} the experimental run, we suspected that Julia may be lazily creating
workers on the spawn invocation. However, since we are throwing out the cold
start trials, we believe this not to be the case. We found this effect
surprising, and we have not yet determined its cause. These results were
also confirmed by other comparisons~\cite{Huang2019JuliaDirichlet} between
inter-process runs and inter-node runs. Note that while the error bars (depicting
one $\sigma$ above and below the mean) are relatively small, we did see several
significantly large outliers in this experiment in Julia's case. The measurements
shown here are {\em after} outlier removal using Tukey's method with a constant of 1.5\footnote{
That is, all measurements not on the interval $[x_{25\%} - 1.5IRQ, x_{75\%} + 1.5IQR]$ have
been removed.}.
Often high-performance applications will need low-latency communication. This
is needed to support explicit communication between tasks (e.g. using explicit
messaging), implicit communication based on access to shared data, and event
notification (e.g. in a parallel runtime system).
Event notification can signify conditions such as the creation of
a task, the availability of data, an error condition, or the completion of
a task. These notifications are
particularly important for processes communicating via event channels.
One of the most common software event notification
mechanisms is a {\em condition variable}. A condition variable is essentially
a queue protected by a mutex, and its interface comprises three functions:
\verb.wait()., \verb.signal()., and \verb.broadcast().. A thread/task calls
\verb.wait(). to block execution until some {\em condition} becomes true (e.g.,
a unit of work is ready to execute, a message delivery has completed). The
thread will then be descheduled and put on the condition variable's queue
until the condition becomes true, thus yielding the processor to another unit of
execution. Some
other thread will then call \verb.signal(). to wake one waiting task and remove
it from the queue, or call \verb.broadcast(). to wake all waiting tasks.
A common model in runtimes/HPC applications is to have worker threads/tasks spread across cores,
where they wait on work to do with condition variables. Unfortunately, Julia's
\verb.Condition(). construct only applies to \verb.Tasks. on the same worker process,
so there is no way to use them for cross-core notifications. To approximate
this behavior, we used Julia's \verb.RemoteChannel. primitive, which is essentially a
producer/consumer queue that can be referenced by \verb.Task.s running on
different Julia worker processes. We create an unbuffered (0-element)
\verb.RemoteChannel. and issue a \verb.take(). operation to simulate a condition
variable \verb.wait()., and have a \verb.Task. on a worker process running on a
remote core issue a \verb.put(). to simulate a condition variable
\verb.signal()..
We compare against a baseline implementation written in pthreads, which creates
a remote thread whose job is to signal a condition that the local thread is
waiting on (using \verb.pthread_cond_signal().). The remote thread takes a time
stamp after signaling, and the local thread records a time stamp after waking
up. We report the latency as the difference between these two timestamps.
The violin plot in Figure~\ref{fig:condvars} shows the results.
We see a median latency of roughly 14$\mu$s for the Julia case, and a few
microseconds for notification using the pthreads condition variable. The Julia
version also shows a significantly higher variance in latency. While the
performance gap is quite large, one should realize that the Julia notification
is occurring across worker processes (OS processes) in separate address spaces,
whereas in the pthreads, case this is happening between kernel threads sharing the
same address space. While the notification latency for Julia is already fairly
small, we expect it will shrink once proper threading support (with the
attendant condition variable implementation) is mature.
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{figs/channels}
\caption{Latency of channel operations (put/get/fetch) using Julia's channel construct
and a simple producer/consumer queue written in C with pthreads.}
\label{fig:channels}
\end{figure}
In addition to simulating condition variables, channels can also be used to
implement both one-sided and two-sided communication models. Julia exposes
three operations on channels: \verb.put(). to place an object on the queue,
\verb.take(). to remove an object from the queue, and \verb.fetch(). to read
the front of the queue (without removing the object). Ideally operations on
channels will just be a read or write protected by a lock or an atomic
operation. We compare Julia's channel against a simple (unoptimized)
producer/consumer queue written in C that is protected by a mutex from the
pthreads library (\verb.pthread_mutex_t.). Figure~\ref{fig:channels}
shows the results. We see here that Julia's channel operations are all very
light-weight, each taking less than 100ns. The differences here with the
C version are not significant, but the operations that Julia provides out of
the box show promising performance.
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{figs/parallel-constructs}
\caption{Mean time to initialize a 50MB
array across execution contexts.}
\label{fig:parallel-constructs}
\end{figure}
An important requirement for any HPC-oriented high-level language is that
it exploit parallelism well. To measure how well Julia maps computations automatically
to execution contexts, we distributed the initialization of a 50MB array among worker
contexts. We use Julia's \verb.pmap(). and \verb.@distributed. features to achieve
this, and compare against an OMP version using the \verb.parallel for. looping
construct and a custom pthreads implementation in C that performs manual decomposition.
Noting the log scale on the y-axis, we see
in Figure~\ref{fig:parallel-constructs} that the mean time to initialize the
array using the parallel constructs in Julia is roughly five times smaller.
\section{Distributed Parallelism}
\label{sec:distributed}
To get a realistic picture of how viable Julia is for communication-intensive
HPC applications, we also measure its performance in a multi-node setting. An
HPC application written in an HLL ideally will have performance close to that
achieved by low-level languages, even when spread across nodes. To do this, we
leverage the \verb$MPI.jl$ package, which provides bindings to the
system-resident MPI library with a set of wrappers that allow overloading of
the typical MPI functions. For all experiments, we use MPICH v3.3a2. The hardware
configuration was previously described in Section~\ref{sec:exp-setup}. Both
Julia and C/C++ applications use the underlying system MPI library. The 14
nodes in our {\em mystic} cluster are connected in a star topology over 10Gb
Ethernet (via Intel X550 NICs) configured with an MTU of 9000. Timings in the distributed
environment are conducted by having rank 0 (the master node) record timestamps (we plan to
incorporate measurements from {\em all nodes}).
\begin{figure}
\centering
\includegraphics[width=0.5\columnwidth]{figs/bsp-topo}
\caption{Connectivity between MPI ranks in our BSP benchmark.}
\label{fig:bsp-topo}
\end{figure}
In our first set of multi-node experiments, we approximate a typical bulk-synchronous
parallel (BSP) code (e.g. a stencil computation) by creating a synthetic
benchmark which consists of compute (memory reads/writes and floating point
ops) and communication (message send/recv) steps.
Each rank will perform a configurable amount of computation between
communications, wherein computation consists of reads (a series of loads
from an array), writes (a series of stores to an array), and flops (a series of
floating point operations). As shown in the algorithm, in the communication
step, ranks send data to their forward neighbor and then receive data from
their backward neighbor. The ranks thus form a ring topology, as depicted in
Figure~\ref{fig:bsp-topo}. We compare the performance of two custom
implementations of this BSP benchmark in both C and Julia. We varied the
number of processes (MPI ranks) per node (1, 2, 4, and 8). The amount of work
performed by each rank is kept constant, thus we are technically showing weak
scaling with this experiment. We gather statistics over 100 iterations.
\begin{figure*}
\centering
\subfigure[10 communication steps.]{
\includegraphics[width=0.32\textwidth]{figs/bsp-runtime-comms-10}
\label{fig:bsp-runtimes-10}
}
\subfigure[100 communication steps.]{
\includegraphics[width=0.32\textwidth]{figs/bsp-runtime-comms-100}
\label{fig:bsp-runtimes-100}
}
\caption{BSP benchmark total execution time.}
\label{fig:bsp-runtimes}
\end{figure*}
Figure~\ref{fig:bsp-runtimes} shows mean execution times (10 communications in
Figure~\ref{fig:bsp-runtimes-10} and 100 in Figure~\ref{fig:bsp-runtimes-100}).
Note that Julia's MPI wrappers actually incur some overhead due from the
polymorphism in the \verb$MPI.jl$ implementation. We suspected this might have
a non-trivial overhead, so we also provided our own custom versions of the MPI
wrappers with static type signatures (denoted with ``opt'' in the figures) and
repeated experiments with those variants. This minor optimization essentially
eliminates string comparisons and runtime type coercion. Communication in
Julia seems to be more expensive than in its lower-level counterpart. Our
optimization of bypassing dynamic dispatch by providing our own low-level
wrappers appears to have no noticeable effect on performance. Perhaps more
interestingly, we see that the performance difference scales with the number of
MPI ranks. Since we are actually increasing the {\em ranks per node} rather
than the absolute node count, the overheads in these experiments is likely
coming from effects {\em within} the node. This would be consistent with what
we saw in the spawn/fetch experiment.
\begin{figure*}
\centering
\subfigure[10 communication steps.]{
\includegraphics[width=0.35\textwidth]{figs/bsp-stacked-comms-10}
\label{fig:bsp-breakdown-10}
}
\subfigure[100 communication steps.]{
\includegraphics[width=0.35\textwidth]{figs/bsp-stacked-comms-100}
\label{fig:bsp-breakdown-100}
}
\caption{BSP experiment broken down per operation.}
\label{fig:bsp-breakdown}
\end{figure*}
To verify that communication is the issue, we took a more detailed look at the
benchmark. Figure~\ref{fig:bsp-breakdown} shows the same experiment, but with
the bars broken down per operation in the benchmark. In this figure the bars
show what percentage of execution each operation comprised.
In both cases, reads are relatively cheap. Interestingly, the C version
writes (again, to an array) are more expensive relatively. Indeed, however,
communications consume a more significant portion of the total execution time
in the Julia version. However, as the {\em number} of communications increases,
the difference between languages diminishes. Because both versions are ultimately
using the same MPI implementation, we can rule out the communications themselves.
While we are still investigating this effect, we suspect the culprit might be
how Julia manages the underlying workers. Note that the figures here exaggerate
the differences since the total execution time of this benchmark is in the tens
of milliseconds.
To further verify that the communications themselves were not the culprit, we
conducted a ping-pong experiment with MPI for both languages, using both the
normal Julia MPI wrappers and our custom implementation. The results are shown
in Figure~\ref{fig:pp}. We measure the ping-pong latency of a round-trip send
between two remote MPI ranks and report the mean latency (top) as we vary
message size (up to 1MB). We also report the harmonic mean of message
throughput (bottom). Outliers have been removed as in our spawn/fetch experiment.
In this case, we see that Julia actually {\em outperforms}
its lower-level counterpart, though we do see overlap in the error bars, and
differences here will be in the $\mu$s range. Again, our minor optimization for
Julia's MPI wrappers appears to have no distinguishable effect. While the
results are at odds with what we saw in our BSP experiment, we note that an
obvious difference in these scenarios is that in this case, we only have two
Julia worker processes, one per node. The improved communication performance in
this scenario further supports our hypothesis above about quirks in worker
management.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/pingpong}
\caption{ Latencies (above) and throughput (below) for an MPI ping-pong experiment
with both Julia and C implementations.}
\label{fig:pp}
\end{figure}
\subsection{HPCG}
\label{sec:hpcg}
The microbenchmarks give us a preliminary idea of how various native Julia
constructs and operatives perform, and what kind of distributed performance we can
expect to achieve. However, it is helpful to have a more realistic application to
perform an evaluation with. Our goal here was to choose an application that is
widely used and represents a broad range of HPC applications. To that end, we chose
a standard, state-of-the-art benchmarking application that is used to evaluate
top supercomputers today. We also wanted our effort to be useful for others to investigate
Julia's large-scale performance (perhaps on a much larger system).
The High-Performance Conjugate Gradients (HPCG)
benchmark~\cite{DONGARRA:HPCG:2013, Dongarra2015HPCG} fits that bill. It is
a tool for ranking computer systems based on a simple additive Schwarz,
symmetric Gauss-Seidel preconditioned conjugate-gradient solver. It constructs
a 3-dimensional, 27-point stencil matrix such that each point, represented by
coordinates $i$, $j$, and $k$ depends on the value of the surrounding 26
points.
The generated matrix ($A$) is sparse and has at least 27 non-zero values for
interior points~\cite{Dongarra2013HPCGSpec}. The boundaries have 7 to 18
non-zero points depending on the boundary position. The Matrix $A$ is symmetric
with positive eigenvalues (symmetric positive definite). It is stored in
Compressed Sparse Row (CSR) format. The global domain of $M_x$, $M_y$, and
$M_z$ in the $x$, $y$, and $z$ dimensions is distributed into a local domain of
$N_x$, $N_y$, and $N_z$. Each local sub-grid is assigned to a process (MPI
rank). The total number of processes is provided as an input parameter. These
processes are distributed into a 3D domain, where the total number of processes
$P = P_x \times P_y \times P_z$. The global domain $M_x$, $M_y$, $M_z$ can be
calculated as $(P_x \times N_x)(P_y \times N_y)(P_z \times N_z)$. A restriction
is also imposed on the domain for accuracy and efficiency purposes. Minimum
grid size in each dimension must be at least
16 and the grid should be a multiple of 8.
We ported HPCG version 3.1 \footnote{Available at
\url{https://hpcg-benchmark.org}} to Julia and ran it varying the local domain
size ($N_x = N_y = N_z = 16,32,64$) and varying the number of MPI ranks per
node (from 1 to 16). The results shown in Figure~\ref{fig:HPCG_EXP} are single
iteration runs. Each
subfigure has a different local domain size, so each is showing strong scaling
for a particular parameter. We compare it to the standard C++/MPI version of
the benchmark (without OpenMP support).
\begin{figure*}
\centering
\subfigure[$N_x = N_y = N_z = 16$]{
\includegraphics[width=.30\textwidth]{figs/hpcg-16}
\label{fig:HPCG_EXP-16}
}
\subfigure[$N_x = N_y = N_z = 32$]{
\includegraphics[width=.30\textwidth]{figs/hpcg-32}
\label{fig:HPCG_EXP-32}
}
\subfigure[$N_x = N_y = N_z = 64$]{
\includegraphics[width=.30\textwidth]{figs/hpcg-64}
\label{fig:HPCG_EXP-64}
}
\caption{Strong scaling of the standard C++ implementation of HPCG compared to our Julia port.}
\label{fig:HPCG_EXP}
\end{figure*}
The results here are consistent with the results from our BSP experiment. With
a small local domain size (see Figure~\ref{fig:HPCG_EXP-16}), computation
comprises a relatively small portion of the execution time, as each processor
will have little work to do before having to communicate. In this setting we
will see the Julia overheads that we have seen in previous experiments. The
remarkable thing to note here is that as the problem size per rank grows (see
Figure~\ref{fig:HPCG_EXP-64}), thus increasing the compute/communication ratio,
Julia actually starts to outperform the C++ version of the benchmark by as much
as 20\%. Julia's scaling also improves (as the C++ version worsens) as the
number of ranks per node is increased for the large local domain size.
Note that this benchmark communicates primarily with point-to-point
operations (\verb.MPI_Send. and \verb.MPI_Recv. in the halo exchange). Though
there are uses of collectives and barriers scattered throughout the code (e.g.
\verb.MPI_AllReduce.), they comprise very little of the total execution time.
The results from HPCG clearly suggest that Julia has significant potential,
even for HPC applications that heavily employ synchronous communication. We
note that our port of HPCG is currently a prototype, and thus we expect its
performance relative to its low-level counterpart will improve further as we
adopt more idiomatic usage of the language and as the language itself improves.
\section{Discussion}
\label{sec:discuss}
We now outline some of the issues we encountered with porting a large benchmark
from a low-level systems language to a Julia, a high-level managed language.
In order to have a fair comparison between C++ and Julia's version of HPCG it
was essential that we adopt it within the confines of Julia's abstractions. One
major difference between C++ and Julia is obviously the use of
managed memory, which made porting HPCG's pointer-heavy manipulation of
sparse matrices a challenge.
Structures in HPCG like \verb.matrixValues., \verb.mtxIndG., \verb.mtxIndL.,
which are pointer to pointer structures, are ported to Julia using 2D arrays.
\verb.matrixdDiagonal. is an array of pointers that refer to the addresses of
diagonal elements in \verb$matrixValues$. In Julia, this is replaced by an array
called \verb.curcols. which stores the column indices of \verb.matrixValues.
where the diagonal elements are stored. Diagonal elements can be fetched for
a given row index by accessing \verb.curcols. on the same index. Pointer
arithmetic is substituted by careful indexing as Julia is 1-indexed (\'{a} la
Fortran) and C++ is 0-indexed.
In the function \verb.FillRandomVector()., the C++ version uses \verb.rand().
from the C standard library to fill a vector with pseudo-random values. Because
Julia uses a different PRNG routine, and because we had to compare outputs
between versions for correctness, we modified this routine to fill vectors with
known, fixed values.
In the procedure \verb.ComputeOptimalShapeXYZ. in the C++ version, a complex
\verb.for. loop uses \verb.MixedBaseCounter. structures and its corresponding
functions to initialize, check bounds, and increment the
\verb.MixedBaseCounter. structure. These are ported using Julia's looping
constructs, which are slightly different. A similar approach is used in
\verb$GenerateProblem_ref$.
Primitive data types in Julia passed as parameters to a function and mutated in
the function body will not retain their altered values on return (as is
standard in high-level languages). However, HPCG uses a common C/C++ idiom of
passing in pointers to primitive types to have the values mutated by the
callee. The Julia versions of such functions must have slightly different
structure to account for this. When possible, we followed Julia's documented
best
practices\footnote{\url{https://docs.julialang.org/en/v1/manual/performance-tips/index.html}}
for performance programming. This included making code type-stable, eliminating
use of any global variables, eliding bounds checks for known-length arrays
(\verb.@inbounds.), and carefully tuning core HPCG math kernels for performance
bottlenecks, e.g. by instrumenting the garbage collector and using Julia's
native profiler. We still, however, believe there is room to improve the
performance.
While not reported in the figures, we did see a notable difference in the
variance of Julia's runs on the larger code. This is not entirely surprising,
as jitter is a known pain point for applying managed languages in HPC settings
(this is, of course, true for the underlying system as
well~\cite{HOEFLER:2010:NOISE, FERREIRA:2008:OSNOISE, MORARI:2011:OSNOISE}).
In Julia's case, two primary sources of non-determinism include the garbage
collector and the JIT compiler. While others have looked at these effects for
other languages~\cite{AKERET:2015:HOPE, PRABHU:2019:MOYA}, we believe there is
opportunity to optimize these features to avoid jitter for Julia (and other
managed, technical computing languages) in synchronous HPC settings.
\section{Related work}
\label{sec:related}
There has long been interest in bringing the convenience of high-level
languages to bear for HPC. High-performance Fortran~\cite{MERLIN:1995:HPF},
Chapel~\cite{CHAMBERLAIN:2007:CHAPEL}, X10~\cite{CHARLES:2005:X10}, and
Erlang~\cite{JOHANSSON:2000:HPE}, and Regent~\cite{SLAUGHTER:2015:REGENT} are
notable examples. While Python was not designed with HPC computing in mind,
there are efforts in that community to adapt it for high-performance
parallelism~\cite{Ross2016MPI4Py,Gomez2015HPCPython,
Hand2017Nbodykit,Paraskevakos2018Molecular}.
Most recent work on Julia focuses on adapting it for practical problems in
technical computing, several of which we described in Section~\ref{sec:julia}.
From the HPC perspective, the most striking example is
Celeste~\cite{Regier2015CelesteVI}, the first large-scale HPC application
written entirely in Julia. It was reported to achieve petascale performance on
a NERSC supercomputer. However, Celeste is not structured to be
communication-intensive, unlike many HPC applications, so its success, while
remarkable, does not yet demonstrate Julia's broad applicability for
large-scale, communication-intensive HPC applications.
Others have evaluated Julia in a multi-core
setting~\cite{Knopp2014JuliaMultiThreading, bean2013ManyCoreJulia}. Hunold and Steiner
study Julia's MPI bindings~\cite{HUNOLD:2020:JULIA}, though in a more focused setting. Our results with the more
substantial HPCG benchmark, however, do line up with their findings.
There also has been extensive work done in programming GPUs with Julia.
In addition to the the foundational GPUArrays package JuliaGPU\footnote{Available at
\url{https://juliagpu.org/}} is a package in Julia
which encompasses support for both NVIDIA and AMD GPUs. The NVIDIA CUDA
boasts mature and full-featured packages for both low-level kernel programming as well as working with high-level operations on arrays. All versions of Julia are supported, and the functionality is actively used by a variety of applications and libraries.
\section{Conclusions and future work}
\label{sec:conc}
In this paper, we evaluated Julia using a range of microbenchmarks and
macrobenchmarks both for shared-memory parallelism within the node and for
distributed parallelism in a cluster computing setting. We showed that while
Julia has slight overheads for communication between processes, these overheads
are often outweighed by gains over its lower-level counterparts (C/C++). We
demonstrated using a custom synthetic BSP benchmark and the first ever Julia
port of a large-scale, HPC benchmark application (HPCG) that Julia can
outperform C++/MPI even in a multi-node setting with synchronous communication.
We believe these findings open up several lines of investigation into Julia as
a first-class HPC language. First, we plan to extend our experiments to add
comparisons to LLVM-compiled C/C++ code. We also plan to further investigate
the source of communication anomalies between Julia's worker processes.
Finally, we hope to explore optimizations for sources of non-determinism in the
Julia runtime and to scale our experiments to hundreds or even thousands of
nodes.
\section*{Acknowledgements}
This work is made possible with support from the United States National Science
Foundation (NSF) via awards CNS-1718252 and CNS-1730689.
\balance
\bibliographystyle{IEEEtran}
|
\section*{Background \& Summary}
Deep learning has been successfully applied to medical image analysis, including abnormality detection in \glspl*{CXR}~\cite{chexnet,radreview}. However, even though, in the medical image context, very large published \gls*{CXR} datasets are available~\cite{padchest,chexpert,chestxray8,mimiccxr}, with sizes in the hundreds of thousands of images, they are much smaller than some natural images datasets~\cite{imagenet,tencentdataset}. These large \gls*{CXR} datasets were labeled through the mining of clinical reports. However, given the relatively small size of the datasets, additional labels may contribute to the training of data-demanding deep convolutional neural networks. We present a dataset named REFLACX containing eye-tracking data collected from radiologists while they dictate reports. This dataset was built as a proof-of-concept for a data collection method that expands the labels of a medical dataset, providing additional supervision during deep learning training.
Li et al. have shown that bounding boxes localizing abnormalities on \glspl*{CXR} can be used to supervise a \gls*{CNN} to improve accuracy and localization scores~\cite{cvpr2018}. They used the manually annotated bounding boxes provided for 880 images of the 112,200 \glspl*{CXR} from the ChestX-ray14 dataset~\cite{chestxray8}. Even though other datasets provide similar labels~\cite{pneumoniadataset,vindrcxr}, such localization labels are rare, and when present, are usually provided in limited quantities, showing the difficulty of scaling up the manual labeling. To try to solve this issue, and in accordance with the prioritization of the development of automated image labeling and annotation methods of the NIH's roadmap for AI in medical imaging~\cite{roadmap}, we collected eye-tracking data from radiologists for implicit localization of anomalies. We believe that the proposed collection method has the potential to be scaled up and used as a nonintrusive annotation approach deployed in a radiology reading room.
Other works have used eye-tracking data to support models in training or inference. Templier et al. and Stember et al. used eye tracking for interactive segmentation of biological/medical images, intending to achieve faster labeling~\cite{segmic,segtumor}. Khosravan et al. proposed a method of integrating an eye tracker into the reading room to combine a radiologist's reading with \gls*{CAD} systems~\cite{falsenegativect}. Gecer et al. used the navigation behavior of pathologists during readings to improve the detection of cancer in histopathology images~\cite{pathologyzoom}. Stember et al. showed that the gaze of a radiologist when dictating standardized sentences for presence/absence of brain tumors in MRIs could be used to localize lesions~\cite{mridictation}. In parallel to our project, Karargyris et al. built a dataset of eye-tracking data and respective dictation of \gls*{CXR} reports, containing 1,083 readings by only one radiologist~\cite{ibm}. Saab et al. built a dataset of eye-tracking data for the task of identifying pneumothoraces for 1,170 \glspl*{CXR}~\cite{miccai2021}. Still, their dataset does not contain dictations.
To collect the proposed dataset, as described in Figure~\ref{fig:overview}, five board-certified subspecialty-trained thoracic radiologists, who closely worked on aspects of the study design, used a custom-built user interface mimicking some of the functionalities from clinical practice. They dictated reports for images sampled from the MIMIC-CXR dataset~\cite{mimiccxr} while eye-tracking data were collected, including gaze location and pupil area. The audio of the dictation was automatically transcribed and manually corrected. The transcriptions included word timestamps for synchronization with the rest of the data. During dictation, we also recorded the zooming, panning, and windowing state of the \gls*{CXR} at all times. After dictating each case, radiologists provided a set of manual labels: the abnormalities they found, selected from a list, and the location of those abnormalities. These manual labels were collected to validate the data collection method and any automatically extracted labels generated from the dataset. They are not easily scalable and would not be collected in more extensive implementations of the collection method. Radiologists also drew a bounding box around the heart and lungs for normalization of chest position.
The collection of data was separated into three phases. The two first preliminary phases were used to adjust minor data collection details and to estimate inter-rater scores for the labels of the dataset, with radiologists reading a shared set of images. For the third and main phase, the sets of \glspl*{CXR} for each radiologist were independent. Considering all phases, REFLACX contains 3,052 cases, for a total of 2,616 unique \glspl*{CXR}. Of the 3,052 cases, 3,032 contain eye-tracking data.
\section*{Methods}
Data collection was divided into three phases. Phases 1 and 2 were preliminary phases where five radiologists read the same set of 59 and 50 \glspl*{CXR}, respectively. The set of possible image-level labels was chosen after a discussion among radiologists. After estimating the inter-rater reliability for phase 1, as shown in Table \ref{tab:fleiss}, a meeting was organized where radiologists discussed the labeling differences for the five cases that had the most negative impact on the reliability scores. The set of labels for phase 2 was slightly modified to clarify labeling and reduce its complexity. An electronic document, included as Supplementary File 1, was distributed to all radiologists with labeling instructions. A glossary by Hansell et al. provided some of the labeling definitions for this document~\cite{definitionopacification}. Phase 3 had the same five radiologists reading independent sets of around 500 \glspl*{CXR} each and used the same set of anomaly labels as phase 2. This phase constitutes the main content of the dataset and has a slightly higher quality of eye-tracking data. The set of labels used for each phase is listed in Table \ref{tab:stats}. Phase 1 went from November 11, 2020 to January 4, 2021, phase 2 from March 1, 2021 to March 11, 2021, and phase 3 from March 24, 2021 to June 7, 2021. Data collection sessions took on average 2.21 hours, with a maximum of 3.92 hours and a minimum of 0.2 hours.
This study complied with all relevant ethical regulations. Eye-tracking data collection was exempted from approval by the University of Utah \gls*{IRB}, and no informed consent was necessary. The use of the MIMIC-CXR \glspl*{CXR} was exempted from approval for this study since the images came from a publicly available and de-identified dataset. The MIMIC-CXR dataset was originally approved by the \gls*{IRB} of BIDMC, and patient consent was waived.
\subsection*{Image data}
\Glspl*{CXR} were sampled from the MIMIC-CXR dataset~\cite{mimiccxr,mimiccxrdicom,physionet}, a publicly available deidentified dataset. Before a random sampling of the images shown to radiologists, the dataset was filtered for including only \glspl*{CXR} that contained ``ViewPosition'' metadata with value ``PA'' or ``AP'', were the only frontal \glspl*{CXR} in their study, and were present in the label table from the MIMIC-CXR-JPG dataset~\cite{mimiccxrjpg,mimiccxrarxiv,physionet}. Images were sampled to include 20\% of images from the test set of the MIMIC-CXR dataset and 80\% from the other splits, and uniformly at random from each of these two splits. After the sampling for phase 3, images with anonymizing black boxes that intersected the lungs were manually excluded before presentation to radiologists. After the sessions, we manually excluded outlier images that, according to the radiologists, had major parts of the lung missing from the field of view, were digitally horizontally flipped, or had a rotation of 90\degree.
\subsection*{Data-collection sessions}
Data collection for each \gls*{CXR} involved two main parts: a dictation and transcription of a free-form radiology report while collecting eye-tracking data, and the selection of labels and ellipses to use as evaluation ground-truth for anomalies present in a \gls*{CXR} and their location. The data collection interface was developed in MATLAB R2019a/Psychtoolbox 3.0.17~\cite{ptb1,ptb2,ptb3}. The code for the interface is available at \href{https://github.com/ricbl/eyetracking}{https://github.com/ricbl/eyetracking}~\cite{mycodezenodo}. The interface is shown in Figure \ref{fig:screens} and as a video in Supplementary File 2, where the moving semitransparent red ellipse, with an axis length of 1\degree\ of visual angle, represents stabilized gaze, i.e., fixations, and the moving blue ellipse represents the instantaneous gaze location sample. The cursor was not drawn in the video and the audio is a digitally generated voice representing the timestamped transcription. We did not include the original dictation in the video for anonymization purposes. This interface allowed for:
\begin{itemize}
\itemsep 0em
\item Dictation of reports. A \gls*{CXR} was displayed to radiologists, and they dictated a report using a handheld microphone while eye-tracking data were collected. Radiologists did not have previous access to the \glspl*{CXR} or their reports. Eye-tracking data were collected as soon as the \gls*{CXR} was shown to radiologists, including data containing a reasoning period when they first saw the image and a dictation period. Radiologists were instructed to dictate reports as they would dictate in clinical practice. We also asked them to dictate free-form reports, so radiologists that used templates in clinical practice had to change their style to free-form.
\item Editing of mistakes of the automated transcription. Radiologists were instructed to not add or change report content, and only correct transcription mistakes.
\item Selection of image-level labels and ellipses localizing each label, including assessing the certainty of the presence of anomaly for each ellipse. Following Panicek et al.~\cite{certainty}, the allowed certainties were: Unlikely (<10\%), Less Likely (~25\%), Possibly (~50\%), Suspicious for/Probably (~75\%), Consistent with (>90\%). Radiologists were instructed to add image-level labels even in case they forgot to mention them in the original report.
\item Image windowing, zooming, and panning. These features were available while dictating and labeling, mimicking features available to radiologists in their clinical practice.
\item Drawing of chest bounding boxes that encompass lungs and heart, for normalization of the variations in \glspl*{CXR}.
\item Calibration of pupil size and access to the eye-tracker calibration screens.
\end{itemize}
In more detail, at the beginning of the session, the calibration interaction adopted the following order:
\begin{enumerate}
\item Eye-tracking calibration screen, where 13 circles, distributed throughout the screen, are displayed in random order for the radiologists to fixate.
\item Calibration of pupil size.
\end{enumerate}
For the rest of the session, for each reading of \gls*{CXR}, the interaction of radiologists followed:
\begin{enumerate}
\item Dictation of report. Radiologists were not allowed to return to this screen for changes in dictation.
\item Image-level label selection.
\item Text input for ``Other'' label, in case it was selected as an image-level label.
\item Localization of image-level labels by the drawing of ellipses and choice of certainty for each ellipse.
\item Drawing of a chest bounding box including lung and heart.
\item Text editing to correct transcription errors.
\item Screen allowing proceeding to the subsequent \gls*{CXR}, displaying warnings for low-quality eye-tracking data, and optional return to calibration screens.
\end{enumerate}
\subsubsection*{Equipment}
To collect eye-tracking data, we used an Eyelink 1000 Plus system, which allows for high spatial (less than 1\degree of visual angle) and temporal (1,000 Hz) resolution. It also allows the radiologists to move their heads within a small area while maintaining this high degree of spatial and temporal acuity. Given our interest in ensuring that we had high-quality eye-tracking data, the experimenter calibrated and validated each radiologist multiple times throughout their viewing sessions. This process can be time-consuming, particularly if the clinician wears bifocals, which often lead to poor calibration. One alternative to this setup would have been to use a mobile eye-tracking system. These systems typically have the eye-tracking apparatus embedded within glasses worn by the subject. A good calibration can be achieved much more easily and possibly without an experimenter guiding this process. However, at present, the resultant data are much more difficult and time-consuming to analyze because most mobile eye-tracking systems do not co-register eye movements with precise screen coordinates. In practice, this typically means that the experimenters must hand-code the co-registration from a video output. This process is time-consuming and can suffer from bias, requiring multiple coders to examine the data to ensure reliable coding.
The Eyelink 1000 Plus was equipped with a 25 mm lens and managed by an Eyelink 1000 Host PC recording gaze at 1,000 Hz. The eye tracker was configured in the remote setup, for which the radiologists put a sticker on their forehead. Radiologists had the freedom to move their heads as long as the sticker and tracked eye stayed within the camera's field of view, and the sticker stayed between 55 cm and 65 cm from the camera. The camera was positioned below a 27 inches BenQ PD2700U, 3,840$\times$2,160 pixels, 60 Hz screen, connected to a Display PC running Ubuntu 18.04. For phase 3, the camera was positioned 11 cm in front of the screen. Eye distance to the bottom of the screen was around 71 cm, while it was around 65 cm to the top of the screen.
\subsubsection*{Calibration}
Eye-tracker calibrations are necessary for finding the correspondence between pupil and cornea positions in the image captured by the eye tracker's camera and locations on the screen where the radiologists are looking. During the calibration, radiologists had to look at 13 targets in several locations on the screen for the eye-tracker to register the pupil and cornea positions for each location and interpolate for other intermediate locations on the screen. We performed an eye-tracker calibration at the beginning of each session, every 25 cases, every time radiologists took a break, or when noticeable quality problems with the data could be solved with a new calibration. These quality problems mainly involved moments when the cornea or the pupil was not recognized for specific eye positions or when the eye-tracker gave incorrect locations for the pupil or the cornea, e.g., glasses recognized as the cornea. The experimenter identified them with access to real-time data-collection information on the Eyelink 1000 Host PC. During calibration, radiologists were positioned such that the forehead sticker was between 59 cm and 61 cm away from the camera. For phases 1 and 2, calibrations were not performed at a regular 25-case interval. For phase 3, the calibration was considered successful if the average error was less than 0.6\degree, and the maximum error was less than 1.5\degree. The left eye was tracked by default, and the right eye was only tracked when the left eye calibration was repeatedly faulty. At each session beginning, radiologists were asked to look at the center of the screen for 15 s for measuring a constant used to normalize the pupil size.
\subsection*{Report transcription}
We collected the dictation audio at 48,000 Hz using a handheld PowerMic II microphone. The audio was transcribed using the IBM Watson Speech to Text service, which provides timestamps for each transcribed word, with the ``en-US\_BroadbandModel'' model. Before phase 1, the service's custom language model was trained with sentences from the ``Findings'' and ``Impressions'' sections of the MIMIC-CXR reports, which were filtered to remove sentences that referenced other studies of a patient through the search of keywords. For phase 2 and phase 3, in addition to language training with the filtered MIMIC-CXR reports, models had language and acoustic training with the collected audio and corrected transcriptions from phase 1.
Audio files had silence trimmed from the start and end of the file to speed up transcriptions. Silence was detected using Otsu's thresholding over the average audio level (dBFS) of 500 ms chunks. Word timestamps were adjusted for the trimming of the beginning of the audio. After the report transcription was received from the cloud service, radiologists could make minor changes to it. Several of the common mistakes of the cloud service were programmatically corrected before providing transcriptions to the radiologists.
\subsection*{Postprocessing}
The eye-tracking gaze samples were parsed for fixations, i.e., locations where gaze was spatially stabilized for a certain period; blinks, i.e., moments when the eye tracker did not capture pupil or cornea; and saccades, i.e., fast eye movements between fixations. Parsing was done in real time by the EyeLink 1000 Host PC, using a saccade velocity threshold of $35^{\circ}/s$, a saccade motion threshold of $0.2^{\circ}$, and a saccade acceleration threshold of $9,500^{\circ}/s^2$. Fixation locations were converted from screen space to image space by recording, at the start of the fixation, what image part was shown at what screen section. Fixations were synchronized with the transcriptions and other recorded data by synchronization messages sent by the Display PC to the EyeLink 1000 Host PC.
Pupil area data, captured by the eye tracker, was normalized by the average value of pupil area from the calibration screen from the beginning of the session. Radiologists were asked to look directly at the center of the screen, marked by a cross. The average value was weighted by fixation durations and calculated only for fixations at most 2\degree\ from the screen center.
After all data-collection sessions, transcriptions were checked and corrected by another person, who looked for additions of content during the radiologist correction screen, which were not allowed, and for out-of-context words and other clear mistakes, which were corrected by relistening to the recorded audios of those cases. During this process, spellings of a few words were standardized. The labels listed as ``Other'' were also standardized. Since the transcriptions were corrected for mistakes after receiving the output from the cloud service, timestamps had to be adapted to the new set of words. We used the counting of syllables to perform a linear interpolation between the times of both texts. Interpolations were calculated for each difference found between strings, as given by the difflib Python library. When there was an addition of words with no removal, the new words used the time range defined by the end time of the previous word and the start time of the next one, making it possible for words to have the same timestamp for start and end.
\subsubsection*{Data quality}
Readings were evaluated for the quality of eye-tracking data by measuring the times that data were classified as a blink during collection. Eye-tracking data were discarded in cases with a blink longer than 3 s or when blinks corresponded to more than 15\% of the data. Warnings were shown between cases for blinks longer than 1.5 s or when blinks corresponded to more than 10\% of the data. Cases that had eye-tracking data discarded were not included in the dataset for phase 3 but were included for phases 1 and 2 to make possible the evaluation of inter-rater scores for other labels. The threshold values were chosen by qualitative observation of blink data histograms from phases 1 and 2 before collecting phase 3 data. Eye-tracking data were also discarded when the radiologist unintentionally clicked the ``Next Screen'' button before completing the dictation, when the eye-tracking data were not correctly saved because of various software problems, and when the eye tracker identified the glasses of the radiologist as their eyes. We discarded eye-tracking data for 7 cases for incomplete dictation, 6 for software problems when saving the data, 2 for large parts of the lungs missing, 1 for a horizontally flipped image, 1 for extreme rotation of the MIMIC-CXR image, 41 for low data quality, and 2 for glasses being confused for eyes. The total of discarded eye-tracking data was 10 cases for phase 1, 10 cases for phase 2, and 40 cases for phase 3. Additionally, one \gls*{CXR} in phase 1 had large parts of the lung missing and is not included in the 59 images present in the dataset. The 2,616 unique \glspl*{CXR} for which data are provided in the REFLACX dataset were included after manual image quality exclusions and data quality exclusions.
\section*{Data Records}
For each reading of a MIMIC-CXR image, the labels of this dataset consist of eye-tracking data, formatted as fixations and as gaze position samples, a timestamped report transcription, ellipses localizing anomalies associated with a certainty, and a chest bounding box. For each case/reading, there is a subfolder containing these labels, separated into individual tables. Subfolders from all three phases are grouped in the same folder (main\_data/ and gaze\_data/), and the phase to which each subfolder belongs is listed in metadata tables, one for each data collection phase. The statistics of the resulting dataset are presented in Table \ref{tab:stats}. Genotypical sex information was extracted from the MIMIC-IV dataset~\cite{mimiciv,physionet}, and was missing for around 0.35\% of the 2,199 unique subjects. The dataset is available on Physionet, at \url{https://doi.org/10.13026/e0dj-8498}~\cite{mydataset,physionet}.
\subsection*{Description of tables and their columns}
\begin{itemize}
\item \textbf{main\_data.zip/main\_data/metadata\_phase\_<phase>.csv}: list of all the subfolders/cases corresponding to a specific phase and their metadata.
\begin{itemize}
\item \textbf{id}: subfolder name and unique identifier for a reading of a specific \gls*{CXR} by a specific radiologist.
\item \textbf{split}: the split given by the MIMIC-CXR dataset for that specific image. The possible values are ``train'', ``validate'', ``test''. Images were sampled so that 20\% of the images were from the test set of the MIMIC-CXR dataset.
\item \textbf{eye\_tracking\_data\_discarded}: for phases 1 and 2, even when the eye-tracking data were discarded for low quality, the anomaly labels, localizing ellipses, and chest bounding box were collected and included in the dataset. This column is ``True'' when the eye-tracking data has been discarded, and ``False'' otherwise. Transcriptions are also not included for these cases. Such cases should not be used for analysis if eye-tracking data or transcription are required. For phase 3, no case with discarded eye-tracking data is included.
\item \textbf{image}: path to the DICOM file from the MIMIC-CXR dataset used in this reading, with the same folder structure as provided by that dataset.
\item \textbf{dicom\_id}: unique identifier of the image that can be used to join tables with the metadata from MIMIC-CXR.
\item \textbf{subject\_id}: unique identifier for the patient of that study.
\item \textbf{image\_size\_x}: horizontal size of the \gls*{CXR} in pixels
\item \textbf{image\_size\_y}: vertical size of the \gls*{CXR} in pixels
\item \textbf{Other columns}: the rest of the columns represent the possible presence of anomaly evidence in the image, as selected by the radiologist. Most of these columns contain values between 0 and 5, representing a certainty of the presence of such anomaly, according to the scale:
\begin{itemize}
\item 0: not selected by radiologist,
\item 1: Unlikely,
\item 2: Less Likely,
\item 3: Possibly,
\item 4: Suspicious for/Probably,
\item 5: Consistent with.
\end{itemize}
Certainties were associated with each localizing ellipse in the image, so each label's maximum certainty is reported. Radiologists were asked not to draw ellipses for the anomaly labels ``Support devices,'' ``Quality issue'' and ``Other,'' so there is no certainty associated with these labels.
\item \textbf{Support devices} and \textbf{Quality issue}: the presence of these labels is reported, using ``True'' or ``False.''
\item \textbf{Other}: A list of the other anomalies reported by radiologists not included in the rest of the labels, separated by ``|.'' If empty, no other anomaly was reported.
\end{itemize}
\item \textbf{main\_data.zip/main\_data/<id>/fixations.csv}:
eye-tracking data summarized as fixations and collected during the dictation of the report.
\begin{itemize}
\item \textbf{timestamp\_start\_fixation / timestamp\_end\_fixation}:
the time in seconds when the fixation started/ended, counting from the start of the case.
\item \textbf{average\_x\_position / average\_y\_position}:
average position for the fixation, given in pixels and in the image coordinate space, where $(0,0)$ is the top left corner.
\item \textbf{pupil\_area\_normalized}:
pupil area, normalized by the calibration performed at the beginning of each session.
\item \textbf{window\_level / window\_width}:
average values of the windowing used for the image during a fixation.
\item \textbf{angular\_resolution\_x\_pixels\_per\_degree / angular\_resolution\_y\_pixels\_per\_degree}:
number representing how many image pixels fit in 1\degree\ of visual angle for each axis (x or y). It is dependent on the position of the fixation and the zoom applied to the image.
\item \textbf{xmin\_shown\_from\_image / ymin\_shown\_from\_image / xmax\_shown\_from\_image /\\ymax\_shown\_from\_image}:
bounding box given in image-space representing what part of the image was shown to the radiologist at the start of the fixation. The reading/case always started with the whole image being shown, but zooming and panning were allowed.
\item \textbf{xmin\_in\_screen\_coordinates / ymin\_in\_screen\_coordinates / xmax\_in\_screen\_coordinates /\\ymax\_in\_screen\_coordinates}:
bounding box given in screen space representing where the part of the image was shown.
\end{itemize}
\item \textbf{gaze\_data.zip/gaze\_data/<id>/gaze.csv}: complete eye-tracking data during the dictation of the report, collected at 1,000 Hz. Even though this data are not necessary for accomplishing the main research goals of this dataset, these data are included for any other analyses that need the gaze location in more detail than provided by the fixations.csv table. Compared to the fixations.csv table, the \textbf{timestamp\_start\_fixation} and \textbf{timestamp\_end\_fixation} columns were replaced by the \textbf{timestamp\_sample} column. The remaining columns are the same in both tables, but they represent values when the eye tracker’s camera captured the gaze sample in gaze.csv.
\begin{itemize}
\item \textbf{timestamp\_sample}: timestamps do not start at 0 because audio recording started before gaze recording.
\item \textbf{x\_position / y\_position / pupil\_area\_normalized / angular\_resolution\_x\_pixels\_per\_degree /\\angular\_resolution\_y\_pixels\_per\_degree}: these columns are empty for timestamps when the eye tracker could not find the radiologist's pupil or cornea, making it impossible to calculate gaze at that moment. These rows are usually associated with moments when radiologists blinked.
\end{itemize}
\item \textbf{main\_data.zip/main\_data/<id>/timestamps\_transcription.csv}: timestamped corrected transcriptions of the reports. Radiologists were allowed to delete parts of the report and to modify transcription errors but not to add content.
\begin{itemize}
\item \textbf{word}: the word that was spoken. Periods (.), commas (,) and slashes (/) occupy one row.
\item \textbf{timestamp\_start\_word/timestamp\_end\_word}: the time in seconds when the dictation of each word started\slash ended, counting from the start of the case.
\end{itemize}
\item \textbf{main\_data.zip/main\_data/<id>/transcription.txt}: the transcription in text form, without timestamps.
\item \textbf{main\_data.zip/main\_data/<id>/anomaly\_location\_ellipses.csv}: bounding ellipses drawn by radiologists for each label present in the image. Each label may appear in more than one ellipse, and each ellipse may contain more than one label. Radiologists were instructed to select more than one label for an ellipse when a single image finding may be evidence of one or another label. Each row of the table represents one ellipse.
\begin{itemize}
\item \textbf{xmin / ymin / xmax / ymax}: coordinates representing the extreme points, in image pixels, of the full horizontal and vertical axes of the ellipse. Coordinate $(0,0)$ represents the top left corner of the image.
\item \textbf{certainty}: value from 0 to 5 representing a certainty of the finding presence, according to the same scale as in the metadata table.
\item \textbf{Other columns}: the rest of the columns have a Boolean value representing the presence of evidence for each anomaly label in the ellipse. Since radiologists were asked not to draw ellipses for ``Support devices'', ``Quality issue'', and ``Other,'' most of the rows have the value ``False'' for these labels. For all other labels present in the image, at least one ellipse is drawn.
\end{itemize}
\item \textbf{main\_data.zip/main\_data/<id>/chest\_bounding\_box.csv}: single-row table containing information for the bounding box drawn around the lungs and the heart.
\begin{itemize}
\item \textbf{xmin / ymin / xmax / ymax}: coordinates representing the extreme points, in image pixels, of the bounding box. Coordinate $(0,0)$ represents the top left corner of the image.
\end{itemize}
\end{itemize}
\section*{Technical Validation}
For all reported technical validation values, we also report standard error and sample size when applicable. Standard errors were only reported for $n>10$. For measurements that had more than one score for the same \gls*{CXR}, e.g., \gls*{IoU} calculated for all pairs of radiologists, we averaged the scores for each \gls*{CXR} before calculating the final average. The sample size given is the number of independent \glspl*{CXR} involved in the calculation.
\subsection*{Eye-tracking data}
Considering the errors provided by each of the calibrations used in data collection, we calculated the average and maximum calibration errors for each phase. Phase 1 had an average calibration error of 0.43$\pm$0.02\degree\ (n=25) and a maximum error of 2.79\degree, whereas, for phase 2, it was 0.43$\pm$0.03\degree\ (n=13) and 1.09\degree, respectively. Phase 3 had an average error of 0.44$\pm$0.01\degree\ (n=128)\ and a maximum error of 1.5\degree.
To validate the presence of abnormality location information in the eye-tracking data, we calculated the presence of fixations inside the abnormality ellipses, as exemplified in Figure \ref{fig:heatmaps}. For each reading that had at least one drawn ellipse, we calculated the \gls*{NCC} between a fixation heatmap and a mask generated from the union of ellipses, using
\begin{equation}
\text{NCC}\left( x_f, x_e\right)=\frac{1}{P-1}\sum_{p }{\frac{\left({x_f}\left(p\right)-\mu_{x_f}\right)}{\sigma_{x_f}}\times\frac{\left(x_e\left(p\right)-\mu_{x_e}\right)}{ \sigma_{x_e}}},
\end{equation}
\noindent where $\text{NCC}\left( x_f, x_e\right)$ is the \gls*{NCC} between the fixation heatmap $x_f$ and the ellipse heatmap $x_e$, $\sigma_{x}$ is the standard deviation of a heatmap $x$, $\mu_x$ is the average value of a heatmap $x$, $x(p)$ is the value of a heatmap $x$ at pixel $p$, and $P$ is the number of pixels in the heatmaps. The fixation heatmaps were generated by drawing Gaussians centered on every fixation, with the standard deviation equal to 1\degree\ in each axis and with intensity proportional to the fixation duration.
To check if the fixations heatmap of a specific \gls*{CXR} has more localization information than the heatmaps from unrelated \glspl*{CXR}, we compared the \gls*{NCC} for two types of fixations heatmap: heatmaps generated from the eye-tracking fixations of each \gls*{CXR} reading, as shown in Figure \ref{fig:heatmaps}c, and a baseline heatmap representing the average gaze over all \glspl*{CXR}, as shown in Figure \ref{fig:heatmaps}d.
To calculate the baseline heatmap shown in Figure \ref{fig:heatmaps}d, we normalized all heatmaps to the same location using the labeled chest bounding boxes to compensate for the variation in the location of the lungs for each \gls*{CXR}. We calculated the average chest bounding box, transformed all heatmaps to this same space, and averaged the heatmaps. We finally transformed the average heatmap back to the space of each \gls*{CXR} to calculate the \gls*{NCC} against the labeled ellipses.
For the fixation heatmaps specific to each \gls*{CXR}, the average \gls*{NCC} achieved over all applicable \glspl*{CXR} was 0.380$\pm$0.014 (n=96), against a baseline score of 0.326$\pm$0.013 (n=96). This result shows that there is more abnormality location information in the fixations for each \gls*{CXR} than on a heatmap built from the usual areas looked at by a radiologist.
To analyze if the localization information given by the eye-tracking data correlates with the time that the presence of an abnormality was mentioned during the dictation, we produced the graph shown in Figure \ref{fig:delay}. To develop this analysis, we annotated the location where labels were mentioned in the report for 200 non-test \glspl*{CXR} from Phase 3. For annotating, we used a mix of a modified version of the chexpert labeler~\cite{chexpert} and hand-labeled corrections. \Glspl*{CXR} that had image-level labels not mentioned in the dictation were not included in the 200 randomly selected \glspl*{CXR}.
We separated the time before the end of mentions of a label into bins of same size. For each bin, we calculated the percentage of fixations localized inside an ellipse for the same abnormality and \gls*{CXR}.
The percentage was calculated considering the duration of the fixation inside the bin. Besides analyzing the delay in time units, we also used sentences units. To calculate the sentence units of each timestamp, we separated the dictation into mid-sentence moments and in-between-sentences pause moments. The timestamps of each transition between these were represented by integer numbers. Timestamps in the middle of a sentence or pause had their representation calculated through linear interpolation of the start and end of their sentence or pause. For example, a 12 s timestamp within the second sentence, dictated from 10 s to 15 s, would be associated with 3.4 sentence units. The sentence units shown in Figure \ref{fig:delay} represent the difference between the sentence units of the mentions and the fixations before the mentions.
We divided the full range of data into 75 bins, of width 1.03 s and 0.21 sentences, and only kept bins up to before the first bin with less than or equal to ten fixations inside ellipses. We calculated 95 \% confidence intervals through bootstrapping, randomly sampling with replacement 200 \glspl*{CXR} from the 200 annotated \glspl*{CXR}. We performed this sampling 800 times and display in Figure \ref{fig:delay} the 2.5\% and 97.5\% percentiles for each bin.
As shown in Figure \ref{fig:delay}, there are peaks of correlation between the location of ellipses and the radiologists' fixations at around 2.5 s before the mention of the respective abnormality. The correlation peak was also calculated to be around 0.6 to 1.25 sentence units before the mention. This correlation shows that the transcription timestamps could be used to get label-specific localization information. With the shown delay between label fixation and mention, our data might need alignment algorithms for the correct association between fixations and label. We leave the exploration of the application of such algorithms to future works.
\subsection*{Validation labels}
\subsubsection*{Image-level labels}
The inter-rater reliability was measured through Fleiss' Kappa~\cite{fleiss} for the image-level labels, calculated using the statsmodels library in Python~\cite{statsmodels}. Image-level labels were considered positive when the maximum certainty, among all ellipses of a given label, was ``Possibly'' or higher.
The inter-rater reliability scores for phases 1 and 2 are shown in Table~\ref{tab:fleiss}. The achieved inter-rater reliability scores are relatively low but in line with scores obtained in other similar studies for readings of \gls*{CXR}~\cite{lowrater1,lowrater2}. Some of the low values might be caused by the low prevalence of some of the labels~\cite{prevalence1,prevalence2}.
\subsubsection*{Localization ellipses}
For each \gls*{CXR}, for each label with more than one radiologist selecting certainty ``Possibly'' or higher, we calculated the average paired \gls*{IoU} between all pairs of respective radiologists. We then calculated the average \gls*{IoU} over all \glspl*{CXR} of each phase, with results presented in Table~\ref{tab:fleiss}. Between phases 1 and 2, we discussed a few examples of \gls*{CXR} that had low \gls*{IoU}.
\subsubsection*{Chest bounding boxes}
Similar to the localization ellipses, the quality of the bounding boxes containing the heart and the lungs was measured through \gls*{IoU}. \gls*{IoU} was calculated between all the pairs of radiologists for every \gls*{CXR} of the preliminary phases. For phase 1, the average \gls*{IoU} was 0.917$\pm$0.004 (n=59), and for phase 2, it was 0.920$\pm$0.004 (n=50). We organized no discussion for improvement of this label between phases 1 and 2.
\section*{Usage Notes}
To have access to the \glspl*{CXR} that the radiologists read, access to the MIMIC-CXR dataset~\cite{mimiccxr,mimiccxrdicom,mimiccxrjpg,physionet} is necessary. Both datasets are accessible only on Physionet, requiring the signature of a data use agreement by a logged-in user. Access to the MIMIC-CXR dataset requires free online courses in HIPAA regulations and human research subject protection.
The main uses intended for our data include:
\begin{itemize}
\item combining fixations into heatmaps, for use as an attention label and research on saliency maps and related subjects;
\item using the fixations as a nonuniformly sampled sequence of attention locations;
\item combining the timestamped transcriptions with the fixations for more specific localization to each abnormality;
\item associating the pupil data with the fixations for more information on the cognitive load of each fixation;
\item validating abnormalities parsed from the transcriptions using the image-level labels;
\item validating the locations found from the eye-tracking data through the abnormality ellipses;
\item using the chest bounding boxes for normalizing the location of the lungs while performing other analyses; and
\item using the chest bounding boxes for training a model to output bounding boxes for unseen data.
\end{itemize}
Other possible uses of the data may include using the certainties provided by the radiologists in uncertainty quantification research and the reports and their transcriptions in image captioning for chest x-rays. In \url{https://github.com/ricbl/eyetracking}, we provide examples on how to generate heatmaps, how to normalize the location of heatmaps using the chest bounding box, how to filter the fixations, how to calculate brightness of the shown \gls*{CXR} in any given time during dictation, and how to load and use the tables from the dataset.
\subsection*{Eye-tracking data}
There are uncertainties in the eye-tracking measurement pipeline. To represent them, we suggest following a method used in the generation of heatmaps in the visual attention modeling literature and modeling the location of each fixation as a Gaussian with a standard deviation of 1\degree\ of visual angle~\cite{onedegree}. We provide pixel resolution per visual angle for each axis of the image, so the Gaussian will be slightly anisotropic in the image space. For some applications, e.g., when generating one embedding vector per fixation for sequence analysis, it might be beneficial to filter out fixations that happened outside of the image. Among other reasons, fixations may have happened outside of the image because there were two buttons in the dictation screen: one to indicate that the dictation was over, and one to reset the windowing, zooming, and panning of the image.
\subsection*{Abnormality labels}
Figure~\ref{fig:labels} shows the hierarchy between the labels of our study and the labels from the MIMIC-CXR dataset. Not all labels present in one dataset have an equivalent in other datasets. This hierarchy was produced to the best of their understanding of the MIMIC-CXR labels. Supplementary File 1 provides the definition of the labels agreed upon among the radiologists who participated in the study.
\subsection*{Pupil data}
Bruny\'{e} et al.~\cite{pupilbrightness} showed that the pupil diameter of pathologists reliably increased for more difficult cases, providing an indicator of cognitive engagement. In our dataset, we provide the normalized pupil area, whose square root is equivalent to the normalized pupil diameter. The normalization was performed by a division by the area of the pupil in a standardized screen. However, the variation of the screen brightness, caused by windowing, zooming, and panning, may cause more variation in pupil area than psychological reasons. We suggest including another normalization, using the division by a value representing the screen brightness at each moment, similarly to what was done by Bruny\'{e} et al.~\cite{pupilbrightness}. For the calculation of this value, we suggest summing the intensity of pixels of the shown \gls*{CXR}. The part of the image shown in each moment and the part of the screen where it was shown are provided in the dataset. When considering the windowing of the \gls*{CXR}, images are shown following
\begin{equation}
shown\_image = min\left(max\left(\frac{original\_image-window\_level}{window\_width}+0.5,0\right),1\right),
\end{equation}
where $window\_width$ can have values from 1.5e-05 to 2 and is usually initialized to 1, $window\_level$ can have values from 0 to 1 and is usually initialized to 0.5, $shown\_image$ is the image sent to the screen, and $original\_image$ is the loaded DICOM image normalized so that its possible range is from 0 to 1, which usually means dividing the image intensities by 4,096. The initial $window\_width$, $window\_level$, and the maximum intensity value were loaded from the DICOM tags of each image file.
\subsection*{Limitations of the study}
\begin{itemize}
\item Limited information presented to radiologists: our setup used a single screen, but multiple screens are used in a clinical setting. With multiple screens, the eye-tracker setup is more complex and would have to be validated. We limited the study to show only frontal \gls*{CXR}. Lateral views, past \glspl*{CXR}, and clinical information were not presented.
\item Report: in clinical practice, reports can be modified after a first transcription. We limited the editing to corrections of the transcription of the original dictation and deletions of dictation mistakes. This limitation was needed to assign timestamps to each word and ensure the radiologist saw the finding while the eye tracker was on. Several radiologists use templates for their reports in clinical practice, only dictating small parts of the report. We did not test such a dictation method.
\item Head position: even though we used the remote mode for the eye tracker, which allows for some freedom of movement, the head movement, posture, and the distance from head to the screen were still more limited than in clinical practice, when radiologists can get closer to the screen to see a detail, for example. Because of this limitation, the use of zooming was probably more frequent than in clinical practice. Furthermore, radiologists mentioned that, with the limitations in position, they became fatigued faster than usual.
\item \gls*{CXR} dataset: we collected readings for images of only one dataset, so the current dataset may have ensuing biases. The radiologists also characterized images from the MIMIC-CXR dataset as having lower quality than usual for their practice, in aspects like the field-of-view excluding small parts of the lung and the blurring present in some images.
\item Display: the GPU of the computer used to display the \glspl*{CXR} supported only 8-bit display, so not all intensities of the original DICOM were shown, reducing the image quality and possibly changing the way radiologists interacted with the \gls*{CXR}. This limitation was partially remediated by allowing the windowing of the image to be changed.
\item Calibration cost: calibrations happened every 45 to 60 minutes, and sometimes more than five retries were needed to reach quality thresholds. The clinical implementation of the data collection method described in this paper for the collection of larges quantities of data might cause an undesired cost to the radiologist reading process.
\item Unautomated processes: another person was in the room coordinating calibrations and checking for low data quality. This same person raised the need for a recalibration or a change in position of the radiologist. This person might have to be replaced by automated processes if this data collection method is implemented in clinical practice.
\end{itemize}
\section*{Code availability}
The code used for all automatic processes described in this paper, involving sampling, collection, processing, and validation of data, is available at \url{https://github.com/ricbl/eyetracking}~\cite{mycodezenodo}. The software and versions we used were: MATLAB R2019a, Psychtoolbox 3.0.17~\cite{ptb1,ptb2,ptb3}, Python 3.7.7, edfapi 3.1, EYELINK II CL v5.15, Eyelink GL Version 1.2 Sensor=AC7, EDF2ASC 3.1, librosa 0.8.0~\cite{librosa}, numpy 1.19.1~\cite{numpy}, pandas 1.1.1~\cite{pandas,pandaszenodo}, matplotlib 3.5.1~\cite{matplotlib1,matplotlib2}, statsmodels 0.12.2~\cite{statsmodels}, shapely 1.7.1~\cite{shapely}, scikit-image 0.17.2~\cite{scikit-image}, pyrubberband 0.3.0, pydicom 2.1.2~\cite{pydicom}, pydub 0.24.1, soundfile 0.10.3.post1, pyttsx3 2.90, pillow 8.0.1~\cite{pillow}, scikit-learn 0.23.2~\cite{scikit-learn}, nltk 3.5~\cite{nltk}, syllables 1.0.0, moviepy 1.0.3~\cite{moviepy}, opencv 3.4.2~\cite{opencv_library}, Ubuntu 18.04.5 LTS, espeak 1.48.04, joblib 1.1.0, ffmpeg 3.4.8, and rubberband-cli 1.8.1.
|
\section{Introduction} \label{sec:intro}
Star formation is one of the most important evolutionary processes that shape galaxies over cosmic times. Either from the inter-galactic medium or through galaxy-galaxy interactions, the accretion of gas into a galaxy potential well provides the fuel for future star formation \citep[e.g.,][]{DiMatteo2007,Bournaud2009}. The mechanisms behind the conversion of gas into stars have been investigated in both distant and nearby galaxies \citep[][]{Kennicutt&Evans2012,Madau&Dickinson2014}. The \citet{Kennicutt1989,Kennicutt1998} seminal studies of the galaxy star formation scaling relations in terms of both the star formation rate and neutral gas surface densities ($\Sigma_{\rm SFR}$ and $\Sigma_{\rm gas}$, respectively), showed they are strongly correlated. More recent studies of the scaling laws between gas, stars, and star formation activity show that the latter is most closely related to molecular gas (H$_2$), and focus on the mechanisms that convert H$_2$ into stars, as the main gas reservoir for star formation \citep[][]{WongBlitz2002,Kennicutt2007,Bigiel2008,Leroy2008,Bigiel2011,Leroy2013}.
Stars form in Giant Molecular Clouds (GMCs) in which the molecular gas is the main constituent \citep[e.g.,][]{Sanders1985}. We usually trace molecular gas through observations of the low-$J$ transitions of the carbon monoxide (CO) molecule which provide a good measure of the total molecular mass. The $^{12}$C$^{16}$O($J=1\,-\,0$) transition has been commonly used as a tracer of H$_2$ since it is the second most abundant molecule and it can be easily excited in the cold Interestellar Medium (ISM). The CO(1-0) emission line is usually optically thick, and the conversion of CO luminosity, ${L}'_{\rm CO 1-0}$, into molecular gas mass, $M_{\rm H_2}$, is done through a CO-to-H$_2$ conversion factor $\alpha_{\rm CO}$ \cite[e.g.,][]{Bolatto2013} which appears reasonably constant in the molecular regions of galactic disks but changes at low-metallicities and frequently in galaxy centers in response to environmental conditions \citep[e.g.,][]{Wolfire2010,Narayanan2012}.
In the last decades a sharp increase in optical data on galaxies has enabled the detailed study of structure assembly in the Universe, with the goal of understanding the mechanisms that drive the Universe from the very smooth state imprinted on the cosmic microwave background radiation to the galaxies we observe today. Optical spectroscopic surveys (e.g, zCOSMOS, \citealt{Lilly2007}; Sloan Digital Sky Survey III, \citealt{Shadab2015}; KMOS$^{\rm 3D}$, \citealt{Wisnioski2015}; SINS, \citealt{ForsterSchreiber2009}) have shown the relations between star formation, stellar population, nuclear activity, and metal enrichment for unresolved galaxies in a broad range of redshits. Meanwhile, gas surveys of nearby galaxies have enabled the exploration of the physics behind the star formation relations \citep[e.g.][]{Leroy2008,Leroy2013,Saintonge2011b,Saintonge2017}. These data have revealed that the star formation rate responds to two main factors: the molecular gas content and the stellar potential of the system. An important piece of information is the internal structure of the galaxies. The new generation of Integrated Field Unit (IFU) spectroscopy surveys (e.g., Calar Alto Legacy Integral Field Area, CALIFA, \citealt{Sanchez2012}; SAMI, \citealt{Croom2012}; MaNGA, \citealt{Bundy2015}) have provided detailed spectral imaging data with unprecedented spectral and spatial coverage and good resolution, giving the opportunity to map metallicities, dynamics, extinctions, SFRs, stellar mass density, and other quantities across galaxies. In addition, imaging spectroscopy of the molecular gas from millimeter-wave interferometers \citep{Bolatto2017,Lin2019,Leroy2021} adds invaluable information to understand the baryon cycle in galaxies in the local Universe, where star formation has experienced a drastic decline since the peak of cosmic activity \citep[][]{Madau&Dickinson2014}.
The study of star formation in galaxies demands a holistic approach, since the phenomenon is controlled by multiple processes and it covers a broad range of scales and environments. The analysis of a broad range of galaxy types with multi-waveband datasets is therefore essential to understand the physical conditions that drive star formation activity. The Extragalactic Database for Galaxy Evolution (EDGE) survey is one of the legacy programs completed by the Combined Array for Millimeter-wave Astronomy (CARMA) interferometer \citep{Bock2006}, spanning imaging observations of CO emission in 126 local galaxies. The EDGE survey, combined with the IFU spectroscopy from the CALIFA survey \citep[][]{Sanchez2012}, constitute the EDGE-CALIFA survey \citep[][]{Bolatto2017}, which provides $^{12}$CO and $^{13}$CO ($J=1-0$) images at good sensitivity and angular resolution covering the CALIFA field-of-view.
In this work, we investigate the star formation efficiency (SFE$_{\rm gas}$, where SFE$_{\rm gas}$ [yr$^{-1}$] $=\Sigma_{\rm SFR}$/$\Sigma_{\rm gas}$) in the EDGE-CALIFA survey taking advantage of its large multiwavelength data for 81 local galaxies with low inclinations. In particular, we investigate how the SFE$_{\rm gas}$ depends on physical quantities such as galactocentric radius, stellar surface density, mid-plane gas pressure, orbital timescale, and the stability of the gas disk to collapse. This paper is organized as follows: Section \ref{S2_Observations} explains the main characteristics of the EDGE-CALIFA survey and the sample selection. In section \ref{S4_Methods} we present the methods employed for data analysis, including the CO stacking procedure and the equations we used to derive the basic quantities. Finally, in sections \ref{S5_Results} and \ref{S6_Conclusion} we present our results, discussion, summary and conclusions of this work, respectively.
\section{DATA PRODUCTS}
\label{S2_Observations}
\subsection{The EDGE and CALIFA surveys}
\label{The_sample}
The EDGE-CALIFA survey \citep[][]{Bolatto2017} is based on the optical Integrated Field Spectroscopy (IFS) CALIFA and CO EDGE surveys. In the next paragraphs, we briefly summarize the main features of these two datasets.
The Calar Alto Legacy Integral Field Area survey, CALIFA \citep[][]{Sanchez2012}, comprises a sample of approximately 800 galaxies at $z\approx0$. The data were acquired by using the combination of the PMAS/PPAK IFU instrument \citep[][]{Roth2005} and the 3.5 m telescope from the Calar Alto Observatory. PMAS/PPAK uses 331 fibers each with a diameter of ${2}''.7$ sorted in an hexagonal shape which covers a field-of-view (FoV) of $\sim$ 1 arcmin$^2$. Its average resolution is $\lambda / \Delta \lambda \sim 850$ at ${\rm \sim 5000 \AA}$ with a wavelength range that spans from $3745$ to ${\rm 7300 \AA}$. CALIFA galaxies are selected such that their isophotal diameters, $D_{25}$, match well the PMA/PPAK FoV, and they range from $45$ to $80$ arcsec in the SDSS $r$-band \citep[][]{Walcher2014}. The CALIFA survey uses a data reduction pipeline designed to produce data cubes with more than $5000$ spectra and with a sampling of $1\times1$ arcsec$^2$ per spaxel. For more details, see \cite{Sanchez2012}.
The Extragalactic Database for Galaxy Evolution, EDGE, is a large intereferometric CO and $^{13}$CO $J = 1-0$ survey which comprises 126 galaxies selected from the CALIFA survey. The observations were taken using the Combined Array for Millimeterwave Astronomy \citep[CARMA,][]{Bock2006} in a combination of the E and D configurations for a total of roughly 4.3 hr per source, with a typical resolution of 8 and 4 arcsec, respectively. The observations used half-beam-spaced seven-point hexagonal mosaics giving a half-power power field-of-view of radius $\sim{50}\arcsec$. The data are primary-gain corrected and masked where the primary beam correction is greater than a factor of $2.5$. The final maps, resulting from the combination of E and D array data, have a velocity resolution of 20 km s$^{-1}$ and typical velocity coverage of 860 km s$^{-1}$, a typical angular resolution of $4.5\arcsec$, and a rms sensitivity of 30 mK at the velocity resolution. For more details, see \cite{Bolatto2017}.
\begin{figure*}
\includegraphics[width=18.cm]{NGC0551_6.pdf}
\caption{Example showing effects of spectral stacking. The average CO spectrum within an annulus that spans from $0.65$ to $0.75$ $r_{25}$ in NGC $0551$ is shown. The left panel shows the average of all spectra in the annulus in the observed velocity frame. The right panel shows the average in the velocity frame relative to H$\alpha$, along with the best Gaussian fit profile (green dashed line).}
\label{Stacked_Example}
\end{figure*}
\vspace{1.cm}
\subsection{{\tt{edge}\_\tt{pydb}} database}
\label{Database}
The EDGE-CALIFA survey provides global (integrated) and spatially resolved information about the molecular/ionized gas and stellar components in 126 nearby galaxies, comprising $\sim15000$ individual lines-of-sight. In the context of this work, and to provide easy yet robust access to this large volume of data, we have used one main source of data to perform our analysis.
The {\tt{edge}\_\tt{pydb}} database (Wong et al. in prep.) is a versatile {\tt PYTHON} environment that allows easy access and filtering of the EDGE-CALIFA data in the variety of analyses we aim to perform. {\tt{edge}\_\tt{pydb}} encompasses a combination of global galaxy properties and spatially resolved information, with a special emphasis on estimation of the CO moments from smoothed and masked versions of the CARMA CO datacubes. All data have been convolved to a common angular resolution of ${7}\arcsec$. By using the {\tt{PIPE3D}} data analysis pipeline \citep[see][ for more details]{Sanchez2015,Sanchez2016_2}, the convolved optical datacubes are reprocessed to generate two dimensional maps at 7\arcsec\ resolution. The pipeline fits the stellar continuum to the emission lines for each spaxel in each datacube (adopting a \citealt{Salpeter1955} Initial Mass Function, IMF), generating maps sampled on a square grid with a spacing of ${3}''$ in RA and DEC. To identify a given pixel in the grid, the data are organized by using a reference position (taken from HyperLEDA\footnote{http://leda.univ-lyon1.fr/}) and an offset indicating spatial position. The final database also contains ancillary data, including information from HyperLEDA, NED\footnote{https://ned.ipac.caltech.edu/}, among others.
\section{Methods}
\label{S4_Methods}
\subsection{Stacking of the CO spectra}
\label{Stacking_CO}
Although many EDGE-CALIFA galaxies have high signal-to-noise detections of CO emission in their central regions, emission is generally faint in their outer parts. Typically, the decrease in emission takes place from $r=0.5r_{25}$ outwards (around $1.1R_{\rm e}$, by assuming that $r_{25}\approx 2.7 R_{\rm e}$; \citealt{Sanchez2014}). \cite{Bolatto2017} published maps of velocity-integrated CO emission and discussed various masking techniques for recovering flux and producing maps with good signal to noise; even so, they tend to miss flux in regions of weak emission and to underestimate the CO flux (see Figure 9 in \citealt{Bolatto2017}). Since one of the main goals of this work is to find how the $\rm H_2$ content changes as a function of radius, it is essential to recover low-brightness CO emission line in the outermost parts of galaxies.
Maps with both good spatial coverage and sensitivity are crucial to set thresholds and timescales for these dependencies. In order to cover a broad range of galactocentric radii, we perform spectral stacking of the $^{12}$CO ($J=1-0$) emission using the H$\alpha$ velocities to coherently align the spectra while integrating in rings. The CO spectral stacking helps recover CO flux in the outer parts of our galaxies, improving our ability to probe the SFE$_{\rm gas}$ in a variety of environments. Many of the molecular gas surveys have measured some of these dependencies in a similar fashion (e.g., using the CO [$J=2-1$] spectral stacking; \citealt{Schruba2011}), although they mostly covered a small range of morphological types and/or stellar masses, or were limited to very local volumes that are subject to cosmic variance because they represent our particular local environment. Although the EDGE-CALIFA survey does not yet encompass resolved H$\rm I$ observations, we will explore the efficiency with respect to total gas and compare it to previous results by assuming a prescription for the atomic gas while keeping in mind the limitations of this methodology.
We perform a CO emission line stacking procedure following the methodology described by \cite{Schruba2011}. The method relies on using the IFU H$\alpha$ velocity data to define the velocity range for integrating CO emission. The key assumption of this method is that both the H$\alpha$ and the CO velocities are similar at any galaxy location. This assumption is consistent with results by \cite{Levy2018}, who found a median value for the difference between the CO and H$\alpha$ rotation curve of $\Delta V = V_{\rm rot}({\rm CO}) - V_{\rm rot}({\rm H \alpha})= 14$ km s$^{-1}$ (within our 20 km s$^{-1}$ channel width) when analyzing a sub-sample of 17 EDGE-CALIFA rotation-dominated galaxies. As we will discuss later, after shifting CO spectra to the H$\alpha$ velocity, we integrate over a window designed to minimize missing CO flux. The smaller the velocity differences between CO and H$\alpha$, the better the signal-to-noise. Similarly, the smaller the velocity window we implement, the smaller the noise in the integrated flux estimate.
We constructed an algorithm coded in {\tt PYTHON} that implements this procedure. Since we are interested in radial variations in galactic properties, we stack in radial bins $0.1r_{25}$ wide. In practice, galactocentric radius is usually a well determined observable and it is covariant with other useful local parameters, which makes it a very useful ordinate \citep{Schruba2011}.
We recover the CO line emission by applying radial stacking based on the following steps: We convert H${\alpha}$ velocity from the optical into the radio velocity convention. Then, for each spaxel in an annulus we shift the CO spectrum by the negative H${\alpha}$ velocity. This step aligns the CO spectrum for each line-of-sight at zero velocity if the intrinsic H$\alpha$ and CO velocities are identical. We then average all the velocity-shifted CO spaxels in an annulus, and integrate the resulting average spectrum over a given velocity window to produce the average intensity in the annulus.
Figure \ref{Stacked_Example} shows the usefulness of the stacking procedure in recovering CO emission. As an example, we show the average CO spectrum of NGC 0551 within an annulus that spans from 0.65 to 0.75 $r_{25}$ ($\sim 1.3-1.7 \, R_{\rm e}$). The left panel contains the average CO spectra within the given annulus using the observed velocity frame, while the right panel shows the average CO spectra after shifting by the observed H${\alpha}$ velocity. If the CO and H$\alpha$ velocities are identical for all spaxels, then the resulting CO emission would appear at zero velocity. This procedure allows us to co--add CO intensities coherently and reject noise. Figure \ref{Stacked_Example} also shows the best Gaussian fit for the averaged-stacked spectra. We expect that in an ideal case the total intensity integrated over the full velocity range ($\sim$860 km s$^{-1}$) is exactly the same in both cases, but the noise would be much larger without the spectral stacking. Without performing the stacking procedure the CO line emission is not evident, and the signal-to-noise ratio in the measurement of CO velocity-integrated intensity is lower. Interferometric deconvolution artifacts that produce negative intensities at some velocities, resulting from incomplete $uv$ sampling and spatial filtering, would also get into the integration more easily without stacking and artificially reduce the intensity.
\begin{figure}
\includegraphics[width=7.9cm]{FWHM_vs_r_median.pdf}
\includegraphics[width=7.6cm]{STACK_UNSTACK.pdf}\\
\caption{ {\it Top:} FWHM of CO line as a function of galactocentric radius. Small colored dots show the FWHM of a Gaussian fit to the stacked spectrum in an annulus. Large green squares indicate the FWHM lying above 80\% of the points at that radius, and the dashed green line is the fit to the squares; we use this function to define the window of flux integration as a function of $r_{\rm gal}$. The gray-dashed line marks the limit at which we reject spectra with a Gaussian fit narrower than 40 km s$^{-1}$ (2 channels in the CO datacubes). {\it Bottom:} Ratio between the final stacked and unstacked integrated CO(1--0) line intensity per annulus for annuli located at $r/r_{25} > 0.5$, which include just $2\sigma$ detection spaxels.}
\label{FWHM_radius}
\end{figure}
\subsection{Extracting fluxes from stacked spectra}
\label{Extracting_fluxes}
After we compute the stacked spectra, we extract the total CO fluxes for each annulus as a function of galactocentric radius. To do this in a way that is likely to include all the CO flux but minimizes the noise, we want to select a matched velocity range that is just large enough to include all CO emission and exclude the baseline (which only adds noise). In order to investigate the ideal integration range we fit Gaussian profiles to each averaged-stacked CO spectrum with a detection. We reject fits which have central velocities more than $\pm 80$ km s$^{-1}$ from zero velocity. We also reject spectra with FWHMs narrower than 40 km s$^{-1}$ (2 channels). Results for valid stacked spectra fits are shown in the top panel of Figure \ref{FWHM_radius}, color coded by the reduced chi-squared of the fit and plotted against normalized galactocentric radius.
We use these data to define a velocity window for the integrated CO line emission fluxes in the stacked spectra. For each radial bin, we define an integration range that guarantees that we integrate the CO line profile between $\pm {\rm FWHM}$ in at least 80\% of annuli. This is represented by the green dots in Figure \ref{FWHM_radius}. We assume that this window is sufficient to contain most of the CO flux, and we can use it to compute errors where no CO is detected. To obtain a prescription we fit the best third-order polynomial to the green squares (green dashed-line) as a function of galactocentric radius, ${\rm FWHM}$($r_{\rm gal}$). Finally, we recompute the CO line emission fluxes for the stacked spectra by integrating the CO stacked spectrum over $\pm {\rm FWHM}(r_{\rm gal})$. We extract the integrated flux uncertainties by taking the rms from the emission free part of the stacked CO spectra.
Using spectral stacking we reach a typical deprojected CO intensity $3\sigma$ uncertainty of $I_{\rm CO} \approx 0.25$ K km s$^{-1}$, or a $3\sigma$ surface density sensitivity of $\Sigma_{\rm mol} \approx 1.1\, M_{\odot}$ pc$^{-2}$, which represents the typical sensitivity in the outermost regions of galaxy disks. The bottom panel of Figure \ref{FWHM_radius} shows the ratio between the final stacked and unstacked integrated CO(1-0) line intensity, per annulus, located at $r>0.5 r_{25}$ (or $r>1.3 R_{\rm e}$), and includes just $2\sigma$ detection spaxels. The histogram shows that the distribution peaks at $\log[I_{\rm CO,Stack}/I_{\rm CO,Unstack}]\sim 0.47$, meaning that, in overall, we are recovering $\sim 3$ times more flux with the stacking procedure.
\subsection{Basic equations and assumptions}
\label{Basic_equations}
To compute the extinction-corrected SFRs, we estimate the extinction \citep[based on the Balmer decrement; see][]{Bolatto2017} for each 7\arcsec\ spaxel using
\begin{equation}
A_{\rm H{\alpha}} = 5.86 \log{\left ( \frac{F_{\rm H{\alpha}}}{2.86F_{\rm H{\beta}}}\right )}
\end{equation}
where $F_{\rm H \alpha}$ and $F_{\rm H \beta}$ are the fluxes of the respective Balmer lines, and the coefficients assumes a \cite{Cardelli1989} extinction curve and an unextincted flux ratio of 2.86 for case B recombination. Then, the corresponding SFR (in M$_{\odot}$ yr$^{-1}$) is obtained using \citep{RosaGonzalez2002}
\begin{equation}
{\rm SFR} = 1.61 \times 7.9 \times 10^{-42} F_{\rm H{\alpha}} 10^{\frac{A_{\rm H{\alpha}}}{2.5}},
\label{eq:sfr}
\end{equation}
\noindent which adopts a Salpeter Initial Mass Function (IMF) corrected by a factor of 1.61 to move it to a Kroupa IMF \citep{Speagle2014}. We use this to compute the Star Formation Rate surface density, $\Sigma_{\rm SFR}$ in M$_\odot$\,yr$^{-1}$\,kpc$^{-2}$, by dividing by the face-on area corresponding to a 7\arcsec\ spaxel, given the angular diameter distance to the galaxy.
The gas surface density is computed as $\Sigma_{\rm gas} = \Sigma_{\rm mol} + \Sigma_{\rm atom}$, where $\Sigma_{\rm H_2}$ is derived from the integrated CO intensity, $I_{\rm CO}$, by adopting a Milky-Way constant CO-to-H$_2$ conversion factor, $X_{\rm CO} = 2 \times 10^{20}$ cm$^{-2}$ (K km s$^{-1}$)$^{-1}$ (or $\alpha_{\rm CO} = 4.3$ M$_\odot$ $[\rm K \, km \, s^{-1} \, pc^{-2}]^{-1}$). For the CO $J=1 - 0$ emission line, we use the following expression to obtain $\Sigma_{\rm mol}$ \citep[i.e., ][]{Leroy2008}
\begin{equation}
\Sigma_{\rm mol} = 4.4 \cos i\, I_{\rm CO},
\end{equation}
\noindent where $I_{\rm CO}$ is in K\,km s$^{-1}$, $\Sigma_{\rm mol}$ is in M$_\odot$\,pc$^{-2}$, and $i$ is the inclination of the galaxy. This equation takes into account the mass correction due to the cosmic abundance of Helium.
To include in our calculations $\Sigma_{\rm atom}$ despite the fact that we do not have resolved H$\rm I$ data, we assume a constant $\Sigma_{\rm atom} = 6 $ M$_{\odot}$ pc$^{-2}$ for face-on disks. This is approximately correct (within a factor of 2) for spiral galaxies out to $r \sim r_{25}$ \citep{Walter2008,Leroy2008}. This value is also in agreement with Monte Carlo simulations performed by \cite{Barrera-Ballesteros2021} to test different values of $\Sigma_{\rm atom}$; they obtain a normal distribution of $\Sigma_{\rm atom}=7$ M$_\odot$ pc$^{-2}$, with a standard deviation of $2$ M$_\odot$ pc$^{-2}$. We also test the influence of metallicity in the CO-to-H$_2$ conversion factor, $\alpha_{\rm CO}$, by using the following equation \citep[from Equation 31 in][]{Bolatto2013}:
\begin{equation}
\alpha_{\rm CO} = 2.9 \exp \left ( \frac{+0.4}{Z' \Sigma^{100}_{\rm GMC}} \right ) \left ( \frac{\Sigma_{\rm total}}{100 \, \rm M_\odot \, pc^{-2} } \right )^{-\gamma},
\label{Alpha_CO}
\end{equation}
in M$_\odot$ $\rm (K \, km s^{-1} \, pc^{-2} )^{-1}$, $\gamma \approx 0.5$ for $\Sigma_{\rm total} > 100$ M$_\odot$ pc$^{-2}$ and $\gamma =0$ otherwise. We adopt the empirical calibrator based on the O3N2 ratio from \cite{Marino2013}, and then we use equation 2 from \cite{Marino2013} to obtain the oxygen abundances, $\rm 12 + \log(O/H)$. Finally, we derive the metallicity normalized to the solar value, ${Z}' = \rm [O/H]/[O/H]_\odot$, where $\rm [O/H]_\odot = 4.9\times 10^{-4}$ \citep[][]{Baumgartner&Mushotzky2006}.
Although there are many definitions for star formation efficiency (SFE$_{\rm gas}$), in this work we use SFR surface density per unit neutral gas surface density (atomic and molecular), $\Sigma_{\rm gas}=\Sigma_{\rm mol} +\Sigma_{\rm atom}$, in units of yr$^{-1}$ for each line-of-sight (LoS),
\begin{equation}
{\rm SFE_{\rm gas}} = \frac{\Sigma_{\rm SFR}}{\Sigma_{\rm gas}}.
\label{eq:sfe}
\end{equation}
Midplane gas pressure, $P_{\rm h}$ is computed using the expression by \cite{Elmegreen1989},
\begin{equation}
P_{\rm h} \approx \frac{\pi}{2} \, G \, \Sigma^{2}_{\rm gas} + \frac{\pi}{2} \, G \, \frac{\sigma_{\rm g}}{\sigma_{\star , z} }\Sigma_{\star}\Sigma_{\rm gas},
\label{Midplane_gas_pressure}
\end{equation}
where $\sigma_{\rm g}$ and $\sigma_{\star , z}$ are the gas and stars dispersion velocities, respectively. We correct the $\Sigma_{\star }$ by the same 1.61 factor used for the SFR to translate them to a Kroupa IMF. We assume $ \sigma_{\rm g}=11$ km s$^{-1}$, which has been found to be a typical value in regions where HI is dominant (\citealt{Leroy2008}). This value is also in agreement with the second moments maps included in \cite{Tamburro2009}, and is also consistent with the CO velocity dispersion for a subsample of EDGE-CALIFA galaxies \citep[][]{Levy2018}. $\sigma_{\star , z}$ is the vertical velocity dispersion (in km\,s$^{-1}$) of stars. Although the EDGE-CALIFA database includes $\sigma_{z}$ measurements that could allow us to model $\sigma_{\star , z}$, the instrumental resolution of the survey constrains us to use them just in the central parts of the galaxies (for details, see \citealt{Sanchez2012}). Therefore, and following the assumptions and derivation included in \cite{Leroy2008}, we use the following expression for $\sigma_{\star , z} $:
\begin{equation}
\sigma_{\star , z} = \sqrt{\frac{2 \pi G l_{\star}}{7.3}} \Sigma^{0.5}_{\star},
\label{sigma_star}
\end{equation}
\noindent where $l_{\star}$ is the disk stellar exponential scale length obtained by fitting azimuthally averaged profiles to $\Sigma_{\star}$ in the SDSS $r$-band and $G=4.301\times10^{-3}$ pc \,M$_\odot^{-1}$\,km$^{2}$\,s$^{-2}$. In cases where we do not have $l_{\star}$ measurements, we use the relation $l_{\star} = [0.25\pm 0.01] r_{25}$ since it corresponds to the best linear fit for our data. See Section \ref{Exponential_scale} for more information about how both $l_{\star}$ and the $l_{\star}$-$r_{25}$ relation are derived.
The dynamical equilibrium pressure ($P_{\rm DE}$) is computed following a similar methodology as for $P_{\rm h}$ \citep[e.g][]{ElmegreenPavarro1994,HerreraCamus2017,Fisher2019,Schruba2019}. Assuming that the gas disk scale height is much smaller than the stellar scale height, and neglecting the gravity from dark matter, we write $P_{\rm DE}$ as \citep{Sun2020}:
\begin{equation}
P_{\rm DE} \approx \frac{\pi}{2} \, G \, \Sigma^{2}_{\rm gas} + \Sigma_{\rm gas} \sigma_{\rm gas , z} \sqrt{2\, G \, \rho_{\star}}.
\label{Dynamical_equilibrium_pressure}
\end{equation}
Here, we assume that $\sigma_{\rm gas , z}=\sigma_{\rm g}=11\,{\rm km \, s^{-1}}$, and $\rho_{\star}$ is the mid-plane stellar volume density from the observed surface density in a kpc-size aperture,
\begin{equation}
\rho_\star = \frac{\Sigma_{\star}}{0.54 l_{\star}}.
\label{rho_star}
\end{equation}
\noindent This equation assumes that the exponential stellar scale height, $h_\star$, is related to the stellar scale length, $l_\star$, by $h_\star/l_\star = 7.3 \pm 2.2$ \citep[][]{Kregel2002}.
The orbital timescale, $\tau_{\rm orb}$, is usually used in the analysis of star formation law dependencies since it can be comparable to timescale of the star formation \citep[e.g.,][]{Silk1997,Elmegreen1997}. Following \cite{Kennicutt1998} and \cite{WongBlitz2002}, we compute $\tau_{\rm orb}$ using:
\begin{equation}
\tau^{-1}_{\rm orb} = \frac{v \left ( r_{\rm gal} \right )}{2 \pi r_{\rm gal}}
\label{Orbital_timescale}
\end{equation}
where $v \left ( r_{\rm gal} \right )$ is the rotational velocity at a galactocentric radius $r_{\rm gal}$. We obtain the H$\alpha$ rotation curves for EDGE-CALIFA galaxies from \cite{Levy2018}. We use them to adjust an Universal Rotation Curve \citep[URC, ][]{Persic1996} for each galaxy to avoid the noise in the inner and outer edges of the H$\alpha$ rotation curves.
\begin{figure*}
\begin{center}
\includegraphics[width=8.8cm]{LENGTH.pdf}
\includegraphics[width=8.8cm]{LENGTH_2.pdf}
\end{center}
\caption{{\it Left:} Comparison between the stellar, ${\it l}_{\star}$, and molecular length scales, ${\it l}_{\rm mol}$, computed by fitting exponential profiles to the respective surface densities as a function of galactocentric radius. The colored circles correspond to 61 EDGE-CALIFA galaxies color-coded by stellar mass derived from SED fitting (see Section \ref{The_sample}). The inset panel shows the comparison between ${\it l}_{\rm mol}$ and the isophotal radius $r_{25}$. The triangles represent uncertain results for which measurements are smaller than $3\sigma$. The solid red and dashed blue lines illustrate the 1:1 scaling and the OLS linear bisector fit (forced through the origin) for all the sources, respectively. {\it Right:} Relationship between the radii that enclose 50\% of the molecular gas and the stellar mass, ${\it R}_{\rm 50, mol}$ and ${\it R}_{50, \star}$, respectively. Conventions and symbols are as in the left panel.}
\label{FIG_14_LENGTH}
\end{figure*}
\begin{figure*}
\includegraphics[width=18cm]{Leroy2008_fig12A_Morph2_4panels.pdf}
\caption{SFE$_{\rm gas}$ vs galactocentric radius. Each line indicates the average SFE$_{\rm gas}$ for individual galaxies in $0.1r_{25}$-wide tilted annuli after stacking. The morphological group for the galaxies in each panel is indicated by the legend in that panel. The plot shows that the SFE$_{\rm gas}$ in individual galaxies generally decreases as a function of galactocentric radius and that the dispersion in SFE$_{\rm gas}$ at particular radii is due mostly to differences between galaxies.}
\label{FIG_1_Leroy_2008_1A}
\end{figure*}
\begin{figure}
\includegraphics[width=8cm]{Sigma_H2_Sigma_Hi.pdf}
\caption{Histogram of galactocentric radii at which $\Sigma_{\rm H_2}$ drops to 6 M$_\odot$ pc$^{-2}$, which is the value of $\Sigma_{\rm atom}$ assumed for EDGE-CALIFA galaxies in this work. The dashed vertical black line is the mean value of $r/R_{\rm e}$ at which this occurs, corresponding to $1.1$ ($0.4 r_{25}$; see inset panel). The gray area represents the uncertainty in mean value of $r/R_{\rm e}$. The inset shows a similar histogram for $r/r_{25}$.}
\label{FIG_1_Leroy2008_2}
\end{figure}
\begin{figure}
\includegraphics[width=8.75cm]{Leroy2008_fig1_2B_Morph.pdf}
\caption{{\it Top:}SFE$_{\rm gas}$ vs galactocentric radius for different morphological types of galaxies. SFE is averaged at each radius over all galaxies of the selected morphological type; types are indicated by shaded color as described in the legend. The vertical extent of the shaded area for each morphogical type is the 1$\sigma$ scatter distribution for that type (see Figure \ref{FIG_1_Leroy_2008_1A}). Circular dots indicate the average SFE$_{\rm gas}$ and galactocentric radius in stacked annuli for all EDGE-CALIFA galaxies; the black solid line is the OLS linear bisector fit to those points using the model $A_{\rm SFE} \times \exp{(-r/r_{\rm SFE})}$. The error bars are the uncertainties of the mean SFE values in each bin. The two dashed black lines show the effect of increasing and decreasing $\Sigma_{\rm atom}$ by a factor of two from its assumed value of 6 M$_\odot$ pc$^{-2}$. The shaded gray band indicates the amount by which the binned SFE$_{\rm gas}$ would increase if we use the metallicity dependent prescription for $\alpha_{\rm CO}$. The green squares are the HERACLES data for spiral galaxies. The figure shows that SFE$_{\rm gas}$ depends on radius, stellar mass and morphological type.}
\label{FIG_1_Leroy2001B}
\end{figure}
We compute the Toomre's instability parameter \citep[$Q$]{Toomre1964} including the effect of stars \citep{Rafikov2001}. The Toomre's instability parameter for the stellar component ($Q_{\rm star}$) is
\begin{equation}
Q_{\rm stars} = \frac{\sigma_{\star, r} \kappa }{\pi G \Sigma_{\star}},
\label{Instability_parameter}
\end{equation}
\noindent where $\sigma_{\star, r}$ is the radial velocity dispersion of the stars. We compute it using $\sigma_{\star, r}=1.67\,\sigma_{\star, z}$, valid for most late-type galaxies \citep[][]{Shapiro2003}. The parameter $\kappa$ is the epicyclic frequency and can be computed as
\begin{equation}
\kappa = 1.41 \frac{v(r)}{r} \sqrt{1+\beta},
\end{equation}
where $\beta = \frac{d \log{\, v \left ( r \right )}}{d \log{r}}$. This derivative is computed based on the URC fit to the H$\alpha$ rotation curve. The Toomre's instability parameter for the gas ($Q_{\rm gas}$) is
\begin{equation}
Q_{\rm gas} = \frac{\sigma_{\rm gas} \kappa }{\pi G \Sigma_{\rm gas}} = \frac{(11\,{\rm km \, s^{-1}}) \kappa }{\pi G \Sigma_{\rm gas}}.
\end{equation}
Since $\Sigma_{\star}$ and $\Sigma_{\rm gas}$ are averaged and stacked by annuli, respectively, then both $Q_{\rm stars}$ and $Q_{\rm gas}$ are derived radially. The condition for instability in the gas+stars disk is then given by
\begin{equation}
\frac{1}{Q_{\rm star+gas}}=\frac{2}{Q_{\rm stars}}\frac{q}{1+q^2}+ \frac{2}{Q_{\rm gas}}R\frac{q}{1+q^2 R^2} > 1,
\label{Rafikov}
\end{equation}
where $q= k \sigma_{\star, r}/\kappa$. Here, $k=2\pi/\lambda$ is the wavenumber at maximum instability. Finally, $R=\sigma_{\rm g}$/$\sigma_{\star,r}$.
\section{Results and Discussion}
\label{S5_Results}
\subsection{Exponential scale lengths}
\label{Exponential_scale}
To investigate the spatial relationship between molecular and stellar components, we compute their exponential scale lengths, $l_{\rm mol}$ and $l_{\star}$, respectively, for a subsample of 68 galaxies. Out of the 81 EDGE-CALIFA galaxies with $i < 75^{\circ}$ , these galaxies are selected since their disks are well fitted by exponential profiles and they have at least three annuli available for the fitting. To avoid annuli within the bulge or with significant variations in $\alpha_{\rm CO}$ usually found in central regions of galaxies \citep[e.g.,][]{Sandstrom2013}, we do not include $\Sigma_{\rm mol}$ and $\Sigma_{\star}$ for $r_{\rm gal} \leq 1.5$ kpc.
It is well known that the CO distribution and star formation activity are closely related \citep[e.g., ][]{Leroy2013}. For instance, \cite{Leroy2009} showed that HERACLES spiral galaxies can be well described by exponential profiles for CO emission in the H$_2$-dominated regions of the disk, with similar CO scale lengths to those for old stars and star-forming tracers, and an early study on the EDGE sample found similar results \citep{Bolatto2017}. Here we use the stacking technique to extend the molecular radial profiles and obtain a better measurement of the distribution.
Although molecular clouds have lifetimes spanning a few to several Myr \citep[similar to the stars that give rise to the H$\alpha$ emission used to compute SFR; e.g.,][]{Blitz&Shu1980,Kawamura2009,Gratier2012}, these are quite short compared with lifetimes of the stellar population in galaxies in the EDGE-CALIFA survey \citep[$0.4$ to $3.9$ Gyr;][]{Barrera-Ballesteros2021}. Consequently, it is not necessarily expected to have comparable distributions for the molecular and the stellar components. However, stellar and CO emission distributions can be similar when the process of converting atomic gas to molecular is driven by the stellar potential \citep[][]{Blitz&Rosolowsky2004,Ostriker2010}. For instance, \cite{Schruba2011} showed a clear correspondence between $l_{\rm CO}$ and $r_{25}$; this correlation is maintained even in the HI-dominated regions of the disk, supporting the role that molecular gas plays in a scenario when the stellar potential well is relevant in collecting material for star formation \citep{Blitz&Rosolowsky2006}. Thus, it is interesting to use the CO stacked data to verify if the exponential decay of $\Sigma_{\rm mol}$ holds in the outer parts of EDGE-CALIFA galaxies.
The left panel of Figure \ref{FIG_14_LENGTH} shows the relation between $l_{\rm mol}$ and $l_{\rm \star}$. The $l_{\rm \star}$ values were obtained by fitting exponential profiles to $\Sigma_{\rm *}(r_{\rm gal})$, after averaging it in annuli, while $l_{\rm mol}$ values were determined from $\Sigma_{\rm mol}(r_{\rm gal})$ derived from the CO stacking procedure. The left panel of Figure \ref{FIG_14_LENGTH} also shows the ordinary least-square (OLS) bisector fit weighted by the uncertainties for all scale lengths measured with better than $3\sigma$ significance (blue dashed line); we find that $l_{\rm mol} = [0.89 \pm 0.04] \,l_{\star}$. This result is in agreement with the relation found by \cite{Bolatto2017} for 46 EDGE-CALIFA galaxies, who obtain $l_{\rm mol} = [ 1.05 \pm 0.06 ] \, l_{\star}$. Compared with \cite{Bolatto2017}, however, the CO radial stacking allows us to compute exponential length scales for a larger galaxy sample (68 in our case) and to constrain them better over a broader range of galactocentric radii. Our results are also in agreement with the exponential length scales for HERACLES ($l_{\rm mol} = [0.9 \pm 0.2] \,l_{\star}$; \citealt{Leroy2008}). The inset in the left panel of Figure \ref{FIG_14_LENGTH} shows the relation between $l_{\rm mol}$ and $r_{25}$. Using an OLS bisector fit, we find that $l_{\rm mol} = [0.22 \pm 0.01]\times r_{25}$, which agrees reasonably with \cite{Young1995}, who find $l_{\rm mol} \approx 0.22 r_{25}$.
In general, resolved molecular gas surveys exhibit similarity between the stellar light and the CO distributions. \cite{Regan2001}, using the CO distribution from the BIMA SONG CO survey, showed that when comparing the scale lengths from exponential fits to the CO and the K-band galaxy profile data for 15 galaxies, the typical CO to stellar scale length ratio is $0.88\pm0.14$. Additionally, single-dish CO measurements plus 3.6$\mu$m data from the HERACLES galaxies show a correspondence between the stellar and molecular disk \citep[][]{Leroy2008,Schruba2011}, with an exponential scale length for CO that follows $l_{\rm mol} \approx 0.2\,r_{25}$.
If the radial distributions for molecular gas and stars are similar, we would expect the radii containing 50\%\ of the CO emission and the star light to also be similar. The right panel of Figure \ref{FIG_14_LENGTH} demonstrates that our data confirm this expectation, as it shows the relation between the radii that enclose 50\% of the molecular gas and the stellar mass, $R_{\rm 50, mol}$ and $R_{50, \star}$, respectively. The dashed blue line represents an ordinary least-square bisector fit (weighted by the uncertainties) for all our 3$\sigma$ detections; we find that $R_{\rm 50, mol} = [0.95 \pm 0.03]\times R_{50, \star}$.
Table \ref{Table} summarizes the properties of the 81 EDGE-CALIFA galaxies included in this work and together with the values for $l_{\rm mol}$, $l_{\star}$, $R_{\rm 50, mol}$, and $R_{50, \star}$ for the 68 galaxies analyzed in this section.
\subsection{SFE and Local Parameters}
In this section, we will look at how local physical parameters affect the star formation efficiency of the total gas, SFE$_{\rm gas}$ $=\Sigma_{\rm SFR}/[\Sigma_{\rm atom}+\Sigma_{\rm mol}]$, following methodologies similar to those used by HERACLES \citep{Leroy2008}, against which we will compare results. We compute efficiencies by dividing the star formation rate surface density obtained from H$\alpha$ corrected for extinction using the Balmer decrement (Eq.~\ref{eq:sfr}) by the total gas surface density (Eq.~\ref{eq:sfe}). As discussed in \S\ref{Basic_equations} we assume a constant $\Sigma_{\rm atom}=6$~M$_\odot$\,pc$^{-2}$.
The EDGE-CALIFA galaxies are generally at larger distances ($\sim$23 to 130 Mpc) than the much more local HERACLES sample (3 to 20 Mpc). Both samples have stellar masses spanning a similar range ($\log[M_{\star}/M_{\odot}] = 9.4$--$11.4$), but EDGE a larger representation of more massive disks and bulges as HERACLS includes mostly late Sb and Sc objects and lower mass galaxies. The parent sample CALIFA galaxies are selected in a large volume to allow adequate representation of the $z=0$ population and numbers that allow statistically significant conclusions for all classes of galaxies represented in the survey \citep[][]{Sanchez2012}. The EDGE follow-up selection is biased toward IR-bright objects, but otherwise tries to preserve the variety and volume of the mother sample. CALIFA does not include dwarf galaxies. EDGE otherwise spans a larger range of properties and has a larger sample size than HERACLES, although with lower spatial resolution ($\sim 1.5$~kpc versus $\sim200$~pc).
We correct our calculations by the inclination of the galaxy (with a $\cos i$ factor, where $i$ is the inclination angle) to represent physical ```face-on'' deprojected surface densities (see \S\ref{Basic_equations}). Our typical $1\sigma$ uncertainty in the SFE$_{\rm gas}$ is $0.22$ dex, dominated by the CO line emission uncertainties derived from the stacking procedure after error propagation.
\begin{figure}
\includegraphics[width=8.5cm]{Leroy2008_fig1_2B_Metallicity.pdf}\\
\hspace{-.5cm}\includegraphics[width=9cm]{Leroy2008_fig2B_2.pdf}
\caption{{\it Top:} Sample distribution of the oxygen abundances, $\rm 12 + log(O/H)$, with the O3N2 as metallicity indicator. The dashed-black line is the assumed solar value, which corresponds to $\rm 12 + log(O/H)_{\odot}= 8.69$. {\it Bottom:} The star formation efficiency of the molecular gas, SFE$_{\rm mol}= \Sigma_{\rm SFR}/\Sigma_{\rm mol}$, vs the ratio between the molecular and the atomic gas surface densities, $R_{\rm mol}= \Sigma_{\rm mol}/\Sigma_{\rm atom}$. Colors code for galactocentric radius (in $R_{\rm e}$) are as indicated by the color bar. Black contours are 80\%, 60\%, 40\%, and 20\% of the points just for detections. Large black filled circles show the mean of EDGE-CALIFA data at each stellar surface density bin; the error bars are the uncertainties of the mean SFE$_{\rm mol}$ values in each bin. The black-solid line shows the OLS linear bisector fit for averaged points of SFE$_{\rm mol}$ over annuli by using the model $y = a x +b$. The shaded region represents uncertainty of the slope derived from the OLS linear bisector fit. The horizontal dashed-blue line is the average SFE$_{\rm mol}$, including the 3$\sigma$ detection, for the sample. The inset panel shows the SFE$_{\rm mol}$ for detections only as a function of galactocentric radius. The blue-dashed line is the average SFE$_{\rm mol}$.}
\label{FIG_1_Metallicity}
\end{figure}
\begin{table*}
\resizebox{.8\linewidth}{!}{
\begin{tabular}{cccccccccc}
\hline\hline
Name & Dist. (Mpc) & $\rm \log[M_{\star}/M_{\odot}]$ & Morph. Class & $\rm \log[M_{\rm mol}/M_{\odot}]$ & Nuclear & $l_{\star}$ (kpc) & $l_{\rm mol}$ (kpc) & $R_{50, \star}$ (kpc) & $R_{50, \rm mol}$ (kpc) \\
\hline
ARP220 & 78.0 & 10.91$\pm$0.09 & Sm & 9.72$\pm$0.0 & LINER & 2.76$\pm$0.24 & 1.7$\pm$0.35 & 3.88$\pm$0.24 & 2.86$\pm$0.35 \\
IC0944 & 100.8 & 11.26$\pm$0.1 & Sa & 10.0$\pm$0.02 & SF & 4.41$\pm$0.11 & 4.26$\pm$0.39 & 7.14$\pm$0.11 & 7.85$\pm$0.39 \\
IC1151 & 30.8 & 10.02$\pm$0.1 & SBc & 7.93$\pm$0.14 & ... & 2.66$\pm$0.07 & 2.76$\pm$0.78 & 4.06$\pm$0.07 & 4.61$\pm$0.78 \\
IC1199 & 68.3 & 10.78$\pm$0.1 & Sbc & 9.35$\pm$0.04 & SF & 4.53$\pm$0.06 & 3.99$\pm$0.44 & 7.52$\pm$0.06 & 6.97$\pm$0.44 \\
IC1683 & 69.7 & 10.76$\pm$0.11 & Sb & 9.68$\pm$0.02 & SF & 5.56$\pm$0.79 & 2.49$\pm$0.13 & 8.92$\pm$0.79 & 3.99$\pm$0.13 \\
IC4566 & 80.7 & 10.76$\pm$0.11 & SABb & 9.68$\pm$0.02 & ... & $<$3.55 & $<$4.4 & $<$5.97 & $<$8.35 \\
NGC0447 & 79.7 & 11.43$\pm$0.1 & S0-a & 9.33$\pm$0.05 & ... & 4.56$\pm$0.7 & 10.0$\pm$0.23 & 6.58$\pm$0.7 & 9.09$\pm$0.23 \\
NGC0477 & 85.4 & 10.9$\pm$0.12 & Sc & 9.54$\pm$0.05 & SF & $<$9.01 & $<$21.78 & $<$14.43 & $<$35.98 \\
NGC0496 & 87.5 & 10.85$\pm$0.13 & Sbc & 9.48$\pm$0.04 & SF & 7.35$\pm$0.34 & 4.11$\pm$0.36 & 12.46$\pm$0.34 & 7.13$\pm$0.36 \\
NGC0528 & 68.8 & 11.06$\pm$0.1 & S0 & 8.36$\pm$0.13 & ... & ... & ... & ... & ... \\
NGC0551 & 74.5 & 10.95$\pm$0.11 & SBbc & 9.39$\pm$0.04 & ... & 4.73$\pm$0.07 & 8.17$\pm$1.68 & 8.01$\pm$0.07 & 13.47$\pm$1.68 \\
NGC1167 & 70.9 & 11.48$\pm$0.09 & S0 & 9.28$\pm$0.06 & LINER & ... & ... & ... & ... \\
NGC2253 & 51.2 & 10.81$\pm$0.11 & Sc & 9.62$\pm$0.02 & SF & 2.48$\pm$0.07 & 2.1$\pm$0.28 & 3.77$\pm$0.07 & 4.07$\pm$0.28 \\
NGC2347 & 63.7 & 11.04$\pm$0.1 & Sb & 9.56$\pm$0.02 & LINER & 2.15$\pm$0.06 & 1.99$\pm$0.38 & 3.86$\pm$0.06 & 4.5$\pm$0.38 \\
NGC2486 & 67.5 & 10.79$\pm$0.09 & Sa & $<$9.05 & ... & ... & ... & ... & ... \\
NGC2487 & 70.5 & 11.06$\pm$0.1 & Sb & 9.47$\pm$0.05 & ... & $<$4.23 & $<$16.66 & $<$5.88 & $<$24.85 \\
NGC2639 & 45.7 & 11.17$\pm$0.09 & Sa & 9.36$\pm$0.02 & LINER & 1.78$\pm$0.01 & 2.88$\pm$0.74 & 2.93$\pm$0.01 & 4.29$\pm$0.74 \\
NGC2730 & 54.8 & 10.13$\pm$0.09 & Sd & 9.0$\pm$0.06 & ... & 5.62$\pm$0.62 & 3.79$\pm$0.25 & 9.57$\pm$0.62 & 6.25$\pm$0.25 \\
NGC2880 & 22.7 & 10.56$\pm$0.08 & E-S0 & $<$7.93 & ... & ... & ... & ... & ... \\
NGC2906 & 37.7 & 10.59$\pm$0.09 & Sc & 9.11$\pm$0.03 & INDEF & 1.72$\pm$0.08 & 1.59$\pm$0.4 & 2.71$\pm$0.08 & 3.0$\pm$0.4 \\
NGC2916 & 53.2 & 10.96$\pm$0.08 & Sb & 9.05$\pm$0.06 & AGN & ... & ... & ... & ... \\
NGC3303 & 89.8 & 11.17$\pm$0.1 & Sa & 9.57$\pm$0.04 & LINER & 3.62$\pm$0.23 & 1.99$\pm$0.11 & 4.97$\pm$0.23 & 3.47$\pm$0.11 \\
NGC3381 & 23.4 & 9.88$\pm$0.09 & SBb & 8.11$\pm$0.08 & ... & ... & ... & ... & ... \\
NGC3687 & 36.0 & 10.51$\pm$0.11 & Sbc & $<$8.42 & ... & $<$1.86 & $<$39.56 & $<$2.63 & $<$66.35 \\
NGC3811 & 44.3 & 10.64$\pm$0.11 & SBc & 9.28$\pm$0.03 & ... & 2.36$\pm$0.09 & 2.18$\pm$0.26 & 2.93$\pm$0.09 & 2.96$\pm$0.26 \\
NGC3815 & 53.6 & 10.53$\pm$0.09 & Sab & 9.16$\pm$0.04 & ... & 2.0$\pm$0.16 & 1.68$\pm$0.27 & 3.05$\pm$0.16 & 3.4$\pm$0.27 \\
NGC3994 & 44.7 & 10.59$\pm$0.11 & Sc & 9.26$\pm$0.03 & ... & 1.09$\pm$0.04 & 1.31$\pm$0.08 & 1.78$\pm$0.04 & 2.23$\pm$0.08 \\
NGC4047 & 49.1 & 10.87$\pm$0.1 & Sb & 9.66$\pm$0.02 & SF & 2.37$\pm$0.02 & 1.26$\pm$0.25 & 3.9$\pm$0.02 & 3.11$\pm$0.25 \\
NGC4185 & 55.9 & 10.86$\pm$0.11 & SBbc & 9.08$\pm$0.07 & INDEF & 4.98$\pm$0.23 & 4.45$\pm$0.85 & 8.19$\pm$0.23 & 7.49$\pm$0.85 \\
NGC4210 & 38.8 & 10.51$\pm$0.1 & Sb & 8.86$\pm$0.05 & LINER & ... & ... & ... & ... \\
NGC4211NED02 & 96.9 & 10.53$\pm$0.13 & S0-a & 9.29$\pm$0.06 & ... & $<$6.65 & $<$10.52 & $<$8.93 & $<$17.82 \\
NGC4470 & 33.4 & 10.23$\pm$0.09 & Sa & 8.59$\pm$0.06 & SF & 1.73$\pm$0.05 & 1.25$\pm$0.29 & 3.04$\pm$0.05 & 2.42$\pm$0.29 \\
NGC4644 & 71.6 & 10.68$\pm$0.11 & Sb & 9.2$\pm$0.05 & ... & 2.7$\pm$0.05 & 3.15$\pm$0.8 & 4.91$\pm$0.05 & 5.88$\pm$0.8 \\
NGC4676A & 96.6 & 10.86$\pm$0.1 & S0-a & 9.88$\pm$0.02 & SF & ... & ... & ... & ... \\
NGC4711 & 58.8 & 10.58$\pm$0.09 & SBb & 9.18$\pm$0.05 & SF & 2.83$\pm$0.06 & 6.24$\pm$0.54 & 4.86$\pm$0.06 & 10.44$\pm$0.54 \\
NGC4961 & 36.6 & 9.98$\pm$0.1 & SBc & 8.41$\pm$0.08 & ... & 1.39$\pm$0.08 & 1.59$\pm$0.33 & 2.1$\pm$0.08 & 2.67$\pm$0.33 \\
NGC5000 & 80.8 & 10.94$\pm$0.1 & Sbc & 9.45$\pm$0.04 & SF & 5.16$\pm$0.61 & 1.06$\pm$0.26 & 6.51$\pm$0.61 & 2.31$\pm$0.26 \\
NGC5016 & 36.9 & 10.47$\pm$0.09 & SABb & 8.9$\pm$0.04 & ... & 1.67$\pm$0.02 & 2.32$\pm$0.42 & 2.89$\pm$0.02 & 3.93$\pm$0.42 \\
NGC5056 & 81.1 & 10.85$\pm$0.09 & Sc & 9.45$\pm$0.04 & ... & 4.22$\pm$0.51 & 3.12$\pm$0.48 & 5.44$\pm$0.51 & 5.99$\pm$0.48 \\
NGC5205 & 25.1 & 9.98$\pm$0.09 & Sbc & 8.37$\pm$0.07 & LINER & $<$1.57 & $<$2.13 & $<$2.35 & $<$3.61 \\
NGC5218 & 41.7 & 10.64$\pm$0.09 & SBb & 9.86$\pm$0.01 & ... & 1.65$\pm$0.08 & 1.43$\pm$0.18 & 2.79$\pm$0.08 & 1.93$\pm$0.18 \\
NGC5394 & 49.5 & 10.38$\pm$0.11 & SBb & 9.62$\pm$0.01 & SF & 2.18$\pm$0.27 & 2.7$\pm$0.21 & 3.36$\pm$0.27 & 4.43$\pm$0.21 \\
NGC5406 & 77.8 & 11.27$\pm$0.09 & Sbc & 9.69$\pm$0.04 & LINER & 4.97$\pm$0.26 & 7.54$\pm$1.9 & 7.23$\pm$0.26 & 12.77$\pm$1.9 \\
NGC5480 & 27.0 & 10.18$\pm$0.08 & Sc & 8.92$\pm$0.03 & LINER & 2.41$\pm$0.1 & 1.27$\pm$0.2 & 4.04$\pm$0.1 & 2.35$\pm$0.2 \\
NGC5485 & 26.9 & 10.75$\pm$0.08 & S0 & $<$8.09 & LINER & ... & ... & ... & ... \\
NGC5520 & 26.7 & 10.07$\pm$0.11 & Sb & 8.67$\pm$0.03 & ... & 1.19$\pm$0.07 & 0.9$\pm$0.11 & 1.67$\pm$0.07 & 1.77$\pm$0.11 \\
NGC5614 & 55.7 & 11.22$\pm$0.09 & Sab & 9.84$\pm$0.01 & ... & 2.25$\pm$0.28 & 1.34$\pm$0.16 & 3.67$\pm$0.28 & 3.1$\pm$0.16 \\
NGC5633 & 33.4 & 10.4$\pm$0.11 & Sb & 9.14$\pm$0.02 & SF & 1.36$\pm$0.03 & 1.4$\pm$0.26 & 2.47$\pm$0.03 & 2.61$\pm$0.26 \\
NGC5657 & 56.3 & 10.5$\pm$0.1 & Sb & 9.11$\pm$0.04 & ... & 2.11$\pm$0.07 & 1.88$\pm$0.13 & 3.63$\pm$0.07 & 3.37$\pm$0.13 \\
NGC5682 & 32.6 & 9.59$\pm$0.11 & Sb & $<$8.29 & SF & 2.11$\pm$0.05 & 1.47$\pm$0.34 & 3.57$\pm$0.05 & 2.21$\pm$0.34 \\
NGC5732 & 54.0 & 10.23$\pm$0.11 & Sbc & 8.82$\pm$0.07 & SF & 2.42$\pm$0.09 & 1.78$\pm$0.11 & 3.92$\pm$0.09 & 3.34$\pm$0.11 \\
NGC5784 & 79.4 & 0.0$\pm$0.0 & S0 & 9.4$\pm$0.04 & ... & 2.4$\pm$0.32 & 1.41$\pm$0.13 & 3.28$\pm$0.32 & 3.46$\pm$0.13 \\
NGC5876 & 46.9 & 10.78$\pm$0.1 & SBab & $<$8.56 & ... & ... & ... & ... & ... \\
NGC5908 & 47.1 & 10.95$\pm$0.1 & Sb & 9.94$\pm$0.01 & ... & 2.92$\pm$0.01 & 1.73$\pm$0.34 & 4.98$\pm$0.01 & 4.55$\pm$0.34 \\
NGC5930 & 37.2 & 10.61$\pm$0.11 & SABa & 9.33$\pm$0.02 & ... & 1.57$\pm$0.07 & 0.82$\pm$0.03 & 2.66$\pm$0.07 & 1.98$\pm$0.03 \\
NGC5934 & 82.7 & 10.87$\pm$0.09 & Sa & 9.81$\pm$0.02 & ... & 3.07$\pm$0.18 & 2.5$\pm$0.17 & 5.17$\pm$0.18 & 4.36$\pm$0.17 \\
NGC5947 & 86.1 & 10.87$\pm$0.1 & SBbc & 9.26$\pm$0.06 & AGN & $<$4.15 & $<$4.7 & $<$5.25 & $<$7.83 \\
NGC5953 & 28.4 & 10.38$\pm$0.11 & S0-a & 9.49$\pm$0.01 & ... & 1.16$\pm$0.17 & 0.5$\pm$0.07 & 1.3$\pm$0.17 & 1.23$\pm$0.07 \\
NGC6004 & 55.2 & 10.87$\pm$0.08 & Sc & 9.33$\pm$0.04 & ... & 5.29$\pm$0.23 & 2.82$\pm$0.21 & 8.18$\pm$0.23 & 4.52$\pm$0.21 \\
NGC6027 & 62.9 & 11.02$\pm$0.1 & S0-a0 & 8.01$\pm$0.22 & ... & ... & ... & ... & ... \\
NGC6060 & 63.2 & 10.99$\pm$0.09 & SABc & 9.68$\pm$0.03 & SF & 3.85$\pm$0.11 & 4.08$\pm$0.52 & 6.25$\pm$0.11 & 7.59$\pm$0.52 \\
NGC6063 & 40.7 & 10.36$\pm$0.12 & Sc & $<$8.53 & SF & 2.65$\pm$0.08 & 3.34$\pm$0.83 & 4.67$\pm$0.08 & 5.99$\pm$0.83 \\
NGC6125 & 68.0 & 11.36$\pm$0.09 & E & $<$8.83 & ... & ... & ... & ... & ... \\
NGC6146 & 128.7 & 11.72$\pm$0.09 & E & $<$9.36 & ... & ... & ... & ... & ... \\
NGC6155 & 34.6 & 10.38$\pm$0.1 & Sc & 8.94$\pm$0.03 & SF & 2.03$\pm$0.06 & 1.97$\pm$0.27 & 3.32$\pm$0.06 & 3.5$\pm$0.27 \\
NGC6186 & 42.4 & 10.62$\pm$0.09 & Sa & 9.46$\pm$0.02 & ... & 9.48$\pm$0.45 & 5.9$\pm$0.28 & 14.74$\pm$0.45 & 10.39$\pm$0.28 \\
NGC6301 & 121.4 & 11.18$\pm$0.12 & Sc & 9.96$\pm$0.03 & INDEF & 9.45$\pm$0.39 & 13.32$\pm$3.83 & 15.5$\pm$0.39 & 20.01$\pm$3.83 \\
NGC6314 & 95.9 & 11.21$\pm$0.09 & Sa & 9.57$\pm$0.03 & INDEF & 6.6$\pm$0.57 & 2.41$\pm$0.07 & 7.43$\pm$0.57 & 4.56$\pm$0.07 \\
NGC6394 & 124.3 & 11.11$\pm$0.1 & SBb & 9.86$\pm$0.04 & AGN & 5.0$\pm$0.34 & 5.13$\pm$0.63 & 9.02$\pm$0.34 & 9.25$\pm$0.63 \\
NGC7738 & 97.8 & 11.21$\pm$0.11 & Sb & 9.99$\pm$0.01 & LINER & 2.42$\pm$0.2 & 1.95$\pm$0.02 & 3.95$\pm$0.2 & 3.81$\pm$0.02 \\
NGC7819 & 71.6 & 10.61$\pm$0.09 & Sb & 9.27$\pm$0.04 & SF & 6.91$\pm$1.06 & 2.6$\pm$0.69 & 9.71$\pm$1.06 & 3.15$\pm$0.69 \\
UGC03253 & 59.5 & 10.63$\pm$0.11 & Sb & 8.88$\pm$0.06 & SF & 5.15$\pm$1.22 & 2.91$\pm$0.72 & 5.74$\pm$1.22 & 4.83$\pm$0.72 \\
UGC03973 & 95.9 & 10.94$\pm$0.08 & Sb & 9.51$\pm$0.05 & AGN & 3.78$\pm$0.38 & 2.91$\pm$0.11 & 5.3$\pm$0.38 & 6.14$\pm$0.11 \\
UGC05108 & 118.4 & 11.11$\pm$0.11 & SBab & 9.75$\pm$0.04 & ... & 4.55$\pm$0.16 & 5.08$\pm$0.78 & 7.56$\pm$0.16 & 7.31$\pm$0.78 \\
UGC05359 & 123.2 & 10.86$\pm$0.13 & SABb & 9.65$\pm$0.05 & SF & 5.25$\pm$0.15 & 6.15$\pm$1.09 & 8.95$\pm$0.15 & 11.06$\pm$1.09 \\
UGC06312 & 90.0 & 10.93$\pm$0.12 & Sa & $<$9.08 & ... & 3.29$\pm$0.07 & 4.81$\pm$0.44 & 5.41$\pm$0.07 & 8.73$\pm$0.44 \\
UGC07012 & 44.3 & 11.0$\pm$2.9 & SBc & 9.9$\pm$0.11 & SF & 2.31$\pm$0.1 & 0.99$\pm$0.17 & 3.36$\pm$0.1 & 2.0$\pm$0.17 \\
UGC09067 & 114.5 & 10.96$\pm$0.12 & Sab & 9.83$\pm$0.04 & SF & 3.39$\pm$0.06 & 3.56$\pm$0.34 & 6.11$\pm$0.06 & 6.8$\pm$0.34 \\
UGC09476 & 46.6 & 10.43$\pm$0.11 & SABc & 9.15$\pm$0.04 & SF & $<$3.85 & $<$5.37 & $<$5.95 & $<$9.46 \\
UGC09759 & 49.2 & 10.02$\pm$0.1 & Sb & 9.07$\pm$0.04 & ... & 2.83$\pm$0.16 & 1.09$\pm$0.17 & 4.5$\pm$0.16 & 1.96$\pm$0.17 \\
UGC10205 & 94.9 & 11.08$\pm$0.1 & Sa & 9.6$\pm$0.04 & SF & 5.41$\pm$0.63 & 2.57$\pm$0.39 & 6.09$\pm$0.63 & 5.12$\pm$0.39 \\
\hline
\end{tabular}}
\caption{Main properties of the 81 EDGE-CALIFA galaxies analyzed in this work. The columns Distance, $M_{\star}$, Morphological Class, and $M_{\rm mol}$ are taken from \cite{Bolatto2017}, where $M_{\rm mol}$ is computed using $\alpha_{\rm CO}=4.36$ M$_\odot$ (K km s$^{-1}$ pc$^{2}$)$^{-1}$. The column Nuclear corresponds to the emission-line diagnostic for the optical nucleus spectrum for CALIFA galaxies by \cite{Garcia-Lorenzo2015}, who classify the galaxies (with signal-to-noise larger than three) into star forming (SF), active galactic nuclei (AGN), and LINER-type galaxies. The columns $l_{\star}$, $l_{\rm mol}$, $R_{50,\star}$, and $R_{50,\rm mol}$ are the exponential scale lengths and the radii that enclose 50\% of the molecular gas and the stellar mass computed in Section \ref{Exponential_scale}, respectively.}
\label{Table}
\end{table*}
\subsubsection{SFE and Galactocentric Radius}
\label{SFE_radius}
Figure \ref{FIG_1_Leroy_2008_1A} shows the relation between SFE$_{\rm gas}$ and galactocentric radius; the four different panels show grouping of the 81 galaxies. Following modern studies, we use $R_{\rm e}$ to normalize galactocentric distances, except when we need to compare to published data which use $r_{25}$. Note that for the EDGE galaxies in this sample, $r_{25} \approx 2.1 R_{\rm e}$. In this figure for clarity we split the Sbc, Sc, and Scd galaxies in two groups by choosing the median of stellar masses of the EDGE-CALIFA sample $\log_{10}$[$M_{\star}$]$=10.7$ \citep{Bolatto2017}. In general, there is a decreasing trend for SFE$_{\rm gas}$ with radius. It is important to note that SFE$_{\rm gas}$ is a fairly smooth function of radius for a given galaxy. In fact, variations between galaxies are frequently larger than variations between most annuli in a galaxy, indicating that the radial decrease in SFE$_{\rm gas}$ within in a galaxy is often smooth and that galaxy to galaxy variations are significant.
Figure \ref{FIG_1_Leroy2008_2} shows the radius at which our measured molecular surface density, averaged over an annulus, is the same as our assumed constant surface density in the atomic disk, $\Sigma_{\rm mol}=\Sigma_{\rm atom}=6$~M$_\odot$\,pc$^{-2}$. The typical radius at which this happens is $r/R_{\rm e} \sim [1.1\pm 0.5]$, or $r/r_{25}\sim[0.47\pm0.28]$ (see inset panel), which agrees with the value of $r/r_{25}\sim0.43 \pm 0.18$ found by \citet{Leroy2008}. Note that in Figure \ref{FIG_1_Leroy_2008_1A} the SFE$_{\rm gas}$ is generally smooth across that radius, suggesting that our assumption of a constant $\Sigma_{\rm atom}$ does not play a major role at determining the shape of the total gas SFE$_{\rm gas}$.
Figure \ref{FIG_1_Leroy2001B} shows the average SFE$_{\rm gas}$ as a function of the normalized galactocentric radius for each of the four different groups of morphological classification used in Figure \ref{FIG_1_Leroy_2008_1A}, with $\pm 1 \sigma$ variation indicated by the color bands. We note a systematic increase in the average SFE$_{\rm gas}$ from early type (red shaded area) to late type galaxies (blue shaded area). The SFE$_{\rm gas}$ tend to be lower for the early spirals (i.e., S0 and earlier; ten galaxies), which have a steeper profile when compared with the rest of the morphological groups, and therefore showing a significant anticorrelation between SFE$_{\rm gas}$ and $r_{\rm gal}$ (Pearson correlation coefficient of r$=-0.6$). This steepening may reflect the degree of central concentration seen in earlier-type galaxies. Sd-Ir galaxies show a SFE$_{\rm gas}$ flattening at $r_{\rm gal}<0.45\,r_{25}$; however, their small amount (only 2 galaxies in our sample) does not allow to conclude that this flattening is statistically significant. When looking at the average SFE$_{\rm gas}$ value, over $r_{\rm gal}$ for all the radial profiles (black-circular dots), we find that the SFE$_{\rm gas}$ decreases exponentially even in regions where the gas is mostly molecular. In EDGE we see an continuous exponential profile for the SFE$_{\rm gas}$ averaged over all galaxies (black line in Figure \ref{FIG_1_Leroy2001B}). Although still within the error bars, this is in contrast to HERACLES, which sees a leveling of the SFE$_{\rm gas}$ in the inner regions. The greater range of SFE$_{\rm gas}$s in our sample may be a reflection of the larger range of galaxy spiral types spanned by EDGE compared to HERACLES, which consisted mostly of late types. In fact the Sbc, Sc, and Scd galaxies in EDGE-CALIFA (green band) are very consistent with the measurements of HERACLES. Where the gas is dominated by the atomic component, $r \gtrsim 0.4 r_{25}$, the SFE$_{\rm gas}$ decreases rapidly to the galaxy edge. Because we assume a constant $\Sigma_{\rm atom}$, this is fundamentally a reflection of the rapid decrease of SFR in the atomic disks.
We can describe the behaviour of the SFE$_{\rm gas}$ for our sample using an ordinary least-square (OLS) linear bisector method to fit a simple exponential decay:
\begin{equation}
{\rm SFE_{\rm gas}} = [0.83\pm 0.07\, {\rm (Gyr^{-1}})] \exp \left ( \frac{-r_{\rm gal}}{[0.31\pm 0.02]\,r_{25}}\right ).
\label{SFE_radius_fit}
\end{equation}
We note that we do not see clear breaks in this trend; instead, we find a continuous smooth exponential decline of SFE$_{\rm gas}$ as a function of $r_{\rm gal}$. This is consistent with the rapid decline of star formation activity in the outer parts of galaxies \citep[e.g., ][]{Leroy2008,Kennicutt1989,MartinKennicutt2001}, and also is in agreement with previous results for low-redshift star-forming galaxies \citep[e.g., ][]{Sanchez_1_2020,Sanchez_2_2020}. In particular, our results agree with the inside out monotonic decrease of the SFE$_{\rm gas}$ shown by \cite{Sanchez_1_2020}. \cite{Sanchez_1_2020} also find that galaxies are segregated by morphology; for a given stellar mass, they show that late-type galaxies present larger SFE$_{\rm gas}$ than earlier ones at any $r_{\rm gal}$, which is consistent with the trend we observe in Figure \ref{FIG_1_Leroy2001B}. In the outer parts, our steeper profiles may be influenced by our assumption of constant HI surface density. However, this does not explain our steeper profiles we also observe in the inner galaxy.
The top and bottom dashed lines in Fig. \ref{FIG_1_Leroy2001B} show how SFE$_{\rm gas}$ changes if instead of 6~M$_\odot$\,pc$^{-2}$ we use $\Sigma_{\rm atom}=3$ and $12$~M$_\odot$\,pc$^{-2}$, which are the two extremes of $\Sigma_{\rm atom}$ values found in HERACLES \citep[][]{Leroy2008}. A better match between EDGE and HERACLES would require using $\Sigma_{\rm atom}=3$~M$_\odot$\,pc$^{-2}$, which appears extremely low. Note that these two studies use different SFR tracers: our extinction-corrected H$\alpha$, may behave differently from the GALEX FUV that dominates the SFR estimate in the outer disks of HERACLES \citep[e.g.,][]{Lee2009}.
How sensitive is the SFE$_{\rm gas}$ determination to the CO-to-H$_2$ conversion factor? To test this we adopt a variable CO-to-H$_2$ conversion factor, $\alpha_{\rm CO}$, using equation \ref{Alpha_CO}. This includes changes in the central regions caused by high stellar surface densities, and changes due to metallicity. When comparing the effects of a constant and a variable prescription of $\alpha_{\rm CO}$ (shaded area in Figure \ref{FIG_1_Leroy2001B}) we observe that the central regions present larger SFE$_{\rm gas}$ variations than the outer disks within the range of galactocentric distances we study, as the latter do not exhibit $12+\log{\rm (O/H)}$ significantly below 8.4 according to the O3N2 indicator, as shown in the top panel of Figure \ref{FIG_1_Metallicity}. Therefore, the variations of the CO-to-H$_2$ conversion factor are generally small and consistent with the assumption of a constant $\alpha_{\rm CO}$.
So far, we have analyzed the SFE of the total gas, but it is also interesting to test whether the star formation efficiency responds to the phase of the ISM. The bottom panel of Figure \ref{FIG_1_Metallicity} shows the star formation efficiency of the molecular gas, SFE$_{\rm mol} = \Sigma_{\rm SFR}/\Sigma_{\rm mol}$ (in yr$^{-1}$), as a function of the ratio between the molecular and the atomic surface densities, $R_{\rm mol} = \Sigma_{\rm mol}/\Sigma_{\rm atom}$. Since we assume $\Sigma_{\rm atom}=6$ M$_\odot$ pc$^{-2}$, $R_{\rm mol}$ is a prescription for the $\Sigma_{\rm mol}$ normalized by a factor of 6. Although there is large scatter, the figure shows that the SFE$_{\rm mol}$, averaged by $R_{\rm mol}$ bins (filled-black dots), remains almost constant over the $R_{\rm mol}$ range, with an average log[SFE$_{\rm mol}$] $\sim -9.15$ (blue-dashed line in bottom panel of Fig. \ref{FIG_1_Metallicity}). The inset panel shows that the SFE$_{\rm mol}$ is also fairly constant over the range of galactocentric radii. These results are in agreement with \cite{Kazuyuki2019}, who find a similar flattening in SFE$_{\rm mol}$ for annuli at $r \lesssim0.6r_{25}$ when analyzing 80 nearby-spiral galaxies selected from the CO Multi-line Imaging of Nearby Galaxies survey \citep[COMING;][]{Sorai2019}. Using CO, FUV+24$\mu$m and H$_\alpha$+24$\mu$m data for 33 nearby-spiral galaxies selected from the IRAM HERACLES survey \citep{Leroy2009}, \cite{Schruba2011} found that H$_2$-dominated regions are well parameterized by a fixed SFE$_{\rm mol}$ equivalent to a molecular gas depletion time of $\tau_{\rm dep,mol}={\rm SFE^{-1}_{mol}}\sim 1.4$ Gyr, which is consistent with our average $\tau_{\rm dep,mol} \sim 1.45\pm 0.23$ Gyr. As for previous studies, these results support the idea that the vast majority of the star formation activity takes place in the molecular phase of the ISM instead of the atomic gas \citep[e.g.,][]{MartinKennicutt2001,Bigiel2008,Schruba2011}.
We explore possible trends between SFE$_{\rm gas}$, galactocentric radius, and nuclear activity. We adopt the nuclear activity classification performed by \cite{Garcia-Lorenzo2015}, who classify CALIFA galaxies (with signal-to-noise larger than three) into star forming (SF), active galactic nuclei (AGN), and LINER-type galaxies, and we apply it, when available, for the 81 galaxies analyzed in this work (see column Nuclear in Table \ref{Table}). We do not identify significant trends as a function of galactocentric radius for any of these three categories.
\subsubsection{SFE versus Stellar and Gas Surface Density}
\label{SFE_SSD}
Since in the previous section we show a clear dependence of SFE$_{\rm gas}$ on galactocentric distance, it is expected that SFE$_{\rm gas}$ will also depend on the stellar surface density, $\Sigma_\star$. Indeed, the top panel of Figure \ref{FIG_3_Leroy2008} shows an approximately power-law relationship between SFE$_{\rm gas}$ and $\Sigma_{\star}$. We quantify this relation by using an OLS linear bisector method in logarithmic space to estimate the best linear fit to our data (excluding upper-limits), obtaining
\begin{multline}
\log[{\rm SFE_{\rm gas}}\,({\rm yr^{-1}})] = [0.32\pm 0.27]\times \log[\Sigma_\star \, ({\rm M_\odot \, pc^{-2}})]\\
-[10.13\pm1.75].
\label{OLS_Stellar_surface_density}
\end{multline}
When comparing the EDGE average SFE$_{\rm gas}$, over $\Sigma_\star$ bins (black dots) with similar HERACLES bins (green squares), we find consistently slightly larger efficiencies at $\log[\Sigma_\star (\rm M_\odot\,pc^{-2})]\lesssim 1.4$ although the HERACLES points are still within the error bars of our data. Since these points are in the outer regions of the EDGE galaxies, this result may sensitive to the adoption of $\Sigma_{\rm atom}=6$~M$_\odot$\,pc$^{-2}$. In the inner regions with $\log[\Sigma_\star (\rm M_\odot\,pc^{-2})]\geq 2.6$, our average efficiencies are also higher, although we do not expect these regions to be sensitive to the choice of $\Sigma_{\rm atom}$. Between these two extremes, however, there is good general agreement between the EDGE and HERACLES results.
The middle and bottom panels of Figure \ref{FIG_3_Leroy2008} show the relation between the H$_2$-to-H$_{\rm I}$ ratio ($R_{\rm mol}=\Sigma_{\rm mol}/\Sigma_{\rm atom} = \Sigma_{\rm mol}/6$ M$_{\odot}$ pc$^{-2}$), $\Sigma_{\rm *}$, and the gas surface density, $\Sigma_{\rm gas}=\Sigma_{\rm mol}+\Sigma_{\rm atom}$, respectively. In the middle panel, we observe a tight correlation between $R_{\rm mol}$ and $\Sigma_{\star}$. The relation is well described by a power-law and there is overall reasonable consistency between EDGE and HERACLES. Our measurements are also consistent with the resolved Molecular Gas Main Sequence relation (rMGMS, $\Sigma_{\rm gas}$-$\Sigma_{\star}$; \citealt{Lin2019}) found for EDGE-CALIFA galaxies by \cite{Barrera-Ballesteros2021}. The bottom panel shows very good agreement between the EDGE and HERACLES results in the range $0.9 \lesssim \log[\Sigma_{\rm gas}] \lesssim 1.5$; outside this range there are small differences, although there is still consistency within the error bars. Therefore, the discrepancies seen in the top panel are not the result of differences in efficiency at a given H$_2$-to-H$_{\rm I}$ ratio nor gas surface density, but likely reflect small systematic differences in the relation between gas and stellar surface density in HERACLES and EDGE. Since we have both a broader morphological and a more numerous sample selection than HERACLES (particularly in the H$_{\rm I}$-dominated regions), our results reflect on a more general power-law dependence of the SFE$_{\rm gas}$ on $\Sigma_{\star}$.
Observations have shown that the fraction of gas in the molecular phase in which star formation takes place depends on the pressure in the medium \citep[][]{Elmegreen1993,Blitz&Rosolowsky2006}. These results suggest that high stellar densities in the inner regions of EDGE-CALIFA galaxies are helping self-gravity to compress the gas, resulting in H$_2$ dominated regions. Once the gas is predominantly molecular, our data suggests that a dependence of the SFE$_{\rm gas}$ on $\Sigma_\star$ persists even in high $\Sigma_\star$, predominantly molecular regions.
\begin{figure}
\begin{tabular}{c}
\hspace{-.5cm}\vspace{-.35cm}\includegraphics[width=8.7cm]{Leroy2008_fig3A.pdf} \\
\hspace{-.5cm}\vspace{-.35cm}\includegraphics[width=8.3cm]{SFE_vs_Rmol.pdf} \\
\hspace{-.9cm}\includegraphics[width=8.7cm]{Leroy2008_fig3A_gas.pdf}
\end{tabular}
\vspace{-0.5cm}
\caption{{\it Top:} SFE$_{\rm gas}$ as a function of stellar surface density, $\Sigma_{\star}$. Circular dots indicate the average SFE$_{\rm gas}$ and $\Sigma_{\star}$ in stacked annuli for the EDGE-CALIFA data. The `r' term represents the Pearson correlation coefficient, including the binned annuli, for the relation between the SFE$_{\rm gas}$ and $\Sigma_{\star}$. Conventions are as in bottom panel of Figure \ref{FIG_1_Metallicity}. {\it Middle:} The H$_2$-to-H$_{\rm I}$ ratio, $R_{\rm mol}$, as a function of $\Sigma_{\rm \star}$; we use $\Sigma_{\rm atom}= 6$~M$_\odot$\,pc$^{-2}$. Conventions are as in top panel. {\it Bottom:} SFE$_{\rm gas}$ as a function of gas surface density. The grey dashed line marks the point at which $\Sigma_{\rm gas}=\Sigma_{\rm atom}= 6$~M$_\odot$\,pc$^{-2}$. Conventions are as in top panel.}
\label{FIG_3_Leroy2008}
\end{figure}
Other studies have given different insights of the relation between star formation activity and the stellar surface density. For instance, analyzing 34 galaxies selected from the ALMA-MaNGA Quenching and STar formation \citep[ALMaQUEST;][]{Lin2019}, \cite{Ellison2020} find that $\Sigma_{\rm SFR}$ is mainly regulated by $\Sigma_{\rm mol}$, with a secondary dependence on $\Sigma_{\star}$. Conversely, analyzing 39 galaxies selected from EDGE-CALIFA, \cite{Dey2019} find a strong correlation between $\Sigma_{\rm SFR}$ and $\Sigma_{\rm mol}$; they show that the $\Sigma_{\rm SFR}-\Sigma_{\star}$ relation is statistically more significant. \cite{Sanchez2021}, however, used the {\tt{edge}\_\tt{pydb}} database to show
that secondary correlations can be driven purely by errors in correlated parameters, and it is necessary to be particularly careful when studying these effects. Errors in $\Sigma_{\rm gas}$, for example, will tend to flatten the relation between SFE$_{\rm gas}$ and $\Sigma_{\rm gas}$ because of the intrinsic correlations between the axes, and will have the same effect on the relation between SFE$_{\rm gas}$ and $\Sigma_*$ because of the positive correlation between $\Sigma_*$ and $\Sigma_{\rm gas}$.
\subsubsection{SFE, Pressure and SFR}
\label{SFE_Midplane_Gas_Pressure}
We explore the dependency of SFE$_{\rm gas}$ on the dynamical equilibrium pressure, $P_{\rm DE}$. While the midplane gas pressure, $P_{\rm h}$ \citep{Elmegreen1989}, is a well studied pressure prescription in a range of previous works \citep[e.g., ][]{Elmegreen1993,Leroy2008}, $P_{\rm DE}$ has been extensively discussed recently (e.g., \citealt{Kim2013,HerreraCamus2017,Sun2020,Barrera-Ballesteros2021}). In both pressure prescriptions, it is assumed that the gas disk scale height is much smaller than the stellar disk scale height and the gravitational influence from dark matter is neglected. $P_{\rm h}$ and $P_{\rm DE}$ have an almost equivalent formulation, although they slightly differ in the term related to the gravitational influence from the stellar component (second term in equations \ref{Midplane_gas_pressure} and \ref{Dynamical_equilibrium_pressure}; see section \ref{Basic_equations}). We quantify this difference by computing the mean $P_{\rm DE}$-to-$P_{\rm h}$ ratio averaged in annuli for our sample, obtaining $P_{\rm DE}/P_{\rm h}\approx 1.51\pm 0.19$. We use this value to convert the $P_{\rm h}$ from HERACLES into $P_{\rm DE}$, since we perform our qualitative analysis using the dynamical equilibrium pressure.
\begin{figure}
\begin{tabular}{c}
\hspace{-1.2cm}\includegraphics[width=9.5cm]{Leroy2008_fig6A_PDE_Ph.pdf} \\
\hspace{-1.2cm}\includegraphics[width=8.8cm]{Leroy2008_fig6A_PDE.pdf}
\end{tabular}
\caption{{\it Top:} SFE$_{\rm gas}$ as a function of dynamical equilibrium pressure, $P_{\rm DE}$. The dashed-blue line corresponds to 1\% of gas converted to stars per disk free-fall time. {\it Bottom:} Star formation rate surface density, $\Sigma_{\rm SFR}$, as a function of $P_{\rm DE}$. The black dashed line is the OLS linear bisector fit for all points. The 'r' term is the Pearson correlation coefficient of this subsample. The shaded region represents uncertainty of the slope derived from the OLS linear bisector fit. Conventions are as in Figure \ref{FIG_3_Leroy2008}.}
\label{FIG_6_Leroy2008}
\end{figure}
The top panel of Figure \ref{FIG_6_Leroy2008} shows the SFE$_{\rm gas}$ as a function of $P_{\rm DE}$ (in units of ${\rm K\, cm^{-3}})$. The slope of the SFE$_{\rm gas}$ vs $P_{\rm DE}$ relation (averaged over $P_{\rm DE}$ bins (black dots) has a break at $\log[P_{\rm DE}]\sim 3.7$. Below $\log[P_{\rm DE}]\lesssim 3.7$ (i.e., where the ISM is H$\rm I$-dominated) we do not see a clear correlation between SFE$_{\rm gas}$ and $P_{\rm DE}$. This is at the sensitivity limit existing data for EDGE, but it is also consistent with the overall behaviour seen in HERACLES corresponding to a steepening of their mean relation. Above this pressure we find a clear linear trend in log-log space. For higher $P_{\rm DE}$ values (e.g., H$_{\rm 2}$-dominated regions) the EDGE average efficiencies are somewhat higher than observed in HERACLES, which flatten out at high $P_{\rm DE}$) although with a scatter that is within the respective 1$\sigma$ error bars. For $\log[P_{\rm DE}]\gtrsim 3.7$ the EDGE average efficiencies are well described by the blue-dashed line, which corresponds to $1\%$ of the gas converted to stars per disk free-fall time, $\tau_{\rm ff} = (G\rho)^{-1/2}$. To quantify this relation, we use an OLS linear bisector method to estimate the best linear fit to our data, obtaining
\begin{multline}
\log[{\rm SFE_{\rm gas}} \,({\rm yr^{-1}})] = [0.41\pm 0.29] \times \log[P_{\rm DE}/k \,({\rm K cm^{-3}})]\\
- [11.32 \pm 2.24].
\label{OLS_SFE_P_DE}
\end{multline}
The bottom panel of Figure \ref{FIG_6_Leroy2008} shows the $\Sigma_{\rm SFR}$ versus $P_{\rm DE}$, color-coded by galactocentric radius. When compared with other recent measurements (e.g., KINGFISH, \citealt{HerreraCamus2017}; PHANGS, \citealt{Sun2020}), our annuli have the advantage of covering a somewhat wider dynamic range in both $\Sigma_{\rm SFR}$ and $P_{\rm DE}$. We find a strong correlation between $\Sigma_{\rm SFR}$ and $P_{\rm DE}$ that is approximately linear for annuli at $\log[P_{\rm DE}/k] \gtrsim 3.7$, although below this limit we observe a break in the trend. As shown by the color coding of the symbols, indicating $r_{\rm gal}$ in Figure \ref{FIG_6_Leroy2008}, this limit is apparently related to the $r_{\rm gal}$ at which the transition from H$_2$-dominated to HI-dominated annuli happens. This transition may be due to the large range of physical properties covered by our sample, which span from molecular dominated to atomic dominated regimes. Where the ISM weight is higher (e.g., H$_2$-dominated regions), the SFR is stabilized by the increasing feedback from star formation to maintain the pressure that counteracts the $P_{\rm DE}$ \citep{Sun2020}. The lack of correlation we observe at $\log[P_{\rm DE}/k] \lesssim 3.7$ ($r\gtrsim 0.7$) is mainly because we are reaching our CO sensitivity in the HI-dominated regions. To quantify the correlation, we estimate the best linear fit by using an OLS linear bisector method in logarithmic space for annuli at $r\gtrsim 0.7$,
\begin{multline}
\log[\Sigma_{\rm SFR} \,({\rm M_\odot \,yr^{-1}})] = [1.10\pm 0.11] \times \log[P_{\rm DE}/k \, ({\rm K\, cm^{-3}})] \\
- [7.28 \pm 0.65].
\label{OLS_P_DE}
\end{multline}
\noindent Note that these results are potentially sensitive to the method we employ for the fitting. Nonetheless, using an orthogonal distance regression (ODR) to fit the same subsample, we obtain very comparable values $\log[\Sigma_{\rm SFR} \,({\rm M_\odot \,yr^{-1}})] = [1.09\pm 0.05] \times \log[P_{\rm DE}/k \, ({\rm K cm^{-3}})] - [7.25 \pm 0.25]$. \cite{Barrera-Ballesteros2021} analyze 4260 resolved star-forming regions of kpc size located in 96 galaxies from the EDGE-CALIFA survey, using a similar sample selection (e.g., inclination, $\sigma_{\rm gas}$ and $\Sigma_{\rm atom}$ constant values, among others) but they just consider Equivalent Widths for the H$\alpha$ line emission $\rm EW(H\alpha)>20 \AA$. Using an ODR fitting technique, they obtain $\log[\Sigma_{\rm SFR} \,({\rm M_\odot \,yr^{-1}})] = [0.97\pm 0.05] \times \log[P_{\rm DE}/k \, ({\rm K cm^{-3}})] - [7.88 \pm 0.48]$, which is in agreement with the distribution shown in the bottom panel of Figure \ref{FIG_6_Leroy2008}. The figure also shows that the correlation agrees with hydrodynamical simulations performed by \cite{Kim2013} (green dashed line), in which they obtain a slope of $1.13$. These results are also consistent with measurements obtained in other galaxy samples. \cite{Sun2020} obtain a slope of $0.84\pm0.01$ for 28 well-resolved CO galaxies ($\sim {1}''.5$, corresponding to $\sim100$ pc) selected from the ALMA-PHANGS sample by using a methodology very similar to ours. Smaller slopes have been referenced in local very actively star-forming galaxies (e.g., local ultra luminous infra-red galaxies, ULIRGs), which at the same time may resemble some of the conditions in high-redshift sub-millimeter galaxies \citep[e.g.,][]{Ostriker&Shetty2011}. \cite{HerreraCamus2017} analyzed the [CII] emission in atomic-dominated regions of 31 KINGFISH galaxies to determine the thermal pressure of the neutral gas and related it to $P_{\rm DE}$, obtaining a slope of $1.3$ (dotted blue line). Our results bridge these two extremes; the strong correlation between $\Sigma_{\rm SFR}$ and $P_{\rm DE}$ and its linearity supports the idea of a feedback-regulated scenario, in which star formation feedback acts to restore balance in the star-forming region of the disk \citep{Sun2020}.
\subsubsection{SFE and Orbital Timescale}
\label{Orbital_timescale_Section}
In the next two sections, we exclude 21 galaxies (out of the 81) since their H$\alpha$ rotation curves \citep[taken from][]{Levy2018} are either too noisy or not well fitted by the universal rotation curve parametric form. The top panel of Figure \ref{FIG_7_Leroy2008} shows SFE$_{\rm gas}$ versus $\tau_{\rm orb}$, the orbital timescale (in units of yr), color-coded by galactocentric radius. When analyzing our efficiencies averaged over orbital timescale bins (black symbols), we note that there is a slightly flattening of the SFE$_{\rm gas}$ at $\log [\tau_{\rm orb}]\sim 7.9-8.1$. We also note that annuli at $\log [\tau_{\rm orb}]\lesssim 8.1$ are usually within the bulge radius in the SDSS $i$-band (reddish-stars symbols). However, the error bars are consistent with SFE$_{\rm gas}$ decreasing as a function of $\tau_{\rm orb}$ including at $\log[\tau_{\rm orb}] < 8.1$. These results are in agreement with what is found in other spatially resolved galaxy samples \citep[e.g.,][]{WongBlitz2002,Leroy2008}. The average gas depletion time for our subsample, $\tau_{\rm dep}=\Sigma_{\rm gas}/\Sigma_{\rm SFR} \approx 2.8^{1.1}_{-1.0}$ Gyr, which agrees farily with the depletion time $\tau_{\rm dep} = 2.2$ Gyr found for HERACLES \citep[not including early-type galaxies;][]{Leroy2013}. \cite{Utomo2017} computed the depletion times for 52 EDGE-CALIFA galaxies using annuli in the region within 0.7 $r_{25}$ (just considering the molecular gas); their average $\tau_{\rm dep} \approx 2.4$ Gyr is in good agreement with our results.
The orbital timescale has a strong correlation with radius, and theoretical arguments expect SFE$_{\rm gas}$ to be closely related to orbital timescale in typical disks \citep{Silk1997,Elmegreen1997,Kennicutt1998}. A correlation between SFE$_{\rm gas}$ and $\tau_{\rm orb}$ is based on the ``Silk-Elmegreen'' relation, which states that $\Sigma_{\rm SFR}=\epsilon_{\rm orb}\,\Sigma_{\rm gas}/\tau_{\rm orb}$, where $\epsilon_{\rm orb}$ is the fraction of the gas converted into stars per orbital time (also called ``orbital efficiency''). Therefore, because ${\rm SFE_{\rm gas}}=\Sigma_{\rm SFR}/\Sigma_{\rm gas}$, SFE$_{\rm gas}$ and $\tau_{\rm orb}$ are related by
\begin{equation}
{\rm SFE_{\rm gas}} = \frac{\epsilon_{\rm orb}}{\tau_{\rm orb}}.
\label{SE_relation}
\end{equation}
\begin{figure}
\hspace{-0.6cm}
\begin{tabular}{c}
\includegraphics[width=8.5cm]{Leroy2008_fig7A.pdf}\\
\hspace{-.3cm}\includegraphics[width=8.5cm]{Leroy2008_fig7A_Morph.pdf}
\end{tabular}
\caption{{\it Top:} SFE$_{\rm gas}$ as a function of the orbital timescale, $\tau_{\rm orb}$. Color coding and symbols are as described in Figure \ref{FIG_3_Leroy2008}. The black-dashed line is the best fit of the binned data and shows 5\% of gas converted into stars per $\tau_{\rm orb}$. The `r' term represents the Pearson correlation coefficient, including the binned annuli, for the relation between the SFE$_{\rm gas}$ and $\tau_{\rm orb}$. {\it Bottom:} SFE$_{\rm gas}$ averaged over $\tau_{\rm orb}$ bins over all galaxies of selected morphological types as in Figure \ref{FIG_1_Leroy2001B}. Black-dashed lines, from top to bottom, represent the 50\%, 17\%, 5\%, 1.7\%, and 0.5\% efficiency of gas converted into stars per $\tau_{\rm orb}$.}
\label{FIG_7_Leroy2008}
\end{figure}
It is interesting to analyze the relations between the different timescales since they can give intuition about the physical processes underlying the star formation activity \citep[e.g.,][]{SemenovKravtsovGnedin2017,Colombo2018}. Equation \ref{SE_relation} shows that the timescale to deplete the gas reservoir and the orbital timescale are related through $\epsilon_{\rm orb}$. Although there is large scatter, the median values of $\tau_{\rm orb}$ and $\tau_{\rm dep}$ for our sample are $(2.0^{+0.9}_{-0.7}) \times 10^{8}$ yr and $(2.8^{+1.1}_{-1.0}) \times 10^{9}$ yr, respectively. These values are in good agreement with previous EDGE-CALIFA sample results found by \cite{Colombo2018}, who analyze a more limited subsample of 39 galaxies without the benefit of CO line stacking and more constrained to inclination below $65^{\circ}$, with $\tau_{\rm orb} = (3.2^{+2.0}_{-1.2})\times 10^{8}$ yr and $\tau_{\rm dep} = (2.8^{+2.3}_{-1.2})\times 10^{9}$ yr. The black-dashed line in the top panel of Figure \ref{FIG_7_Leroy2008} corresponds to the best fit to our binned data (black symbols); our fit excludes lower limits (shown as triangles in the figure), and it shows that $\epsilon_{\rm orb}\approx 5\%$ of the total gas mass is converted to stars per $\tau_{\rm orb}$. This average efficiency is lower but similar to the $\epsilon_{\rm orb}\approx7\%$ of efficiency found by \cite{WongBlitz2002} and \cite{Kennicutt1998}, and the $\epsilon_{\rm orb}\approx6\%$ efficiency for HERACLES \citep{Leroy2008}. Also, this efficiency is the same to the average molecular gas orbital efficiency found by \cite{Colombo2018} for their subsample of EDGE-CALIFA galaxies ($\epsilon_{\rm orb}\approx5\%$). Similar to our results, all of these studies did not find a clear correlation between SFE$_{\rm gas}$ and $\tau_{\rm orb}$ in the inner regions of disks, where the ISM is mostly molecular.
Like \cite{Colombo2018}, however, we find that a constant $\epsilon_{\rm orb}$ is not a good approximation for the data. The efficiency per orbital time depends on the Hubble morphological type, with $\epsilon_{\rm orb}$ increasing from early- to late-types. This is shown in the bottom panel of Figure \ref{FIG_7_Leroy2008}, which shows the data grouped according to the same four morphological classes used in Figure \ref{FIG_1_Leroy2001B}. Our results show that annuli from Sbc, Sc, and Scd galaxies, which are the most numerous in our sample, seem to group around $\epsilon_{\rm orb}\sim5\%$. This value is also representative of the typical $\epsilon_{\rm orb}$ seen for the morphological bins comprised by Sa, Sab, and Sb and Sd, Sdm, Sm, and Ir types in the range $8.0<\log[\tau_{\rm orb}]<8.4$. However, these groups also show $\epsilon_{\rm orb}\lesssim5\%$ in the ranges $\log[\tau_{\rm orb}]<8.0$ and $\log[\tau_{\rm orb}] > 8.4$. However, early-type galaxies (with admittedly limited statistics, 21 annuli in total) show substantially lower $\epsilon_{\rm orb}$, with a median of $\epsilon_{\rm orb}=1.2\%$. These values are in agreement with previous results for EDGE-CALIFA galaxies by \cite{Colombo2018}, even though sample selection and processing were different. They observe a $\epsilon_{\rm orb}\sim 10\%$ for Sbc galaxies (most numerous in their sub-sample), and a systematic decrease in orbital efficiencies from late- to early-type galaxies.
As concluded in \cite{Colombo2018}, our results support the idea of a non-universal efficiency per orbit for the ``Silk-Elmegreen'' law. Figure \ref{FIG_7_Leroy2008} shows that $\epsilon_{\rm orb}$ depends not just on morphological type, but the behavior also varies with galactocentric radius: at short orbital time scales ($\log[\tau_{\rm orb}]\lesssim8.3$), or small radii ($\log[r/R_{\rm e}]\lesssim1.1-1.3$) the efficiency per unit time SFE$_{\rm gas}$ tends to be constant, and as a consequence the observed $\epsilon_{\rm orb}$ tends to systematically decrease as $\tau_{\rm orb}$ decreases. This is best seen in the top panel in the departure of the binned data (black symbols) from the dashed line of constant $\epsilon_{\rm orb}$. Note that this is also the approximate radius of the molecular disk, the region where molecular gas dominates the gaseous disk (Figure~\ref{FIG_1_Leroy2008_2}).
Other studies have also reported SFE$_{\rm gas}$ deviations as a function of morphology. \cite{Koyama2019} analyze CO observations of 28 nearby galaxies to compute the C-index $= R90_{petro,r}/R50_{petro,r}$ as an indicator of the bulge dominance in galaxies (where $R90_{petro,r}$ and $R50_{petro,r}$ are the radius containing the 90\% and 50\% of Petrossian flux for SDSS $r$-band photometric data, respectively). Although they do not detect a significant difference in the SFE$_{\rm gas}$ for bulge- and disk-dominated galaxies, they identify some CO-undetected bulge-dominated galaxies with unusual high SFE$_{\rm gas}$s. Their results may reflect the galaxy population during the star formation quenching processes caused by the presence of a bulge component, and they could explain the flattening shown in top panel (mostly dominated by annuli within bulges) and bottom panel (mainly due by early-type and Sb-Scd galaxies annuli) of Figure \ref{FIG_7_Leroy2008}.
\begin{figure*}
\begin{center}
\includegraphics[width=8.cm]{Leroy2008_fig12A_Qgas.pdf}
\includegraphics[width=8.cm]{Leroy2008_fig12A.pdf}\\
\vspace{-0.25cm}
\includegraphics[width=7.8cm]{Leroy2008_fig12B_Qgas.pdf}
\includegraphics[width=7.8cm]{Leroy2008_fig12B.pdf}\\
\vspace{-3.15cm}\hspace{-.2cm}\includegraphics[width=9.3cm]{Leroy2008_fig12B_Qgas_Morph.pdf}
\hspace{-1.4cm}\includegraphics[width=9.3cm]{Leroy2008_fig12B_Morph.pdf}
\end{center}
\vspace{-3.3cm}
\caption{SFE$_{\rm gas}$ as a function of Toomre's gravitational instability Q parameter for two different scenarios. {\it{Left:}} The SFE$_{\rm gas}$ is plotted as a function of the Toomre Q parameter for gas, $Q_{\rm gas}$. {\it{Right:}} The SFE$_{\rm gas}$ is plotted as a function of the Toomre Q parameter redefined by \cite{Rafikov2001} to include the contribution of stars and gas, $Q_{\rm stars+gas}$. Green squares in the upper left and right panels correspond to HERACLES spiral galaxies; black dashed line sets the limit where the gas is unstable or stable. The center left and right panels show the same points included in the upper ones but this time divided into low and high galaxy stellar mass sets; red points correspond to binned annuli which belong to galaxies with $\log_{10}$($M_{\star}$)$>10.7$, while blue points belong to galaxies with stellar masses below this limit. Blue and red contours are the 66\% and 33\% of the points for each mass set. The bottom left and right panels provide information about the morphological type of the host galaxy for a given annulus. The crosses correspond to the center of mass of the $\log_{10}$SFE$_{\rm gas}$ vs center of mass of $\log_{10} Q$ points for each set of morphological types.}
\label{FIG_12_Leroy2008}
\end{figure*}
\subsubsection{Gravitational instabilities}
\label{Gravitational_instabilities}
The formulation of the Toomre $Q$ gravitational stability parameter \citep[][see Section \ref{Basic_equations} for more details]{Toomre1964} provided a useful tool to quantify the stability of a thin disk disturbed by axisymmetric perturbations. Some studies have shown that the star formation activity is widespread where the gas disk is $Q$-unstable against large-scale collapse \citep[e.g., ][]{Kennicutt1989,MartinKennicutt2001}.
First we examine the case where only gas gravity is considered; the top left panel of Figure \ref{FIG_12_Leroy2008} considers this case, showing the SFE$_{\rm gas}$ as a function of both the Toomre instability parameter for a thin disk of gas (x-bottom axis), $Q_{\rm gas}$, and galactocentric radius (indicated by dot color). The vertical black-dashed line marks the limit where the gas becomes unstable to axisymmetric collapse. The vast majority of our points are in stable (or marginally stable) annuli with an average $Q_{\rm gas} = 3.2$. There is no apparent correlation of SFE$_{\rm gas}$ with $Q_{\rm gas}$ (Pearson correlation coefficient of 0.17), and that is independent of galaxy mass (middle left panel) or type (bottom left panel). In other words, SFE$_{\rm gas}$ does not decrease as stability increases (i.e as $Q_{\rm gas}$ increases). This is in agreement with similar results reported in previous studies. For example, using H$_{\rm I}$ observation for 20 dwarf Irregular galaxies selected from the Local Irregulars That Trace Luminosity Extremes, The H$_{\rm I}$ Nearby Galaxy Survey \citep[LITTLE THINGS; ][]{Hunter2012}, \cite{Elmegreen2015} find that dIrr galaxies are $Q_{\rm gas}$-stable, with a mean $Q_{\rm gas}\sim4$. They also find their galaxies have relatively thick disk, with typical (atomic) gas scale heights of $h_{\rm gas}\sim0.3-1.5$kpc. Consequently, they are more stable than the infinitely thin disks for which the $Q_{\rm gas}=1$ criterion is derived.
Stars represent the dominant fraction of mass in disks at galactocentric radii with active star formation. Thus, it makes sense to account for their gravity when determining the stability of the ISM in these regions. The top right panel of Figure \ref{FIG_12_Leroy2008} shows the SFE$_{\rm gas}$ as a function of Toomre's instability parameter modified by \citealt{Rafikov2001} to include the effects of both gas and stars, $Q_{\rm stars+gas}$, again galactocentric radius is indicated by color. As expected, we find that disks become more unstable when stellar gravity is included in addition to gas with a few points appearing in the nominally unstable region for thin disks. The bulk of the annuli, however, are found at around $Q_{\rm stars+gas}\approx1.6$. This is roughly consistent with calculations of $Q$ in other samples \citep{Romeo2020}. There is, however, no correlation of SFE$_{\rm gas}$ with $Q$.
\begin{figure}
\hspace{-0.6cm}
\begin{tabular}{c}
\includegraphics[width=8.cm]{Ratio_VS_Q.pdf}\\
\includegraphics[width=7.7cm]{Ratio_VS_Q_2.pdf}
\end{tabular}
\caption{{\it Top:} Molecular gas surface density, $\Sigma_{\rm mol}$, as a function of Toomre's instability parameter for gas and stars, $Q_{\rm stars+gas}$, for annuli with galactocentric radii within $0.3r_{25}$ (red points), between $0.3r_{25}$ and $0.5r_{25}$ (yellow points), and outside $0.5r_{25}$ (blue points). Each point represents the value of $\Sigma_{\rm mol}$ averaged over a $r/r_{25}$-wide annuli. Points are color-coded by galactocentric radius (in $r_{25}$), as indicated by the color bar on the right side. Solid-lines are PCA major and minor axes for which each of the groups vary most. The axes are normalized to fit the major and minor axes of the elliptical contours that enclose 50\% of the annuli within a given range. The ratio between the major and the minor axes from the PCA, $R_{\rm maj/min}= r_{\rm maj}/r_{\rm min}$, is in the upper left. Typical 1-$\sigma$ error bars are shown in bottom left. The horizontal black-dotted line represents the assumed $\Sigma_{\rm atom} = 6$ M$_{\odot}$ pc$^{-2}$. {\it Bottom:} Distribution of the Spearman rank correlation coefficients for the three $r$ ranges in the top panel after randomizing the $\Sigma_{\rm mol}$ data, per range, in 200 realizations to test for the degree of internal correlation of the axes. The horizontal dashed-red, dashed-yellow, and dashed-blue lines are the Spearman rank coefficients for the actual data, from inner to outer ranges, respectively. This shows that the correlations observed in the top panel are completely consistent with being a result of the definition of $Q_{\rm gas+stars}$ (see discussion in the text), and thus (although tantalizing) are not particularly meaningful.}
\label{Ratio_vs_Q}
\end{figure}
The center panels of Figure \ref{FIG_12_Leroy2008} show the SFE$_{\rm gas}$s versus $Q_{\rm gas}$ and $Q_{\rm stars+gas}$ but this time splitting the points in two groups of different galaxy stellar mass; as in Section \ref{SFE_radius}, we choose $\log_{10}$[$M_{\star}$]$=10.7$ to split the groups. Although the two groups separate in $Q_{\rm gas}$, with annuli from galaxies with $\log_{10}$[$M_{\star}$]$<10.7$ tending to be in general more stable, the separation disappears once the stars are taken into account in the $Q$ calculation.
In one of the ideas on how stars relate to SFE$_{\rm gas}$, \cite{Dib2017} show that star-formation may be associated with the fastest growing mode of instabilities. In that case, the relation between SFR and gas in spiral galaxies may be modulated by the stellar mass, which will contribute to the gravitational instability and regulation of star formation \citep[like in the case of NGC 628; ][]{Dib2017}. Also, the $\Sigma_{\rm SFE_{\rm gas}}$-$\Sigma_{\star}$ relation, known as the ``extended Schmidt law'', suggests a critical role for existing stellar populations in ongoing star formation activity, and it may be a manifestation of more complex physics where $\Sigma_\star$ is a proxy for other variables or processes \citep[][]{Shi2011}. Our results may reflect the importance of instabilities in enhancing the SFE$_{\rm gas}$ due to the strong gravitational influence from stars, particularly in galaxies with $\log_{10}$[$M_{\star}$]$>10.7$. But in the aggregate there is no apparent evidence for a trend showing that annuli with more unstable $Q$ have higher star formation efficiencies.
The bottom panels of Figure \ref{FIG_12_Leroy2008} show the same relations as upper panels but this time the data are grouped in four bins by morphological type. In both panels crosses correspond to the "center of mass" for each morphological group. Although annuli in early-type galaxies are more ``Toomre stable", the statistics are very sparse and the Toomre calculation may not apply (since these are not thin disks). Otherwise, we do not find a clear trend between morphology and stability based on the Toomre parameter for stars and gas. Previous studies have reported that $Q_{\rm star+gas}$ increases towards the central parts of spirals. For example, \cite{Leroy2008} found that although molecular gas is the dominant component of the ISM in the central regions, HERACLES galaxies seems to be more stable there than near the H$_{\rm 2}$-to-H$_{\rm I}$ transition. If the type of gravitational instability that $Q$ is sensitive to plays a role in star formation in galaxies, we would expect to see some links between $Q$ and molecular gas abundance. It is therefore interesting to test if there is dependence of the H$_2$-to-H$_{\rm I}$ ratio, $R_{\rm mol}=\Sigma_{\rm mol}$/$\Sigma_{\rm atom}$ on the degree of gravitational instability in EDGE galaxies. Since we assume a constant $\Sigma_{\rm atom}$, however, for us $R_{\rm mol}$ is simply a normalized molecular gas surface density, $\Sigma_{\rm mol}$. We use the typical H$_2$-to-H$_{\rm I}$ transition radius found in \S\ref{SFE_radius} to split the annuli into three groups: i) annuli at $r<0.3 r_{25}$ ($r < 0.6 R_{\rm e}$; red points) which should be strongly molecular, ii) annuli between $0.3 r_{25}<r$ and $r<0.5 r_{25}$ ($0.6 R_{\rm e} < r < 1.4 R_{\rm e}$; yellow points) which should be around the molecular to atomic transition region, and iii) annuli at $r > 0.5 r_{25}$ ($r > 1.4 R_{\rm e}$; blue points) which should be dominated by atomic gas. The top panel of Figure \ref{Ratio_vs_Q} shows that $\Sigma_{\rm mol}$ has a large scatter and does not seem to depend strongly on $Q_{\rm star+gas}$. Within the each range, however, we find that annuli with smaller galactocentric radii tend to be slightly more stable.
A suggestive trend emerges when we limit the range of galactocentric radii. We compute a Principal Component Analysis (PCA; \citealt{Pearson1901}) to find the main axis along which the three populations vary most. The top panel of Figure \ref{Ratio_vs_Q} shows the PCA major and minor axis for annuli in the three defined zones. The axes have been normalized to fit the minor of major axes of the elliptical contours that enclose 50\% of the annuli over a given range. The figure suggest that, within a given range, we tend to find more plentiful molecular gas in regions where annuli are more Toomre unstable. A concern, however, is that the axes in this plot have a degree of intrinsic correlation since the computation of $Q_{\rm star+gas}$ includes $\Sigma_{\rm mol}$. Therefore to assert that the correlation we observe is physically meaningful we need to show that it is stronger than that imposed by the mathematics of the computation. We quantify the strength of the correlations using the Spearman rank correlation coefficient, which is a non-parametric measure of the monotonicity of the observed correlations. To investigate the degree to which the axes are internally correlated, we randomize the $\Sigma_{\rm mol}$ data (within each range) and recompute $Q_{\rm star+gas}$ in 200 realizations, to obtain the distributions of the Spearman rank correlation coefficient for each randomized group. Clearly in the randomized data we would expect only the degree of correlation caused by the mathematical definition of the quantities.
The bottom panel of Figure \ref{Ratio_vs_Q} shows that the Spearman rank correlation coefficients for the actual data (dashed-red, dashed-yellow, and dashed-blue vertical lines) are consistent with the distributions seen in the randomized histograms. These results suggest that the correlation between $\Sigma_{\rm mol}$ and $Q_{\rm star+gas}$ seen in the top panel of Figure \ref{Ratio_vs_Q} is purely driven by the implementation of equation \ref{Rafikov}, in which $Q_{\rm star+gas}$ depends on $\Sigma_{\rm mol}$.
\section{Summary and conclusions}
\label{S6_Conclusion}
We present a systematic study of the star formation efficiency and its dependence on other physical parameters in 81 galaxies from the EDGE-CALIFA survey. We analyse CO 1-0 datacubes which have ${7}\arcsec$ angular resolution and 20 km s$^{-1}$ channel width, along with H$\alpha$ velocities extracted from the EDGE database, {\tt edge\_pydb} (Wong et al. in prep.). We implement a spectral stacking procedure for CO spectra shifted to the H$\alpha$ velocity to enable detection of faint emission and obtain surface densities averaged over annuli of width $0.1r_{25}$ ($\sim {3.5}\arcsec$), and measure $\Sigma_{\rm mol}$ out to typical galactocentric radii of $r\approx1.2\,r_{25}$ ($r\sim 3\,R_{\rm e}$). We assume a constant \citep{Walter2008}, a Milky-Way constant conversion factor of $\alpha_{\rm CO} = 4.3$ M$_\odot$ $[\rm K \, km \, s^{-1} \, pc^{-2}]^{-1}$, and a constant $\sigma_{\rm g}=11$ km s$^{-1}$ \citep{Leroy2008,Tamburro2009}. We perform a systematic analysis to explore molecular scale lengths and the dependence of the star formation efficiency SFE$_{\rm gas}$=$\Sigma_{\rm SFR}/(\Sigma_{\rm mol}+\Sigma_{\rm atom})$ on various physical parameters. Our main conclusions are as follows:
\begin{enumerate}
\item We determine the molecular and stellar exponential disk scale lengths, $l_{\rm mol}$ and $l_{\star}$, by fitting the radial $\Sigma_{\rm mol}$ and $\Sigma_{\star}$ profiles, respectively. We also obtain the radii that encloses 50\% of the total molecular mass, $R_{50,\rm mol}$, and stellar mass, $R_{50,\star}$ (see Fig. \ref{FIG_14_LENGTH}). To quantify the relations, we use an OLS linear bisector method to fit all our $3\sigma$ detections beyond $r > 1.5$ kpc. We find that $l_{\rm mol} = [0.86\pm0.07] \times l_{\star}$, $l_{\rm mol} = $ [$0.24 \pm 0.01$]$ \times r_{25}$, and $R_{50,\rm mol } = [0.93\pm 0.05$]$ \times R_{50,\star}$. These results are in agreement with values from the current literature, and indicate that on average the molecular and stellar radial profiles are similar.
\item We find that on average the SFE$_{\rm gas}$ exhibits a smooth exponential decline as a function of galactocentric radius, without a flattening towards the centers of galaxies seen in some previous studies (see Fig. \ref{FIG_1_Leroy2001B}), in agreement with recent results \citep[e.g.,][]{Sanchez2020_1,Sanchez2020_2}. We note a systematic increase in the average SFE$_{\rm gas}$ from early to late type galaxies. In HI-dominated regions, this conclusion depends strongly on our assumption of a constant HI surface density for the atomic disk. The EDGE-CALIFA survey encompasses a galaxy sample that has not been well represented by prior studies, which includes a larger number of galaxies with a broader range of properties and morphological types. This may explain the differences we observe when we compare our result with previous work.
\item The SFE$_{\rm gas}$ has a clear dependence on $\Sigma_{\star}$ (see Fig. \ref{FIG_3_Leroy2008}), a relation that holds for both the atomic-dominated and the molecular-dominated regimes. The SFE$_{\rm gas}$ has a comparatively flatter dependence on $\Sigma_{\rm gas}$ for high values of the gas surface density. This suggests that the stellar component has a strong effect on setting the gravitational conditions to enhance the star formation activity, not just converting the gas from HI to H$_2$. However, statistical tests, which are beyond the scope of this work, may be required to demonstrate that this secondary relation is not induced by errors \citep{Sanchez2021}.
\item There is a clear relationship between SFE$_{\rm gas}$ and the dynamical equilibrium pressure, $P_{\rm DE}$, particularly in the innermost regions of galactic disks. Moreover, we find a strong correlation between $\Sigma_{\rm SFR}$ and $P_{\rm DE}$. We identify a transition at $\log[P_{\rm DE}/k ({\rm K \, cm^{-3}})] \sim$ 3.7, above which we find a best-linear-fit slope of $1.11\pm0.15$. Our results are in good agreement with the current literature and support a self-regulated scenario in which the star formation acts to restore the pressure balance in active star-forming regions.
\item We find a power-law decrease of SFE$_{\rm gas}$ as a function of orbital time $\tau_{\rm orb}$ (see Fig. \ref{FIG_7_Leroy2008}). The average $\tau_{\rm orb}$ within $0.7 r_{25}$ for our galaxies is $2.6\pm0.2$ Gyr, with a typical efficiency for converting gas into stars of $\sim5\%$ per orbit. Note, however, that there are systematic trends in this efficiency. In particular, we note that there is a flattening of the SFE$_{\rm gas}$ for $\log [\tau_{\rm orb} ({\rm yr})]\sim 7.9-8.1$ which may reflect star formation quenching due to the presence of a bulge component. Although our methodology is different, our findings support the conclusion that the star formation efficiency per orbital time is a function of morphology \citep{Colombo2018}.
\item Finally, under the assumption of a constant velocity dispersion for the gas, we do not find clear correlations between the SFE$_{\rm gas}$ and $Q_{\rm gas}$ or $Q_{\rm stars+gas}$. It is possible that larger samples of galaxies may be required to confidently rule out any trends. Our typical annulus has $Q_{\rm stars+gas}\sim1.6$, independent of galaxy mass or morphological type. The range of $\Sigma_{\rm mol}$ is very broad, and we do not find any meaningful trends.
\end{enumerate}
Future VLA HI and ALMA CO data may improve the spatial coverage and sensitivity, allowing us remove some limitations and extend this analysis to fainter sources (e.g., earlier galaxy types), contributing to a more extensive and representative sample of the local universe.
\section{acknowledgments}
V. Villanueva acknowledges support from the scholarship ANID-FULBRIGHT BIO 2016 - 56160020 and funding from NRAO Student Observing Support (SOS) - SOSPA7-014. A. D. Bolatto, S. Vogel, R. C. Levy, and V. Villanueva, acknowledge partial support from NSF-AST1615960. J.B-B acknowledges support from the grant IA-100420 (DGAPA-PAPIIT, UNAM) and funding from the CONACYT grant CF19-39578. R.H.-C. acknowledges support from the Max Planck Society under the Partner Group project "The Baryon Cycle in Galaxies" between the Max Planck for Extraterrestrial Physics and the Universidad de Concepci{\'o}n. Support for CARMA construction was derived from the Gordon and Betty Moore Foundation, the Kenneth T. and Eileen L. Norris Foundation, the James S. McDonnell Foundation, the Associates of the California Institute of Technology, the University of Chicago, the states of California, Illinois, and Maryland, and the NSF. CARMA development and operations were supported by the NSF under a cooperative agreement and by the CARMA partner universities. This research is based on observations collected at the Centro Astron\'{o}mico Hispano-Alem\'{a}n (CAHA) at Calar Alto, operated jointly by the Max-Planck Institut f\"{u}r Astronomie (MPA) and the Instituto de Astrofisica de Andalucia (CSIC). M. Rubio acknowledge support from ANID(CHILE) Fondecyt grant No 1190684 and partial support from ANID project Basal AFB-170002. AST-1616199 for Illinois (TW/YC/YL) and AST-1616924 for Berkeley (LB/DU). This research has made use of NASA’s Astrophysics Data System.
\software{ Astropy \citep{AstropyCollaboration2018}, MatPlotLib \citep{Hunter2007},
NumPy \citep{Harris2020}, SciPy \citep{2020SciPy-NMeth}, seaborn \citep{Waskom2021}, Scikit-learn \citep{scikit-learn}.}
|
\section{Basic Assumptions}
This section lays out the assumptions about the structure and the property of the physical universe and of the concept of causality in order to prepare us to work in the mathematical domain. The reason one must articulate these hypotheses is that it is the only way to move the subject matter into the mathematical domain so as to develop some logico-mathematical results about them. Although the assumptions given in this paper may be disputable, the results we show will be rigorous and clear. One can discuss whether any or all of these hypotheses are not representative of the physical reality that we are living in, however we tried to keep these assumptions on the most fundamental level and in accordance with up-to-date cosmological understanding to the best of our abilities.
We begin by giving the following assumptions about the universe and then move onto giving some axioms concerning causality and events. In the final part we give our definition of worldline and how it can be represented in tree structures. The order between these items is irrelevant. However we shall list them starting from the simplest assumption and then get more specific.
\vspace{0.5cm}
\noindent {\bf I. Assumptions about the universe.}
\begin{enumerate}
\item [(a)] The universe is expanding indefinitely.\footnote{Whether the universe has a beginning or not will not be relevant to our analysis and will not affect the results for reasons that will become clear later. However, since we will be concerned with spacetime light cones of events in its general form, we may assume without loss of generality that the past and future cones are both {\em unbounded} and so we will make our analysis under the assumption that the future light cone of any event is potentially unbounded. Nevertheless we will be primarily concerned with the future light cone of events.}
\item[(b)] (i) There are finitely many atoms in the universe at any fixed moment. (ii) Every atom consists of finitely many indivisible parts.
\item[(c)] By (a) and (b), it follows that the observable universe is finite.
\item[(d)] Physical space is not dense.\footnote{In fact, whether space is discrete or continuous is not a settled question in physics as both are consistent with different physical theories. See Forrest \cite{Forrest1995} for a discussion.} That is, for any two distinct points $x$ and $y$, there are at most finitely many points in-between $x$ and $y$ in the Euclidean sense. That is,
\[
|\{z:bet(z,x,y)\}|<\aleph_0,
\]
where $bet(z,x,y)$ denotes that the point $z$ is between $x$ and $y$.
\item[(e)] No-signalling principle \cite{masanes2006} is valid (i.e. events are bounded by the speed of light in their future light cones).
\end{enumerate}
\noindent {\bf II. Assumptions about causality and events}
\begin{enumerate}
\item [(a)] Let $x$ and $y$ be two events.\footnote{By {\em event} we mean an instantaneous situation or action that is associated with a point in spacetime. Events are primitive objects of the domain of discourse in the theory of causal sets.} We let $x\leq y$ mean ``event $x$ is a cause of event $y$" (or ``event $y$ is the effect of $x$). We let $x<y$ mean $x\leq y$ and $x\neq y$. Any given event $x$ defines a double light cone consisting of lower and upper parts, where the {\em lower cone} is defined as the collection of all possible past events that end up with $x$. That is, the lower cone of $x$ is
\[
\{y:y\leq x\}.
\]
The {\em upper cone} represents all possible future events that stems from $x$, which is defined as the set
\[
\{y:x\leq y\}.
\]
\item[(b)] Causality is a transitive. That is, if $x\leq y$ and $y\leq z$, then $x\leq z$.
\item[(c)] Causality is not dense, i.e., if $x\leq y$ then there do not exist infinitely many events occurring between the event $x$ and the event $y$.\footnote{This is also known as the {\em locally finiteness} condition in causal sets.}
\item[(d)] We say that two events $x$ and $y$ are {\em incomparable} (or {\em spacelike separated}) if neither $x\leq y$ nor $y\leq x$. Two incomparable events $x$ and $y$ cannot be seen as a single event. Similarly, no event can be decomposed into smaller incomparable sub-events. In other words, all events are `atomic' (see Section 2.2 for details).
\item[(e)] For every event $x$, since there are finitely many atoms---by I(b) and I(d)---there can only be finitely many events $y_1,\ldots,y_n$ such that $y_i$ is the {\em immediate successor} (or {\em immediate effect}) of $x$, meaning that $x<a<y_i$ for no event $a$ (See Section 2.2 for details).
\item[(f)] A {\em closed timelike curve} is a sequence of events that happen between two distinct events $x$ and $y$ such that $x\leq y$ and $y\leq x$. There exist {\em closed timelike curves} iff the {\em anti-symmetric property} does not hold for causality. In other words, there are closed timelike curves iff it is not the case that $x=y$ whenever $x\leq y$ and $y\leq x$.\footnote{In most theories of causal sets, the anti-symmetric property is implicitly assumed. We follow the same tradition.}
\item[(g)] It follows from I(b), II(c), and II(e) that given any event $x$, the collection of all future events that stem from $x$ forms a subset of $2^{<\mathbb{N}}$ (see Section 2.2 for details).
\item[(h)] The collection of all events which end up with $x$ gives the lower cone of $x$ and it can be similarly represented by a subset of $2^{<\mathbb{N}}$ (see Section 2.2 for details).
\end{enumerate}
\noindent {\bf III. Assumptions on worldlines}
\begin{enumerate}
\item[(a)] Let $x$ be an event. A {\em worldline} of $x$ is sequence of events
\[\ldots,p_2(x),p_1(x),x,f_1(x),f_2(x),\ldots
\]
where each $p_i(x)$ denotes the past events such that $p_i(x)\leq x$ and that $p_{i+1}(x)\leq p_i(x)$ for every $i$, whereas $f_i(x)$ denotes the future events satisfying that $x\leq f_i(x)$ and $f_i(x)\leq f_{i+1}(x)$ for every $i$.
\end{enumerate}
Note that there may exist many worldlines in the possible world interpretation, i.e., if the universe is non-deterministic, whereas there can only be one worldline associated with a given event in a deterministic picture. We do not endorse a particular position here, but we will later see what follows if we assume each position.
We shall now discuss about the consequences of these basic assumptions and also argue what worldlines actually correspond to in our presentation.
\section{Worldlines as real numbers}
We first give some basic notions and facts about effectively closed subsets of Cantor space and computable trees. These notions can be found in more detail in \cite{Diamondstone}, and \cite{CenzerPi01classes}. We then elaborate more on the basic assumptions given in Section 1 to associate these notions with the physical concepts. Finally we give some results and corollaries that will apply to causal sets.
\subsection{Computable subsets of Cantor space}
\label{sec:cantorspace}
When we say a {\em computable set} we mean a set $A\subseteq\mathbb{N}$ for which there is an algorithm such that for any given natural number $n$, the algorithm can decide whether or not $n\in A$.\footnote{We will use the terms {\em computable} and {\em decidable} interchangably.} Of course any countable set is in one-to-one correspondence with the set of natural numbers under an appropriate pairing function. The definition then generalizes to any countable set under a suitable pairing function. Let us start by giving the notation for strings and trees.
\vspace{0.5cm}
\noindent{\bf Strings.}
A {\em string} is a finite sequence of symbols. We denote finite strings with lowercase Greek letters like $\sigma,\tau,\eta,\rho,\pi,\upsilon$. We let $\sigma*\tau$ denote the concatenation of $\sigma$ followed by $\tau$. We let $\sigma\subseteq\tau$ denote that $\sigma$ is an initial segment of $\tau$. For example, $1001\subseteq 10011$. We say a string $\sigma$ is {\em incompatible with $\tau$} if neither $\sigma\subseteq\tau$ nor $\tau\subseteq\sigma$. Otherwise we say that $\sigma$ is {\em compatible} with $\tau$. Similarly, we say that $\sigma$ is an {\em extension} of $\tau$ if $\tau\subseteq\sigma$.
\vspace{0.5cm}
\noindent {\bf Trees and $\Pi^0_1$ classes.}
A set $T$ of strings is {\em upward closed} if $\sigma\in T$ and $\sigma\subseteq\tau$ then $\tau\in T$. A {\em tree} is an upward closed set of strings. For $n\in\mathbb{N}$, an {\em $n$-ary tree} is a tree with at most $n$ many branchings for each node. We say that a set $A$ {\em lies} on $T$ if there exist infinitely many $\sigma$ in $T$ such that $\sigma\subseteq A$. A set $A$ is a {\em path} on $T$ if $A$ lies on $T$. So if $A$ is a path on $T$, then every initial segment of $A$ is in $T$. We denote the set of infinite paths of $T$ by $[T]$.
We say that a string $\sigma\in T$ is {\em infinitely extendible} if there exists some $A\supseteq\sigma$ such that $A\in[T]$. A tree $T$ is {\em perfect} if every infinitely extendible string in $T$ has at least two incompatible extensions in $T$.
If $\sigma,\tau\in T$ and $\sigma\subseteq\tau$ and there does not exist $\sigma'$ with $\sigma\subseteq\sigma'\subseteq\tau$ then we say that $\tau$ is an {\em immediate successor} of $\sigma$ in $T$ and $\sigma$ is the {\em immediate predecessor} of $\tau$ in $T$.
\begin{defn}
We say that a tree $T$ is {\em computable} if for any string $\sigma$, there is an algorithm which decides whether or not $\sigma\in T$.
\end{defn}
\begin{defn}
\begin{enumerate}
\item[(i)] Let $T\subseteq n^{<\mathbb{N}}$ be a tree. The set of infinite paths through $T$ is defined as
\begin{center}
$\left[T\right]=\{A:\forall n(A\upharpoonright n\in T)\}$,
\end{center}
where $A\upharpoonright n$ denotes the initial segment of the path $A$ up to length $n$.
\item[(ii)] A class $\mathcal{A}\subseteq n^\mathbb{N}$ is $\Pi^0_1$ (or {\em effectively closed}) if there exists a computable relation $\varphi(n,X)$ such that $X\in \mathcal{A}$ if and only if $\forall n \ \varphi(n,X)$, where $n$ ranges over $\mathbb{N}$ and $X$ ranges over $n^\mathbb{N}$.\footnote{Effectively closed means computably closed.}
\end{enumerate}
\end{defn}
Now from the definitions, it is easy to see the following.
\begin{thm}
Let $\mathcal{A}\subseteq n^\mathbb{N}$ be a class for some natural number $n$. The following are known to be equivalent.
\begin{enumerate}
\item[(i)] $\mathcal{A}=\left[T\right]$ for some computable tree $T$.
\item[(ii)] $\mathcal{A}$ is effectively closed.
\end{enumerate}
\end{thm}
Since we are working in Cantor space, we shall mention the compactness property of it. Although compactness can be provided by various forms, the most well known form is given by {\em K\"{o}nig's lemma} \cite{Konig}.
\begin{lem}[K\"{o}nig's Lemma]
If $T$ is a finitely branching infinite tree, then $T$ has an infinite path.
\end{lem}
We shall now give some topological properties about trees.
\begin{defn}
Let $n\in\mathbb{N}$ and let $T\subseteq n^{<\mathbb{N}}$ be a tree.
\begin{enumerate}
\item[(i)] For any given $\sigma\in T$, we let $T_\sigma$ be the {\em subtree of nodes compatible with $\sigma$} and be defined as
\begin{center}
$T_\sigma=\{\tau\in T:\sigma\textrm{ is compatible with }\tau\}$.
\end{center}
\item[(ii)] A path $A\in\left[T\right]$ is said to be {\em isolated} if there exists a string $\sigma\in T$ such that $\left[T_\sigma\right]=\{A\}$, in which case we say that $\sigma$ {\em isolates} $A$. Otherwise $A$ is called a {\em limit point}.
\end{enumerate}
\begin{defn}
We say that $\sigma\in T$ is {\em infinitely extendible} in $T$ if there exists some $A\supseteq\sigma$ such that $A\in\left[T\right]$.
\end{defn}
Note that when $\sigma$ isolates $A$, there are no incompatible infinite extensions of $\sigma$ in $T$.
\end{defn}
\subsection{Translation between mathematical and physical contexts}
Next we elaborate on some of the assumptions given in Section 1 and then argue that given any event $x$, the future light cone of $x$ can be seen as an effectively closed subset of Cantor space, i.e., a $\Pi^0_1$ class. So the worldlines form members of such classes. Hence one can treat strings in a computable tree as spacetime events in light cones, and treat worldlines as infinite paths on the tree through which the effectively closed set is represented.
\vspace{0.5cm}
\noindent {\bf Elaboration of II(d):}
Given an event $x$, a split of at least two immediate effects of $x$ defines incomparable spacetime events (see Figure \ref{fig:split}). In the MWI, thus, a split is usually considered as a way of defining alternate universes.
\begin{figure}[ht]
\begin{center}
\includegraphics{split.eps}
\caption{A split defines alternate worldlines.}
\label{fig:split}
\end{center}
\end{figure}
Hence the assumption given in II(d) ensures the existence of alternate universes relative to any event. One may ask what would happen if there were mergings in the light cone of an event. It may be possible that two incomparable events may later merge into a single event and continue along the same worldline (see Figure \ref{fig:causality}). Even if separate universes merge into a single universe at some point in the future---in which case the structure of causal sets could possibly be represented by a directed acyclic graph---in our case one can just use trees to represent light cones. Furthermore, splitting and merging events will not have any effect on the cardinality of worldlines. The reason is the following. The number of paths from the beginning of the split to the merge always is finite. Let us take Figure \ref{fig:causality} as an illustration. A split occurs at event $b$ and a merging occurs at event $l$. Now there can only be finitely many events that are in the past of $l$ {\em and} the future of $b$. That is, there are finitely many $z$ such that $b\leq z$ and $z\leq l$. We shall call the collection of such $z$'s a {\em split-merge region}. The number of paths that start with $b$ and end up with $l$ is clearly finite. This then has no effect on the number of worldlines in the future of $b$ if the future of $b$ contains infinitely many paths.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1.3]{causality.eps}
\caption{An example of a split-merge region between the events $b$ and $l$, and of $b$ and $k$.}
\label{fig:causality}
\end{center}
\end{figure}
\noindent {\bf Elaboration of II(e):} We think of {\em event} as an indivisible entity in our discrete spacetime conception. By I(c), since the observable universe is finite and there are finitely many entities in the universe, there can only be finitely many immediate effects of an event. This is due to that there are finitely many positions a particle can be in at each next time click, where the number of positions is bounded by the size of the observable universe and the speed of light.\footnote{This discussion can possibly be extended to show that under Lorentz transformation, there exist a suitable pairing function that can translate the order of events/branches between different reference frames while maintaining the overall causal structure of forward and backward branching trees, however, we did not delve further into this line of inquiry.} This means that there are finitely many possible effects that emerge from every event. This also guarantees that the future light cone of an event (and similarly the past light cone, as these two light cones are symmetric with each other) can be represented by a finitely branching tree. The elements of the tree consists of spacetime events partially ordered by causation. So our first observation is the following: Given any event $x$, the future (and the past) light cone of $x$ can be thought of as a finitely branching tree of strings. Since the universe is expanding indefinitely, the tree is potentially infinite in the limit.
\vspace{0.5cm}
\noindent{\bf Elaboration of II(f):} In causal sets, usually the anti-symmetric property is assumed outright, and this is also what we adopt as well. However, one may ask what follows if one chooses not to add the anti-symmetric property as an axiom. The anti-symmetric property, it seems, ensures that the light cone of an event does not contain any cycles. In fact this corresponds to the concept of closed timelike curves in the casual structure. If $x\neq y$ whenever $x\leq y$ and $y\leq x$, then this means that the spacetime region that lays in-between the events $x$ and $y$ is a closed timelike curve. So the anti-symmetric property fails iff there exists a closed timelike curve. We leave it as an open question that how many possible universes there would be if one decided to reject the anti-symmetric property. Note that the representation of light cones changes in case one assumes the existence of closed timelike curves.
\noindent{\bf Elaboration of II(g):} It follows from I(b), II(c), II(e) and from the previous paragraph that given any event $x$, the collection of all future events from $x$ forms a subset of $n^{<\mathbb{N}}$. II(c) ensures that if $y$ is an immediate successor of $x$, then there is no spacetime event $z$ satisfying that $x<z<y$. So given the future light cone of $x$ as a tree, all nodes in the tree are separated due to this condition. So II(c) makes the concept of {\em immediate effect} available for use. And so whenever $y$ is an immediate effect of $x$, this can be simply shown by a directed branch from $x$ to $y$, where $x$ is put just below $y$ in the tree. Combining this with II(e) and I(d), the future light cone of an event $x$ forms a collection of events which are the immediate effects of $x$, the immediate effects of effects of $x$, and so on. This defines a subset of $n^{<\mathbb{N}}$ for some natural $n$ since there are only finitely many immediate effects of $x$.
\vspace{0.5cm}
\noindent{\bf Elaboration of II(h):} Since we work with cardinal arithmetic and we want to calculate the number of universes relative to a given event $x$, we may imagine the past light cone as the symmetry of future light cone. But we need to look at two possibilities where they might differ from each other.
Possibility 1: The universe has no beginning. Assuming that the past of $x$ is eternal, future and past light cones of $x$ can be taken as two separate tree structures where one is just an inverted version of the other. So the same structure holds for both past and future light cones of an event $x$. That is, they are both infinite subsets of $n^{<\mathbb{N}}$.
Possibility 2: The universe has a beginning. In this case, the number of events that end up with $x$ is finite.
Assuming that the universe is indefinitely expanding, the number of future events of $x$ is greater than or equal to the number events in the past of $x$. So this means that, in the worst case, we can just work with the future light cone of $x$ to calculate the cardinality of universes relative to $x$. The dominant factor without loss of generality can thus be assumed to come from the number of worldlines that extend $x$.
\vspace{0.5cm}
\noindent{\bf Elaboration of III(a):}
The definition given in III(a) guarantees that a worldline of a given event $x$ is a sequence of successive events in the past and the future of $x$. A future worldline of $x$ is just a path that lies on the future light cone of $x$. Since we argued in the elaboration of II(h) that it is sufficient to work with the future light cone, it is worth noting that the future light cone of $x$ is simply a subset of $n^{<\mathbb{N}}$ as argued in the elaboration of II(g). But since the future light cone is represented by a finitely branching tree $T$ of strings, where $x$ is the root of the tree and every successor of $x$ is an effect of $x$, the infinite branches that lie on $T$ give us the future worldlines of $x$, that is, possible universes that emerge from $x$.
\vspace{0.5cm}
We have not quite argued about how to form a tree of strings from events. Let $x$ be a spacetime event. The future light cone of $x$ can be formed as a tree $T\subseteq n^{<\mathbb{N}}$ for some natural number $n$ inductively as follows: We let $x$ be the empty string, hence the root of the tree. Let $\sigma\in T$ represents an event $y$. Every immediate effect $z_i$ of $y$ is defined by $\sigma*i$ in $T$.
Now one may also ask if the tree defined by the effects of an event $x$ constitutes a computable tree. This is in fact trivial since we really take the full tree $n^{<\mathbb{N}}$ to begin with, where $n$ is bounded by the space that a particle can travel from $x$ with a speed of light. So the tree $n^{<\mathbb{N}}$ consists of all immediate effects of $x$, all immediate effects of immediate effects of $x$, and so on. The future light cone of $x$ is then always computable and it basically corresponds to the full $n$-ary tree. By I(e), the immediate effects of any given event is bounded by the region of space through which light can travel in a least possible time tick. It is clear that the number of infinite paths on a full $n$-ary tree is uncountable since it is perfect.
Given a spacetime event $x$, it should now be clear that the future light cone of $x$ is a full $n$-ary tree. The class of all infinite paths (worldlines) on this tree gives us an effectively closed sets. So now properties about members of effectively closed sets can be translated into a property of worldlines extending the event $x$. Each future worldline is an infinite path on the tree of effects of $x$. So every infinite path can be thought of as a possible universe starting from a given spacetime event $x$. Of course $x$ can always be defined as the first cause (such as the Big Bang event) if we assume that the universe has a beginning.
\subsection{Alternate universes in causal sets}
\label{sec:alternateuniverses}
The axioms of causal sets given in Section 1 describes a discrete spacetime structure on which a partial order is defined by causality. Given any event $x$, the future light cone of $x$, which is represented by a full $n$-ary tree $T$, defines all alternate universes that originate from $x$. These universes are defined by the collection of all infinite paths on $T$, and this set is necessarily an uncountable set since it is perfect. But we ask under what conditions there could be less universes. Clearly if all alternative universes have an end, then there must be only finitely many universes since there will always be a point in each universe after which no event occurs. A more interesting worldline is that which is isolated by an event. An isolated path on $T$ describes a worldline which eventually stablizes and stops to split further. We observe that this corresponds to the following notion.
\begin{defn}
Let $x$ be an event. A worldline $p$ of $x$ is called {\em eventually deterministic} if $p$ is an isolated path on the future light cone of $x$.
\end{defn}
Now we will make use of the following results.
\begin{thm}\label{thm:theoremtrees}
\begin{enumerate}
\item [(i)] Any isolated member of an effectively closed set is computable.
\item[(ii)] An effectively closed set is called {\em special} if it does not contain a computable member. Every special effectively closed set has $2^{\aleph_0}$ many members. By contrapositive, if an effectively closed set is countable, then it contains a computable member.
\item[(iii)] Any countable closed subset of Cantor space contains an isolated member.
\end{enumerate}
\end{thm}
\begin{prf}
\begin{enumerate}
\item [(i)] Let $T$ be a computable tree and let $A$ be a path on $T$. Assume that $A\in[T]$ isolated. Then there is a string $\sigma\subseteq A$ such that no path on $T$ except $A$ extends $\sigma$. But then by K\"{o}nig's Lemma, for any $n>|\sigma|$, there exists a unique $\tau\supseteq\sigma$ such that the subtree of $T$ above $\tau$ is infinite. To compute $A\upharpoonright n$ for $n>|\sigma|$, we find $m\geq n$ such that exactly one $\tau\supseteq\sigma$ of length $n$ has an extension of length $m$ in $T$. Then, $A\upharpoonright n = \tau$.
\item[(ii)] By (i), since there is no computable member in $[T]$, every member of $T$ splits. Therefore the number of elements of $[T]$ is $2^{\aleph_0}$.
\item[(iii)] Let $[T]$ be a closed set for a tree $T$. If $[T]$ has no isolated points, then $T$ is perfect and thus $[T$] is uncountable.
\end{enumerate}
\end{prf}
One consequence of this theorem concerns classes with finitely many members.
\begin{col}
Let $\mathcal{P}$ be a finite $\Pi^0_1$ class. Then every member of $\mathcal{P}$ is computable.
\end{col}
The rest is just translating the results into the physical domain. We can argue that given any event $x$, if all worldlines extending $x$ are eventually deterministic, then there are countably many alternate universes with respect to $x$. Consider the future light cone of $x$, which is described by a tree $T$. This is simply due to the fact that if every member of $[T]$ is eventually deterministic---and so all isolated---then $[T]$ defines a discrete set. Thus $[T]$ is necessarily countable.
It also follows from Theorem \ref{thm:theoremtrees} that if there are countably many universes extending a given event $x$, then at least one of these universes is decidable.
Another corollary is that there can only be countably many universes which have an ending. To see this, let $x$ be an event. Consider the future light cone of $x$. But there cannot be uncountably many finite paths extending $x$ since each would have to be an isolated path.
\section{Discussion and Conclusion}
In this work we tried to explore the relationship between the many-worlds interpretation of quantum mechanics and the cardinality of the possible worlds in different circumstances in the parallel universes under the given interpretation. Initially, we started by defining what an event is with respect to the given interpretation of spacetime, and showed how worldlines could be constructed based on this. We then associated worldlines with the infinite paths of computable trees, and the collection of paths of these trees with effectively closed subsets of Cantor space, (in fact $n^\mathbb{N}$). Using some known topological properties of the Cantor space and computability, under the assumptions we made with respect to the nature of spacetime (i.e., that being discrete), we derived some results through finitely branching trees containing all possible worldlines as infinite paths, irrespective of whether there is an initial event (like Big Bang) or an eternal universe model. Our observation revealed the following results; if all worldlines extending $x$ are `eventually deterministic', in the sense that there exists an event above which there is no more split, then the number of alternate universes with respect to $x$ is exactly $\aleph_0$. Furthermore, if there are countably many alternate universes with respect to $x$, then at least one of them must be decidable in the sense that there is an effective uniform method, i.e. an algorithm, which determines whether or not any given event belongs to the given worldline. We also pointed out that the number of universes which have an end can only be countable, and so the number of universes with no end is uncountable.
The summary and discussion above can also be understood in the following manner. In the MWI of quantum mechanics, there exist all the possible universes with all possible configurations. Universes with no beginning, universes with singular beginnings, universes with no end, or multiple ends, eternal universes with no ends, universes with no beginning but that will end, and so on. In our analysis, we identified a relationship that states, under certain discrete spacetime assumptions, given an event $x$, if all worldlines extending that event lead to 'eventually deterministic' futures, then at least one of them must be decidable. The inverse of this conditional statement may not hold. Hence, our work actually identifies that for a certain type of a universe---where events lead to 'eventually deterministic' futures---there is a decidable worldline and that there are exactly $\aleph_0$ many eventually deterministic universes. We believe that, the property of decidability is an important distinction for an interpretation of quantum mechanics, that is, models that are unable to produce decidable worldlines should not be considered as physically possible.
We consider this work as a minor step toward the bigger goal of creating rigorous mathematical tools to apply on the problem of distinguishing between different interpretations of quantum mechanics. As the field of quantum technologies progress rapidly, the need to address the still unresolved questions from the early days of quantum mechanics should be identified as a worthy endeavour, and we want to strongly encourage the readers of this work to contemplate on these issues. Formulation of clear experiments to rule out certain interpretations can only come from delving deep into the assumptions and their implications, and it is a long and arduous journey that awaits the community of researchers both in foundations of quantum mechanics and philosophy of physics.
\bibliographystyle{abbrv}
|
\section{Introduction}
The \emph{reciprocity conjecture} in the Langlands program predicts a relationship
between pure motives\footnote{Here (in light of the standard conjectures~\cite{MR1265519}) one may take pure motives up to numerical or homological equivalence.
Conjecturally, one can also substitute
(for irreducible motive) the notion of an irreducible weakly compatible system
of Galois representations~\cite{TaylorICM} or an irreducible geometric Galois representation in the sense of Fontaine--Mazur~\cite{FM}.}
and automorphic representations. The simplest version (as formulated by Clozel~\cite[Conj~2.1]{Clozel}) states that
there should be a bijection
between irreducible motives~$M$ over a number field~$F$ with coefficients in~$\overline{\Q}$ and
cuspidal algebraic representations~$\pi$ of~$\mathrm{GL}_n(\mathbf{A}_F)$ satisfying a number of explicit additional compatibilities,
including the equality of algebraic and analytic~$L$-functions~$L(M,s) = L(\pi,s)$. In light of multiplicity one theorems~\cite{MR623137}, this pins down
the correspondence uniquely.
There is also a version of this conjecture
for more general reductive groups, although its formulation requires some care (as was done by Buzzard and Gee~\cite{BG}).
Beyond the spectacular application by Wiles to Fermat's Last Theorem~\cite[Theorem~0.5]{W},
the Taylor--Wiles method~\cite{W,TW}
gave a completely new technique --- and to this date
the most successful one --- for studying the problem of reciprocity. The ideas in these two papers have sustained progress in the field
for almost\footnote{Wiles in~\cite{W} dates the completion of the proof to September~19, 1994.}
30 years.
In this survey, we explain how the Taylor--Wiles method has evolved over this period
and where it stands today. \textbf{One warning:} the intended audience for this document is
entirely complementary to the audience for my talk ---
I shall assume more than a passing familiarity with the arguments of~\cite{W,TW}.
Moreover, this survey is as much a personal and historical\footnote{A whiggish history, naturally.
Even with this caveat, it should be clear
that
the narrative arc of progress presented here at best represents my own interpretation of events. I have added a few quotes from first hand sources when I felt
they conveyed a sense of what the experts were thinking in a manner not easily obtainable from other sources.
For other survey articles on similar topics, see~\cite{BuzzardE,BreuilE}.}
discussion as a mathematical one --- giving anything more than hints
on even a fraction of what is discussed here would be close to impossible given the space constraints and the competence of the author.
Even with the absence of any real mathematical details in this paper, the sheer amount of activity in this field has led me to discard any discussion
of advances not directly related to~$R=\mathbf{T}$ theorems, which necessitates the omission of a lot of closely related beautiful mathematics.
\subsection{The Fontaine--Mazur Conjecture}
Let~$F$ be a number field. The Fontaine--Mazur conjecture\footnote{Fontaine told me
(over a \emph{salad de g\'{e}siers} in Roscoff in 2009) that he and Mazur formulated their conjecture in the mid-80s.
(Colmez pointed me towards these notes~\cite{FontaineNotes} from a talk given by Fontaine at the 1988 Mathematische Arbeitstagung in Bonn.)
He noted that Serre had originally been skeptical,
particularly of the claim that any everywhere unramified representation inside~$\mathrm{GL}_n(\overline{\Q}_p)$
must have finite image, and set off to find a counterexample (using the construction of Golod--Shavarevich~\cite{GS}).
He (Serre) did not succeed!}~\cite{FM} predicts that any continuous irreducible~$p$-adic Galois representation
\begin{equation*}
\rho: G_F \rightarrow \mathrm{GL}_n(\overline{\Q}_p)
\end{equation*}
which is both unramified outside finitely many primes and potentially semi-stable (equivalently, de Rham~\cite{ColmezB}) at all places~$v|p$ should be associated to a motive~$M/F$ with coefficients in~$\overline{\Q}$.
Any such~$\rho$ is automatically conjugate to a representation in~$\mathrm{GL}_n(E)$ for some finite field~$E/\mathbf{Q}_p$ and further stabilizes an~$\mathcal{O}_E$-lattice. The
corresponding
residual representation~$\overline{\rho}: G_F \rightarrow \mathrm{GL}_n(k)$ where~$k = \mathcal{O}_E/\pi_E$ is the residue field of~$E$ is unique up to semi-simplification.
Let us assume here for expositional convenience that~$\overline{\rho}$ is absolutely irreducible.
Following Mazur~\cite{Mazur}, one may define a universal deformation ring which parameterizes
all deformations of~$\overline{\rho}$ unramified outside a finite set~$S$.
One can then further impose local conditions to define deformation rings~$R$ whose~$\overline{\Q}_p$-valued
points are associated to Galois representations which are de Rham at~$v|p$ with fixed Hodge--Tate weights.
Assuming the Fontaine--Mazur conjecture, these~$\overline{\Q}_p$-valued points correspond to all pure motives~$M$ unramified outside~$S$
whose~$p$-adic realizations are Galois representations with the same local conditions at~$p$ and the same
fixed residual representation~$\overline{\rho}$.
Assuming the reciprocity conjecture, these motives should then be associated to a finite dimensional space of automorphic forms. This leads to the extremely
non-trivial prediction that~$R$ has finitely many~$\overline{\Q}_p$-valued points.
The problem of reciprocity is now to link these~$\overline{\Q}_p$-valued points of~$R$ to automorphic forms.
\subsection{\texorpdfstring{$R = \mathbf{T}$}{R eq T} theorems}
Associated to the (conjectural) space of automorphic forms corresponding to~$\overline{\Q}_p$-valued points of~$R$
is a ring of endomorphisms generated by Hecke operators. The na\"{\i}ve version of~$\mathbf{T}$ is defined to be the completion of this ring with respect to a maximal ideal~$\mathfrak{m}$ defined in terms of~$\overline{\rho}$. The mere existence of~$\mathfrak{m}$ is itself conjectural, and amounts --- in the special case of odd absolutely irreducible~$2$-dimensional representations~$\overline{\rho}$ of~$G_{\mathbf{Q}}$ --- to
Serre's conjecture~\cite{MR885783}. Hence, in the Taylor--Wiles method, one usually assumes the existence of a suitable~$\mathfrak{m}$ as a hypothesis.
The usual shorthand way of describing what comes out of the Taylor--Wiles method is then an~``$R = \mathbf{T}$ theorem.''
Proving an~$R =\mathbf{T}$ theorem can more or less be divided into three different problems:
\begin{enumerate}
\item Understanding~$\mathbf{T}$. Why does there exist\footnote{At the time of Wiles' result, this was seen as the easier
direction (if not easy), although, in light of the success of the Taylor--Wiles method, it
may well be the harder direction in general.}
a map~$R \rightarrow \mathbf{T}$?
This is the problem of the ``existence of Galois representations.'' Implicit here is the problem of showing that those Galois representations not only exist but have
the ``right local properties'' at the ramified primes, particularly those dividing~$p$.
\item Understanding~$R$. Wiles introduced a mechanism for controlling~$R$ via its
tangent space using Galois cohomology (in particular Poitou--Tate duality~\cite{Milne}),
and this idea has proved remarkably versatile.
What has changed, however, is our understanding of local Galois representations and how this
information can be leveraged to understand the structure of~$R$.
\item Understanding why the map~$R \rightarrow \mathbf{T}$ is an isomorphism.
\end{enumerate}
We begin by summarizing the original~$R=\mathbf{T}$ theorem from this
viewpoint (or more precisely, the modification by Faltings which appears as an appendix to~\cite{TW}). We only discuss for now the so-called
``minimal case\footnote{The case when the Galois representations attached to~$R$ and~$\mathbf{T}$ have minimal level~$N$ as determined by the residual representation.}'' since this is most relevant for subsequent generalizations (see~\S\ref{section:kisin}).
Our summary is cursory, but see~\cite{MR1605752,CSS} for excellent expositional sources on early versions of the Taylor--Wiles method.
We start with a representation~$\overline{\rho}: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\overline{\mathbf{F}}_p)$ for~$p > 2$ which (say) comes from a semistable elliptic curve~$E$
and which we assume to be modular.
By a theorem of Ribet~\cite{Ribet100}, we may assume it is modular of level either~$N = N(\overline{\rho})$ or~$N = N(\overline{\rho})p$ where~$N(\overline{\rho})$ is the Serre
weight~\cite{MR885783} of~$\overline{\rho}$.
\begin{enumerate}
\item Understanding~$\mathbf{T}$: The construction of Galois representations associated to modular forms has its own interesting history (omitted here),
but (in the form originally needed by Wiles) was
more or less complete for modular forms (and even Hilbert modular forms) by 1990. The required local properties at primes different from~$p$ followed from work of Carayol~\cite{Carayol},
and the local properties at~$p$ were well understood either by Fontaine--Laffaille theory~\cite{FL}, or, in the ordinary case,
by Mazur--Wiles~\cite{MW} (see also work of Hida~\cite{MR848685,MR1463699}).
\item Understanding~$R$: Here~$R$ is a deformation ring of~$\overline{\rho}$ subject to precise local deformation conditions at~$p$
and the primes dividing~$N(\overline{\rho})$. For the prime~$p$,
the local conditions amount either to an ``ordinary'' or ``finite-flat'' restriction.
One then interprets the dual of the reduced tangent space~$\mathfrak{m}_R/(\mathfrak{m}^2_R,p)$ of~$R$ in terms of Galois cohomology,
in particular as a subgroup (Selmer group) of classes in~$H^1(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))$ satisfying local conditions. This can be thought
of as analogous to a class group, and one does not have any \emph{a priori} understanding of how large it can be although it has some finite dimension~$d$.
Using the Greenberg--Wiles formula, the obstructions in~$H^2(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))$ can
be related to the reduced tangent space, and allow one to realize~$R$ as a quotient
of~$W(k) \llbracket x_1, \ldots, x_d \rrbracket$ by $d$ relations. In particular, if~$R$
was finite and free as a~$W(k)$-module (as would be the case if~$R = \mathbf{T}$) then~$R$ would be a complete intersection.
\item Understanding why the map~$R \rightarrow \mathbf{T}$ is an isomorphism. Here lies the heart of the Taylor--Wiles method.
The ring~$\mathbf{T}$ acts on a natural module~$M$ of modular forms. One shows --- under a mild hypothesis on~$\overline{\rho}$
--- the existence of (infinitely many) sets~$Q = Q_N$ for any natural number~$N$ of cardinality~$|Q| = d$ --- so-called Taylor--Wiles primes --- with a number of pleasant properties:
\begin{enumerate}
\item The primes~$q \in Q$ are congruent to~$1 \bmod p^N$.
\item Let~$R_Q$ be the deformation ring capturing the same local properties as~$R$ but modified so that the representations at primes in~$Q$
may now be ramified of degree~$p^N$.
There is naturally a surjection~$R_{Q} \rightarrow R$, but for Taylor--Wiles primes, this modification does not increase the size of the tangent space.
In particular, for a fixed ring~$R_{\infty} = W(k) \llbracket x_1, \ldots, x_d \rrbracket$ there are surjections~$R_{\infty} \rightarrow R_Q \rightarrow R$ for every~$Q$.
\item The corresponding rings~$\mathbf{T}$ and~$\mathbf{T}_{Q}$ act naturally on spaces of modular forms~$M$ and~$M_Q$ respectively.
Using multiplicity one theorems, Wiles proves (see~\cite[Theorem~2.1]{W}) that~$M$ and~$M_Q$ are free of rank one over~$\mathbf{T}$ and~$\mathbf{T}_Q$ respectively. The space~$M$ can be interpreted as a space of modular forms for
a particular modular curve~$X$. The second key property of Taylor--Wiles primes is that there are no new modular forms associated to~$\overline{\rho}$ at level~$X_0(Q)$, and hence~$M$ can \emph{also}
be interpreted as a space of modular forms for~$X_0(Q)$.
There is a Galois cover~$X_1(Q) \rightarrow X_0(Q)$ with Galois group~$(\mathbf{Z}/Q \mathbf{Z})^{\times}$, and hence an intermediate cover~$X_H(Q) \rightarrow X_0(Q)$
with Galois group~$\Delta_N = (\mathbf{Z}/p^N \mathbf{Z})^d$ acting via diamond operators. The space~$M_Q$ is essentially a localization of a certain space of modular forms for~$X_H(Q)$
(with some care taken at the Hecke operators for primes dividing~$Q$).
Since the cohomology of modular curves (localized at the maximal ideal corresponding to~$\mathfrak{m}$)
is concentrated in degree one, the module~$M_Q$ turns out to be free over an auxiliary ring~$S_N = W(k)[\Delta_N]$ of diamond operators,
and the quotient~$M_Q/\mathfrak{a}_Q$ for the augmentation ideal~$\mathfrak{a}_Q$ of~$S_N$ is isomorphic to~$M$. It follows that~$\mathbf{T}_Q/\mathfrak{a}_Q = \mathbf{T}$.
\item The diamond operators have an interpretation on the Galois deformation side, and there is a identification~$R_Q/\mathfrak{a}_Q = R$ where~$R_Q$ and~$\mathbf{T}_Q$ can be viewed compatibly as~$S_{N}$-modules.
\end{enumerate}
\item Finally, one ``patches'' these constructions together for larger and larger~$Q$. This is somewhat counterintuitive, since for different~$Q$
the Galois representations involved are not compatible. However, one forgets the Galois representations and only remembers the structures relative
to both the diamond operators~$S_N$ and~$R_{\infty}$, giving the data of a surjection
$$R_{\infty} \rightarrow \mathbf{T}_{\infty}$$
with a compatible action of~$S_{\infty} = \projlim S_N \simeq W(k)\llbracket t_1, \ldots, t_d \rrbracket$.
Using the fact that~$\mathbf{T}_{\infty}$ is free of finite rank over~$S_{\infty}$,
and that~$R_{\infty}$ and~$S_{\infty}$ are formally smooth of the same dimension, one deduces that~$R_{\infty}
=\mathbf{T}_{\infty}$ and then~$R = \mathbf{T}$ after quotienting out by the augmentation ideal of~$S_{\infty}$.
\end{enumerate}
\section{The early years}
\subsection{The work of Diamond and Fujiwara}
Wiles
made essential uses of multiplicity one theorems in order to deduce that~$M_Q$ was free over~$\mathbf{T}_{Q}$.
Diamond~\cite{DiamondMult} and Fujiwara~\cite{Fujiwara} (independently)
had the key insight that one could instead patch the modules~$M_Q$ directly --- and then argue directly with the resulting object~$M_{\infty}$ as a module over~$R_{\infty}$ which was also
free over~$S_{\infty}$.
Using the fact that~$R_{\infty}$ is formally smooth, this allowed
one to deduce \emph{a posteriori} that~$M_{\infty}$ was free over~$R_{\infty}$
using the Auslander--Buchsbaum
formula~\cite{ABF}. This not only removed the necessity of
proving difficult multiplicity one results but gave new proofs of these
results\footnote{There is an intriguing result of Brochard~\cite{Brochard} which weakens the hypotheses of Diamond's freeness criterion even further,
although this idea has not yet been fully exploited.}
which could then be generalized to situations where the known methods
(often using the~$q$-expansion principle) were unavailable\footnote{The history of the subject involves difficult theorems in the arithmetic geometry
of Shimura varieties being replaced by insights from commutative algebra, paving the way to generalizations where further insights from the arithmetic
geometry of Shimura varieties are required.}. Diamond had the following to say about how he came up with the idea to patch
modules rather than use multiplicity one theorems:
\smallskip
\begin{footnotesize}
My vague memory is that I was writing down examples of ring homomorphisms and modules, subject to some constraints imposed by a Taylor--Wiles setup, and I couldn't break ``$M$ free over the group ring implies~$M$ free over~$R$.'' (I still have the notebook with the calculations somewhere, mostly done during a short trip with some friends to Portugal.) I didn't know \emph{what} commutative algebra statement I needed, but I knew I needed to learn more commutative algebra and found my way to Bruns and Herzog's ``Cohen-Macaulay Rings'' \cite{MR1251956} (back in the library in Cambridge UK by then). When I saw the statement of Auslander--Buchsbaum, it just clicked.
\end{footnotesize}
\smallskip
Diamond made a second improvement~\cite{MR1405946,MR1638490} dealing with primes away from~$p$ in situations where the corresponding minimal
local deformation problem was not controlled by the Serre level~$N(\overline{\rho})$ alone.
\subsection{Integral \texorpdfstring{$p$}{p}-adic Hodge Theory, part I: Conrad--Diamond--Taylor} \label{CDT}
One early goal after Fermat was the resolution of the full Taniyama--Shimura conjecture, namely,
the modularity of all elliptic curves over~$\mathbf{Q}$.
After the improvements of Diamond, the key remaining problem was understanding
deformation rings associated to local Galois representations at~$p$ coming from elliptic curves with bad reduction at~$p$.
Since Wiles' method (via Langlands--Tunnell~\cite{MR574808,MR621884})
was ultimately reliant on working with the prime~$p = 3$, this meant understanding deformations at~$p$ of level~$p^2$
and level~$p^3$, since any elliptic curve over~$\mathbf{Q}$ has a twist such that the largest power of~$3$ dividing the conductor is at most~$27$.
Ramakrishna in his thesis~\cite{RamakrishnaThesis} had studied the local deformation problem for finite flat representations (the case when~$(N,p) = 1$) and
proved that the corresponding local deformation rings were formally smooth.
The case when~$p$ exactly divides~$N$ was subsumed into the ordinary case, also treated by Wiles.
In level~$p^2$, one can show that the Galois representations associated to the relevant
modular forms\footnote{This is not true for all modular forms of level~$p^2$ and weight~$2$,
but only for those whose conductor at~$p$ remains divisible by~$p^2$ after any quadratic twist.}
of level~$p^2$ become finite flat
after
passing to a finite extension~$L/\mathbf{Q}_p$ with ramification degree~$e \le p- 1$.
In this range, Conrad~\cite{Conrad1,Conrad2} was able to adapt ideas of Fontaine~\cite{Fontaine} to give an equivalence between the local Galois
deformations (assuming~$\overline{\rho} |_{G_{\mathbf{Q}_p}}$ was irreducible) and linear algebra data. In particular,
as in the work of Ramakrishna, one can show that the relevant local deformation rings are formally smooth, and so
Conrad, Diamond, and Taylor were able to adapt
the Taylor--Wiles method to this setting~\cite{CDT}.
\subsection{Integral \texorpdfstring{$p$}{p}-adic Hodge Theory, part II: Breuil--Conrad--Diamond--Taylor} \label{BCDT}
A central technical ingredient in all of the arguments so far has been some use of integral $p$-adic Hodge Theory,
and in particular the theory of finite flat group schemes and Barsotti--Tate groups as developed by Fontaine and others.
All integral versions of this theory required a hypothesis on either the weight or the ramification index~$e$ relative to the bound~$p-1$.
However, around this time, Christophe Breuil made a
breakthrough\footnote{Much of the development of integral~$p$-adic Hodge theory over the last~$20$ years since~\cite{BCDT}
has been inspired by its use in the Taylor--Wiles method. However, the timing of Breuil's work was more of a happy coincidence,
although Breuil was certainly aware of the fact that a computable theory of finite flat group schemes over highly ramified bases could well have
implications in the Langlands program.}
by finding a new way to understand the integral theory
of finite flat group schemes over arbitrarily ramified bases~\cite{BreuilAnnals}. This was just the technical tool required
to push the methods of~\cite{CDT} to level~$p^3$.
Using these results, Breuil, Conrad, Diamond and Taylor~\cite{BCDT} were able to show that enough suitably chosen local deformation
rings were formally smooth to prove the modularity of all elliptic curves.
\subsection{Higher weights, totally real fields, and base change}
Many of the methods which worked for modular forms were directly adaptable to the case both of general rank~$2$
motives over~$\mathbf{Q}$ with distinct Hodge--Tate weights (corresponding to modular forms of weight~$k \ge 2$ rather than~$k = 2$)
and also to such motives over totally real fields (which are related to Hilbert modular forms), see
in particular the work of Fujiwara~\cite{Fujiwara} (and more recently Frietas--Le Hung--Siksek~\cite{BAOBAO}).
Another very useful innovation was a base change idea of Skinner--Wiles~\cite{SWBC}
which circumvented the need to rely on Ribet's level
lowering theorem. The use of cyclic base change (\cite{MR574808} in this case and~\cite{MR1007299} in general) subsequently became a standard tool in the subject.
For example, it meant that one could always reduce to a situation where the ramification at all primes~$v \nmid p$ was unipotent.
The paper~\cite{SWBC} was
related to a more ambitious plan by Wiles to prove modularity for all totally real fields:
\smallskip
\begin{footnotesize}
After Fermat I started to work with Taylor and then Diamond on the general case but decided very soon that I would rather try to do the totally real case for~$\mathrm{GL}(2)$.
I think this was while I was getting back into other kinds of problems but I thought I should still earn my bread and butter. One lunch time at the IAS in 1996 Florian Pop spoke to me and explained to me about finding points over fields totally split at some primes (e.g. real places) as he had written a paper~\cite{POP} about this with some others.
Was this any use for the Tate--Shavarevich group? I immediately saw that whether or not it was any use for TS (I doubted it) it should certainly give potential modularity. This gave some kinds of lifting so I worked on the other half (i.e. descent) thinking that just needed a similar insight. At some point I suggested to Chris that we try to do Ribet's theorem using cyclic base change as that would be useful in proving modularity and was buying time while I waited to get the right idea. Unfortunately I completely misjudged the difficulty of descent and the problem is still there. I think it is both much harder than I thought and also more important. I hope still to prove it! Of course Taylor found potential modularity and then, what I had assumed was much harder, a way to think about~$\mathrm{GL}(n)$.
\end{footnotesize}
\section{Reducible representations: Skinner--Wiles}
\label{section:SW}
One of the key hypotheses in the Taylor--Wiles method concerns restrictions
on the representation~$\overline{\rho}$, in particular the hypothesis that~$\overline{\rho} |_{G_{\mathbf{Q}(\zeta_p)}}$
is absolutely irreducible. In~\cite{SW1,SW2,SW3}, Skinner and Wiles introduced
a new argument in which this hypothesis was relaxed, at least assuming
the representations were ordinary at~$p$. In the ordinary setting, one can replace the
rings~$R$ and~$\mathbf{T}$ (which in the original setting are finite over~$W(k)$) by rings which
are finite (and typically flat) over Iwasawa algebras~$\Lambda = W(k) \llbracket (\mathbf{Z}_p)^d \rrbracket$ for some~$d$
which arise as weight spaces,
the point being that the ordinary deformations of varying weight admit a good integral theory.
The first innovation (in part) involves making a base
change so that the reducible locus is (relatively) ``small,'' (measured in terms of the
codimension over~$\Lambda$). The second idea is then to apply a variant of the Taylor--Wiles
method to representations~$\varrho: G_F \rightarrow \mathrm{GL}_2(\mathbf{T}/\mathfrak{p})$
for non-maximal prime ideals~$\mathfrak{p}$ of~$F$\footnote{Representations~$\varrho$
to infinite quotients~$\mathbf{T}/\mathfrak{p}$ had also arisen in Wiles' paper on
Galois representations associated to ordinary modular forms~\cite{WilesOrdinary}
where the concept of pseudo-deformation was also first introduced.}. Wiles again:
\smallskip
\begin{footnotesize}
We had worked out a few cases we could do without big Hecke rings in some other papers and I would say it was more a feat of stamina and technique to work through it. Of course the use of these primes was much more general and systematic than anything that went before. There is also an amusing point in this paper where we use a result from commutative algebra. It seemed crucial then though I don't know if it still is. This is proposition A.1 of Raynaud~\cite{MR0407021}. I had thought at some point during the work on Fermat that this result might be needed and had asked Michel Raynaud about it. He said he would think about it. A week later he came back to me, somewhat embarrassed that he had not known right away, to say that it was a result in his wife's thesis. So the reference to M.Raynaud is actually to his wife, Mich\`{e}le Raynaud, though he gave the reference.
\end{footnotesize}
Allen~\cite{Allen} was later able to adapt these arguments to the~$p=2$ dihedral case, which
(in a certain sense) realized the original
desire\footnote{As far as primary historical sources go, the introduction of Wiles' paper~\cite{W}
is certainly worth reading.} of Wiles to work at the prime~$p=2$.
\section{The Artin Conjecture}
\label{section:artin}
While the approach of~\cite{W,TW} applied (in principle) to all Galois representations
associated to modular forms of weight~$k \ge 2$, the case of modular forms of weight~\mbox{$k = 1$}
is qualitatively quite different (see also~\S\ref{section:calegari}).
It was therefore quite surprising when Buzzard--Taylor~\cite{MR1709306}
proved weight one modularity lifting theorems for odd continuous representations~$\rho: G_{\mathbf{Q}}
\rightarrow \mathrm{GL}_2(\overline{\Q}_p)$ which were unramified at~$p$.
Using this, Buzzard--Dickinson--Shepherd-Barron--Taylor~\cite{MR1845181} proved the Artin conjecture for a positive proportion of all odd~$A_5$
representations, which had previously only been known in a finite number of cases\footnote{In a computational \emph{tour de force} for the time, Buhler~\cite{Buhler} in his thesis
had previously established the modularity of an explicit odd projective~$A_5$ representation of conductor~$800$.} up to twist.
Standard ordinary modularity theorems showed the existence of ordinary modular forms
associated to such representations~$\rho$ --- however, the classicality theorems of
Hida~\cite{MR848685} do not apply (and are not true!) in weight one.
The main idea of~\cite{MR1709306} was to exploit the fact that~$\rho$ is unramified to construct
\emph{two} ordinary modular forms each corresponding to a choice
of eigenvalue of~$\rho(\mathrm{Frob}_p)$ assuming these eigenvalues are
distinct\footnote{This argument can be modified to deal with the case when
the eigenvalues of~$\overline{\rho}(\mathrm{Frob}_p)$ coincide by modifying~$R$ and~$\mathbf{T}$
to include operators corresponding (on the Hecke side) to~$U_p$.
Geraghty and I discovered an integral version of this idea ourselves (``doubling,'' following Wiese's paper~\cite{Wiese})
during the process of writing~\cite{CG}, although it turned out that, at least in characteristic zero, Taylor already
had the idea in his back pocket in the early 2000s.}.
One then has to argue~\cite{MR1709306} that these two ordinary forms are the oldforms associated
to a classical eigenform of weight one, which one can do by exploiting
both the rigid geometry of modular curves and the~$q$-expansion principle.
Although the original version of this argument required a number of improvements
to the usual Taylor--Wiles method (Dickinson overcame some technical issues when~$p=2$~\cite{MR1845182} and
Shepherd-Barron--Taylor proved some new cases of Serre's conjecture for~$\mathrm{SL}_2(\mathbf{F}_4)$ and~$\mathrm{SL}_2(\mathbf{F}_5)$-representations
in~\cite{MR1491977}), it was ripe
for generalization to totally real fields\footnote{The proof all that finite odd~$2$-dimensional representations over~$\mathbf{Q}$ are modular
was completed by Khare and Wintenberger
as a consequence of their proof of Serre's conjecture, see~\S\ref{sec:Serre}.}
After a key early improvement by
Kassaei~\cite{MR2219265}, the~$n = 2$ Artin conjecture for totally real fields is now
completely resolved under the additional assumption that the representation is
odd by a number of authors, including Kassaei--Sasaki--Tian and Pilloni--Stroh~\cite{MR3581174,MR2983010,MR3294624,SasakiInvent,MR3639600,MR3581178}.
On the other hand, the reliance on~$q$-expansions in this
argument
has proved
an obstruction to extending this to other
groups.
(See also~\S\ref{section:GSp4}).
\section{Potential Modularity}
\label{section:potential} \label{section:pq}
One new idea which emerged in Taylor's paper~\cite{MR1954941} was the concept of \emph{potential modularity}. Starting with a representation~$\rho: G_F \rightarrow \mathrm{GL}_2(\overline{\Q}_p)$ for a totally
real field~$F$,
one could sidestep the (difficult) problem of proving the modularity of~$\overline{\rho}$ by proving
it was modular over some finite totally real extension~$F'/F$. In the original paper~\cite{W},
Wiles employed a~$3$-$5$ switch to deduce the modularity of certain mod~$5$ representations
from the modularity of mod~$3$ representations. More generally,
one can prove the modularity of a mod~$p$ representation~$\overline{\rho}_p$
from the modularity of a mod~$q$ representation~$\overline{\rho}_q$ if one can find both of them
occurring as the residual representation of a compatible family where the Taylor--Wiles hypotheses
apply to~$\overline{\rho}_q$. For example, if~$\overline{\rho}_p$ and~$\overline{\rho}_q$ are
representations valued in~$\mathrm{GL}_2(\mathbf{F}_p)$ and~$\mathrm{GL}_2(\mathbf{F}_q)$ respectively, one can try to
find the compatible family by finding an elliptic curve with a given mod~$p$ and mod~$q$ representation.
The obstruction to doing such a~$p$-$q$ switch over~$F$
is
that the corresponding moduli spaces (which in this case are twists of the modular curve~$X(pq)$) are not in general rational, and hence have no
reason to admit rational points. However, exploiting an idea
due to Moret-Bailly~\cite{MB}, Taylor showed that these moduli spaces at least had many points over totally
real fields where one could additionally ensure that the Taylor--Wiles hypothesis applies at the prime~$q$.
At the cost of proving a weaker result, this gives a huge amount of extra flexibility that has proved
remarkably useful. Taylor's first application of this idea was to prove the Fontaine--Mazur
conjecture for many~$2$-dimensional representations, since the potential modularity of these representations
was enough to prove (for example) that they come from compatible families of Galois representations
(even over the original field~$F$!), and that they satisfy purity (which is known for Hilbert modular
forms of regular weight). The concept of potential modularity, however, has proved crucial for other applications, not least of which is the proof of the Sato--Tate conjecture (see~\S\ref{section:satotate}).
\section{The work of Kisin}
A key ingredient in the work of Breuil--Conrad--Diamond--Taylor (\S\ref{CDT},\S\ref{BCDT})
(and subsequent work of Savitt~\cite{MR2004122,Savitt})
was the fact that a certain local deformation ring~$R^{\mathrm{fl}}$ defined in terms of integral~$p$-adic
Hodge theory was formally smooth. The calculations of~\cite{BCDT,Savitt}, however, applied only to some
(very) carefully chosen situations sufficient for elliptic curves but certainly not for all~$2$-dimensional
representations. In the 2000s, Kisin made a number of significant contributions, both to the understanding
of local deformation rings but also to the structure of the Taylor--Wiles argument itself~\cite{Kisin1,Kisin2,Kisin3,Kisin4,Kisin5,Kisin6}.
\subsection{Local deformation rings at \texorpdfstring{$v = p$}{v eq p}}
One difficulty with understanding local deformation rings~$R^{\mathrm{fl}}$ associated to finite flat group schemes
over highly ramified bases is that the group schemes themselves are not uniquely defined by their generic fibres.
Kisin~\cite{Kisin4} had the idea that one could also define the moduli space of the group schemes
themselves, giving a projective resolution~$\mathscr{G}\kern-0.1em{\mathscr{R}} \rightarrow \mathrm{Spec}(R^{\mathrm{fl}})$ (this map is
an isomorphism after inverting~$p$).
Kisin further realized that the geometry of~$\mathscr{G}\kern-0.1em{\mathscr{R}}$ was related to local models
of Shimura varieties, for which one had other available techniques to analyze their structure and singularities.
Later, Kisin was also able~\cite{Kisin6} to construct local deformation rings~$R$ capturing deformations
of a fixed local representation~$\overline{\rho}$ which become semi-stable over a fixed extension~$L/\mathbf{Q}_p$
and had Hodge--Tate weights in any fixed finite range~$[a,b]$, absent a complete integral theory of such representations.
(There are also are constructions where one fixes the inertial type of the corresponding representation.)
Kisin further proved that the generic fibres of these rings were indeed of the expected dimension and often
formally smooth.
\subsection{Kisin's modification of Taylor--Wiles} \label{section:kisin}
Beyond analyzing the local deformation rings themselves, Kisin crucially found a way~\cite{Kisin4} to modify the Taylor--Wiles method to
avoid the requirement that these rings are formally smooth, thus greatly expanding the scope of the method.
First of all, Kisin reimagined the global deformation ring~$R$ as an algebra
over a (completed tensor product)
\begin{equation*}
R^{\mathrm{loc}} = \widehat{\bigotimes}_{v \in S} R_v
\end{equation*}
of local deformation rings~$R_v$ for sets of places~$v \in S$, in particular including the
prime~$p$\footnote{Since the local residual representations are typically reducible,
Kisin also introduced the notion of \emph{framed} deformation rings which are always well-defined,
and which (properly taking into account the extra variables) are compatible with the Taylor--Wiles
argument.}.
Now, after a Taylor--Wiles
patching argument, one constructs a big module~$M_{\infty}$ over~$R_{\infty}$
(and free over the auxiliary ring of diamond operators~$S_{\infty}$)
but where~$R_{\infty}$ is no longer a power series ring over~$W(k)$
but a power series ring over~$R^{\mathrm{loc}}$. If the algebras~$R_v$ for~$v \in S$ are themselves power series rings, one is reduced
precisely to the original Taylor--Wiles setting as modified by Diamond. On the other hand, if the~$R_v$ are (for example)
not power series rings but are
integral domains over~$W(k)$ of the expected dimension,
then Kisin explained how one could still deduce that~$M[1/p]$ was a faithful~$R[1/p]$-module, which proves
that~$R[1/p]=\mathbf{T}[1/p]$ and suffices for applications to modularity.
More generally, assuming only that the~$R_v$ are flat over~$W(k)$ and that the generic fibre~$R_v[1/p]$ is
equidimensional of the expected dimension, the modularity of any point of~$R$
reduces to showing that there is at least one modular point which
lies on the same component of~$R_v[1/p]$\footnote{There are some subtleties to understanding~$R[1/p]$ for complete local Noetherian~$W(k)$-algebras
that are not obvious on first consideration. The first and most obvious blunder to avoid is to recognize that~$R[1/p]$
is usually far from being a local ring. Similarly, the ring~$R[1/p]$ can be regular
and still have multiple components,
as can be seen in an example as simple as~$R=\mathbf{Z}_p\llbracket X \rrbracket/X(X-p)$.
Perhaps more importantly, however, the ring~$R[1/p]$ ``behaves'' in some important ways like a finitely generated algebra over a field.}.
In the original modularity lifting arguments, one treated the minimal case first and then deduced the non-minimal cases
using a subtle commutative algebra criterion which detected isomorphisms between complete intersections.
From the perspective of Kisin's modification, all that is required is to show that there exists a single modular point
with the right non-minimal local properties. In either case, both Wiles and Kisin used Ihara's Lemma
to establish the
existence of congruences between old and new forms, but Kisin's argument is much softer and thus more generalizable
to other situations\footnote{In particular, Wiles' numerical criterion~\cite[Thm~5.3]{MR1605752} relies on certain rings being complete
intersections, and Kisin's local deformation rings are not complete intersections (or even Gorenstein) in general --- see~\cite{Snowden}.}.
Kisin had the following to say about his thought process:
\smallskip
\begin{footnotesize}
The idea of thinking of~$R$ as an~$R^{\mathrm{loc}}$ algebra just popped into my head, after I'd been thinking about the Wiles--Poitou--Tate formula, and how it fit into the Taylor--Wiles patching argument. This was in Germany, I think in 2002. I had the idea about moduli of finite flat group schemes in the Fall of 2003, after I arrived in Chicago.
It was entirely motivated by modularity. I had been trying to compute these deformation rings, by looking at deformations of finite flat group schemes. For~$e<p-1$, the finite flat model is unique, so I knew this gave the deformation ring in this case; this already gave some new cases. However I was stuck about the meaning of these calculations in general for quite some time. At some point I thought I'd better write up what I had, but as soon as I started thinking about that --- within a day --- I realized what the correct picture was with the families of finite flat group schemes resolving the deformation ring. I already knew about Breuil's unpublished note~\cite{Breuilnote}, and quite quickly was able to prove the picture was correct. It was remarkable that prior to coming to Chicago, I didn't even know the definition of the affine Grassmannian, but within a few months of arriving, it actually showed up in my own work.
To me the whole project was incredibly instructive. If I had known more about what was (thought to be) essential in the Taylor--Wiles method, I never would have started the project. Not having fixed ideas gave me time to build up intuition. I also should have gotten the idea about moduli of finite flat group schemes much sooner if I'd been more attentive to what the geometry was trying to tell me.
\end{footnotesize}
\vspace{-1em}
\section{\texorpdfstring{$p$}{p}-adic local Langlands}
\subsection{The Breuil--M\'{e}zard conjecture}
Prior to Kisin's work, Breuil and M\'{e}zard~\cite{BM} undertook a study of certain low weight potentially semi-stable
deformation rings, motivated by~\cite{BCDT}. They discovered (in part conjecturally) a crucial link between the geometry of these Galois deformation rings
(in particular, the Hilbert--Samuel multiplicities of their special fibres) with the mod-$p$ reductions (and corresponding
irreducible constituents) of lattices inside locally algebraic~$p$-adic representations of~$\mathrm{GL}_2(\mathbf{Z}_p)$.
In the subsequent papers~\cite{Breuil1,Breuil2}, Breuil
raised the hope that there could exist a~$p$-adic Langlands correspondence relating certain mod-$p$ (or~$p$-adic Banach space)
representations of~$\mathrm{GL}_2(\mathbf{Q}_p)$
to geometric~$2$-dimensional~$p$-adic representations of~$G_{\mathbf{Q}_p}$\footnote{The starting observation~\cite{BreuilIntro} is as follows: if~$\pi = \bigotimes' \pi_v$
is the automorphic representation associated to a modular form~$f$, then~$\pi_v$ determines (and is determined by)~$\rho_f |_{G_{\mathbf{Q}_v}}$ for all~$v \ne p$ (at least
up to Frobenius semi-simplification). On the other hand, $\pi_p$ does not determine the~$p$-adic representation~$\rho_f |_{G_{\mathbf{Q}_p}}$
(except in the exceptional setting where~$\pi_p$ is spherical and~$a_p$ is not a~$p$-adic unit), raising the question of what extra~$\mathrm{GL}_2(\mathbf{Q}_p)$ structure associated
to~$f$ should
determine (and be determined by)~$\rho_f |_{G_{\mathbf{Q}_p}}$.}. Breuil recounts the origins of these conjectures as follows:
\medskip
{\footnotesize
The precise moment I became 100\% sure that there would be a non-trivial~$p$-adic correspondence for~$\mathrm{GL}_2(\mathbf{Q}_p)$ was in the computations of~\cite{Breuil2}. In these computations, I reduced mod~$p$ certain~$\overline{\mathbf{Z}}_p$--lattices in certain locally algebraic representations of~$\mathrm{GL}_2(\mathbf{Q}_p)$, and at some point, I found out that this reduction mod~$p$ had a really nice behaviour, so nice that clearly, it was predicting (via the mod-$p$ correspondence) what the reduction mod-$p$ would be on the~$\mathrm{Gal}(\overline{\Q}_p/\mathbf{Q}_p)$-side.
}
These ideas were further developed by Colmez in~\cite{MR2493219,MR2642407,MR2642409} amongst other papers\footnote{In~\cite{KisinEarly}, Kisin had shown
that the~$p$-adic representations~$V$ associated to non-classical finite slope overconvergent modular forms with~$U_p$-eigenvalue~$a_p$ satisfied~$\dim D_{\mathrm{cris}}(V) = 1$, and moreover that crystalline
Frobenius acted on this space by~$a_p$.
(This
paper was itself apparently motived by the goal of \emph{disproving} the Fontaine--Mazur conjecture!)
On the way to the~$2004$ Durham symposia on~$L$-functions and Galois representations, Fontaine
raised the question to Colmez to
what extent this determined the corresponding Galois representation.
Colmez worked out the answer the evening before his talk
and incorporated it into his lecture the following day, ultimately leading to the notion of trianguline representations~\cite{MR2493219}.}:
Colmez studied various Banach space completions defined by Breuil and proved they
were non-zero using the theory of~$(\varphi,\Gamma)$-modules.
Since the theory of~$(\varphi,\Gamma)$-modules applies to all Galois representations and not
just potentially semi-stable ones, this led Colmez to propose a~$p$-adic local Langlands correspondence for
\emph{arbitrary}~$2$-dimensional representations~$G_{\mathbf{Q}_p} \rightarrow \mathrm{GL}_2(E)$, and he was ultimately able to construct a functor
from suitable~$\mathrm{GL}_2(\mathbf{Q}_p)$-representations to Galois representations of~$G_{\mathbf{Q}_p}$.
Colmez gave a talk on his construction at a conference in Montreal in September 2005. At the same conference, Kisin gave a talk
presenting a proof of the Breuil--M\'{e}zard conjecture by relating it directly to~$R=\mathbf{T}$ theorems and the Fontaine--Mazur conjecture for
odd~$2$-dimensional
representations of~$G_{\mathbf{Q}}$ with distinct Hodge--Tate weights. While Kisin's argument exploited results of Berger--Breuil~\cite{BB} and Colmez, it was realized
by the key participants (perhaps in real time) that Colmez' $p$-adic local Langlands correspondence should be viewed as taking place over the entire local deformation ring.
Subsequently Colmez was able to construct the inverse functor\footnote{To add some further confusion to the historical chain of events,
the published version of~\cite{KisinFM} incorporates some of these subsequent developments. Note also that the
current state of affairs is that the proof of the full~$p$-adic local Langlands correspondence for~$\mathrm{GL}_2(\mathbf{Q}_p)$ (for example as proved in~\cite{MR3272011}
but see also~\cite[Remarque VI.6.51]{MR2642409})
still relies on the global methods of~\cite{emerton2010local}, which in turn relies on~\cite{MR2642409}.
These mutual dependencies, however, are not circular! The difficulty arises in the supercuspidal case. One philosophical reason that global methods
are useful here is that all global representations are yoked together by an object (the completed cohomology group~$\widetilde{H}^1(\mathbf{Z}_p)$) with
good finiteness properties. One can then exploit the fact
that crystabeline representations (for which the~$p$-adic local Langlands correspondence is known by~\cite{MR2642409})
are Zariski dense inside unrestricted global deformation rings
(\cite[Theorem~1.2.3]{emerton2010local}, using arguments going back to B\"{o}ckle~\cite{MR1854117}).}.
Colmez writes:
\medskip
{\footnotesize
I received a paper of Breuil (a former version of~\cite{MR2667887}) during my stay at the Tata Institute in December 2003---January 2004. In December,
I was spending Christmas under Goa's palm trees with my daughter when Breuil's paper arrived in my email. That paper contained a conjecture (in the semi-stable case) that I was sure I could prove using~$(\varphi,\Gamma)$-modules (if it was true\ldots). I spent January 2004 working on it and after 15 days of computations in the dark, I finally found a meaning to some part of a painful formula (you can find some shadow of all of this in (iii) of Remark 0.5 of my unpublished~\cite{ColmezUnpublished}). By the end of the month, I was confident that the conjecture was proved and I told so to Breuil who adapted the computations to the crystalline case, and wrote them down with the help of Berger (which developed into~\cite{BB}). (One thing that makes computations easier and more conceptual in the crystalline case is that you end up with the universal completion of the locally algebraic representation you start with; something that is crucial in Matthew [Emerton]'s proof of the FM conjecture.) Durham was in August of that year and Berger--Breuil had notes from a course they had given in China~\cite{BBnotes}.
Those notes were instrumental in my dealing with trianguline representations at Durham (actually, I did some small computation and the theory just developed by itself during the night before my talk which was supposed to be on something else\ldots I think I came up with the concept of trianguline representations later, to justify the computations, I don't remember what language I used in my talk which had some part on Banach--Colmez spaces as far as I can remember.
}
\subsection{Local--global compatibility for completed cohomology}
From a different perspective, Emerton had introduced the completed cohomology groups~\cite{MR2207783}
as an alternative means for constructing the Coleman--Mazur eigencurve~\cite{CM}. Inspired by Breuil's work, Emerton
formulated~\cite{EmertonOne}
a local--global compatibility conjecture for completed cohomology in the language of the then nascent~$p$-adic Langlands correspondence.
After the construction of the correspondence for~$\mathrm{GL}_2(\mathbf{Q}_p)$ by Colmez and Kisin, Emerton was able to prove most of his conjecture, leading to a new proof of (many cases of) the Fontaine--Mazur conjecture.
The results of Kisin~\cite{KisinFM} and Emerton fell short of proving the full version of this conjecture for two reasons. The first was related to some technical
issues with the~$p$-adic local Langlands correspondence, both at the primes~$p=2$ and~$3$ but also when the residual representation
locally had the shape~$1 \oplus \overline{\varepsilon}$ for the cyclotomic character~$\varepsilon$. (The local issues have now more
or less all been resolved~\cite{MR3272011}. The most general global results for~$p=2$ are currently due to Tung~\cite{MR4257080}.)
A second restriction was the Taylor--Wiles
hypothesis that~$\overline{\rho}$ was irreducible. Over the intervening years, a number of key improvements to the local story have been found,
in particular by Colmez, Dospinescu, Hu, and Pa{\v{s}}k{\=u}nas~\cite{MR3150248,MR3272011,PH}.
Very recently, Lue Pan~\cite{Pan1} found a way to marry techniques from Skinner--Wiles in the reducible case (\S\ref{section:SW}) to techniques from~$p$-adic local Langlands to
completely prove the modularity (up to twist) of any geometric representation~$\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\overline{\Q}_p)$ for~$p \ge 5$ only assuming the hypotheses
that~$\rho$ has distinct Hodge--Tate weights and that~$\rho$ is odd\footnote{The assumption on the Hodge--Tate weights is almost certainly removable
using recent progress on the ideas discussed in~\S\ref{section:artin} (Sasaki has announced such a result).
Moreover, Pan has found a different
approach to this case as well, see~\cite[Theorem 1.0.5]{Pan2} and the subsequent comments.
The hypothesis that~$\rho$ is odd is more troublesome --- see~\S\ref{section:even}.}.
\section{Serre's Conjecture} \label{sec:Serre}
In Wiles' original lectures in Cambridge in 1993, he introduced
his method with the statement that it was~\emph{orthogonal} to Serre's conjecture~\cite{MR1322785}.
In some senses, this viewpoint turned out to be the opposite of prophetic,
in that the ultimate resolution of Serre's conjecture used the Taylor--Wiles method as its central core.
The proof of Serre's conjecture by Khare and Wintenberger~\cite{SerreKW1,SerreKW2,KW1,KW2}
introduced a new technique
for lifting residual Galois representations to characteristic zero (see~\S\ref{section:KW}) which
has proved very useful for subsequent modularity lifting theorems.
\subsection{Ramakrishna lifting}
Ramakrishna, in a series of papers in the late 90s~\cite{RK1,RK2}, studied the question of lifting an odd Galois representation
\begin{equation*}
\overline{\rho}:G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\overline{\mathbf{F}}_p)
\end{equation*}
to a global potentially semistable representation in characteristic zero unramified outside finitely many primes. This is a trivial consequence
of Serre's conjecture\footnote{Trivial only assuming the results of Tsuji~\cite{Tsuji} and Saito~\cite{Saito}, of course.}
but is highly non-obvious without such an assumption. Ramakrishna succeeded in proving the existence of lifts by an ingenious
argument involving adding auxiliary primes and modifying the local deformation problem to a setting where there all global obstructions
vanished. The resulting lifts had the added property that they were valued in~$\mathrm{GL}_2(W(k))$ whenever~$\overline{\rho}$ was valued in~$\mathrm{GL}_2(k)$.
Adaptations of Ramakrishna's method had a number of important applications even under the assumption of residual modularity, including in~\cite{CHT} where it was used to produce
characteristic zero lifts with Steinberg conditions at some auxiliary primes.
There is also recent work of Fakhruddin, Khare, and Patrikis~\cite{FKP} which considerably
extends these results in a number of directions.
\subsection{The Khare--Wintenberger method} \label{section:KW}
One disadvantage of Ramakrishna's method was that it required allowing auxiliary ramification which (assuming Serre's
conjecture) should not be necessary\footnote{If one insists on finding a lift is valued in~$\mathrm{GL}_2(W(k))$ rather than~$\mathrm{GL}_2(\mathcal{O}_E)$
for some ramified extension~$E/W(k)[1/p]$, then some auxiliary ramification is necessary in general, at least in fixed weight.}.
Khare and Wintenberger found a new powerful method for avoiding this. The starting point is the idea
that, given an odd representation~$\overline{\rho}: G_F \rightarrow \mathrm{GL}_2(\overline{\mathbf{F}}_p)$ for a totally real field~$F$ satisfying the Taylor--Wiles hypotheses,
one could find a finite extension~$H/F$ where~$\overline{\rho}$ is modular (exactly as in~\S\ref{section:potential}). Then, using an~$R=\mathbf{T}$ theorem over~$H$,
one proves that the corresponding deformation ring~$R_H$ of~$\overline{\rho} |_{G_H}$ is finite over~$W(k)$. However, for formal reasons, there is a map~$R_H \rightarrow R_F$
(where~$R$ is the deformation ring corresponding to the original representation~$\overline{\rho}$) which is a finite morphism, and hence the ring~$R_F/p$ is Artinian.
Then, by Galois cohomological arguments, one proves the ring~$R_F$ has dimension at least one, from which one deduces that~$R_F$ has~$\overline{\Q}_p$-valued points.
Even more can be extracted from this argument, however --- the~$\overline{\Q}_p$-valued point of~$R_F$ certainly comes from a~$\overline{\Q}_p$-valued point of~$R_H$,
and hence comes from a compatible family of Galois representations over~$H$. Using the fact that one member of the family extends to~$G_F$, it can be argued that
the entire family descends to a compatible family over~$F$. This one can then hope to prove is modular by working at a different (possibly smaller) prime, where (hopefully)
one can prove the associated residual representation is modular. In this way, one can inductively
reduce Serre's conjecture~\cite{MR885783} to the case~$p=2$ and~$N(\overline{\rho})=1$, where Tate had previously proved in a letter to Serre~\cite[July 2, 1973]{ColmezTate2} (also~\cite{Tate}) that all such absolutely irreducible
representations are modular by showing that no such representations exist. The entire idea is very clean, although in practice the difficulty reduces to the step
of proving modularity lifting theorems knowing either that~$\overline{\rho}$ is either modular and absolutely irreducible or is reducible. Khare and Wintenberger's
timing was such that the automorphy lifting technology was just good enough for the proof to work, although this required some extra effort at the prime~$p = 2$
(both in their own work and in a key assist by Kisin~\cite{Kisin3}).
As with Ramakrishna's method, the Khare--Wintenberger lifting method has also been systematically exploited for modularity lifting applications
(for example in~\cite{BLGGT} (see~\S\ref{section:BLGGT}) building on ideas of Gee~\cite{gee061}).
\section{Higher dimensions}
Parallel to the developments of~$p$-adic Langlands for~$n = 2$, the first steps were made to generalize the theory to higher dimensional
representations.
Unlike in the case of modular forms, substantially less was known about the existence of Galois representations until the 90s.
\subsection{Construction of Galois representations, part I: Clozel--Kottwitz}
The first general construction\footnote{Clozel's paper is from 1991 and thus not strictly ``post--Fermat'' as is the remit of this survey.
However, it can be considered a natural starting point for the ``modern'' arithmetic theory of automorphic forms for~$\mathrm{GL}(n)$ and so
it seems
reasonable to mention it here.}
of Galois representations in dimension~$n > 2$ was made by Clozel~\cite{Clozel} (see also the work of Kottwitz~\cite{Kottwitz}).
Clozel's theorem
applies to certain automorphic forms for~$\mathrm{GL}_n(\mathbf{A}_L)$ for CM fields~$L/L^{+}$.
The construction requires
three important hypotheses on~$\pi$: The first is
that~$\pi$ is conjugate self-dual, that is~$\pi^{\vee} \simeq \pi^c$. If~$\pi$ is a base change from an algebraic representation of~$L^{+}$ and~$n = 2$ then
this condition is automatic\footnote{At least after a twist which is always possible to achieve in practice,
see~\cite[Lemma~4.1.4]{CHT}.
More generally, one can work with unitary similitude groups and consider~$\pi$ with~$\pi^{\vee} \simeq \pi^c \otimes \chi$
for suitable characters~$\chi$.},
but it is far from automatic when~$n > 2$. The second condition is an assumption
on the infinitesimal character which (in the case of modular forms) is equivalent to the condition that the weight~$k$ is~$\ge 2$.
Finally, there is a technical condition that for some finite place~$x$ the representation~$\pi_x$
is square integrable. A number of improvements (particularly at the bad primes) were made by Harris--Taylor
in~\cite[Theorem~C]{ht}, and later by Taylor--Yoshida and Caraiani~\cite{TY,MR2972460,MR3272276}, bringing the theory roughly in line with that of modular forms at the time of Wiles,
and in particular primed for possible generalizations of the Taylor--Wiles method to higher dimensions.
\subsection{The Sato--Tate conjecture, part I}
\label{section:satotate}
Harris, and Taylor (as early as 1996) started the work of generalizing the Taylor--Wiles machinery to the setting of~$n$-dimensional representations.
They quickly understood that the natural generalization of these ideas in~$n$-dimensions required the hypothesis
that the Galois representations were self-dual up to a twist.
This meant that one should not consider general automorphic forms on the group~$\mathrm{GL}_n(\mathbf{A}_{\mathbf{Q}})$ but rather groups of symplectic or orthogonal type
depending on the parity of~$n$. If one replaced Galois representations over totally real fields by Galois representations over imaginary CM
fields and then further imposed the condition that the Galois representations are conjugate self-dual, the relevant automorphic forms should then come
from unitary groups.
There were two benefits of working with these hypotheses. First of all, the relevant automorphic representations for unitary groups were, as with modular forms,
associated to cohomology classes on Shimura varieties. In particular, under the assumption that there existed
an auxiliary prime~$x$ such that~$\pi_x$ was square integrable, they could
be seen inside the ``simple'' Shimura varieties of type~$U(n-1,1)$ considered by Kottwitz~\cite{Kottwitz}. On the other hand, the same
Hecke eigenclasses (if not Galois representations) also came from a compact form of the group and thus inside the cohomology
of zero-dimensional varieties\footnote{Inside~$H^0$, of course}. The advantage of working in this setting is that the freeness of~$M_Q$
over the ring of diamond operators is immediate\footnote{In more general contexts, the freeness of~$M_Q$ is closely related to the vanishing of cohomology
localized at~$\mathfrak{m}$ in all but one degree.}.
In the fundamental paper~\cite{CHT}, Clozel, Harris, and Taylor succeeded in overcoming many of the technical difficulties generalizing the arguments
of~\cite{W,TW} to these representations. Although the argument in spirit was very much the same, there are a number of points for~$\mathrm{GL}_2$
where things are much easier.
One representative example of this phenomenon is understanding Taylor--Wiles primes. While the Galois side generalizes readily, the automorphic side requires many new ideas
and some quite subtle arguments concerning the mod~$p$ structure of certain~$\mathrm{GL}_n(\mathbf{Q}_q)$-representations of conductor~$1$ and conductor~$q$.
In order to prove the Sato--Tate conjecture for a modular form~$f$, it was already observed by Langlands that it sufficed to prove the modularity
of all the symmetric powers of~$f$. However, it turns out that the weaker assumption that each of these symmetric powers is \emph{potentially modular}
suffices, and by some subterfuge only the even powers are required~\cite{Harristwist}. In order to prove potential modularity theorems, one
needs to be able to carry out some version of the~$p$-$q$ switch (\S\ref{section:pq}). In order to do \emph{this}, one needs a source
of motives which both generate Galois representations of the right shape (conjugate self dual and with distinct Hodge--Tate weights) and yet also come in positive dimensional families.
It turned out that there already existed such motives in the literature, namely, the so-called Dwork family. However, given the strength of the automorphy
lifting theorems in~\cite{CHT}, considerable effort had to be made in studying the geometry of the Dwork family to ensure that the~$p$-$q$ switch
would produce geometric Galois representations with the right local properties. These issues were precisely addressed in the companion
paper by Harris, Shepherd-Barron and Taylor~\cite{Shep}. Taken together, these papers contained all the ingredients to prove the potential modularity of higher symmetric
powers of modular forms (satisfying a technical square integrable condition at some auxiliary prime) with one exception.
As mentioned earlier, the work of Kisin had simplified the passage from the minimal case to the non-minimal case --- ``all'' that was required
was to produce congruences between the original form and forms of higher level rather than to compute a precise congruence number as in~\cite{W}.
However, even applying Kisin's approach seemed to require Ihara's Lemma, and despite several years of
effort, the authors of~\cite{CHT} were not able to
overcome this obstacle\footnote{The issue remains unresolved to this day.}. Here is Michael Harris' recollection of the process:
\medskip
{\footnotesize
In the spring of 1995, I was at Brandeis, Richard was at MIT, and I wanted
to understand the brand new proof of Fermat's Last Theorem. So I asked Richard
if he would help me learn by collaborating on modularity for higher-dimensional groups. The collaboration
took off a year later, when Richard wrote to tell me about the Diamond--Fujiwara
argument and suggested that we work out the Taylor--Wiles method for unitary groups. This
developed over the next 18 months or so into the early version of what eventually
became the IHES paper with Clozel. But it had no punch line.
I was hoping to work out
some non-trivial examples of tensor product functoriality for~$\mathrm{GL}(n)\times \mathrm{GL}(m)$, where one of the
two representations was congruent mod~$l$ to one induced from a CM Hecke character.
This would have required some numerical verification. In the meantime we got sidetracked into
proving the local Langlands conjecture~\cite{ht}.
The manuscript on automorphy lifting went through several drafts and was circulated;
you can still read it on my home page~\cite{Harrissecret}. Genestier and Tilouine~\cite{GT} quoted it when they proved
modularity lifting for Siegel modular forms. When Clozel saw the draft he told me we should try to prove
the Sato--Tate Conjecture. Although this was in line with my hope for examples of tensor product functoriality,
it seemed completely out of reach, because I saw no way to prove residual modularity of symmetric powers.
When I heard about the Skinner--Wiles paper I came up with a quixotic plan to prove symmetric power
functoriality for Eisenstein representations, using the main conjecture of Iwasawa theory to
control the growth of the deformation rings. This was in the spring of 2000, at the IHP special
semester on the Langlands program, where I first met Chris Skinner.
One day Chris told me that Richard had invented potential modularity. This led me to a
slightly less hopeless plan to prove potential symmetric power functoriality by proving it for~$2$-dimensional
representations congruent to potentially abelian representations, as in the potential modularity
argument. I told Richard about this idea, probably the day he arrived in Paris. He asked: why
apply potential modularity to the~$2$-dimensional representation; why not instead apply it to
the the symmetric power representations directly? I then replied: that would require a variation
of Hodge structures with a short list of properties: mainly, the correct~$h^{p,q}$'s and large monodromy groups.
We checked that potential modularity was sufficient for Sato--Tate. We then resolved to ask our
contacts if they knew of VHS with the required properties. The whole conversation lasted about 20 minutes.
I asked a well-known algebraic geometer, who said he did not know of any such VHS.
Richard asked Shepherd-Barron, who immediately told him about the Calabi-Yau hypersurfaces
that had played such an important role in the mirror symmetry program.
(And if my algebraic
geometer hadn't wanted to be dismissive, for whatever reason, he would have realized this as well.)
The~$h^{p,q}$'s were fine but we didn't know about the monodromy. However, Richard was staying at the
IHES, and by a happy accident so was Katz, and when Richard asked Katz about the monodromy
for this family of hypersurfaces Katz told him they were called the Dwork family and gave him the
page numbers in one of his books.
So within a week or two of our first conversation, we found ourselves needing only one more
result to complete the proof of Sato--Tate. This was Ihara's lemma, which occupied our attention over
the next five years. In the meantime, Clozel had written a manuscript on symmetric powers,
based on the reducibility mod ell of symmetric powers. The argument was incomplete but he had
several ideas that led to his joining the project, and he also hoped to use ergodic theory to prove Ihara's lemma.
In the summer of 2003 Clozel and I joined Richard in ``old'' Cambridge to try to work this out. The rest
you know. We finally released a proof conditional on Ihara's lemma in the fall of 2005. A few months
later Richard found his local deformation argument, and the proof was complete.
}
\subsection{Taylor's trick: Ihara Avoidance}
\label{section:iharaavoid}
Shortly after the preprints~\cite{CHT,MR1491977} appeared, Taylor found a way to overcome the problem
of Ihara's lemma. Inspired by Kisin's formulation of the Taylor--Wiles method (\S\ref{section:kisin}), Taylor
had the idea of comparing two global deformation rings~$R^1$ and~$R^2$. Here (for simplicity) the local deformation
problems associated to~$R^1$ and~$R^2$ are formally smooth at all but a single prime~$q$. At the prime~$q$, however,
the local deformation problem associated to~$R^1$ consists of tamely ramified representations where a generator~$\sigma$
of tame inertia has characteristic polynomial~$(X-1)^n$, and for~$R^2$ the
characteristic polynomial has the shape~$(X-\zeta_1)\ldots (X - \zeta_n)$
for some fixed distinct roots of unity~$\zeta_i \equiv 1 \bmod p$. On the automorphic side, there are two patched modules~$H_1$ and~$H_2$,
and there is an equality~$H_1/p = H_2/p$. The local deformation ring~$R^1_q$ associated to~$R^1$ at~$q$ is reducible
and has multiple components
in the generic fibre, although the components in characteristic zero are in bijection to the components in the special fibre.
On the other hand, the local deformation ring associated to~$R^2$ at~$q$ consists of a single component, and so using Kisin's
argument one deduces that~$H_2$ has full support. Now a commutative algebra argument using the identity~$H_1/p=H_2/p$
and the structure of~$R^1_q$ implies that~$H_1$ has sufficiently large support over~$R_1$, enabling one to deduce the modularity
of every~$\overline{\Q}_p$-valued point of~$R_1$\footnote{Taylor's argument proves theorems of the form~$R[1/p]^{\mathrm{red}} = \mathbf{T}[1/p]$ rather than~$R = \mathbf{T}$. This
is still perfectly sufficient for proving modularity lifting results, but not always other interesting corollaries associated to~$R=\mathbf{T}$ theorems
like finiteness of the corresponding adjoint Selmer groups (though see~\cite{PDuke,1912.11265}).}.
\subsection{The Sato--Tate conjecture, part II}
After Taylor's trick, one was \emph{almost} in a position to complete the proof of Sato--Tate for all classical modular forms.
A few more arguments were required. One was the tensor product trick due to Harris which enabled one to pass
from conjugate self-dual motives with weights in an arithmetic progression to conjugate self-dual motives with consecutive
Hodge--Tate weights by a judicious twisting argument using CM characters. A second ingredient was the analysis of
the ordinary deformation ring by Geraghty~\cite{ger}. One of the requirements of the~$p$-$q$ trick was the condition that
certain moduli spaces (the Dwork family in this case) had points over various local extensions~$E$ of~$\mathbf{Q}_p$, in order to construct
a motive~$M$ over a number field~$F$ with~$F_v = E$ for~$v|p$. For the purposes of modularity
lifting, one wants strong control over the local deformation ring at~$p$, and the choice of local deformation ring is more or less forced by
the geometric properties of the~$p$-adic representations associated to~$M$ . One way to achieve this would be to work in
the Fontaine--Laffaille range where the local deformation rings were smooth. But this requires both that~$M$ is smooth at~$p$
and that the ramification degree~$e$
of~$E/\mathbf{Q}_p$ is one. It is not so clear, however, that the Dwork family contains suitable points (for a fixed residual representation~$\overline{\rho}$)
which lie in any unramified extension of~$\mathbf{Q}_p$. What Geraghty showed, however, was that certain ordinary deformation rings\footnote{In Geraghty's setting,
the residual representations~$\overline{\rho}$ were locally trivial. Hence the definition of ``ordinary'' was not something that could be defined
on the level of Artinian rings, and the construction (as with Kisin's construction of local deformation rings associated to certain types) is therefore indirect.}
were connected over arbitrarily ramified bases.
The final piece, however, was the construction of Galois
representations for all conjugate self-dual regular algebraic cuspidal~$\pi$ without the extra condition that~$\pi_q$
was square integrable for some~$q$. This story merits it own separate discussion; suffice to say
that it required the combined efforts of many people and the resolution of many difficult problems, not least of which
was the fundamental lemma by Laumon and Ng\^{o}~\cite{LN,N} (see also the Paris book project~\cite{MR2742611,book2},
the work of Shin~\cite{Shin},
and many more references which if I attempted to make complete
would weigh down the bibliography and still contain grievous omissions).
\subsection{Big image conditions}
The original arguments in~\cite{W,TW} required a ``big image'' hypothesis, namely that~$\overline{\rho}$ was absolutely irreducible after restriction to the Galois group of~$\mathbf{Q}(\zeta_p)$.
Wiles' argument also required the vanishing of certain cohomology groups associated to the adjoint representation of the image of~$\overline{\rho}$.
These assumptions had natural analogues in~\cite{CHT} (so-called ``big image'' hypotheses) although they were quite restrictive, and it wasn't clear that they would
even apply to most residual representations coming from some irreducible compatible family.
In the setting of~$2$-dimensional representations, the Taylor--Wiles
hypothesis guarantees the existence of many primes~$q$ such that~$q \equiv 1 \bmod p$ and such that~$\overline{\rho}(\mathrm{Frob}_q)$ has distinct eigenvalues.
This ensures, for example, that there cannot be any Steinberg deformations at~$q$ because the ratio of the eigenvalues of any Steinberg deformation must be~$q$.
In dimension~$n$, one natural way to generalize this might be to say that~$\overline{\rho}(\mathrm{Frob}_q)$ has distinct eigenvalues, although this is not always possible
to achieve for many irreducible representations~$\overline{\rho}$. A weaker condition is that~$\overline{\rho}(\mathrm{Frob}_q)$ has
an eigenvalue~$\alpha$ with multiplicity one. For such~$q$, there will be no deformations which are unipotent on inertia at~$q$ for which the generalized~$\alpha$
eigenspace is not associated to a~$1$-dimensional block. The translation of this into an automorphic condition on~$U_q$-eigenvalues is precisely what is done in~\cite{CHT}
(there are additional technical conditions on~$\mathrm{Frob}_q$ with respect to the adjoint representation~$\mathrm{ad}(\overline{\rho})$ which we omit here).
In~\cite{jack}, however, Thorne finds a way to allow~$\overline{\rho}(\mathrm{Frob}_q)$ to have an eigenvalue~$\alpha$ with higher multiplicity, and yet still cut out (integrally) the space
of automorphic forms whose Galois representations decompose at~$q$ as an unramified representation plus a one dimensional representation which is tamely ramified of~$p$-power
order. This technical improvement is very important because (as proved in the appendix by Guralnick, Herzig, Taylor, and Thorne~\cite{jack}) it imposes no restrictions on~$\overline{\rho}$ when~$p \ge 2n+1$
beyond the condition that~$\overline{\rho}$ is
absolutely irreducible after restriction to~$G_{\mathbf{Q}(\zeta_p)}$. This improvement is very useful for applications.
\subsection{Potentially diagonalizable representations} \label{section:BLGGT}
After the proof of Sato--Tate for modular forms, Barnet-Lamb, Gee, and Geraghty turned their attention to proving the analogous theorem for Hilbert modular forms of regular weight.
The methods developed so far were well-suited both to representations~$\rho$ which were either ordinary or when~$\rho$ was not ordinary but still Fontaine--Laffaille.
(The latter implies that~$\overline{\rho} |_{G_{\mathbf{Q}_p}}$ is absolutely irreducible of some particular shape.)
For a modular form over~$\mathbf{Q}$, one easily sees that~$\rho$ takes one of one of these forms for any sufficiently large~$p$.
For Hilbert modular forms, one certainly expects that the ordinary hypothesis should hold for all~$v|p$ and infinitely many~$p$, but this remains open. The difficulty arises when, for some prime~$p$
(that splits completely, say)
the~$p$-adic representation is ordinary at some~$v|p$ but
non-ordinary other~$v|p$.
The reason that this causes issues
is that, when applying the Moret-Bailly argument in the~$p$-$q$ switch, one wants to avoid any ramification at~$p$ for the non-ordinary case, and yet have large ramification
at the ordinary case to make~$\overline{\rho}$ locally trivial, and these desires are not compatible. The resolution in~\cite{HilbertST} involved a clever refinement of the Harris
tensor product trick. These ideas were further refined in~\cite{BLGGT} and led to the concept of a \emph{potentially diagonalizable}
representation~$\rho: G_E \rightarrow \mathrm{GL}_n(\overline{\Q}_p)$ for some finite extension of~$E/\mathbf{Q}_p$. Recall from~\S\ref{section:kisin} that, in the modified form of the Taylor--Wiles
method, proving modularity of some lift of~$\overline{\rho}$ often comes down to showing the existence of a modular lift lying on a smooth point of the corresponding component of
the generic fibre of~$R^{\mathrm{loc}}$. In light of Taylor's Ihara avoidance trick (\S\ref{section:iharaavoid}), the difficulty in this problem is mostly at the prime~$p$, and in particular
the fact that one knows very little about the components of general Kisin potentially crystalline deformation rings. A potentially diagonalizable representation is one for which,
after some finite (necessarily solvable!) extension~$E'/E$, the representation~$\rho |_{G_E'}$ is crystalline and lies on the same generic irreducible component as a diagonal representation.
This notion has a number of felicitous properties. First, it includes Fontaine--Laffaille representations and ordinary potentially crystalline representations. Second, it is clearly invariant under base change.
Third, it is compatible with the tensor product trick of Harris. These features make it supremely well-adapted to the current forms of the Taylor--Wiles method.
By combining this notion with methods of~\cite{HilbertST,blght}, as well as extensive use of Khare--Wintenberger lifting (\S\ref{section:KW}), Barnet--Lamb, Gee, Geraghty, and Taylor in~\cite{BLGGT}
proved the potential automorphy of \emph{all} conjugate self-dual irreducible\footnote{One variant proved shortly thereafter by Patrikis--Taylor~\cite{PT} replaced
the irreducibility condition by a purity condition (which is automatically satisfied by representations coming from pure motives).} odd\footnote{Although
there is no longer a non-trivial complex conjugation in the Galois group of a CM field, there is still an oddness condition related to the conjugate self-duality of the representation
and the fact that there are two ways for an irreducible representation to be self-dual (orthogonal and symplectic).}
compatible systems of Galois representations over a totally real field.
\subsection{Even Galois Representations}
\label{section:even}
The Fontaine--Mazur conjecture for geometric Galois representations~$\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\overline{\Q}_p)$ predicts that, up to twist, either~$\rho$
is modular \emph{or} $\rho$ is even with finite image. The methods of~\cite{W,TW} required the assumption that~$\overline{\rho}$ was modular
and so \emph{a priori} the assumption that~$\rho$ was odd (at least when~$p > 2$).
Nothing at all was known about the even case before the papers~\cite{Cal1,Cal2} in which a very simple trick made the problem accessible
to modularity lifting machinery under the assumption that the Hodge--Tate weights are distinct. The punch line is that, for any CM field~$F/F^{+}$,
the restriction~$\mathrm{Sym}^2(\rho): G_{F} \rightarrow \mathrm{GL}_3(\overline{\Q}_p)$ is conjugate self-dual and
no longer sees the ``evenness'' of~$\rho$\footnote{The representation~$\rho$ itself restricted to~$G_F$ will not be odd in the required sense ---
one exploits the fact here that~$3$ is odd whereas symplectic representations are always even dimensional.}. Hence one can hope to prove it is potentially modular for some CM extension~$L/L^{+}$,
and then by cyclic base change~\cite{MR1007299} potentially modular for the totally real field~$L^{+}$. But Galois representations coming from
regular algebraic automorphic forms for totally real fields will not be
even\footnote{I managed to twist Taylor's arm into writing the paper~\cite{TaylorSign} which proved this for odd~$n$, which
sufficed for my purposes where~$n$ was either~$3$ or~$9$. This is now also known for general~$n$, see Caraiani--Le Hung~\cite{AnaBao}.},
and thus one obtains a contradiction.
These ideas are already enough to deduce the main result of~\cite{Cal1} directly from~\cite{BLGGT}, although in contrast~\cite{Cal2} uses (indirectly) the full strength of the~$p$-adic local Langlands correspondence
via theorems of Kisin~\cite{KisinFM}. The papers~\cite{Cal1,Cal2} still fall short of completely resolving the Fontaine--Mazur in this case even for~$p > 7$,
since there remain big image hypotheses on~$\overline{\rho}$.
On the other hand, this trick has nothing to say about the case when the Hodge--Tate weights are equal (see~\S\ref{section:hopeless}).
\subsection{Modularity of higher symmetric powers}
Another parallel development in higher dimensions was the extension of Skinner--Wiles (\S\ref{section:SW}) to higher dimensions.
Many of the arguments of Skinner--Wiles relied heavily on the fact that any proper submodule of a~$2$-dimensional representation must
have dimension~$1$, and one-dimensional representations are very well understood by class field theory.
Nonetheless, in~\cite{MR3327536}, Thorne proved a residually reducible modularity theorem for higher dimensional
representations. In order to overcome the difficulty of controlling
reducible deformations, he imposed a Steinberg condition at some auxiliary prime. Although this is a definite restriction,
it does apply (for example) to the Galois representation coming from the symmetric power of a modular form which also satisfies this condition.
In a sequence of papers~\cite{CT1,CT2,CT3}, Clozel and Thorne applied this modularity lifting theorem to prove new cases of
symmetric power functoriality (see also the paper of Dieulefait~\cite{Dieulefait}). A key difficulty here is again the absence of Ihara's lemma in order to find automorphic forms
with the correct local properties. Very recently (using a number of new ideas), Newton and Thorne~\cite{NT1,NT2} were able to (spectacularly!) complete this program and prove the
full modularity of all symmetric powers of all modular forms.
\section{Beyond self-duality and Shimura varieties}
All the results discussed so far --- with the exception of those discussed in~\S\ref{section:artin} ---- apply only to Galois representations which
are both regular and satisfy some form of self-duality. Moreover, they all correspond to automorphic forms which can be detected by the (\'{e}tale) cohomology
of Shimura varieties. Once one goes beyond these representations, many of the established methods begin to break
immediately\footnote{I should warn the reader that this section and the next (even more than the rest of this paper) is filtered through the lens of my own
personal research journey --- \emph{caveat lector}!}.
An instructive case to consider is the case of~$2$-dimensional geometric Galois representations of an imaginary quadratic field~$F$ with distinct Hodge--Tate weights.
The corresponding automorphic forms for~$\mathrm{GL}_2(\mathbf{A}_F)$ contribute to the cohomology of locally symmetric spaces~$X$ which are arithmetic
hyperbolic~$3$-manifolds\footnote{Already by 1970, Serre (following ideas of Langlands) was trying to link Mennicke's computation that~$\mathrm{GL}_2(\mathbf{Z}[\sqrt{-109}])^{\mathrm{ab}}$ is infinite to the possible existence of elliptic curves over~$\mathbf{Q}(\sqrt{-109})$ with good reduction everywhere~\cite[Jan 14, 1970]{ColmezTate1}.}.
These spaces are certainly not algebraic varieties and their cohomology is hard to access via algebraic methods.
One of the first new questions to arise in this context is the relationship between torsion classes and Galois representations.
Some speculations about this matter were made by Elstrodt, Grunewald, and Mennicke at least as far back as 1981~\cite{EGM}, but the most influential conjecture was due to Ash~\cite{Ash},
who conjectured that eigenclasses in the cohomology of congruence subgroups of~$\mathrm{GL}_n(\mathbf{Z})$ over~$\overline{\mathbf{F}}_p$ (which need not lift to characteristic zero)
should give rise to~$n$-dimensional Galois
representations over finite fields.
Later, conjectures were made~\cite{AshSerre,AshExtra} in the converse conjecture in the spirit of Serre~\cite{MR885783} linking Galois representations to classes in cohomology modulo~$p$.
Certainly around~2004, however, it was not at all clear what exactly one should expect the landscape to be\footnote{I recall conversations with a number of experts at the 2004 Durham conference, where nobody seemed quite sure even what the dimension of the ordinary deformation ring~$R$ of a~$3$-dimensional representation~$\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_3(\overline{\mathbf{F}}_p)$ should be. Ash, Pollack, and Stevens had computed numerical examples where a regular algebraic ordinary cuspidal form for~$\mathrm{GL}_3(\mathbf{A}_{\mathbf{Q}})$
not twist-equivalent to a symmetric square did not appear to admit classical deformations. (I learnt about this example from Stevens at a talk at Banff in
December~$2003$.)
This would be easily explained if~$R$ had (relative) dimension~$0$ over~$\mathbf{Z}_p$
but be more mysterious otherwise.}, and so it was around this time I decided to start thinking about this question\footnote{One great benefit to me at the time of thinking about
Galois representations over imaginary quadratic fields was that it didn't require me to understand the geometry of Shimura varieties which I have always found
too complicated to understand. The irony of course is that the results of~\cite{10author,BCGP} ultimately rely on extremely intricate properties of Shimura varieties.}
in earnest. I became convinced very soon (for aesthetic reasons if not anything else) that if one modified~$\mathbf{T}$ to be the
ring of endomorphisms acting on integral cohomology (so that it would see not only the relevant automorphic forms but also the torsion classes) then there should still
be an isomorphism~$R=\mathbf{T}$. Moreover, this equality would not only be a form of reciprocity which moved beyond the conjecture linking motives to automorphic forms,
but it suggested that the integral cohomology of arithmetic groups (including the torsion classes) were themselves the fundamental object of interest.
Various developments served only to confirm this point of view. In my paper with Mazur~\cite{CalMazur}, we gave some theoretical evidence for why ordinary
families of Galois representations of imaginary quadratic fields might on the one hand be positive dimensional and explained completely by torsion classes and
yet not contain any classical automorphic points at all. During the process of writing~\cite{CalDunfield}, Dunfield (numerically) compared the torsion classes
in the cohomology of inner forms of~$\mathrm{GL}_2$ and the data was in perfect agreement with a conjectural Jacquet--Langlands correspondence for torsion
(later taken up in joint work with Venkatesh~\cite{CAVE}). Emerton and I had the idea of working with completed cohomology groups both to construct Galois representations and
even possibly to approach questions of modularity. The first idea was to exploit the well-known relationship between the cohomology of these manifolds
and the cohomology of the boundary of certain Shimura varieties. We realized that if we could control the co-dimension of the completed cohomology
groups over the non-commutative Iwasawa algebra, the Hecke eigenclasses would be forced to be seen by eigenclasses coming from the middle
degree of these Shimura varieties where one had access to Galois representations\footnote{Unfortunately, these conjectures~\cite[Conj~1.5]{MR2905536} remain all open in more
or less all cases except for Scholze's results in the case of certain Shimura varieties~\cite[Cor~4.2.3]{scholze-torsion}.}.
On the automorphy lifting side,
we had even vaguer ideas~\cite[\S1.8]{MR2905536}\footnote{Pan's remarkable paper~\cite{Pan1} turned some of these pipe dreams
into reality.} on how to proceed. A different (and similarly unsuccessful) approach\footnote{This is taken from my 2006 NSF proposal, and I believe
influenced by my conversations with Taylor at Harvard around that time.} was to work with ordinary deformations over a partial weight space for a split prime~$p=vw$ in an imaginary quadratic field~$F$.
That is, deformations of~$\overline{\rho}$ which had an unramified quotient at~$v$ and~$w$ but with varying weight at~$v$ and fixed weight at~$w$.
Here the yoga of Galois deformations suggested that~$R$ should be finite flat over~$W(k)$ in this case (and even a complete intersection). Moreover,
one had access to~$\mathbf{T}$ using an overlooked\footnote{One should never overlook results of Hida. I only learnt about this paper when Hida pointed it out to me
(with a characteristic smile on his face)
after my talk in Montreal in 2005. I was pleased at least that the idea that these families were genuinely non-classical was not anticipated either
in~\cite{HidaIQF} or in~\S4 of Taylor's thesis~\cite{taylorthesis}.}
result of Hida~\cite{HidaIQF}, and in particular one could deduce that~$\mathbf{T}$ has dimension at least one.
If one could show that~$\mathbf{T}$ was flat over~$W(k)$, then one could plausibly apply (assuming the existence of Galois representations) the original argument of~\cite{W,TW}.
The flatness of~$\mathbf{T}$, however, remains an open problem\footnote{One might even argue that there is no compelling argument to believe it is true ---
the problem is analogous to the vanishing of the~$\mu$-invariant in Iwasawa theoretic settings.}.
\subsection{The Taylor--Wiles method when \texorpdfstring{$l_0 > 0$}{l g 0}, part I: Calegari--Geraghty}
\label{section:calegari}
Shortly before (and then during) the special year on Galois representations at the IAS in 2010-2011, I started to work with Geraghty in earnest on the problem
of proving~$R=\mathbf{T}$ in the case of imaginary quadratic fields, assuming the existence of a surjection~$R \rightarrow \mathbf{T}$.
A computation in Galois cohomology shows that the expected ``virtual'' dimension of~$R$ over~$W(k)$ should be~$-1$, and hence the patched module~$M_{\infty}$
should have codimension~$1$ over the ring of diamond operators~$S_{\infty}$.
We realized this was a consequence of the fact that, after localizing the cohomology
at a non-Eisenstein maximal ideal, the cohomology should be non-zero in exactly two degrees. More precisely, patching the presentations
of these~$S_N$-modules
would result in a balanced presentation of~$M_{\infty}$ as an~$S_{\infty}$-module with the same (finite) number
of generators as relations.
We then
realized that the same principle held more generally for~$n$-dimensional representations over any number field. In characteristic zero, the
localized cohomology
groups were non-zero exactly in a range~$[q_0,q_0+l_0]$ (with~$q_0$ and~$l_0$ as defined in~\cite{BM}) where~$-l_0$ coincided with the expected virtual dimension of~$R$ over~$W(k)$ coming
from Galois cohomology. We could thus show --- assuming the localized torsion cohomology also vanished in this range --- that by
patching \emph{complexes} $P_Q$ (rather than modules~$M_Q$),
one arrives at a complex~$P_{\infty}$ of free~$S_{\infty}$ modules in degrees~$[q_0,q_0 + l_0]$.
Because the ring~$R_{\infty}$ of dimension~$\dim R_{\infty} = (\dim S_{\infty}) - l_0$ acts by patching on~$H^*(P_{\infty})$, a
simple commutative
algebra lemma then shows that~$M_{\infty} = H^*(P_{\infty})$ has codimension~$l_0$ over~$S_{\infty}$
and must be concentrated in the final degree. In particular, the Taylor--Wiles method (as modified by Diamond) could be happily adapted
to this general setting\footnote{David Hansen came up with a number of these ideas independently~\cite{Hansen}.}.
Moreover, the arguments were compatible with all the other improvements, including Taylor's Ihara avoidance argument amongst other things\footnote{These methods
only prove~$R[1/p]^{\mathrm{red}} = \mathbf{T}[1/p]^{\mathrm{red}}$, of course. In situations where~$\mathbf{T} \otimes \mathbf{Q} = 0$, the methods of~\cite{CG} in the minimal
case prove not only that~$R=\mathbf{T}$ but also that (both) rings are complete intersections. Moreover, one also has access to level raising (on the level of complexes)
and Ihara's lemma~\cite[\S4]{CAVE}, and I tried for some time (unsuccessfully) to adapt the original minimal $\Rightarrow$ non-minimal arguments
of~\cite{W} to this setting. There certainly seems to be some rich ideas in commutative algebra in these situations to explore, see for example
recent work of Tilouine--Urban~\cite{TU}.}.
We also realized that the same idea applied to Galois representations coming from the coherent cohomology of Shimura varieties even when
the corresponding automorphic forms were not discrete series. While our general formulation involved a number of conjectures we considered hopeless,
the coherent case had at least one setting in which many more results were available, namely the case of modular forms of weight one,
where the required vanishing conjecture was obvious, and where we were able to establish the existence of the required map~$R \rightarrow \mathbf{T}$
with all the required local properties by direct arguments. Although the state of knowledge concerning Galois representations
increased tremendously between the original conception of~\cite{CG} and its final publication, by early 2016 it still seemed out of reach
to make any of the results in~\cite{CG} unconditional.
\subsection{Construction of Galois representations, part II}
Before one can hope to prove~$R=\mathbf{T}$ theorems, one needs to be able to associate Galois representations to the corresponding automorphic forms.
There are two contexts in which one might hope to make progress. The first is in situations where the automorphic forms contribute to the
Betti cohomology of some locally symmetric space --- for example, tempered algebraic cuspidal automorphic representations for~$\mathrm{GL}_n(\mathbf{A}_F)$
and any~$F$. The second is in situations where the automorphic forms contribute to the coherent cohomology of some Shimura variety.
Here the first and easiest case corresponds to weight one modular forms, where the Galois representations were first constructed
by Deligne and Serre~\cite{deligne-serre}.
In work of Harris--Soudry--Taylor~\cite{HTS,HTST}, Galois representations were constructed for regular algebraic forms for~$\mathrm{GL}_2(\mathbf{A}_F)$ for an imaginary quadratic field~$F$ and satisfying a further
restriction on the central character. Harris, Soudry, and Taylor exploited (more or less) the fact that the automorphic induction of such forms
are self-dual (although not regular) and still contribute to coherent cohomology,
so one can construct Galois representations using a congruence argument as in the paper of Deligne and Serre~\cite{deligne-serre}.
On the other hand, this does not prove the expected local properties of the Galois representation at~$v|p$.
It was well-known for many years that the Hecke eigenclasses associated to regular algebraic cuspidal automorphic forms for~$\mathrm{GL}_n(\mathbf{A}_F)$
for a CM field~$F$ could be realized as eigenclasses coming from the boundary of certain unitary Shimura varieties of type~$U(n,n)$. It was, however,
also well-known that the corresponding \'{e}tale cohomology classes did not realize the desired Galois
representations\footnote{For a more basic example of what can go wrong, note that the Hecke eigenvalues of~$T_{l}$
on~$H^0(X,\mathbf{Q}_p)$ of a modular curve are~$1+l$, which corresponds to the Galois representation~$\mathbf{Q}_p \oplus \mathbf{Q}_p(1)$. However, only the piece~$\mathbf{Q}_p$
occurs inside~$H^0$.}. Remarkably, this problem was completely and unexpectedly resolved in 2011 in~\cite{HLTT} by Harris--Lan--Taylor--Thorne.
Richard Taylor writes:
\smallskip
\begin{footnotesize}
For~\cite{HLTT} I knew that the Hecke eigenvalues we were interested in contributed to Betti cohomology of~$U(n,n)$. The problem was to show that they contributed to overconvergent $p$-adic cusp forms. I was convinced on the basis of Coleman's paper ``classical and overconvergent modular forms'' \cite{MR1369416} that this must be so. I can't now reconstruct exactly why Coleman's paper convinced me of this, and it is possible, even probable, that my reasoning didn't really make any sense. However it was definitely this that kept me working at the problem, when we weren't really getting anywhere.
\end{footnotesize}
Amazingly, this breakthrough immediately inspired the next
development:
\subsection{Construction of Galois representations, part III: Scholze} \label{section:scholze}
In~\cite{scholze-torsion}, Scholze succeeded in constructing Galois representations associated to torsion classes
in the setting of~$\mathrm{GL}_n(\mathbf{A}_F)$ for a CM field~$F$.
Scholze had the idea after seeing some lectures on~\cite{HLTT}:
\smallskip
\begin{footnotesize}
During a HIM trimester at Bonn, Harris and Lan gave some talks about
their construction of Galois representations. At the time, I had some
ideas in my head that I didn't have any use for: That Shimura varieties
became perfectoid at infinite level, and that there is a Hodge--Tate
period map defined on them. The only consequence I could draw from this
were certain vanishing results for completed cohomology as conjectured
in your work with Emerton; so at least I knew that the methods were able
to say something nontrivial about torsion classes in the cohomology.
After hearing Harris' and Lan's talks, I was trying to see whether these
ideas could help in extending their results to torsion classes. After a
little bit of trying, I found the fake-Hasse invariants, and then it was
clear how the argument would go.
\end{footnotesize}
Even after this breakthrough, Scholze's construction still fell short of the conjectures in~\cite{CG}
in two ways. The first was that the Galois representation (ignoring here issues of pseudo-representations) was valued not in~$\mathbf{T}$
but in~$\mathbf{T}/I$ for some ideal~$I$ of fixed nilpotence. This is not a crucial obstruction to the methods of~\cite{CG}.
The second issue, however, was that the Galois representations were constructed (in the end) via $p$-adic congruences,
and thus one did not have
control over their local properties at~$p$ which \emph{are} crucial for modularity applications.
\subsection{The Taylor--Wiles method when \texorpdfstring{$l_0 > 0$}{l g 0}, part II: DAG}
Although not directly related to new~$R=\mathbf{T}$ theorems, one new recent idea in the subject has been
the work of Galatius--Venkatesh~\cite{GalV} on derived deformation rings in the context of Venkatesh's conjectures over~$\mathbf{Z}$.
This work (in part) reinterprets the arguments of~\cite{CG} in terms of a derived Hecke action.
The authors define a derived version~$\mathcal{R}$ of~$R$ with~$\pi_0(\mathcal{R}) = R$.
Under similar hypotheses to~\cite{CG}, the higher
homotopy groups of~$\mathcal{R}$ are shown to exist precisely in degrees~$0$ to~$l_0$.
One viewpoint of the minimal case of~\cite{CG} is that one constructs a (highly non-canonical) formally smooth ring~$R_{\infty}$
of dimension~$n - l_0$
with an action of a formally smooth ring~$S_{\infty}$ of dimension~$n$ such that the minimal deformation ring~$R$ is~$R_{\infty} \otimes_{S_{\infty}}
S_{\infty}/\mathfrak{a}$ for the augmentation ideal~$\mathfrak{a}$. Moreover, the ring~$R$ is identified both with the action of~$\mathbf{T}$ on the entire cohomology and simultaneously
on the cohomology in degree~$q_0 + l_0$.
On the other hand, when~$l_0 > 1$, the intersection of~$R_{\infty}$ and~$S_{\infty}/\mathfrak{a}$ over~$S_{\infty}$ is never
transverse\footnote{When~$l_0 = 1$, the intersection \emph{can} be transverse when~$R$ is a finite ring. In this setting, the relevant
cohomology is also non-zero and finite in exactly one degree.
On the other hand, as soon as~$\mathrm{Hom}(R,\overline{\Q}_p)$ is non-zero (for example, when there exists an associated motive) and~$l_0 > 0$, the intersection will
always be non-transverse.},
and homotopy groups of the derived intersection recover the cohomology in all degrees
(under the running assumptio2n one also knows that the patched cohomology is free).
On the other hand, there is a more canonical way to define~$\mathcal{R}$, namely to take the unrestricted global deformation ring~$R^{\mathrm{glob}}$ (which has
no derived structure) and intersect it with a suitable local crystalline deformation ring as algebras over the unrestricted local deformation ring. The expected dimension
of this intersection is also~$-l_0$ over~$W(k)$, although this is not so clear from this construction. Hence~\cite{GalV} can be viewed
as giving an intrinsic definition of~$\mathcal{R}$ independent of any choices of Taylor--Wiles primes and showing that its homotopy groups are related
(as with~$R_{\infty} \otimes^{\mathbf{L}}_{S_{\infty}} S_{\infty}/\mathfrak{a}$) to the cohomology\footnote{There are some subtleties as to what the precise statement
should be in the presence of global congruences, but already this author gets confused at the best of times between homology and cohomology, so I will not try to unentangle these issues here.}. These ideas have hinted at a closer connection between the Langlands program in the arithmetic case and the function field case than was previously
anticipated\footnote{Not anticipated by many people, at least; Michael Harris has been proselytizing the existence of a connection for quite some time.}, see for example work of Zhu~\cite{Zhuarxiv}.
\section{Recent Progress}
\subsection{Avoiding conjectures involving torsion I: the 10-author paper}
As mentioned in~\S\ref{section:scholze}, even after the results of~\cite{scholze-torsion} there remained a significant gap
to make the results of~\cite{CG} unconditional, namely, the conjecture that these Galois representations had the right local
properties at~$p$ and a second conjecture predicting the vanishing of (integral) cohomology localized at a non-maximal
ideal~$\mathfrak{m}$ outside a certain precise range (corresponding to known results in characteristic zero).
It should be noted that the second conjecture was still open (in all but the easiest cases)
in the simpler setting of Shimura varieties. The first hints that one could possibly make progress on this second conjecture (at least for Shimura
varieties) was given in an informal talk by Scholze in
Bellairs\footnote{I was invited to give the lecture series in Bellairs after Matthew Emerton didn't respond to his emails. Through
some combination of the appeal of my own work and the fact that the lectures were given on a beach in Barbados,
I managed to persuade Patrick Allen, George Boxer, Ana Caraiani, Toby Gee, Vincent Pilloni, Peter Scholze, and Jack Thorne to come,
all of whom are now my co-authors, and all of whom (if they weren't already at the time)
are now more of an expert in this subject than I am. The thought that I managed to teach any of them something about the subject
is pleasing indeed.} in~2014. This very quickly led to a long term collaboration between Scholze and Caraiani~\cite{CS1,CS2}, which
Caraiani describes as follows:
\smallskip
\begin{footnotesize}
At the Barbados conference in May 2014, Peter gave a lecture on how one might compute the cohomology of compact unitary Shimura varieties with torsion coefficients. The key was to have some control for~$R\pi_{HT*} \mathbb{F}_{\ell}$ restricted to any given Newton stratum. He was expressing this in terms of a conjecture that had grown out of his work on local Langlands using the Langlands--Kottwitz method. After his talk, I went to ask him some questions about this conjecture and it sounded like there were some things that still needed to be made precise. He asked if I wanted to help him make his strategy work. After some hesitation (because I didn’t think I knew enough or was strong enough to work with him), I accepted. Later that evening, I suggested switching from the Langlands--Kottwitz approach to understanding~$R\pi_{HT*}\mathbb{F}_{\ell}$ to an approach more in the style of
Harris--Taylor. This relies on the beautiful Mantovan product formula that describes Newton strata in terms of Rapoport--Zink spaces and Igusa varieties. Maybe something like this could help illuminate the geometry of the Hodge--Tate period morphism? Peter immediately saw that this should work and we made plans for me to visit Bonn that summer to continue the collaboration.
As Peter and I were finishing writing up the compact case, it became clear to us that the vanishing theorem would give a way to construct Galois representations associated to generic mod~$p$ classes that preserves the desired information at~$p$. Peter started thinking about the non-compact case and how that might apply to the local-global compatibility needed for Calegari--Geraghty. I remember discussing this with him at the Clay Research conference in Oxford in September 2015. By spring 2016, Richard started floating the idea of a working group on Calegari--Geraghty and found out that Peter and I had an approach to local-global compatibility. Around June 2016, Richard suggested to me to organize the working group with him. Peter was very excited about the idea, but wasn’t sure he would be able to attend for family reasons. In the end, we found a date in late October 2016 that worked for everyone.
\end{footnotesize}
The working group met under the auspices of the first ``emerging topics'' workshop\footnote{Although
later described as
a ``secret'' workshop, it was an ``invitation-only working group.''}
at the IAS to determine the extent to which the expected consequences could be applied to modularity lifting:
A clear stumbling point
was the vanishing of integral cohomology after localization outside the range of degrees~$[q_0,q_0 + l_0]$. On the other hand,
Khare and Thorne had already observed in~\cite{KT} by a localization argument that this could sometimes be avoided in certain
minimal cases. It was this argument we were able to modify for the general case,
thus avoiding the need to prove the (still open) vanishing conjectures for torsion classes\footnote{I regard my main contribution to~\cite{10author} as explaining how the arguments in~\cite{CG} using Taylor's
Ihara avoidance (\S\ref{section:iharaavoid}) were incompatible with any characteristic zero localization argument in the absence of (unknown)
integral vanishing results in cohomology. The objection (even in the case $l_0 = 0$) was that it was easy to construct complexes~$P^1$
and~$P^2$ of free~$S_{\infty}$ modules so that the support of~$H^*(P^1/p)$ and~$H^*(P^2/p)$ coincided (as they must) but
that (for example) $H^*(P^1)[1/p]$ was zero even though~$H^*(P^2)[1/p]$ was not. The objection to this objection, however,
which was resolved during the workshop (and which to
be clear I played no part in resolving!) is to not merely to compare the support of the complexes~$P^i/p$
but to consider the entire complex in the derived category. In particular, even (say) for a finite~$\mathbf{Z}_p$-module~$M$,
the module~$M[1/p]$ is non-zero exactly when~$M \otimes^{\mathbf{L}} \mathbf{F}_p$ has non-zero Euler characteristic.}.
The result of the workshop was a success beyond what we could have reasonably
anticipated
--- we ended up with
more or less\footnote{It is worth emphasizing that an incredible amount of work was required to turn these ideas into reality,
and that this intellectual effort was by and large
carried out by the younger members of the collaboration.}
the outline of a plan to prove all the main modularity lifting theorems which finally appeared in~\cite{10author},
namely the Ramanujan conjecture for regular algebraic automorphic forms for~$\mathrm{GL}_2(\mathbf{A}_F)$ of weight zero for any CM field~$F$,
and potential modularity (and the Sato--Tate conjecture)
for elliptic curves over CM fields.
There have already been a number of advancements beyond~\cite{10author} including in particular by
Allen, Khare, and Thorne~\cite{AKT} proving the modularity of many elliptic curves over~CM fields and
a potential automorphy theorem for ordinary representations by Qian~\cite{2104.09761}. It does not seem completely implausible
that results of the strength of~\cite{BLGGT} for~$n$-dimensional regular Galois representations of~$G_{\mathbf{Q}}$
are within reach.
\subsection{Avoiding conjectures involving torsion II: abelian surfaces}
\label{section:GSp4}
A second example that Geraghty and I had considered during the 2010--2011 IAS special year was the case of abelian surfaces,
corresponding to low (irregular) weight Siegel modular forms of genus~$g = 2$.
It was clear that a key difficulty was proving the vanishing of~$H^2(X,\omega^2)_{\mathfrak{m}}$ where~$X$
was a (compactified) Siegel~$3$-fold with good reduction at~$p$, where $\mathfrak{m}$ is maximal ideal of the Hecke algebra
corresponding to an absolutely irreducible representation, and where~$\omega |_{Y} = \det \pi_* \Omega^1_{\mathcal{A}/Y}$
on the open moduli space~$Y \subset X$ admitting a corresponding
universal abelian surface~$\mathcal{A}/Y$.
In other irregular weights (corresponding to motives
with Hodge--Tate weights~$[0,0,k-1,k-1]$ for~$k \ge 4$) the vanishing of the corresponding cohomology groups
was known by Lan and Suh~\cite{LS}. The vanishing of~$H^2(X,\omega^2)_{\mathfrak{m}}$ was more subtle, however,
because the corresponding group does \emph{not} vanish in general before localization in contrast to the previous cases.
In~\cite{CGGSp4}, we proved a minimal modularity theorem for these higher weight representations and a minimal
modularity theorem in the abelian case contingent on the vanishing conjecture above which we did not manage to resolve (and which remains unresolved).
I finished and then submitted the paper after I had moved to Chicago and Geraghty had moved to Facebook in 2015.
I then started working with Boxer and Gee\footnote{George Boxer had also arrived at Chicago in 2015, and was collaborating with Gee on companion form
results for Siegel modular forms, with the hope (in part) of deducing the modularity of abelian surfaces from Serre's
conjecture for~$\mathrm{GSp}_4$ in a manner analogous to the deduction by of the Artin conjecture
from Serre's conjecture for~$\mathrm{GL}_2$ in~\cite{KArtin,KW1}. They usually worked together at Plein Air cafe. Since I had
thought about similar questions with Geraghty and frequently went to Plein Air for 6oz cappuccinos, it was not entirely
surprising for us to start working together.}
on this vanishing question under certain supplementary local hypotheses.
(By this point, Galois representations associated to torsion classes in coherent cohomology had been constructed by Boxer~\cite{1507.05922}
and Goldring--Koskivirta~\cite{Wushi}.)
But then in November of 2016 (one week after the
IAS workshop!), Pilloni's paper on higher Hida theory~\cite{pilloniHidacomplexes} was first posted.
It was apparent to us that Pilloni's ideas would be extremely useful, and the four of us began a collaboration
almost immediately. Just as in~\cite{10author}, we were ultimately able to avoid proving any vanishing conjectures.
However, unlike~\cite{10author}, the way around this problem was not purely by commutative algebra, but instead by
working with ideas from~\cite{pilloniHidacomplexes}. Namely, instead of working with the cohomology of the full Siegel modular variety~$X$,
one could work with
the coherent cohomology of a certain open variety of~$X$ with cohomological dimension one whose (infinite dimensional)
cohomology could still be tamed using the methods of higher Hida theory~\cite{pilloniHidacomplexes}
in a way analogous to how Hida theory controls the (infinite dimensional) cohomology of the affine variety (with cohomological dimension zero)
corresponding to the ordinary locus. Generalizing this to a totally real field, one could then \emph{combine} these
ideas with the Taylor--Wiles method as modified in~\cite{CG}
to prove the potential modularity of abelian surfaces over totally real fields~\cite{BCGP}.
This coincidentally gives a second proof of the potential modularity of elliptic curves over CM fields proven in~\cite{10author}.
(The papers~\cite{10author} and~\cite{BCGP} both were conceived of and completed
within a week or so of each other.)
\section{The depths of our ignorance}
\label{section:hopeless}
Despite what can reasonably be considered significant progress in proving many cases of modularity since 1993, it remains the case that
many problems appear just as hopeless as they did then\footnote{Or perhaps harder, since there has been almost 30 years without
any progress whatsoever.}. Perhaps most embarrassing is the case of even Galois representations~$G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\mathbf{C})$ with
non-solvable image (equivalently, projective image~$A_5$). For example, we cannot establish the Artin conjecture for a \emph{single}
Galois representation whose image is the binary icosahedral group~$\mathrm{SL}_2(\mathbf{F}_5)$ of order~$120$.
The key problem is that the automorphic forms (Maass forms with
eigenvalue~$\lambda = 1/4$ in this case) are very hard to access --- given an even (projective)~$A_5$ Galois
representation, we don't even know how to prove that there exists a corresponding Maass form with the right Laplacian eigenvalue,
let alone one whose Hecke eigenvalues correspond to the Galois representation\footnote{Motives can be divided according to a tetrachotomy.
The first form are the Tate (and potentially Tate) motives, whose automorphy was known to Riemann and Hecke.
The second form are the motives (conjecturally) associated to automorphic representations which
are discrete series at infinity and thus amenable to the Taylor--Wiles method.
The third form are the motives (conjecturally) associated to automorphic representations
which are at least seen by some flavour of cohomology, either by the Betti cohomology of locally symmetric spaces or the coherent cohomology
of Shimura varieties (possibly in degrees greater than zero) which are amenable in principle to the modified Taylor--Wiles method. The fourth form consist
of the rest, which (besides a few that can be accessed
by cyclic base change) are a complete mystery.}.
In many ways, we have made no real progress on this question.
The case of curves of genus~$g > 2$ whose Jacobians have no extra endomorphisms seems equally hopeless.
One can only take solace in the fact that the Shimura--Taniyama conjecture seemed equally out of reach before Wiles' announcement in Cambridge
in 1993.
\bibliographystyle{emss}
|
\section{Introduction}
The space of \emph{stability conditions} on a triangulated category $\mathcal{T}$ was introduced by Bridgeland \cite{Bridgeland2007}, with inspiration from work by Douglas \cite{douglas2002dirichlet} in string theory. Roughly speaking, a stability condition $\sigma$ on a triangulated category $\mathcal{T}$ consists of data $(Z,\mathcal{P})$ where $Z$ is called the \emph{central charge}, and $\mathcal{P}$ is a slicing on $\mathcal{T}$. There are many examples and conjectures about the constructions of the stability conditions, we refer to surveys \cite{Bridgeland2009,macri2021,macri2017} for more information and applications.
Based on \cite{bondal1989representation,bondal1989representable,bernstein1982faisceaux}, Macr\`i \cite{macri2007stability,macri2004} gave a procedure generating the stability conditions on triangulated categories generated by an exceptional collection which was developed further by Collins and Polishchuk \cite{collins2010}. To a fixed complete exceptional collection $\mathcal{E}=\hkh{E_0,\cdots,E_n}$ on a triangulated category $\mathcal{T}$ , Macr\`i \cite{macri2007stability,macri2004} obtained an subspace of stability conditions, denoted by $\Theta_{\mathcal{E}}$, generated by $\mathcal{E}$.
Motivated by the work of Macr\`i \cite{macri2007stability,macri2004}, Dimitrov and Katzarkov \cite{DimitrovGeorge2016,DimitrovGeorge2016imrn} introduced the notation $\sigma$-\emph{exceptional collection} and reformulated Macr\`i's construction of $\Theta_{\mathcal{E}}$. For length 3 complete exceptional collection $\mathcal{E}=\hkh{E_0,E_1,E_2}$, Dimitrov and Katzarkov \cite{DimitrovGeorge2016} gave an exact description of $\Theta_{\mathcal{E}}$. For longer length cases, it is difficult to give an exact description of $\Theta_{\mathcal{E}}$ following the approach in \cite{DimitrovGeorge2016}. As explained in \cite[Remark 2.5]{DimitrovGeorge2016}, the main difficulty is that the similarly equality in \cite[Lemma 2.4]{DimitrovGeorge2016} does not hold for $n\geq 3$.
In this paper we give a description of $\Theta_{\mathcal{E}}$ for length 4 exceptional collection $\mathcal{E}$ as follows
\begin{thm}\label{thmmain}
Let $\mathcal{T}$ be a finite linear triangulated category over $\mathbb{C}$. Let $\mathcal{E}=\hkh{E_0,E_1,E_2,E_3}$ be a complete exceptional collection, denoted by
\[
k_{ij}=\min\hkh{k:\operatorname{Hom}^{k}(E_i,E_j)\neq 0}\in\mathbb{Z}.
\]
\begin{equation*}
\begin{split}
\mathcal{K}_{02}=\min\hkh{k_{02},k_{01}+k_{12}-1}, \quad \mathcal{K}_{13}=\min\hkh{k_{13},k_{12}+k_{23}-1},\\
\mathcal{K}_{03}=\min\hkh{k_{01}+k_{12}+k_{23}-2, k_{01}+k_{13}-1,k_{02}+k_{23}-1,k_{03}},
\end{split}
\end{equation*}
and
\[
\Theta_1=\hkh{
(y_0,y_1,y_2,y_3)\in\mathbb{R}^4\Bigg|\begin{array}{cl}
&y_0-y_1<k_{01},\quad y_1-y_2<k_{12}\\
&y_2-y_3<k_{23}, \quad y_1-y_3<\mathcal{K}_{13}\\
&y_0-y_3<\mathcal{K}_{03},\quad y_0-y_2<\mathcal{K}_{02}
\end{array}
}
\]
Then $\Theta_{\mathcal{E}}$ is homeomorphic with the set
\[
\mathbb{R}_{>0}^4\times\ykh{ \Theta_1\setminus \Delta}.
\]
where $\Delta$ is the union of following five sets
\begin{eqnarray*}
&\Delta^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\mathcal{R}_{23}^1,\quad\mathcal{R}_{13}^1\\
&\xqz{x_0-x_1}=k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_2-x_3}=k_{23}-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}\\
&\Delta^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\mathcal{R}_{23}^1,\quad\mathcal{R}_{13}^0,\quad \xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_2-x_3}=k_{23}-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}\\
&\Delta^3=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\mathcal{R}_{12}^1,\quad\mathcal{R}_{13}^0, \quad \xqz{x_0-x_1}<k_{01}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,x_2-x_3=\xqz{x_2-x_3}
\end{array}
}\\
&\Delta^4=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\mathcal{R}_{12}^1,\quad\mathcal{R}_{13}^0, \quad \xqz{x_0-x_1}<k_{01}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_0-x_3} =\mathcal{K}_{03}-1
\end{array}
}\\
&\Delta^5=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\mathcal{R}_{12}^0,\quad\mathcal{R}_{13}^1,\quad\xqz{x_1-x_3}=\mathcal{K}_{13}-1,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}= \mathcal{K}_{02}-1
\end{array}
}
\end{eqnarray*}
where $\mathcal{R}_{ij}^{\alpha},\alpha\in\hkh{0,1}, 1\leq i,j\leq 3$ are the relations defined by $\eqref{relationR}$.
\end{thm}
{\textbf{Proof idea.}} Our method is different from \cite{DimitrovGeorge2016}. We refer to Section \ref{mainproof} for complete notations. Our proof based on two ideas. First, we can see that $x_0=y_0$, then we need to $z_i\in (x_0-1,x_0+1)$, then the possible choices of $p_i$ are $\xqz{x_0-x_i}$ or $\sqz{x_0-x_i}$, then we reduce to check conditions $(z_0,z_1,z_2,z_3)\in S^3(-1,1)$ and $(0,p_1,p_2,p_3)\in A_0$, secondly, our main technical observation is the properties of floor function $\xqz{x}$ (See Lemma \ref{pair1}, Lemma \ref{triple1}, Lemma \ref{triple3}, Corollary \ref{triple2} and Corollary \ref{pijcor}) which can classify $\Theta_1$ into different cases, and check them case by case (See Lemma \ref{lemmacase1}-\ref{lemmacase4}). This classification can give all elements of $\Delta$ and avoid the more technical arguments of \cite[Lemma 2.4]{DimitrovGeorge2016}.
The paper is structured as follows. In the next section we review the basic notations and constructions. In Section \ref{mainproof} we give the complete proof. In Section \ref{example} we give the examples and some remarks of author's original motivation.
\section{Review: basic definitions and constructions}
In this section we review the basic notations and facts concerning the construction of the stability conditions following Macr\`i \cite{macri2007stability,macri2004}, Dimitrov--Katzarkov \cite{DimitrovGeorge2016,DimitrovGeorge2016imrn} and Bridgeland \cite{Bridgeland2007}. For details the reader is referred to the original papers \cite{macri2007stability,macri2004,DimitrovGeorge2016,DimitrovGeorge2016imrn,Bridgeland2007,bondal1989representation}.
Let $\mathcal{T}$ be a linear triangulated category and of finite type over $\mathbb{C}$. We denoted by $\jkh{E}$ the extension closed subcategory of $\mathcal{T}$ generated by subcategory $E$ of $\mathcal{T}$ and $K(\mathcal{T})$ the Grothendieck group of $\mathcal{T}$. Similarly, the Grothendieck group of an abelian category $\mathcal{A}$ is denoted $K(\mathcal{A})$.
\subsection{Stability conditions}
\begin{defn}[\cite{Bridgeland2007}]
A \emph{stability condition} $\sigma=(Z,\mathcal{P})$ on a triangulated category $\mathcal{T}$ consists of a group homomorphism $Z:K(\mathcal{T})\to \mathbb{C}$, and full additive subcategories $\mathcal{P}(\phi)\subset \mathcal{T}$ for each $\phi\in\mathbb{R}$, satisfying the following axioms:
\begin{enumerate}[(a)]
\item if $E\in\mathcal{P}(\phi)$ then $Z(E)=m(E)\exp(i\pi\phi)$ for some $m(E)\in\mathbb{R}_{>0}$,
\item for all $\phi\in\mathbb{R}$, $\mathcal{P}(\phi+1)=\mathcal{P}(\phi)[1]$,
\item if $\phi_1>\phi_2$ and $A_j\in\mathcal{P}(\phi_j)$ then $\operatorname{Hom}_{\mathcal{T}}(A_1,A_2)=0$,
\item for each nonzero object $E\in\mathcal{T}$ there are a finite sequence of real numbers
\[
\phi_1 > \phi_2 > \cdots>\phi_n
\]
and a collection of triangles
\[
\xymatrix@C=.5em{
0_{\ } \ar@{=}[r] & E_0 \ar[rrrr] &&&& E_1 \ar[rrrr] \ar[dll] &&&& E_2
\ar[rr] \ar[dll] && \ldots \ar[rr] && E_{n-1}
\ar[rrrr] &&&& E_n\ar[dll] \ar@{=}[r] & E_{\ } \\
&&& A_1 \ar@{-->}[ull] &&&& A_2 \ar@{-->}[ull] &&&&&&&& A_n \ar@{-->}[ull]
}
\]
with $A_j\in \mathcal{P}(\phi_j)$ for all $j$.
\end{enumerate}
\end{defn}
Nonzero objects in $\mathcal{P}(\phi)$ are said to be \emph{semistable} of phase $\phi$. For interval $I\subseteq\mathbb{R}$, we denote $\mathcal{P}(I)$ the extension closed subcategory of $\mathcal{T}$ generated by the subcategories $\mathcal{P}(\phi),\phi\in I$. If there exists some $\varepsilon>0$ such that each $\mathcal{P}((\phi-\varepsilon,\phi+\varepsilon))$ is of finite length, then we call the stability condition is \emph{locally finite}.
By \cite[Theorem 1.2, Corollary 1.3]{Bridgeland2007}, the space of locally finite stability conditions, denoted by $\operatorname{Stab}(\mathcal{T})$, is a complex manifold. By \cite[Lemma 8.2]{Bridgeland2007}, $\operatorname{Stab}(\mathcal{T})$ carries a right action of the group $\widetilde{\operatorname{GL}}^+(2,\mathbb{R})$, the universal covering space of $\operatorname{GL}^+(2,\mathbb{R})$, and a left action by isometries of the group of exact autoequivalences of $\mathcal{T}$.
For an abelian category $\mathcal{A}$, Bridgeland \cite[Definition 2.1]{Bridgeland2007} defined the \emph{stability function} $Z: K(\mathcal{A})\to \mathbb{C}$, and \cite[Definition 2.2]{Bridgeland2007} introduced the \emph{Harder-Narasimhan property} for the stability function. Then Bridgeland \cite[Proposition 5.3]{Bridgeland2007} showed that to give a stability condition on a triangulated category $\mathcal{T}$ is equivalent to giving a bounded $t$-structure on $\mathcal{T}$ and a stability function on its heart with the Harder-Narasimhan property.
\subsection{Exceptional objects}
The theory of exceptional collections and helix theory developed in the Rudakov seminar \cite{rudakov1990helices}. In order to get a heart of a bounded $t$-structure on $\mathcal{T}$ generated by the exceptional collection $\mathcal{E}$, Macr\`i \cite{macri2007stability,macri2004} introducted the notation \emph{Ext} exceptional collection.
Now, we recall the basic definitions. Following \cite{bondal1989representation} we denote
\[
\operatorname{Hom}^{\bullet}(A,B)=\bigoplus_{k\in\mathbb{Z}}\operatorname{Hom}^k(A,B)[-k],
\]
where $A,B\in\mathcal{T},\operatorname{Hom}^k(A,B)=\operatorname{Hom}(A,B[k])$. An object $E\in\mathcal{T}$ is called \emph{exceptional} if it satisfies
\begin{equation*}\operatorname{Hom}^i(E,E)=
\begin{cases}
\mathbb{C} & \text{ if } i=0,\\
0 & \text{ otherwise. }
\end{cases}
\end{equation*}
An ordered collection of exceptional objects $\{E_0,\cdots,E_n\}$ is called \emph{exceptional} in $\mathcal{T}$ if it satisfies
\[
\operatorname{Hom}^{\bullet}(E_i,E_j)=0,\text{ for } i>j.
\]
We call $\{E_0,\cdots,E_n\}$ the \emph{length} $n+1$ exceptional collection, and we call an exceptional collection of two objects an \emph{exceptional pair}.
Let $(E,F)$ an exceptional pair. We define objects $\mathcal{L}_E F$ and $\mathcal{R}_F E$ (which we call \emph{left mutation} and \emph{right mutation} respectively) by the distinguished triangles
\begin{eqnarray*}
\mathcal{L}_E F\longrightarrow \operatorname{Hom}^{\bullet}(E,F)\otimes E\longrightarrow F,\\
E\longrightarrow \operatorname{Hom}^{\bullet}(E,F)^*\otimes F\longrightarrow \mathcal{R}_F E,
\end{eqnarray*}
where $V[k]\otimes E$ denotes an object isomorphic to the direct sum of $\dim V$ copies of the object $E[k]$. A \emph{mutation} of an exceptional collection $\mathcal{E}=\{E_0,\cdots,E_n\}$ is defined as a mutation of a pair of adjacent objects in this collection:
\[
\mathcal{R}_i\mathcal{E}=\{E_0,\cdots,E_{i-1},E_{i+1},\mathcal{R}_{E_{i+1}}E_i, E_{i+2},\cdots,E_n\},
\]
\[
\mathcal{L}_i\mathcal{E}=\{E_0,\cdots,E_{i-1},\mathcal{L}_{E_{i}}E_{i+1}, E_{i}, E_{i+2},\cdots,E_n\},
\]
for $i=0,\cdots,n-1$.
\begin{defn}[\cite{macri2007stability}]
Let $\mathcal{E}=\{E_0,\cdots,E_n\}$ be an exceptional collection. We call $\mathcal{E}$
\begin{itemize}
\item strong, if $\operatorname{Hom}^k(E_i,E_j)=0$ for all $i$ and $j$, with $k\neq 0$;
\item Ext, if $\operatorname{Hom}^{\leq 0}(E_i,E_j)=0$ for all $i\neq j$;
\item complete, if $\mathcal{E}$ generates $\mathcal{T}$ by shifts and extensions.
\end{itemize}
\end{defn}
\subsection{Basic construction}
Let $\mathcal{E}=\hkh{E_0,\cdots,E_n}$ be a complete exceptional collection on $\mathcal{T}$. For $\textbf{p}=(p_0,\cdots,p_n)\in\mathbb{Z}^{n+1}$, we denote by $\mathcal{E}[\textbf{p}]:=\hkh{E_0[p_0],\cdots,E_n[p_n]}$ the shift of the exceptional collection $\mathcal{E}$ and $\mathcal{A}_{\mathcal{E}[\textbf{p}]}=\jkh{E_0[p_0],\cdots,E_n[p_n]}$ the extension closed subcategory generated by $\mathcal{E}[\textbf{p}]$. We write
\begin{equation}\label{A0def}
A_0=\hkh{\textbf{p}: \textbf{p}=(p_0,\cdots,p_n)\in\mathbb{Z}^{n+1} \text{ such that } \mathcal{E}[\textbf{p}] \text{ is Ext}}.
\end{equation}
From now, we assume $\textbf{p}\in A_0$. By \cite[Lemma 3.14]{macri2007stability} , $\mathcal{A}_{\mathcal{E}[\textbf{p}]}$ is the heart of a bounded $t$-structure on $\mathcal{T}$. Fix $z_0,\cdots,z_n\in \mathbb{H}:=\hkh{z\in\mathbb{C}: z=|z|\exp(i\pi\phi),0<\phi\leq 1}$. Define a stability function $Z_{\textbf{p}}: K_0(\mathcal{A}_{\mathcal{E}[\textbf{p}]})\to \mathbb{C}$ by
\[
Z_{\textbf{p}}(E_i[p_i])=z_i, \text{ for all } i.
\]
By \cite[Lemma 3.16]{macri2007stability}, $\mathcal{A}_{\mathcal{E}[\textbf{p}]}$ is an abelian category of finite length and with finitely many simple objects, then by \cite[Proposition 2.4]{Bridgeland2007} any stability function $Z_{\textbf{p}}$ on $\mathcal{A}_{\mathcal{E}[\textbf{p}]}$ satisfies the Harder--Narasimhan property. By \cite[Proposition 5.3]{Bridgeland2007}, this extends to a unique stability condition on $\mathcal{T}$ which is locally finite.
Following \cite[Defintion 3.13]{DimitrovGeorge2016imrn}, we denote by $\mathbb{H}^{\mathcal{E}[\textbf{p}]}$ the set of stability conditions on $\mathcal{T}$ constructued above, denote $\Theta_{\mathcal{E}[\textbf{p}]}^{\prime}=\mathbb{H}^{\mathcal{E}[\textbf{p}]}\cdot\widetilde{\operatorname{GL}}^+(2,\mathbb{R})$, and
\begin{equation}
\Theta_{\mathcal{E}}=\bigcup_{\textbf{p}\in A_0} \Theta_{\mathcal{E}[\textbf{p}]}^{\prime}\label{thetaall}.
\end{equation}
For the convenience, Dimitrov--Katzarkov \cite[Definition 2.3]{DimitrovGeorge2016} introduced the notation $S^n(-1,1)$, which defined by
\[
S^n(-1,+1)=\hkh{(y_0,\cdots,y_n)\in\mathbb{R}^{n+1}: -1<y_i-y_j<1,i<j}\subset\mathbb{R}^{n+1}.
\]
By the comment in Dimitrov--Katzarkov \cite[p. 833-834]{DimitrovGeorge2016} and \cite[formula (18)]{DimitrovGeorge2016}, there exists a homeomorphism which we denote by $f_{\mathcal{E}}$:
\begin{eqnarray}\label{homeo}
f_{\mathcal{E}}:\Theta_{\mathcal{E}}\longrightarrow \mathbb{R}_{>0}^{n+1}\times \ykh{\bigcup_{\textbf{p}\in A_0}S^n(-1,+1)-\textbf{p}}.
\end{eqnarray}
\begin{remark}
Dimitrov--Katzarkov \cite[p. 833-834]{DimitrovGeorge2016} defined the homeomorphism $f_{\mathcal{E}}$ by the restriction. The exact definition of $f_{\mathcal{E}}$ is not necessary for our proof. For the details of $f_{\mathcal{E}}$ the reader is referred to the original paper \cite{DimitrovGeorge2016}.
\end{remark}
For $n=2$ case, Dimitrov--Katzarkov \cite{DimitrovGeorge2016} gave an exact description of $\Theta_{\mathcal{E}}$. In the next section we give a complete description of $\Theta_{\mathcal{E}}$ for $n=3$ case using different menthod.
\section{Main Theorem}\label{mainproof}
In this section we prove the main theorem. For $x\in\mathbb{R}$, let us denote \emph{floor function} by
\[
\xqz{x}=\max\hkh{n\in \mathbb{Z}: n\leq x},
\]
\emph{ceiling function} by $\sqz{x}=\xqz{x}+1$ and \emph{fractional part} of $x$ by $\hkh{x}=x-\xqz{x}$.
It is easy to check that for every $x$ and $y$, the following inequality holds
\begin{equation}\label{xqz}
\xqz{x}+\xqz{y}\leq \xqz{x+y}\leq \xqz{x}+\xqz{y}+1.
\end{equation}
For $(x_0,x_1,x_2,x_3)\in\mathbb{R}^4, 1\leq i<j\leq 3$ and $\alpha\in\hkh{0,1}$, by the inequality \eqref{xqz}, we can define $\mathcal{R}_{ij}^{\alpha}$ as the relation
\begin{equation}\label{relationR}
\xqz{x_0-x_j }-\xqz{x_0-x_i }=\xqz{x_i-x_j }+\alpha.
\end{equation}
By the inequality \eqref{xqz}, the following lemma holds.
\begin{lemma}\label{pair1}
For fixed $i, j$, $\mathcal{R}_{ij}^0$ and $\mathcal{R}_{ij}^1$ do not hold at the same time.
\end{lemma}
Moreover, for triple relations $\mathcal{R}_{12}^{\alpha}$, $\mathcal{R}_{13}^{\beta}$ and $\mathcal{R}_{23}^{\gamma}$, the following lemma holds.
\begin{lemma}\label{triple1}
Let $(x_0,x_1,x_2,x_3)\in\mathbb{R}^4$, $\alpha,\beta,\gamma\in\hkh{0,1}$. If relations $\mathcal{R}_{12}^{\alpha}$, $\mathcal{R}_{13}^{\beta}$ and $\mathcal{R}_{23}^{\gamma}$ hold at the same time. Then
\begin{equation}
\beta\leq\alpha+\gamma\leq \beta+1.\label{abcinequ}
\end{equation}
In particular, all possibly values of triple $(\alpha,\beta,\gamma)$ are
\[
(0,0,0),\quad (1,1,0),\quad (1,0,0),\quad (0,0,1), \quad (1,1,1),\quad (0,1,1).
\]
\end{lemma}
\begin{proof}
If relations $\mathcal{R}_{12}^{\alpha}$, $\mathcal{R}_{13}^{\beta}$ and $\mathcal{R}_{23}^{\gamma}$ hold at the same time, i.e.
\begin{align*}
&\xqz{x_0-x_2}-\xqz{x_0-x_1}=\xqz{x_1-x_2}+\alpha\\
&\xqz{x_0-x_3}-\xqz{x_0-x_1}=\xqz{x_1-x_3}+\beta\\
&\xqz{x_0-x_3}-\xqz{x_0-x_2}=\xqz{x_2-x_3}+\gamma
\end{align*}
then we have
\[
\xqz{x_1-x_3}+\beta=\xqz{x_0-x_3}-\xqz{x_0-x_1}=\xqz{x_2-x_3}+\xqz{x_1-x_2}+\alpha+\gamma
\]
By the inequality \eqref{xqz}, we have
\[
\xqz{x_2-x_3}+\xqz{x_1-x_2}\leq \xqz{x_1-x_3}\leq \xqz{x_2-x_3}+\xqz{x_1-x_2}+1
\]
Hence
\[
\beta\leq\alpha+\gamma\leq \beta+1.
\]
Then we can list all possibly values of $(\alpha,\beta,\gamma)$ as follows
\[
(0,0,0),\quad (1,1,0),\quad (1,0,0),\quad (0,0,1), \quad (1,1,1),\quad (0,1,1).
\]
\end{proof}
Then we have following corollary
\begin{coro}\label{triple2}
Relations $\mathcal{R}_{12}^0$, $\mathcal{R}_{13}^1$ and $\mathcal{R}_{23}^0$ do not hold at the some time. Relations $\mathcal{R}_{12}^1$, $\mathcal{R}_{13}^0$ and $\mathcal{R}_{23}^1$ do not hold at the some time.
\end{coro}
Before the main theorem, we recall more properties of $\mathcal{R}_{ij}^{\alpha},\alpha\in\hkh{0,1}$.
\begin{lemma}\label{triple3}
Let $(x_0,x_1,x_2,x_3)\in\mathbb{R}^4$. If $\xqz{x_i-x_j}= x_i-x_j$ hold, then $\mathcal{R}_{ij}^{0}$ hold.
\end{lemma}
\begin{proof}
If $\xqz{x_i-x_j}= x_i-x_j$ holds, then $\hkh{x_i}=\hkh{x_j}$, we also have $\hkh{x_0-x_i}=\hkh{x_0-x_j}$. By the definition of $\hkh{x}$, we have
\[
x_0-x_i-\xqz{x_0-x_i}=x_0-x_j-\xqz{x_0-x_j}
\]
then $\mathcal{R}_{ij}^{0}$ hold.
\end{proof}
Hence we have
\begin{coro}\label{pijcor}
If $\mathcal{R}_{ij}^0$ holds,
then
\[
\ykh{x_i+\xqz{x_0-x_i}}-\ykh{x_j+\xqz{x_0-x_j}}=x_i-x_j-\xqz{x_i-x_j}\in[0,1).
\]
If $\mathcal{R}_{ij}^1$ holds, then $\xqz{x_i-x_j}\neq x_i-x_j$, and
\[
\ykh{x_i+\xqz{x_0-x_i}}-\ykh{x_j+\xqz{x_0-x_j}}=x_i-x_j-\xqz{x_i-x_j}-1\in(-1,0).
\]
\end{coro}
Before the main theorem, we prove the following four lemmas which is useful in the proof of the main theorem.
\begin{lemma}\label{lemmacase1}
If $\xqz{ x_0-x_1}=k_{01}-1$ and $\xqz{ x_0-x_2}=\mathcal{K}_{02}-1$, then for any $(x_0,x_1,x_2,x_3)\in\Theta_1$, there exists $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$ such that $x_0=z_0, x_i=z_i-p_i, i=1,2,3$.
\end{lemma}
\begin{proof}
If $\xqz{ x_0-x_1}=k_{01}-1$ and $\xqz{ x_0-x_2}=\mathcal{K}_{02}-1$, then for any $(x_0,x_1,x_2,x_3)\in\Theta_1$ , we can set
\[
p_i=\xqz{ x_0-x_i}, z_i=x_i+p_i, i=1,2,3.
\]
Then we have
\begin{eqnarray*}
p_2-p_1=\mathcal{K}_{02}-1-k_{01}+1\leq k_{12}-1, \\
p_3-p_1\leq \mathcal{K}_{03}-1-k_{01}+1\leq \mathcal{K}_{13}-1,\\
p_3-p_2 \leq \mathcal{K}_{03}-1-\mathcal{K}_{02}+1\leq k_{23}-1
\end{eqnarray*}
Hence $(0,p_1,p_2,p_3)\in A_0$. By the Corollary \ref{pijcor}, we have $z_i\in S^3(-1,+1)$.
\end{proof}
\begin{lemma}\label{lemmacase2}
Let
\[
\Delta_{\uppercase\expandafter{\romannumeral2}}=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\quad\mathcal{R}_{23}^1,\\
&\xqz{x_0-x_1}=k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_2-x_3}=k_{23}-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}.
\]
If $\xqz{ x_0-x_1}=k_{01}-1\text{ and } \xqz{ x_0-x_2}<\mathcal{K}_{02}-1$, then for any $(x_0,x_1,x_2,x_3)\in\Theta_1\setminus\Delta_{\uppercase\expandafter{\romannumeral2}}$, there exists $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$ such that $x_0=z_0, x_i=z_i-p_i, i=1,2,3$.
\end{lemma}
\begin{proof}
If $\xqz{ x_0-x_1}=k_{01}-1\text{ and } \xqz{ x_0-x_2}<\mathcal{K}_{02}-1$, by the Lemma \ref{pair1}, we can consider following cases
\begin{enumerate}
\item If the relation $\mathcal{R}_{23}^0$ holds, then
we can set
\[
p_i=\xqz{ x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
Then we have
\begin{eqnarray*}
p_2-p_1\leq\mathcal{K}_{02}-1-k_{01}+1\leq k_{12}-1, \\
p_3-p_1\leq \mathcal{K}_{03}-1-k_{01}+1\leq \mathcal{K}_{13}-1,\\
p_3-p_2=\xqz{x_0-x_3}-\xqz{x_0-x_2}=\xqz{x_2-x_3}\leq k_{23}-1
\end{eqnarray*}
Hence $(0,p_1,p_2,p_3)\in A_0$. By the Corollary \ref{pijcor}, we have $z_i\in S^3(-1,+1)$.
\item If the relation $\mathcal{R}_{23}^1$ holds,
\begin{enumerate}[(i)]
\item\label{case221} If $\xqz{ x_2-x_3}\leq k_{23}-2$, we can set
\[
p_i=\xqz{x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
Then we have
\begin{eqnarray*}
p_2-p_1\leq\mathcal{K}_{02}-1-k_{01}+1\leq k_{12}-1, \\
p_3-p_1\leq \mathcal{K}_{03}-1-k_{01}+1\leq \mathcal{K}_{13}-1,\\
p_3-p_2=\xqz{x_0-x_3}-\xqz{x_0-x_2}=\xqz{x_2-x_3}+1\leq k_{23}-1
\end{eqnarray*}
Hence $(0,p_1,p_2,p_3)\in A_0$. By the Corollary \ref{pijcor}, we have $z_i\in S^3(-1,+1)$.
\item If the relation $\mathcal{R}_{12}^0$ and the condition $\xqz{ x_1-x_2}\neq x_1-x_2$ hold,
we set
\[
p_1=\xqz{ x_0-x_1}, \quad p_2=\sqz{ x_0-x_2 } ,\quad p_3=\xqz{ x_0-x_3}, z_i=x_i+p_i, i=1,2,3
\]
Then we have
\begin{eqnarray*}
p_2-p_1\leq\mathcal{K}_{02}-1-k_{01}+1\leq k_{12}-1, \\
p_3-p_1\leq \mathcal{K}_{03}-1-k_{01}+1\leq \mathcal{K}_{13}-1,\\
p_3-p_2=\xqz{x_0-x_3}-\xqz{x_0-x_2}=\xqz{x_2-x_3}+1\leq k_{23}-1
\end{eqnarray*}
Hence $(0,p_1,p_2,p_3)\in A_0$. By the Corollary \ref{pijcor}, we have $z_i\in S^3(-1,+1)$.
\end{enumerate}
\end{enumerate}
If $\mathcal{R}_{23}^1$ and $\mathcal{R}_{12}^1$ hold, then only case \ref{case221} is possible. Then we can write $\Delta_{\uppercase\expandafter{\romannumeral2}}$ by
\[
\Delta_{\uppercase\expandafter{\romannumeral2}}=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\quad\mathcal{R}_{23}^1,\\
&\xqz{x_0-x_1}=k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_2-x_3}=k_{23}-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}
\]
\end{proof}
\begin{lemma}\label{lemmacase3}
Let $\Delta_{\uppercase\expandafter{\romannumeral3}}=\Delta_{\uppercase\expandafter{\romannumeral3}}^1\cup\Delta_{\uppercase\expandafter{\romannumeral3}}^2\cup\Delta_{\uppercase\expandafter{\romannumeral3}}^3$
where
\begin{eqnarray*}
&\Delta_{\uppercase\expandafter{\romannumeral3}}^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}= \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,x_2-x_3=\xqz{x_2-x_3}
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral3}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}= \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_0-x_3} =\mathcal{K}_{03}-1
\end{array}
}
\\
&\Delta_{\uppercase\expandafter{\romannumeral3}}^3=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\quad\mathcal{R}_{12}^0,\mathcal{R}_{13}^1,\xqz{x_1-x_3}=\mathcal{K}_{13}-1,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}= \mathcal{K}_{02}-1
\end{array}
}.
\end{eqnarray*}
If $\xqz {x_0-x_1}< k_{01}-1\text{ and } \xqz{x_0-x_2}=\mathcal{K}_{02}-1$, then for any $(x_0,x_1,x_2,x_3)\in\Theta_1\setminus\Delta_{\uppercase\expandafter{\romannumeral3}}$, there exists $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$ such that $x_0=z_0, x_i=z_i-p_i, i=1,2,3$.
\end{lemma}
\begin{proof}
If $\xqz {x_0-x_1}< k_{01}-1\text{ and } \xqz{x_0-x_2}=\mathcal{K}_{02}-1$, as the similar analysis in Lemma \ref{lemmacase2}, by Lemma \ref{pair1}, we can consider the following cases:
\begin{enumerate}
\item If the relation $\mathcal{R}_{13}^0$ holds,
\begin{enumerate}[(i)]
\item If the relation $\mathcal{R}_{12}^0$ holds, then
we can find
\[
p_i=\xqz{ x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\item If the relation $\mathcal{R}_{12}^1$ and $\xqz{x_1-x_2}\leq k_{12}-2$ hold, then
we can find
\[
p_i=\xqz{ x_0-x_i},z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\item If relations $\mathcal{R}_{12}^1$, $\mathcal{R}_{23}^0$ , $\xqz{x_2-x_3}\neq x_2-x_3$ and $\sqz{x_0-x_3}\leq\mathcal{K}_{03}-1$ hold, then
we can find
\[
p_1=\sqz{ x_0-x_1}, p_2=\xqz{ x_0-x_2},p_3=\sqz{ x_0-x_3},z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\end{enumerate}
\item If the relation $\mathcal{R}_{13}^1$ holds,
\begin{enumerate}[(i)]
\item If $\mathcal{R}_{12}^0$ and $x_1-x_3\neq \xqz{x_1-x_3}\leq \mathcal{K}_{13}-2$ hold, we can find
\[
p_i=\xqz{x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\item If $\mathcal{R}_{12}^1$ holds, we can find
\[
p_1=\sqz{x_0-x_1}, p_2=\xqz{x_0-x_2}, p_3=\xqz{x_0-x_3}, \quad z_i=x_i+p_i, i+1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\end{enumerate}
\end{enumerate}
By Lemma \ref{triple3}, if $\xqz{x_1-x_3} = x_1-x_3$, then the relation $\mathcal{R}_{13}^0$ holds. Hence we can write $\Delta_{\uppercase\expandafter{\romannumeral3}}$ as
\[
\Delta_{\uppercase\expandafter{\romannumeral3}}=\Delta_{\uppercase\expandafter{\romannumeral3}}^1\cup\Delta_{\uppercase\expandafter{\romannumeral3}}^2\cup\Delta_{\uppercase\expandafter{\romannumeral3}}^3
\]
where
\begin{eqnarray*}
&\Delta_{\uppercase\expandafter{\romannumeral3}}^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}= \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,x_2-x_3=\xqz{x_2-x_3}
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral3}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}= \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_0-x_3} =\mathcal{K}_{03}-1
\end{array}
}
\\
&\Delta_{\uppercase\expandafter{\romannumeral3}}^3=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\quad\mathcal{R}_{12}^0,\mathcal{R}_{13}^1,\xqz{x_1-x_3}=\mathcal{K}_{13}-1,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}= \mathcal{K}_{02}-1
\end{array}
}
\end{eqnarray*}
\end{proof}
\begin{lemma}\label{lemmacase4}
Let $\Delta_{\uppercase\expandafter{\romannumeral4}}=\Delta_{\uppercase\expandafter{\romannumeral4}}^1\cup\Delta_{\uppercase\expandafter{\romannumeral4}}^2\cup\Delta_{\uppercase\expandafter{\romannumeral4}}^3$
where
\begin{eqnarray*}
&\Delta_{\uppercase\expandafter{\romannumeral4}}^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\mathcal{R}_{23}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_2-x_3}=x_2-x_3,
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral4}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cl}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\mathcal{R}_{23}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_0-x_3}=\mathcal{K}_{03}-1,
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral4}}^3=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cl}
&\quad\mathcal{R}_{12}^0,\mathcal{R}_{13}^0,\mathcal{R}_{23}^1,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}<\mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=x_1-x_2,\xqz{x_2-x_3}=k_{23}-1,
\end{array}
}
\end{eqnarray*}
If $\xqz {x_0-x_1}< k_{01}-1\text{ and } \xqz{x_0-x_2}<\mathcal{K}_{02}-1$, then for any $(x_0,x_1,x_2,x_3)\in\Theta_1\setminus\Delta_{\uppercase\expandafter{\romannumeral4}}$, there exists $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$ such that $x_0=z_0, x_i=z_i-p_i, i=1,2,3$.
\end{lemma}
\begin{proof}
If $\xqz {x_0-x_1}< k_{01}-1\text{ and } \xqz{x_0-x_2}<\mathcal{K}_{02}-1$, as similar analysis in Lemma \ref{lemmacase2}, we can list all possible choices as follows
\begin{enumerate}[(A)]
\item\label{A120} If the relation $\mathcal{R}_{12}^0$ holds, then we have the following possible choices
\begin{enumerate}[(1)]
\item\label{A1201} $p_1=\xqz{x_0-x_1}$ and $p_2=\xqz{x_0-x_2}$,
\item\label{A1202} $p_1=\sqz{x_0-x_1}$ and $p_2=\sqz{x_0-x_2}$,
\item\label{A1203} If $x_1-x_2\neq \xqz{x_1-x_2} \text{ and } \xqz{x_1-x_2}\leq k_{12}-2$,
then we can choose
\[
p_1=\xqz{x_0-x_1} \text{ and } p_2=\sqz{x_0-x_2}.
\]
\end{enumerate}
\item\label{A121} If the relation $\mathcal{R}_{12}^1$ holds, then we have the following possible choices
\begin{enumerate}[(1)]
\item\label{A1211} $p_1=\sqz{x_0-x_1}$ and $p_2=\xqz{x_0-x_2}$,
\item\label{A1212} If $\xqz{x_1-x_2}\leq k_{12}-2$ then $p_1=\xqz{x_0-x_1}$ and $p_2=\xqz{x_0-x_2}$, or
\item\label{A1213} If $\xqz{x_1-x_2}\leq k_{12}-2$ then $p_1=\sqz{x_0-x_1}$ and $p_2=\sqz{x_0-x_2}$,
\end{enumerate}
\item\label{A130} If the relation $\mathcal{R}_{13}^0$ holds, then we have the following possible choices
\begin{enumerate}[(1)]
\item\label{A1301} $p_1=\xqz{x_0-x_1}$ and $p_3=\xqz{x_0-x_3}$,
\item\label{A1302} If $\sqz{x_0-x_3}\leq k_{03}-1$, then we can choose $p_1=\sqz{x_0-x_1}$ and $p_3=\sqz{x_0-x_3}$,
\item\label{A1303} If $x_1-x_3\neq \xqz{x_1-x_3}, \xqz{x_1-x_3}\leq k_{13}-2,\text{ and } \sqz{x_0-x_3}\leq\mathcal{K}_{03}-1$
then we can choose
\[
p_1=\xqz{x_0-x_1} \text{ and } p_3=\sqz{x_0-x_3}.
\]
\end{enumerate}
\item\label{A131} If the relation $\mathcal{R}_{13}^1$ holds, then we have the following possible choices
\begin{enumerate}[(1)]
\item\label{A1311} $p_1=\sqz{x_0-x_1}$ and $p_3=\xqz{x_0-x_3}$,
\item\label{A1312} If $\xqz{x_1-x_3}\leq k_{13}-2$ then we can choose $p_1=\xqz{x_0-x_1}$ and $p_3=\xqz{x_0-x_3}$, or
\item\label{A1313} If $\xqz{x_1-x_3}\leq k_{13}-2$ and $\sqz{x_0-x_3}\leq\mathcal{K}_{03}-1$ then we can choose $p_1=\sqz{x_0-x_1}$ and $p_3=\sqz{x_0-x_3}$,
\end{enumerate}
\item\label{A230} If the relation $\mathcal{R}_{23}^0$ holds, then we have the following possible choices
\begin{enumerate}[(1)]
\item\label{A2301} $p_2=\xqz{x_0-x_2}$ and $p_3=\xqz{x_0-x_3}$,
\item\label{A2302} If $\sqz{x_0-x_3}\leq\mathcal{K}_{03}-1$, then we can choose $p_2=\sqz{x_0-x_2}$ and $p_3=\sqz{x_0-x_3}$,
\item\label{A2303} If $x_2-x_3\neq \xqz{x_2-x_3}, \xqz{x_2-x_3}\leq k_{23}-2, \text{ and } \sqz{x_0-x_3}\leq\mathcal{K}_{03}-1$
then we can choose
\[
p_2=\xqz{x_0-x_2} \text{ and } p_3=\sqz{x_0-x_3}.
\]
\end{enumerate}
\item\label{A231} If the relation $\mathcal{R}_{23}^1$ holds, then we have the following possible choices
\begin{enumerate}[(1)]
\item\label{A2311} $p_2=\sqz{x_0-x_2}$ and $p_3=\xqz{x_0-x_3}$,
\item\label{A2312} If $\xqz{x_2-x_3}\leq k_{23}-2$, then we can choose $p_2=\xqz{x_0-x_2}$ and $p_3=\xqz{x_0-x_3}$, or
\item\label{A2313} If $\xqz{x_2-x_3}\leq k_{23}-2$ and $\sqz{x_0-x_3}\leq\mathcal{K}_{03}-1$, then we can choose $p_2=\sqz{x_0-x_2}$ and $p_3=\sqz{x_0-x_3}$.
\end{enumerate}
\end{enumerate}
By Lemma \ref{triple1}, Corollary \ref{triple2} , Corollary \ref{pijcor} and above analysis \eqref{A120}-\eqref{A121}-\eqref{A130}-\eqref{A131}-\eqref{A230}-\eqref{A231}, for this case, we have the following choices
\begin{enumerate}[(a)]
\item If $\mathcal{R}_{12}^0$, $\mathcal{R}_{13}^0$ and $\mathcal{R}_{23}^0$ hold, we can choose \eqref{A1201}-\eqref{A1301}-\eqref{A2301}
we can find
\[
p_i=\xqz{x_0-x_i}, z_i=x_i+p_i, i=1,2, 3,
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\item If $\mathcal{R}_{12}^0$, $\mathcal{R}_{13}^1$ and $\mathcal{R}_{23}^1$ hold, we can choose \eqref{A1202}-\eqref{A1311}-\eqref{A2311}, we can find
\[
p_1=\sqz{ x_0-x_1},\quad p_2=\sqz{ x_0-x_2} , \quad p_3=\xqz{x_0-x_3}, \quad z_i=x_i+p_i, i=1,2, 3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\item If $\mathcal{R}_{12}^1$, $\mathcal{R}_{13}^1$ and $\mathcal{R}_{23}^0$ hold, we can choose \eqref{A1211}-\eqref{A1311}-\eqref{A2301}, we can find
\[
p_1=\sqz{x_0-x_1}, \quad p_2=\xqz{ x_0-x_2}, \quad p_3=\xqz{ x_0-x_3} \quad z_i=x_i+p_i, i=1, 2,3,
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\item
If $\mathcal{R}_{12}^1$, $\mathcal{R}_{13}^0$ and $\mathcal{R}_{23}^0$ hold,
\begin{enumerate}[(i)]
\item if $\xqz{x_2-x_3}\leq k_{23}-2, x_2-x_3\neq \xqz{x_2-x_3} \text{ and } \xqz{x_0-x_3}\leq \mathcal{K}_{03}-2,$
we can choose \eqref{A1211}-\eqref{A1302}-\eqref{A2303},
we can find
\[
p_1=\sqz{x_0-x_1}, \quad p_2=\xqz{ x_0-x_2}, \quad p_3=\sqz{ x_0-x_3}, z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$,
\item if $\xqz{x_1-x_2}\leq k_{12}-2$, we can choose \eqref{A1212}-\eqref{A1301}-\eqref{A2301} ,we can find
\[
p_i=\xqz{x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
or moreover $\xqz{x_0-x_3}\leq \mathcal{K}_{03}-2$, we can also choose \eqref{A1213}-\eqref{A1302}-\eqref{A2302}, we can find
\[
p_i=\sqz{x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$,
\end{enumerate}
\item
If $\mathcal{R}_{12}^0$, $\mathcal{R}_{13}^0$ and $\mathcal{R}_{23}^1$ hold,
\begin{enumerate}[(i)]
\item
if $\xqz{x_2-x_3}\leq k_{23}-2$, we can choose \eqref{A1201}-\eqref{A1301}-\eqref{A2312},
we can find
\[
p_i=\xqz{x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
or moreover if $\sqz{x_0-x_3}\leq \mathcal{K}_{03}-1$, we can also choose \eqref{A1202}-\eqref{A1302}-\eqref{A2313},
we can find
\[
p_i=\sqz{x_0-x_i}, z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$,
\item
if $x_1-x_2\neq\xqz{x_1-x_2}$ , $\xqz{x_1-x_2}\leq k_{12}-2$,
we can choose \eqref{A1203}-\eqref{A1301}-\eqref{A2311},
we can find
\[
p_1=\xqz{x_0-x_1}, p_2=\sqz{x_0-x_2}, p_3=\xqz{x_0-x_3}, z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$,
\end{enumerate}
\item
If $\mathcal{R}_{12}^1$, $\mathcal{R}_{13}^1$ and $\mathcal{R}_{23}^1$ hold,
\begin{enumerate}[(i)]
\item
if $\xqz{x_2-x_3}\leq k_{23}-2$, we can choose \eqref{A1211}-\eqref{A1311}-\eqref{A2312}, we can find
\[
p_1=\sqz{x_0-x_1}, p_2=\xqz{x_0-x_2}, p_3=\xqz{x_0-x_3}, z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$,
\item
if $\xqz{x_1-x_2}\leq k_{12}-2$, we can choose \eqref{A1212}-\eqref{A1311}-\eqref{A2311}, we can find
\[
p_1=\sqz{x_0-x_1},p_2=\sqz{x_0-x_2},p_3=\xqz{x_0-x_3} z_i=x_i+p_i, i=1,2,3
\]
such that $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$.
\end{enumerate}
\end{enumerate}
If $\mathcal{R}_{12}^1$, $\mathcal{R}_{13}^0$ and $\mathcal{R}_{23}^0$ hold, i.e.
\begin{align*}
&\xqz{x_0-x_2}-\xqz{x_0-x_1}=\xqz{x_1-x_2}+1\\
&\xqz{x_0-x_3}-\xqz{x_0-x_1}=\xqz{x_1-x_3}\\
&\xqz{x_0-x_3}-\xqz{x_0-x_2}=\xqz{x_2-x_3}
\end{align*}
then we have
\[
\xqz{x_1-x_3}=\xqz{x_1-x_2}+\xqz{x_2-x_3}+1
\]
If moreover,
\[
\xqz{x_1-x_2}=k_{12}-1,\xqz{x_2-x_3}=k_{23}-1
\]
then we have
\[
\xqz{x_1-x_3}=\xqz{x_1-x_2}+\xqz{x_2-x_3}+1=k_{12}+k_{23}-1
\]
But
\[
\xqz{x_1-x_3}\leq \mathcal{K}_{13}-1\leq k_{12}+k_{23}-2
\]
This is a contraction. Hence, if $\xqz{x_1-x_2}=k_{12}-1$ and $\xqz{x_2-x_3}=k_{23}-1$ hold then $\mathcal{R}_{12}^1$, $\mathcal{R}_{13}^0$ and $\mathcal{R}_{23}^0$ does not hold at the same time. And the same reason, if $\xqz{x_1-x_2}=k_{12}-1$ and $\xqz{x_2-x_3}=k_{23}-1$ hold then
\begin{itemize}
\item
$\mathcal{R}_{12}^1$, $\mathcal{R}_{13}^1$ and $\mathcal{R}_{23}^1$ does not hold at the same time.
\item
$\mathcal{R}_{12}^0$, $\mathcal{R}_{13}^0$ and $\mathcal{R}_{23}^1$ does not hold at the same time.
\end{itemize}
By above analysis, we can write $\Delta_{\uppercase\expandafter{\romannumeral4}}$ by
\[
\Delta_{\uppercase\expandafter{\romannumeral4}}=\Delta_{\uppercase\expandafter{\romannumeral4}}^1\cup\Delta_{\uppercase\expandafter{\romannumeral4}}^2\cup\Delta_{\uppercase\expandafter{\romannumeral4}}^3
\]
where
\begin{eqnarray*}
&\Delta_{\uppercase\expandafter{\romannumeral4}}^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\mathcal{R}_{23}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_2-x_3}=x_2-x_3,
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral4}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cl}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\mathcal{R}_{23}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_0-x_3}=\mathcal{K}_{03}-1,
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral4}}^3=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cl}
&\quad\mathcal{R}_{12}^0,\mathcal{R}_{13}^0,\mathcal{R}_{23}^1,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}<\mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=x_1-x_2,\xqz{x_2-x_3}=k_{23}-1,
\end{array}
}
\end{eqnarray*}
\end{proof}
Using inequality \eqref{xqz}, Lemma \ref{triple1}, Lemma \ref{lemmacase1}, Lemma \ref{lemmacase2}, Lemma \ref{lemmacase3}, Lemma \ref{lemmacase4}, Corollary \ref{triple2} and Corollary \ref{pijcor}, we will prove the Theorem \ref{thmmain} of this paper:
\begin{proof}
We write
\[
\Theta_2=
\bigcup_{(0,p_1,p_2,p_3)\in A_0}\hkh{(x_0,x_1,x_2,x_3)\Big|
\begin{array}{cll}
-1+p_i&<x_0-x_i<&1+p_i, i=1,2,3\\
-1+p_k-p_j&<x_j-x_k<&1+p_k-p_j ,1\leq j<k\leq 3
\end{array}
}.
\]
By the definition of $A_0$, the following inequalities hold
\begin{align}
&p_1\leq k_{01}-1, \label{pinequ1}\\
& p_2\leq \mathcal{K}_{02}-1, \label{pinequ2}\\
& p_3\leq\mathcal{K}_{03}-1,\label{pinequ3}\\
&p_2-p_1\leq k_{12}-1, \label{pinequ4}\\
&p_3-p_2\leq k_{23}-1,\label{pinequ5}\\
&p_3-p_1\leq \mathcal{K}_{13}-1\label{pinequ6}
\end{align}
By the above inequalities \eqref{pinequ1}-\eqref{pinequ6}, we have
\[
\Theta_2\subseteq \Theta_1
\]
By the homeomorphism \eqref{homeo}, $\Theta_\mathcal{E}$ is homeomorphism to $\mathbb{R}_{>0}^3\times \Theta_2$. To prove the theorem, we need to show that for every $(x_0,x_1,x_2,x_3)\in\Theta_1\setminus \Delta$, there exist $(0,p_1,p_2,p_3)\in A_0$ and $z_i\in S^3(-1,+1)$ such that $(x_0,x_1,x_2,x_3)=(z_0,z_1,z_2,z_3)-(0,p_1,p_2,p_3)$ and $\Theta_2\cap\Delta=\emptyset$. By the definition of $S^3(-1,+1)$, we need to choose $z_i$ such that $z_i\in (x_0-1,x_0+1)$, hence the possible choices of $p_i$ are $p_i=\xqz{x_0-x_i}$ or $p_i=\sqz{x_0-x_i}$ and $z_i=x_i+p_i$. We need to check the inequalities \eqref{pinequ1}-\eqref{pinequ6} for $p_i$ and $z_i\in S^3(-1,+1)$.
By relations $\mathcal{R}_{ij}^0$, $\mathcal{R}_{ij}^1$, the inequalities \eqref{pinequ1}-\eqref{pinequ6} and Corollary \ref{pijcor}, we need to consider the following four cases.
\begin{enumerate}[(I)]
\item \label{case1}$\xqz{ x_0-x_1}=k_{01}-1$ and $\xqz{ x_0-x_2}=\mathcal{K}_{02}-1$
\item \label{case2}$\xqz{ x_0-x_1}=k_{01}-1$ and $\xqz{ x_0-x_2}<\mathcal{K}_{02}-1$
\item \label{case3}$\xqz{ x_0-x_1}<k_{01}-1$ and $\xqz{ x_0-x_2}=\mathcal{K}_{02}-1$
\item \label{case4} $\xqz{ x_0-x_1}<k_{01}-1$ and $\xqz{ x_0-x_2}<\mathcal{K}_{02}-1$
\end{enumerate}
By the above analysis, we can describe the set $\Delta$ as follows.
\begin{itemize}
\item For Case \ref{case1}, by Lemma \ref{lemmacase1}, there are no elements in $\Delta$.
\item For Case \ref{case2}, by Lemma \ref{lemmacase2}, Then the set of elements of $\Delta$ in this case is $\Delta_{\uppercase\expandafter{\romannumeral2}}$
\item For Case \ref{case3}, by Lemma \ref{lemmacase3}, Then the set of elements of $\Delta$ in this case is $\Delta_{\uppercase\expandafter{\romannumeral3}}$
\item For Case \ref{case4}, by Lemma \ref{lemmacase4} , Then the set of elements of $\Delta$ in this case is $\Delta_{\uppercase\expandafter{\romannumeral4}}$
\end{itemize}
By the Corollary \ref{triple2}, we can rewrite $\Delta_{\uppercase\expandafter{\romannumeral4}}^1,\Delta_{\uppercase\expandafter{\romannumeral4}}^2,\Delta_{\uppercase\expandafter{\romannumeral4}}^3$ as
\begin{eqnarray*}
&\Delta_{\uppercase\expandafter{\romannumeral4}}^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_2-x_3}=x_2-x_3,
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral4}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cl}
&\quad\mathcal{R}_{12}^1,\mathcal{R}_{13}^0,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_0-x_3}=\mathcal{K}_{03}-1,
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral4}}^3=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cl}
&\quad\mathcal{R}_{13}^0,\mathcal{R}_{23}^1,\\
&\xqz{x_0-x_1}<k_{01}-1,\xqz{x_0-x_2}<\mathcal{K}_{02}-1, \\
&\xqz{x_1-x_2}=x_1-x_2,\xqz{x_2-x_3}=k_{23}-1,
\end{array}
}
\end{eqnarray*}
By the Corollary \ref{triple2}, we can rewrite $\Delta_{\uppercase\expandafter{\romannumeral2}}$ as $\Delta_{\uppercase\expandafter{\romannumeral2}}=\Delta^1\cup\Delta_{\uppercase\expandafter{\romannumeral2}}^2$,
where
\begin{eqnarray*}
&\Delta^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\quad\mathcal{R}_{23}^1,\quad\mathcal{R}_{13}^1\\
&\xqz{x_0-x_1}=k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_2-x_3}=k_{23}-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}\\
&\Delta_{\uppercase\expandafter{\romannumeral2}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\quad\mathcal{R}_{23}^1,\quad\mathcal{R}_{13}^0\\
&\xqz{x_0-x_1}=k_{01}-1,\xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_2-x_3}=k_{23}-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}
\end{eqnarray*}
Then we hvae
\begin{eqnarray*}
&\Delta^2:=\Delta_{\uppercase\expandafter{\romannumeral4}}^3\cup\Delta_{\uppercase\expandafter{\romannumeral2}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\mathcal{R}_{23}^1,\quad\mathcal{R}_{13}^0,\quad \xqz{x_0-x_2}< \mathcal{K}_{02}-1, \\
&\xqz{x_2-x_3}=k_{23}-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}\\
&\Delta^3:=\Delta_{\uppercase\expandafter{\romannumeral3}}^1\cup\Delta_{\uppercase\expandafter{\romannumeral4}}^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\mathcal{R}_{12}^1,\mathcal{R}_{13}^0, \quad \xqz{x_0-x_1}<k_{01}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,x_2-x_3=\xqz{x_2-x_3}
\end{array}
}\\
&\Delta^4:=\Delta_{\uppercase\expandafter{\romannumeral3}}^2\cup\Delta_{\uppercase\expandafter{\romannumeral4}}^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\mathcal{R}_{12}^1,\mathcal{R}_{13}^0, \quad \xqz{x_0-x_1}<k_{01}-1, \\
&\xqz{x_1-x_2}=k_{12}-1,\xqz{x_0-x_3} =\mathcal{K}_{03}-1
\end{array}
}
\end{eqnarray*}
For the convenience we rewrite $\Delta_{\uppercase\expandafter{\romannumeral3}}^3$ by $\Delta^5$.
Hence we prove that
\[
\Delta=\Delta_{\uppercase\expandafter{\romannumeral2}}
\cup \Delta_{\uppercase\expandafter{\romannumeral3}}\cup\Delta_{\uppercase\expandafter{\romannumeral4}}=\bigcup_{i=1}^5\Delta^i.
\]
By the construction of $\Delta$, we have $\Theta_2\cap \Delta=\emptyset$.
We complete the proof.
\end{proof}
\section{Examples}\label{example}
In this section we give an application of the Theorem \ref{thmmain}. Let $X$ be a quadric surface, which isomorphism to $\mathbb{P}^1\times \mathbb{P}^1$, and $\operatorname{\mathcal{D}}^b(X)$ is the bounded derived category of coherent sheaves on $X$. We use the notation
\[
\mathcal{O}(a,b)=\pi_1^*\mathcal{O}_{\mathbb{P}^1}(a)\otimes\pi_2^*\mathcal{O}_{\mathbb{P}^1}(b)
\]
where $\pi_1,\pi_2: X\to\mathbb{P}^1$ are the projections.
Then $\mathcal{E}=\ykh{\mathcal{O},\mathcal{O}(1,0),\mathcal{O}(0,1),\mathcal{O}(1,1)}$ is a complete strong exceptional collection on $\operatorname{\mathcal{D}}^b(X)$. For $\mathcal{E}$, we have
\[
k_{ij}=0,\mathcal{K}_{02}=\mathcal{K}_{13}=-1, \text{ and } \mathcal{K}_{03}=-2.
\]
Then we can rewrite the Theorem \ref{thmmain} as follwes:
\begin{prop}
Let $X$ be a quadric surface, $\mathcal{E}=\ykh{\mathcal{O},\mathcal{O}(1,0),\mathcal{O}(0,1),\mathcal{O}(1,1)}$ be a complete strong exceptional collection on $\operatorname{\mathcal{D}}^b(X)$.
Then
\[
\Theta_1=\hkh{
(y_0,y_1,y_2,y_3)\Bigg|\begin{array}{cl}
&y_0<y_1<y_2<y_3, \quad y_1-y_3<-1\\
&y_0-y_3<-2,\quad y_0-y_2<-1
\end{array}
}.
\]
and we can express $\Theta_{\mathcal{E}}$ as followes:
\[
\mathbb{R}_{>0}^4\times \ykh{\Theta_1\setminus\Delta}
\]
where $\Delta$ is the union of following five sets
\begin{eqnarray*}
&\Delta^1=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
&\xqz{x_1-x_3}=\xqz{x_0-x_3},\\
&\xqz{x_0-x_1}=-1,\xqz{x_0-x_2}=\xqz{x_0-x_3}<-2, \\
&\xqz{x_2-x_3}=-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}\\
&\Delta^2=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{cll}
& \xqz{x_0-x_3}=\xqz{x_0-x_2}< -2, \\
&\xqz{x_0-x_3}-\xqz{x_0-x_1}=\xqz{x_1-x_3},\\
&\xqz{x_2-x_3}=-1,x_1-x_2=\xqz{x_1-x_2}
\end{array}
}\\
&\Delta^3=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\xqz{x_0-x_1}=\xqz{x_0-x_2}<-1, \\
&\xqz{x_0-x_3}-\xqz{x_0-x_1}=\xqz{x_1-x_3},\\
&\xqz{x_1-x_2}=-1,x_2-x_3=\xqz{x_2-x_3}
\end{array}
}\\
&\Delta^4=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\xqz{x_0-x_1}=\xqz{x_0-x_2}<-1, \\
&\xqz{x_0-x_1}+\xqz{x_1-x_3}+3=0,\\
&\xqz{x_1-x_2}=-1,\xqz{x_0-x_3}=-3
\end{array}
}\\
&\Delta^5=\hkh{(x_0,x_1,x_2,x_3)\in\Theta_1\Bigg|
\begin{array}{clll}
&\xqz{x_0-x_1}<-1,\\
&\xqz{x_0-x_2}= -2,\xqz{x_1-x_3}= -2\\
&\xqz{x_0-x_1}=-\xqz{x_1-x_2}-2=\xqz{x_0-x_3}+1
\end{array}
}.
\end{eqnarray*}
\end{prop}
\begin{remark}
The author's original motivation was to give a detail description of stability conditions on $X$ and excepted to get a contractible connected component of the spaces of stability conditions on $\operatorname{\mathcal{D}}^b(X)$.
Compared to the $\mathbb{P}^2$ case \cite{lichunyi2017,macri2004}, $\mathbb{P}^1\times \mathbb{P}^1$ case is more difficult. As compute in the Theorem \ref{thmmain}, $\Delta$ is the union of five sets. We can not give a simple description of $\Delta$ now. We wish to prove that $\Theta_2$ is contractible. Another problem is that the strong exceptional collection on $\mathbb{P}^1\times \mathbb{P}^1$ is not stong after the mutation. The space of stability condition generated by exceptional collectioon on $\mathbb{P}^1\times \mathbb{P}^1$ is still not clear.
\end{remark}
\subsection*{Acknowledgements} I would like to thank to Professor Sen Hu and Dr. Teng Huang for their encouragement and valuable comments.
\renewcommand{\refname}{Reference}
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{sec:intro}
The popular ``pre-training + fine-tuning'' paradigm has revolutionized NLP ~\cite{devlin-etal-2019-bert,liu2019roberta,yang2019xlnet,lewis-etal-2020-bart} and CV ~\cite{chen2020simple,he2020momentum} over the last few years.
Although models trained this way can achieve impressive performance, they still require task-specific annotated data and fine-tuning for each end task.
Recent work adopt pre-training for zero-shot transfer to end tasks without fine-tuning, including GPT~\cite{radford2018improving,radford2019language,brown2020language} for NLP tasks and CLIP~\cite{radford2021learning} for image classification.
\begin{figure}[!t]
\centering
\vspace{-10pt}
\includegraphics[width=\linewidth]{img/CROP_merge_.pdf}
\vspace{-15pt}
\caption{VideoCLIP aims for zero-shot video understanding via learning fine-grained association between video and text in a transformer using a contrastive objective with two key novelties: (1) for \textit{positive} pairs, we use video and text clips that are \textit{loosely} temporarily overlapping instead of enforcing strict start/end timestamp overlap; (2) for \textit{negative} pair
, we employ a retrieval based sampling technique that uses video clusters to form batches with mutually harder videos.} \vspace{-15pt}
\label{fig:overview}
\end{figure}
This paper focuses on pre-training for zero-shot transfer to \textit{video}-text understanding tasks.
Our approach pre-trains a Transformer model~\cite{vaswani2017attention,devlin-etal-2019-bert} with a contrastive objective~\cite{oord2018representation,chen2020simple} using pairs of video-text clips.
Different from CLIP that scales pre-training data for zero-shot transfer to image classification on an explicitly assembled dataset using a simple contrastive objective~\cite{chen2020simple}, this paper uses a \textit{publicly established} pre-training dataset, HowTo100M~\cite{miech2019howto100m}, for~\textit{zero-shot video understanding}.
We show that the resulting pre-trained model can be either directly applied to, or fine-tuned on, a series of video-text tasks at both the global sequence and local clip/token level.
We find that straightforward objectives~\cite{chen2020simple} lead to poor results, and hypothesize that learning fine-grained associations between video and text is crucial for success of zero-shot transfer to end tasks. Since end tasks may require different granularities of video-text correspondence. The granularity can be about sequence length (such as long video versus short text (\textit{e}.\textit{g}. classification), token level or sequence level) and semantics (``apple'' vs ``banana'' or ``apple'' vs ``car''). Previous efforts sample short, temporally aligned video and text clips with contrastive learning within a random batch, falling short on learning the fine-grained association between video frames and word tokens.
We present VideoCLIP
that
aims to pre-train a \textit{unified} video-text representation with contrastive learning using two key techniques (see~Fig.~\ref{fig:overview}) to compute the training objective.
First,
we aim to improve the association of video and text with different sequence lengths.
Although the majority of video clips and text transcriptions are not semantically aligned~\cite{miech2019howto100m}, current video-text models are trained with exact temporal alignment.
As a result, multiple or longer text clips may have better alignment with a video clip~\cite{miech2020end} and many
clips may not have any corresponding captions (see a detailed discussion of issues in \S \ref{sec:overlap}).
To address these issues, we pre-train with temporally \textbf{overlapped} pairs of video and text clips (of varying length), thereby greatly increasing the quality and quantity of the video-text alignment.
We show in experiments that this simple and general approach significantly improves performance.
Second, we learn fine-grained video-text similarity from a contrastive loss with a new method for gathering (implicitly) harder negative pairs.
Although existing works contrast intra-video clips via sampling multiple clips from the same video \cite{miech2019howto100m,miech2020end}, we find that mining clips from other videos can provide much more challenging negatives.
We propose a \textbf{retrieval augmented pre-training} approach to retrieve a cluster of videos that are similar to each other for each training batch.
Retrieval-augmented pre-training alternatively performs retrieving video clusters and uses the retrieved video clusters for pre-training
(see \S~\ref{sec:retri} for details).
After pre-training, we apply our model for zero-shot transfer \textit{without} any fine-tuning on target dataset labels. We directly use our pre-trained model on a diverse set of~\textit{four} tasks in ~\textit{five} datasets, including text-video retrieval (for text-to-video similarity), VideoQA (for video-to-text similarity),
action localization (for video frame to text label similarity) and segmentation (for video token to text label similarity with rejection) (see \S\ref{sec:zs}).
Our experiments reveal that VideoCLIP has strong performance, even compared to supervised approaches which use human-annotated labels on the downstream tasks. For example, in text-video retrieval on Youcook2~\cite{zhou2017towards}, VideoCLIP outperforms all existing zero-shot methods and even outperforms fully supervised pre-training + fine-tuning methods, but without using any labels.
In summary, the main contributions of this paper include:
\textit{(i}) we propose to pre-train a \textit{unified} model that is capable of zero-shot transfer to \textit{multiple} end tasks for video-text understanding, even surpassing fully-supervised methods in some cases,
and (\textit{ii}) we introduce two novel techniques to improve the learning of fine-grained video-text association.
\vspace{5pt}
\section{Related Work}
\label{sec:related_work}
\paragraph{Pre-training for Zero-shot Transfer.}
Recently, the paradigm of pre-training has made impressive progress with the scale of training data and computational power.
For example, in NLP, the paradigm has shifted from learning word embeddings for task-specific architecture \cite{mikolov2013distributed,bojanowski2017enriching,peters2018deep}, to pre-training+fine-tuning~\cite{devlin-etal-2019-bert,liu2019roberta,lewis-etal-2020-bart} and few-shot/zero-shot transfer~\cite{radford2018improving,radford2019language,brown2020language,alayrac2020self,ramesh2021zero}
that have task-agnostic architecture.
One line of pre-training for zero-shot transfer focuses on generative (auto-regressive) models~\cite{radford2018improving,radford2019language,brown2020language}, where examples and prompts of an end task are used as context for a language model to respond properly to that task~\cite{brown2020language}; the other line of studies focuses on discriminative models~\cite{alayrac2020self,miech2020end}, where a similarity search or ranking model learns a joint space (\textit{e}.\textit{g}. ~via contrastive learning~\cite{chen2020simple,he2020momentum}) and later transfer to a particular task.
Recently, CLIP~\cite{radford2021learning} transfers image-text similarity to many image classification tasks, where the text branch serves as supervision for learning a general image representation and subsequently serves as a hyper network for downstream vision tasks.
Our effort aligns with the latter line of work, but is the first to transfer a pre-trained discriminative model to a broad range of tasks in multi-modal video understanding.
\paragraph{Multi-modal Video-Text Pre-training.}
Multi-modal models have also adopted the pre-training+fine-tuning paradigm.
One line of work
adopts multiple unimodal encoders for retrieval tasks.
For example, \cite{miech2019howto100m,miech2020end,ging2020coot,gabeur2020multi,alayrac2020self,patrick2021supportset,huang21naacl} adopt contrastive learning for pre-training and shows the possibility of zero-shot transfer to text-video retrieval tasks.
CBT \cite{sun2019contrastive}, HERO \cite{li-etal-2020-hero}, VideoAsMT \cite{korbar2020video} and UniVL \cite{luo2020univilm} adopt multi-task learning (MTL) for pre-training on retrieval tasks.
HERO \cite{li-etal-2020-hero} and UniVL \cite{luo2020univilm} further adopt a cross-encoder to further learn the fusion of different modalities.
The other line of work adopts a single \textit{cross-modal} encoder and concatenates the vision and text sequences as inputs, including
VideoBERT \cite{sun2019videobert}, Unicoder-VL \cite{li2020unicoder}, VL-BERT \cite{Su2020VL-BERT:}, UNITER \cite{chen2020uniter}, VLP \cite{zhou2018end},
ActBERT \cite{zhu2020actbert} and VLM \cite{xu-etal-2021-vlm}.
Although this approach is intuitive, it limits the capability of zero-shot transfer. For example, it is non-trivial to perform retrieval tasks on a single encoder as feeding vision and text in a pairwise manner is not flexible and data efficient~\cite{luo2020univilm}.
\noindent\textbf{Retrieval Augmented Training.}
Augmenting traditional training with a non-parametric retrieval component has recently shown impressive results in pre-training
\cite{khandelwal2019generalization,guu2020realm,lewis2020pre}
and QA
\cite{izacard2020leveraging,karpukhin2020dense}.
We find that contrastive learning and retrieval augmented training can have good synergy because the former aims to discriminate examples and the latter aims to find harder examples for discrimination.
To the best of our knowledge, there is no existing work of retrieval augmented training for video, perhaps because videos exhibit unique challenges for data-efficient training (see \S\ref{sec:retri}).
\section{VideoCLIP Pre-training}
\label{sec:pretrain}
In the paradigm of multi-modal video-text pre-training for zero-shot transfer, the key challenge is to learn fine-grained association in-between video and text to cover the diverse needs of end tasks.
We cover VideoCLIP pre-training in this section, and discuss the needs of zero-shot transfer to different end tasks in the next section.
We first describe video and text model backbone and contrastive loss;
then we propose overlapped video and text clips to improve the association of positive pairs; lastly, we describe retrieval augmented pre-training to improve the mining of negative examples.
\subsection{Video and Text Encoding}
VideoCLIP consumes pairs of video and text clips $(v, t)$ as inputs.
It makes no assumptions on the encoder architectures and can work with any video and text backbone.
We use Transformer~\cite{vaswani2017attention} model for both the video and text.
The video features, extracted by a convolutional neural network (CNN), are first projected to \textit{video tokens} before fed into our video transformer, as described next.
\paragraph{Video and Text Transformers.}
Let $\bm{c}_v$ be a video clip of a sequence of continuous frames (we use \textit{bold} symbols to indicate sequences). We feed $\bm{c}_v$ into a (frozen) pre-trained video encoder $f_{\theta_\text{CNN}}$ and then apply a trainable MLP, $f_{\theta_\text{MLP}}$, with weights $\theta_\text{MLP}$ to obtain \textit{video token}s $\bm{x}_v \in \mathbb{R}^d$ with the same embedding dimension, $d$, as for word embeddings in our architecture:
\begin{equation}
\begin{split}
\bm{x}_v = f_{\theta_\text{MLP}}(\texttt{stopgrad}(f_{\theta_\text{CNN}}(\bm{c}_v))),
\end{split}
\end{equation}
where $\texttt{stopgrad}$ is a stop-gradient operation, to reflect that the video CNN is frozen.
Similarly, vectors for text tokens $\bm{x}_t$ are obtained via embedding lookup as in BERT~\cite{devlin-etal-2019-bert}.
Then
$\bm{x}_v$ and $\bm{x}_t$ are feed into
two separate trainable Transformers,
$f_{\theta_\text{v}}$ and $f_{\theta_\text{t}}$, respectively,
to obtain the hidden states for video and text tokens
\begin{equation}
\label{eq:forward}
\begin{split}
\bm{h}_v = f_{\theta_\text{v}}(\bm{x}_v), \bm{h}_t = f_{\theta_\text{t}}(\bm{x}_t).
\end{split}
\end{equation}
To obtain the hidden states (\textit{i.e.}~global features) of video and text clips, we apply average pooling over the sequence of tokens for video and text, respectively
\begin{equation}
\begin{split}
{z}_v = \texttt{AvgPool}(\bm{h}_v), {z}_t = \texttt{AvgPool}(\bm{h}_t).
\end{split}
\end{equation}
We use average pooling (instead of using the \texttt{[CLS]} token) to encourage $f_{\theta_\text{v}}$ and $f_{\theta_\text{t}}$ to learn token-level representations that may benefit token-level tasks, such as action localization and action segmentation (see Section~\ref{sec:zs}).
VideoCLIP aims at pre-training the \textit{unified} video-text representation, captured by the
Transformer model parameters $\theta_\text{v}$ and $\theta_\text{t}$ for video and text, and consequently use it for zero-shot downstream tasks.
In appendix, we also explore \textit{shared weights} for video and text, $\theta_\text{v} \equiv \theta_\text{t}$, and our ablations show that separate video/text transformers yields slightly better performance.
Notably, using a frozen video backbone ($f_{\theta_\text{CNN}}$) enables us to go beyond short-term visual input (typical video CNNs~\cite{xie2018rethinking,feichtenhofer2019slowfast} only capture temporal windows of \app3 seconds), and allows us to model \textit{long-term visual-textual correspondences} spanning \app32 seconds. We describe our training methodology next.
\subsection{Contrastive Loss}
We use a contrastive loss (\mbox{InfoNCE} \cite{oord2018representation} objective) to learn the correspondence between video and text.
In particular, we minimize the sum of two multi-modal contrastive losses:
\begin{equation}
\small
\label{eq}
\mathcal{L}=-\sum_{(v, t)\in B}\Big(\log\text{NCE}(z_v, z_t)
+ \log \text{NCE}(z_t, z_v)\Big),
\end{equation}
where $B$ is
the batch
that contains sampled video-text pairs and $\text{NCE}(z_v, z_t)$ and $\text{NCE}(z_t, z_v)$ corresponds to the contrastive loss on video-to-text similarity and text-to-video similarity. Specifically, the video-to-text contrastive loss is given by
\begin{equation}\label{eq:nce}
\text{NCE}(z_v, z_t) = {\frac{ \exp\left({z_v \cdot z_t^+ / \tau}\right)}{ {\sum_{z \in \{z_t^+, z_t^-\}}} {\exp\left({z_v \cdot z / \tau}\right)} }},
\end{equation}
with $\tau$ being a temperature hyper-parameter and $z_{t}^{+}$ are \textit{positive} embedded text clips overlapping with video clip embedding $z_v$, and $\{z_{t}^{-}\}$ are \textit{negative} embedded text clips that are implicitly formed by other text clips in the training batch.
The text-to-video loss $\text{NCE}(z_t, z_v)$ is defined symmetrically.
The next sections (\S\ref{sec:overlap} and \S\ref{sec:retri}) describe how we construct the positive, $z_t^{+}$, and negatives, $\{z_t^{-}\}$, in our pre-training objective \eqref{eq:nce}.
\subsection{Overlapped Video-Text Clips}
\label{sec:overlap}
To build overlapping \textit{positive} video/text pairs, we
(\textit{i}) sample a text clip (because sampling a video clip first may not have nearby corresponding text);
(\textit{ii}) sample a timestamp within the boundary of text clip as the \textit{center} for a video clip;
(\textit{iii}) grow a video clip with random duration (up to \app32 seconds) from this center timestamp.
Our empirical results show this simple method works well in practice, and we discuss its benefits w.r.t.~prior efforts next.
\noindent \textbf{Low Relevance Temporal Alignment.}
Existing video-text pre-training methods, \textit{e}.\textit{g}.,~\cite{miech2019howto100m}, consider temporally exactly aligned clips (video and text clips sharing the same start/end timestamps).
Although strict alignment seems natural, it is less likely that temporally aligned video and text clips are also semantically close in short clips.
For example, a video clip of ``\textit{a person speaking}'' may have a low relevance\footnote{We use the term low relevance instead of noisy alignment because temporally aligned clips may still have low relevance on certain perspectives, such as positive emotions, an opened mouth with any transcription popping up, and ``going to'' in transcription indicates visual contents may show up later.}
with the exact temporally aligned transcription ``\underline{I am going to show you how to cook fried rice}''.
However, a later video clip showing ``\textit{rice in wok}'' may have a better semantic visual alignment.
One explanation for this low relevance of temporal alignment is that humans are less likely to speak and perform actions simultaneously.
Using exact temporal alignment limits the examples considered in the contrastive loss.
Taking the previous $\text{NCE}(z_v, z_t)$ term as an example, the low relevance (positive) pair could be in the numerator of the objective \eqref{eq:nce}, whereas higher relevance pairs (\textit{e}.\textit{g}.~\textit{rice in wok} appearing later in a video with an introductionary text clip of ``{I am going to show you how to cook fried rice}'') are possibly used as negative pairs, under exact temporal alignment for constructing positive/negative samples.
Although existing work \cite{miech2020end} aligns multiple nearby text clips with one (short) video clip of fixed 3.2 seconds duration, this only partially solves the low relevance problem and can attenuate noise, as the text clips may only partially correspond to the visuals and might have no temporal overlap with the short-duration video clip per se.
\noindent \textbf{Better Video-Text Association.}
As such, we believe a (self-supervised) method that can curate higher relevance video-text pairs at a large-scale is crucial for effective learning.
Our approach to sample video and text pairs $(v, t)$ of different lengths while requiring \textit{temporal overlap} improves video-text relevance and encourages fine-grained association.
As such, a video (or text clip) can have a better chance to be aligned or supervised by nearby text and vice versa. By contrast, video clips without any temporally aligned text are never contributing as a positive video-text pair in our objective.
\subsection{Retrieval Augmented Training}
\label{sec:retri}
Our intention is to learn to model more fine-grained video-text similarity by using difficult examples in our contrastive pre-training objective \eqref{eq:nce}.
We construct \textit{negatives} in our training batch by using hard pairs $\{z_t^{-}\}$, which are semantically to the pairs in the numerator, using retrieval based sampling.
Recall that contrastive loss (\textit{e}.\textit{g}. in equation \eqref{eq:nce}) uses positive pairs in a batch $B$, and typically negative pairs are implicitly induced from other positive pairs in the same batch.
\noindent \textbf{Dense Video Cluster Retrieval.}
Our approach aims to find video clusters to construct a batch of training samples.
We formulate this as a dense retrieval process on the latent space of a video, derived from the video/text embeddings of our transformer that is trained by the contrastive loss~\eqref{eq:nce}.
Our overall training process can be described as a two-stage method that alternatively performs \textit{retrieval} and \textit{training} in each epoch, and is summarized in Algorithm~\ref{alg:simulator}.
\begin{algorithm}[!t]
\LinesNumbered
\DontPrintSemicolon
\caption{Retrieval Augmented Training}
\label{alg:simulator}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\SetKwProg{Def}{def}{:}{}
\Input{$\mathcal{V}$ is video set; $M$ is model.}
\BlankLine
\ForEach{epoch}{
infer global features for all videos $\mathcal{V}$ on $M$: each video $V\in\mathcal{V}$'s global feature is computed as $z_V=\frac{1}{2|B_V|}\sum_{(v, t) \in B_V} (z_v + z_t)$, where $B_V$ indicates all clip pairs of $V$;\;
build dense index on all videos' $z_V$;\;
retrieve $|\mathcal{C}|$ video clusters, where each cluster $c\in\mathcal{C}$ is sampled as $c \sim k\text{NN}(z_V, 2k)$, $|c| = k$ from a random video $V$;\;
sample overlapped video-text pairs from $c\in \mathcal{C}$ to train $M$.\;
}
\end{algorithm}
For each epoch, Line 2-4 corresponds to the retrieval stage and Line 5 corresponds to the training stage. Specifics are as folows.
Line 2 computes the global features $z_V$ for each video by averaging the embeddings of all of its video-text clips. An ablation (in appendix) shows that this is better than using the starting clip of a video to infer the representative video embedding.
Line 3 constructs the dense index\footnote{We use FAISS: \url{https://github.com/facebookresearch/faiss}.} for all videos to be used in our retrieval-based training.
Line 4 first finds $|\mathcal{C}|$ (corresponds to the number of overall batches in the training set) random videos, where each video $V$ yields a video cluster $c$ as follows.
We sample $|c|$ videos from $k$ neighboring videos of $V$. Instead of searching $k$ nearest videos directly (see ablation in Table~\ref{tbl:abl_retri}), we sample $k$ videos from the $2k$ nearest videos.
This is because we want videos in a cluster to be mutually closer to each other (not all close to video $V$).
In this way, all video/text clips sampled from one video can serve as negative examples for clips sampled from another video.
\section{Zero-shot Transfer to End Tasks}
\label{sec:zs}
We present methods for zero-shot transfer of \mbox{VideoCLIP} to a variety of end tasks (\textit{without} using any labels).
For each task, we specify requirements that highlight the aspect of pre-training.
\noindent\textbf{Text$\rightarrow$Video Retrieval.}
Text$\rightarrow$video retrieval tests the text-to-video similarity computed on the learned video-text representation.
$\text{NCE}(z_t, z_v)$ in Equation \ref{eq} contributes to this task as it discriminates different video clips in the numerator and denominator for a given text clip.
It also tests the distribution of hard negative examples in the denominator given it reports multiple recall metrics.
\noindent\textbf{Multiple-choice VideoQA.}
In multiple-choice VideoQA~\cite{yu2018joint}, the model aligns each video with one out of several text candidate answers.
It tests video$\rightarrow$text similarities with a pre-trained model. We formulate this task as ranking candidate textual answers for a given video question query.
This corresponds to the $\text{NCE}(z_v, z_t)$ term in Equation~\ref{eq}, where the subtle differences in texts are discriminated against each other.
\noindent\textbf{Action Segmentation.}
\label{sec:seg}
Action segmentation assigns each token (or frame) of a video with one of the pre-defined labels to separate meaningful segments of videos from the rest tokens (or frames).
This is similar to sequence labeling (\textit{e}.\textit{g}.~named entity recognition (NER)) in NLP.
Inspired by the setup of CLIP~\cite{radford2021learning}, the text encoder of VideoCLIP can serve as self-supervision for videos during pre-training and as a hyper network to provide hidden states of segment textual labels for a video token.
As such, the hidden state of each video token can have a distribution of similarity over segment labels.
This task tests video token to text similarities.
One challenge in action segmentation is that it contains an \underline{O}utside label that does not exist in transcription during pre-training.
This \underline{O}utside label is task-dependent because it means a token does not belong to any of the pre-defined labels.
This is similar to open set recognition \cite{scheirer2012toward} or out-of-domain intent detection \cite{lane2006out}, where the \textit{rejection} label is not presented during training but all new classes during inference (not shown in training) should be covered by the \textit{rejection} label.
Let $t \in L$ be one label in the set of all labels $L$ excluding the \underline{O}utside label.
We apply the following conditions to each video token $u$ to curate the prediction with the \underline{O}utside label $\hat{y}_u$:
\begin{equation}
\begin{split}
\begin{cases}
\argmax_{t\in L} (h_u z{_t}^T)&\text{if}\ \max_{t\in L}(h_u z{_t}^T)>\gamma, \\
\text{\underline{O}utside}&\text{otherwise,}
\end{cases}
\end{split}
\end{equation}
where $\gamma$ is a threshold.
Note that in zero-shot transfer, there is no access to training or validation data to decide a threshold as a hyper-parameter.
Thus, we estimate $\gamma$ as the maximum of dot products of intra-labels:
$\gamma = \max(z_{t} z_{t'}^T)$, where $t\in L, t'\in L$ and $t\neq t'$.
\noindent\textbf{Action Step Localization.}
In this task, each video is associated with a ``task'' with multiple steps $S$, where each step $t\in S$ is described as a short text.
Action step localization is to assign each video token to one or multiple steps in the associated task.
This is similar to action segmentation except that the label set is not pre-defined and does not contain the \underline{O}utside label.
As such, we first obtain the hidden states for each video frame (or token) $h_u$ from transformer. Then we separately forward text labels into the text backbone to obtain the hidden states of step labels $z_S$.
The distribution of each video token over steps is predicted as $\text{Softmax}(h_u z{_S}^T)$.
\section{Experiments}
\label{sec:exp}
\subsection{VideoCLIP Pre-training}
For pre-training, we use HowTo100M \cite{miech2019howto100m} that contains instructional videos via searching keywords from wikihow\footnote{\url{www.wikihow.com}} in YouTube.
We use 1.1M videos after filtering out videos which are not available or cannot be decoded.
We randomly sample 4K videos as the validation set and use the rest for pre-training.
On average, the duration of each video is \app6.5 minutes with \app110 clip-text pairs.
After removing repeated words from ASR, we end up with \app7.7 GB of text transcriptions, with 2.4 tokens per second on average.
\subsection{End Task Setups}
\paragraph{Text$\rightarrow$Video Retrieval.} We use Youcook2, MSR-VTT and DiDeMo to evaluate zero-shot transfer to text-video retrieval.
Youcook2 \cite{zhou2017towards} has 2K cooking videos with a total duration of 176 hours and 5.26 minutes on average per video.
It shows about 89 recipes in 14K video clips.
Each video clip is annotated with one sentence.
We follow the splits of~\citet{miech2019howto100m} to make sure there is no overlap between pre-training and evaluation data.
We have 3,305 test clip-text pairs from 430 videos for zero-shot evaluation.
MSR-VTT~\cite{xu2016msr} is a well-known dataset for text-video retrieval, question answering etc.
Following JSFusion~\cite{yu2018joint,miech2019howto100m}, we randomly sampled 1K clip-text pairs as test data for evaluation of zero-shot transfer.
DiDeMo \cite{anne2017localizing} has 10,000 videos annotated with 40,000 sentences on Flicker videos. We evaluate video-paragraph retrieval on 4021 available testing examples\footnote{\url{https://github.com/LisaAnne/LocalizingMoments/blob/master/utils/eval.py}}.
\paragraph{VideoQA.} We further use the QA test data \cite{yu2018joint} for MSR-VTT to evaluate multiple-choice VideoQA.
Recall that this task can be formulated as a video-text retrieval task except the candidate textual answers are associated with each video and only one answer is correct (most relevant).
On average, VideoQA for MSR-VTT has 5 candidate answers per video.
\paragraph{Action Segmentation.} We use COIN \cite{tang2019coin} to evaluate action segmentation.
It has 11,827 videos (476 hours) in total and the testing set has 2797 videos, where each video is labeled with 3.91 segments per video on average.
There are 778 segment labels and we feed these textual labels into the text backbone to obtain their latent space.
As a reminder of Section \ref{sec:seg}, we do not model the \underline{O}utside label explicitly and determine an \underline{O}utside label only when all other 778 labels reject a video token.
Note that videos in COIN can last for several minutes, we apply a sliding window with a step size of 16 seconds and a window size of 32 seconds.
During inference, we average the logits for overlapped tokens from multiple windows.
\paragraph{Action Step Localization.} We use CrossTask \cite{zhukov2019cross} to evaluate action localization.
It contains 83 different tasks and 4.7K videos.
Each task has a set of steps in the form of text descriptions and each frame of video is annotated with one or multiple steps as a distribution.
We use the testing data split via the official code\footnote{ \url{https://github.com/DmZhukov/CrossTask}}, which contains 1690 annotated videos.
We leave details of fine-tuning data to appendix.
\subsection{Implementation Details}
\paragraph{Video Encoder.}
We use a S3D~\cite{xie2018rethinking} for video encoder $f_{\theta_\text{CNN}}$. It is pre-trained on HowTo100M~\cite{miech2020end} to extract video tokens of dimension 512.
We use 30fps and extract one video token per second. This can be pre-computed for efficiency.
\paragraph{Transformers.}
For the video and text Transformers, $f_{\theta_\text{v}}$ and $f_{\theta_\text{t}}$, we initialize their weights with the pre-trained $\text{BERT}_{\text{BASE-uncased}}$~\cite{devlin-etal-2019-bert}. Using the same type of transformer further allows us to perform ablation study on sharing video and text backbones (see Table \ref{tbl:abl_retri}).
We only use the first 6 Transformer layers for the video input and all 12 layers for the text input.
Please note that the video/text encoders in VideoCLIP is generally applicable to other pre-trained Transformers.
We use a single layer MLP $f_{\theta_\text{MLP}}$ with GELU activation~\cite{hendrycks2016gaussian} to map the S3D outputs to the 768-dimensional inputs of the video Transformer.
We limit the maximum number of video tokens to be 32.
For video transformer, its input sequence is 34 with \texttt{[CLS]} and \texttt{[SEP]} tokens.
For text transformer, we have 61 text tokens plus \texttt{[CLS]} and \texttt{[SEP]} tokens (63 in total).
The number of text tokens roughly doubling in the number of video tokens because text comes at \app2.4 tokens per second (on average) in the HowTo100M data, while our video tokens are extracted at 1 token per second.
A text clip has a random length between 8 and 61 tokens, whereas a video clip has 3 to 32 seconds.
We sample 16 video/text pairs from each video and use $k{=}$32 videos to form batches of size $|B|{=}$512.
\paragraph{Training Details.}
We pre-train our model on 8 NVIDIA Tesla V100 GPUs (each with 32 GB memory) for 25 epochs using fp16 precision for \app1 day.
We use Adam \cite{kingma2014adam} as optimizer with betas of (0.9, 0.98), an initial learning rate of 5e-5, 1000 steps of warm-up, and a polynomial decay learning rate schedule.
Gradients are clipped at 2.0. The softmax temperature in objective \eqref{eq:nce} is set to $\tau=1.0$.
\begin{table}[!t]
\centering
\setlength\tabcolsep{0.1pt}
\scalebox{.83}{
\begin{tabular}{l c c c c}
\textit{Youcook2} dataset & R@1 $\uparrow$ & R@5 $\uparrow$ & R@10 $\uparrow$ \\%& Median R $\downarrow$ \\
\shline
\textsc{Supervised}\\
HGLMM\cite{klein2015associating} & 4.6 & 14.3 & 21.6\\% & 75 \\
Coot\cite{ging2020coot} & 16.7 & 40.2 & 52.3\\% & 9\\
UniVL (FT-Joint)\cite{luo2020univilm} & 22.2 & 52.2 & 66.2\\% & 5 \\
\textbf{VideoCLIP} (Fine-tuned) & \textbf{32.2} & \textbf{62.6} & \textbf{75.0} \\ \shline
\textsc{Zero-shot}\\
Random & 0.0 & 0.2 & 0.3 \\% & 1675 \\
HowTo100M\cite{miech2019howto100m} & 6.1 & 17.3 & 24.8 \\%& 46 \\
MIL-NCE\cite{miech2020end} & 15.1 & 38.0 & 51.2 \\%& 10 \\
\textbf{VideoCLIP} (Zero-shot) & \textbf{22.7} & \textbf{50.4} & \textbf{63.1}\\
\\
\textit{MSR-VTT} dataset & R@1 $\uparrow$ & R@5 $\uparrow$ & R@10 $\uparrow$ \\%& Median R $\downarrow$ \\
\shline
\textsc{Supervised}\\
UniVL (FT-Joint) \cite{luo2020univilm} & 20.6 & 49.1 & 62.9\\% & 6 \\
ClipBERT \cite{lei2021less} & 22.0 & 46.8 & 59.9\\
MMT \cite{gabeur2020multi} & 25.8 & 57.2 & 69.3\\
Support Set\cite{patrick2021supportset} & 30.1 & \textbf{58.5} & \textbf{69.3}\\% & 3.0\\
\textbf{VideoCLIP} (Fine-tuned) & \textbf{30.9} & 55.4 & 66.8 \\
\shline
\textsc{Zero-shot}\\
Random & 0.1 & 0.5 & 1.0 \\%& 500 \\
HowTo100M\cite{miech2019howto100m} & 7.5 & 21.2 & 29.6 \\%& 38\\
MIL-NCE\cite{miech2020end} & 9.9 & \textbf{24.0} & \textbf{32.4} \\%& 29.5 \\
SupportSet\cite{patrick2021supportset} & 8.7 & 23.0 & 31.1 \\%& 31.0\\
\textbf{VideoCLIP} (Zero-shot) & \textbf{10.4} & 22.2 & 30.0\\
\end{tabular}
}
\caption{\textit{Text$\rightarrow$video retrieval} on Youcook2 and VTT. }
\label{tbl:youcook}
\end{table}
\subsection{Main Results}
We evaluate VideoCLIP on various end tasks and compare it with other zero-shot and supervised methods that use labels on the target datasets.
\paragraph{Text-video Retrieval.} The results on Youcook2 and MSR-VTT are shown in Table~\ref{tbl:youcook}. The result on DiDeMo is shown in Table~\ref{tbl:didemo}.
On Youcook2 (Table~\ref{tbl:youcook}, top), VideoCLIP shows impressive performance gains and has much better accuracy than traditional supervised methods. The zero-shot transfer performance is even close to the performance level of supervised baselines with pre-training.
With fine-tuning, VideoCLIP reaches state-of-the-art on Youcook2.
On MSR-VTT (Table~\ref{tbl:youcook}, bottom), VideoCLIP shows solid improvements but with a larger zero-shot to supervised gap than on Youcook2. The major reason could be domain shift from HowTo100M to MSR-VTT. The captions in MSR-VTT are more descriptive (\textit{e}.\textit{g}., ``a basketball player is playing basketball'' and are less likely to appear in the transcriptions of HowTo100M).
After fine-tuning, VideoCLIP reaches state-of-the-art R@1. Note that this is achieved \textit{without using any supervised data} such as ImageNet or \textit{large-scale external data} (\textit{i.e.}, 65 million Instagram data) used by the second best method, Support Set~\cite{patrick2021supportset}.
\begin{table}[h]
\centering
\setlength\tabcolsep{0.1pt}
\scalebox{1.}{
\begin{tabular}{l c c}
\textit{DiDeMo} dataset & R@1 $\uparrow$ & R@5 \\% $\uparrow$ & R@10 $\uparrow$ \\%& Median R $\downarrow$ \\
\shline
\textsc{Supervised}\\
S2VT \cite{venugopalan2014translating} & 11.9 & 33.6 \\
FSE \cite{zhang2018cross} & 13.9 & 44.5\\
CE \cite{liu2019use} & 16.1 & 41.1\\
ClipBERT \cite{lei2021less} & 20.4 & 48.0\\
\shline
\textsc{Zero-shot}\\
\textbf{VideoCLIP} (Zero-shot) & \textbf{16.6} & \textbf{46.9}\\
\end{tabular}
}
\caption{\textit{Text$\rightarrow$video retrieval} on DiDeMo.}
\label{tbl:didemo}
\end{table}
On DiDeMo (Table~\ref{tbl:didemo}), VideoCLIP has better performance than most supervised methods. Note that ClipBERT\cite{lei2021less} has image pre-training before video+text fine-tuning.
\begin{table}[h]
\centering
\scalebox{0.83}{
\begin{tabular}{l c}
\textit{MSR-VTT} dataset & Accuracy $\uparrow$ \\
\shline
\textsc{Supervised}\\
LSTM-fusion \cite{yu2018joint} & 38.3 \\
C+LSTM+SA-FC7 \cite{torabi2016learning} & 60.2 \\
SNUVL \cite{yu2016video} & 65.4 \\
EITanque \cite{kaufman2017temporal} & 65.5 \\
CT-SAN \cite{yu2017end} & 66.4 \\
VSE-LSTM \cite{kiros2014unifying} & 67.3 \\
MLB \cite{kim2016hadamard} & 76.1 \\
JSFusion\cite{yu2018joint} & 83.4 \\
ActBERT\cite{zhu2020actbert} & 85.7\\
ClipBERT\cite{lei2021less} & 88.2\\
\textbf{VideoCLIP} (Fine-tuned) & \textbf{92.1}\\
\hline
\textsc{Zero-shot}\\
\textbf{VideoCLIP} (Zero-shot) & \textbf{73.9}\\
\end{tabular}
}
\caption{
\textit{VideoQA} on MSR-VTT.}
\label{tbl:videoqa}
\end{table}
\begin{table}[!htbp]
\setlength\tabcolsep{0.1pt}
\centering
\scalebox{0.8}{
\begin{tabular}{l c}
\textit{COIN} dataset & Frame Accuracy $\uparrow$ \\
\shline
\textsc{Supervised}\\
NN-Viterbi \cite{richard2018neuralnetwork} & 21.2\\
VGG \cite{simonyan2014very} & 25.8\\
TCFPN-ISBA \cite{ding2018weakly} & 34.3\\
CBT \cite{sun2019contrastive} & 53.9\\
ActBERT \cite{zhu2020actbert} & 57.0\\
MIL-NCE \cite{miech2020end} & 61.0\\
\textbf{VideoCLIP} (Fine-tuned) & \textbf{68.7} \\
\shline
\textsc{Zero-shot}\\
\textbf{VideoCLIP} (Zero-shot) & \textbf{58.9} \\
\end{tabular}
}
\caption{\textit{Action segmentation} on COIN.}
\label{tbl:coin}
\end{table}
\paragraph{Video Question Answering.}
In Table~\ref{tbl:videoqa}, zero-shot VideoCLIP outperforms most supervised methods but similarly suffers from domain shift from HowTo100M to MSR-VTT.
After fine-tuning, it reaches the best performance, indicating VideoCLIP also provides strong features for fine-tuning.
\paragraph{Action Segmentation.} We report the results of action segmentation on COIN in Table~\ref{tbl:coin}. Zero-shot transfer of VideoCLIP to COIN outperforms all supervised methods, \textit{without using any labels on this dataset}.
This indicates that VideoCLIP also learns good token-level video representations.
Fine-tuning VideoCLIP further yields a \app10\% accuracy gain, indicating potential room for improvement.
\paragraph{Action Step Localization.}
Lastly, we report VideoCLIP's performance on CrossTask in Table~\ref{tbl:crosstask}.
It shows a small gap to supervised methods when using zero-shot action step localization.
Fine-tuning leads to a \app10\% gain, outperforming all prior work on this dataset.
\begin{table}[!t]
\centering
\setlength\tabcolsep{2.0pt}
\scalebox{0.85}{
\begin{tabular}{l c}
\textit{CrossTask} dataset & Average Recall $\uparrow$\\
\shline
\textsc{Supervised}\\
Alayrac \cite{alayrac2016unsupervised} & 13.3\\
Zhukov \cite{zhukov2019cross} & 22.4\\
Supervised \cite{zhukov2019cross} & 31.6\\
ActBERT \cite{zhu2020actbert} & 41.4\\
UniVL \cite{luo2020univilm} & 42.0\\
\textbf{VideoCLIP} (Fine-tuned) & \textbf{47.3}\\
\shline
\textsc{Zero-shot}\\
HowTo100M \cite{miech2019howto100m} & 33.6\\
MIL-NCE \cite{miech2020end} & 40.5\\
\textbf{VideoCLIP} (Zero-shot) & \textbf{33.9} \\
\end{tabular}
}
\caption{\textit{Action step localization} on CrossTask.}
\label{tbl:crosstask}
\end{table}
\begin{table*}[!htbp]
\small
\centering
\scalebox{0.85}{
\begin{tabular}{l l l}
Query Text & Text of Top-1 video from \textbf{VideoCLIP} (Zero-shot) & Text of Top-1 video from \textbf{VideoCLIP} (Fine-tuned)\\
\shline
pick the ends off the verdalago &
\shortstack[l]{put chickpeas parsley chopped onion chili powder\\ ground cumin in food processor}
& pick the ends off the verdalago\\
\hline
add the fried pita to the salad and mix & toss the salad & add the dressing and bread pieces the the salad\\
\hline
\shortstack[l]{place chicken in hot oil\\ and fry until golden brown} & fry the chicken in oil & fry the chicken wings in deep oil\\
\hline
\shortstack[l]{fry dark meats together and\\ white meats together} & add the mutton to the pan & add the diced beef meat to it and roast it\\
\hline
rub salt and pepper onto the chicken & season them with salt and pepper & rub salt and pepper onto the chicken\\
\hline
\end{tabular}
}
\caption{Qualitative error analysis of \textit{text$\rightarrow$video retrieval} on Youcook2.
}
\label{tbl:err_youcook}
\end{table*}
\subsection{Discussion on Work that Fine-tunes CLIP Model}
There are concurrent works \cite{luo2021clip4clip,portillo2021straightforward} about using image+text model \cite{radford2021learning} for video+text downstream tasks.
Note that \cite{luo2021clip4clip} and \cite{portillo2021straightforward} use image pre-training (no video pre-training) and transfer to videos, whereas our focus is about improving video pre-training using a novel pre-training objective.
Besides this conceptual difference
\cite{luo2021clip4clip,portillo2021straightforward} are using a pre-trained image CLIP\cite{radford2021learning} model from OpenAI which is trained on huge, semi-curated web image+text pairs that provides exceptional zero-shot performance on many datasets (\textit{e}.\textit{g}. ImageNet); however, the CLIP pre-training data is sourced from web-search engines (which on their own use fully supervised neural networks trained on ImageNet and other datasets); therefore, is not fair to compare to our approach which only trains on HowTo100M instructional videos.
\subsection{Ablation Study}
In Table~\ref{tbl:abl_retri},
we perform an ablation study on zero-shot transfer for text$\rightarrow$video retrieval on Youcook2 to quantify the
the contribution of overlapping clips and retrieval augmented pre-training.
In the first group, we study the effectiveness of the two proposed methods.
VideoCLIP without retrieval augmented training significantly drops performance by over \textbf{4}\% in R@1 and additionally using \textit{exact alignment} positives,~\textit{i.e.}, the same start/end timestamp for a pair of video and text clips, has another \textbf{4}\% drop in R@1. Therefore, both techniques combined lead to a \app50\% relative improvement in recall.
Further, by \textit{using MIL-NCE clips and loss} we evaluate the potential benefit of using the training objective from MIL-NCE~\cite{miech2020end} (which uses multiple temporally adjacent clips as positives) in our architecture. This ablation isolates the pre-training objective from model and data.
We observe that the MIL-NCE loss can improve the direct alignment objective but performs significantly worse than our objective (16.1 \textit{vs.}~22.7 R@1).
\begin{table}[t!]
\centering
\setlength\tabcolsep{2.0pt}
\scalebox{0.84}{
\begin{tabular}{l c c c}
\textit{Youcook2} dataset & R@1 $\uparrow$ & R@5 $\uparrow$ & R@10 $\uparrow$ \\% & Median R $\downarrow$ \\
\shline
\textbf{VideoCLIP} (Zero-shot) & \textbf{22.7} & \textbf{50.4} & \textbf{63.1}\\
$-$ w/o retrieval & 18.5 & 42.8 & 54.6 \\
$-$ w/o retrieval and w/o overlap & 12.4 & 30.2 & 40.7 \\
$-$ using MIL-NCE clips and loss & 16.1 & 38.6 & 51.1 \\
\shline
$-$ shared video/text transformer & 21.9 & 48.1 & 60.6 \\
$-$ retrieve $k$ & 22.5 & 49.3 & 61.4\\
$-$ use first 32 sec for retrieval & 20.1 & 46.3 & 58.7\\
$-$ use \texttt{[CLS]} & 22.1 & 47.1 & 59.6\\
\end{tabular}
}
\caption{Ablation on \textit{text$\rightarrow$video retrieval} (Youcook2).
}
\label{tbl:abl_retri}
\end{table}
In the second group, we further study the design choices of modeling.
\textit{shared video/text transformer} indicates $ f_{\theta_\text{v}}$ is the same as $ f_{\theta_\text{t}}$, which only decreases performance slightly. This suggests that using a joint backbone for video and text is effective.
\textit{retrieve $k$} indicates direct searching $k$ nearest neighbors instead of sampling $k$ videos from $2k$ nearest neighbors (used by VideoCLIP) in Line 4 of Algorithm 1.
Sampling from nearest neighbors yields video clusters of better quality.
\textit{use starting 32 sec for retrieval} indicates using the first 32 secs of a video as representation for video retrieval, which is an inferior representation of the whole video.
Unlike employing \texttt{Avgpool}, using \texttt{[CLS]} token only prevents VideoCLIP from exploiting token-level information and thus yields worse performance.
\subsection{Qualitative Analysis}
We examine errors for text-video retrieval of Youcook2 in both zero-shot transfer and fine-tuning setting in Table~\ref{tbl:err_youcook}.
We observe that in zero-shot transfer, VideoCLIP has no prior knowledge about a particular task/dataset on how long a text and video clip should be paired together for the text-retrieval task.
Fine-tuning allows to correct this type of error.
Further, we observe that VideoCLIP tends to mix objects of similar color/shape together.
We leave incorporating such type of knowledge into pre-training to future work.
\vspace{5pt}
\section{Conclusion}
We have presented VideoCLIP, an approach to pre-train a video-text model for zero-shot transfer to end tasks that require fine-grained association between video and language.
VideoCLIP uses an objective that contrasts temporally overlapping positives with hard negatives stemming from nearest neighbor retrieval.
In evaluation this approach outperforms prior work on a variety of tasks, without any supervision on downstream datasets, and in some cases VideoCLIP is competitive or better than prior work that uses full supervision; nevertheless, we still observe gains for fine-tuning our model. We hope that our code and model will foster future research in multi-modal video understanding.
\vspace{5pt}
\section*{Code}
Code and models are made available at \url{https://github.com/pytorch/fairseq/tree/main/examples/MMPT}.
\vspace{5pt}
\section*{Acknowledgments}
We thank Licheng Yu for in-depth discussion and feedback, as well as Huaishao Luo and Mandela Patrick for supporting baseline implementation.
\bibliographystyle{acl_natbib}
|
\section{Introduction}
One of the main goals of the Beam Energy Scan (BES) Program at the Relativistic Heavy Ion Collider (RHIC) is to explore the phase diagram of the Quantum Chromodynamics (QCD)~\cite{STAR-BESI}. At vanishing and small values of the chemical potentials for conserved charges, such as net-baryon, net-charge and net-strangeness, Lattice QCD has predicted that it occurs a smooth crossover from the hadronic phase to the QGP phase~\cite{crossover}. Based on Lattice QCD calculations with physical values of light and strange quark masses, the negative signal of ratio of the sixth order to the second order ($C_6/C_2$) is observed in the crossover region~\cite{C6-karsch-v1,C6-karsch-v2,C6-karsch-v3,C6-karsch-v4}. The QCD-assisted low-energy effective theory and the QCD based models, such as the Polyakov loop extended quark-meson (PQM) and the Nambu-Jona-Lasinio (PNJL) models, also support that the sixth order cumulant is negative near the chiral crossover transition~\cite{PQM-C6-v0,PQM-C6-v1,PQM-C6-v2, PNJL-C6-v1, PNJL-C6-v2}. Consequently, $C_6/C_2$ is a powerful observable to study the QCD phase diagram in experiment.
The STAR Collaboration had reported the preliminary results of net-proton $C_6/C_2$~\cite{STAR-C6C2-v1,STAR-C6C2-v2, STAR-C6C2-v3}. In 0-40\% centrality, $C_6/C_2$ is negative at $\sqrt{s_{NN}}$ = 200 GeV, which is consistent with the Lattice QCD calculation. However, $C_6/C_2$ is positive and close to unity at $\sqrt{s_{NN}}$ = 54.4 GeV, unlike the QCD predicted negative behavior. In the future, with the sufficient statistics accumulated, the analysis of the net-proton $C_6/C_2$ is possible by the ALICE collaboration at LHC~\cite{LHC}.
Before illustrating the physics of measured $C_6/C_2$, the contributions of non-phase transition related influences should be subtracted, such as the conservation of the total baryon number, experimental acceptances in terms of kinematic variables, efficiency corrections, the difference of $C_6/C_2$ in net-proton and net-baryon and so on~\cite{C6C2-koch, C6C2-influence-1, C6C2-influence-2, C6C2-influence-3}. It was shown that $C_6/C_2$ can be negative for $\sqrt{s_{NN}} \le $ 40 GeV due to the baryon number conservation~\cite{C6-fu}. In this paper, we will focus on discussions of the BD contribution and influence of the statistics used for data analysis at $\sqrt{s_{NN}}$ = 200 GeV.
In nuclear collisions, some basic statistical distributions, such as the Poisson distribution, BD and Negative Binomial distribution (NBD), are frequently used to describe the shape of the multiplicity distributions~\cite{HRG-1,HRG-2,bd-1,bd-2}.
Theoretically, in the free hadron gas in equilibrium, the resonance gas in the hadron phase obeys the Poisson distributions. It was also argued that the proton number is given by the superposition of the binomial distribution of the baryon number due to the isospin randomization~\cite{Asakawa}. Consequently, studies of the cumulants in different statistical distributions can extract information related to the nature of the particle production mechanism.
Recently, cumulants in these statistical distributions were widely taken as baselines to help us understand the experimentally measured cumulants, such the the cumulants of the net-proton, net-charge, net-kaon and net-$\Lambda$~\cite{NBD-gary,Xiaofeng-baseline-proton, STAR-proton-v1, STAR-charge,PHENIX-charge,STAR-kaon, STAR-kappa}. To date, the studies of cumulants in BD/NBD have been up to the fourth order. It is found that the BD/NBD baselines can quantitatively explain the cumulants of net-charge, net-kaon and net-$\Lambda$ in experiment. The BD can also describe the fourth order cumulant of the net-proton number at $\sqrt{s_{NN}}\ge$ 39 GeV. Consequently, it is also interesting to study the net-proton $C_6/C_2$ in BD.
With the data accumulated, STAR has about 850M events for analysis of $C_6/C_2$ at $\sqrt{s_{NN}}$ = 200 GeV~\cite{STAR-C6C2-v2}. However, the data used in the analysis include three independent parts: (a) around 420M events for 0-80\% centrality collected from the year 2010 minimum bias (MB) trigger, (b) around 110M events for 0-10\% centrality from the year 2010 central trigger, and (c) around 320M events for 0-80\% centrality from the year 2011 MB trigger. In principle, $C_6/C_2$ in each centrality should be firstly calculated by these three independent parts separately. For each part, $C_6/C_2$ is obtained by the method of centrality bin width correction (CBWC)~\cite{cbwc-star,cbwc-v2,cbwc-v3}. It means $C_6/C_2$ needs to be calculated first in each Refmult3 bin ($N_{ch}$) to reduce the initial size fluctuation. Then, $C_6/C_2$ is averaged over all $N_{ch}$ in a given centrality.
A critical issue is that the statistics in each of $N_{ch}$ are still considerably limited even with 320M MB events. Our previous studies showed that we must check whether the statistics in each of $N_{ch}$ are sufficient to satisfy the Central Limit Theorem (CLT) when applying the CBWC method in cumulants analysis~\cite{statistics-chenlz-v1,statistics-chenlz-v2,statistics-chenlz-v3}.
It showed that 90M MB events are not sufficient for analysis of net-proton $C_6/C_2$ at $\sqrt{s_{NN}} = 11.5$ GeV in the UrQMD model. These statistics are of the same order of magnitude as that in each data collection at $\sqrt{s_{NN}} = $ 200 GeV at RHIC/STAR.
In addition, the required statistics are also related to the detail shape and width of the net-proton multiplicity distributions. Therefore, we need to re-examine if we can obtain reliable $C_6/C_2$ with 320M MB events.
\begin {figure*}[htbp]
\begin{center}
\includegraphics[width=2.73in]{epsilon_ppbar_final.pdf}
\includegraphics[width=2.73in]{c6c2-bd-expectations.pdf}
\figcaption{\label{C6-bd-expectations} (Color Online) The left panel shows $\varepsilon_p$ and $\varepsilon_{\bar{p}}$ in different centralities at $\sqrt{s_{NN}} =$ 200 GeV in Au + Au collisions measured by RHIC/STAR~\cite{cbwc-v3}. The protons and anti-protons are identified at mid-rapidity $(|y|<0.5)$ within $0.4<p_{T}< 2.0$ GeV/$c$. The right panel is the $\varepsilon$ dependence of $C_6/C_2$.}
\end{center}
\end{figure*}
In this paper, we will start off the discussions from behavior of net-proton $C_6/C_2$ in BD.
At $\sqrt{s_{NN}}$ = 200 GeV, it is found that the negative $C_6/C_2$ can be obtained with the experimentally measured parameters.
To study influence of the statistics on $C_6/C_2$, a toy simulation about validation of the CLT in $C_6/C_2$ is demonstrated in Section 3. To make a direct comparison with experimental data, more detail procedures of simulations are described
in Section 4.
With the experimentally collected statistics and extracted BD parameters, we show that there is a marked drop in $C_6/C_2$ in 0-10\% centrality in Section 5. The baseline of $C_6/C_2$ is also demonstrated. Finally, the summary is presented in Section 6.
\section{$C_6/C_2$ in BD}
In probability theory and statistics, the BD is the discrete probability distribution of the number of successes in a sequence of $n$ independent binary experiments, where the results of each experiment is true with probability $p$ and false with probability $q=1-p$~\cite{bimomial-distribution}. The probability that the binomial random variable $x$ takes on values in its range can be expressed using the binomial probability function:
\begin{equation}\label{bd-formula}
P(x) = \left(\begin{array}{c}
n\\x \end{array} \right)p^{x}(1-p)^{n-x}=\frac{n!}{x!(n-x)!}p^{x}(1-p)^{n-x},
\end{equation}
where $x$ corresponds to the number of protons or anti-protons in each event.
Experimentally, if we know the mean $\mu$ and variance $\sigma^2$ ($\sigma^2<\mu$ for a BD, while $\sigma^2>\mu$ for a NBD) of the distribution, then the input parameters of $p$ and $n$ are:
\begin{equation}\label{bd-probability}
p=1-\frac{\sigma^2}{\mu}=1-\varepsilon,
\end{equation}
and
\begin{equation}\label{bd-probability}
n=\frac{\mu}{p}=\frac{\mu}{1-\varepsilon},
\end{equation}
where $\varepsilon=\frac{\sigma^2}{\mu}$.
With given $\mu$ and $\varepsilon$, the expectations of cumulants from the second to the sixth order can be written as:
\begin{eqnarray}
C_2 &=& n\left(p-p^2\right) = \varepsilon\mu\label{2nd-exp}\\
C_3 &=& n\left(p-3p^2+2p^3\right) = \varepsilon\mu \left(-1+2\varepsilon\right)\label{3rd-exp}\\
C_4 &=& n\left(p-7p^2+12p^3-6p^4\right)= \varepsilon\mu \left(1-6\varepsilon + 6\varepsilon^2\right)\label{4th-exp}\\
C_5 &=& n\left(p-15p^2+50p^3-60p^4+24p^5\right)\nonumber\\ &=& \varepsilon\mu \left(-1 + 14\varepsilon - 36 \varepsilon^2 + 24 \varepsilon^3\right)\label{5th-exp}\\
C_6 &=& n\left(p-31p^2+180p^3-390p^4+360p^5-120p^6\right)\nonumber\\ & =& \varepsilon\mu \left(1-30\varepsilon + 150\varepsilon^2 - 240\varepsilon^3 + 120 \varepsilon^4\right)\label{6th-exp}
\end{eqnarray}
If the numbers of protons and anti-protons are independently produced as BD, the net-proton $C_6/C_2$ can be expressed as:
\begin{equation}\label{C6C2-bd}
C_6/C_2 = \frac{C_6^{p}+C_6^{\bar{p}}}{C_{2}^{p}+C_{2}^{\bar{p}}}.
\end{equation}
Generally, the expected $C_6/C_2$ is related to $\varepsilon_p$, $\varepsilon_{\bar{p}}$, $\mu_p$, and $\mu_{\bar{p}}$. Based on these four parameters, one can obtain the expected $C_6/C_2$ in each centrality. In contrast, the experimental studies had shown that $\varepsilon_p$ and $\varepsilon_{\bar{p}}$ are close to each other at $\sqrt{s_{NN}} = 200$ GeV in Au + Au collisions shown in Fig.~\ref{C6-bd-expectations}(a) ~\cite{cbwc-v3}. The error contains the statistical and systematical uncertainties, which is performed by $\sigma_{\varepsilon} =\sqrt{\sigma^2_{stat} + \sigma^2_{sys}}$. The protons and anti-protons are selected at mid-rapidity $(|y|<0.5)$ within $0.4<p_{T}< 2.0$ GeV/c.
It shows $\varepsilon_p$ and $\varepsilon_{\bar{p}}$ extracted from STAR are consistent with each other. Within 1$\sigma_{\varepsilon}$ of uncertainty, the centrality dependence of $\varepsilon$ is weak.
To make an appropriate approximation, we assume $\varepsilon = \varepsilon_p = \varepsilon_{\bar{p}}$ in this paper. In this case,
the expectation of net-proton $C_6/C_2$ can be written as:
\begin{equation}\label{C6C2-bd}
C_6/C_2 = 1-30\varepsilon + 150\varepsilon^2 - 240\varepsilon^3 + 120 \varepsilon^4.
\end{equation}
Eq.~(\ref{C6C2-bd}) shows the expectation of $C_6/C_2$ is only dependent on $\varepsilon$. The effects of $\mu_p$ and $\mu_{\bar{p}}$ are cancelled. The detailed $\varepsilon$ dependence of $C_6/C_2$ is shown in Fig.~\ref{C6-bd-expectations}(b).
It shows $C_6/C_2$ drops drastically with the decrease of $\varepsilon$.
It is already negative with $\varepsilon<0.958$, which is within the experimentally measured range.
In addition, $C_6/C_2$ has a broad value range and it can change from positive to negative,
within 1$\sigma_{\varepsilon}$ uncertainty of $\varepsilon$.
Consequently, it is not suitable to directly give the expectation only based on the unique measured value of $\varepsilon$, without considering its uncertainty. Instead, we must set an interval of $\varepsilon$ to study the behavior of $C_6/C_2$. Fig.~\ref{C6-bd-expectations}(a) shows the upper and lower values of the error bar touch $\varepsilon$ at approximately 0.99 and 0.94, respectively. If we assume the range of $\varepsilon$ is between 0.94 and 0.99, then the expected values of $C_6/C_2$ are from -0.31 to 0.71. The negative $C_6/C_2$ can be obtained in the pure statistical BD. Only the negative signal is not enough to be taken as an indication of a smooth crossover transition.
Here the obtained $C_6/C_2$ is the ideal theoretical expectation. In the experiment, the statistics are still a critical issue for the analysis of $C_6/C_2$. The satisfaction of the CLT requires to be carefully checked before the data analysis.
\section{Central Limit Theorem in $C_6/C_2$}
\begin {figure*}
\begin{center}
\includegraphics[width=2.7in]{mean_clt.pdf}
\includegraphics[width=2.7in]{mean_c6c2.pdf}
\figcaption{\label{mean-clt} (Color Online) Statistics dependence of $\left<Mean\right>$ and $\left<C_6/C_2\right>$, respectively.}
\end{center}
\end{figure*}
\begin{figure*}[b]
\begin{center}
\includegraphics[width=2.1in]{urqmd-refmult3.pdf}
\includegraphics[width=2.1in]{mean-proton-urqmd.pdf}
\includegraphics[width=2.1in]{mean-antiproton-urqmd.pdf}
\figcaption{\label{UrQMD-refmult3} (Color Online) The left panel shows normalized probability distribution of Refmult3 in UrQMD model in Au + Au collisions at $\sqrt{s_{NN}} = $ 200 GeV. The middle and right panels show $\left<N_{p}\right>$ and $\left<N_{\bar{p}}\right>$ as a function of Refmult3, respectively. }
\end{center}
\end{figure*}
It is known that if the statistics are sufficient for validation of CLT~\cite{CLT}, then the experimentally measured results should always be consistent with the true value within 1$\sigma$, 2$\sigma$ and 3$\sigma$ of uncertainties with probability of 68.3$\%$, 95.5$\%$ and 99.7$\%$, respectively. For most of the measurements, such as the mean of the net-proton number, the required statistics for CLT are easily achieved in experiments. In this case, within the uncertainties, the mean of the measured observable needs to be a real value and independent of the statistics. i.e.,
\begin{equation}\label{mean}
\left<X\right>_{n_{1}} = \left<X\right>_{n_{2}} = \left<X\right>_{n_{3}},
\end{equation}
where $X$ is the measured observable. The subscripts $n_1$, $n_2$ and $n_3$ denote different statistics. As an example, Fig.~\ref{mean-clt}(a) shows the simulated $\left<Mean\right>$ of the net-proton number as a function of the statistics. The BD is applied in simulations with parameters: $\mu_{p} = 12.66, \mu_{\bar{p}} = 7.5, \varepsilon_{p} = \varepsilon_{\bar{p}} = 0.97$. $\mu_{p}$ and $\mu_{\bar{p}}$ are mean values of proton and anti-proton in $0-10\%$ centrality with $0.4<p_{T}<2.0$ GeV/c and $|y|<0.5$. For each data point, we randomly and independently generate 50 sub-samples with the fixed statistics to calculate $\left<Mean\right>$ as:
\begin{equation}\label{mean-simulation}
\left<Mean\right> = \frac{\sum_{i=1}^{i=N}(Mean)_i}{N},
\end{equation}
where $(Mean)_i$ is the averaged number of net-proton in $i_{th}$ sub-sample. There are two methods to estimate error($\left<Mean\right>$). One is obtained by the formula of the error propagation:
\begin{equation}\label{mean-error}
error(\left<Mean\right>) = \frac{\sqrt{\sum_{i=1}^{i=N}error(Mean)_i^{2}}}{N}.
\end{equation}
We can also first calculate the width of $\left<Mean\right>$ based on these $N$ results.
Then error($\left<Mean\right>$) is the width of $\left<Mean\right>$ divided by $\sqrt{N}$.
A Good agreement is obtained based on these two methods. In this paper, the formula of error propagation is used to measure the statistical uncertainty.
The simulated statistics in Fig.~\ref{mean-clt}(a) are from 50 to 5,000 events. The black dashed line is the theoretical expectation. For mean of the net-proton number analysis, Fig.~\ref{mean-clt}(a) clearly shows that 50 events are sufficient for the validation of the CLT. That is why we do not require to check whether the statistics are sufficient for most of the observables.
However, it is a challenge for analysis of the high-order cumulants which are up to the fifth order, sixth order, or even the eighth order~\cite{statistics-chenlz-v1, statistics-chenlz-v2, statistics-chenlz-v3}. When the CBWC method is applied in cumulants calculations, the statistics in each $N_{ch}$ are significantly limited in $0-10\%$ centrality even with a few hundred million MB events. As a crude estimation, supposing there are 1000 $N_{ch}$ bins in 0-10$\%$ centrality, the averaged events are only approximately 10,000 in each $N_{ch}$ with 100M MB events.
If 10, 000 events are not sufficient for CLT in $C_6/C_2$ calculations, the value obtained by the CBWC method is not reliable in 0-10\% centrality with 100M MB events.
By using the same simulated parameters as Fig.~\ref{mean-clt}(a), while the simulated sub-sample $N$ is significantly larger than 50, the statistics dependence of $\left<C_6/C_2\right>$ from 10,000 to 3M events are demonstrated in Fig.~\ref{mean-clt}(b). Below 0.1M events in each $N_{ch}$, $\left<C_6/C_2\right>$ is systematically smaller than 0 and theoretical BD expectations. Up to 0.5M events in each $N_{ch}$, $\left<C_6/C_2\right>$ is consistent with theoretical expectations within statistical uncertainties.
As we mentioned,
the required statistics are also related to the detail shape
and width of the net-proton multiplicity distributions~\cite{statistics-chenlz-v1, statistics-chenlz-v2, statistics-chenlz-v3}. The behavior of the statistical dependency in different centralities requires a careful case-by-case study.
To directly compare, we study the statistics dependence of $C_6/C_2$ in BD with the same calculation method as RHIC/STAR.
\section{Simulation Setup}
\begin {figure*}
\begin{center}
\includegraphics[width=2.85in]{c6c2-simulation-method-bd-0to10-final.pdf}
\includegraphics[width=2.85in]{c6c2-simulation-method-bd-30to40.pdf}
\figcaption{\label{C6C2-bd-example} (Color Online) $C_6/C_2$ and their statistical errors of 200 samples in 0-10\% and 30-40\% centrality, respectively. Each of $C_6/C_2$ is calculated by the CBWC method with 300M MB events. The protons and anti-protons are independently and randomly generated from the BD with $\varepsilon = 0.97$. The input parameters for simulations are taken from UrQMD model at $\sqrt{s_{NN}} = $ 200 GeV in Au + Au collisions.}
\end{center}
\end{figure*}
To study the statistics dependence of $C_6/C_2$ measurement based on the CBWC method, the input parameters for simulations are taken from the UrQMD model at $\sqrt{s_{NN}} = $ 200 GeV in Au + Au collisions~\cite{UrQMD-1, UrQMD-2}.
Fig.~\ref{UrQMD-refmult3}(a) shows the normalized probability distribution of Refmult3, which is used for the centrality selection. To avoid the auto-correlation and improve the centrality resolution, Refmult3 is defined by using the number of charged $\pi$ and $K$ in the final state within pseudo-rapidity $|\eta|<$1.0~\cite{cbwc-v3}.
Figs.~\ref{UrQMD-refmult3}(b) and (c) show the means of proton and anti-proton numbers, $\left<N_{p}\right>$ and $\left<N_{\bar{p}}\right>$, as a function of Refmult3, respectively. As a comparable study to STAR measurements, the protons and anti-protons in the UrQMD model are carried out at mid-rapidity $(|y|<0.5)$ in the transverse momentum range $0.4<p_T<2.0$ GeV/$c$. These values of $\left<N_{p}\right>$ and $\left<N_{\bar{p}}\right>$ are taken as the input parameters in the following simulations.
For a straightforward comparison between the simulated results and theoretical expectations, we assume that the numbers of protons and anti-protons independently follow BD with $\varepsilon = \varepsilon_{p} = \varepsilon_{\bar{p}} = 0.97 $ in each Refmult3 bin. Finally, the influence of statistics on $C_6/C_2$ with the whole interval of $\varepsilon$, from 0.94 to 0.99, is studied. The required number of MB events is 300 million (M).
Under these assumptions, the simulated procedures are implemented as follows:
(a) Multiplying the normalized probability of Refmult3 by 300M. The corresponding number of events in $j_{th}$ Refmult3 bin is $n_j$.
(b) According to $n_{j}$, the values of $\left<N_{p}\right>$ and $\left<N_{\bar{p}}\right>$ in $j_{th}$ Refmult3 bin, the numbers of protons and anti-protons are independently and randomly generated event-by-event by BD in each Refmult3 bin.
(c) Calculating $\left(C_6/C_2\right)_{j}$ and its error in each Refmult3 bin. Its error is estimated by the delta theorem method~\cite{delta-error}. Then the centrality dependence of $C_6/C_2$ and error($C_6/C_2$) in each centrality are obtained based on the CBWC method.
(d) Repeating the procedures of (b) and (c) $N$ times. Here, $N$ should be large enough to reduce the uncertainty of averaged $C_6/C_2$, $\left<C_6/C_2\right>$. We set $N = 1000$ in the following.
(e) Calculating the centrality dependence of $\left<C_6/C_2\right>$ and error($\left<C_6/C_2\right>$). The formula of error propagation is used to estimate the statistical uncertainty.
To demonstrate more details about the simulations, Figs.~\ref{C6C2-bd-example}(a) and (b) show 200 results of 1000 samples in 0-10\% and 30-40\% centrality, respectively.
To reduce the statistical error and get a more stable result, the value of $\left<C_6/C_2\right>$, shown by the black solid line in each panel, is derived from the total 1000 samples. The black dashed lines are 1$\sigma$ limits of $C_6/C_2$.
Both of these two plots clearly demonstrate that $C_6/C_2$ randomly fluctuates around $\left<C_6/C_2\right>$. In Fig.~\ref{C6C2-bd-example}(a), the probabilities of $C_6/C_2$ that lie outside 1$\sigma$ and 2$\sigma$ of $\left<C_6/C_2\right>$ are about 31.0\% and 5.0\%, respectively. Meanwhile, 30.0\% and 4.0\% of the observations lie outside of the 1$\sigma$ and 2$\sigma$ standard deviations shown in Fig.~\ref{C6C2-bd-example}(b). Theoretically, for $N$ independent observations $(x_{1}, x_{2}, \cdots, x_{N})$ with the same expectation, they are approximated with a normal distribution. In this case, 31.7\% and 4.5\% of the observations are outside 1$\sigma$ and 2$\sigma$ standard deviations of the expectation, respectively. Consequently, our results are consistent with the theoretical studies. It confirms the validity of our simulations.
\begin{figure*}[htp]
\begin{center}
\includegraphics[width=4.8in]{c6c2-bd97-all-centrality.pdf}
\caption{\label{C6-bd-centrality} Statistical dependence of $\left<C_6/C_2\right>$ in BD with $\varepsilon = 0.97$. The statistical error is calculated from formula of the error propagation. The black solid lines are the corresponding theoretical BD expectations derived from Eq.~(\ref{C6C2-bd}).}
\end{center}
\end{figure*}
The red solid line in each panel is the theoretical BD expectation. It is about 0.233 with $\varepsilon = 0.97$. In principle, if the statistics in each Refmult3 bin are sufficient, the final $\left<C_6/C_2\right>$ should be consistent with 0.233 in all centralities. However, Figs.~\ref{C6C2-bd-example}(a) and (b) show the disagreement of $C_6/C_2$ between the simulated $\left<C_6/C_2\right>$ and theoretical BD expectation. The difference of $C_6/C_2$ in 0-10\% centrality is more significant than that in 30-40\% centrality. Comprehensive studies of the statistics dependence of $C_6/C_2$ are shown in the following section.
\section{Statistics dependence of $\left<C_6/C_2\right>$ in BD}
As we mentioned, the experimental data collected for $C_6/C_2$ analysis include three different runs: 300M MB events, 420M MB events, and 110M from central trigger. Here 100M events from central trigger can be reference to 1000M MB events in our simulations.
As a direct comparison, we can study the statistical dependence of $\left<C_6/C_2\right>$ with the magnitude of a few hundred million.
By considering $\varepsilon=0.97$ to the simulations, Figs.~\ref{C6-bd-centrality}(a) to (f) show statistical dependence of $\left<C_6/C_2\right>$ in different centralities from 100M to 1000M MB events. In 0-10\% centrality, Fig.~\ref{C6-bd-centrality}(a) shows that all of the results are systematically smaller than the black solid line, which is the theoretical expectation.
Below 500M MB events, the values of $\left<C_6/C_2\right>$ first increase as the statistics increase and they are all negative. From 600M to 1000M MB events, $\left<C_6/C_2\right>$ is consistent with zero with 2$\sigma$ of the statistical uncertainty. Consequently, with the current statistics at RHIC/STAR, $C_6/C_2$ may be under-estimated in 0-10\% centrality at $\sqrt{s_{NN}} = $ 200 GeV.
In 10-20\% centrality, Fig.~\ref{C6-bd-centrality}(b) shows that the values of $\left<C_6/C_2\right>$ are still systematically smaller than that in BD. However, it becomes positive above 200M MB events. It suggests that the effect of the statistics becomes weaker.
In 20-30\% and 30-40\% centralities, $\left<C_6/C_2\right>$ is only slightly smaller than that in BD, as shown in Figs.~\ref{C6-bd-centrality}(c) and (d). Finally, they are in agreement with each other in peripheral collisions shown in Figs.~\ref{C6-bd-centrality}(e) and (f).
Fig.~\ref{C6-bd-centrality} shows that the influence of statistics on $\left<C_6/C_2\right>$ is strongest in central collisions, comparing to the mid-central and peripheral collisions.
In 0-10\% centrality, $\left<C_6/C_2\right>$ is still smaller than the expectation even with 1000M MB events.
$\left<C_6/C_2\right>$ is still about 0 and we do not observe the convergence even with 1,000M MB events.
It is due to the wider distribution of net-proton number and the smaller statistics in each $N_{ch}$ in $0-10\%$ centrality. It can be further understood from Fig.~\ref{mean-clt}(b), in which the simulated parameters are also extracted from mean values of proton and anti-proton in $0-10\%$ centrality in the UrQMD model. As the same estimation as in section 2, with 1,000M MB events, the averaged statistics in each $N_{ch}$ is around 0.1M, assuming there are 1,000 $N_{ch}$ bins in $0-10\%$ centrality. Fig.~\ref{mean-clt}(b) also shows that $\left<C_6/C_2\right>$ is also about 0 with 0.1M events in each $N_{ch}$, which is consistent with that shown in Fig.~\ref{C6-bd-centrality}(a). In contrast, Fig.~\ref{mean-clt}(b) also shows that $\left<C_6/C_2\right>$ is consistent with the theoretical expectations above 0.5M events, which suggests that 5,000M MB events are required if the CBWC method in each $N_{ch}$ is applied. That is why the convergence is not observed in Fig.~\ref{C6-bd-centrality}(a).
One possible alternative is to reduce the number of bins when applying CBWC method, such as the centrality bin width of each $\delta 0.5\%$, $\delta 1.0\%$, and $\delta2.5\%$ and so on. It is beyond the scope of this paper to examine through the transport model or the experimental data directly.
\begin{center}
\includegraphics[width=3.0in]{c6c2-band-simulation_final.pdf}
\figcaption{\label{C6-bd-band} (Color Online) Centrality dependence of $\left<C_6/C_2\right>$ in BD with $\varepsilon$ from 0.94 to 0.99 and in SD, respectively. Each $\left<C_6/C_2\right>$ is averaged from 10,000 randomly and independently simulated samples, with 300M MB events in each sample.}
\end{center}
Because the influence of the statistics cannot be neglected, it is necessary to study $C_6/C_2$ in BD with the current statistics at RHIC/STAR.
With 300M MB events, the blue filled band of Fig.~\ref{C6-bd-band} shows centrality dependence of $\left<C_6/C_2\right>$ with $\varepsilon$ from 0.94 to 0.99. The solid and dashed lines are theoretically expected $C_6/C_2$ with $\varepsilon=0.99$ and 0.94, respectively.
In 0-10\% and 10-20\% centralities, the simulated results are systematically smaller than those with the ideal theoretical expectations. Particularly, there is a marked drop in $\left<C_6/C_2\right>$ in 0-10\% centrality. With $\varepsilon=0.99$, the simulated $\left<C_6/C_2\right>$ is still negative, while the theoretical expectation is about 0.71. Meanwhile, this phenomenon is also observed with the same simulation method in Skellam distribution (SD), shown by circular solid points in Fig.~\ref{C6-bd-band}. The simulated $\left<C_6/C_2\right>$ is consistent with zero while it must be unity in SD. It further confirm that $C_6/C_2$ can be under-estimated in 0-10\% centrality. With insufficient statistics, $C_6/C_2$ can vary from a positive value to a negative one.
Together with the same statistics as RHIC/STAR, the negative $C_6/C_2$ is more significant in BD. The blue filled band shown in Fig.~\ref{C6-bd-band} almost covers the current preliminary measured $C_6/C_2$ within experimental large uncertainties~\cite{STAR-C6C2-v1}. It can better describe the experimental measured $C_6/C_2$, comparing to the SD baseline. Nonetheless, it cannot completely reproduce the experimentally measured results.
First, the experimental data show $C_6/C_2$ decreases from peripheral to central collisions~\cite{STAR-C6C2-v1}. In our simulations, the variation range of $\left<C_6/C_2\right>$ is too large, although $\varepsilon$ is just from 0.94 to 0.99. It is difficult to study the feature of centrality dependence of $C_6/C_2$. Second, besides the statistics and BD, $C_6/C_2$ can still be affected by many other complex contributions. We must correct those contributions well before relating the experimental measurements to our calculations. Third, we assume the proton and anti-proton are produced independently. However, they must have correlation, such as from resonance decay, which is not considered in this paper.
A direct comparison between the experimentally measured results and this baseline requires further studies. In contrast, our studies also imply that we must exclude factors of non-phase transition related influences well before connecting the measured $C_6/C_2$ to the theoretical Lattice QCD calculations.
\section{Summary}
We studied the behavior of the net-proton $C_6/C_2$ in BD in Au + Au collisions at $\sqrt{s_{NN}} = $ 200 GeV. By a reasonable approximation of $\varepsilon = \varepsilon_p = \varepsilon_{\bar{p}}$, the net-proton $C_6/C_2$ is only dependent on $\varepsilon$ in BD. With protons and anti-protons carried out in the mid-rapidity $|y|<0.5$ and $0.4<p_{T}<2.0$ GeV/$c$ in experiment, $\varepsilon_{p}$ and $\varepsilon_{\bar{p}}$ are close to each other and the values are between 0.94 and 0.99 in all centralities. In this region, $\varepsilon$ has a very significant effect on $C_6/C_2$. The values of $C_6/C_2$ decrease from 0.71 to -0.31 with $\varepsilon$ from 0.99 down to 0.94. The negative
$C_6/C_2$ is observed in the pure statistical BD. It suggests that the negative $C_6/C_2$ is not sufficient to be considered as an indication of a smooth crossover transition.
We also simulated the statistics dependence of $\left<C_6/C_2\right>$ based on CBWC calculation method.
Our method can compare the simulated results and theoretical expectations in a straightforward manner. The significantly dropped signal is observed in 0-10\% centrality. With 300M MB events, the simulated $C_6/C_2$ is negative, while it is about 0.71 with $\varepsilon=0.99$ in BD. This phenomenon is also observed with the same method of simulation in SD. The simulated $\left<C_6/C_2\right>$ is consistent with zero while the theoretical expectation is unity. Moreover, with 1000M MB events, the simulated $\left<C_6/C_2\right>$ is consistent with zero, while the theoretical expectation is about 0.233 with $\varepsilon = 0.97$ in BD.
Consequently, with the current statistics at RHIC/STAR, $C_6/C_2$ may be under-estimated in 0-10\% centrality at $\sqrt{s_{NN}} = $ 200 GeV.
With the experimental collected statistics, the negative $C_6/C_2$ is more significant in BD. It shows the values of $C_6/C_2$ have a broad range which can change from positive to negative. Comparing the baselines of $C_6/C_2$ in SD and BD, the BD could better describe the experimental measured $C_6/C_2$. Within large uncertainties at RHIC/STAR, the obtained baseline almost covers the current preliminary measured $C_6/C_2$. Consequently, the exclusions of the non-phase transition related influences are required when using $C_6/C_2$ to study the chiral phase transition.
\end{multicols}
\vspace{-1mm}
\centerline{\rule{80mm}{0.1pt}}
\vspace{2mm}
\begin{multicols}{2}
|
\section{Introduction} \label{intro}
This article deals with the problem of uniqueness of the local-in-time weak solutions to a model of superfluidity governed by the nonlinear Schr\"odinger equation (NLS) coupled with the inhomogeneous Navier-Stokes equations (NSE) for incompressible fluids. The NLS is used to describe the dynamics of the superfluid phase, while the NSE is employed to describe the evolution of the normal Helium liquid. \medskip
We present a result that is analogous to weak-strong uniqueness, but for less regular strong solutions (since our weak solutions are more regular than usual). Motivated by the analytical results in \cite{Jayanti2022LocalSuperfluidity}, we investigate the issue of uniqueness in Section \ref{local uniqueness of weak solutions} presenting a new class of weak solutions with additional regularity properties. Weak solutions usually lack the regularity that is needed to make them unique, a notable exception being the incompressible NSE in 2D \cite{Lions1959Un2}. Two different solutions to this impediment have been established in the literature. One is the conditional regularity theorems, where the weak solutions are shown to be smooth as soon as they belong to a critical regularity class. There are many results in this direction, mostly devoted to the incompressible NSE $-$ see Caffarelli et al. \cite{Caffarelli1982PartialEquations}, Escauriaza et al. \cite{EscauriazaL2003L3-solutionsUniqueness}, Prodi \cite{Prodi1959UnNavier-Stokes}, Serrin \cite{Serrin1963TheEquations}, or more recently, Neustupa et al. \cite{Neustupa2002AnVelocity,Neustupa2000AnEquations}. In the case of the compressible NSE, an analogous result was established in \cite{Feireisl2011SuitableFluids}. \medskip
The other approach, used in this work, is to show weak-strong uniqueness, i.e., that a strong solution is unique in the class of weak solutions (see \cite{Wiedemann2017Weak-strongDynamics} for a survey, including the case of the incompressible NSE). In the context of compressible fluids, related results can be found in Feireisl, Novotn\'{y} and collaborators \cite{Feireisl2011SuitableFluids,Feireisl2012RelativeSystem,Feireisl2012WeakstrongSystem}, and in Mellet and Vasseur \cite{Mellet2008AsymptoticEquations}. Intermediate on the spectrum of complexity associated with compressibility are the incompressible inhomogeneous fluids, which is the model considered in this paper. There has been work done for weak-strong uniqueness in this regard too \cite{Germain2008StrongSystem}. Similar methods have been used to obtain conditional uniqueness for related systems, like fluid-solid interactions \cite{Chemetov2019Weak-strongCondition,Kreml2020Weak-strongInteraction}, fluid-particle interactions \cite{Ballew2013WeaklySystem}, compressible magnetohydrodynamics \cite{Yan2013OnFlows}, and Euler-Korteweg-Poisson systems \cite{Donatelli2015Well/IllProblems}. \medskip
Our analysis is motivated by the pioneering work of Dafermos \cite{Dafermos1979TheStability} and DiPerna \cite{DiPerna1979UniquenessLaws}, the results of Germain \cite{Germain2011WeakstrongSystem}, the analysis of Mellet and Vasseur \cite{Mellet2008AsymptoticEquations}, as well as the approach of Feireisl et al. \cite{Feireisl2011SuitableFluids}. Usually, weak-strong uniqueness theorems involve strong solutions that are very regular: bounded (in $x$) first-order space-time derivatives. This is required so that the strong solution bears the brunt of the smoothness requirements, allowing us to work with rather irregular weak solutions (for instance, those satisfying a simple $L^2_x$ based energy estimate). A Gronwall’s argument is employed on the difference between the weak and strong solutions to conclude uniqueness (see Theorem \ref{weak-strong uniqueness} below), yielding that a weak solution agrees with a classical solution starting from the same initial data when such a classical solution exists. We also refer the reader to Section 2.5 in \cite{Lions1996MathematicalMechanics} where a related result is presented for incompressible inhomogeneous fluids. It must be noted that the scheme, originally developed for weak solutions of the Navier-Stokes equations (with regularity only at the level of the energy bound, i.e., $u\in L^2_x$), does not work here because of added interactions with the Schr\"odinger equation. We will necessarily need to consider solutions with smoother velocity fields. We will discuss this briefly in Section \ref{weak-strong uniqueness proof} below, after sketching the steps used in the treatment of the momentum equation. \medskip
The weak solutions constructed in \cite{Jayanti2022LocalSuperfluidity} to a model of superfluidity required more a priori estimates than just the energy, and this led to better-behaved solutions (albeit not regular enough to prove uniqueness). We can, however, trade the increased regularity of our weak solution for a less smooth strong solution. A \textit{weak-moderate} uniqueness theorem is presented in Section \ref{local uniqueness of weak solutions}. Apart from the use of ``moderate" solutions (which belong to a lower regularity class), another feature is that the full regularity of the existence result is not utilized. This means that this uniqueness result will continue to hold even if weaker solutions (than in \cite{Jayanti2022LocalSuperfluidity}) are shown to exist for the model in question. \medskip
\subsection{Notation} \label{notation}
Let $\mathfrak{D}(\Omega)$ be the space of smooth, compactly-supported functions on $\Omega$. Then, $H^s_0(\Omega)$ is the completion of $\mathfrak{D}$ under the Sobolev norm $H^s$. The more general Sobolev spaces are denoted by $W^{s,p}(\Omega)$, where $s\in\mathbb{R}$ is the derivative index and $1\le p\le\infty$ is the integrability index. A dot on top, like $\Dot{H}^s(\Omega)$ or $\Dot{W}^{s,p}$, is used when referring to the homogeneous Sobolev spaces. \medskip
Consider a 3D vector-valued function $u\equiv (u_1,u_2,u_3)$, where $u_i\in \mathfrak{D}(\Omega), i=1,2,3$. The collection of all divergence-free functions $u$ defines $\mathfrak{D}_d(\Omega)$. Then, $H^s_d(\Omega)$ is the completion of $\mathfrak{D}_d(\Omega)$ under the $H^s$ norm. In addition, to say that a complex-valued wavefunction $\psi\in H^s(\Omega)$ means that its real and imaginary parts are the limits (in the $H^s$ norm) of functions in $\mathfrak{D}(\Omega)$. \medskip
The $L^2$ inner product, denoted by $\langle \cdot,\cdot \rangle$, is sesquilinear (the first argument is complex conjugated, indicated by an overbar) to accommodate the complex nature of the Schr\"odinger equation. Thus, for example, $\langle \psi,B\psi \rangle = \int_{\Omega} \Bar{\psi}B\psi \ dx$. Needless to say, since the velocity and density are real-valued functions, we will ignore the complex conjugation when they constitute the first argument of the inner product. \medskip
We use the subscript $x$ on a Banach space to denote the Banach space is defined over $\Omega$. For instance, $L^p_x$ stands for the Lebesgue space $L^p(\Omega)$, and similarly for the Sobolev spaces: $H^s_{d,x} := H^s_d(\Omega)$. For spaces/norms over time, the subscript $t$ will denote the time interval in consideration, such as $L^p_t := L^p_{[0,T]}$, where $T$ stands for the local existence time unless mentioned otherwise. The Bochner spaces $L^p(0,T;X)$ and $C([0,T];X)$ have their usual meanings, as ($L^p$ and continuous, respectively) maps from $[0,T]$ to a Banach space $X$.
\medskip
We also use the notation $X\lesssim Y$ to imply that there exists a positive constant $C$ such that $X\le CY$. The dependence of the constant on various parameters (including the initial data), will be denoted using a subscript as $X\lesssim_{k_1,k_2} Y$ or $X\le C_{k_1,k_2}Y$. \bigskip
\subsection{Organization of the paper}
In Section \ref{mathematical model}, we describe the mathematical model and recall the existence results of \cite{Jayanti2022LocalSuperfluidity}, followed by the statement of the main theorem. Section \ref{weak-strong uniqueness proof} is devoted to an explanation of how the standard weak-strong uniqueness is difficult to establish for this model unless the weak solutions are sufficiently regular (at which point weak-moderate uniqueness itself applies). Finally, the main result concerning the moderate solutions is proved in Section \ref{local uniqueness of weak solutions}, using the energy structure of the system. \bigskip
\section{Mathematical model and main results} \label{mathematical model}
\subsection{Model and recap of existence results}
The nonlinear Schr\"odinger equation (NLS) is used to describe the dynamics of the superfluid phase, while the incompressible inhomogeneous Navier-Stokes equations (NSE) govern the normal Helium liquid (see \cite{Jayanti2021GlobalEquations} for a discussion on the physics of superfluidity and other related models). The ``Pitaevskii model" that we consider in this work is as follows. Henceforth, all statements are in the context of a smooth, bounded domain $\Omega\subset\mathbb{R}^3$. \medskip
\begin{align}
\partial_t \psi + \Lambda B\psi &= -\frac{1}{2i}\Delta\psi + \frac{\mu}{i}\lvert\psi\rvert^2 \psi \tag{NLS} \label{NLS} \\
B = \frac{1}{2}\left(-i\nabla - u \right)^2 + \mu \lvert \psi \rvert^2 &= -\frac{1}{2}\Delta + iu\cdot\nabla + \frac{1}{2}\lvert u \rvert^2 + \mu \lvert \psi \rvert^2 \tag{CPL} \label{coupling} \\
\partial_t \rho + \nabla\cdot(\rho u) &= 2\Lambda\Re(\Bar{\psi}B\psi) \tag{CON} \label{continuity} \\
\partial_t (\rho u) + \nabla\cdot (\rho u \otimes u) + \nabla p - \nu \Delta u &= \!\begin{multlined}[t]
-2\Lambda \Im(\nabla \Bar{\psi}B\psi) + \Lambda\nabla \Im(\Bar{\psi}B\psi) \tag{NSE} \label{NSE} + \frac{\mu}{2}\nabla\lvert\psi\rvert^4
\end{multlined} \\
\nabla\cdot u &= 0 \tag{DIV} \label{divergence-free}
\end{align}\medskip
We supplement the equations with initial and boundary conditions\footnote{For a justification of the exclusion of $t=0$ in the boundary conditions for the wavefunction, see Remark 2.4 in \cite{Jayanti2022LocalSuperfluidity}.} as follows.
\begin{equation*} \tag{INI} \label{initial conditions}
\psi(0,x) = \psi_0(x) \qquad u(0,x) = u_0(x) \qquad \rho(0,x) = \rho_0(x) \quad a.e. \ x\in\Omega
\end{equation*}\medskip
\begin{equation} \tag{BC} \label{boundary conditions}
\begin{aligned}
u = \frac{\partial u}{\partial n} = 0 \quad &a.e. \ (t,x)\in[0,T]\times\partial\Omega \\
\psi = \frac{\partial \psi}{\partial n} = \frac{\partial^2 \psi}{\partial n^2} = \frac{\partial^3 \psi}{\partial n^3} = 0 \quad &a.e. \ (t,x)\in(0,T]\times\partial\Omega \notag
\end{aligned}
\end{equation}\medskip
\noindent where $n$ is the outward normal direction on the boundary $\partial\Omega$, and $T$ is the local existence time. For more details about the model, refer to the discussion in \cite{Jayanti2022LocalSuperfluidity} or the derivation in \cite{Pitaevskii1959PhenomenologicalPoint}. Weak solutions that we seek are those that satisfy the governing equations in the sense of distributions, for a certain class of test functions. \medskip
\begin{mydef}[Weak solutions\footnote{See Remark \ref{strong or weak solutions?}.}] \label{definition of weak solutions}
Let $\Omega \subset \mathbb{R}^3$ be a bounded set with a smooth boundary $\partial\Omega$. For a given time $T>0$, consider the following test functions:
\begin{enumerate}
\item a complex-valued scalar field $\varphi \in H^1(0,T;L^2(\Omega))\cap L^2(0,T;H_0^1(\Omega))$,
\item a real-valued, divergence-free (3D) vector field $\Phi \in H^1(0,T;L^2_d(\Omega)) \cap L^2(0,T;H^1_d(\Omega))$, and
\item a real-valued scalar field $\sigma \in H^1(0,T;L^2(\Omega))\cap L^2(0,T;H^1(\Omega))$.
\end{enumerate}
A triplet $(\psi,u,\rho)$ is called a \textbf{weak solution} to the Pitaevskii model if:
\begin{enumerate} [(i)]
\item
\begin{equation}
\begin{gathered}
\psi\in L^2(0,T;H^{\frac{7}{2}+\delta}_0(\Omega)) \\
u\in L^2(0,T;H^{\frac{3}{2}+\delta}_d(\Omega)) \\
\rho \in L^{\infty}([0,T]\times\Omega)
\end{gathered}
\end{equation}\medskip
\item and they satisfy the governing equations in the sense of distributions for all test functions, i.e.,
\begin{multline} \label{weak solution wavefunction}
-\int_0^T \int_{\Omega} \left[ \psi\partial_t\Bar{\varphi} + \frac{1}{2i}\nabla\psi\cdot\nabla\Bar{\varphi} - \Lambda\Bar{\varphi}B\psi - i\mu\Bar{\varphi}\lvert \psi \rvert^2\psi \right] dx \ dt \\ = \int_{\Omega} \left[ \psi_0\Bar{\varphi}(t=0) - \psi(T)\Bar{\varphi}(T) \right] dx
\end{multline}
\begin{multline} \label{weak solution velocity}
-\int_0^T \int_{\Omega} \left[ \rho u\cdot \partial_t \Phi + \rho u\otimes u:\nabla\Phi - \nu\nabla u:\nabla\Phi - 2\Lambda\Phi\cdot\Im(\nabla\Bar{\psi}B\psi) \right] dx \ dt \\ = \int_{\Omega} \left[ \rho_0 u_0 \Phi(t=0) - \rho(T)u(T)\Phi(T) \right] dx
\end{multline}
\begin{equation} \label{weak solution density}
-\int_0^T \int_{\Omega} \left[ \rho \partial_t \sigma + \rho u\cdot\nabla\sigma + 2\Lambda \sigma \Re(\Bar{\psi}B\psi) \right] dx \ dt = \int_{\Omega} \left[ \rho_0 \sigma(t=0) - \rho(T)\sigma(T) \right] dx
\end{equation}
\noindent where (the initial data) $\psi_0 \in H^{\frac{5}{2}+\delta}_0(\Omega)$, $u_0 \in H^{\frac{3}{2}+\delta}_d(\Omega)$ and $\rho_0 \in L^{\infty}(\Omega)$. \medskip
\end{enumerate}\medskip
\end{mydef}\medskip
In \cite{Jayanti2022LocalSuperfluidity}, the following existence result was proven. \medskip
\begin{thm} [Local existence] \label{local existence}
For any $\delta \in (0,\frac{1}{2})$, let $\psi_0 \in H^{\frac{5}{2}+\delta}_0(\Omega)$ and $u_0\in H^{\frac{3}{2}+\delta}_d(\Omega)$. Suppose $\rho_0$ is bounded both above and below a.e. in $\Omega$, i.e., $0<m\le\rho_0\le M<\infty$. Then, there exists a local existence time $T$ and at least one weak solution $(\psi,u,\rho)$ to the Pitaevskii model. In particular, the weak solutions have the following regularity:
\begin{gather}
\psi\in C([0,T];H^{\frac{5}{2}+\delta}_0(\Omega)) \cap L^2(0,T;H^{\frac{7}{2}+\delta}_0(\Omega)) \label{weak solution psi regularity} \\
u\in C([0,T];H^{\frac{3}{2}+\delta}_d(\Omega)) \cap L^2(0,T;H^{2}_d(\Omega)) \label{weak solution u regularity} \\
\rho\in L^{\infty}([0,T]\times \Omega)\cap C([0,T];L^2(\Omega)) \label{weak solution rho regularity}
\end{gather
\noindent where $T$ depends (inversely) on:
\begin{enumerate}[(i)]
\item $\varepsilon\in (0,m)$, the allowed infimum of the density field, and
\item $X_0 = 1 + \norm{\Delta\psi_0}_{L^2_x}^2 + \nu\norm{\nabla u_0}_{L^2_x}^2$
\end{enumerate}
\end{thm}\medskip
\begin{rem} \label{strong or weak solutions?}
The regularity of the solutions seem to suggest that the wavefunction and velocity are strong solutions. Indeed this is true, as they are strongly continuous in their topologies. On the other hand, the density is truly a weak solution and is the reason for referring to the triplet as a weak solution.
\end{rem} \medskip
\begin{rem}
For fixed values of $\nu$ and $\Lambda$, it was argued in Section 2.1 of \cite{Jayanti2022LocalSuperfluidity} that the local existence time $T$ grows with decreasing $\varepsilon$ (up to a finite maximum value).
\end{rem}\medskip
The solutions also satisfy some bounds/estimates (some derived a priori), summarized here for convenience. \medskip
\begin{enumerate}
\item \textit{Superfluid mass bound}:
\begin{equation} \label{superfluid mass bound}
\norm{\psi}_{L^2_x}(t) \le \norm{\psi_0}_{L^2_x} \quad a.e. \ t \in [0,T]
\end{equation}\medskip
\item \textit{Normal fluid density bound}: Given our choice of $T$, we know that
\begin{equation} \label{density upper lower bound}
0<\varepsilon<\rho(t,x)<M' := M+m-\varepsilon \ , \ a.e. \ (t,x)\in [0,T]\times\Omega
\end{equation}\medskip
\item \textit{Energy equation}:
\begin{equation} \label{energy equality}
\begin{multlined}
\left( \frac{1}{2}\norm{\sqrt{\rho}u}_{L^2_x}^2 + \frac{1}{2}\norm{\nabla\psi}_{L^2_x}^2 + \frac{\mu}{2}\norm{\psi}_{L^4_x}^4 \right)(t) + \nu\norm{\nabla u}_{L^2_{[0,t]}L^2_x}^2 + 2\Lambda\norm{B\psi}_{L^2_{[0,t]}L^2_x}^2 \\ = \frac{1}{2}\norm{\sqrt{\rho_0}u_0}_{L^2_x}^2 + \frac{1}{2}\norm{\nabla\psi_0}_{L^2_x}^2 + \frac{\mu}{2}\norm{\psi_0}_{L^4_x}^4 =: E_0 \quad a.e. \ t\in [0,T]
\end{multlined}
\end{equation}\medskip
\item \textit{Higher-order energy estimate}:
\begin{equation} \label{higher-order energy estimate}
\begin{gathered}
X(t) \le 2X_0 \quad a.e. \ t\in[0,T] \qquad , \qquad \int_0^T Y(\tau) d\tau \le 31X_0 \\
where \ X(t) = 1 + \norm{\Delta\psi}_{L^2_x}^2 + \nu\norm{\nabla u}_{L^2_x}^2 \quad , \quad X_0 = X(0) \\
Y = \Lambda\norm{\nabla(B\psi)}_{L^2_x}^2 + \norm{\sqrt{\rho}\partial_t u}_{L^2_x}^2 + \frac{\nu^2}{M'}\norm{\Delta u}_{L^2_x}^2
\end{gathered}
\end{equation}\medskip
\item \textit{Highest-order energy estimate}:
\begin{equation} \label{highest-order energy estimate}
\begin{gathered}
\norm{\psi}_{H^{\frac{5}{2}+\delta}_x}^2 (t) \lesssim \norm{\psi_0}_{H^{\frac{5}{2}+\delta}_x}^2 e^{c Q_T} \quad a.e. \ t\in[0,T] \\
\norm{B\psi}_{L^2_{[0,T]}H^{\frac{3}{2}+\delta}_x} \lesssim \Lambda^{-\frac{1}{2}} (Q_T^{\frac{1}{2}} + 1) e^{c Q_T} \norm{\psi_0}_{H^{\frac{5}{2}+\delta}_x} \\
\norm{u}_{H^{\frac{3}{2}+\delta}_x}^2(t) \lesssim \norm{u_0}_{H^{\frac{3}{2}+\delta}_x}^2 + \frac{1}{\nu}\sqrt{\frac{M'}{\varepsilon}}X_0 \\
where \ Q_T = \Lambda\frac{M'}{\nu^2}X_0 + \left(\Lambda \frac{M'}{\nu^2\varepsilon} + \gamma\right)X_0^2 T + \Lambda E_1^2 T \quad , \quad E_1 = \norm{u_0}_{H^{\frac{3}{2}+\delta}_x}^2 + \norm{\psi_0}_{H^{\frac{5}{2}+\delta}_x}^2
\end{gathered}
\end{equation}\medskip
\item \textit{Time-derivative bounds}:
\begin{equation} \label{time-derivative bounds}
\begin{gathered}
\norm{\partial_t \psi}_{L^2_{[0,T]}L^2_x} \lesssim \norm{B\psi}_{L^2_{[0,T]}L^2_x} + T^{\frac{1}{2}}\norm{\psi}_{L^{\infty}_{[0,T]}H^2_x} + \mu T^{\frac{1}{2}} \norm{\psi}_{L^{\infty}_{[0,T]}H^1_x}^3 \\
\norm{\partial_t u}_{L^2_{[0,T]}L^2_x} \lesssim \frac{1}{\sqrt{\varepsilon}} X_0 \\
\norm{\partial_t \rho}_{L^2_{[0,T]}H^{-1}_x} \lesssim \left( T E_0 M' \right)^{\frac{1}{2}} + \left( \Lambda^{-1} E_0 X_0 \right)^{\frac{1}{2}}
\end{gathered}
\end{equation}\medskip
\end{enumerate} \medskip
In the derivation of \eqref{energy equality}, we make use of the \textit{non-conservative form} of \eqref{NSE}. Note that $\Tilde{p}$ is simply a modified pressure that includes gradient terms on the RHS of the momentum equation. \medskip
\begin{equation} \label{NSE'}
\rho\partial_t u + \rho u\cdot\nabla u + \nabla \Tilde{p} - \nu \Delta u = -2\Lambda \Im(\nabla \Bar{\psi}B\psi) - 2\Lambda u\Re(\Bar{\psi}B\psi) \tag{NSE'}
\end{equation}\bigskip
\subsection{Uniqueness results}
Given that weak solutions exist, it is instructive to ask whether they are unique and under what additional assumptions, if any. The main result of this article is a ``weak-moderate" uniqueness theorem. But first, we will discuss the more standard (and desirable) weak-strong uniqueness statement, i.e., starting from the same initial conditions, if we have a weak solution and a strong solution (in the precise sense described below), then they are both identical almost everywhere. \medskip
\begin{conj} [Weak-strong uniqueness] \label{weak-strong uniqueness}
Let $(\psi,u,\rho)$ be a weak solution to the Pitaevskii model, whose regularity is governed by \eqref{superfluid mass bound}-\eqref{energy equality}, i.e.,
\begin{gather}
\psi\in C([0,T];H^1_0(\Omega)) \cap L^2(0,T;H^2_0(\Omega)) \label{weak solution psi regularity least} \\
u\in L^{\infty}([0,T];L^2_d(\Omega)) \cap L^2(0,T;H^1_d(\Omega)) \label{weak solution u regularity least} \\
\rho\in L^{\infty}([0,T]\times \Omega)\cap C([0,T];L^2(\Omega)) \label{weak solution rho regularity least}
\end{gather
Suppose there exists a strong solution $(\Tilde{\psi},\Tilde{u},\Tilde{\rho})$ starting from the same initial conditions in $H^1_0 \times L^2_d \times L^{\infty}$ (such that both these solutions exist up to some time $T$). Specifically, it is sufficient that the strong solution has the following additional regularity in addition to \eqref{weak solution psi regularity least}-\eqref{weak solution rho regularity least}:
\begin{equation} \label{increased regularity for weak-strong uniqueness}
\begin{gathered}
\Tilde{\rho}\in L^2(0,T;W^{1,3}(\Omega)) \quad , \quad \Tilde{u} \in L^2(0,T;W^{1,\infty}(\Omega)) \\
\partial_t \Tilde{u} \in L^2(0,T;L^{\infty}(\Omega))
\end{gathered}
\end{equation}\medskip
Then, the solutions are identical a.e. in $[0,T]\times\Omega$.
\end{conj}\medskip
\begin{rem} \label{global weak solutions}
If we have weak solutions whose regularity is determined purely by the energy equation, then it is tempting to assume that such solutions would be global in time. Unfortunately, this is not the case, since we do not have a way of guaranteeing a positive lower bound for the density, globally in time. At the moment, it is not clear at all how to approach such a problem, due to the presence of the sign-indefinite term on the RHS of \eqref{continuity}. A more elaborate discussion on this issue can be found in \cite{Jayanti2022LocalSuperfluidity}, and this is an interesting open question worth exploring in the future.
\end{rem}\medskip
Based on the regularity in \eqref{weak solution psi regularity least}, one can easily see that $\partial_t \psi$ is already in $L^2(0,T;L^2(\Omega))$, which is why no improvement is needed for the smoothness of $\Tilde{\psi}$. In fact, the wavefunction has enough regularity to be classified as a strong solution; it is the velocity and density that make the solutions ``weak". While this conjecture works for the Navier-Stokes equations, we run into a problem here: the coupling to the NLS. This gives rise to terms that cannot be handled at the level of regularity of the weak solutions. To this end, we will outline the main steps in the analysis of the momentum equation, to clarify exactly where the difficulties arise. \medskip
Furthermore, in the above conjecture, we observe that $\Tilde{u}$ needs to have bounded (in space) spatio-temporal derivatives. As much as this is not unusual to expect from strong solutions, it is possible to trade this rather high regularity for slightly smoother weak solutions. This is stated in the next theorem, and will be referred to as ``weak-moderate uniqueness" since the notion of strong solution is significantly relaxed. \medskip
\begin{thm} [Weak-moderate uniqueness] \label{local uniqueness}
Let $(\psi,u,\rho)$ be a weak solution to the Pitaevskii model, with regularity governed by \eqref{superfluid mass bound}-\eqref{higher-order energy estimate}, i.e.,
\begin{gather}
\psi\in C([0,T];H^{2}_0(\Omega)) \cap L^2(0,T;H^{3}_0(\Omega)) \label{weak solution psi regularity reduced} \\
u\in C([0,T];H^{1}_d(\Omega)) \cap L^2(0,T;H^{2}_d(\Omega)) \label{weak solution u regularity reduced} \\
\rho\in L^{\infty}([0,T]\times \Omega)\cap C([0,T];L^2(\Omega)) \label{weak solution rho regularity reduced}
\end{gather
Suppose there exists a ``moderate" solution $(\Tilde{\psi},\Tilde{u},\Tilde{\rho})$ starting from the same initial data (in $H^2_0 \times H^1_d \times L^{\infty}$), such that $\Tilde{\rho} \in L^2(0,T;W^{1,3}(\Omega)), \ \partial_t\Tilde{u}\in L^2(0,T;L^3(\Omega))$ in addition to the regularity in \eqref{weak solution psi regularity reduced}-\eqref{weak solution rho regularity reduced}. Then, the solutions are identical a.e. in $[0,T]\times\Omega$. \medskip
\end{thm}\medskip
It is to be noted that in the proof of Theorem \ref{local existence}, we already established that for a weak solution, $\partial_t u\in L^2(0,T;L^2)$ $-$ evident from \eqref{density upper lower bound} and \eqref{higher-order energy estimate}. So, as far as the (time derivative of the) ``moderate" velocity goes, we have only assumed a slight increase in the spatial integrability, from $L^2$ to $L^3$ (as opposed to $L^{\infty}$ in weak-strong uniqueness). Also noteworthy is the absence of any additional (over the weak solution) bound on the spatial derivatives of $\Tilde{u}$. The regularity of the ``moderate" density remains unchanged from Conjecture \ref{weak-strong uniqueness}. \medskip
\begin{rem} \label{lower regularity of weak solutions in weak-moderate result}
It is important to recognize that the regularity of the weak solutions in Theorem \ref{local uniqueness} is lower than those in Theorem \ref{local existence}. As mentioned before, this allows wiggle room for enhanced (less regular) existence results without compromising uniqueness.
\end{rem}\bigskip
\bigskip
Summarized below are some well-known properties of Sobolev and Lebesgue spaces (see \cite{Temam1977Navier-StokesAnalysis} and Chapter 5 of \cite{Evans2010PartialEquations}) that will be repeatedly utilized in the calculations that follow. We will begin with a discussion on the difficulties of proving Conjecture \ref{weak-strong uniqueness} in the next section. \medskip
\begin{lem}[Sobolev embeddings] \label{sobolev embeddings}
For $\Omega$ a smooth, bounded subset of $\mathbb{R}^3$,
\begin{enumerate}
\item $H^1(\Omega)\subset L^6(\Omega)$
\item $H^s(\Omega) \subset L^{\infty}(\Omega) \ \forall \ s>\frac{3}{2}$ \ ; \ $H^s(\mathbb{R}) \subset L^{\infty}(\mathbb{R}) \ \forall \ s>\frac{1}{2}$
\item $H^s(\Omega) \Subset H^{s'}(\Omega) \ \forall \ s,s'\in\mathbb{R}, s>s'$
\end{enumerate}
\end{lem}\medskip
\begin{lem}[Inequalities and interpolations] \label{inequalities and interpolations}
For $\Omega$ a smooth, bounded subset of $\mathbb{R}^3$,
\begin{enumerate}
\item Poincar\'e's inequality: $\norm{f}_{L^p} \lesssim \norm{\nabla f}_{L^p} \ \forall \ 1<p<\infty, f\in W^{1,p}_0(\Omega) \Rightarrow \norm{f}_{W^{1,p}} \equiv \norm{\nabla f}_{L^p}$
\item Ladyzhenskaya's inequality: $\norm{f}_{L^4} \lesssim \norm{f}_{L^2}^{\frac{1}{4}} \norm{\nabla f}_{L^2}^{\frac{3}{4}} \ \forall \ f\in H^1_0(\Omega)$
\item Agmon's inequality: $\norm{f}_{L^{\infty}} \lesssim \norm{f}_{H^1}^{\frac{1}{2}} \norm{f}_{H^2}^{\frac{1}{2}} \ \forall \ f\in H^1_0(\Omega)\cap H^2(\Omega)$
\item Lebesgue interpolation: $\norm{f}_{L^3} \lesssim \norm{f}_{L^2}^{\frac{1}{2}} \norm{f}_{L^6}^{\frac{1}{2}} \ \forall \ f\in L^2 \cap L^6$
\end{enumerate}
\end{lem}\bigskip
\section{Remark on weak-strong uniqueness (Conjecture \ref{weak-strong uniqueness})} \label{weak-strong uniqueness proof}
In this section, we will investigate Conjecture \ref{weak-strong uniqueness}, and our modus operandi is motivated by the classical results of Section 2.5 in \cite{Lions1996MathematicalMechanics}. Our goal is to demonstrate that the regularity in \eqref{weak solution psi regularity least}-\eqref{weak solution rho regularity least} is not sufficient to show a weak-strong uniqueness result, even if the strong solutions are smooth. \medskip
Consider the momentum equation. In what follows, $\Psi = -2\Lambda u \Re(\overline{\psi}B\psi)$ and $\Psi' = -2\Lambda\Im(\nabla\overline{\psi}B\psi)$. We begin from \eqref{weak solution velocity}, setting $\Phi = \Tilde{u}$. This way, we can arrive at an expression for $\int_{x} \rho u\cdot\Tilde{u}$. \medskip
\begin{equation} \label{rho u.Tilde(u) 1}
\int_{\Omega} \rho u\cdot\Tilde{u} + \nu\int_{t,\Omega} \nabla u:\nabla\Tilde{u} = \int_{\Omega} \rho_0 \abs{u_0}^2 + \int_{t,\Omega} \rho u\cdot\left( \partial_t \Tilde{u} + \Tilde{u}\cdot\nabla\Tilde{u} \right) + \int_{t,\Omega} \Psi' \cdot\Tilde{u}
\end{equation}\medskip
We write \eqref{NSE'} for the strong solution, and rearrange it to get:
\begin{equation} \label{NSE strong solution - rearranged}
\rho\partial_t \Tilde{u} + \rho u\cdot\nabla\Tilde{u} + \nabla \Tilde{p} -\nu\Delta\Tilde{u} = \Tilde{\Psi} + \Tilde{\Psi'} + (\rho - \Tilde{\rho})\partial_t \Tilde{u} + (\rho - \Tilde{\rho})\Tilde{u}\cdot\nabla\Tilde{u} + \rho (u - \Tilde{u})\cdot\nabla\Tilde{u}
\end{equation}\medskip
Multiplying \eqref{NSE strong solution - rearranged} by $u$ and integrating, \medskip
\begin{equation} \label{NSE strong solution - rearranged, multiplied by u}
\begin{multlined}
\int_{t,\Omega}\rho u\partial_t \Tilde{u} + \int_{t,\Omega}\rho u\otimes u:\nabla\Tilde{u} +\nu\int_{t,\Omega}\nabla u:\nabla\Tilde{u} = \int_{t,\Omega}\Tilde{\Psi}\cdot u + \int_{t,\Omega}\Tilde{\Psi'}\cdot u \\ + \int_{t,\Omega}(\rho - \Tilde{\rho})u\cdot\left(\partial_t \Tilde{u} + \Tilde{u}\cdot\nabla\Tilde{u}\right) + \int_{t,\Omega}\rho (u - \Tilde{u})\otimes u:\nabla\Tilde{u}
\end{multlined}
\end{equation}\medskip
Adding \eqref{rho u.Tilde(u) 1} and \eqref{NSE strong solution - rearranged, multiplied by u}, \medskip
\begin{equation} \label{rho u.Tilde(u)}
\begin{multlined}
\int_{\Omega} \rho u\cdot\Tilde{u} + 2\nu\int_{t,\Omega} \nabla u:\nabla\Tilde{u} = \int_{\Omega} \rho_0 \abs{u_0}^2 + \int_{t,\Omega}(\rho - \Tilde{\rho})u\cdot\left(\partial_t \Tilde{u} + \Tilde{u}\cdot\nabla\Tilde{u}\right) \\ + \int_{t,\Omega}\rho (u - \Tilde{u})\otimes u:\nabla\Tilde{u} + \int_{t,\Omega} \left[ \Psi' \cdot\Tilde{u} + \Tilde{\Psi} \cdot u + \Tilde{\Psi'} \cdot u \right]
\end{multlined}
\end{equation}\medskip
Next, we take the inner product of both \eqref{NSE} and \eqref{NSE'} with $u$, use incompressibility, and add them, to arrive at the energy equation for the normal fluid alone.
\begin{equation} \label{normal fluid energy equation}
\frac{d}{dt}\frac{1}{2}\norm{\sqrt{\rho}u}_{L^2_x}^2 + \nu\norm{\nabla u}_{L^2_x}^2 = - 2\Lambda\int_{\Omega}u\cdot\Im(\nabla \Bar{\psi}B\psi) - \Lambda\int_{\Omega}\abs{u}^2\Re(\Bar{\psi}B\psi)
\end{equation}\medskip
This obviously holds as an a priori estimate, valid for the approximate fields $(\psi^N,u^N,\rho^N)$, but given the regularity of the solution, we can easily pass to the limit to see that it is accurate for the weak solutions too. From this equation, integrating on $[0,t]$ for $0\le t\le T$, we can obtain an equation for $\int_{\Omega} \rho u\cdot u$. \medskip
\begin{equation} \label{rho u.u}
\frac{1}{2}\int_{\Omega} \rho \abs{u}^2 + \nu\int_{t,\Omega} \abs{\nabla u}^2 = \frac{1}{2}\int_{\Omega} \rho_0 \abs{u_0}^2 + \int_{t,\Omega} \left[ \frac{1}{2}\Psi\cdot u + \Psi' \cdot u \right]
\end{equation}\medskip
Finally, multiply \eqref{NSE strong solution - rearranged} by $\Tilde{u}$ and integrate over $[0,t]\times \Omega$. Since $\Tilde{u}$ is a strong solution, $\frac{1}{2}\abs{\Tilde{u}}^2$ can work as a test function. Therefore, we use \eqref{weak solution density} to simplify, and the resulting equation is: \medskip
\begin{equation} \label{rho Tilde(u).Tilde(u)}
\begin{multlined}
\frac{1}{2}\int_{\Omega} \rho \abs{\Tilde{u}}^2 + \nu\int_{t,\Omega} \abs{\nabla \Tilde{u}}^2 = \frac{1}{2}\int_{\Omega} \rho_0 \abs{u_0}^2 + \int_{t,\Omega}(\rho-\Tilde{\rho})\Tilde{u}\cdot\left( \partial_t \Tilde{u} + \Tilde{u}\cdot\nabla\Tilde{u} \right) \\ + \int_{t,\Omega}\rho (u-\Tilde{u})\otimes \Tilde{u}:\nabla\Tilde{u} + \int_{t,\Omega} \left[ \Tilde{\Psi} + \Tilde{\Psi'} \right]\cdot\Tilde{u} + \Lambda\Re\int_{t,\Omega}\overline{\psi}B\psi \abs{\Tilde{u}}^2
\end{multlined}
\end{equation}\medskip
Add \eqref{rho u.u} and \eqref{rho Tilde(u).Tilde(u)}, and subtract \eqref{rho u.Tilde(u)}. Then, differentiate the result with respect to $t$. \medskip
\begin{equation} \label{weak-strong uniqueness for velocity - preliminary}
\begin{aligned}
\renewcommand{\arraystretch}{2}
\left. \begin{array}{c}
\frac{d}{dt}\frac{1}{2}\norm{\sqrt{\rho}(u - \Tilde{u})}_{L^2_x}^2 \\
+ \ \nu\norm{\nabla(u-\Tilde{u})}_{L^2_x}^2
\end{array} \right\} = &-\underbrace{\int_{\Omega} (\rho-\Tilde{\rho})(u-\Tilde{u})\cdot\left( \partial_t \Tilde{u} + \Tilde{u}\cdot\nabla\Tilde{u} \right)}_{\circled{I-1}} - \underbrace{\int_{\Omega} \rho(u-\Tilde{u})\otimes(u-\Tilde{u}):\nabla\Tilde{u}}_{\circled{I-2}} \\
&- \underbrace{\Lambda\Re\int_{\Omega} (\overline{\psi}B\psi - \overline{\Tilde{\psi}}\Tilde{B}\Tilde{\psi})(\abs{u}^2 - \abs{\Tilde{u}}^2)}_{\circled{I-3}} - \underbrace{\Lambda\Re\int_{\Omega} \overline{\Tilde{\psi}}\Tilde{B}\Tilde{\psi}\abs{u-\Tilde{u}}^2}_{\circled{I-4}} \\
&- \underbrace{2\Lambda\Im\int_{\Omega} (\nabla\overline{\psi}B\psi - \nabla\overline{\Tilde{\psi}}\Tilde{B}\Tilde{\psi})\cdot(u-\Tilde{u})}_{\circled{I-5}}
\end{aligned}
\end{equation}\medskip
Of course, the next step in this process is to reduce the above equation to a form that is amenable to the Gr\"onwall's inequality, and this is usually achieved by employing H\"older's and Young's inequalities to extract quadratic terms of interest. Since the strong solutions are sufficiently smooth, it is easy to see (see Section 2.5 in \cite{Lions1996MathematicalMechanics} for details) that \circled{I-1} and \circled{I-2} can be handled in a straightforward manner, by extracting terms like $\norm{\rho-\Tilde{\rho}}_{L^2_x}^2$ and $\norm{\sqrt{\rho}(u-\Tilde{u})}_{L^2_x}^2$. In the usual Navier-Stokes setting, these are the only terms. However, here we have additional coupling terms. While \circled{I-4} can be handled in the same way as the previous ones, the other integrals pose problems. For instance, during the analysis of \circled{I-5}, one would encounter a term of the form: \medskip
\begin{equation*}
\int_{\Omega} B\psi \ (u-\Tilde{u}) \cdot \nabla\overline{\varphi}
\end{equation*}\medskip
\noindent where $\varphi = \psi-\Tilde{\psi}$. Recall that based on our assumptions in Conjecture \ref{weak-strong uniqueness}, we only have $\psi\in L^{\infty}_t H^1_x \cap L^2_t H^2_x$, $B\psi\in L^2_{t,x}$ and $u\in L^{\infty}_t L^2_x \cap L^2_t H^1_x$. It is clear that no splitting of Lebesgue indices (using H\"older's inequality) will be satisfactory: $B\psi$ has to be in $L^2_x$, which means the other two factors can be in something like $L^6_x$ and $L^3_x$, say. Then, the one in $L^6_x$ is extracted as a dissipative factor, and leaves $B\psi$ and the other factor squared. We can then use interpolation to replace the $L^3_x$ factor by $L^2_x$ and $L^6_x$. Once again, we have to extract out another dissipative factor, which results in $\norm{B\psi}_{L^2_x}^4$. Other combinations of Lebesgue indices also lead to these kinds of dead-ends. \medskip
We run into a very similar wall with the corresponding term in \circled{I-3}. Mirroring the above argument, since $B\psi\in L^2_x$, the only option we are left with is to have each of the remaining factors in $L^6_x$. Extracting out a dissipative velocity contribution of the form $\norm{\nabla(u-\Tilde{u})}_{L^2_x}^2$ yields: \medskip
\begin{align*}
\int_{\Omega} B\psi \ (u+\Tilde{u}) \cdot (u-\Tilde{u}) \overline{\varphi} &\lesssim \norm{B\psi}_{L^2_x} \norm{u+\Tilde{u}}_{L^6_x} \norm{u-\Tilde{u}}_{L^6_x} \norm{\varphi}_{L^6_x} \\
&\lesssim \norm{B\psi}_{L^2_x}^2 \norm{u+\Tilde{u}}_{H^1_x}^2 \norm{\nabla\varphi}_{L^2_x}^2 + \norm{\nabla(u-\Tilde{u})}_{L^2_x}^2
\end{align*}\medskip
Evidently, this too is inadequate, since when the Gr\"onwall inequality is applied, the time integral has to distribute over $\norm{B\psi}_{L^2_x}^2 \norm{u+\Tilde{u}}_{H^1_x}^2$, which is impossible due to the regularity assumed. \medskip
In conclusion, we note that the standard weak-strong uniqueness proof will not work here, even for strong solutions that are smooth. It appears that Conjecture \ref{weak-strong uniqueness} may not hold true. This naturally leads us to consider slightly more regular weak solutions, as is done in Theorem \ref{local uniqueness}. \bigskip
\section{Weak-moderate uniqueness (Proof of Theorem \ref{local uniqueness})} \label{local uniqueness of weak solutions}
We will use the difference between the weak and moderate solutions as the test functions, so as to obtain a Gr\"onwall-type estimate for the $L^{\infty}_{[0,T]} L^2_x$ norm ($L^{\infty}_{[0,T]} H^1_x$ for the wavefunction) of the difference of two solutions. This is possible in the case of the velocity and wavefunction simply because of their regularities. In the case of the density, it is possible to achieve it by working with smooth approximations and passing to the limit (recall that $\rho\in C^0_t L^2_x$). Finally, observe that the increased spatial regularity of $\Tilde{\rho}$ results in an increased temporal regularity too.
\begin{equation} \label{additional time regularity of density}
\begin{aligned}
\norm{\partial_t \Tilde{\rho}}_{L^2_{[0,T]} L^2_x} &\lesssim \norm{\Tilde{u}\cdot\nabla\Tilde{\rho}}_{L^2_{[0,T]} L^2_x} + \norm{\overline{\Tilde{\psi}}\Tilde{B}\Tilde{\psi}}_{L^2_{[0,T]} L^2_x} \\
&\lesssim \norm{\Tilde{u}}_{L^{\infty}_{[0,T]} L^6_x} \norm{\Tilde{\rho}}_{L^2_{[0,T]} W^{1,3}_x} + \norm{\Tilde{\psi}}_{L^{\infty}_{[0,T]} L^{\infty}_x} \norm{\Tilde{B}\Tilde{\psi}}_{L^2_{[0,T]} L^2_x} < \infty
\end{aligned}
\end{equation}\medskip
The basic idea of the proof is to subtract the governing equations for the weak and moderate fields, and use the difference of the fields as the test functions itself, i.e., $\varphi = \psi - \Tilde{\psi}$, $\Phi = u - \Tilde{u}$, and $\sigma = \rho - \Tilde{\rho}$. In what follows, $\kappa$ is a positive real number used to extract out the dissipative terms $\norm{\nabla\Phi}_{L^2_x}^2$ and $\norm{D^2\varphi}_{L^2_x}^2$, and it is small enough that they may be absorbed into the dissipative terms on the LHS of the momentum and Schr\"odinger equations, respectively. \medskip
\subsection{The wavefunction equation} \label{weak uniqueness for wavefunction}
We will begin by acting with the gradient operator on \eqref{NLS} for each of the two solutions $\psi$ and $\Tilde{\psi}$. Then, we will subtract the equations and integrate against $\nabla(\overline{\psi} - \overline{\Tilde{\psi}})$. The real part of the resulting equation is given below. The analysis is similar to the calculations in Section 3.2 of \cite{Jayanti2022LocalSuperfluidity}, in that we will be looking at the energy ($H^1_x$ norm) difference between the two wavefunctions. \medskip
\begin{equation} \label{wavefunction gradient difference step 1}
\frac{d}{dt}\frac{1}{2} \norm{\nabla(\psi - \Tilde{\psi})}_{L^2_x}^2 = -\Lambda\Re\int_{\Omega} \nabla(\overline{\psi}-\overline{\Tilde{\psi}})\cdot\nabla(B\psi - \Tilde{B}\Tilde{\psi}) + \mu\Im \int_{\Omega} \nabla(\overline{\psi}-\overline{\Tilde{\psi}})\cdot\nabla\left(\abs{\psi}^2\psi - \abs{\Tilde{\psi}}^2\Tilde{\psi} \right)
\end{equation}\medskip
Integrating by parts on the RHS, we get a Laplacian term in each of the integrals. In the first integral, we observe that $B\psi - \Tilde{B}\Tilde{\psi} = -\frac{1}{2}\Delta(\psi - \Tilde{\psi}) + \dots$, so that this gives us a ``dissipation" term along with other terms. In the second term of the RHS, we extract the $\Delta(\psi - \Tilde{\psi})$ in $L^2_x$, and absorb it into the aforementioned dissipation term. \medskip
\begin{align*}
-\Lambda\Re\int_{\Omega} \nabla\overline{\varphi}\cdot\nabla(B\psi - \Tilde{B}\Tilde{\psi}) &\begin{multlined}[t]= \Lambda\Re\int_{\Omega} \Delta\varphi \left[-\frac{1}{2}\Delta\varphi + \frac{1}{2}\abs{u}^2\varphi + \frac{1}{2}\Tilde{\psi}(u+\Tilde{u})\cdot\Phi + i\Phi\cdot\nabla\psi \right. \\
\left. + \Tilde{u}\cdot\nabla\varphi + \mu\left( \abs{\psi}^2\varphi + \Tilde{\psi}\abs{\varphi}^2 + \abs{\Tilde{\psi}}^2\varphi + \Tilde{\psi}^2 \overline{\varphi} \right) \right] \end{multlined} \\
&\begin{multlined}[t]
\lesssim -\frac{\Lambda}{2}\norm{D^2\varphi}_{L^2_x}^2 + \frac{\kappa}{2}\Lambda\norm{D^2\varphi}_{L^2_x}^2 \\ + \kappa^{-1}\Lambda\left\Vert\abs{u}^2\varphi + \Tilde{\psi}(u+\Tilde{u})\cdot\Phi + i\Phi\cdot\nabla\psi + \Tilde{u}\cdot\nabla\varphi \right. \\ \left. + \mu\left( \abs{\psi}^2\varphi + \Tilde{\psi}\abs{\varphi}^2 + \abs{\Tilde{\psi}}^2\varphi + \Tilde{\psi}^2 \overline{\varphi} \right)\right\Vert_{L^2_x}^2 \end{multlined}
\end{align*}
\medskip
The last term in \eqref{wavefunction gradient difference step 1} is also handled in a similar manner, and after absorbing the ``dissipation'' term, we arrive at: \medskip
\begin{multline} \label{varphi equation initial}
\frac{d}{dt}\norm{\nabla\varphi}_{L^2_x}^2 + \frac{\Lambda}{2}\norm{D^2\varphi}_{L^2_x}^2 \lesssim \kappa^{-1}\Lambda \left[ \underbrace{\norm{\abs{u}^2\varphi}_{L^2_x}^2}_{\circled{A-1}} + \underbrace{\norm{\Tilde{\psi}(u+\Tilde{u})\cdot\Phi}_{L^2_x}^2}_{\circled{A-2}} + \underbrace{\norm{\Phi\cdot\nabla\psi}_{L^2_x}^2}_{\circled{A-3}} + \underbrace{\norm{\Tilde{u}\cdot\nabla\varphi}_{L^2_x}^2}_{\circled{A-4}} \right. \\ \left. + \mu^2 (1+\Lambda^{-1}) \underbrace{\norm{\abs{\psi}^2\varphi + \Tilde{\psi}\abs{\varphi}^2 + \abs{\Tilde{\psi}}^2\varphi + \Tilde{\psi}^2 \overline{\varphi}}_{L^2_x}^2}_{\circled{A-5} + \circled{A-6} + \circled{A-7} + \circled{A-8}} \right] + \kappa\Lambda \norm{D^2\varphi}_{L^2_x}^2
\end{multline}\medskip
Now, we will look at each of the terms in \eqref{varphi equation initial}.
\begin{enumerate}[(i)]
\item
\begin{equation*}
\circled{A-1} \lesssim \Lambda\norm{u}_{L^6_x}^4\norm{\varphi}_{L^6_x}^2 \lesssim \Lambda\norm{u}_{H^1_x}^4 \norm{\nabla\varphi}_{L^2_x}^2
\end{equation*}
\item
\begin{align*}
\circled{A-2} &\lesssim \kappa^{-1}\Lambda \norm{\Tilde{\psi}}_{L^{\infty}_x}^2 \norm{u+\Tilde{u}}_{L^6_x}^2 \norm{\Phi}_{L^3_x}^2 \\
&\lesssim \kappa^{-3}\nu^{-1}\Lambda^2 \norm{u+\Tilde{u}}_{H^1_x}^4 \norm{\Tilde{\psi}}_{H^2_x}^4 \norm{\Phi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}
\item\begin{align*}
\circled{A-3} &\lesssim \kappa^{-1}\Lambda \norm{\nabla\psi}_{L^6_x}^2 \norm{\Phi}_{L^3_x}^2 \\
&\lesssim \kappa^{-3}\nu^{-1}\Lambda^2 \norm{\psi}_{H^2_x}^4 \norm{\Phi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{A-4} &\lesssim \kappa^{-1}\Lambda \norm{\Tilde{u}}_{L^6_x}^2 \norm{\nabla\varphi}_{L^3_x}^2 \\
&\lesssim \kappa^{-3}\Lambda \norm{\Tilde{u}}_{H^1_x}^4 \norm{\nabla\varphi}_{L^2_x}^2 + \kappa\Lambda\norm{D^2\varphi}_{L^2_x}^2
\end{align*}\medskip
\noindent where we have used $L^3$ interpolation. Note that the last term above, combined with the last term on the RHS of \eqref{varphi equation initial}, can be absorbed into the LHS of \eqref{varphi equation initial}. This is the ``dissipation'' term we seek to extract out whenever appropriate. \medskip
\item
\begin{align*}
\circled{A-5} &\lesssim \kappa^{-1}(1+\Lambda)\mu^2 \norm{\psi}_{L^6_x}^4 \norm{\varphi}_{L^6_x}^2 \\
&\lesssim \kappa^{-1}(1+\Lambda)\mu^2 \norm{\psi}_{H^1_x}^4 \norm{\nabla\varphi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{A-6} &\lesssim \kappa^{-1}(1+\Lambda)\mu^2 \norm{\Tilde{\psi}}_{L^6_x}^2 \norm{\varphi}_{L^6_x}^2 \norm{\varphi}_{L^6_x}^2 \\
&\lesssim \kappa^{-1}(1+\Lambda)\mu^2 \norm{\Tilde{\psi}}_{H^1_x}^2 \left( \norm{\psi}_{H^1_x}^2 + \norm{\Tilde{\psi}}_{H^1_x}^2 \right) \norm{\nabla\varphi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{A-7} + \circled{A-8} &\lesssim \kappa^{-1}(1+\Lambda)\mu^2 \norm{\Tilde{\psi}}_{L^6_x}^4 \norm{\varphi}_{L^6_x}^2 \\
&\lesssim \kappa^{-1}(1+\Lambda)\mu^2 \norm{\Tilde{\psi}}_{H^1_x}^4 \norm{\nabla\varphi}_{L^2_x}^2
\end{align*}\medskip
\end{enumerate}
Combining these estimates into \eqref{varphi equation initial} gives us:
\begin{equation} \label{varphi equation final}
\frac{d}{dt} \norm{\nabla\varphi}_{L^2_x}^2 + \frac{\Lambda}{2} \norm{D^2\varphi}_{L^2_x}^2 \le h_1(t) \left[ \norm{\nabla\varphi}_{L^2_x}^2 + \norm{\Phi}_{L^2_x}^2 \right] + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 + \kappa\Lambda\norm{D^2\varphi}_{L^2_x}^2
\end{equation}\medskip
\noindent where $h_1 \in L^1_{[0,T]}$. \bigskip
\subsection{The momentum equation} \label{the momentum equation uniqueness}
The notion of the weak solution holds for all times $0\le t\le T$, where $T$ is the chosen local existence time. Considering \eqref{weak solution velocity} and \eqref{weak solution density} up to a time $t \le T$, and differentiating with respect to $t$, we arrive at:
\begin{equation} \label{weak solution velocity differentiated wrt time}
-\int_{\Omega} \left[ \rho u\cdot \partial_t \Phi + \rho u\otimes u:\nabla\Phi - \nu\nabla u:\nabla\Phi - 2\Lambda\Phi\cdot\Im(\nabla\Bar{\psi}B\psi) \right] = -\frac{d}{dt}\int_{\Omega} \rho(t)u(t)\cdot\Phi(t)
\end{equation}\medskip
\begin{equation} \label{weak solution density differentiated wrt time}
-\int_{\Omega} \left[ \rho \partial_t \sigma + \rho u\cdot\nabla\sigma + 2\Lambda \sigma \Re(\Bar{\psi}B\psi) \right] = -\frac{d}{dt}\int_{\Omega} \rho(t)\sigma(t)
\end{equation}\medskip
Subtracting \eqref{weak solution velocity differentiated wrt time} written for $u$ and $\Tilde{u}$, we get:
\begin{equation} \label{weak solution velocity difference}
\begin{multlined}
-\int_{\Omega} \rho \Phi\cdot \partial_t \Phi - \int_{\Omega}\sigma \Tilde{u}\cdot\partial_t \Phi - \int_{\Omega} \left[ \rho\Phi\otimes u + \rho \Tilde{u}\otimes\Phi + \sigma\Tilde{u}\otimes\Tilde{u} \right]:\nabla\Phi + \nu\norm{\nabla\Phi}_{L^2_x}^2 \\ + 2\Lambda\int_{\Omega} \Phi\cdot \Im\left( \nabla\overline{\psi}B\psi - \nabla\overline{\Tilde{\psi}}\Tilde{B}\Tilde{\psi} \right) = -\frac{d}{dt}\int_{\Omega} \rho\abs{\Phi}^2 - \frac{d}{dt} \int_{\Omega} \sigma \Tilde{u}\cdot\Phi
\end{multlined}
\end{equation}\medskip
The first term on the LHS can be rewritten since we see that $\frac{\abs{\Phi}^2}{2}$ satisfies the requirements to be a test function for the density field in \eqref{weak solution density differentiated wrt time}.
\begin{equation*}
-\int_{\Omega} \rho \Phi\cdot \partial_t \Phi = -\int_{\Omega} \rho \partial_t \frac{\abs{\Phi}^2}{2} = \int_{\Omega} \rho u\cdot \nabla \frac{\abs{\Phi}^2}{2} + \Lambda\Re\int_{\Omega} \abs{\Phi}^2 \overline{\psi}B\psi - \frac{d}{dt}\int_{\Omega} \rho \frac{\abs{\Phi}^2}{2}
\end{equation*}\medskip
Further simplification using H\"older's and Young's inequalities (to appropriately absorb the dissipation $\norm{\nabla\Phi}_{L^2_x}^2$ from the RHS into the LHS) yields,
\begin{equation} \label{phi equation initial}
\begin{aligned}
\renewcommand{\arraystretch}{2}
\left. \begin{array}{c}
\frac{d}{dt}\frac{1}{2}\norm{\sqrt{\rho}\Phi}_{L^2_x}^2 \\
+ \ \frac{\nu}{2}\norm{\nabla\Phi}_{L^2_x}^2
\end{array} \right\} \le &\frac{1}{\nu} \underbrace{\left[ \norm{\rho\Phi\otimes u}_{L^2_x}^2 + \norm{\rho \Tilde{u}\otimes\Phi}_{L^2_x}^2 + \norm{\sigma\Tilde{u}\otimes\Tilde{u}}_{L^2_x}^2 \right]}_{\circled{B-1} + \circled{B-2} + \circled{B-3}} \\
&- 2\Lambda \underbrace{\int_{\Omega} \Phi\cdot\Im\left( \nabla\overline{\varphi}B\psi + \nabla\overline{\Tilde{\psi}}B\varphi + \nabla\overline{\Tilde{\psi}}(B-\Tilde{B})\Tilde{\psi} \right) }_{\circled{B-4} + \circled{B-5} + \circled{B-6}} \\
&- \underbrace{\int_{\Omega}\partial_t (\sigma \Tilde{u})\cdot\Phi}_{\circled{B-7}} - \underbrace{\int_{\Omega}\rho u\cdot \nabla\frac{\abs{\Phi}^2}{2}}_{\circled{B-8}} - \Lambda\underbrace{\int_{\Omega}\abs{\Phi}^2 \Re\overline{\psi}B\psi }_{\circled{B-9}}
\end{aligned}
\end{equation}\medskip
Each of the terms on the RHS can be bounded from above as follows (using the H\"older and Young inequalities, and the results listed in Lemmas \ref{sobolev embeddings} and \ref{inequalities and interpolations}).
\begin{enumerate}[(i)]
\item
\begin{equation*}
\circled{B-1} \lesssim \nu^{-1}\norm{\rho}_{L^{\infty}_x}^2 \norm{u}_{L^{\infty}_x}^2 \norm{\Phi}_{L^2_x}^2 \lesssim \nu^{-1}\norm{\rho}_{L^{\infty}_x}^2 \norm{u}_{H^2_x}^2 \norm{\Phi}_{L^2_x}^2
\end{equation*}\medskip
\item
\begin{equation*}
\circled{B-2} \lesssim \nu^{-1}\norm{\rho}_{L^{\infty}_x}^2 \norm{\Tilde{u}}_{L^{\infty}_x}^2 \norm{\Phi}_{L^2_x}^2 \lesssim \nu^{-1}\norm{\rho}_{L^{\infty}_x}^2 \norm{\Tilde{u}}_{H^2_x}^2 \norm{\Phi}_{L^2_x}^2
\end{equation*}\medskip
\item Using Agmon's inequality to bound the $L^{\infty}$ norm,
\begin{equation*}
\circled{B-3} \lesssim \nu^{-1}\norm{\Tilde{u}}_{L^{\infty}_x}^4 \norm{\sigma}_{L^2_x}^2 \lesssim \nu^{-1}\norm{\Tilde{u}}_{H^1_x}^2 \norm{\Tilde{u}}_{H^2_x}^2 \norm{\sigma}_{L^2_x}^2
\end{equation*}\medskip
\item We interpolate the $L^3$ norm (Lemma \ref{inequalities and interpolations}):
\begin{align*}
\circled{B-4} &\lesssim \Lambda\norm{\nabla\varphi}_{L^6_x} \norm{B\psi}_{L^3_x} \norm{\Phi}_{L^2_x} \\
&\lesssim \kappa^{-1}\Lambda \norm{B\psi}_{L^3_x}^2 \norm{\Phi}_{L^2_x}^2 + \kappa\Lambda \norm{D^2\varphi}_{L^2_x}^2 \\
&\lesssim \kappa^{-1}\Lambda \norm{B\psi}_{L^2_x}\norm{B\psi}_{H^1_x} \norm{\Phi}_{L^2_x}^2 + \kappa\Lambda \norm{D^2\varphi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{B-5} &= \Lambda\Im\int_{\Omega} \Phi\cdot\nabla\overline{\Tilde{\psi}} \left( -\frac{1}{2}\Delta\varphi + \frac{1}{2}\abs{u}^2\varphi + iu\cdot\nabla\varphi + \mu\abs{\psi}^2\varphi \right) \\
&= \circled{B-5.1} + \circled{B-5.2} + \circled{B-5.3} + \circled{B-5.4}
\end{align*}\medskip
\begin{itemize}
\item
\begin{align*}
\circled{B-5.1} &\lesssim \Lambda\norm{\Phi}_{L^3_x} \norm{\nabla\Tilde{\psi}}_{L^6_x} \norm{\Delta\varphi}_{L^2_x} \\
&\lesssim \kappa^{-1}\Lambda \norm{\Tilde{\psi}}_{H^2_x}^2 \norm{\Phi}_{L^3_x}^2 + \kappa\Lambda\norm{\Delta\varphi}_{L^2_x}^2 \\
&\lesssim \kappa^{-3}\nu^{-1}\Lambda \norm{\Tilde{\psi}}_{H^2_x}^4 \norm{\Phi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 + \kappa\Lambda\norm{D^2\varphi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{B-5.2} &\lesssim \Lambda\norm{u}_{L^6_x}^2 \norm{\nabla\Tilde{\psi}}_{L^6_x} \norm{\Phi}_{L^3_x} \norm{\varphi}_{L^6_x} \\
&\lesssim (\kappa\nu)^{-1}\norm{u}_{H^1_x}^4 \norm{\Phi}_{L^2_x}^2 + \Lambda^2 \norm{\Tilde{\psi}}_{H^2_x}^2 \norm{\varphi}_{L^2_x}^2 + \kappa\nu \norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{B-5.3} &\lesssim \Lambda\norm{u}_{L^6_x} \norm{\nabla\Tilde{\psi}}_{L^6_x} \norm{\Phi}_{L^6_x} \norm{\nabla\varphi}_{L^2_x} \\
&\lesssim (\kappa\nu)^{-1}\Lambda^2\norm{u}_{H^1_x}^2 \norm{\Tilde{\psi}}_{H^2_x}^2 \norm{\nabla\varphi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{B-5.4} &\lesssim \Lambda \norm{\psi}_{L^{\infty}_x}^2\norm{\nabla\Tilde{\psi}}_{L^6_x} \norm{\Phi}_{L^3_x} \norm{\varphi}_{L^6_x} \\
&\lesssim (\kappa\nu)^{-1} \norm{\Tilde{\psi}}_{H^2_x}^4 \norm{\Phi}_{L^2_x}^2 + \Lambda^2\norm{\psi}_{H^2_x}^4 \norm{\nabla\varphi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
\end{itemize}
\item First we expand $B-\Tilde{B}$ as: \medskip
\begin{align} \label{B - Tilde(B)}
B - \Tilde{B} &= \frac{1}{2}\left( \abs{u}^2 - \abs{\Tilde{u}}^2 \right) + i(u-\Tilde{u})\cdot\nabla + \mu\left( \abs{\psi}^2 - \abs{\Tilde{\psi}}^2 \right) \notag \\
&= \frac{1}{2} (u + \Tilde{u})\cdot\Phi + i\Phi\cdot\nabla + \mu \left( \abs{\varphi}^2 + \overline{\varphi}\Tilde{\psi} + \varphi\overline{\Tilde{\psi}} \right)
\end{align}\medskip
Then,
\begin{align*}
\circled{B-6} &= \int_{\Omega} \Phi\cdot\Im \left[ \nabla\overline{\Tilde{\psi}} \left\{ \frac{1}{2}(u+\Tilde{u})\cdot\Phi + i\Phi\cdot\nabla + \mu(\abs{\varphi}^2 + \overline{\varphi}\Tilde{\psi} + \varphi \overline{\Tilde{\psi}})\Tilde{\psi} \right\} \Tilde{\psi} \right] \\
&\begin{multlined} \lesssim \norm{\Tilde{\psi}}_{L^{\infty}_x} \norm{\nabla\Tilde{\psi}}_{L^6_x} \norm{u+\Tilde{u}}_{L^6_x} \norm{\Phi}_{L^3_x}^2 + \norm{\nabla\Tilde{\psi}}_{L^6_x}^2 \norm{\Phi}_{L^3_x}^2 + \norm{\Tilde{\psi}}_{L^{\infty}_x}^3 \norm{\nabla\Tilde{\psi}}_{L^3_x} \norm{\Phi}_{L^2_x} \norm{\varphi}_{L^6_x} \\ + \norm{\Tilde{\psi}}_{L^{\infty}_x}^2 \norm{\nabla\Tilde{\psi}}_{L^6_x} \norm{\Phi}_{L^2_x} \left(\norm{\psi}_{L^6_x} + \norm{\Tilde{\psi}}_{L^6_x}\right) \norm{\varphi}_{L^6_x} \end{multlined} \\
&\begin{multlined} \lesssim (\kappa\nu)^{-1}\norm{\Tilde{\psi}}_{H^2_x}^4 \left( 1 + \norm{u+\Tilde{u}}_{H^1_x}^2 \right) \norm{\Phi}_{L^2_x}^2 + \norm{\Tilde{\psi}}_{H^2_x}^2 \left( \norm{\nabla\psi}_{L^2_x}^2 + \norm{\nabla\Tilde{\psi}}_{L^2_x}^2 \right) \norm{\nabla\varphi}_{L^2_x}^2 \\ + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 \end{multlined}
\end{align*}\medskip
\item
\begin{equation*}
\circled{B-7} = -\underbrace{\int_{\Omega}\partial_t \Tilde{u} \cdot\sigma\Phi}_{\circled{B-7.1}} - \underbrace{\int_{\Omega} \partial_t \sigma (\Tilde{u}\cdot\Phi)}_{\circled{B-7.2}}
\end{equation*}\medskip
\begin{itemize}
\item
\begin{align*}
\circled{B-7.1} &\lesssim \norm{\partial_t \Tilde{u}}_{L^3_x} \norm{\sigma}_{L^2_x} \norm{\Phi}_{L^6_x} \\
&\lesssim (\kappa\nu)^{-1} \norm{\partial_t \Tilde{u}}_{L^3_x}^2 \norm{\sigma}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
\item By subtracting the continuity equations for $\rho$ and $\Tilde{\rho}$, we obtain
\begin{equation*}
\partial_t \sigma + \Phi\cdot\nabla\Tilde{\rho} +u\cdot\nabla\sigma = 2\Lambda \Re \left( \overline{\psi}B\psi - \overline{\Tilde{\psi}}\Tilde{B}\Tilde{\psi} \right)
\end{equation*}
which allows us to rewrite $\circled{B-7.2}$ as
\begin{align*}
\circled{B-7.2} = \underbrace{\int_{\Omega} \left(\Tilde{u}\cdot\Phi\right) \left[ -\Phi\cdot\nabla\Tilde{\rho} - u\cdot\nabla\sigma + 2\Lambda \Re \left( \overline{\psi}B\psi - \overline{\Tilde{\psi}}\Tilde{B}\Tilde{\psi} \right) \right]}_{\circled{B-7.2.1}+\circled{B-7.2.2}+\circled{B-7.2.3}}
\end{align*}\medskip
Each of these terms will now be estimated. For the first two terms, we will integrate by parts and interpolate the $L^3$ norm: \medskip
\begin{align*}
\circled{B-7.2.1} &= \int_{\Omega} \Tilde{\rho}\Phi\cdot\nabla(\Phi\cdot\Tilde{u}) = \int_{\Omega} \Tilde{\rho}\Phi\cdot(\nabla\Phi)\cdot\Tilde{u} + \int_{\Omega} \Tilde{\rho}\Phi\otimes\Phi:\nabla\Tilde{u}) \\
&\lesssim \norm{\Tilde{\rho}}_{L^{\infty}_x} \norm{\Phi}_{L^3_x} \norm{\nabla\Phi}_{L^2_x} \norm{\Tilde{u}}_{L^6_x} + \norm{\Tilde{\rho}}_{L^{\infty}_x} \norm{\Phi}_{L^3_x} \norm{\Phi}_{L^6_x} \norm{\nabla\Tilde{u}}_{L^2_x} \\
&\lesssim (\kappa\nu)^{-1}\norm{\Tilde{\rho}}_{L^{\infty}_x}^2 \norm{\Tilde{u}}_{H^1_x}^2 \norm{\Phi}_{L^3_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 \\
&\lesssim (\kappa\nu)^{-3}\norm{\Tilde{\rho}}_{L^{\infty}_x}^4 \norm{\Tilde{u}}_{H^1_x}^4 \norm{\Phi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*} \medskip
\begin{align*}
\circled{B-7.2.2} &= \int_{\Omega}\sigma u\cdot\nabla(\Phi\cdot\Tilde{u}) = \int_{\Omega}\sigma u\cdot(\nabla\Phi)\cdot\Tilde{u} + \int_{\Omega}\sigma u\cdot(\nabla\Tilde{u})\cdot\Phi \\
&\lesssim \norm{u}_{L^{\infty}_x} \norm{\Tilde{u}}_{L^{\infty}_x} \norm{\sigma}_{L^2_x} \norm{\nabla\Phi}_{L^2_x} + \norm{u}_{L^6_x} \norm{\nabla\Tilde{u}}_{L^6_x} \norm{\sigma}_{L^2_x} \norm{\Phi}_{L^6_x} \\
&\lesssim (\kappa\nu)^{-1}\left( \norm{u}_{L^{\infty}_x}^2 \norm{\Tilde{u}}_{L^{\infty}_x}^2 + \norm{u}_{H^1_x}^2 \norm{\Tilde{u}}_{H^2_x}^2 \right) \norm{\sigma}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 \\
&\begin{multlined} \lesssim (\kappa\nu)^{-1}\left( \norm{u}_{H^1_x} \norm{u}_{H^2_x} \norm{\Tilde{u}}_{H^1_x} \norm{\Tilde{u}}_{H^2_x} + \norm{u}_{H^1_x}^2 \norm{\Tilde{u}}_{H^2_x}^2 \right) \norm{\sigma}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 \end{multlined}
\end{align*} \medskip
\noindent where we used Agmon's inequality to bound the $L^{\infty}$ terms. Finally,
\begin{align} \label{B-7.2.3}
\circled{B-7.2.3} = \underbrace{2\Lambda\int_{\Omega} (\Tilde{u}\cdot\Phi) \Re \left[ \overline{\varphi}\Tilde{B}\Tilde{\psi} + \overline{\psi}\Tilde{B}\varphi + \overline{\psi} (B-\Tilde{B}) \psi \right]}_{\circled{B-7.2.3.1}+\circled{B-7.2.3.2}+\circled{B-7.2.3.3}}
\end{align}\medskip
These are, in turn, estimated as follows. \medskip
\begin{align*}
\circled{B-7.2.3.1} &\lesssim \Lambda\norm{\Tilde{u}}_{L^6_x} \norm{\Phi}_{L^6_x} \norm{\varphi}_{L^6_x} \norm{\Tilde{B}\Tilde{\psi}}_{L^2_x} \\
&\lesssim (\kappa\nu)^{-1}\Lambda^2\norm{\Tilde{u}}_{H^1_x}^2 \norm{\Tilde{B}\Tilde{\psi}}_{L^2_x}^2 \norm{\nabla\varphi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
The middle term requires to be broken down further, much like $\circled{B-5}$.
\begin{align*}
\circled{B-7.2.3.2} &= \underbrace{\Lambda\Re\int_{\Omega} (\Tilde{u}\cdot\Phi) \overline{\psi} \left[ -\frac{1}{2}\Delta\varphi + \frac{1}{2}\abs{\Tilde{u}}^2\varphi + i\Tilde{u}\cdot\nabla\varphi + \mu \abs{\Tilde{\psi}}^2 \varphi \right]}_{\circled{B-7.2.3.2.1}+\circled{B-7.2.3.2.2}+\circled{B-7.2.3.2.3}+\circled{B-7.2.3.2.4}}
\end{align*}\medskip
Just as in \circled{B-5}, we extract a dissipative factor, and make use of interpolation. \medskip
\begin{align*}
\circled{B-7.2.3.2.1} &\lesssim \Lambda\norm{\Tilde{u}}_{L^6_x} \norm{\Phi}_{L^3_x} \norm{\psi}_{L^{\infty}_x} \norm{\Delta\varphi}_{L^2_x} \\
&\lesssim \kappa^{-1}\Lambda \norm{\Tilde{u}}_{H^1_x}^2 \norm{\psi}_{H^2_x}^2 \norm{\Phi}_{L^2_x} \norm{\Phi}_{L^6_x} + \kappa\Lambda \norm{D^2\varphi}_{L^2_x}^2 \\
&\lesssim \kappa^{-3}\nu^{-1}\Lambda^2 \norm{\Tilde{u}}_{H^1_x}^4 \norm{\psi}_{H^2_x}^4 \norm{\Phi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 + \kappa\Lambda \norm{D^2\varphi}_{L^2_x}^2
\end{align*
\begin{align*}
\circled{B-7.2.3.2.2} &\lesssim \Lambda\norm{\Tilde{u}}_{L^6_x}^3 \norm{\psi}_{L^{\infty}_x} \norm{\varphi}_{L^6_x} \norm{\Phi}_{L^6_x} \\
&\lesssim (\kappa\nu)^{-1}\Lambda^2 \norm{\Tilde{u}}_{H^1_x}^6 \norm{\psi}_{H^2_x}^2 \norm{\nabla\varphi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*
\begin{align*}
\circled{B-7.2.3.2.3} &\lesssim \Lambda\norm{\Tilde{u}}_{L^6_x}^2 \norm{\psi}_{L^{\infty}_x} \norm{\nabla\varphi}_{L^6_x} \norm{\Phi}_{L^2_x} \\
&\lesssim \kappa^{-1}\Lambda \norm{\Tilde{u}}_{H^1_x}^4 \norm{\psi}_{H^2_x}^2 \norm{\Phi}_{L^2_x}^2 + \kappa\Lambda\norm{D^2\varphi}_{L^2_x}^2
\end{align*
\begin{align*}
\circled{B-7.2.3.2.4} &\lesssim \Lambda \norm{\Tilde{u}}_{L^6_x} \norm{\psi}_{L^6_x} \norm{\Tilde{\psi}}_{L^{\infty}_x}^2 \norm{\Phi}_{L^2_x} \norm{\varphi}_{L^6_x} \\
&\lesssim \Lambda \norm{\Tilde{u}}_{H^1_x}^2 \norm{\psi}_{H^1_x}^2 \norm{\Phi}_{L^2_x}^2 + \Lambda \norm{\Tilde{\psi}}_{H^2_x}^4 \norm{\nabla\varphi}_{L^2_x}^2
\end{align*}\medskip
The final term in \eqref{B-7.2.3} is dealt with in a manner mirroring that of $\circled{B-6}$.
\begin{align*}
\circled{B-7.2.3.3} &= 2\Lambda\Re\int_{\Omega} \Tilde{u}\cdot\Phi \overline{\psi} \left[ \frac{1}{2}\Phi\cdot(u+\Tilde{u}) \psi + i\Phi\cdot\nabla\psi + \mu\left( 2\Re(\Tilde{\psi}\overline{\varphi}) + \abs{\varphi}^2 \right)\psi \right] \\
&\begin{multlined} \lesssim \Lambda \norm{\Tilde{u}}_{L^6_x} \left[ \norm{u+\Tilde{u}}_{L^6_x} \norm{\psi}_{L^{\infty}_x}^2 + \norm{\psi}_{L^{\infty}_x} \norm{\nabla\psi}_{L^6_x} \right] \norm{\Phi}_{L^3_x}^2 \\ + \Lambda \norm{\Tilde{u}}_{L^6_x} \norm{\psi}_{L^{\infty}_x}^2 \left[ \mu \norm{\Tilde{\psi}}_{L^6_x} + \mu\norm{\varphi}_{L^6_x} \right] \norm{\varphi}_{L^6_x} \norm{\Phi}_{L^2_x} \end{multlined} \\
&\begin{multlined} \lesssim (\kappa\nu)^{-1}\Lambda^2 \norm{\Tilde{u}}_{H^1_x}^2 \norm{\psi}_{H^2_x}^4 \left[ 1+\norm{u+\Tilde{u}}_{H^1_x}^2 + \kappa\nu \right] \norm{\Phi}_{L^2_x}^2 \\ + \mu^2\left[ \norm{\Tilde{\psi}}_{H^1_x}^2 + \norm{\psi}_{H^1_x}^2 \right] \norm{\nabla\varphi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{multlined}
\end{align*}\medskip
\end{itemize}
\item Using \eqref{additional time regularity of density}, and the vector identity $\nabla\frac{\abs{\Phi}^2}{2} = \Phi\cdot\nabla\Phi - (\nabla\times\Phi)\times\Phi$,
\begin{align*}
\circled{B-8} &\lesssim \norm{\rho}_{L^{\infty}_x} \norm{u}_{L^6_x} \norm{\Phi}_{L^3_x} \norm{\nabla\Phi}_{L^2_x} \\
&\lesssim (\kappa\nu)^{-1}\norm{\rho}_{L^{\infty}_x}^2 \norm{u}_{H^1_x}^2 \norm{\Phi}_{L^3_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2 \\
&\lesssim (\kappa\nu)^{-3}\norm{\rho}_{L^{\infty}_x}^4 \norm{u}_{H^1_x}^4 \norm{\Phi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
\item
\begin{align*}
\circled{B-9} &\lesssim \norm{\psi}_{L^{\infty}_x} \norm{B\psi}_{L^3_x} \norm{\Phi}_{L^2_x} \norm{\Phi}_{L^6_x} \\
&\lesssim (\kappa\nu)^{-1}\norm{\psi}_{H^2_x}^2 \norm{B\psi}_{L^2_x} \norm{B\psi}_{H^1_x} \norm{\Phi}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{align*}\medskip
\end{enumerate}
Returning to \eqref{phi equation initial}, and using the ensuing estimates, we arrive at:
\begin{equation} \label{phi equation final}
\varepsilon\frac{d}{dt}\norm{\Phi}_{L^2_x}^2 + \frac{\nu}{2} \norm{\nabla\Phi}_{L^2_x}^2 \le h_2(t) \left[ \norm{\nabla\varphi}_{L^2_x}^2 + \norm{\Phi}_{L^2_x}^2 + \norm{\sigma}_{L^2_x}^2 \right] + \kappa\nu \norm{\nabla\Phi}_{L^2_x}^2 + \kappa\Lambda \norm{D^2\varphi}_{L^2_x}^2
\end{equation}\medskip
\noindent where $h_2 \in L^1_{[0,T]}$. Also, note that we have replaced the density in the first term of the LHS by its minimum value. \medskip
\subsection{The continuity equation} \label{weak uniqueness for density}
Finally, in the case of the continuity equation, we take the difference of \eqref{continuity} written for each of the solutions, multiply by $\rho-\Tilde{\rho}$ and integrate over $\Omega$. \medskip
\begin{equation} \label{sigma equation initial}
\frac{d}{dt} \frac{1}{2} \norm{\sigma}_{L^2_x}^2 = \underbrace{\int_{\Omega} \Tilde{\rho}\Phi\cdot\nabla\sigma}_{\circled{C-1}} - 2\Lambda\Re\underbrace{\int_{\Omega} \sigma\overline{\varphi}B\psi + \sigma\overline{\Tilde{\psi}}(B-\Tilde{B})\Tilde{\psi}}_{\circled{C-2} + \circled{C-3}}
\end{equation}\medskip
\begin{rem}
The above calculations can be rigorously justified by considering the difference between the equations for $\rho^N$ (the approximate densities in Section 4.3 of \cite{Jayanti2022LocalSuperfluidity}) and $\Tilde{\rho}$. Passing to the limit $N\rightarrow\infty$ leaves us with the desired equation for $\sigma = \rho-\Tilde{\rho}$, since we know that $\rho\in C(0,T;L^2_x)$.
\end{rem}\medskip
\begin{enumerate}[(i)]
\item Once again, recalling that we have assumed in Theorem \ref{local uniqueness} that $\Tilde{\rho}\in L^2_{[0,T]}W^{1,3}_x$. Thus,
\begin{equation*}
\circled{C-1} = -\int_{\Omega}\nabla\Tilde{\rho}\cdot\Phi\sigma \lesssim \norm{\nabla\Tilde{\rho}}_{L^3_x} \norm{\Phi}_{L^6_x} \norm{\sigma}_{L^2_x} \lesssim (\kappa\nu)^{-1}\norm{\nabla\Tilde{\rho}}_{L^3_x}^2 \norm{\sigma}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{equation*}\medskip
\noindent Observe that $\sigma$ should always be considered in the $L^2_x$ norm while using H\"older's, since there is no dissipation term for the density. \medskip
\item For the second term, apart from the above, we also interpolate the $L^3_x$ norm between the $L^2_x$ and $L^6_x$ norms.
\begin{align*}
\circled{C-2} \lesssim \Lambda\norm{\sigma}_{L^2_x} \norm{\varphi}_{L^6_x} \norm{B\psi}_{L^3_x} \lesssim \Lambda^2 \norm{B\psi}_{L^2_x} \norm{\nabla B\psi}_{L^2_x} \norm{\sigma}_{L^2_x}^2 + \norm{\nabla\varphi}_{L^2_x}^2
\end{align*}\medskip
\item Similar analysis on the third term yields
\begin{align*}
\circled{C-3} &\lesssim \Lambda\norm{\sigma}_{L^2_x} \norm{\Tilde{\psi}}_{L^{\infty}_x} \norm{(B-\Tilde{B})\Tilde{\psi}}_{L^2_x} \\
&\begin{multlined} \lesssim \Lambda\norm{\sigma}_{L^2_x} \norm{\Tilde{\psi}}_{H^2_x} \left[ \norm{\Tilde{\psi}}_{L^{\infty}_x} \norm{u+\Tilde{u}}_{L^6_x} \norm{\Phi}_{L^3_x} + \norm{\nabla\Tilde{\psi}}_{L^6_x} \norm{\Phi}_{L^3_x} \right. \\ \left. + \norm{\Tilde{\psi}}_{L^6_x}^2 \norm{\varphi}_{L^6_x} + \norm{\Tilde{\psi}}_{L^{\infty}_x} \norm{\varphi}_{L^3_x} \norm{\varphi}_{L^6_x} \right] \end{multlined} \\
&\begin{multlined} \lesssim \Lambda\left[ \norm{\Tilde{\psi}}_{H^2_x}^4 + \norm{\Tilde{\psi}}_{H^1_x}^2\norm{\Tilde{\psi}}_{H^2_x}^2 \right] \norm{\sigma}_{L^2_x}^2 + (\kappa\nu)^{-1}\Lambda\left[ 1 + \norm{u+\Tilde{u}}_{H^1_x}^4 \right]\norm{\Phi}_{L^2_x}^2 \\
+ \Lambda\left[ \norm{\Tilde{\psi}}_{H^1_x}^2 + \norm{\psi}_{H^1_x}^2 \right] \norm{\nabla\varphi}_{L^2_x}^2 + \kappa\nu \norm{\nabla\Phi}_{L^2_x}^2 \end{multlined}
\end{align*}\medskip
In going to the last step, we tackle the $\norm{\varphi}_{L^3_x}^2$ term using interpolation and the Poincar\'e inequality as follows:
\begin{equation*}
\norm{\varphi}_{L^3_x}^2 \lesssim \norm{\varphi}_{L^2_x}\norm{\varphi}_{L^6_x} \lesssim \norm{\varphi}_{L^2_x} \norm{\varphi}_{H^1_x}
\lesssim \norm{\nabla\varphi}_{L^2_x}^2 \lesssim \norm{\nabla\psi}_{L^2_x}^2 + \norm{\nabla\Tilde{\psi}}_{L^2_x}^2
\end{equation*}\bigskip
\end{enumerate}
Using the above estimates, we can simplify \eqref{sigma equation initial} to read: \medskip
\begin{equation} \label{sigma equation final}
\frac{d}{dt}\norm{\sigma}_{L^2_x}^2 \le h_3(t) \norm{\sigma}_{L^2_x}^2 + \kappa\nu\norm{\nabla\Phi}_{L^2_x}^2
\end{equation}\medskip
\noindent where $h_3 \in L^1_{[0,T]}$. \bigskip
\bigskip
\bigskip
We now proceed to add \eqref{varphi equation final}, \eqref{phi equation final} and \eqref{sigma equation final}. By choosing $\kappa$ sufficiently small, we can ensure all the dissipation terms are absorbed by the LHS. What remains is:
\begin{equation}
\frac{d}{dt} \left[ \norm{\varphi}_{L^2_x}^2 + \norm{\Phi}_{L^2_x}^2 + \norm{\sigma}_{L^2_x}^2 \right] \lesssim \left( h_1 + h_2 + h_3 \right)(t) \left[ \norm{\varphi}_{L^2_x}^2 + \norm{\Phi}_{L^2_x}^2 + \norm{\sigma}_{L^2_x}^2 \right]
\end{equation}\medskip
Using Gr\"onwall's inequality, and the fact that $\norm{\varphi}_{L^2_x} = \norm{\Phi}_{L^2_x} = \norm{\sigma}_{L^2_x} = 0$ at $t=0$ completes the proof of Theorem \ref{local uniqueness}.
\qed
\addtocontents{toc}{\protect\setcounter{tocdepth}{0}}
\section*{Acknowledgments}
Both the authors are grateful to the anonymous referees for their comments and suggestions which significantly helped to improve the original manuscript. P.C.J. was partially supported by the Ann Wylie Fellowship at UMD. Both P.C.J. and K.T. gratefully acknowledge the support of the National Science Foundation under the awards DMS-1614964 and DMS-2008568. \bigskip
\addtocontents{toc}{\protect\setcounter{tocdepth}{2}}
\bibliographystyle{alpha}
|
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{T}{his} document is a template for \LaTeX. If you are
reading a paper or PDF version of this document, please download the
electronic file, trans\_jour.tex, from the IEEE Web site at \underline
{http://www.ieee.org/authortools/trans\_jour.tex} so you can use it to prepare your manuscript. If
you would prefer to use LaTeX, download IEEE's LaTeX style and sample files
from the same Web page. You can also explore using the Overleaf editor at
\underline
{https://www.overleaf.com/blog/278-how-to-use-overleaf-with-}\discretionary{}{}{}\underline
{ieee-collabratec-your-quick-guide-to-getting-started\#.}\discretionary{}{}{}\underline{xsVp6tpPkrKM9}
If your paper is intended for a conference, please contact your conference
editor concerning acceptable word processor formats for your particular
conference.
IEEE will do the final formatting of your paper. If your paper is intended
for a conference, please observe the conference page limits.
\subsection{Abbreviations and Acronyms}
Define abbreviations and acronyms the first time they are used in the text,
even after they have already been defined in the abstract. Abbreviations
such as IEEE, SI, ac, and dc do not have to be defined. Abbreviations that
incorporate periods should not have spaces: write ``C.N.R.S.,'' not ``C. N.
R. S.'' Do not use abbreviations in the title unless they are unavoidable
(for example, ``IEEE'' in the title of this article).
\subsection{Other Recommendations}
Use one space after periods and colons. Hyphenate complex modifiers:
``zero-field-cooled magnetization.'' Avoid dangling participles, such as,
``Using \eqref{eq}, the potential was calculated.'' [It is not clear who or what
used \eqref{eq}.] Write instead, ``The potential was calculated by using \eqref{eq},'' or
``Using \eqref{eq}, we calculated the potential.''
Use a zero before decimal points: ``0.25,'' not ``.25.'' Use
``cm$^{3}$,'' not ``cc.'' Indicate sample dimensions as ``0.1 cm
$\times $ 0.2 cm,'' not ``0.1 $\times $ 0.2 cm$^{2}$.'' The
abbreviation for ``seconds'' is ``s,'' not ``sec.'' Use
``Wb/m$^{2}$'' or ``webers per square meter,'' not
``webers/m$^{2}$.'' When expressing a range of values, write ``7 to
9'' or ``7--9,'' not ``7$\sim $9.''
A parenthetical statement at the end of a sentence is punctuated outside of
the closing parenthesis (like this). (A parenthetical sentence is punctuated
within the parentheses.) In American English, periods and commas are within
quotation marks, like ``this period.'' Other punctuation is ``outside''!
Avoid contractions; for example, write ``do not'' instead of ``don't.'' The
serial comma is preferred: ``A, B, and C'' instead of ``A, B and C.''
If you wish, you may write in the first person singular or plural and use
the active voice (``I observed that $\ldots$'' or ``We observed that $\ldots$''
instead of ``It was observed that $\ldots$''). Remember to check spelling. If
your native language is not English, please get a native English-speaking
colleague to carefully proofread your paper.
Try not to use too many typefaces in the same article. You're writing
scholarly papers, not ransom notes. Also please remember that MathJax
can't handle really weird typefaces.
\subsection{Equations}
Number equations consecutively with equation numbers in parentheses flush
with the right margin, as in \eqref{eq}. To make your equations more
compact, you may use the solidus (~/~), the exp function, or appropriate
exponents. Use parentheses to avoid ambiguities in denominators. Punctuate
equations when they are part of a sentence, as in
\begin{equation}E=mc^2.\label{eq}\end{equation}
Be sure that the symbols in your equation have been defined before the
equation appears or immediately following. Italicize symbols ($T$ might refer
to temperature, but T is the unit tesla). Refer to ``\eqref{eq},'' not ``Eq. \eqref{eq}''
or ``equation \eqref{eq},'' except at the beginning of a sentence: ``Equation \eqref{eq}
is $\ldots$ .''
\subsection{\LaTeX-Specific Advice}
Please use ``soft'' (e.g., \verb|\eqref{Eq}|) cross references instead
of ``hard'' references (e.g., \verb|(1)|). That will make it possible
to combine sections, add equations, or change the order of figures or
citations without having to go through the file line by line.
Please don't use the \verb|{eqnarray}| equation environment. Use
\verb|{align}| or \verb|{IEEEeqnarray}| instead. The \verb|{eqnarray}|
environment leaves unsightly spaces around relation symbols.
Please note that the \verb|{subequations}| environment in {\LaTeX}
will increment the main equation counter even when there are no
equation numbers displayed. If you forget that, you might write an
article in which the equation numbers skip from (17) to (20), causing
the copy editors to wonder if you've discovered a new method of
counting.
{\BibTeX} does not work by magic. It doesn't get the bibliographic
data from thin air but from .bib files. If you use {\BibTeX} to produce a
bibliography you must send the .bib files.
{\LaTeX} can't read your mind. If you assign the same label to a
subsubsection and a table, you might find that Table I has been cross
referenced as Table IV-B3.
{\LaTeX} does not have precognitive abilities. If you put a
\verb|\label| command before the command that updates the counter it's
supposed to be using, the label will pick up the last counter to be
cross referenced instead. In particular, a \verb|\label| command
should not go before the caption of a figure or a table.
Do not use \verb|\nonumber| inside the \verb|{array}| environment. It
will not stop equation numbers inside \verb|{array}| (there won't be
any anyway) and it might stop a wanted equation number in the
surrounding equation.
If you are submitting your paper to a colorized journal, you can use
the following two lines at the start of the article to ensure its
appearance resembles the final copy:
\smallskip\noindent
\begin{small}
\begin{tabular}{l}
\verb+\+\texttt{documentclass[journal,twoside,web]\{ieeecolor\}}\\
\verb+\+\texttt{usepackage\{\textit{Journal\_Name}\}}
\end{tabular}
\end{small}
\section{Units}
Use either SI (MKS) or CGS as primary units. (SI units are strongly
encouraged.) English units may be used as secondary units (in parentheses).
This applies to papers in data storage. For example, write ``15
Gb/cm$^{2}$ (100 Gb/in$^{2})$.'' An exception is when
English units are used as identifiers in trade, such as ``3\textonehalf-in
disk drive.'' Avoid combining SI and CGS units, such as current in amperes
and magnetic field in oersteds. This often leads to confusion because
equations do not balance dimensionally. If you must use mixed units, clearly
state the units for each quantity in an equation.
The SI unit for magnetic field strength $H$ is A/m. However, if you wish to use
units of T, either refer to magnetic flux density $B$ or magnetic field
strength symbolized as $\mu _{0}H$. Use the center dot to separate
compound units, e.g., ``A$\cdot $m$^{2}$.''
\section{Some Common Mistakes}
The word ``data'' is plural, not singular. The subscript for the
permeability of vacuum $\mu _{0}$ is zero, not a lowercase letter
``o.'' The term for residual magnetization is ``remanence''; the adjective
is ``remanent''; do not write ``remnance'' or ``remnant.'' Use the word
``micrometer'' instead of ``micron.'' A graph within a graph is an
``inset,'' not an ``insert.'' The word ``alternatively'' is preferred to the
word ``alternately'' (unless you really mean something that alternates). Use
the word ``whereas'' instead of ``while'' (unless you are referring to
simultaneous events). Do not use the word ``essentially'' to mean
``approximately'' or ``effectively.'' Do not use the word ``issue'' as a
euphemism for ``problem.'' When compositions are not specified, separate
chemical symbols by en-dashes; for example, ``NiMn'' indicates the
intermetallic compound Ni$_{0.5}$Mn$_{0.5}$ whereas
``Ni--Mn'' indicates an alloy of some composition
Ni$_{x}$Mn$_{1-x}$.
\begin{figure}[!t]
\centerline{\includegraphics[width=\columnwidth]{fig1.png}}
\caption{Magnetization as a function of applied field.
It is good practice to explain the significance of the figure in the caption.}
\label{fig1}
\end{figure}
Be aware of the different meanings of the homophones ``affect'' (usually a
verb) and ``effect'' (usually a noun), ``complement'' and ``compliment,''
``discreet'' and ``discrete,'' ``principal'' (e.g., ``principal
investigator'') and ``principle'' (e.g., ``principle of measurement''). Do
not confuse ``imply'' and ``infer.''
Prefixes such as ``non,'' ``sub,'' ``micro,'' ``multi,'' and ``ultra'' are
not independent words; they should be joined to the words they modify,
usually without a hyphen. There is no period after the ``et'' in the Latin
abbreviation ``\emph{et al.}'' (it is also italicized). The abbreviation ``i.e.,'' means
``that is,'' and the abbreviation ``e.g.,'' means ``for example'' (these
abbreviations are not italicized).
A general IEEE styleguide is available at \underline{http://www.ieee.org/authortools}.
\section{Guidelines for Graphics Preparation and Submission}
\label{sec:guidelines}
\subsection{Types of Graphics}
The following list outlines the different types of graphics published in
IEEE journals. They are categorized based on their construction, and use of
color/shades of gray:
\subsubsection{Color/Grayscale figures}
{Figures that are meant to appear in color, or shades of black/gray. Such
figures may include photographs, illustrations, multicolor graphs, and
flowcharts.}
\subsubsection{Line Art figures}
{Figures that are composed of only black lines and shapes. These figures
should have no shades or half-tones of gray, only black and white.}
\subsubsection{Author photos}
{Head and shoulders shots of authors that appear at the end of our papers. }
\subsubsection{Tables}
{Data charts which are typically black and white, but sometimes include
color.}
\begin{table}
\caption{Units for Magnetic Properties}
\label{table}
\setlength{\tabcolsep}{3pt}
\begin{tabular}{|p{25pt}|p{75pt}|p{115pt}|}
\hline
Symbol&
Quantity&
Conversion from Gaussian and \par CGS EMU to SI $^{\mathrm{a}}$ \\
\hline
$\Phi $&
magnetic flux&
1 Mx $\to 10^{-8}$ Wb $= 10^{-8}$ V$\cdot $s \\
$B$&
magnetic flux density, \par magnetic induction&
1 G $\to 10^{-4}$ T $= 10^{-4}$ Wb/m$^{2}$ \\
$H$&
magnetic field strength&
1 Oe $\to 10^{3}/(4\pi )$ A/m \\
$m$&
magnetic moment&
1 erg/G $=$ 1 emu \par $\to 10^{-3}$ A$\cdot $m$^{2} = 10^{-3}$ J/T \\
$M$&
magnetization&
1 erg/(G$\cdot $cm$^{3}) =$ 1 emu/cm$^{3}$ \par $\to 10^{3}$ A/m \\
4$\pi M$&
magnetization&
1 G $\to 10^{3}/(4\pi )$ A/m \\
$\sigma $&
specific magnetization&
1 erg/(G$\cdot $g) $=$ 1 emu/g $\to $ 1 A$\cdot $m$^{2}$/kg \\
$j$&
magnetic dipole \par moment&
1 erg/G $=$ 1 emu \par $\to 4\pi \times 10^{-10}$ Wb$\cdot $m \\
$J$&
magnetic polarization&
1 erg/(G$\cdot $cm$^{3}) =$ 1 emu/cm$^{3}$ \par $\to 4\pi \times 10^{-4}$ T \\
$\chi , \kappa $&
susceptibility&
1 $\to 4\pi $ \\
$\chi_{\rho }$&
mass susceptibility&
1 cm$^{3}$/g $\to 4\pi \times 10^{-3}$ m$^{3}$/kg \\
$\mu $&
permeability&
1 $\to 4\pi \times 10^{-7}$ H/m \par $= 4\pi \times 10^{-7}$ Wb/(A$\cdot $m) \\
$\mu_{r}$&
relative permeability&
$\mu \to \mu_{r}$ \\
$w, W$&
energy density&
1 erg/cm$^{3} \to 10^{-1}$ J/m$^{3}$ \\
$N, D$&
demagnetizing factor&
1 $\to 1/(4\pi )$ \\
\hline
\multicolumn{3}{p{251pt}}{Vertical lines are optional in tables. Statements that serve as captions for
the entire table do not need footnote letters. }\\
\multicolumn{3}{p{251pt}}{$^{\mathrm{a}}$Gaussian units are the same as cg emu for magnetostatics; Mx
$=$ maxwell, G $=$ gauss, Oe $=$ oersted; Wb $=$ weber, V $=$ volt, s $=$
second, T $=$ tesla, m $=$ meter, A $=$ ampere, J $=$ joule, kg $=$
kilogram, H $=$ henry.}
\end{tabular}
\label{tab1}
\end{table}
\subsection{Multipart figures}
Figures compiled of more than one sub-figure presented side-by-side, or
stacked. If a multipart figure is made up of multiple figure
types (one part is lineart, and another is grayscale or color) the figure
should meet the stricter guidelines.
\subsection{File Formats For Graphics}\label{formats}
Format and save your graphics using a suitable graphics processing program
that will allow you to create the images as PostScript (PS), Encapsulated
PostScript (.EPS), Tagged Image File Format (.TIFF), Portable Document
Format (.PDF), Portable Network Graphics (.PNG), or Metapost (.MPS), sizes them, and adjusts
the resolution settings. When
submitting your final paper, your graphics should all be submitted
individually in one of these formats along with the manuscript.
\subsection{Sizing of Graphics}
Most charts, graphs, and tables are one column wide (3.5 inches/88
millimeters/21 picas) or page wide (7.16 inches/181 millimeters/43
picas). The maximum depth a graphic can be is 8.5 inches (216 millimeters/54
picas). When choosing the depth of a graphic, please allow space for a
caption. Figures can be sized between column and page widths if the author
chooses, however it is recommended that figures are not sized less than
column width unless when necessary.
There is currently one publication with column measurements that do not
coincide with those listed above. Proceedings of the IEEE has a column
measurement of 3.25 inches (82.5 millimeters/19.5 picas).
The final printed size of author photographs is exactly
1 inch wide by 1.25 inches tall (25.4 millimeters$\,\times\,$31.75 millimeters/6
picas$\,\times\,$7.5 picas). Author photos printed in editorials measure 1.59 inches
wide by 2 inches tall (40 millimeters$\,\times\,$50 millimeters/9.5 picas$\,\times\,$12
picas).
\subsection{Resolution }
The proper resolution of your figures will depend on the type of figure it
is as defined in the ``Types of Figures'' section. Author photographs,
color, and grayscale figures should be at least 300dpi. Line art, including
tables should be a minimum of 600dpi.
\subsection{Vector Art}
In order to preserve the figures' integrity across multiple computer
platforms, we accept files in the following formats: .EPS/.PDF/.PS. All
fonts must be embedded or text converted to outlines in order to achieve the
best-quality results.
\subsection{Color Space}
The term color space refers to the entire sum of colors that can be
represented within the said medium. For our purposes, the three main color
spaces are Grayscale, RGB (red/green/blue) and CMYK
(cyan/magenta/yellow/black). RGB is generally used with on-screen graphics,
whereas CMYK is used for printing purposes.
All color figures should be generated in RGB or CMYK color space. Grayscale
images should be submitted in Grayscale color space. Line art may be
provided in grayscale OR bitmap colorspace. Note that ``bitmap colorspace''
and ``bitmap file format'' are not the same thing. When bitmap color space
is selected, .TIF/.TIFF/.PNG are the recommended file formats.
\subsection{Accepted Fonts Within Figures}
When preparing your graphics IEEE suggests that you use of one of the
following Open Type fonts: Times New Roman, Helvetica, Arial, Cambria, and
Symbol. If you are supplying EPS, PS, or PDF files all fonts must be
embedded. Some fonts may only be native to your operating system; without
the fonts embedded, parts of the graphic may be distorted or missing.
A safe option when finalizing your figures is to strip out the fonts before
you save the files, creating ``outline'' type. This converts fonts to
artwork what will appear uniformly on any screen.
\subsection{Using Labels Within Figures}
\subsubsection{Figure Axis labels }
Figure axis labels are often a source of confusion. Use words rather than
symbols. As an example, write the quantity ``Magnetization,'' or
``Magnetization M,'' not just ``M.'' Put units in parentheses. Do not label
axes only with units. As in Fig. 1, for example, write ``Magnetization
(A/m)'' or ``Magnetization (A$\cdot$m$^{-1}$),'' not just ``A/m.'' Do not label axes with a ratio of quantities and
units. For example, write ``Temperature (K),'' not ``Temperature/K.''
Multipliers can be especially confusing. Write ``Magnetization (kA/m)'' or
``Magnetization (10$^{3}$ A/m).'' Do not write ``Magnetization
(A/m)$\,\times\,$1000'' because the reader would not know whether the top
axis label in Fig. 1 meant 16000 A/m or 0.016 A/m. Figure labels should be
legible, approximately 8 to 10 point type.
\subsubsection{Subfigure Labels in Multipart Figures and Tables}
Multipart figures should be combined and labeled before final submission.
Labels should appear centered below each subfigure in 8 point Times New
Roman font in the format of (a) (b) (c).
\subsection{File Naming}
Figures (line artwork or photographs) should be named starting with the
first 5 letters of the author's last name. The next characters in the
filename should be the number that represents the sequential
location of this image in your article. For example, in author
``Anderson's'' paper, the first three figures would be named ander1.tif,
ander2.tif, and ander3.ps.
Tables should contain only the body of the table (not the caption) and
should be named similarly to figures, except that `.t' is inserted
in-between the author's name and the table number. For example, author
Anderson's first three tables would be named ander.t1.tif, ander.t2.ps,
ander.t3.eps.
Author photographs should be named using the first five characters of the
pictured author's last name. For example, four author photographs for a
paper may be named: oppen.ps, moshc.tif, chen.eps, and duran.pdf.
If two authors or more have the same last name, their first initial(s) can
be substituted for the fifth, fourth, third$\ldots$ letters of their surname
until the degree where there is differentiation. For example, two authors
Michael and Monica Oppenheimer's photos would be named oppmi.tif, and
oppmo.eps.
\subsection{Referencing a Figure or Table Within Your Paper}
When referencing your figures and tables within your paper, use the
abbreviation ``Fig.'' even at the beginning of a sentence. Do not abbreviate
``Table.'' Tables should be numbered with Roman Numerals.
\subsection{Checking Your Figures: The IEEE Graphics Analyzer}
The IEEE Graphics Analyzer enables authors to pre-screen their graphics for
compliance with IEEE Transactions and Journals standards before submission.
The online tool, located at
\underline{http://graphicsqc.ieee.org/}, allows authors to
upload their graphics in order to check that each file is the correct file
format, resolution, size and colorspace; that no fonts are missing or
corrupt; that figures are not compiled in layers or have transparency, and
that they are named according to the IEEE Transactions and Journals naming
convention. At the end of this automated process, authors are provided with
a detailed report on each graphic within the web applet, as well as by
email.
For more information on using the Graphics Analyzer or any other graphics
related topic, contact the IEEE Graphics Help Desk by e-mail at
<EMAIL>.
\subsection{Submitting Your Graphics}
Because IEEE will do the final formatting of your paper,
you do not need to position figures and tables at the top and bottom of each
column. In fact, all figures, figure captions, and tables can be placed at
the end of your paper. In addition to, or even in lieu of submitting figures
within your final manuscript, figures should be submitted individually,
separate from the manuscript in one of the file formats listed above in
Section \ref{formats}. Place figure captions below the figures; place table titles
above the tables. Please do not include captions as part of the figures, or
put them in ``text boxes'' linked to the figures. Also, do not place borders
around the outside of your figures.
\subsection{Color Processing/Printing in IEEE Journals}
All IEEE Transactions, Journals, and Letters allow an author to publish
color figures on IEEE Xplore\textregistered\ at no charge, and automatically
convert them to grayscale for print versions. In most journals, figures and
tables may alternatively be printed in color if an author chooses to do so.
Please note that this service comes at an extra expense to the author. If
you intend to have print color graphics, include a note with your final
paper indicating which figures or tables you would like to be handled that
way, and stating that you are willing to pay the additional fee.
\section{Conclusion}
A conclusion section is not required. Although a conclusion may review the
main points of the paper, do not replicate the abstract as the conclusion. A
conclusion might elaborate on the importance of the work or suggest
applications and extensions.
\appendices
Appendixes, if needed, appear before the acknowledgment.
\section*{Acknowledgment}
The preferred spelling of the word ``acknowledgment'' in American English is
without an ``e'' after the ``g.'' Use the singular heading even if you have
many acknowledgments. Avoid expressions such as ``One of us (S.B.A.) would
like to thank $\ldots$ .'' Instead, write ``F. A. Author thanks $\ldots$ .'' In most
cases, sponsor and financial support acknowledgments are placed in the
unnumbered footnote on the first page, not here.
\section*{References and Footnotes}
\subsection{References}
References need not be cited in text. When they are, they appear on the
line, in square brackets, inside the punctuation. Multiple references are
each numbered with separate brackets. When citing a section in a book,
please give the relevant page numbers. In text, refer simply to the
reference number. Do not use ``Ref.'' or ``reference'' except at the
beginning of a sentence: ``Reference \cite{b3} shows $\ldots$ .'' Please do not use
automatic endnotes in \emph{Word}, rather, type the reference list at the end of the
paper using the ``References'' style.
Reference numbers are set flush left and form a column of their own, hanging
out beyond the body of the reference. The reference numbers are on the line,
enclosed in square brackets. In all references, the given name of the author
or editor is abbreviated to the initial only and precedes the last name. Use
them all; use \emph{et al.} only if names are not given. Use commas around Jr.,
Sr., and III in names. Abbreviate conference titles. When citing IEEE
transactions, provide the issue number, page range, volume number, year,
and/or month if available. When referencing a patent, provide the day and
the month of issue, or application. References may not include all
information; please obtain and include relevant information. Do not combine
references. There must be only one reference with each number. If there is a
URL included with the print reference, it can be included at the end of the
reference.
Other than books, capitalize only the first word in a paper title, except
for proper nouns and element symbols. For papers published in translation
journals, please give the English citation first, followed by the original
foreign-language citation See the end of this document for formats and
examples of common references. For a complete discussion of references and
their formats, see the IEEE style manual at
\underline{http://www.ieee.org/authortools}.
\subsection{Footnotes}
Number footnotes separately in superscript numbers.\footnote{It is recommended that footnotes be avoided (except for
the unnumbered footnote with the receipt date on the first page). Instead,
try to integrate the footnote information into the text.} Place the actual
footnote at the bottom of the column in which it is cited; do not put
footnotes in the reference list (endnotes). Use letters for table footnotes
(see Table \ref{table}).
\section{Submitting Your Paper for Review}
\subsection{Final Stage}
When you submit your final version (after your paper has been accepted),
print it in two-column format, including figures and tables. You must also
send your final manuscript on a disk, via e-mail, or through a Web
manuscript submission system as directed by the society contact. You may use
\emph{Zip} for large files, or compress files using \emph{Compress, Pkzip, Stuffit,} or \emph{Gzip.}
Also, send a sheet of paper or PDF with complete contact information for all
authors. Include full mailing addresses, telephone numbers, fax numbers, and
e-mail addresses. This information will be used to send each author a
complimentary copy of the journal in which the paper appears. In addition,
designate one author as the ``corresponding author.'' This is the author to
whom proofs of the paper will be sent. Proofs are sent to the corresponding
author only.
\subsection{Review Stage Using ScholarOne\textregistered\ Manuscripts}
Contributions to the Transactions, Journals, and Letters may be submitted
electronically on IEEE's on-line manuscript submission and peer-review
system, ScholarOne\textregistered\ Manuscripts. You can get a listing of the
publications that participate in ScholarOne at
\underline{http://www.ieee.org/publications\_standards/publications/}\discretionary{}{}{}\underline{authors/authors\_submission.html}.
First check if you have an existing account. If there is none, please create
a new account. After logging in, go to your Author Center and click ``Submit
First Draft of a New Manuscript.''
Along with other information, you will be asked to select the subject from a
pull-down list. Depending on the journal, there are various steps to the
submission process; you must complete all steps for a complete submission.
At the end of each step you must click ``Save and Continue''; just uploading
the paper is not sufficient. After the last step, you should see a
confirmation that the submission is complete. You should also receive an
e-mail confirmation. For inquiries regarding the submission of your paper on
ScholarOne Manuscripts, please contact <EMAIL> or call +1 732
465 5861.
ScholarOne Manuscripts will accept files for review in various formats.
Please check the guidelines of the specific journal for which you plan to
submit.
You will be asked to file an electronic copyright form immediately upon
completing the submission process (authors are responsible for obtaining any
security clearances). Failure to submit the electronic copyright could
result in publishing delays later. You will also have the opportunity to
designate your article as ``open access'' if you agree to pay the IEEE open
access fee.
\subsection{Final Stage Using ScholarOne Manuscripts}
Upon acceptance, you will receive an email with specific instructions
regarding the submission of your final files. To avoid any delays in
publication, please be sure to follow these instructions. Most journals
require that final submissions be uploaded through ScholarOne Manuscripts,
although some may still accept final submissions via email. Final
submissions should include source files of your accepted manuscript, high
quality graphic files, and a formatted pdf file. If you have any questions
regarding the final submission process, please contact the administrative
contact for the journal.
In addition to this, upload a file with complete contact information for all
authors. Include full mailing addresses, telephone numbers, fax numbers, and
e-mail addresses. Designate the author who submitted the manuscript on
ScholarOne Manuscripts as the ``corresponding author.'' This is the only
author to whom proofs of the paper will be sent.
\subsection{Copyright Form}
Authors must submit an electronic IEEE Copyright Form (eCF) upon submitting
their final manuscript files. You can access the eCF system through your
manuscript submission system or through the Author Gateway. You are
responsible for obtaining any necessary approvals and/or security
clearances. For additional information on intellectual property rights,
visit the IEEE Intellectual Property Rights department web page at
\underline{http://www.ieee.org/publications\_standards/publications/rights/}\discretionary{}{}{}\underline{index.html}.
\section{IEEE Publishing Policy}
The general IEEE policy requires that authors should only submit original
work that has neither appeared elsewhere for publication, nor is under
review for another refereed publication. The submitting author must disclose
all prior publication(s) and current submissions when submitting a
manuscript. Do not publish ``preliminary'' data or results. The submitting
author is responsible for obtaining agreement of all coauthors and any
consent required from employers or sponsors before submitting an article.
The IEEE Transactions and Journals Department strongly discourages courtesy
authorship; it is the obligation of the authors to cite only relevant prior
work.
The IEEE Transactions and Journals Department does not publish conference
records or proceedings, but can publish articles related to conferences that
have undergone rigorous peer review. Minimally, two reviews are required for
every article submitted for peer review.
\section{Publication Principles}
The two types of contents of that are published are; 1) peer-reviewed and 2)
archival. The Transactions and Journals Department publishes scholarly
articles of archival value as well as tutorial expositions and critical
reviews of classical subjects and topics of current interest.
Authors should consider the following points:
\begin{enumerate}
\item Technical papers submitted for publication must advance the state of knowledge and must cite relevant prior work.
\item The length of a submitted paper should be commensurate with the importance, or appropriate to the complexity, of the work. For example, an obvious extension of previously published work might not be appropriate for publication or might be adequately treated in just a few pages.
\item Authors must convince both peer reviewers and the editors of the scientific and technical merit of a paper; the standards of proof are higher when extraordinary or unexpected results are reported.
\item Because replication is required for scientific progress, papers submitted for publication must provide sufficient information to allow readers to perform similar experiments or calculations and
use the reported results. Although not everything need be disclosed, a paper
must contain new, useable, and fully described information. For example, a
specimen's chemical composition need not be reported if the main purpose of
a paper is to introduce a new measurement technique. Authors should expect
to be challenged by reviewers if the results are not supported by adequate
data and critical details.
\item Papers that describe ongoing work or announce the latest technical achievement, which are suitable for presentation at a professional conference, may not be appropriate for publication.
\end{enumerate}
\section{Reference Examples}
\begin{itemize}
\item \emph{Basic format for books:}\\
J. K. Author, ``Title of chapter in the book,'' in \emph{Title of His Published Book, x}th ed. City of Publisher, (only U.S. State), Country: Abbrev. of Publisher, year, ch. $x$, sec. $x$, pp. \emph{xxx--xxx.}\\
See \cite{b1,b2}.
\item \emph{Basic format for periodicals:}\\
J. K. Author, ``Name of paper,'' \emph{Abbrev. Title of Periodical}, vol. \emph{x, no}. $x, $pp\emph{. xxx--xxx, }Abbrev. Month, year, DOI. 10.1109.\emph{XXX}.123456.\\
See \cite{b3}--\cite{b5}.
\item \emph{Basic format for reports:}\\
J. K. Author, ``Title of report,'' Abbrev. Name of Co., City of Co., Abbrev. State, Country, Rep. \emph{xxx}, year.\\
See \cite{b6,b7}.
\item \emph{Basic format for handbooks:}\\
\emph{Name of Manual/Handbook, x} ed., Abbrev. Name of Co., City of Co., Abbrev. State, Country, year, pp. \emph{xxx--xxx.}\\
See \cite{b8,b9}.
\item \emph{Basic format for books (when available online):}\\
J. K. Author, ``Title of chapter in the book,'' in \emph{Title of
Published Book}, $x$th ed. City of Publisher, State, Country: Abbrev.
of Publisher, year, ch. $x$, sec. $x$, pp. \emph{xxx--xxx}. [Online].
Available: \underline{http://www.web.com}\\
See \cite{b10}--\cite{b13}.
\item \emph{Basic format for journals (when available online):}\\
J. K. Author, ``Name of paper,'' \emph{Abbrev. Title of Periodical}, vol. $x$, no. $x$, pp. \emph{xxx--xxx}, Abbrev. Month, year. Accessed on: Month, Day, year, DOI: 10.1109.\emph{XXX}.123456, [Online].\\
See \cite{b14}--\cite{b16}.
\item \emph{Basic format for papers presented at conferences (when available online): }\\
J.K. Author. (year, month). Title. presented at abbrev. conference title. [Type of Medium]. Available: site/path/file\\
See \cite{b17}.
\item \emph{Basic format for reports and handbooks (when available online):}\\
J. K. Author. ``Title of report,'' Company. City, State, Country. Rep. no., (optional: vol./issue), Date. [Online] Available: site/path/file\\
See \cite{b18,b19}.
\item \emph{Basic format for computer programs and electronic documents (when available online): }\\
Legislative body. Number of Congress, Session. (year, month day). \emph{Number of bill or resolution}, \emph{Title}. [Type of medium]. Available: site/path/file\\
\textbf{\emph{NOTE: }ISO recommends that capitalization follow the accepted practice for the language or script in which the information is given.}\\
See \cite{b20}.
\item \emph{Basic format for patents (when available online):}\\
Name of the invention, by inventor's name. (year, month day). Patent Number [Type of medium]. Available: site/path/file\\
See \cite{b21}.
\item \emph{Basic format}\emph{for conference proceedings (published):}\\
J. K. Author, ``Title of paper,'' in \emph{Abbreviated Name of Conf.}, City of Conf., Abbrev. State (if given), Country, year, pp. \emph{xxxxxx.}\\
See \cite{b22}.
\item \emph{Example for papers presented at conferences (unpublished):}\\
See \cite{b23}.
\item \emph{Basic format for patents}$:$\\
J. K. Author, ``Title of patent,'' U.S. Patent \emph{x xxx xxx}, Abbrev. Month, day, year.\\
See \cite{b24}.
\item \emph{Basic format for theses (M.S.) and dissertations (Ph.D.):}
\begin{enumerate}
\item J. K. Author, ``Title of thesis,'' M.S. thesis, Abbrev. Dept., Abbrev. Univ., City of Univ., Abbrev. State, year.
\item J. K. Author, ``Title of dissertation,'' Ph.D. dissertation, Abbrev. Dept., Abbrev. Univ., City of Univ., Abbrev. State, year.
\end{enumerate}
See \cite{b25,b26}.
\item \emph{Basic format for the most common types of unpublished references:}
\begin{enumerate}
\item J. K. Author, private communication, Abbrev. Month, year.
\item J. K. Author, ``Title of paper,'' unpublished.
\item J. K. Author, ``Title of paper,'' to be published.
\end{enumerate}
See \cite{b27}--\cite{b29}.
\item \emph{Basic formats for standards:}
\begin{enumerate}
\item \emph{Title of Standard}, Standard number, date.
\item \emph{Title of Standard}, Standard number, Corporate author, location, date.
\end{enumerate}
See \cite{b30,b31}.
\item \emph{Article number in~reference examples:}\\
See \cite{b32,b33}.
\item \emph{Example when using et al.:}\\
See \cite{b34}.
\end{itemize}
\section{Introduction}
\label{sec:introduction}
The virulent, fast spreading, and mutating nature of SARS-CoV-$2$ caused a worldwide pandemic with more than $229$ million people infected and $4.7$ million lives taken in the span of $1.5$ years since emerging \cite{whodash}. This pandemic has challenged the ability of healthcare systems to triage patients and overwhelmed capacities. Beyond healthcare, societal impacts have been apparent and major sectors of the global economy have been affected \cite{sohrabi2020world}. With this and the devastating nature of the pandemic, curbing the infection caused by the virus has become a major concern for all nations. Several protocols have been implemented such as mask mandates and lockdowns to reduce the spread of the virus. Frequent and rapid surveillance testing, however remains the most contributing factor in reducing the transmission of the virus \cite{nicola2020health}. The current method-of-choice for COVID-$19$ testing is based on reverse transcriptase-PCR (RT-PCR) platforms. These RT-PCR-based test procedures can be expensive, use complex reagents, and require a relatively high level of experience~\cite{carter2020assay}.
Alternatively, MALDI-ToF mass spectrometry (MS) methods provide advantages over RT-PCR-based methods such as the ability to rapidly process a large number of samples and yield highly-accurate results in COVID-$19$ diagnostic tests \cite{tran2021novel,nachtigall2020detection}. The adoption of MALDI-ToF MS approach also sees a reduced demand for complex reagents like those required in molecular assay techniques such as RT-PCR. Apart from the equipment used, the specimen used in the analysis also plays a vital role in the applicability of a diagnostic test. Current procedures use a nasopharyngeal (NP) swab or a saliva sample. Among these two specimens, saliva has been reported to contain higher viral loads for longer periods of time~\cite{burki2020testing}. Moreover, saliva collection is less invasive and does not incite a cough, sneeze or a gag reflex, which could produce infectious aerosols \cite{pasomsub2021saliva}.
For these reasons, in order to add to the number of diagnostic tests for COVID-$19$, we adopt the collection of saliva through gargle samples and leverage a MALDI-ToF MS-based approach for the analysis. This approach has already been established as an effective workflow, reported to agree with results from NP swab RT-PCR analysis for the diagnosis of COVID-$19$~\cite{iles2020development, chivte2021maldi}. The output of such MALDI-ToF MS-based approach is a mass spectrum, which plots the mass-to-charge ratio (m/z) versus ion intensity as shown in Figure~\ref{fig:raw}. This reflects the masses of biomolecules found within the saliva specimens and their relative abundance. These spectra can be complex, containing hundreds of signals that differ from specimen to specimen, and could include critical information regarding the presence of virus or the immune response of the patient.
Therefore, we adopt an AI-based approach to learn the relationships between the intensity of m/z peaks and the outcome of a RT-PCR test. Often, when an AI system learns a complex pattern, their decision making process is unclear and termed a black-box system. These black-box systems are uninterpretable and pose an impossible task for domain experts and health practitioners to understand the contributing factors to the outcome of a test. We tackle this problem using X-AI to present the rationale behind decision making of our AI-based approach. The X-AI approach not only provides valuable diagnostic information but also shows the ability of reproducing our results in clinical settings. The ability to analyze the response from the model instills trust in our AI system.
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{figures/rawfinal.png}
\caption{Comparison of COVID-$19$ negative and a COVID-$19$ positive gargle MALDI-ToF spectra after preprocessing with baseline correction and normalization. Biomarkers ranges of interest in the spectra are highlighted with a gray hue. Zoom-in panels for each range show differences between the disease statuses.}
\label{fig:raw}
\end{figure*}
\section{LITERATURE REVIEW}
\label{sec:literaturereview}
Following the recent success of AI solutions in the healthcare domain, several researchers have leveraged AI techniques to design efficient COVID-$19$ testing methods using mass spectrometry datasets~\cite{nachtigall2020detection, tran2021novel, deulofeu2021detection}. These three studies were conducted on different datasets but the mode of sample collection, in this case NP swab, remains consistent. These studies differ, however, in sample processing and choice of algorithm. Nachtigall et al.~\cite{nachtigall2020detection} have used Cary-Blair transport medium to collect $362$ samples, extracted and utilized peaks from the mass spectra as features. Further, they applied correlation-based\cite{hall2000correlation} and information-gain-based~\cite{appavu2011bayes} feature selection methods to identify $88$ important peaks in the spectra. Although, they achieved $93.9\%$ accuracy with the support vector machine (SVM)~\cite{noble2006support} algorithm, their results were not consistent across the population. In other work, Tran et al.~\cite{tran2021novel} used a saline solution to collect $199$ samples. While optimizing their AI solution, they used Machine Intelligence Learning Optimizer (MILO)~\cite{milo} for hyperparameter searching. As a result, their extreme gradient boosting (XgBoost)~\cite{chen2016xgboost} and deep neural network (DNN)~\cite{lecun2015deep} algorithms achieved $98.3\%$ accuracy. In a more recent study, Deulofeu et al.~\cite{deulofeu2021detection} used nasal swabs and collected $148$ samples using DeltaSwab-ViCUM viral transport medium (VTM-$1$) and $66$ samples using DeltaSwab-Virus viral transport medium (VTM-$2$). Deulofeu et al.~\cite{deulofeu2021detection} pre-processed raw data using principal component analysis (PCA)~\cite{wold1987principal} and evaluated several machine learning (ML) algorithms with a $10$-fold cross-validation. The authors reported the accuracies for VTM-$1$ and VTM-$2$ datasets as $70.3\%$ with XgBoost~\cite{chen2016xgboost} and $88.4\%$ with SVM~\cite{noble2006support} respectively.
The AI algorithms employed in these studies are complex and often require several pre- and post-processing steps. The absence of these details make it difficult to compare their results. Additionally, these solutions only provide a simple, binary decision: positive or negative. In order for medical practitioners and domain experts to further analyze their decisions and make more informed conclusions, it is crucial for testing methods to provide rationale on its decision making process. This ultimately makes the solution and decision-making process more trustworthy~\cite{ribeiro2016should} for its application in clinical settings and more acceptable to end users. X-AI is one of the more recent advancements in AI\cite{linardatos2021explainable} and is capable of accurately approximating the prediction mechanisms of AI algorithms. Explainable algorithms such as Shapely Additive Explanations (SHAP) \cite{lundberg2017unified} have gained popularity in the AI interpretability domain and AI in healthcare \cite{lundberg2020local} but have never been used in COVID-$19$ diagnostic applications. In this paper, our goal is two fold: $1)$ to train an AI model to achieve high accuracy in COVID-$19$ diagnosis and $2)$ to leverage X-AI techniques to explain the reasoning behind the outcomes from AI models.
\section{OUR METHODOLOGY AND CONTRIBUTIONS}
\label{sec:contributions}
The facts mentioned in section~\ref{sec:introduction} emphasize the need for additional COVID-$19$ diagnostic methods that are relatively inexpensive, non-invasive, highly accurate, and are interpretable. Several studies since the onset of the pandemic have proposed more platforms and methods for COVID-$19$ testing, however, these have limitations based on the experiment procedures or the analysis. In this work, we address some of these limitations by interpreting our AI algorithm to provide more informed explanations of diagnosis, leveraging a cost-effective and easy-to-use MALDI-ToF MS approach, and using a non-invasive specimen type (saliva). After assessing multiple AI approaches, we built our AI solution by training a tree-based ensemble algorithm (random forest~\cite{breiman2001random}) on MALDI-ToF MS spectra data. Additionally, we interpret the predictions from trained random forest model using X-AI techniques such as SHAP, impurity-based feature importance, and permutation feature importance.
\section{DATASET PREPARATION}
\subsection{Data Collection}
We collected gargle samples from students at Northern Illinois University Athletics (Illinois, USA) in the months of August and September of $2020$. Students who consented to participate in our study were asked to provide a water gargle sample and a simultaneous NP swab sample. The gargle sample was processed as described in~\cite{chivte2021maldi} and the NP swabs were confirmed COVID-$19$ positive/negative by RT-PCR conducted by the Dekalb County Health Department (Illinois, USA). For each gargle sample, the outcome of the RT-PCR test on NP swab was used as the ground truth in our analyses. In total, $152$ gargle samples were analyzed, including $60$ COVID-$19$ positive and $92$ COVID-$19$ negative.
In brief, sample handling included participants gargling $10$ mL of bottled spring water for $30$ seconds which was then deposited in a $50$ mL conical centrifuge tube. The collected specimen was filtered through a $\leq0.45\mu$m syringe filter, acetone precipitated to concentrate proteins, centrifuged, and resuspended in a reconstitution buffer. After incubation, we spotted each sample on a target plate, air-dried, and transferred the plate to a Shimadzu AXIMA Performance~\cite{massspec} mass spectrometer, which was calibrated with a protein standard daily. Ions were detected in positive-ion, linear detection mode over a range of $2,000$ -- $200,000$ m/z. Finally, we corrected the baseline of all the generated spectra using the Shimadzu Launchpad software obtained from the manufacturer~\cite{massspec} and normalized the ion count intensities with the parent peak of the protein standard calibrant.
\subsection{Potential Biomarker Selection}
\label{sec:biomarkeridentification}
The presence or absence of peaks and their relative intensities in the collected mass spectra can be correlated to the presence or abundance of proteins within each specimen. The spectra we collected represent a multitude of proteins (after acetone precipitation) that are in the range of $2,000$ -- $200,000$ m/z. Since our goal was to differentiate between a positive and negative sample, we carefully selected only those m/z ranges that were observed to change given the COVID-$19$ status by RT-PCR. An overlay of mass spectra highlighting these differences for a COVID-$19$ positive and negative sample is shown in Figure~\ref{fig:raw}. Note that Figure~\ref{fig:raw} only shows a segment of the spectra where potential COVID-$19$ biomarkers are present. An attempt to identify the potential biomarkers was made in our previous work~\cite{chivte2021maldi}. These biomarker peaks were selected to potentially include both human proteins (immunoglobulin heavy chain, amylase, asymptomatic biomarker) and SARS-CoV-$2$ viral proteins (Spike protein subunits S$2$ and S$1$), summarized in Table \ref{table:bioRanges}. For the sake of simplicity, here, these ranges were named Human Protein-$1$, Human Protein-$2$, Human Protein-$3$, and Viral Protein-$1$, Viral Protein-$2$, respectively.
\begin{table}[htb]
\begin{center}
\begin{tabular}[h!t]{||c|c|c||}
\hline
\makecell{\textbf{Biomarker}} & \textbf{Mass Range} & \textbf{Potential Identity}\\ \hline \hline
$B_0$ & $27,900$ -- $29,400$ & { \makecell{Human Protein-$1$}}\\ \hline
$B_1$ & $55,500$ -- $59,000$ & {\makecell{Human Protein-$2$}}\\ \hline
$B_2$ & $66,400$ -- $68,100$ & {\makecell{Viral Protein-$1$}} \\ \hline
$B_3$ & $78,600$ -- $80,500$ & {\makecell{Viral Protein-$2$}}\\ \hline
$B_4$ & $111,500$ -- $115,500$ & {\makecell{Human Protein-$3$}} \\ \hline
\end{tabular}
\\ \rule{0mm}{0mm}
\end{center}
\captionof{table}{Potential biomarker ranges and their descriptions.}
\label{table:bioRanges}
\end{table}
\vspace{-1em}
\subsection{Feature Engineering}
\label{sec:featureengineering}
The performance of supervised ML algorithms depends on the quality of the features used in training. Typically, the prediction power of any ML classification algorithm increases with an increase in the number of features and languishes after reaching a certain number~\cite{mclachlan2004discriminant}. If we had directly used the raw readings of all five chosen m/z ranges within the spectra, each sample would have $12,600$ features (the combined bandwidth of all $5$ ranges); this could potentially overfit the ML models for our limited dataset of $152$ samples. We thus extracted features that balanced being easy to compute but also have enough information to discriminate between the positive and negative samples. After iterating through many statistical and intuitive features, we decided to compare the area under the curves (AUC) \cite{tallarida1987area} for the five biomarkers ($B_0$ -- $B_4$). Although the AUC captures the intensity of a biomarker, a wide variance from one sample to another was observed. This happens mainly due to two reasons. First, each sample has potentially different amounts of human/viral protein present and second, information regarding exposure to the virus or time course of infection was not collected. Such differences in immune response and viral load would be reflected in the biomarker intensities. Regardless, we observed that the ratio of AUC between pairs of biomarkers showed a consistent pattern for positive and negative samples. Therefore, we computed $10$ features for each sample as the ratio of each possible combination of different biomarker pairs (shown in Table \ref{table:features}). These features were computed for the MALDI-ToF dataset of size $152$ with $60$ COVID-$19$ positive and $92$ COVID-$19$ negative samples.
\vspace{-1em}
\subsection{Data Preparation}
We partitioned $152$ samples from our MALDI-ToF dataset using a train-test-split strategy with $70:30$ train-test-split ratio to form two disjoint and independent splits. This prevents data leak from classifier training into performance assessment. Within both train- and test-splits, the ratio of the count of positive to negative samples was kept constant to maintain class balance. Our classifier uses the train-split dataset ($41$ COVID-$19$ positive and $65$ COVID-$19$ negative) and learns the COVID-$19$ prediction task. We then assess the performance of the trained classifier on the unseen test-split dataset ($19$ COVID-$19$ positive and $27$ COVID-$19$ negative samples).
\section{METHODS}
\label{sec:methods}
After evaluating several tree-based algorithms, we selected random forest (RF)~\cite{breiman2001random} classifier due to its superior performance among tree-based algorithms. The goal of RF algorithm is to train a classifier function ($f$) that learns the relationships between the ratio features generated from training samples, ($x_{train}$) and the outcome of the associated COVID-$19$ RT-PCR test ($y_{train}$). After completion of training, the algorithm was tested on new/unseen saliva samples ($x_{test}$) to evaluate the precision and recall metrics of the trained classifier ($f(x_{test})=y_{test}$) as compared to the corresponding RT-PCR result.
Further, we present the complex relationships learned by the trained RF algorithm between the input ratio features and the outcome of a COVID-$19$ RT-PCR test in a human interpretable manner with the aid of X-AI algorithms. By leveraging the X-AI algorithms, explanations for the outcomes of RF algorithm are possible on both a global (all-samples) and local (per-sample) basis.
\begin{table}[ht]
\begin{center}
\begin{tabular}{||c|c||}
\hline
\textbf{Feature} & \textbf{Description} \\ \hline \hline
$R_0$ & $AUC_{B_0} \thickspace/\thickspace AUC_{B_1}$ \\ \hline
$R_1$ & $AUC_{B_0}\thickspace/\thickspace AUC_{B_2}$ \\ \hline
$R_2$ & $AUC_{B_0}\thickspace/\thickspace AUC_{B_3}$ \\ \hline
$R_3$ & $AUC_{B_0}\thickspace/\thickspace AUC_{B_4}$ \\ \hline
$R_4$ & $AUC_{B_1}\thickspace/\thickspace AUC_{B_2}$ \\ \hline
$R_5$ & $AUC_{B_1}\thickspace/\thickspace AUC_{B_3}$ \\ \hline
$R_6$ & $AUC_{B_1}\thickspace/\thickspace AUC_{B_4}$ \\ \hline
$R_7$ & $AUC_{B_2}\thickspace/\thickspace AUC_{B_3}$ \\ \hline
$R_8$ & $AUC_{B_2}\thickspace/\thickspace AUC_{B_4}$ \\ \hline
$R_9$ & $AUC_{B_3}\thickspace/\thickspace AUC_{B_4}$ \\ \hline
\end{tabular}
\\ \rule{0mm}{0mm}
\end{center}
\captionof{table}{Ratio features generated from AUC values of five potential biomarkers.}
\label{table:features}
\end{table}
\vspace{-1em}
\subsection{Random Forest (RF) Classifier}
\label{sec:rf}
RF is an ensemble classification algorithm that combines several tree predictors (Decision Trees~\cite{safavian1991survey}) to build a forest classifier. Each tree in the forest is trained from a randomly sampled vector from the input features. These random vectors are sampled independently from the same distribution in the predictor space~\cite{breiman2001random}. The outcome of a trained forest is the majority vote taken in the form of mean or median of the predictions from all the independently trained trees. Additionally, in RF algorithm, we can fine-tune the parameters such as the number of trees (also known as estimators) in the forest, methods of bootstrapping, and the properties of trees. If bootstrapping is enabled, random samples are taken from the training set with replacement to train decision trees and take the majority vote. We can tune the core properties of each decision tree such as maximum depth, maximum number of leaves, splitting criterion, minimum samples per split, and minimum samples per leaf to optimize the performance.
The first step in training the RF classifier is to find optimized hyperparameters by leveraging a grid search technique that uses $5$-fold cross-validation on the training dataset. The grid search algorithm conducts exhaustive try-outs on all possible sets of hyperparameter configurations wherein each set of configuration is catalogued with its cross-validation accuracy score. The hyperparameter configuration with the best score (shown in Table~\ref{table:hyperparameters}) was selected to train the RF algorithm on full training data which is further evaluated and interpreted with the test-split data.
\begin{table}[h!t]
\begin{center}
\begin{tabular}{||c|c||}
\hline
\textbf{Hyperparameter} & \textbf{Value} \\ \hline \hline
Number of estimators & $100$ \\ \hline
Bootstrap & True \\ \hline
Splitting criterion & Gini impurity \\ \hline
Maximum depth of tree & $3$ \\ \hline
Minimum samples per split & $2$ \\ \hline
Minimum samples per leaf & $1$ \\ \hline
\end{tabular}
\\ \rule{0mm}{0mm}
\end{center}
\captionof{table}{Hyperparameters of our RF classifier.}
\label{table:hyperparameters}
\end{table}
\vspace{-1em}
\subsection{Random Forest Interpretation Using Explainable-AI}
\label{sec:interpret}
To instill trust in the prediction of the model, we interpret the RF model both globally and locally using X-AI techniques. While global interpretation explains the importance of each feature in overall model prediction, local interpretation provides the reasoning on a single instance level. For global interpretation, analyzing the importance of features determined by a trained model gives information about its prediction process. In this section, we leverage multiple techniques to investigate feature importance of our trained RF model.
\subsubsection{Impurity-based feature importance (IFI)}
\label{sec:ifi}
First, we interpreted the trained RF classifier with a model-specific, IFI technique. This technique computes the feature importance as the average decrease in Gini impurity~\cite{ceriani2012origins} from a feature node to its child nodes. Generally, except for the leaf nodes that occur at the bottom of the tree, all other nodes have a threshold condition placed on a feature whose decision splits the node into two child nodes. The drop in impurity in this split is then calculated as the difference of Gini impurity between the node and the sum of its child nodes, weighted by the number of samples in each node. The final feature importance is the mean of drop in impurity at every prevailing feature node across the forest. Following these procedures, we calculated the IFI for all $10$ features as shown in panel-A of Figure~\ref{fig:featureimportance} with the drop in Gini impurity on the y-axis and features on the x-axis. Although this method gives a general understanding of feature rankings, one disadvantage is inflation of importance for features with high cardinality in the forest. Therefore, we investigated other feature importance algorithms to validate the results.
\subsubsection{Permutation Feature Importance (PFI)}
\label{sec:pfi}
The PFI technique~\cite{breiman2001random}
calculates the drop in performance when the trained RF algorithm is validated with permuted feature values which essentially makes the feature a noise. This method analyzes one feature at a time by permuting their values to break the association between the feature and the outcome. Clearly, permutation of important features results in a larger performance drop. Therefore, the feature importance is gauged as the drop in permuted feature accuracy from the base accuracy (where every feature was used in its original form). Bias from random value assignment to feature values was prevented by permuting values and averaging the drop in accuracy across $100$ iterations. However, this algorithm falls short in capturing feature interaction effects as it permutes only one feature at a time. Panel-B in Figure~\ref{fig:featureimportance} shows the permutation importance scores for RF classifier with overall feature importance printed on the y-axis and the features on the x-axis.
\subsubsection{Shapely Additive Explanations (SHAP)}
\label{sec:shap}
SHAP algorithm, introduced by Lundberg and Su-In~\cite{lundberg2017unified} can be formalized as an additive feature attribution technique that unifies concepts from local model-agnostic explanations (LIME)~\cite{ribeiro2016should} and Shapely values~\cite{shapley1953value}. In general, linear models are easy to interpret as the feature importance of the model are reflected by the coefficients of the model itself. This is not the case for the RF algorithm as a simple linear model does not have the capability to approximate a more complex RF algorithm. However, output of a single sample could be explained by generating some neighbourhood samples followed by fitting a linear model locally on the newly generated neighborhood samples. The LIME algorithm follows this approach: it generates a local point of view per sample by applying perturbations to its feature values and generating some artificial data in the neighbourhood of a sample. LIME then fits a surrogate model on the artificial data and presents the coefficients of the surrogate model as the model interpretation. The downside of LIME lies in its generation of artificial samples using a heuristic kernel function. Artificially replicating clinical data is a difficult task and depending on kernel function heuristics remains unreliable. SHAP adopts LIME procedures and bridges its limitations with the Shapely values which borrow principles from cooperative game theory. Shapely values rationalize the distribution of credit or payoff among the players in a cooperative game by inspecting the contribution of each player individually as well as in a collaborative team effort~\cite{shapley1953value}.
In theory, the feature attributions assigned by our trained RF classifier ($f$) to the features $R_i$ (where $i\in(0,9)$ for $10$ ratio features) in a given datasample ($x$) is computed as ($\phi_{R_i}(f,x)$) using Equation~\eqref{eq:shaplyeq}. This equation facilitates SHAP to compute a unique optimal solution for feature attribution by combining individual as well as team efforts. This can be understood in light of PFI where we extracted feature attributions by masking each feature at a time and computing the drop in performance of the classifier. Clearly, PFI captures individual contributions but fails to grasp collaborative contributions. The more intuitive approach of SHAP allows for collaborative contributions in conjunction to individual contributions to be considered by iterating over all possible subsets $S$ in our feature space having $F$ number of features excluding feature $R_i$ (where $F=9$) as $S \subseteq F /\{R_i\}$. In each subset, the marginal contribution ($C$) of feature $R_i$ is extracted as the difference in outcome of function $f$ with and without $R_i$ in the subset as shown in Equation~\eqref{eq:shaplyCont} where $f_{S\cup\{R_i\}}$ and $f_{S}$ are two retrained functions with and without $R_i$ respectively that are marginalized over features absent in the subset. The resultant sum of $C$ is weighted with $W$ (refer to Equation~\eqref{eq:shaplyWeight}) whose numerator is the product of different number of ways a subset can be formed ($S!$) and the number of ways to choose features excluding the features in a subset ($(|F|-|S|-1)!$). The numerator is normalized by $F!$, the total possible ways of choosing the features. Finally, for each feature in $x$, SHAP values are computed using Equation~\ref{eq:shaplyeq} that iterates over all possible subsets $S$ and calculates SHAP values. The final SHAP values have a magnitude that shows the importance of the feature and a direction where positive sign or negative sign corresponds to their contribution towards the COVID-$19$ positive or negative decision.
Since the idea behind Shapely values is that there exists only one solution to allocate the attributions, the results are more reliable than previous methods such as LIME~\cite{ribeiro2016should}. In addition to this, SHAP has other desirable axiomatic properties. The local accuracy property entails that the sum of all feature attributions of a model should approximate the original model meaning $f(x)=\phi_{R_n}+\sum_{i=0}^{9}\phi_{R_i}$, where $R_n$ is the feature attribution when all features are toggled off. The missingness property ensures that the features that are toggled of have no impact in model's output. Lastly, the consistency property ensures for any feature $R_i$, $\phi_{R_i}$ remains consistent with respect to the impact of feature values of on any retrained model. For example, in two functions trained on different subsets including the feature $R_i$, if the feature values of $R_i$ have larger impact on $f1$ than $f2$, then $\phi_{R_i}(f1, x) \geq \phi_{R_i} (f2, x)$.
\begin{equation}
C = f_{S\cup\{R_{i}\}}(x_{S\cup\{R_{i}\}})-f_{S}(x_{S})
\label{eq:shaplyCont}
\end{equation}
\begin{equation}
W = \frac{|S|!(|F|-|S|-1)!}{|F|!}
\label{eq:shaplyWeight}
\end{equation}
\begin{equation}
\phi_{R_{i}}(f,x)=\sum_{S \subseteq F /\{R_{i}\}} W \times C
\label{eq:shaplyeq}
\end{equation}
One can assume that classical Shapely value approach in Equation~\eqref{eq:shaplyeq} is a promising way to obtain feature attributions, nevertheless, the computational complexity reaches non-polynomial time which makes it difficult to apply to our dataset. In $2020$, Lundberg~\cite{lundberg2020local} introduced tree explainer, a model that leverages the underlying structure of tree-based models and optimizes the computational complexity to polynomial time. Tree explainer approximates the model while adhering to the natural properties of Shapely values~\cite{shapley1953value} mentioned previously. Therefore, we used tree explainer~\cite{lundberg2020local} to generate a SHAP value matrix of dimensions $[46,10]$, for $46$ testing samples having $10$ ratio features. Using the SHAP value matrix, we then interpret our trained RF model both globally and locally. For global model interpretation, we extract the magnitude of feature importances in two steps: $1$) take the absolute SHAP values which nullifies the effect of direction and $2$) compute the mean of absolute SHAP values across all samples. This results in condensing the $[46,10]$ matrix to a single array of $10$ feature importances which are shown in panel-C of Figure~\ref{fig:featureimportance}. By comparing all $3$ feature importance algorithms in Table \ref{table:globalFI}, it is conclusive that the two most important features are $R_9$ and $R_5$ and the least important are $R_2$ and $R_4$.
In addition to global (all-sample) explanations, we present local (per-sample) explanations using SHAP value matrix as shown in Figure~\ref{fig:shaplocal}. In these plots, the x-axis represents the SHAP value per-sample and features are numbered on the y-axis. Additionally, we embedded feature values in color ranging from blue (low value) to red (high value) and assigned each feature to a quartile based on its value as shown at origin of each bar. This enables us to interpret the SHAP values and observe their effects on the feature values. Details on global and local explanation are provided in sections~\ref{sec:globalinterpretation} and~\ref{sec:localinterpretation}, respectively.
\begin{figure}[h!t]
\centering
\includegraphics[width=\linewidth, keepaspectratio]{figures/allimppanel.png}\\
\caption{Feature importance for RF classifier using three feature interpretation techniques: Impurity-based feature importance (panel-A, top), Permutation Feature Importance (panel-B, middle), and Shapely Additive Explanations (panel-C, bottom). The y-axis shows the feature importance for each feature shown on the x-axis.}
\label{fig:featureimportance}
\end{figure}
\section{Results}
\subsection{Results of Random Forest Classifier}
\label{sec:rfresults}
We trained the RF algorithm with the hyperparameters (shown in Table \ref{table:hyperparameters}) extracted through an exhaustive grid search. A total of $106$ samples were employed in training and $46$ in testing the model. Although data augmentation techniques could enrich the dataset to create more training samples, we avoided doing this because it could have made the dataset noisy by adding false patterns. The trained RF algorithm achieved $91.30\%$ accuracy on the test-split, whereas $86.79\%$ accuracy was achieved on the train-split. Similar high accuracies on both training and test sets assure that the model is not overfitted. Confusion matrix for both splits are shown in Figure \ref{fig:rfconf}. Overall, only $4$ and $14$ samples are classified falsely in the test and train dataset, respectively. As mentioned earlier, our dataset had more negative samples than positive ones, making it crucial to examine the accuracy of the model for both positive and negative categories. Therefore, we evaluated precision, recall and F$1$ metrics on both the test and train dataset for positive and negative samples independently. As shown in the evaluation report (in Figure~\ref{fig:rfreport}), the train-split had a high precision, recall and F$1$-score of $0.86$, $0.94$, $0.90$ for the COVID-$19$ negative samples and $0.89$, $0.76$, $0.82$ for positive samples, respectively. On the other hand, precision, recall and F$1$-scores for COVID-$19$ negative and positive for the test-split were $0.87$, $1.00$, $0.93$ and $1.00$, $0.79$, $0.88$, respectively. In case of test-split, there were $0$ false positives, $4$ false negatives and other $42$ samples had true predictions. In train-split, the false predictions were slightly higher with $4$ false positives and $10$ false negatives and a total of $92$ correct predictions. Although these metrics portray high efficiency of the trained model in COVID-$19$ testing, it still does not provide transparency on its prediction mechanisms. This motivated us to further dissect the trained RF model and understand its prediction mechanisms using X-AI techniques.
\begin{table*}[htbp]
\begin{center}
\begin{tabular}{||c|c||}
\hline
Method & Order of feature importances \\ \hline \hline
Impurity-based feature importance & $R_9 > R_5 > R_8 > R_7 > R_0 > R_3 > R_1 > R_6 > R_4 > R_2$ \\ \hline
Permutation feature importance & $R_9 > R_5 > R_6 > R_8 > R_0 > R_7 > R_1 > R_3 > R_4 > R_2$ \\ \hline
Shapely additive explanations & $R_9 > R_5 > R_8 > R_7 > R_0 > R_6 > R_3 > R_1 > R_4 > R_2$ \\ \hline
\end{tabular}
\\ \rule{0mm}{0mm}
\captionof{table}{Global feature importances shown in descending order for different feature interpretation techniques.}
\label{table:globalFI}
\end{center}
\end{table*}
\begin{figure}[h!t]
\centering
\includegraphics[width=0.48\linewidth, keepaspectratio]{figures/rfconfTrain.png}
\includegraphics[width=0.48\linewidth, keepaspectratio]{figures/rfconfTest.png}\\
\caption{Confusion matrices for the RF classifier on the train-split and test-split.}
\label{fig:rfconf}
\end{figure}
\begin{figure}[h!t]
\centering
\includegraphics[width=\linewidth, keepaspectratio]{figures/prfe.png}\\
\caption{Precision, recall and F$1$-score for test-split dataset for COVID-$19$ negative (on left) and COVID-$19$ positive (on right).}
\label{fig:rfreport}
\end{figure}
\subsection{Global Interpretation of Random Forest Classifier}
\label{sec:globalinterpretation}
The idea of global interpretation of a model goes beyond the analysis of feature importance. In fact, it can elaborate on how the distribution of each feature is affecting the model prediction. As shown earlier, each of the three feature importance algorithms concluded similar feature ranking: the top two features were consistently $R_9$ and $R_5$ while the bottom two were $R_2$ and $R_4$ (Table \ref{table:globalFI}). On the other hand, $R_7$, $R_8$, $R_1$, and $R_3$ had slightly varying importances among the three algorithms. Only one feature, $R_6$, was seen to be varying by two positions across them. It is fairly common to observe slight differences in the feature importances across the three algorithms due to underlying procedural differences. However, on the whole, these rankings reveal that nine out of ten feature importances assigned by the three algorithms were fairly consistent. The consistent nature of feature attributions show that our algorithm can be understood when viewed from the different perspectives of the X-AI algorithms. Although these are important findings, it doesn't explain how the model prediction is correlated with the feature values. While IFI and PFI algorithms are good at generating feature rankings, they do not provide enough information to explore the relationship between model prediction and the distribution of individual feature. Therefore, we extended the SHAP algorithm that is capable to analyze the feature distribution and the model dependency on it.
The SHAP algorithm generated a violin plot, which is a fine-grained view of SHAP (feature importance) values and feature distribution as shown in Figure~\ref{fig:shap}. This plot illustrates the distribution of SHAP values across all test-split samples, allowing us to observe trends of SHAP values in conjunction with feature values. To interpret these plots, the higher the SHAP value, the larger is the importance; the positive SHAP values support positive COVID-$19$ prediction, negative values support negative output; and the feature value is color coded where blue represents the values in lower ranges, red represents the values is in higher ranges. For example, the feature $R_9$ has SHAP values extending up to $+0.165$ towards a positive COVID-$19$ decision, where the feature value is low as it is colored in blue. On the negative side, it can be seen that mostly high values of $R_9$ contribute towards a negative COVID-$19$ prediction. On the contrary for feature $R_5$, smaller and larger feature values contribute towards COVID-$19$ negative and positive tests, respectively. Features $R_6$ and $R_7$ follow similar patterns as $R_5$ while $R_0$ and $R_1$ have similar patterns of $R_9$.
\begin{figure}[h!t]
\centering
\includegraphics[width=\linewidth,keepaspectratio]{figures/violin.png}\\
\caption{Violin plot showing for global interpretation of RF classifier using SHAP values.}
\label{fig:shap}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[width=0.31\linewidth,keepaspectratio]{figures/TP0.png}
\includegraphics[width=0.31\linewidth,keepaspectratio]{figures/TN3.png}
\includegraphics[width=0.31\linewidth,keepaspectratio]{figures/FN25.png}\\
\caption{Local explanations generated using SHAP algorithm for three records, true positive sample (on the left), true negative sample (in the middle) and false negative sample (on the right).}
\label{fig:shaplocal}
\end{figure*}
The violin plot provides an approximate idea about the model output for any sample based on the feature distribution. This can be immensely helpful to cross examine the model in the case that the model prediction is not following the trend of the violin plot. For example, if the model is predicting COVID-$19$ positive for a sample where the $R_9$ feature value is very high, we can easily identify that the model is not following the trend and hence, it needs further validation. This kind of information is critical to instill trust in users so that they can question outputs from the model.
\subsection{Local Interpretation of Random Forest Classifier}
\label{sec:localinterpretation}
For domain experts and clinicians handling COVID-$19$ testing, local explanations of the model unveil rich diagnostic information as they provide rationale behind predictions on a per-sample basis. For this purpose, we have further leveraged the SHAP algorithm for interpreting one sample at a time. To do this, we take a holistic view on the classifier performance by viewing different cases of predictions such as a true positive (TP), a true negative (TN), and a false negative (FN) (shown in Figure~\ref{fig:shaplocal}). Since, there were no cases of false positives in the test dataset, we could not explore them in examples. In Figure~\ref{fig:shaplocal}, each bar is colored based on its feature value, ranging from a small value (blue) to large value (red). The labels at the origin of each bar indicate the condition on each feature value to support either category. The feature importance, similar to the global SHAP plots, is shown on the x-axis.
The local interpretation helps us to understand how, for any individual sample, different features interact with each other to sway the final decision of the model. In the case of a true positive sample (left plot in Figure \ref{fig:shaplocal}), $R_9$ and $R_5$ are the most contributing features towards COVID-$19$ positive outcome. followed by $R_8$, $R_4$, $R_2$ and $R_1$ with lower importance. Although $R_0$, $R_3$ and $R_7$ are directing towards a negative prediction, there is more support for the positive case based on the SHAP values. In the true negative record (middle plot in Figure \ref{fig:shaplocal}), the $R_9$ feature value is mid-ranged followed by features $R_7$, $R_8$, $R_5$, and $R_0$. We also observe in the true negative sample that all features except $R_9$ have relatively smaller values compared to a true positive. Although not any single feature is directing very strongly towards negative test, all of them together support a negative prediction. We finally show explanation of a false negative case (right plot in Figure \ref{fig:shaplocal}). In the false negative sample, the $R_9$ feature had a smaller value compared to the true negative sample. However, all other features except $R_6$ had similar values compared to the true negative sample. These high similarities of COVID-$19$ positive sample with a negative one potentially alluded the model to classify a COVID-$19$ positive as a negative test. Perhaps a medical practitioner in control of making decision based on the outcome of a COVID-$19$ test can benefit from a local analysis of the test and inspect the features that contributed most towards a positive or negative test. This in conjunction with consideration of a patient's symptoms can lead to more informed conclusions and diagnosis.
\section{Discussion}
\label{sec:discussion}
While analyzing our results through various interpretation techniques, we observed quite remarkably that all three methods of global feature interpretation consistently resulted in the most important features as $R_9$ and $R_5$ (Table~\ref{table:globalFI}). The meaning of these features is of particular interest because of the biomarker ranges making up the features, $B_3$/$B_4$ for $R_9$ and $B_1/B_3$ for $R_5$. First, the features $R_9$ and $R_5$ are host-to-virus and virus-to-host features, respectively, and represent the interaction between host immune response and viral antigen. This is as described in our previous work~\cite{chivte2021maldi}. Second, the viral protein fragment that corresponds to $B_3$ is tentatively suggested to be the S$1$ subunit of the viral spike protein (see Table~\ref{table:bioRanges}). S$1$ is a key viral biomarker in the mass spectra because the observed m/z ratio does not overlap significantly with other proteins or fragments from the gargle sample.
The least important features, $R_4$ and $R_2$, were decidedly the same by each method as well. $R_4$ contains $B_2$, which is ambiguous in that the AUC value for this peak may consist of \textit{both} human and viral proteins (a small amount of human serum albumin and spike protein S$2$ subunit, respectively). Therefore, the presence of $B_2$ could render $R_4$ less important, as $B_2$ would be present in both disease and non-disease states. Considering the nature of the biomarkers, it was odd that $R_2$ was not found to be as important as $R_5$: both ratio features should technically represent the same proteins, as $B_0$ is the double charged ion of $B_1$. To investigate this, the correlation plot between each ratio feature value was plotted and, not surprisingly, $R_5$ and $R_2$ showed a linear relationship (as shown in Figure~\ref{fig:correlation}). This is most likely the reason why $R_2$ was deemed less important than $R_5$ and the least important feature altogether: the algorithm avoids repetitive predictions by not placing importance on features equivalent in meaning. It should be pointed out that each biomarker has been assigned a putative identity based on preliminary studies only and have yet to be unambiguously characterized.
\begin{figure}
\centering
\includegraphics{figures/r5r2.png}
\caption{Correlation plot showing positive correlation between feature values of $R_5$ and $R_2$.}
\label{fig:correlation}
\end{figure}
Turning to local feature interpretation, these could offer a powerful diagnostic view: instead of a binary decision by RT-PCR or algorithm with a percent confidence, the local explanations generated using SHAP values show the contribution of each feature in each sample. This could allow domain experts and health officials administering such a test to make more informed decisions based on the putative host-to-viral or viral-to-host ratios, for case-by-case bases as seen in Figure~\ref{fig:shaplocal}. If the algorithm decides positive or negative, the local feature explanation would provide additional information as to \textit{why} the test output its decision. However, due to the small size of the dataset and the lack of information regarding time points of infection/exposure when the samples were collected, it is not possible to draw definitive conclusions from the local explanations presented here.
Nonetheless, not only in this diagnostic approach is the viral aspect of COVID-$19$ considered (as is only considered in PCR-based tests), but the human immune response is also given due consideration. ML algorithms coupled with the utility of X-AI interpretability can be implemented into healthcare to address the limitations of black box systems and improve the diagnostic process, from test output to informed care.
\section{Conclusion}
\label{sec:conclusion}
We have analyzed the application of machine learning models on the MALDI-ToF based data of gargle samples and inferred the performance of RF to be the best among the other algorithms. The RF classifier was evaluated on a $70\%$-$30\%$ train-test-split strategy where the accuracy on the test dataset was $91.30\%$. Our observations strongly support that high accuracy for COVID-$19$ diagnosis can also be achieved using saliva samples as compared to NP swabs. Given the ease-of-use and the advantages of saliva as a specimen over NP swabs, gargle samples can be employed in clinical settings. Further, we utilized the concepts of explainable-AI and interpreted the RF algorithm using SHAP and feature importance techniques such as permutation feature importance and impurity-based feature importance. By using these interpretation models, we showed both global and local explanations to the model's performance. This step is essential as it lets the domain experts and medical practitioners to understand the mechanisms of the black-box AI models when using in a real-time basis. Our approach will help developing faster, cheaper diagnostic methods for COVID-$19$ which are not only precise but also interpretable and trustworthy.
\section{Acknowledgements}
\label{sec:acknowledgements}
\begin{itemize}
\item This research used computing resources of data, devices and interaction Laboratory (ddiLab) at Northern Illinois University.
\item The authors gratefully acknowledge instrumentation support from Shimadzu Scientific Instruments, Inc.
\end{itemize}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Entanglement is an important quantum resource used in quantum information theory, quantum communication, quantum cryptography, and quantum computation \cite{HHHH}. Therefore, distinguishing between separable and entangled states is of utmost importance. A quantum state represented by a density operator $\rho$ on the Hilbert space $\mathcal{H}_1\otimes\mathcal{H}_2$ is separable if and only if it can be decomposed into $\rho = \sum_k p_k \rho^{(1)}_k \otimes \rho^{(2)}_k$, where $\rho^{(1)}_k$ and $\rho^{(2)}_k$ are density operators of two subsystems and $p_k$ is a probability distribution. It turns out that $\rho$ is separable if and only if $(\oper\otimes\Phi)[\rho]\geq 0$ for any positive map $\Phi$ \cite{Horodeccy}. Any quantum state that violates this condition is therefore entangled. Every entangled state can be detected via an entanglement witness \cite{Terhal1,Terhal2}, which is a block-positive operator with at least one negative eigenvalue. Entanglement witnesses $W$ are related to positive but not completely positive maps $\Phi$ via the Choi-Jamio{\l}kowski isomorphism,
\begin{equation}\label{W}
W=\sum_{i,j=0}^{d-1}|i\>\<j|\otimes\Phi[|i\>\<j|],
\end{equation}
where $|k\>$ is an orthonormal basis in $\mathcal{H}\simeq\mathbb{C}^d$. Any such bipartite operator is Hermitian and positive on separable states ($\< \psi \otimes \phi|W|\psi \otimes \phi\>\geq 0$). Entanglement witnesses define a universal mathematical tool to analyze quantum entangled states. For any entangled state $\rho$, there exists an entanglement witness $W$ (which is not unique) such that $\Tr(W \rho ) <0$ (cf. \cite{Guhne,TOPICAL} for recent reviews).
There are several proposals for constructing such operators. A special class consists in decomposable witnesses, which can be represented as $W=A+B^\Gamma$, where $A,B \geq 0$ and $B^\Gamma$ denotes a partial transposition. However, such witnesses cannot be used to detect bound entanglement; i.e., entangled states that are PPT (positive under partial transposition). Construction of indecomposable witnesses is notoriously hard. There is one class related to the well-known realignment or computable cross-norm (CCNR) separability criterion \cite{R1,R2,R3}, and covariance matrix criterion \cite{CMM1,CMM2,CMM3}. These were recently generalized in Refs. \cite{ESIC-Darek,ESIC-Darek2}. Another interesting class of indecomposable witnesses is constructed in terms of mutually unbiased bases \cite{Spengler,MUBs,Junu1,Junu2,Durt}. This approach was then generalized to mutually unbiased measurements (MUMs) \cite{Kalev} in Refs. \cite{EW-SIC,EW-2MUB}.
In this paper, we further develop the construction of entanglement witnesses (in particular, indecomposable ones) in terms of MUMs, which generalize the notion of mutually unbiased bases (MUBs) to non-projective operators. Contrary to MUBs, of which there are $d+1$ for power prime dimensions $d$ \cite{Wootters,Ivonovic} and at least three otherwise \cite{MUB-2}, one can always construct the maximal number of $d+1$ MUMs. The applications of mutually unbiased measurements range from entropic uncertainty relations \cite{ChenFei,Rastegin,Rastegin2}, through separability criteria for $d$-dimensional bipartite \cite{ChenMa,Shen,ShenLi} and multipartite \cite{Liu,ChenLi} systems, to $k$-nonseparability detection of multipartite qudit systems \cite{Liu3}, and finally to entanglement detection of bipartite states \cite{MUM_purity}. Also, Li et al. used the MUMs to introduce new positive quantum maps and entanglement witnesses \cite{Li} that generalize the construction from Ref. \cite{MUBs}.
In the following sections, we recall the definition of mutually unbiased measurements as well as the method of their construction. Next, we use a set of $N\leq d+1$ MUMs to introduce a family of positive, trace-preserving maps and the corresponding entanglement witnesses. We illustrate our results with several examples. By using mutually unbiased measurement to construct orthonormal Hermitian bases, we prove that these witnesses do not depend on the parameter $\kappa$ that characterizes the MUM properties. However, it turns out that there exists a relation between indecomposability of witnesses and the optimal value of $\kappa$. Finally, we show how our family of entanglement witnesses is related to a large class of witnesses based on the CCNR separability criterion \cite{YuLiu}.
\section{Mutually unbiased measurements}
In quantum information theory, any measurement is represented by a positive, operator-valued measure (POVM) $\{E_\alpha\,|\,E_\alpha\geq 0,\,\sum_\alpha E_\alpha=\mathbb{I}_d\}$. The probability of obtaining the outcome labeled by $\alpha$ is $p_\alpha=\Tr(E_\alpha\rho)$, where $\rho$ is a density operator. A special class of POVMs consists in measurement operators that are orthogonal projectors. Symmetric projective measurements can be performed using mutually unbiased bases. Recall that orthonormal bases $\{\psi_k^{(\alpha)},k=0,\ldots,d-1\}$ in $\mathbb{C}^d$, numbered by $\alpha=1,\ldots,N$, are mutually unbiased if and only if $|\<\psi_k^{(\alpha)}|\psi_l^{(\beta)}\>|^2=1/d$ for $\alpha\neq\beta$. Now, the corresponding rank-1 projectors $P_k^{(\alpha)}=|\psi_k^{(\alpha)}\>\<\psi_k^{(\alpha)}|$, which satisfy the properties
\begin{equation}\label{MUB}
\begin{split}
&\Tr(P_k^{(\alpha)})=1,\\
&\Tr(P_k^{(\alpha)}P_l^{(\beta)})=\delta_{\alpha\beta}\delta_{kl}+\frac 1d (1-\delta_{\alpha\beta}),
\end{split}
\end{equation}
forms a set of $N$ mutually unbiased projective measurements. Kalev and Gour generalized this notion to non-projective measurement operators \cite{Kalev}. Indeed, the measurements $\{P_k^{(\alpha)}|P_k^{(\alpha)}\geq 0,\sum_{k=0}^{d-1}P_k^{(\alpha)}=\mathbb{I}_d\}$ are mutually unbiased if and only if
\begin{equation}\label{MUM}
\begin{split}
&\Tr(P_k^{(\alpha)})=1,\\
&\Tr(P_k^{(\alpha)}P_l^{(\beta)})=\frac 1d +\frac{d\kappa-1}{d-1}\delta_{\alpha\beta}\left(\delta_{kl}
-\frac 1d \right),
\end{split}
\end{equation}
where $1/d<\kappa\leq 1$. For $\kappa=1$, one reproduces eq. (\ref{MUB}). The maximal number of $d+1$ MUMs forms an informationally complete set and can be constructed using an orthonormal basis $\{\mathbb{I}_d/\sqrt{d},G_{\alpha,k}\}$ of traceless Hermitian operators $G_{\alpha,k}$. The relation between $G_{\alpha,k}$ and $P_k^{(\alpha)}$ is given by the formula
\begin{equation}\label{mum}
P_k^{(\alpha)}=\frac 1d \mathbb{I}_d+tF_k^{(\alpha)},
\end{equation}
where
\begin{equation}\label{f}
F_k^{(\alpha)}=\left\{\begin{aligned}
\sum_{l=1}^{d-1}G_{\alpha,l}-\sqrt{d}(\sqrt{d}+1)G_{\alpha,k},&\qquad k\neq 0,\\
(\sqrt{d}+1)\sum_{l=1}^{d-1}G_{\alpha,l},&\qquad k=0.
\end{aligned}\right.
\end{equation}
The parameter $t$ relates to $\kappa$ via
\begin{equation}\label{t}
\kappa=\frac 1d +(d-1)t^2(1+\sqrt{d})^2,
\end{equation}
and it is chosen in such a way that $P_k^{(\alpha)}\geq 0$. The optimal value $\kappa_{\rm opt}$ is the highest possible value of $\kappa$ for which the condition $P_k^{(\alpha)}\geq 0$ holds. Notably, $\kappa_{\rm opt}$ depends on the choice of the operator basis $G_{\alpha,k}$. For example, $\kappa_{\rm opt}=\frac{d+2}{d^2}$ for the Gell-Mann matrices (see Appendix A), and $\kappa_{\rm opt}=1$ for the basis that gives rise to the mutually unbiased basis.
\section{Positive maps and entanglement witnesses}
Let us consider the trace-preserving map
\begin{equation}\label{PTP}
\Phi=\frac{1}{d\kappa-1}\left[(d\kappa-1-N+2L)\Phi_0+\sum_{\alpha=L+1}^N\Phi_\alpha
-\sum_{\alpha=1}^L\Phi_\alpha\right],
\end{equation}
where $\Phi_0[X]=\mathbb{I}_d\Tr(X)/d$ is the completely depolarizing channel and
\begin{equation}
\Phi_\alpha[X]=\sum_{k,l=0}^{d-1}\mathcal{O}^{(\alpha)}_{kl}P_k^{(\alpha)}\Tr(P_l^{(\alpha)}X).
\end{equation}
The maps $\Phi_\alpha$ are constructed from mutually unbiased measurements $P_k^{(\alpha)}$ and orthogonal rotations $\mathcal{O}^{(\alpha)}$ that preserve the vector $\mathbf{n}_\ast=(1,\ldots,1)/\sqrt{d}$. Note that the greater the value of $L$ (i.e., the more we subtract), the higher the coefficient that stands before the identity operator $\mathbb{I}_d$.
\begin{Proposition}
The trace-preserving map defined by eq. (\ref{PTP}) is positive.
\end{Proposition}
\begin{proof}
Take an arbitrary rank-1 projector $P$. We prove that
\begin{equation}\label{pos}
\Tr(\Phi[P])^2\leq\frac{1}{d-1},
\end{equation}
which is a sufficient positivity condition for $\Phi$ \cite{MUBs}.
For simplicity, we consider the map
\begin{equation}
\widetilde{\Phi}=a\Phi_0+\sum_{\alpha=L+1}^N\Phi_\alpha-\sum_{\alpha=1}^L\Phi_\alpha,
\end{equation}
where $a=d\kappa-1-N+2L$ and $\widetilde{\Phi}=(d\kappa-1)\Phi$. Now, let us calculate
\begin{equation}\label{PTP1}
\begin{split}
\Tr(\widetilde{\Phi}[P])^2=\Tr\Bigg\{&a^2\Phi_0[P]^2+\sum_{\alpha,\beta=1}^L
\Phi_\alpha[P]\Phi_\beta[P]+\sum_{\alpha,\beta=L+1}^N\Phi_\alpha[P]\Phi_\beta[P]
+2a\sum_{\alpha=L+1}^N\Phi_0[P]\Phi_\alpha[P]\\&-2a\sum_{\alpha=1}^L\Phi_0[P]\Phi_\alpha[P]
-2\sum_{\alpha=L+1}^N\sum_{\beta=1}^L\Phi_\alpha[P]\Phi_\beta[P]\Bigg\}.
\end{split}
\end{equation}
Observe that the subsequent terms can be simplified as follows,
\begin{equation}
\Tr(\Phi_0[P]^2)=\Tr(\Phi_0[P]\Phi_\alpha[P])=\Tr(\Phi_\alpha[P]\Phi_\beta[P])=\frac 1d,\quad \alpha\neq\beta,
\end{equation}
and
\begin{equation}
\Tr(\Phi_\alpha[P]^2)=\frac{1-\kappa}{d-1}+\frac{d\kappa-1}{d-1}\sum_{m=0}^{d-1}
\left[\Tr(P_m^{(\alpha)}P)\right]^2,
\end{equation}
where we used the trace properties of MUMs from eq. (\ref{MUM}), as well as the properties of the orthogonal rotation matrices,
\begin{equation}
\sum_{k=0}^{d-1}\mathcal{O}^{(\alpha)}_{kl}=\sum_{l=0}^{d-1}\mathcal{O}^{(\alpha)}_{kl}=1,\qquad
\sum_{k=0}^{d-1}\mathcal{O}^{(\alpha)}_{kl}\mathcal{O}^{(\alpha)}_{km}=\delta_{lm}.
\end{equation}
Hence, eq. (\ref{PTP1}) reduces to
\begin{equation}\label{PTP2}
\begin{split}
\Tr(\widetilde{\Phi}[P])^2=&\,\frac 1d \Big[a^2+(N-L)(N-L-1)+L(L-1)+2a(N-L)-2aL
-2L(N-L)\Big]\\&+\frac{1-\kappa}{d-1}N+\frac{d\kappa-1}{d-1}\sum_{\alpha=1}^N\sum_{m=0}^{d-1}
\left[\Tr(P_m^{(\alpha)}P)\right]^2,\\
&=\frac 1d \Big[(a+N-2L)^2-N\Big]+\frac{1-\kappa}{d-1}N+\frac{d\kappa-1}{d-1}\sum_{\alpha=1}^N\sum_{m=0}^{d-1}
\left[\Tr(P_m^{(\alpha)}P)\right]^2
\end{split}
\end{equation}
The mutually unbiased measurements satisfy the following property \cite{Rastegin,ChenFei},
\begin{equation}
\sum_{\alpha=1}^N\sum_{k=0}^{d-1}\left[\Tr\left(P_k^{(\alpha)}P\right)\right]^2
\leq \frac{N-1}{d}+\kappa.
\end{equation}
Applying this inequality to eq. (\ref{PTP2}), together with the definition of $a$, results in
\begin{equation}
\begin{split}
\Tr(\widetilde{\Phi}[P])^2\leq\frac{(d\kappa-1)^2-N}{d} +\frac{1-\kappa}{d-1}N+\frac{d\kappa-1}{d-1}\left(\frac{N-1}{d}+\kappa\right)=
\frac{(d\kappa-1)^2}{d-1},
\end{split}
\end{equation}
which finally proves that condition (\ref{pos}) holds.
\end{proof}
\begin{Remark}
In the proof to Proposition 1, out of all the defining properties of MUMs, the positivity condition $P_k^{(\alpha)}\geq 0$ is the only one that is never used. Hence, one can take $P_k^{(\alpha)}\ngeq 0$ to construct $\Phi$ using eq. (\ref{PTP}), and this map is still positive. In other words, any operators $P_k^{(\alpha)}$ that sum up to the identity and satisfy eq. (\ref{MUM}) for an arbitrary real parameter $\kappa$ give rise to a positive, trace-preserving map $\Phi$.
\end{Remark}
Note that the map $\Phi$ generalizes several positive maps already known in the literature:
\begin{itemize}
\item when no inversions are present ($L=N$), one recovers the map considered in ref. \cite{Li};
\item for $L=N$ and $\kappa=1$, $\Phi$ reduces to the map constructed from MUBs \cite{MUBs};
\item if $L=N=d+1$ and there are no rotations ($\mathcal{O}^{(\alpha)}=\mathbb{I}_d$), one arrives at the maps of the type analyzed in ref. \cite{MUM_GPC};
\item if $L=N=d+1$, $\mathcal{O}^{(\alpha)}=\mathbb{I}_d$, and $\kappa=1$, we obtain the generalized Pauli map \cite{ICQC}.
\end{itemize}
Positive maps find important applications in the theory of quantum entanglement, where they are used to detect entangled (non-separable) states.
From definition in eq. (\ref{W}), we find that the entanglement witness corresponding to the positive map $\Phi$ reads
\begin{equation}\label{W2}
W=\frac{d\kappa-1-N+2L}{d}\mathbb{I}_{d^2}
+\sum_{\alpha=L+1}^NH_\alpha-\sum_{\alpha=1}^LH_\alpha,
\end{equation}
where
\begin{equation}
H_\alpha=\sum_{k,l=0}^{d-1}\mathcal{O}_{kl}^{(\alpha)}
\overline{P}_l^{(\alpha)}\otimes P_k^{(\alpha)}.
\end{equation}
Recall that in any dimension $d$ one can always construct the maximal set of $d+1$ MUMs using an orthonormal basis $\{\mathbb{I}_d/\sqrt{d},G_{\alpha,k}\}$ of traceless Hermitian operators $G_{\alpha,k}$. Hence, whenever one knows the full set of $d+1$ mutually unbiased measurements, the entanglement witness is equivalently given by
\begin{equation}\label{WW}
\widetilde{W}=\frac{d(d-1)(\sqrt{d}+1)^2}{d\kappa-1}W=(d-1)(\sqrt{d}+1)^2
\mathbb{I}_{d^2}+\sum_{\alpha=L+1}^NJ_\alpha-\sum_{\alpha=1}^LJ_\alpha,
\end{equation}
with
\begin{equation}\label{J}
J_\alpha=\sum_{k,l=0}^{d-1}\mathcal{O}_{kl}^{(\alpha)}
\overline{F}_l^{(\alpha)}\otimes F_k^{(\alpha)}.
\end{equation}
In the above equation, we used the one-to-one correspondence between $P_k^{(\alpha)}$ and $G_{\alpha,k}$ found by Kalev and Gour \cite{Kalev} to reverse engineer the Hermitian basis from a known complete set of MUMs.
Observe that there is now no dependence of the witness $\widetilde{W}$ on the parameter $\kappa$. This is due to the fact that $\kappa$ characterizes mutually unbiased measurements and not operator bases.
Now, let us propose several examples of entanglement witnesses that fall into the category established by $\widetilde{W}$.
\begin{Example}
First, let us take the maximal values for $N=L=d+1$. Also, assume that there are no rotations, so that $\mathcal{O}^{(\alpha)}=\mathbb{I}_d$ for $\alpha=1,\ldots,d+1$. Finally, fix the operator basis $G_{\alpha,k}$ to be the Gell-Mann matrices (see Appendix A). In this case, we have
\begin{equation}
\sum_{\alpha=1}^dJ_\alpha=d(\sqrt{d}+1)^2\left[\sum_{k,m=0}^{d-1}|k\>\<m|\otimes|k\>\<m|
-\sum_{k=0}^{d-1}|k\>\<k|\otimes|k\>\<k|\right],
\end{equation}
\begin{equation}
J_{d+1}=(\sqrt{d}+1)^2\left[d\sum_{k=0}^{d-1}|k\>\<k|\otimes|k\>\<k|-\mathbb{I}_{d^2}\right].
\end{equation}
This allows us to write the formula for $\widetilde{W}$ from eq. (\ref{WW}) in the form
\begin{equation}
\widetilde{W}=d(\sqrt{d}+1)^2\left[\mathbb{I}_{d^2}-dP_+\right],
\end{equation}
where $P_+=\frac 1d \sum_{i,j=0}^{d-1}|i\>\<j|\otimes|i\>\<j|$ is the maximally entangled state. This is exactly the entanglement witness corresponding to the reduction map \cite{EW_reduction}.
\end{Example}
\begin{Example}\label{Ex2}
Now, consider the dimension $d=3$ and take, as in the previous example, $N=L=4$ as well as $\mathcal{O}^{(\alpha)}=\mathbb{I}_3$ for $\alpha=1,2,3$. However, assume that the final rotation matrix $\mathcal{O}^{(4)}=S_k$ describes a permutation, where
\begin{equation}
S_1=\begin{pmatrix}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{pmatrix},\qquad
S_2=\begin{pmatrix}
0 & 1 & 0 \\
0 & 0 & 1 \\
1 & 0 & 0
\end{pmatrix}.
\end{equation}
Denote the witness corresponding to the choice $\mathcal{O}^{(4)}=S_k$ by $\widetilde{W}_k$. For the Gell-Mann matrices, one finds
\begin{equation}
\widetilde{W}_1=6(2+\sqrt{3})\left[\begin{array}{c c c|c c c|c c c}
1 & \cdot & \cdot & \cdot & -1 & \cdot & \cdot & \cdot & -1 \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
-1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & -1 \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot \\
-1 & \cdot & \cdot & \cdot & -1 & \cdot & \cdot & \cdot & 1
\end{array}\right],\qquad
\widetilde{W}_2=6(2+\sqrt{3})\left[\begin{array}{c c c|c c c|c c c}
1 & \cdot & \cdot & \cdot & -1 & \cdot & \cdot & \cdot & -1 \\
\cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
-1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & -1 \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
-1 & \cdot & \cdot & \cdot & -1 & \cdot & \cdot & \cdot & 1
\end{array}\right],
\end{equation}
which belong to the Choi-type maps analyzed in ref. \cite{HaKye2}. For clarity, all zeros are represented by dots.
The same witnesses can also be obtained from the mutually unbiased bases \cite{MUBs}.
\end{Example}
Unfortunately, these simple forms of entanglement witnesses constructed from the Gell-Mann matrices are not preserved for $d>3$. However, this can be remedied if one modifies the diagonal matrices in the Gell-Mann basis.
\begin{Example}\label{EX}
In what follows, we generalize Example \ref{Ex2} to an arbitrary finite dimension $d$. Once again, we take the maximal $N=L=d+1$, $\mathcal{O}^{(\alpha)}=\mathbb{I}_d$ for $\alpha=1,\ldots,d$, and the permutation matrix $\mathcal{O}^{(d+1)}=S^{(r)}$, $S^{(r)}|i\>=|i+r\>$. Instead of the Gell-Mann matrices, in the construction of the entanglement witness, we use the operator basis introduced in Appendix B. Due to this change, $J_{d+1}$ can be simplified to
\begin{equation}
J_{d+1}=(\sqrt{d}+1)^2\left[d\sum_{k=0}^{d-1}|k-r\>\<k-r|\otimes|k\>\<k|-\mathbb{I}_{d^2}\right].
\end{equation}
Now, the associated witness $\widetilde{W}$ is given by
\begin{equation}
\widetilde{W}=d(\sqrt{d}+1)^2\left[\mathbb{I}_{d^2}-dR\right],
\end{equation}
where
\begin{equation}
R=\frac 1d \left[\sum_{k,m=0}^{d-1}|k\>\<m|\otimes|k\>\<m|-\sum_{k=0}^{d-1}
\big(|k\>\<k|-|k-r\>\<k-r|\big)\otimes|k\>\<k|\right].
\end{equation}
\end{Example}
Interestingly, even though the parameter $\kappa$ that characterizes mutually unbiased measurements is not present in the formula for $\widetilde{W}$,
the properties of entanglement witnesses depend on the optimal (maximal) value of $\kappa$. Indeed, for the MUMs constructed from the Gell-Mann matrices with $\kappa_{\rm opt}=\frac{d+2}{d^2}$, there are less indecomposable witnesses than for the MUBs, where $\kappa_{\rm opt}=1$.
\begin{Example}
In dimension $d=3$, assuming the maximal value of $N=4$ and no rotations ($\mathcal{O}^{(\alpha)}=\mathbb{I}_3$ for $\alpha=1,\ldots,4$), there are five indecomposable witnesses that can be constructed from the mutually unbiased bases. Some examples for $L=2$ are
\begin{equation}
\widetilde{W}_1=2(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
\cdot & \cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & 1 \\
\cdot & 3 & \cdot & \cdot & \cdot & -2 & -2 & \cdot & \cdot \\
\cdot & \cdot & 3 & -2 & \cdot & \cdot & \cdot & -2 & \cdot \\
\hline
\cdot & \cdot & -2 & 3 & \cdot & \cdot & \cdot & -2 & \cdot \\
1 & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 1 \\
\cdot & -2 & \cdot & \cdot & \cdot & 3 & -2 & \cdot & \cdot \\
\hline
\cdot & -2 & \cdot & \cdot & \cdot & -2 & 3 & \cdot & \cdot \\
\cdot & \cdot & -2 & -2 & \cdot & \cdot & \cdot & 3 & \cdot \\
1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & \cdot
\end{array}\right],
\qquad
\widetilde{W}_2=2(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
4 & \cdot & \cdot & \cdot & -1 & \cdot & \cdot & \cdot & -1 \\
\cdot & 1 & \cdot & \cdot & \cdot & 2 & 2 & \cdot & \cdot \\
\cdot & \cdot & 1 & 2 & \cdot & \cdot & \cdot & 2 & \cdot \\
\hline
\cdot & \cdot & 2 & 1 & \cdot & \cdot & \cdot & 2 & \cdot \\
-1 & \cdot & \cdot & \cdot & 4 & \cdot & \cdot & \cdot & -1 \\
\cdot & 2 & \cdot & \cdot & \cdot & 1 & 2 & \cdot & \cdot \\
\hline
\cdot & 2 & \cdot & \cdot & \cdot & 2 & 1 & \cdot & \cdot \\
\cdot & \cdot & 2 & 2 & \cdot & \cdot & \cdot & 1 & \cdot \\
-1 & \cdot & \cdot & \cdot & -1 & \cdot & \cdot & \cdot & 4
\end{array}\right],
\end{equation}
which can be used to detect the positive partial transpose (PPT) states
\begin{equation}
\rho_1=\frac{1}{24}
\left[\begin{array}{c c c|c c c|c c c}
4 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & 1 \\
\cdot & 2 & \cdot & \cdot & \cdot & 2 & 2 & \cdot & \cdot \\
\cdot & \cdot & 2 & 2 & \cdot & \cdot & \cdot & 2 & \cdot \\
\hline
\cdot & \cdot & 2 & 2 & \cdot & \cdot & \cdot & 2 & \cdot \\
1 & \cdot & \cdot & \cdot & 4 & \cdot & \cdot & \cdot & 1 \\
\cdot & 2 & \cdot & \cdot & \cdot & 2 & 2 & \cdot & \cdot \\
\hline
\cdot & 2 & \cdot & \cdot & \cdot & 2 & 2 & \cdot & \cdot \\
\cdot & \cdot & 2 & 2 & \cdot & \cdot & \cdot & 2 & \cdot \\
1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & 4
\end{array}\right],\quad
\rho_2=\frac{1}{3(3+\sqrt{3})}
\left[\begin{array}{c c c|c c c|c c c}
\sqrt{3}-1 & \cdot & \cdot & \cdot & \sqrt{3}-1 & \cdot & \cdot & \cdot & \sqrt{3}-1 \\
\cdot & 2 & \cdot & \cdot & \cdot & -1 & -1 & \cdot & \cdot \\
\cdot & \cdot & 2 & -1 & \cdot & \cdot & \cdot & -1 & \cdot \\
\hline
\cdot & \cdot & -1 & 2 & \cdot & \cdot & \cdot & -1 & \cdot \\
\sqrt{3}-1 & \cdot & \cdot & \cdot & \sqrt{3}-1 & \cdot & \cdot & \cdot & \sqrt{3}-1 \\
\cdot & -1 & \cdot & \cdot & \cdot & 2 & -1 & \cdot & \cdot \\
\hline
\cdot & -1 & \cdot & \cdot & \cdot & -1 & 2 & \cdot & \cdot \\
\cdot & \cdot & -1 & -1 & \cdot & \cdot & \cdot & 2 & \cdot \\
\sqrt{3}-1 & \cdot & \cdot & \cdot & \sqrt{3}-1 & \cdot & \cdot & \cdot & \sqrt{3}-1
\end{array}\right],
\end{equation}
respectively.
On the contrary, all witnesses that arise from the Gell-Mann basis are decomposable, including
\begin{equation}
\widetilde{W}_3=6(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 1 \\
\cdot & 1 & \cdot & -1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & -1 & \cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & 1 & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & 1 & \cdot \\
1 & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot
\end{array}\right],
\qquad
\widetilde{W}_4=2(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
4 & \cdot & \cdot & \cdot & -3 & \cdot & \cdot & \cdot & \cdot \\
\cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & 1 & \cdot & \cdot & \cdot & 3 & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
-3 & \cdot & \cdot & \cdot & 4 & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & 3 & \cdot \\
\hline
\cdot & \cdot & 3 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 3 & \cdot & 1 & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 4
\end{array}\right].
\end{equation}
Recall that a witness $W$ is decomposable if it can be written as $W=A+B^\Gamma$, where $A,B\geq 0$ and $\Gamma$ denotes the partial transposition with respect to the second subsystem. In our example, $\widetilde{W}_3$ can be decomposed into $\widetilde{W}_3=A_3+B_3^\Gamma$ with the positive operators
\begin{equation}
A_3=6(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & 1 & \cdot & -1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & -1 & \cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & 1 & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & 1 & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot
\end{array}\right],\qquad
B_3=6(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & 1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & \cdot & 1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot
\end{array}\right],
\end{equation}
whereas $\widetilde{W}_4$ is decomposable into $\widetilde{W}_4=A_4+B_4^\Gamma$
with positive
\begin{equation}
A_4=2(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & 1 & \cdot & -1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & 1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot \\
\hline
\cdot & -1 & \cdot & 1 & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & 1 & \cdot \\
\hline
\cdot & \cdot & 1 & \cdot & \cdot & \cdot & 1 & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & 1 & \cdot & 1 & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot
\end{array}\right],\qquad
B_4=2(2+\sqrt{3})
\left[\begin{array}{c c c|c c c|c c c}
4 & \cdot & \cdot & \cdot & -2 & \cdot & \cdot & \cdot & 2 \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
-2 & \cdot & \cdot & \cdot & 4 & \cdot & \cdot & \cdot & 2 \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
2 & \cdot & \cdot & \cdot & 2 & \cdot & \cdot & \cdot & 4
\end{array}\right].
\end{equation}
\end{Example}
Now, observe that $J_\alpha$ from eq. (\ref{J}) can be expressed directly through the elements of the operator basis $G_{\alpha,k}$, as the operators $F_k^{(\alpha)}$ depend directly on $G_{\alpha,k}$. Indeed, after writing out $F_k^{(\alpha)}$ using eq. (\ref{f}), one arrives at
\begin{equation}
J_\alpha=\sum_{k,l=1}^{d-1}\mathcal{Q}_{kl}^{(\alpha)}\overline{G}_{\alpha,l}
\otimes G_{\alpha,k}
\end{equation}
with
\begin{equation}
\mathcal{Q}_{kl}^{(\alpha)} = d(\mathcal{O}_{00}^{(\alpha)}-1) + d(\sqrt{d}+1)^2 \mathcal{O}_{kl}^{(\alpha)} - d(\sqrt{d}+1) (\mathcal{O}_{0l}^{(\alpha)}+\mathcal{O}_{k0}^{(\alpha)}).
\end{equation}
In the above formula, $\mathcal{Q}^{(\alpha)}$ are rescaled orthogonal matrices since $\mathcal{Q}^{(\alpha)T}\mathcal{Q}^{(\alpha)}
=\mathcal{Q}^{(\alpha)}\mathcal{Q}^{(\alpha)T}=d^2(\sqrt{d}+1)^4\mathbb{I}_{d-1}$. Assume for now that $N=L=d+1$. Then, the corresponding entanglement witness $\widetilde{W}$ is given by
\begin{equation}
\widetilde{W}=d(\sqrt{d}+1)^2\left[\mathbb{I}_{d^2}- d(\sqrt{d}+1)^2 G_0\otimes G_0-\frac{1}{d(\sqrt{d}+1)^2}\sum_{\alpha=1}^{d+1}J_\alpha\right],
\end{equation}
where $G_0=\mathbb{I}_d/\sqrt{d}$. After a simple relabelling of indices, $(\alpha,k)\longmapsto\mu$, it follows that
\begin{equation}
\widetilde{W}=d(\sqrt{d}+1)^2\left[\mathbb{I}_{d^2}-\sum_{\mu,\nu=0}^{d^2-1}Q_{\mu\nu} G_\mu^T\otimes G_\nu\right]
\end{equation}
with a block-diagonal orthogonal matrix
\begin{equation}
Q=\frac{1}{d(\sqrt{d}+1)^2}
\left[\begin{array}{c c c c c}
d(\sqrt{d}+1)^2 & & & & \\
& \mathcal{Q}^{(1)} & & & \\
& & \mathcal{Q}^{(2)} & & \\
& & & \ddots & \\
& & & & \mathcal{Q}^{(d+1)}
\end{array}\right].
\end{equation}
Therefore, the entanglement witnesses constructed from $d+1$ MUMs belong to a larger class of witnesses \cite{YuLiu}
\begin{equation}\label{WW2}
W^\prime=\mathbb{I}_{d^2}-\sum_{\mu,\nu=0}^{d^2-1}Q_{\mu\nu} G_\mu^T\otimes G_\nu
\end{equation}
defined with the use of any orthonormal Hermitian bases $G_\mu$ and an orthogonal matrix $Q_{\mu\nu}$. Actually, it is enough that $Q^TQ\leq\mathbb{I}_{d^2}$, so the above consideration is also true for both $N\leq d+1$ and $L\leq d+1$ if one allows for the rotation matrices $\widetilde{\mathcal{O}}^{(\alpha)}$ to change the sign of $\mathbf{n}_\ast$ ($\widetilde{\mathcal{O}}^{(\alpha)}\mathbf{n}_\ast=\pm\mathbf{n}_\ast$).
\section{Conclusions}
In this paper, we constructed a family of positive, trace-preserving maps using $N\leq d+1$ mutually unbiased measurements and orthogonal matrices. We showed that these maps give rise to entanglement witnesses regardless of the value of the parameter $\kappa$, which means that the positivity of mutually unbiased measurements is not required. We provided several interesting examples of witnesses for $d=3$ as well as an arbitrary finite dimension using the construction of MUMs from two different orthonormal Hermitian bases. We also showed that there is a relation between indecomposability of witnesses and the optimal value of $\kappa$. At last, we proved that our construction belongs to the family of witnesses based on the CCNR separability criterion \cite{YuLiu}, where the Hermitian basis consists in the identity and traceless operators, and the orthogonal matrix is block-diagonal. It would be interesting to provide a multipartite generalization of this construction.
\section{Acknowledgements}
This paper was supported by the Polish National Science Centre project No. 2018/30/A/ST2/00837. K.S. was also supported by the Foundation for Polish Science (FNP).
|
\section*{Acknowledgements}
This research was partially supported by the National Science Foundation (NSF) via the grant numbers: 2008837, 2007210, 1755946, I2040950 and 2006889.
\section{Conclusion Remarks}
\noindent \textbf{Summary.}
In this paper, we study the problem of improving the training efficiency of reinforced feature selection (RFS). We propose a traverse strategy to simplify the multi-agent formulation of the RFS to a single-agent framework, an implementation of Monte Carlo method under the framework, and two strategies to improve the efficiency of the framework.
\noindent\textbf{Theoretical Implications.}
The single-agent formulation reduces the requirement of computational resources, the early stopping strategy improves the training efficiency, the decision history based traversing strategy diversify the training process, and the interactive reinforcement learning accelerates the training process without changing the optimal policy.
\noindent \textbf{Practical Implications.}
Experiments show that the Monte Carlo method with the traverse strategy can significantly reduce the hardware occupation in practice, the decision history based traverse strategy can improve performance of the traverse strategy, the interactive reinforcement learning can improve the training of the framework.
\noindent \textbf{Limitations and Future Work.}
Our method can be further improved from the following aspects: 1) The framework can be adapted into a parallel framework, where more than one (but much smaller than the feature number) agents work together to finish the traverse; 2) Besides reward level, the interactive reinforcement learning can obtain advice from action level and sampling level.
3) The framework can be implemented on any other reinforcement learning frameworks, e.g., deep Q-network, actor critic and proximal policy optimization (PPO).
\iffalse
In this paper, we study the problem of improving the training efficiency of reinforced feature selection (RFS). We propose a traverse strategy to simplify the multi-agent formulation of the RFS to single-agent, which greatly reduces the computational burden and make it more applicable in real-world scenarios. We develop an off-policy Monte Carlo method and propose an early stopping criteria to im-prove the training efficiency. We design extensive experiments and we find that: (1) The proposed single-agent framework can significantly alleviate the high computational resource occupation; (2) The proposed early stopping criteria can not only improve the training efficiency, but also enhance the predictive accuracy; (3) The early stopping criteria is sensitive to the pre-defined threshold, and the threshold differs from different datasets. (4) The proposed $\epsilon$-greedy behavior policy can help
to improve the performance of the Monte Carlo reinforced feature selection method.
\fi
\section{Experimental Results}
We conduct extensive experiments on real-world datasets to study: (1) the overall performance of early stopping Monte Carlo based reinforced feature selection (\textbf{ES-MCRFS}); (2) the training efficiency of the early stopping criteria; (3) the sensitivity of the threshold in the early stopping criteria; (4) the computational burden of the traverse strategy; (5) the decision history based traverse strategy; (6) the behavior policy in the ES-MCRFS.
\subsection{Experimental Setup}
\subsubsection{Data Description}
We use four publicly available datasets on classification task to validate our methods, i.e., Forest Cover (FC) dataset \cite{forest_cover_dataset}, Spambase (Spam) dataset \cite{Dua_2019}, Insurance Company Benchmark (ICB) dataset \cite{van2000coil} and Arrhythmia (Arrhy) dataset \cite{647926}. The statistics of the datasets are in Table \ref{dataset}.
\begin{table}[h]
\begin{center}
\caption{Statistics of datasets.}
\label{dataset}
\begin{tabular}{|p{1.0cm}|p{1.0cm}|p{1.0cm}|p{1.0cm}|p{1.2cm}|}
\hline
& FC &Spam &ICB & Arrhy \\\hline
Features& 54& 57& 86 & 274 \\ \hline
Samples& 15120& 4601& 5000 & 452 \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{table*}[!th]
\centering
\caption{Overall performance.}
\label{overall_perf}
\setlength{\tabcolsep}{5mm}{
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|}
\hline
\multicolumn{2}{|l|}{\multirow{2}{*}{}} & \multicolumn{2}{c|}{FC} & \multicolumn{2}{c|}{Spam} & \multicolumn{2}{c|}{ICB} & \multicolumn{2}{c|}{Arrhy} \\ \cline{3-10}
\multicolumn{2}{|l|}{} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} \\ \hline
\multirow{7}{*}{\begin{sideways}{{Algorithms}}\end{sideways}} & K-Best & 0.7904 & 0.8058 & 0.9207 & 0.8347 & 0.8783 & 0.8321 & 0.6382 & 0.6406 \\ \cline{2-10}
& LASSO & 0.8438 & 0.8493 & 0.9143 & 0.8556 & 0.8801 & 0.8507 & 0.6293 & 0.6543 \\ \cline{2-10}
& GFS & 0.8498 & 0.8350 & 0.9043 & 0.8431 & 0.9099 & 0.637 & 0.6406 & 0.6550 \\ \cline{2-10}
& mRMR & 0.8157 & 0.8241 & 0.8980 & 0.8257 & 0.8998 & 0.8423 & 0.6307 & 0.6368 \\ \cline{2-10}
& RFE & 0.8046 & 0.8175 & 0.9351 & 0.8480 & 0.9045 & 0.8502 & 0.6452 & 0.6592 \\ \cline{2-10}
& MARFS & 0.8653 & 0.8404 & 0.9219 & 0.8742 & 0.8902 & 0.8604 & 0.7238 & 0.6804 \\ \cline{2-10}
& MCRFS & 0.8688 & 0.8496 & 0.9256 & 0.8738 & 0.8956 & 0.8635 & 0.7259 & 0.7152 \\ \cline{2-10}
& ES-MCRFS & \textbf{0.8942} & \textbf{0.8750} & \textbf{0.9402} & \textbf{0.9067} & \textbf{0.9187} & \textbf{0.8803} & \textbf{0.7563} & \textbf{0.7360} \\ \hline
\end{tabular}
}
\end{table*}
\subsubsection{Evaluation Metrics}
In the experiments, we have classification as the downstream task for feature selection problem, therefore we use the two most popular evaluation metrics for classification task:
\noindent\textbf{Accuracy} is given by $Acc = \frac{TP+TN}{TP+TN+FP+FN}$, where $TP,TN, FP, FN$ are true positive, true negative, false positive and false negative for all classes.
\noindent\textbf{F1-score} is given by $F1=\frac{2*P*R}{P+R}$, where $P=\frac{TP}{TP+FP}$ is precision and $R=\frac{TP}{TP+FN}$ is recall.
\subsubsection{Baseline Algorithms}
We compare our proposed ES-MCRFS method with the following baselines: (1) \textbf{K-Best} ranks features by unsupervised scores with the label and selects the top $k$ highest scoring features \cite{yang1997comparative}. In the experiments, we set $k$ equals to half of the number of input features.(2) \textbf{LASSO} conducts feature selection via $l1$ penalty \cite{tibshirani1996regression}. The hyper parameter in LASSO is its regularization weight $\lambda$ which is set to 0.15 in the experiments.
(3) \textbf{GFS} selects features by calculating the fitness level for each feature to generate better feature subsets via crossover and mutation \cite{leardi1996genetic}.
(4) \textbf{mRMR} ranks features by minimizing feature’s redundancy and maximizing their relevance with the label\cite{peng2005feature}.
(5) \textbf{RFE} selects features by recursively selecting smaller and smaller feature subsets \cite{granitto2006recursive}.(6)\textbf{MARFS} is a multi-agent reinforcement learning based feature selection method \cite{liu2019automating}. It uses $M$ feature agents to control the selection/deselection of the $M$ features. Besides, we also compare our method with its variant without early stopping strategy, i.e., Monte Carlo based reinforced feature selection \textbf{MCRFS}.
\subsubsection{Implementation}
In the experiments, for all deep networks, we set mini-batch size to 16 and use AdamOptimizer with a learning rate of 0.01. For all experience replays, we set memory size to 200. We set the $Q$ network in our methods as a two-layer ReLU with 64 and 8 nodes in the first and second layer. The classification algorithm we use for evaluation is a random forest with 100 decision trees. The stop time is set to 3000 steps. The state representation method in reinforced feature selection is an auto-encoder method whose encoder/decoder network is a two-layer ReLU with 128 and 32 nodes in the first and second layer.
\subsubsection{Environmental Setup} The experiments were carried on a server with an I9-9920X 3.50GHz CPU, 128GB memory and a Ubuntu 18.04 LTS operation system.
\subsection{Overall Performance}
We compare the proposed ES-MCRFS method with baseline methods and its variant with regard to the predictive accuracy. As Table \ref{overall_perf} shows, the MCRFS, which simplify the reinforced feature selection into a single-agent formulation, achieves similar performance with the multi-agent MARFS. With the help of traverse strategy and early stopping criteria, the ES-MCRFS outperforms all the other methods.
\subsection{Sensitivity Study of Early Stopping Criteria}
We study the threshold sensitivity in the early stopping criteria by differing the threshold $v$ and evaluate the predictive accuracy. Figure \ref{thresh_sens} shows that the optimal threshold for the four datasets are 0.4, 0.5, 0.7, 0.7. It reveals that the early stopping criteria is sensitive to the pre-defined threshold, and the optimal threshold varies on different datasets.
\begin{figure}[htbp]
\centering
\subfigure[FC]{
\includegraphics[width=4.1cm]{figures/experiments_1/threshold/FC.eps}
}
\hspace{-5mm}
\subfigure[Spam]{
\includegraphics[width=4.1cm]{figures/experiments_1/threshold/Spam.eps}
}
\vspace{-4mm}
\subfigure[ICB]{
\includegraphics[width=4.1cm]{figures/experiments_1/threshold/ICB.eps}
}
\hspace{-5mm}
\subfigure[Arrhy]{
\includegraphics[width=4.1cm]{figures/experiments_1/threshold/Arrhy.eps}
}
\caption{Threshold sensitivity of early stopping criteria.}
\label{thresh_sens}
\end{figure}
\subsection{Training Efficiency of Early Stopping Criteria}
We compare the predictive accuracy with different numbers of training episodes to study the training efficiency of the early stopping. Figure \ref{effiency_early_stop} shows that with early stopping criteria, the Monte Carlo reinforced feature selection can achieve convergence more quickly, and the predictive accuracy can be higher after convergence.
\begin{figure}[p]
\centering
\subfigure[FC]{
\includegraphics[width=4.1cm]{figures/experiments_1/effiency_stopping_criteria/FC.eps}
}
\hspace{-5mm}
\subfigure[Spam]{
\includegraphics[width=4.1cm]{figures/experiments_1/effiency_stopping_criteria/Spam.eps}
}
\vspace{-4mm}
\subfigure[ICB]{
\includegraphics[width=4.1cm]{figures/experiments_1/effiency_stopping_criteria/ICB.eps}
}
\hspace{-5mm}
\subfigure[Arrhy]{
\includegraphics[width=4.1cm]{figures/experiments_1/effiency_stopping_criteria/Arrhy.eps}
}
\caption{Predictive accuracy on training step.}
\label{effiency_early_stop}
\end{figure}
\subsection{Study of the Behavior Policy}
We study the difference between random behavior policy and the $\epsilon$-greedy policy presented in Equation \ref{pro_b}. We combine the two policies with MCRFS and ES-MCRFS respectively. Figure \ref{behaviro_policy} shows that the $\epsilon$-greedy policy outperforms the random behavior policy on all datasets.
\begin{figure}[p]
\centering
\subfigure[FC]{
\includegraphics[width=4cm]{figures/experiments_1/behavior_policy/FC.eps}
}
\hspace{-4mm}
\subfigure[Spam]{
\includegraphics[width=4cm]{figures/experiments_1/behavior_policy/Spam.eps}
}
\vspace{-4mm}
\subfigure[ICB]{
\includegraphics[width=4.1cm]{figures/experiments_1/behavior_policy/ICB.eps}
}
\hspace{-4mm}
\subfigure[Arrhy]{
\includegraphics[width=4.1cm]{figures/experiments_1/behavior_policy/Arrhy.eps}
}
\caption{Predictive accuracy on different training strategies. RB for random behavior policy and GB for $\epsilon$-greedy behavior policy.}
\label{behaviro_policy}
\end{figure}
\subsection{Computational Burden of Traverse Strategy}
We compare the computational burden of the MCRFS which uses single-agent and the traverse strategy to substitute the multi-agent strategy in the MARFS. Table \ref{CPU_Mem_Occ} shows that the CPU and memory cost when implementing the two methods. Our method MCRFS requires less computational resources than the multi-agent MARFS.
\begin{table}[h]
\centering
\scriptsize
\caption{ \small CPU and memory (in MB) occupation.}
\label{CPU_Mem_Occ}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{} & \multicolumn{2}{c|}{FC} & \multicolumn{2}{c|}{Spam} & \multicolumn{2}{c|}{ICB} & \multicolumn{2}{c|}{Arrhy} \\ \cline{2-9}
& CPU & Mem & CPU & Mem & CPU & Mem & CPU & Mem \\ \hline
MARFS & 72\% & 1531 & 75\% & 1502 & 86\% & 1797 & 97\% & 4759 \\ \hline
MCRFS & 57\% & 1429 & 54\% & 1395 & 59\% & 1438 & 55\% & 1520 \\ \hline
\end{tabular}
\end{table}
\subsection{Decision History Based Traverse Strategy}
We study the decision history based traverse strategy by comparing its performance with the vanilla traverse strategy on ES-MCRFS. Table \ref{Deci_His} shows that the decision history can significantly improve performance of the traverse strategy.
\begin{table}[h]
\setlength{\arraycolsep}{1.5pt}
\scriptsize
\centering
\caption{\small Traverse strategy ablation. DH for decision history.}
\label{Deci_His}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{} & \multicolumn{2}{c|}{FC} & \multicolumn{2}{c|}{Spam} & \multicolumn{2}{c|}{ICB} & \multicolumn{2}{c|}{Arrhy} \\ \cline{2-9}
& Acc & F1 & Acc & F1 & Acc & F1 & Acc & F1 \\ \hline
\multicolumn{1}{|l|}{No DH} & \multicolumn{1}{l|}{0.75} & \multicolumn{1}{l|}{0.82} & \multicolumn{1}{l|}{0.83} & \multicolumn{1}{l|}{0.79} & \multicolumn{1}{l|}{0.75} & \multicolumn{1}{l|}{0.81} & \multicolumn{1}{l|}{0.59} & \multicolumn{1}{l|}{0.56} \\ \hline
With DH & 0.89 & 0.88 & 0.94 & 0.91 & 0.92 & 0.88 & 0.76 & 0.74 \\ \hline
\end{tabular}
\end{table}
\begin{table*}[t!]
\centering
\caption{Performance with different utility function}
\label{utility_func}
\setlength{\tabcolsep}{4mm}{
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|}
\hline
\multicolumn{2}{|l|}{\multirow{2}{*}{}} & \multicolumn{2}{c|}{FC} & \multicolumn{2}{c|}{Spam} & \multicolumn{2}{c|}{ICB} & \multicolumn{2}{c|}{Arrhy} \\ \cline{3-10}
\multicolumn{2}{|l|}{} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} & \multicolumn{1}{c|}{Acc} & \multicolumn{1}{c|}{F1} \\ \hline
\multirow{3}{*}{\begin{sideways}{{Utility}}\end{sideways}} & $Rd$ & 0.8689 & 0.8433 & 0.9250 & 0.8749 & 0.8997 & 0.8788 & 0.7340 & 0.6955 \\ \cline{2-10}
& $Rv$ & 0.8703 & 0.8507 & 0.9317 & 0.8831 & 0.9001 & 0.8793 & 0.7393 & 0.7143 \\ \cline{2-10}
& $Rv-Rd$ & \textbf{0.8842} & \textbf{0.8650} & \textbf{0.9402} & \textbf{0.8949} & \textbf{0.9117} & \textbf{0.8903} & \textbf{0.7492} & \textbf{0.7258} \\ \hline
\end{tabular}
}
\end{table*}
\vspace{-0.1cm}
\subsection{Training Efficiency of reward-level interactive strategy}
We compare the predictive accuracy with different numbers of training episodes to study the training efficiency of the reward-level interactive (RI) strategy. Figure \ref{reward_interactive} shows that with RI, the Monte Carlo reinforced feature selection can achieve convergence more quickly. However, as the ES-MCRFS already achieves good performance, the RI can not improve its final performance.
\begin{figure}[h]
\centering
\subfigure[FC]{
\includegraphics[width=4.1cm]{figures/experiments_1/RI/FC.eps}
}
\hspace{-5mm}
\subfigure[Spam]{
\includegraphics[width=4.1cm]{figures/experiments_1/RI/Spam.eps}
}
\vspace{-4mm}
\subfigure[ICB]{
\includegraphics[width=4.1cm]{figures/experiments_1/RI/ICB.eps}
}
\hspace{-5mm}
\subfigure[Arrhy]{
\includegraphics[width=4.1cm]{figures/experiments_1/RI/Arrhy.eps}
}
\caption{Predictive accuracy on training step.}
\label{reward_interactive}
\end{figure}
\vspace{-0.1cm}
\subsection{Study of the Utility Function}
We define the utility function $\mathcal{U}$ as the combination of relevance ($Rv$) function and redundancy ($Rd$) function. Here we study the impact of the two components for the utility function. Table \ref{utility_func} shows that when we use $Rv$ independently as the utility function, its performance is better than the $Rd$. This is because $Rv$ evaluates the relationship between features and the label, which is directly related to the classification task, while $Rd$ evaluates the relationship among features, which is an indirect evaluation to the classification task. The combination of the two functions ($Rv-Rd$) as the utility function significantly outperforms each of the independent functions, revealing the $Rd$ and $Rv$ coordinate and make up each other's shortage.
\section{Introduction}
In general data mining and machine learning pipelines, before proceeding with machine learning tasks, people need to preprocess the data first. Preprocessing technologies include data cleaning, data transformation and feature engineering. As one of the most important feature engineering technique, feature selection aims to select the optimal feature subset from the original feature set for the downstream task.
Traditional feature selection methods can be categorized into three families: (i) filter methods, in which features are ranked by a specific score (e.g., univariate feature selection \cite{yang1997comparative,forman2003extensive}, correlation based feature selection \cite{hall1999feature,yu2003feature});
(ii) wrapper methods, in which optimal feature subset is identified by a search strategy that collaborates with predictive tasks (e.g., evolutionary algorithms \cite{yang1998feature,kim2000feature}, branch and bound algorithms \cite{narendra1977branch,kohavi1997wrappers}); (iii) embedded methods, in which feature selection is part of the optimization objective of predictive tasks (e.g., LASSO \cite{tibshirani1996regression}, decision tree \cite{sugumaran2007feature}). However, these studies have shown not just strengths but also some limitations.
For example, filter methods ignore the feature dependencies and interactions between feature selection and predictors.
Wrapper methods have to directly search a very large feature space of $2^N$ feature subspace candidates, where $N$ is the number of features.
Embedded methods are subject to the strong structured assumptions of predictive models, i.e., in LASSO, the non-zero weighted features are considered to be important.
\begin{figure}[htbp]
\centering
\subfigure[Feature subspace candidates]{
\includegraphics[width=4.1cm]{figures/feature_subsapce.eps}
}
\hspace{-5mm}
\subfigure[Reinforced feature selection]{
\includegraphics[width=4.1cm]{figures/reinforced_feature_selection.eps}
}
\caption{Reinforced feature selection explores the feature subspace by assigning each feature one agent, and the agent's policy decides the selection of its corresponding feature.}
\label{fig_FSC_RFS}
\end{figure}
Recently, reinforcement learning has been incorporated with feature selection and produces an emerging feature selection method, called reinforced feature selection \cite{liu2019automating,fan2020autofs}. In the reinforced feature selection, there are multiple agents to control the selection of features, one agent for one feature. All agents cooperate to generate the optimal feature set. It has been proved to be
superior to traditional feature selection methods due to its powerful global search ability. However, each agent adapts a neural network as its policy. Since the agent number equals the feature number ($N$ agents for $N$ features), when the feature set is extremely large, we need to train a large number of neural networks, which is computationally high and not applicable for large-scale datasets. Our research question is: Can we design a more practical and efficient method to address the feature selection problem while preserving the effectiveness of reinforced fearure selection? To answer this questions, there are three challenges.
The first challenge is to reformulate the feature selection problem with smaller number of agents. Intuitively, we can define the action of the agent as the selected feature subset. For a given feature set, we input it to the agent's policy and the agent can directly output the optimal subset. However, the feature subset space is as large as $2^N$, where $N$ is the feature number. When the dataset is large, the action space is too large for the agent to explore directly
To tackle this problem, we design a traverse strategy, where one single agent visit each feature one by one to decide its selection (to select or deselect). After traversing all the feature set, we can obtain the selected feature subset.
We adapt the off-policy Monte Carlo method to our framework. In the implementation, we design two policies, i.e., one behavior policy and one target policy. The behavior policy is to generate the training data and the target policy is to generate the final feature subset. In each training iteration, we use the behavior policy to traverse the feature set, and generates one training episode. The training episode consists of a series of training samples, each of which contains the state, the action and the reward. Similar with \cite{liu2019automating}, we regard the selected feature subset as the environment, and its representation as the state. The action 1/0 denotes selection/deselection, and the reward is composed of predictive accuracy, feature subset relevance and feature subset redundancy. Using the training episode, we evaluate the target policy by calculating its Q value with importance sampling, and improve it by the Bellman equation. After more and more iterations, the target policy becomes better and better. After the training is done, we use the target policy to traverse the feature set and can derive the optimal feature subset. Besides, the behavior policy is supposed to cover the target policy as much as possible so as to generate more high-quality training data, and should introduce randomness to enable exploration \cite{sutton2018reinforcement}. We design an $\epsilon$-greedy behavior policy, to better balance the coverage and the diversity.
The second challenge is to improve the training efficiency of the proposed traverse strategy. In this paper, we improve the efficiency from two aspects. One improvement is to conduct the importance sampling in an incremental way, which saves repeated calculations between samples. In the off-policy Monte Carlo method, since the reward comes from the behavior policy, when we use it to evaluate the target policy, we need to multiply it by an importance sampling weight. We decompose the sampling weight into an incremental format, where the calculation of the sampling weight can directly use the result of previous calculations. The other improvement is to propose an early stopping criteria to assure the quality of training samples as well as stopping the meaningless traverse by behavior policy. In Monte Carlo method, if the behavior policy is too far away from the target policy, the samples from the behavior policy are considered harmful to the evaluation of the target policy. As the traverse method is continuous and the importance sampling weight calculation depends on the previous result, once the sample at time $t$ is skew, the following samples are skew. We propose a stopping criteria based on the importance sampling weight, and re-calculate a more appropriate weight to make the samples from the behavior policy more close to the target policy.
The third challenge is how to improve the training efficiency by external advice. In classic interactive reinforcement learning, the only source of reward is from the environment, and the advisor does have access to the reward function. However, in many cases, the advisor can not give direct advice on action, but can evaluate the state-action pair. In this paper, we define a utility function $\mathcal{U}$ which can evaluate state-action pair and provide feedback to the agent just like the environment reward does.
When integrating the advisor utility $\mathcal{U}$ with the environment reward $\mathcal{R}$ to a more guiding reward $\mathcal{R}^\prime$, we should not change the optimal policy, namely the optimal policy guided by $\mathcal{R}^\prime$ should be identical to the optimal policy guided by $\mathcal{R}$. In this paper, we propose a state-based reward integration strategy, which leads to a more inspiring integrated reward as well as preserving the optimal policy.
To summarize, the contributions of this paper are: (1) We reformulate the reinforced feature selection into a single-agent framework by proposing a traverse strategy; (2) We design an off-policy Monte Carlo method to implement the proposed framework; (3) We propose an early stopping criteria to improve the training efficiency. (4) We propose a reward-level interactive strategy to improve the training efficiency. (5) We design extensive experiments to reveal the superiority of the proposed method.
\section{Proposed Method}
In this section, we first propose a single-agent Monte Carlo based reinforced feature selection method. And then, we propose an episode filtering method to improve the sampling efficiency of the Monte Carlo method. In addition, we apply the episode filtering Mote Carlo method to the reinforced feature selection scenario. Finally, we design a reward shaping strategy to improve the training efficiency.
\begin{figure*}[!th]
\centering
\includegraphics[width=18cm]{figures/single_agent_RFS.eps}
\vspace{-2.5mm}
\caption{Single-agent Reinforced feature selection with traverse strategy. At each step, the agent transverses \\ features one by one to decide their selection. The traverse data are stored in the memory to form a training episode. }
\vspace{-3mm}
\label{fig_SRFS}
\end{figure*}
\subsection{Monte Carlo Based Reinforced Feature Selection}\label{MCRFS}
The MARFS method has proved its effectiveness, however, the multi-agent strategy greatly increases the computational burden and hardware cost. Here, we propose a single-agent traverse strategy and use Mote Carlo method as the reinforcement learning algorithm.
\subsubsection{Traverse strategy}
As Figure \ref{fig_SRFS} shows, rather than using $N$ agents to select their corresponding features in the multi-agent strategy, we design one agent to traverse all features one at a time.
In the $i$-th episode, beginning from time $t=1$, the behavior policy $b^i$ firstly decides the selection decision (select or not select) for feature 1, and then, at time $t=2$, $b^i$ decides the selection decision for feature 2. With time going on, the features are traversed one by one, and the selected features forms a selected feature subset $\mathcal{F}^i_t$. Meanwhile, this process also generates an episode $\mathbf{x}^i_N = \{x^i_1, x^i_2, \dots, x^i_t, \dots x^i_N\}$, where $x^i_t = (s^i_t, a^i_t, r^i_t)$ is a tuple of state, action and reward. The action $a^i_t = 1/0$ is the selection/deselection decision of the $t$-th feature, the next state $s^i_{t+1}$ is derived by $represent(\mathcal{F}^i_t)$ and the reward $r^i_t$ is derived by $eval(\mathcal{F}^i_t)$.
\subsubsection{Monte Carlo Method for Reinforced Feature Selection}
With the episode generated by the behavior policy $b^i$, we can evaluate our target policy $\pi^i$ and improve $\pi^i$. Both the behavior policy $b^i$ and the target policy $\pi^i$ provide the probability of taking action $a$ given a specific state $s$.
Specifically, We generate an episode $\mathbf{x}^i_N$ by $b^i$. Then, we calculate the accumulated reward by:
\begin{equation}
G^i(s^i_t, a^i_t) = \sum_{j=0}^t \gamma^{(t-j)} \cdot r^i_j
\end{equation}
where $0 \leq \gamma \leq 1$ is a discount factor.
As the state space is extremely large, we use a neural network $Q(s, a)$ to approximate $G(s, a)$.
The target policy $\pi$ is different from the behavior policy $b$, and the reward comes from samples derived from policy $b$, therefore the accumulated reward of $\pi$ should be calculated by multiplying an importance sampling weight:
\begin{equation}\label{overall_weight}
\rho_t^i = \frac{\Pi_{j=0}^{t}\pi^i(a^i_j|s^i_j)}{\Pi_{j=0}^{t}b^i(a^i_j|s^i_j)}
\end{equation}
The $Q_{\pi^i}(s, a)$ can be optimized by minimizing the loss:
\begin{equation}\label{loss_pi}
\mathcal{L}_{\pi^i} = || Q_{\pi^i}(s^i_t, a^i_t) - \rho_t^i*G^i(s^i_t, a^i_t)||^2
\end{equation}
The probability of taking action $a$ for state $s$ under policy $\pi$ in the next iteration can be calculated by:
\begin{equation}\label{pro_pi}
\pi^{i+1}\{a|s\} = \frac{exp(Q_{\pi^i}\{a, s\})}{exp(Q_{\pi^i}\{a = 0, s\}) + exp(Q_{\pi^i}\{a = 1, s\})}
\end{equation}
We develop an $\epsilon$-greedy policy of $b$ based on the Q value from $\pi$:
\begin{equation}\label{pro_b}
b^{i+1}\{a|s\}=
\begin{cases}
1-\epsilon& a = argmax_a Q_{\pi^i}(s,a);\\
0 & otherwise;
\end{cases}
\end{equation}
Algorithm \ref{MCRFS} shows the process of Monte Carlo based feature selection (MCRFS) with traverse strategy.
\begin{algorithm}[t]
\caption{Monte Carlo Based Reinforced Feature Selection with Traverse Strategy}
\label{MCRFS}
\LinesNumbered
\KwIn{Feature set $\mathcal{F} = \{f_1, f_2, ..., f_N \}$, downstream machine learning task $\mathcal{T}$.}
\KwOut{Optimal feature subset $\mathcal{F}^\prime$.}
Initialize the behavior policy $b^1$, target policy $\pi^1$, exploration number $M$, $\mathcal{F}^\prime = \Phi$.
\For{$i=1$ to $M$}
{
Initialize state $s^i_1$.\\
\For{$t=1$ to $N$}
{
Derive action $a^i_t$ with behavior policy $b^i(s^i_t)$.\\
Perform $a^i_t$, getting selected feature subset $\mathcal{F}^i_t$.\\
Obtain the next state $s^i_{t+1}$ by $represent(\mathcal{F}^i_t)$ and reward $r^i_t$ by $eval(\mathcal{F}^i_t)$.\\
}
Update target policy $\pi^{i+1}$ by Equation \ref{pro_pi} and behavior policy $b^{i+1}$ by Equation \ref{pro_b}.\\
\If{$eval(\mathcal{F}^i_N)>eval(\mathcal{F}^\prime)$}{$\mathcal{F}^\prime = \mathcal{F}^i_N$.}
}
Return $\mathcal{F}^\prime$.
\end{algorithm}
\subsection{Early Stopping Monte Carlo Based Reinforced Feature Selection} \label{EFMC}
In many cases, the feature set size $N$ can be very large, meaning that there can be a large number of samples in one episode $\mathbf{x}^i_N$. The problem is, if the sample at time $T$ is bad (the Chi-squared distance between $b^i(s^i_t)$ and $\pi^i(s^i_t)$ is large), all the subsequent samples (from $T$ to $N$) in the episode are skew \cite{maceachern1999sequential}. The skew samples not only are a waste time time to generate, but also do harm to the policy evaluation, therefore we need to find some way to stop the sampling when the episode becomes skew.
\subsubsection{Incremental Importance Sampling}
Rather than calculating the importance sampling weight for each sample directly by Equation \ref{overall_weight}, we here decompose it into an incremental format. Specifically, in the $i$-th iteration, we define the weight increment:
\begin{equation}
w_t^i = \frac{\pi^i(a^i_t|s^i_t)}{b^i(a^i_t|s^i_t)}
\end{equation}
and the importance sampling weight can be calculated by:
\begin{equation}
\rho_t^i = \rho_{t-1}^i \cdot w_t^i
\end{equation}
Thus, at each time, we just need to calculate a simple increment to update the weight.
\subsubsection{Early Stopping Monte Carlo Method for Reinforced Feature Selection}
We first propose the stopping criteria, and then propose a decision history based traversing strategy to enhance diversity.
\noindent \textbf{\textsl{Early stopping criteria}}. We stop the traverse by probability:
\begin{equation}\label{stop_pro}
p_t^i = max(0, 1 -\rho_t^i/v)
\end{equation}
where $0 \leq v \leq 1$ is the stopping threshold.
And for the acquired episode, we recalculate the importance sampling weight for each sample by:
\begin{equation}\label{pvm}
w_t^i = p_v^i\cdot \rho_t^i/p_t^i
\end{equation}
where the $p_v$ can be calculated by:
\begin{equation}
p_v^i = \int max (0, 1 -\rho_t^i/v) b^i(\mathbf{s}_t) \; d\mathbf{s}_t
\end{equation}
As $p_v^i$ is identical for all samples in the $i$-th episode regardless of $t$, the calculation of Equation \ref{pvm} does almost no increase to the computation.
\noindent \textbf{\textsl{Decision history based traversing strategy}}. In the $i$-th iteration, the stopping criteria stops the traverse at time $t$, and the features after $t$ are not traversed. With more and more traverses, the front features (e.g., $f_1$ and $f_2$) are always selected/deselected by the agent, while the backside features (e.g., $f_N$ and $f_{N-1}$) get very few opportunity to be decided. To tackle this problem, we record the decision times we made on each feature, and re-rank their orders to diversify the decision process in the next traverse episode. For example, in the past $5$ episodes, if the decision times of feature set \{$f_1$, $f_2$, $f_3$\} are \{5,2,4\}, then in the 6-th episode, the traverse order is $f_2 \rightarrow f_3 \rightarrow f_1$.
Algorithm \ref{MCRFS_ES} shows the process of Monte Carlo based feature selection (MCRFS) with early stopping traverse strategy. specifically, we implement the early stopping Monte Carlo based reinforced feature selection method as follows:
\begin{itemize}
\item[1.] Use a random behavior policy $b^0$ to traverse the feature set. Stop the traverse with the probability in Equation \ref{stop_pro} and get an episode $\mathbf{x}_{N_0}^0$.
\item [2.] Evaluate the policy $\pi^0$ to get the Q value $Q^0$ by minimizing Equation \ref{loss_pi}, and derive the updated policy $\pi^1$ and $b^1$ from Equation \ref{pro_pi} and \ref{pro_b} respectively.
\item [3.] Update the record of traverse times for each feature. Re-rank feature order. The smaller times one feature was traversed, the more forward order it should get.
\item[4.] Use the updated policy $\pi^1$ and $b^1$ to traverse the re-ranked feature set for the next $M$ steps. Derive the policy $\pi^M$ and $b^M$. Use $\pi^M$ to traverse the feature set without stopping criteria, and derive the final feature subset.
\end{itemize}
\begin{algorithm}[t]
\caption{Monte Carlo Based Reinforced Feature Selection with Early Stopping Traverse Strategy}
\label{MCRFS_ES}
\LinesNumbered
\KwIn{Feature set $\mathcal{F} = \{f_1, f_2, ..., f_N \}$, downstream machine learning task $\mathcal{T}$.}
\KwOut{Optimal feature subset $\mathcal{F}^\prime$.}
Initialize the behavior policy $b^1$, target policy $\pi^1$, exploration number $M$, $\mathcal{F}^\prime = \Phi$.
\For{$i=1$ to $M$}
{
Initialize state $s^i_1$.\\
Rank features with their decision history. \\
\For{$t=1$ to $N$}
{
Derive action $a^i_t$ with behavior policy $b^i(s^i_t)$.\\
Perform $a^i_t$, getting selected feature subset $\mathcal{F}^i_t$.\\
Obtain the next state $s^i_{t+1}$ by $represent(\mathcal{F}^i_t)$ and reward $r^i_t$ by $eval(\mathcal{F}^i_t)$.\\
Break the loop with probability $p_t^i$ derived from Equation \ref{stop_pro};
}
Update target policy $\pi^{i+1}$ by Equation \ref{pro_pi} and behavior policy $b^{i+1}$ by Equation \ref{pro_b}.\\
\If{$eval(\mathcal{F}^i_N)>eval(\mathcal{F}^\prime)$}{$\mathcal{F}^\prime = \mathcal{F}^i_N$.}
}
Return $\mathcal{F}^\prime$.
\end{algorithm}
\subsection{Interactive Reinforcement Learning}
As all the steps in this section belong to the same iteration, we omit the superscript $i$ in each denotation for simplicity.
Reinforcement learning is proposed to develop the optimal policy $\pi_\mathcal{M}^*(s) = argmax_{a} Q_{\mathcal{M}}^*(s,a)$ for an MDP $\mathcal{M}$.
The optimal $Q$-value can be updated by Bellman equation \cite{bellman1957dynamic}:
\begin{equation}\label{Bellman_Equ}
Q_{\mathcal{M}}^*(s_t,a_t) = \mathcal{R}(s_t, a_t) + \gamma * max_{a_{t+1}}Q_{\mathcal{M}}^*(s_{t+1},a_{t+1})
\end{equation}
Interactive reinforcement learning (IRL) is proposed to accelerate the learning process of reinforcement learning (RL) by providing external action advice to the RL agent \cite{suay2011effect}. As Figure \ref{fig_act_IRL} shows, for selected advising states, the action of RL agent is decided by the advisor's action advice instead of its own policy. The algorithm to select advising states varies with the problem setting. Typical algorithms for selecting advising states include early advising, importance advising, mistake advising and predictive advising \cite{torrey2013teaching}. To better evaluate the utility of the state-action pair $(s_t, a_t)$, we define a utility function $\mathcal{U}(s_t, a_t)$. The utility function can give a feedback of how the action benefits from the state from the advisor's point of view.
\begin{figure}[h]
\centering
\includegraphics[width=8.5cm]{figures/Action_Level_IRL.eps}
\caption{Classic interactive reinforcement learning. The advisor gives the agent advice at the action level.}
\label{fig_act_IRL}
\end{figure}
\noindent \textbf{\textsl{Reward-Level Interactive Reinforcement Learning}}.
In reinforcement learning (RL), we aim to obtain the optimal policy for the MDP $\mathcal{M} = \{\mathcal{S},\mathcal{A},\mathcal{R}, \gamma, \mathcal{P}\}$. However, in IRL, when we change the reward function $\mathcal{R}$ to a more inspiring reward function $\mathcal{R}^\prime$, the original MDP $\mathcal{M}$ is changed to a new MDP $\mathcal{M}^\prime = \{\mathcal{S},\mathcal{A},\mathcal{R}^\prime, \gamma, \mathcal{P}\}$. Without careful design , the optimal policy derived from $\mathcal{M}^\prime$ would be different from the optimal policy for $\mathcal{M}$.
Here we give a universal form of reward advice without limitation on the form of utility function $\mathcal{U}(s,a)$:
\begin{equation}\label{reward_advice}
\mathcal{R}^\prime (a_t,s_t) = \mathcal{R}(a_t,s_t) + c * (\gamma * \mathcal{U}(s_{t+1}) - \mathcal{U}(s_t))
\end{equation}
where $\mathcal{U}(s_t) = E_{a_t}[\mathcal{U}(a_t, s_t)]$, $c$ is the weight to balance the proportion of the utility function.
We prove that the optimal policies of $\mathcal{M}$ and $\mathcal{M}^\prime$ are identical when the reward advice is Equation \ref{reward_advice}:
We firstly subtract $c * \mathcal{U}(s_t)$ from both sides of Equation \ref{Bellman_Equ}, and we have:
\begin{equation}
\begin{split}
Q_{\mathcal{M}}^*(s_t,a_t) &- c * \mathcal{U}(s_t)= \mathcal{R}(s_t, a_t) \\
&+\gamma * max_{a_{t+1}}Q_{\mathcal{M}}^*(s_{t+1},a_{t+1}) - c * \mathcal{U}(s_t)
\end{split}
\end{equation}
\begin{algorithm}[t]
\caption{Reward-Level Interactive Reinforcement Learning}
\label{alg_reward_level_IRL}
\LinesNumbered
Initialize replay memory $\mathcal{D}$; Initialize the $Q$-value function with random weights; Initialize the advising state number $N_a$, stop time $T$;
\For{$t=1$ to $T$}
{
$a_t=
\begin{cases}
\text{random action}& \text{with probability} \epsilon;\\
max_{a_t}Q(s_t, a_t)& \text{with probability} 1-\epsilon;
\end{cases}$
Perform $a_t$, obtaining reward $\mathcal{R}(a_t,s_t)$ and next state $s_{t+1}$;
$\mathcal{R}^\prime (s_t,a_t)=
\begin{cases}
\mathcal{R}(s_t,a_t)& t> N_a;\\
\mathcal{R}(s_t,a_t) + c * (\gamma * \mathcal{U}(s_{t+1}) - \mathcal{U}(s_t))& t \leq N_a;
\end{cases}$
Store transition $(s_t, a_t, \mathcal{R}^\prime (s_t,a_t), s_{t+1})$ in $\mathcal{D}$;
Randomly sample mini-batch of data from $\mathcal{D}$;
Update $Q(s,a)$ with the sampled data;
}
\end{algorithm}
We add and subtract $c* \gamma * \mathcal{U}(s_{t+1})$ on the right side:
\begin{equation}\label{subtract_two_sides}
\begin{split}
&Q_{\mathcal{M}}^*(s_t,a_t) - c * \mathcal{U}(s_t)= \mathcal{R}(s_t, a_t)\\
&+\gamma * max_{a_{t+1}}Q_{\mathcal{M}}^*(s_{t+1},a_{t+1}) - c * \mathcal{U}(s_t)\\
&+ c* \gamma * \mathcal{U}(s_{t+1}) - c* \gamma * \mathcal{U}(s_{t+1})\\
&=\mathcal{R}(s_t, a_t) + c* \gamma * \mathcal{U}(s_{t+1})- c * \mathcal{U}(s_t) \\
&+\gamma * max_{a_{t+1}}[Q_{\mathcal{M}}^*(s_{t+1},a_{t+1}) - c * \mathcal{U}(s_{t+1})]\\
\end{split}
\end{equation}
We define:
\begin{equation}\label{def_of_partial}
Q_{\mathcal{M}^\prime}^{\partial}(s_t,a_t) = Q_{\mathcal{M}}^*(s_t,a_t) - c * \mathcal{U}(s_t)
\end{equation}
Then Equation \ref{subtract_two_sides} has the new form:
\begin{equation}
\begin{split}
&Q_{\mathcal{M}^\prime}^{\partial}(s_t,a_t)= \mathcal{R}(s_t, a_t) + c* [\gamma * \mathcal{U}(s_{t+1})- \mathcal{U}(s_t)] \\
&+\gamma * max_{a_{t+1}}Q_{\mathcal{M}^\prime}^{\partial}(s_{t+1},a_{t+1})\\
& = \mathcal{R^\prime}(s_t, a_t)+\gamma * max_{a_{t+1}}Q_{\mathcal{M}^\prime}^{\partial}(s_{t+1},a_{t+1})
\end{split}
\end{equation}
which is the Bellman equation of $Q_{\mathcal{M}^\prime}^{\partial}(s_t,a_t)$ with reward $\mathcal{R}^\prime$, meaning $Q_{\mathcal{M}^\prime}^{\partial}(s_t,a_t)$ is the optimal policy $Q$-value for MDP $\mathcal{M}^\prime$, i.e.,
\begin{equation}\label{partial_eqv_opt}
Q_{\mathcal{M}^\prime}^*(s_t,a_t) = Q_{\mathcal{M}^\prime}^{\partial}(s_t,a_t)
\end{equation}
We combine Equation \ref{def_of_partial} and Equation \ref{partial_eqv_opt} and have:
\begin{equation}
Q_{\mathcal{M}^\prime}^*(s_t,a_t) = Q_{\mathcal{M}}^*(s_t,a_t) - c * \mathcal{U}(s_t)
\end{equation}
Obviously,
\begin{equation}
\begin{split}
argmax_{a_t} Q_{\mathcal{M}^\prime}^*(s_t,a_t) & = argmax_{a_t}[Q_{\mathcal{M}}^*(s_t,a_t) - c * \mathcal{U}(s_t)]\\
& = argmax_{a_t}Q_{\mathcal{M}}^*(s_t,a_t)
\end{split}
\end{equation}
which reveals the optimal policy of MDP $\mathcal{M}^\prime$ with reward $\mathcal{R}^\prime$ is identical to the optimal policy of MDP $\mathcal{M}$ with reward $\mathcal{R}$.
As the reward advice $\mathcal{R}^\prime$ consists of more information than the original reward $\mathcal{R}$, it can help the reinforcement learning agent explore the environment more efficiently. We give a detailed description of reward-level IRL in Algorithm \ref{alg_reward_level_IRL}. Specifically, we adapt the early advising strategy \cite{torrey2013teaching} to select the advising states, i.e., the advisor gives advice for the first $n$ states the IRL agent meets.
\begin{figure}[t]
\centering
\includegraphics[width=8.5cm]{figures/Reward_Level_IRL.eps}
\caption{Reward-level interactive reinforcement learning. The advisor gives advice at the reward level.}
\label{fig_overview_IRL}
\end{figure}
\subsection{Comparison with Prior Literature}
Compared with filter methods, our methods capture feature interactions; Compared with wrapper methods, our methods reduce the search space; Compared with embedded methods, our methods don't rely on strong structured assumptions; Compared with multi-agent reinforcement learning feature selection, our methods achieve parallel performance with lower computational cost.
\section{Preliminaries}\label{Preliminaries}
We first introduce some preliminary knowledge about the Markov decision process(MDP) and the Monte Carlo method to solve MDP, then we give a brief description of multi-agent reinforced feature selection.
\subsection{Markov Decision Process}
Markov decision process (MDP) is defined by a tuple $\mathbf{M} = \{\mathcal{S},\mathcal{A},\mathcal{R}, \gamma, \mathcal{P}\}$, where state space $\mathcal{S}$ is finite, action space $\mathcal{A}$ is pre-defined, reward function $\mathcal{R}: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R} $ is a mapping function from state-action pair to a scalar, $\gamma \in [0,1]$ is a discount factor and $\mathcal{P}: \mathcal{S} \times \mathcal{A} \times \mathcal{S} \rightarrow \mathbb{R}$ is the transition probability from state-action pair to the next state. In this paper, we study the most popular case when the environment is deterministic and thus $\mathcal{P} \equiv 1$.
We use superscripts to discriminate different episode, and use subscripts to denote the time step inside the episode, e.g., $s^i_t$, $a^i_t$ denote the state and action at time $t$ in the $i$-th episode.
\subsection{Monte Carlo for Solveing MDP}
Monte Carlo method can take samples from the MDP to evaluate and improve its policy. Specifically, at the $i$-th iteration, with a behavior (sampling) policy $b^i$, we can derive an episode $\mathbf{x}^i = \{x^i_1, x^i_2, \dots, x^i_t, \dots x^i_N\}$, where $x^i_t = (s^i_t, a^i_t, r^i_t)$ is a sample consisting state, action and reward. With the episode, we can evaluate the Q value $Q_{\pi^i}(s^i_t, a^i_t)$ over our policy (detailed in Section \ref{MCRFS}), and improve it by Bellman optimality:
\begin{equation}\label{Bellman_Eq}
\pi^{i+1}(s) = argmax_a Q_{\pi^i}(s, a)
\end{equation}
With the evaluation-improvement process going on, the policy $\pi$ becomes better and better, and can finally converge to the optimal policy. The general process is:
\begin{equation}\label{improve}
\pi^0 \stackrel{E}{\rightarrow} Q_{\pi^0} \stackrel{I}{\rightarrow} \pi^1 \stackrel{E}{\rightarrow} Q_{\pi^1} \stackrel{I}{\rightarrow} \dots \pi^M \stackrel{I}{\rightarrow} Q_{\pi^M}
\end{equation}
where $\stackrel{E}{\rightarrow}$ denotes the policy evaluation and $\stackrel{I}{\rightarrow}$ denotes the policy improvement. After $M$ iterations, we can achieve an optimal policy.
As Equation \ref{improve} shows, the policy evaluation and improvement need many iterations, and each iteration needs one episode $\mathbf{x}$ ($\mathbf{x}^i$ for the $i$-th iteration) consisting $N$ samples.
\subsection{Multi-Agent Reinforced Feature Selection}
Feature selection aims to find an optimal feature subset $\mathcal{F}^\prime$ from the original feature set $\mathcal{F}$ for a downstream machine learning task $\mathcal{M}$.
Recently, the emerging multi-agent reinforced feature selection (MARFS) method \cite{liu2019automating} formulates the feature selection problem into a multi-agent reinforcement learning task, in order to automate the selection process.
As Figure \ref{fig_MARL} shows, in the MARFS method, each feature is assigned to a feature agent, and the action of feature agent decides to select/deselect its corresponding feature. It should be noted that the agents simultaneously select features, meaning that there is only one time step inside an iteration, and thus we omit the subscript here. At the $i$-th iteration, all agents cooperate to select a feature subset $\mathcal{F}^i$. The next state $s^{i+1}$ is derived by the representation of selected feature subset $\mathcal{F}^i$:
\begin{equation}
s^{i+1} = represent(\mathcal{F}^i)
\end{equation}
where $\mathcal{F}^i$ is the selected feature subset at time $t$. $represent$ is a representation learning algorithm which converts the dynamically changing $\mathcal{F}^i_t$ into a fixed-length state vector $s^{i+1}$. The $represent$ method can be meta descriptive statistics, autoencoder based deep representation and dynamic-graph based GCN in \cite{liu2019automating}.
The reward $r^i$ is an evaluation of the selected feature subset $\mathcal{F}^i$:
\begin{equation}
r^i = eval(\mathcal{F}^i)
\end{equation}
where $eval$ is evaluations of $\mathcal{F}^i$, which can be a supervised metric with the machine learning task $\mathcal{F}$ taking $\mathcal{F}^i$ as input, unsupervised metrics of $\mathcal{F}^i$, or the combination of supervised and unsupervised metrics in \cite{liu2019automating}. The reward is assigned to each of the feature agent to train their policies. With more and more steps' exploration and exploitation, the policies become more and more smart, and consequently they can find better and better feature subsets.
\begin{figure}[t]
\centering
\includegraphics[width=6cm]{figures/MAFS_overview.eps}
\caption{Multi-agent Reinforced feature selection. Each feature is controlled by one feature agent.}
\label{fig_MARL}
\end{figure}
\section{Related Work}\label{related_work}
\noindent \textbf{Efficient Sampling in Reinforcement Learning.}
Reinforcement learning is a trial-and-error based method, which requires high-quality samples to train its policy. It is always a hot topic to pursue efficient sampling for reinforcement learning. One research direction is to generate training samples with high quality based on the importance sampling technology,
such as rejection control \cite{liu1998rejection} and marginalized importance sampling \cite{xie2019towards}.
These methods basically control the sampling process based on the importance sampling weight. Another research direction is to sample diversified sample from different policy parameters. The diversity partially contributes to the exploration and thus have better performance on some specific tasks \cite{fortunato2017noisy}.
However, these methods suffer from slow convergence and no theoretical guarantee \cite{yu2018towards}. Besides, there are other attempts to develop sample efficient reinforcement learning, such as curiosity-driven exploration and hybrid optimization \cite{raginsky2017non,wang2020defending}.
\noindent \textbf{Feature Selection.}
Feature selection can be categorized into three types, i.e., filter methods, wrapper methods and embedded methods \cite{liu2021automated,zhao2020simplifying}.
Filter methods rank features only by relevance scores and only top-ranking features are selected. The representative filter methods is the univariate feature selection \cite{forman2003extensive}
The representative wrapper methods are branch and bound algorithms \cite{narendra1977branch,kohavi1997wrappers}.
Wrapper methods are supposed to achieve better performance than filter methods since they search on the whole feature subset space. Evolutionary algorithms \cite{yang1998feature,kim2000feature} low down the computational cost but could only promise local optimum results.
Embedded methods combine feature selection with predictors more closely than wrapper methods. The most widely used embedded methods are LASSO \cite{tibshirani1996regression} and decision tree \cite{sugumaran2007feature}.
\noindent \textbf{Interactive Reinforcement Learning.}
Interactive reinforcement learning (IRL) is proposed to accelerate the learning process of reinforcement learning. Early work on the IRL topic can be found in \cite{lin1991programming}, where the authors presents a general approach to making robots which can improve their performance from experiences as well as from being taught. Unlike the imitation learning which intends to learn from an expert other than the environment \cite{schaal1999imitation,ho2016generative}, IRL sticks to learning from the environment and the advisor is only an advice-provider in its apprenticeship \cite{knox2013teaching,wang2021reinforced}.
As the task for the advisor is to help the agent pass its apprenticeship, the advisor has to identify which states belong to the apprenticeship. In \cite{torrey2013teaching}, the authors study the advising state selection and propose four advising strategies, i.e., early advising, importance advising, mistake correcting and predictive advising.
|
\section{Introduction}
A rotation-powered pulsar converts most of its rotational energy losses into a highly relativistic magnetized wind \citep[e.g.][]{GJ69,KC84a}. The collision of the pulsar wind with the ambient supernova (SN) ejecta and/or interstellar material (ISM) results in a termination shock and creates a pulsar wind nebula (PWN), a
region of up to tens of pc, filled with relativistic electrons and positrons \citep{Rees74, Reynolds84}. It is widely believed that lower energy radiation in the PWNe is produced by the synchrotron radiation of leptons in the magnetic
field and the higher energy
part of the spectrum is produced by leptons in the inverse Compton scattering of the ambient photons \citep[e.g.][]{KC84b, deJager92, Atoyan96}
The Crab nebula is a unique representative of PWNe, characterized by a very broad spectral energy distribution (SED) that spans over 21 decades, from MHz radio wavelengths to ultra high-energy (UHE, $E>100\,$TeV)
gamma-rays. It is powered by the most energetic pulsar PSR~J0534+2200 (or the Crab pulsar) found in our Galaxy with a current spindown luminosity of $L_s=4.5\times10^{38}\rm erg~s^{-1}$. The pulsar has a characteristic age of $\tau_c=1260$\,years while the record in Chinese chronicles shows its true age to be 967\,years \citep{Lundmark1921}. In the Crab nebula, several photon fields serve as targets for the inverse Compton (IC) radiation of electrons \citep{Atoyan96}. Three
dominant IC components are contributed by the far-infrared
(FIR), cosmic microwave background (CMB) radiation , and synchrotron photons. Up-scattering of
synchrotron photons through the synchrotron-self-Compton
(SSC) channel provides the major contribution at TeV energies. The synchrotron target is however characterized by
a relatively high photon energy, therefore in the UHE band
the SSC process is significantly suppressed because of the
Klein-Nishina (KN) effect and the IC scattering process of CMB photons is the dominant component.
The remarkable discovery of PeV photons from Crab Nebula recently by the Large High Altitude Air Shower Observatory (LHAASO) is the first step towards the opening of the PeV window in the cosmic electromagnetic spectrum \citep{LHAASO21_sci,LHAASO21_nat}.
Although the current data of of Crab nebula measured by LHAASO
is consistent with a single log-parabola type spectrum with a photon index close to $-3.7$ at PeV energy, the highest-energy spectrum shows a possible hardening, implying the existence of a second spectral component of either leptonic or hadronic origin as discussed in \citet{LHAASO21_sci}. A two-component scenario with a main electron component and a secondary proton component has been proposed to explain the whole spectrum of the Crab Nebula\citep{LHAASO21_sci}.
In fact, it has been suggested that a fraction of the spin-down power
of pulsars can be converted into a wind of nuclei \citep{Hoshino92, Arons94, Gallant94}. Nuclei can be accelerated in pulsar magnetospheres as discussed by \citet{ChengKS86, Bednarek97}. These nuclei may suffer partial photo-disintegration in the non-thermal radiation fields of the pulsar's outer magnetosphere. The products (protons and neutrons) of the photo-disintegration and surviving heavier nuclei are injected into the PWN, and then interact with the ambient matter. Thus, $\gamma$-rays are produced via inelastic collisions, and this hadronic origin of gamma-rays has been suggested as an alternative or
an additional component to the IC gamma-rays in the literature \citep[see e.g.][]{ChengKS90, Atoyan96, Aharonian98, Bednarek97, Horns06, YangX09,ZhangL09, LiH10}.
Recently, \citet{ZhangX20} obtained an upper limit of 0.5\% on the energy fraction of relativistic protons contained in the Crab Nebula by modelling the broadband spectrum including the data of the Tibet AS$\gamma$ experiment \citep{ASgamma19}. It may be worth noting that, without considering the escape of protons from the nebula, the constraint does not reflect the true fraction of the energy channelled into relativistic protons from the pulsar wind. \citet{LHAASO21_sci} found that to account for PeV emission via the hadronic process, it needs a power of $10^{36}\rm erg~s^{-1}$ for 10\,PeV protons and an order of
magnitude more for a broad $E^{-2}$ spectrum assuming the most effective confinement for protons. This power is about 0.2-2\% of the pulsar's current spindown power.
In this paper, we will study the proton content in pulsar wind using the LHAASO data of the Crab Nebula and considering the diffusive escape effect. The rest of the paper is organized as follows. We study the constraint on the proton content in the Crab Nebula in Section 2. In Section 3, we will discuss the implications of the result and give a summary in Section 4.
\section{The effect of diffusive escape on the hadronic emission}
The possible hardening of the Crab Nebula spectrum at PeV energies could, in principle, be either leptonic or hadronic. In this section, we will focus on the hadronic scenario, while a brief discussion on the leptonic scenario is presented in the discussion section.
In the hadronic scenario, the PeV photons from the Crab nebula arise from $pp$ collisions with the matter in the nebula. The energy of protons that produce $E_\gamma=1\,$PeV is $E_p=10E_\gamma=10\,$PeV. The
collision loss time for relativistic protons interacting with the matter in the nebula is
\begin{equation}
t_{\rm pp}=\frac{1}{\xi_{\rm pp} n \sigma_{\rm pp} c}=10^7{\rm yr} \left(\frac{n}{\rm 10 cm^{-3}}\right)^{-1} \left(\frac{\xi_{\rm pp}}{0.2}\right)^{-1},
\end{equation}
where $n$ is the average gas density in the nebula, $\sigma_{\rm pp}$ is the cross section for proton-proton interaction and $\xi_{\rm pp}$ is the inelasticity of pion production. The mass in the Crab
filaments is estimated to be $7.2\pm0.5 M_\odot$ \citep{Owen15}, so the gas density is $10 \,{\rm cm^{-3}}$ for a volume of $30 \,{\rm pc^3}$.
The luminosity of the PeV gamma-ray emission in the energy range of 0.5-1.1 PeV is about $L_\gamma\simeq 5\times10^{31}{\rm erg s^{-1}}$ \citep{LHAASO21_sci}. Then, the energy of protons producing these PeV gamma-rays is about
\begin{equation}
W_{p}\simeq L_\gamma t_{pp}=1.5\times10^{46}{\rm erg} \left(\frac{n}{\rm 10 cm^{-3}}\right)^{-1} .
\end{equation}
Assuming a $E^{-2}$ spectrum for protons, the total energy in relativistic protons will be larger by a factor of several. This energy is only a very small fraction of the total spindown energy of the Crab pulsar, which is about $10^{49}{\rm erg}$.
Note, however, that $W_p$ is the energy of protons that remained in the Crab nebula at present. The real energy of protons that were injected into the nebula could be much larger, since protons/nuclei injected into the nebula at earlier time could have escaped out of the nebula.
The modelling of PWNe as the result of the magnetohydrodynamic (MHD) downstream flow from a shocked, relativistic pulsar wind has been successful in reproducing many features of the nebulae observed close to the central pulsars \citep{KC84a}. In the model, the particles are advected with a toroidal magnetic field. Cross field scattering of particles is expected to be small \citep[e.g.][]{deJager09}, so that diffusion of particles can be neglected. However, although the core region has a clear toroidal magnetic structure, the outer nebula has a complex structure that includes a radial component to the magnetic field.
In the Crab, the prominent toroidal structure observed at X-ray and optical wavelengths extends only to 40'' from the pulsar \citep{Weisskopf00}, while the nebular radius is 200'' (corresponding to a physical size of approximately $r_{\rm crab}=2\,$pc at a nominal distance of 2\,kpc). In addition, the radial variation of spectral index due to synchrotron losses is smoother than expected in the MHD flow model \citep{Tang12}. Thus, diffusion should be the dominant process for particle transport in the outer nebula. Indeed, \citet{Tang12} found that
the diffusion models can reproduce the basic data on the nebular size and the spectral index variation of some PWNe including the Crab nebula.
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{diff_coeff_10PeV}
\includegraphics[width=1\columnwidth]{diff_coeff_100PeV}
\caption{{\bf Left:} Allowable parameter space for the index of the diffusion coefficient ($\delta$) and the turbulence injection scale ($\lambda_{\rm inj}$). {\bf Right}: Corresponding diffusion coefficient as a function of particle energy. The upper panels are for $E_{p,\rm max}=10^{16}\,$eV while the lower panels are for $E_{p,\rm max}=10^{17}\,$eV.}
\label{fig:diff}
\end{figure}
The diffusive escape time from the Crab Nebula
with an effective radius $R$ is given by $t_{\rm diff}=R^2/[4D(E)]$ assuming isotropic diffusion, where $D(E)$ is the diffusion coefficient. \citet{Tang12} obtained a diffusion coefficient of $D_0=2.4\times 10^{26}{\rm cm^2 s^{-1}}$ at the energy of $E=0.6 {\rm TeV}$. The diffusion coefficient at 10\,PeV is, however, not known. The particle gyroradius for 10\,PeV protons is $r_g=E/eB=3\times10^{17}{\rm cm} \left(E/10 {\rm PeV}\right)\left(B/100\mu {\rm G}\right)^{-1}$, which is larger than the mean free path ($\simeq 3D_0/c$) of particles at low energies, invalidating the constant diffusion coefficient assumption. We therefore employ a general form for the diffusion coefficient, $D(E)\propto \left(E/0.6{\rm TeV}\right)^{\delta}$. The parameter $\delta$ depends on the power spectrum of the turbulence (i.e., $W(k)\propto k^{\delta-2}$ with $k$ being the wave number) and is usually considered to be $0-1$ in literature, with $\delta=0$ representing the hard-sphere approximation while $\delta=1$ representing the Bohm-type approximation for the turbulence. The CR diffusion is due to the scattering by the plasma wave with wavelength ($\lambda \sim 1/k$) comparable to the gyroradius of the CR. The diffusion coefficient can be given by $D(E)=(c/3)(r_g/kW(k))$ where $kW(k)$ representing the ratio of turbulence energy density at scale $r_g\simeq 1/k$ to the total magnetic energy density. Assuming a strong turbulence with $kW(k)\sim 1$ at the turbulence injection scale $\lambda_{\rm inj}$, we can rewrite the diffusion coefficient to be $D(E)=(c/3)\lambda_{\rm inj}\left(E/\lambda_{\rm inj}eB\right)^\delta$. Note that the injection scale of the turbulence $\lambda_{\rm inj}$ cannot exceed the size of the nebula $r_{\rm crab}=2\,$pc while it should not be smaller than the gyroradius of the maximum proton energy considered here. Such a condition can put a constraint on the value of $\delta$, if we further require the employed diffusion coefficient to be consistent with the inferred one by \citet{Tang12} to certain extent, i.e., $D_0/2<D(0.6\rm TeV)<3D_0/2$ given a 50\% error range. The allowable parameter space for $\lambda_{\rm inj}$ and $\delta$ is shown in Fig.~\ref{fig:diff} for $E_{p,\rm max}=10\,$PeV and 100\,PeV respectively. Also shown are the corresponding diffusion coefficients. The magnetic field is taken to be $B=112\,\mu$G which is the best-fit parameter found by \citet{LHAASO21_sci}. Note that the mean free path of particle, $\sim 3D/c$, is $\lesssim 1\,$pc, which is smaller than $r_{\rm crab}$ even for 100\,PeV in the considered scenario here, and hence a diffusion description for the particle transport in the nebula is roughly valid up to 100\,PeV. A more strict treatment should include the ballistic propagation of the highest-energy particle in the core region of the nebula (see e.g. \citealt{Aloisio09, Prosekin15}), which may change the expected gamma-ray flux around 10\,PeV.
For an instant injection of particle, the spatial distribution of particle density at an epoch $t$ after the injection is proportional to $(D(E)t)^{-3/2}\exp\left(-r^2/4D(E)t\right)$ assuming spherically symmetric diffusion \citep{Atoyan95}. This means that the particle distribution approximately keeps constant at $r<r_{\rm diff}\equiv 2\left[D(E)t\right]^{1/2}$ whereas declines quickly at $r>r_{\rm diff}$. Therefore, the fraction of particles remaining inside the nebula approximately reads $f_p(E_p,t)={\rm min}[1,(r_{\rm crab}/r_{\rm diff})^3]$. We assume that the CR injection luminosity is a fraction of $\eta_p$ of the spindown luminosity of the pulsar which evolves with time as $(1+t/\tau_0)^{-\sigma}$ \citep{Pacini73}, where $\tau_0\simeq 680\,$yr is the initial spindown timescale and $\sigma=2.33$ given Crab pulsar's braking index being 2.5 \citep{Lyne88}. The cumulative CR protons that are still confined inside the nebula can be given by
\begin{equation}\label{eq:npspec}
\frac{dN_p}{dE_p}=\int_0^{\tau_{\rm crab}} \frac{Q_0(E_p)f_p(E_p,t)}{\left[1+(\tau_{\rm crab}-t)/\tau_0\right]^{\sigma}}dt
\end{equation}
with $\tau_{\rm crab}=967\,$yr is the age of Crab pulsar at today. $Q_0(E_p)=N_0E_p^{-\alpha_p}\exp(-E_p/E_{p,\rm max})$ is the proton injection spectrum where $N_0$ is the normalization factor. The value of $N_0$ can be found by $\int E_pQ_0(E_p)dE_p=\eta_pL_{s,0}$ with $L_{s,0}=L_s(1+\tau_{\rm crab}/\tau_0)^{\sigma}$ being the initial spindown luminosity of the pulsar.
We can find a critical energy $E_c$ for protons by $\tau_{\rm crab}=r_{\rm crab}^2/4D(E_c)$. For protons remaining in the nebula, the spectrum above this energy is modified by the energy-dependent escape and becomes softer than the injection spectrum. On the contrary, the spectrum below this energy simply follows the injection spectrum. We therefore expect a softening in the proton spectrum beyond the energy $E_c=E_0(r_{\rm crab}^2/4\tau_{\rm crab}D_0)^{1/\delta}$, and subsequently a corresponding softening in the pionic gamma-ray spectrum beyond $\approx 0.1E_c$. We calculate the expected spectrum of the hadronic emission from the proton spectrum given by Eq.~\ref{eq:npspec}, following the semi-analytical method developed by \citet{Kafexhiu14}. The attenuation to gamma-ray photons by the pair-production process on CMB is taken into account. We consider all the possible diffusion coefficients shown in Fig.~\ref{fig:diff}, and normalize the resulting 1\,PeV gamma-ray flux to $10^{-13}\, \rm erg~cm^{-2}s^{-1}$ as constrained by the LHAASO observation. As such, a soft injection proton spectrum (e.g., $\alpha>2$) is not favored as it would over-produce the measured gamma-ray spectrum below PeV.
\begin{figure}[htbp]
\includegraphics[width=1\columnwidth]{Crab_proton_fraction_emax.eps}
\includegraphics[width=1\columnwidth]{Crab_proton_fraction_slope.eps}
\caption{{\bf Upper panel:} Possible hadronic gamma-ray components for a fixed proton injection spectral index of $\alpha_p=2$ for the Crab nebula. The blue, yellow, red bands represent the cases of $E_{p,\rm max}=10\,$PeV, 30\,PeV, and 100\,PeV with $\eta_p=(10-57)\%$, $(13-39)\%$, and $(21-36)\%$, respectively. The solid black curve represents the electron IC flux, obtained by LHAASO Collaboration (2021) for the multiwavelength emission modelling. {\bf Lower panel:} Similar to the upper panel but for a fixed maximum proton energy of $E_{p,\rm max}=100\,$PeV. The green, magenta and red bands show the cases of $\alpha_p=1.6$, 1.8 and 2.0 with $\eta_p=(8-14)\%$, $(11-18)\%$, and $(26-43)\%$, respectively. The Fermi-LAT data is taken from \citet{Arakawa20} while the LHAASO data is taken from \citet{LHAASO21_sci}. The setup of the electron component as well as the electromagnetic environment are the same with those in Figure S4 or S5 in \citet{LHAASO21_sci}, i.e., assuming a spectral of $\frac{dN}{dE_e}\propto E_e^{-\alpha}\left[1+(E_e/E_b)^{-\Delta \alpha}\right]^{-1}\exp\left[-(E_e/E_0)^2\right]$ where $\alpha=3.42$, $\Delta \alpha=1.76$, $E_b=0.76\,$TeV, $E_0=450$\,TeV. The target radiation field includes the synchrotron radiation in a magnetic field of $B=112\,\mu$G, a dust emission of 70\,K with an energy density of $0.5\,\rm eV~cm^{-3}$, the interstellar radiation field and the CMB.}
\label{fig:hadronic}
\end{figure}
The results are shown in Fig.~\ref{fig:hadronic} for a spectral index of $\alpha_p=2$ but with different maximum energies of accelerated protons (the upper panel) and for a fixed maximum energy of protons at 100\,PeV but with different spectral indexes $\alpha_p$ (the lower panel). A leptonic component accounting for the multiwavelength emission is also shown with employing the same parameters given by \citet{LHAASO21_sci} in their leptonic-hadronic scenario. The hadronic emission component is harder than the leptonic emission above PeV even if an extremely large $E_{e,max}$ is invoked. Thus we can diagnose the presence of the hadronic component with the measured gamma-ray spectrum above PeV. Currently, the large statistical uncertainties of the LHAASO measurement at PeV energies do not allow a strong statement about
the spectral hardening. In a more conservative way, the LHAASO data in the Figures can be regarded as the upper limit of the contribution from the possible hadronic component. This translates to an upper limit on the total energy in relativistic protons from the pulsar wind (i.e., $\eta_p$). In the case of the proton spectral slope being fixed at $\alpha_p=2.0$, we found $\eta_p=(10-57)\%$, $(13-39)\%$, and $(21-36)\%$ for $E_{p,\rm max}=10\,$PeV, 30\,PeV and 100\,PeV respectively. In the case of the maximum proton energy being fixed at $E_{p,\rm max}=100\,$PeV, we found $\eta_p=(8-14)\%$, $(11-18)\%$, and $(26-43)\%$ for $\alpha_p=1.6$, 1.8 and 2.0, respectively. The range of $\eta_p$ is due to the uncertainty of the diffusion coefficient considered in the calculation. In general, we see that the current LHAASO data allows a fraction of $10\%-50\%$ spindown energy to be converted to the relativistic proton energy in our model.
The above discussions have assumed that the hadronic component in the pulsar wind is pure protons. However, the composition is not well-known and heavier nuclei could show up in the nebula if the photo-disintegration effect is not important. For nuclei with a mass number $A$, the cross section of the hadronuclear interaction becomes $A$ times larger but the energy fraction lost into created pions in one collision is also reduced by $1/A$. As a result, even if heavier nuclei are presented in the nebula, it would not change the above result significantly.
\section{Discussions}
\subsection{Possible Contribution to Cosmic Rays above the Knee}
In our scenario, a considerable fraction of Crab pulsar's spindown energy could be converted into relativistic protons, but most of them have already escape the Crab Nebula, becoming cosmic rays. Considering that the age of the pulsar is about 1000 years and it is located at 6500 light years ($\simeq 2\,$kpc) away from Earth, protons that escape at the earliest epoch have travelled a distance of $2\sqrt{D_{\rm ISM}\times 7500\rm yr}$. Assuming the cosmic-ray diffusion coefficient in the ISM $D_{\rm ISM}\simeq 4\times10^{28}(E/1\,{\rm GeV})^{1/3}\rm cm^2s^{-1}$ \citep[e.g.][]{Trotta11}, those protons can travel to a distance of $l=0.9(E_p/10{\rm PeV})^{1/6}\,$kpc, and have not arrived at Earth. On the other hand, if all the pulsars are cosmic-ray proton accelerators, they might contribute to the measured cosmic ray flux. According to the ATNF Pulsar Catalogue \citep{Manchester05}, the sum of spindown luminosities of Galactic pulsars is about $L_{\rm s, tot}\approx 10^{39}\rm erg~s^{-1}$. Assuming that such a luminosity is stable over million years and {that $\eta_p$ is the same for all pulsars}\footnote{Using the same $\eta_p$ for all pulsars is a naive assumption. The nuclei extracted from the surface of a neutron star are accelerated, at first when passing the outer gap of the inner pulsar magnetosphere and later in the pulsar wind. Due to the lack of knowledge about the details of the above processes, we follow previous assumption that a same fraction of the pulsar rotational energy is lost on acceleration of iron nuclei (and then to relativistic protons due to photo-disintegration) for all galactic pulsars \citep{Bednarek97,Bednarek04}.}, we may roughly estimate the flux of cosmic rays produced by pulsars, in the framework of the leaky box model, by
\begin{equation}
\begin{split}
F(E_p)&=\frac{c}{4\pi}\frac{\eta_p L_{\rm s,tot}t_{\rm esc}}{2\pi R_{\rm Gal}^2H_{\rm CR}\ln (E_{p,\rm max}/E_{p,\rm min})}\\
&\approx 2\times 10^3 \left(\frac{f_{\rm pul}\eta_pL_{\rm s, tot}}{10^{39}\,\rm erg~s^{-1}}\right)\left(\frac{E_p}{10\,{\rm PeV}}\right)^{-1/3}\\
&\times \left(\frac{H_{\rm CR}}{4\,\rm kpc}\right)\left(\frac{R_{\rm Gal}}{15\,\rm kpc}\right)^{-2}\, \rm eV~cm^{-2}s^{-1}sr^{-1}
\end{split}
\end{equation}
where $t_{\rm esc}=H_{\rm CR}^2/4D_{\rm ISM}$ is the cosmic-ray residence time in the Galaxy with $H_{\rm CR}$ being the scale height of the cosmic-ray halo. $R_{\rm Gal}$ is the radius of the Galaxy. $f_{\rm pul}(\geq 1)$ accounts for the contribution of off-beamed pulsars that we cannot observe, and its value could be a factor of a few according to the study by \citet{Tauris98}. Here we have assumed a flat proton spectrum with $\ln (E_{p,\rm max}/E_{p,\rm min})\simeq 20$. The KASCADE-Grande experiment has revealed a cosmic-ray proton flux of $\sim 3000 \,\rm eV~cm^{-2}s^{-1}sr^{-1}$ at 10\,PeV and $\sim 100 \,\rm eV~cm^{-2}s^{-1}sr^{-1}$ at 100\,PeV \citep{KASCADE13b}. It implies that Galactic pulsars may have a significant contribution to the cosmic-ray proton flux between $10-100\,$PeV provided an approximate choice of the value of $\eta_p$, which is consistent with the previous suggestion \citep{Bednarek04}.
\subsection{The Second Leptonic Component}
The spectral hardening of the Crab Nebula at PeV energies could also be due to a second electron component. \citet{LHAASO21_sci} studied such a scenario with the assumption of a Maxwellian-type spectrum \citep[see also][]{Aharonian98} and an $E^{-1.5}$ spectrum for the second electron component. We here explore more general cases with the hard electron spectrum, considering the maximum energy $E_{e,\rm max}$ and the slope of the electron spectrum $\alpha$ as free parameters to study their influences, where the electron spectrum is defined as $dN_e/dE_e\propto E_e^{-\alpha}{\rm exp}(-E_e/E_{e,\rm max})$ (for details in Appendix.~\ref{sec:app}).
The results of the spectra are shown in Fig.~\ref{fig:KN}. We find that, although the electron IC radiation suffers the KN effect at such high energies, it could still account for a spectral hardening up to several PeV (or even higher) energy provided that the electron spectrum is sufficiently hard and the maximum energy is sufficiently high. The critical point is that whether the required maximum electron energy and the hard spectrum can be generated. The former one can be determined by the balance between the acceleration and the synchrotron energy lose rates, which reads \citep{LHAASO21_sci}
\begin{equation}\label{eq:emax}
E_{e,\rm max}=6 \left(\frac{B}{100\mu G}\right)^{-1/2}{\rm PeV},
\end{equation}
in the most efficient acceleration case. Although $B\approx 100\mu$G is obtained via the SED fitting in the one-zone model, the magnetic field strength inside the nebula could be weaker at the termination shock (i.e., the acceleration site) than at larger distance \citep{KC84a, Atoyan96}. Thus, acceleration of $\sim 10\,$PeV electrons in the nebula may still be possible. On the other hand, the hard spectrum may be achieved in the magnetic reconnection process in optimistic conditions \citep{Sironi14, Guo14, Werner16}. Therefore, it may not be easy to distinguish between the hadronic and the leptonic origin for the spectral hardening (if confirmed) solely from the UHE spectrum measurement. Future sensitive neutrino telescopes (e.g. IceCube Gen-2) might be able to distinguish between the leptonic and hadronic scenarios for the spectral hardening \citep{Aartsen2021}, since the hadronic scenario predict a neutrinos flux comparable to the PeV gamma-ray flux.
\section{Summary}
To summarize, based on the latest measurement of PeV emission of the Crab Nebula by LHAASO, we put a constraint on the proton content accelerated in pulsar wind by taking into account the particle escape effect. We found that the current LHAASO data allows up to $(10-50)\%$ of the spindown energy being converted to relativistic protons. This is much larger that previous limits obtained without considering the diffusive escape effect \citep{ZhangX20}. Future observations with the full LHAASO will be able to determine reliably whether there is a hard component in the spectrum of the Crab Nebula. If this hadronic component is confirmed in future, our results may imply that a significant fraction of spindown power is converted into relativistic ions in the pulsar wind. In addition, Galactic pulsars may make a considerable contribution to the cosmic-ray proton flux at $10-100$\,PeV provided the proton spectrum is not too soft. A second electron component might also explain the spectral hardening beyond PeV energy provided a sufficiently high maximum electron energy and a sufficiently hard spectrum. It'd therefore be difficult to distinguish the hadronic and leptonic origins for the spectral hardening (if confirmed in the future) solely from the spectrum measurement, but the future neutrino observation could provide an important test.
\section*{Acknowledgments}
We thank Felix Aharonian for invaluable discussions. This work is supported by the National Key R \& D program of China under the grant 2018YFA0404203 and the NSFC grants
11625312, 11851304 and U2031105.
\newpage
|
\section{Introduction}
\begin{figure}
\includegraphics[width=\textwidth]{frontfigure.png}
\centering
\caption{\textsc{RAFT}{} includes naturally occurring classification datasets, mimicking work that is usually given to human research assistants. Each task comes with natural language instructions and labels in addition to 50 training examples.}
\end{figure}
\begin{comment}
The problem
\end{comment}
Few-shot learning, the capacity to complete a task given a small number of demonstrations \citep{FeiFei2006OneshotLO}, is one of the hallmarks of human intelligence \citep{Tenenbaum2011HowTG, Lake2011OneSL}. As researchers, we leverage this capacity when we delegate work on crowdsourcing platforms or give a task with examples to a human research assistant.
\citet{brown2020gpt3} show that large pre-trained language models exhibit few-shot learning capabilities for a wide range of natural language tasks.
If those capabilities were comparable to people on economically relevant tasks, this would be important to know: a single model could be used across multiple real-world tasks, with low per-task data labeling cost.
However, these models have also been shown to have inconsistent few-shot performance depending on the exact setup and task being solved \citep[e.g.][]{mishra2021natural, perez2021true}.
The mixed evidence suggests that it would be valuable to measure and track few-shot performance on a set of tasks that is representative of what appears in practice.
Natural language tasks coarsely split into generation, classification, and retrieval. We focus on classification tasks because they support high-quality automated evaluation, cover a wide range of economically valuable tasks, and yet don't have existing real-world benchmarks.
\begin{comment}
Why existing work doesn't solve it
\end{comment}
Existing few-shot classification benchmarks are typically designed to highlight areas where models fall short \citep{schick2021its} or to study particular model abilities \citep{bragg2021flex, ye2021crossfit, mishra2021natural}. The tasks and evaluation setup aren't optimized to measure progress in applied settings:
\begin{itemize}
\item \textit{Tasks} that are generated or chosen specifically to test language models may not represent some of the challenges found when applying these models in real-world settings. For example, SuperGLUE \citep{wang2019superglue} and the few-shot equivalent FewGLUE \citep{schick2021its} mainly include short texts. Doing well on applied tasks sometimes requires reasoning over long texts. Existing systems struggle with long texts due to a limited context window, especially in the few-shot setting where some systems learn from examples presented in context.
\item The \textit{evaluation} does not closely mirror deployment, and may both under- and overestimate models' capabilities.
It may underestimate model capability by restricting models to the closed-book setting (e.g., no retrieval from online sources) and using uninformative labels (e.g., 0/1 instead of ``about literature'' vs.\ ``about movies'').
It may overestimate model capability by using many more than a few examples for setting hyperparameters during validation \citep{perez2021true}.
\end{itemize}
\begin{comment}
How we're solving it
\end{comment}
\textsc{RAFT}{} is a real-world few-shot text classification benchmark designed to measure how much recent and upcoming NLP advances benefit applications:
\begin{itemize}
\item The \textit{tasks} are naturally occurring tasks. Their labeling is inherently valuable to someone, and they may have challenges that are not reflected in synthetic tasks. Inherent value means that, if it were sufficiently fast and cheap, it would be desirable to outsource the task to human research assistants or crowd workers. Challenges refers to the need for information retrieval, domain expertise, parsing long documents, and making use of instructions. Table \ref{tab:basic-dataset} shows the real-world challenges presented by \textsc{RAFT}{}, including 4 datasets with long input texts.
\item The \textit{evaluation} closely mirrors deployment. For each task, we release a public training set with 50 examples and a larger unlabeled test set\footnote{Datasets are at \href{https://raft.elicit.org/datasets}{https://raft.elicit.org/datasets}{}}. We encourage unsupervised pre-training on the unlabelled examples and open-domain information retrieval. We keep the test-set labels private and provide automated evaluation through a Hugging Face leaderboard\footnote{Instructions for submission are at \href{https://raft.elicit.org/submit}{https://raft.elicit.org/submit}}.
\end{itemize}
\begin{comment}
Evidence that our solution is good
\end{comment}
In addition to the gold-standard human labels, we collect automatic and crowdsourced baselines. The automatic baselines reveal areas where current techniques struggle, such as reasoning over long texts and tasks with many classes. The crowdsourced baseline reveals that \textsc{RAFT}{} includes a mix of moderate to difficult tasks. We also observe difficulties in collecting human crowdsourced baselines on some datasets, particularly when domain expertise is important, which suggests that real-world value often depends on domain knowledge.
The \textsc{RAFT}{} datasets and leaderboard can be viewed and submitted to at \href{https://raft.elicit.org}{https://raft.elicit.org}{}.
\begin{comment}
Old Table 1
\begin{table}[h]
\centering
\begin{tabular}{m{6em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em}}
\toprule
\textbf{Description} & \textit{ADE} & \textit{B77} & \textit{NIS} & \textit{OSE} & \textit{Over} & \textit{SOT} & \textit{SRI} & \textit{TAI} & \textit{ToS} & \textit{TEH} & \textit{TC} \\
\midrule
Long inputs & -- & -- & \checkmark & \checkmark & -- & -- & \checkmark & \checkmark & -- & -- & -- \\
Domain expertise & \checkmark & -- & -- & -- & \checkmark & -- & -- & \checkmark & \checkmark & -- & -- \\
Detailed instructions & -- & -- & -- & -- & -- & -- & \checkmark & \checkmark & \checkmark & \checkmark & -- \\
\bottomrule
\end{tabular}
\caption{Real-world challenges presented by multiple \textsc{RAFT}{} datasets. More information on \textsc{RAFT}{} datasets is in Section \ref{sec:dataset-description}.}
\label{tab:dataset-challenges}
\end{table}
\end{comment}
\section{Related Work}
We briefly review few-shot learning in NLP, then the benchmarks that are most similar to \textsc{RAFT}{}.
\subsection{Few-shot learning in NLP}
Pre-trained language models (PLMs) such as BERT \citep{devlin2019bert} and GPT-3 \citep{brown2020gpt3} can learn to do some NLP tasks when prompted with a few demonstrations, including some classification tasks. The two primary approaches to few-shot classification using PLMs are in-context learning and prompt-based fine-tuning.
\textbf{In-context learning.} A PLM is primed with labeled examples in its prompt. It classifies the example included at the end of the prompt by predicting the classification conditioned on the priming. GPT-3 \citep{brown2020gpt3} used in-context learning to achieve promising results on a variety of classification tasks. UniFew \citep{bragg2021flex} similarly achieved strong results on classification tasks via in-context learning, converting classification tasks into a multiple-choice question answer format for prompting.
\textbf{Prompt-based fine-tuning}. A PLM is fine-tuned with masked-language modeling objectives to learn from few examples. This is also known as Pattern-exploiting training (PET) \citep{schick2021exploiting}. While PET requires task-specific prompts, it achieves better performance than GPT-3 in-context with smaller models \citep{schick2021its}. LM-BFF \citep{gao2021making} improves prompt-based fine-tuning by dynamically constructing prompts.
\subsection{Few-shot NLP benchmarks}
The most closely related few-shot NLP benchmarks are FLEX \cite{bragg2021flex}, FewGLUE \cite{schick2021its}, CrossFit \cite{ye2021crossfit}, and NaturalInstructions \cite{mishra2021natural}. Each of these benchmarks includes at least some classification tasks with meaningful textual labels.
These benchmarks are designed to study transfer between tasks \citep{bragg2021flex, ye2021crossfit}, pinpoint where NLP models fall short \citep{schick2021its}, and evaluate ability of models to follow instructions \citep{mishra2021natural}, whereas \textsc{RAFT}{} is designed to be representative of real-world classification tasks. This difference in goals is reflected in selection of tasks and evaluation:
\textbf{Tasks.} FLEX, FewGLUE and NaturalInstructions test on traditional NLP tasks. CrossFit tests on 160 tasks from the Hugging Face Datasets\footnote{\href{https://huggingface.co/datasets}{https://huggingface.co/datasets}} and includes some naturally occurring datasets, including the TweetEval dataset \citep{barbieri-etal-2020-tweeteval} that \textsc{RAFT}{} uses as well. CrossFit excludes tasks that leverage external sources or information retrieval techniques, need domain knowledge (e.g. COVID-19 datasets), and long documents (e.g. scientific papers). Like \textsc{RAFT}{}, FLEX includes tasks with strong class imbalance.
\textbf{Evaluation.} None of the existing benchmarks allow open-domain information retrieval. Like FLEX, \textsc{RAFT}{} provides no extra validation data beyond the training examples. \citet{perez2021true} argue that the performance of state-of-the-art few-shot methods has been overestimated by most other existing benchmarks because they use labeled examples beyond the few training instances provided for model and parameter selection.
\begin{comment}
Additional refs to consider:
- han_2021: Ptr: Prompt tuning with rules for text classification
- hendrycks_2020: Dan Hendrycks, Collin Burns, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.
- hendrycks_2021: Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset.
- minaee_2021: Deep learning based text classification: A comprehensive review.
- schick_2020: Few shot text generation with pattern exploiting training.
- min_2021: sewon min, michael lewis, hannaneh hajishirzi, and luke zettlemoyer. Noisy channel language model prompting for few shot text classification.
- xu_2021: Liang Xu, Xiaojing Lu, Chenyang Yuan, Xuanwei Zhang, Hu Yuan, Huilin Xu, Guoao Wei, Xiang Pan, and Hai Hu. Fewclue: A chinese few-shot learning evaluation benchmark.
- Hou2020FewJointAF: FewJoint: A Few-shot Learning Benchmark for Joint Language Understanding
Trapit Bansal, Rishikesh Jha, and Andrew McCallum. 2020. Learning to Few-Shot Learn Across Diverse Natural Language Classification Tasks. In COLING
\end{comment}
\section{Benchmark Description}
\textsc{RAFT}{} is a few-shot {\em classification} benchmark. We focus on classification primarily because automatic evaluation is more reliable than for generation tasks. We believe (as our results will later confirm) that there still is a substantial gap between even non-expert humans and automated systems in the few-shot classification setting.
Both tasks (datasets and metadata) and evaluation (rules for submission, metrics) are chosen to mirror real-world classification problems.
\subsection{Tasks}
A classification task is a dataset with labeled natural language entries. Each label corresponds one-to-one with a natural language class name. Each task has instructions for labeling.
\subsubsection{Dataset selection criteria}
\label{sec:dataset-selection}
We selected datasets based on the following criteria (``non-trivial real world tasks''):
\textbf{Naturally occurring.} We focus on data that are naturally occurring, rather than being synthetically generated to test and improve language models.
\textbf{Intrinsic value.} We select datasets for which the correct labeling inherently provides real-world value. \textsc{RAFT}{} includes tasks like hate-speech detection, medical case report parsing, and literature review automation, where better performance translates into practical benefits. This criterion involves subjectivity, but we aimed to select tasks that approximate the distribution of valuable classification tasks well.
\textbf{Realistic class distribution.} We did not exclude datasets with heavily imbalanced classes.
\textbf{Open-domain feasibility.} As we provide an open-domain setting where information retrieved from the web may be used to augment predictions, we excluded tasks for which the correct label is extremely easily discoverable through a Google search. For example, we considered including the LIAR \citep{wang-2017-liar} dataset which includes Politifact statements and their veracity. We decided against including it since it would be trivial to get 100\% accuracy by running a site search on \href{https://www.politifact.com/}{https://www.politifact.com/}.
In order to gather datasets meeting the above requirements, we put out a collaboration request. We also reached out to users of classification on Elicit \citep{elicit}. Lastly, we conducted a search of existing datasets on the Hugging Face Hub\footnote{\href{https://huggingface.co/datasets}{https://huggingface.co/datasets}} and PapersWithCode \footnote{\href{https://paperswithcode.com}{https://paperswithcode.com}}.
\subsubsection{Dataset preparation}
In cases where the test set was over 5,000 data points, we randomly selected 5,000 to serve as a test set in order to keep the test set sizes manageable. When the dataset didn't already have textual labels, we added textual labels according to our best understanding of the task.
\subsubsection{Selected \textsc{RAFT}{} datasets}
\label{sec:dataset-description}
\begin{comment}
Add table with info for each task. Some subset of:
Name
Example
Real-world explanation
Real-world challenges
Description
Test set size
Num classes
Average length
\end{comment}
\begin{table}[ht]
\small
\centering
\begin{tabular}{m{15em}M{4.5em}M{4.5em}M{5em}M{4.5em}M{4.5em}}
\toprule
\textbf{Dataset Name} & \textit{Long inputs} & \textit{Domain expertise} & \textit{Detailed instructions} & \textit{Number of classes} & \textit{Test set size} \\
\midrule
ADE Corpus V2 (\textit{ADE}) & -- & \checkmark & -- & 2 & 5,000 \\
Banking77 (\textit{B77}) & -- & -- & -- & 77 & 5,000 \\
NeurIPS impact statement risks (\textit{NIS}) & \checkmark & -- & -- & 2 & 150 \\
OneStopEnglish (\textit{OSE}) & \checkmark & -- & -- & 3 & 516 \\
Overruling (\textit{Over}) & -- & \checkmark & -- & 2 & 2,350 \\
Semiconductor org types (\textit{SOT}) & -- & -- & -- & 3 & 449 \\
Systematic review inclusion (\textit{SRI}) & \checkmark & -- & \checkmark & 2 & 2,243 \\
TAI safety research (\textit{TAI}) & \checkmark & \checkmark & \checkmark & 2 & 1,639 \\
Terms of Service (\textit{ToS}) & -- & \checkmark & \checkmark & 2 & 5,000 \\
TweetEval Hate (\textit{TEH}) & -- & -- & \checkmark & 2 & 2,966 \\
Twitter complaints (\textit{TC}) & -- & -- & -- & 2 & 3,399 \\
\bottomrule
\end{tabular}
\caption{Overview of the tasks in \textsc{RAFT}{}. \textit{Long inputs}, \textit{Domain expertise}, and \textit{Detailed instructions} are some of the real-world challenges posed by \textsc{RAFT}{}.}
\label{tab:basic-dataset}
\end{table}
We selected 11 datasets, accessible at \href{https://raft.elicit.org/datasets}{https://raft.elicit.org/datasets}{}. Table \ref{tab:basic-dataset} presents an overview of the datasets. More details are available in the Appendix.
\textbf{ADE Corpus V2 (\textit{ADE}).} The ADE corpus V2 \citep{GURULINGAPPA2012885} contains sentences from medical case reports annotated for relation to adverse drug effects. We focus on the binary classification task of whether a sentence is related to an adverse drug effect (ADE).
\textbf{Banking77 (\textit{B77}).} Banking77 \citep{casanueva2020efficient} contains online banking customer service queries annotated with their intents.
\textbf{NeurIPS impact statement risks (\textit{NIS}).} We include the broader impact statements from NeurIPS 2020 papers collected in the dataset from \citet{ashurst2021aiethics}. We annotate these based on whether they mention possibly harmful applications of the research done in the paper .\footnote{The raw scraped NeurIPS impact statements can be found at \href{https://raft.elicit.org/neurips-impact}{https://raft.elicit.org/neurips-impact}.}
\textbf{OneStopEnglish (\textit{OSE}).} OneStopEnglish \citep{vajjala-lucic-2018-onestopenglish} contains articles sourced from The Guardian newspaper and rewritten by teachers to suit three levels of adult English as a Second Language (ESL) learners.
\textbf{Overruling (\textit{Over}).} Overruling \citep{zheng2021does} contains statements from a law corpus annotated based on whether they are overruling, defined as nullifying a previous case decision as a precedent.
\textbf{Semiconductor org types (\textit{SOT}).} We collect a dataset of institutions that have contributed to semiconductor conferences in the last 25 years, then classify these institutions into organization types: ``university", ``company", and ``research institute".
\textbf{Systematic review inclusion (\textit{SRI}).} We use data from a systematic meta-review studying interventions to increase charitable donations \citep{noetel_slattery_saeri_lee_houlden_farr_gelber_stone_huuskes_timmons_2020}. The task is to predict whether a paper advances past the screening stage.
\textbf{TAI safety research (\textit{TAI}).} We include data from the formation of a bibliographic database for research on the safety of transformative artificial intelligence (TAI) \citep{riedel_deibel_2020}. We choose the binary task of predicting whether a work is classified as TAI safety research.
\textbf{Terms of Service (\textit{ToS}).} The Terms of Service dataset \citep{Lippi_2019} contains clauses from Terms of Services, annotated by whether they are potentially unfair to consumers.
\textbf{TweetEval Hate (\textit{TEH}).} We include the hate-speech detection task from the TweetEval dataset \citep{barbieri-etal-2020-tweeteval}, which was curated from \citet{basile-etal-2019-semeval}.
\textbf{Twitter complaints (\textit{TC}).} We include a dataset of tweets annotated by whether they contain a complaint \citep{preotiuc-pietro-etal-2019-automatically}.
\subsection{Evaluation}
\subsubsection{Setting and rules}
The \textsc{RAFT}{} evaluation replicates real-world few-shot classification problems by restricting to 50 labeled examples without validation set, providing meaningful instructions and labels, and using a no-holds-barred setting:
\textbf{50 labeled examples.} We provide 50 labeled examples per task (not per class). In the authors' experience with users of the classification tool Elicit \citep{elicit}, this is approximately the number of examples people are willing to label for a task with a few thousand unlabeled examples. The 50 examples are chosen randomly, mirroring the applied setting in which one can't easily choose a balanced set. No examples beyond the chosen 50 are available for validation.\footnote{By only releasing 50 labelled examples, we make it difficult to cheat by using more than 50 examples for validation. For the NeurIPS impact statement risks and Semiconductor org type datasets, the test set labels aren't available publicly. For other datasets, the test set labels are available publicly but it is non-trivial and discouraged to seek them out.}
\textbf{Task-specific instructions.} As an important replacement for large amounts of labeled data, instructions can specify how a task should be done. Therefore, we provide the instructions we give to human labelers so that they can be used in instructing automatic systems. The level of detail of the instructions varies. We write the instructions based on information from publications (for datasets published elsewhere) or in consultation with the dataset creator (for new datasets).
\textbf{Meaningful label names.} Similar to instructions, textual labels are an important aspect of few-shot and especially zero-shot learning. We create default textual labels for each dataset as recommended by FLEX \citep{bragg2021flex}.
\textbf{Transfer learning permitted.} Transfer and meta-learning using other datasets is permitted, including further pre-training on other corpora.
\textbf{Unlabeled data permitted.} Use of the unlabeled \textsc{RAFT}{} test sets is permitted, as unlabeled data are usually available in the applied setting.
\textbf{Open-domain retrieval permitted.} Models may be augmented with information retrieved from the internet, e.g. via automated web searches.\footnote{Ideally, we'd only allow information from before the time each dataset needed to be labeled. This isn't currently feasible, so we settle for a fully open-domain setting.}
\textbf{Submission requires only labels.} Submission on the test set is open to the public and only requires upload of test set labels. This is in line with benchmarks like GLUE \citep{wang2019glue} and SuperGLUE \citep{wang2019superglue}, but is in contrast to the few-shot benchmark FLEX \citep{bragg2021flex}. By only requiring labels, we give submission creators maximal flexibility in what models to set up.
\textbf{Weekly evaluation.} Evaluation is run on a weekly basis to minimize information gained from frequent repeated submissions.
\begin{table}[ht]
\small
\centering
\begin{tabular}{m{9.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em} M{1.5em}}
\toprule
\textbf{Baseline} & \textbf{Avg} & \textit{ADE} & \textit{B77} & \textit{NIS} & \textit{OSE} & \textit{Over} & \textit{SOT} & \textit{SRI} & \textit{TAI} & \textit{ToS} & \textit{TEH} & \textit{TC} \\
\midrule
Human (crowdsourced) & \textbf{.735} & \textbf{.830} & \textbf{.607} & \textbf{.857} & \textbf{.646} & .917 & \textbf{.908} & .468 & .609 & \textbf{.627} & \textbf{.722} & \textbf{.897} \\
GPT-3 (175B) & .627 & .686 & .299 & .679 & .431 & \textbf{.937} & .769 & \textbf{.516} & \textbf{.656} & .574 & .526 & .821 \\
AdaBoost & .514 & .543 & .023 & .626 & .475 & .838 & .455 & .506 & .556 & .560 & .443 & .625 \\
GPT-Neo (2.7B) & .481 & .452 & .149 & .408 & .343 & .681 & .406 & .493 & .605 & .565 & .554 & .636 \\
GPT-2 (1.6B) & .458 & .600 & .121 & .561 & .245 & .498 & .380 & .492 & .612 & .498 & .311 & .723 \\
BART MNLI Zero-shot & .382 & .234 & .332 & .615 & .360 & .462 & .644 & .026 & .469 & .122 & .543 & .400 \\
Plurality class & .331 & .446 & .000 & .353 & .164 & .337 & .271 & .493 & .344 & .471 & .366 & .391 \\
GPT-3 Zero-shot & .292 & .163 & .000 & .572 & .323 & .378 & .628 & .027 & .362 & .164 & .303 & .290 \\
\bottomrule
\end{tabular}
\caption{Performance of \textsc{RAFT}{} baselines (F1)}
\label{tab:baseline-results}
\end{table}
\subsection{Metrics}
Since some \textsc{RAFT}{} datasets have substantial class imbalances, we use F1 as our evaluation metric. We compute macro-averaged F1 scores, even for binary datasets. To get an overall score, we average across all datasets.
\section{Baselines}
The code for all automatic baselines is open-sourced at \href{https://raft.elicit.org/baselines}{https://raft.elicit.org/baselines}{}.
\subsection{GPT-3 baseline}
We provide a simple automatic baseline using GPT-3 \citep{brown2020gpt3}, accessed through the OpenAI API\footnote{\href{https://beta.openai.com}{https://beta.openai.com}}. As in \citet{brown2020gpt3}, we use in-context learning, adding labeled examples to the prompt to prime GPT-3. We also run a zero-shot version with no training examples included in the prompt.
\subsubsection{Prompt construction}
We build a prompt consisting of:
\begin{enumerate}
\item Task-specific instructions
\item $N$ labeled examples, with $N$ selected on a per-task basis
\item The target example to classify
\end{enumerate}
Example prompts for all datasets are available at \href{https://raft.elicit.org/baseline-prompts}{https://raft.elicit.org/baseline-prompts}.
\textbf{Truncation.} GPT-3's context window support up to $2,048$ tokens. The experiments in \citet{brown2020gpt3} include as many complete labeled examples as would fit in the context, reporting that typically 10 to 100 examples fit. However, datasets such as \textit{OSE} have very long inputs so that only 1-2 complete labeled examples would fit, suggesting that another approach may be better.
We select $N$ training examples to include in a given prompt and then truncate the examples. In a given task, the instructions take up $I$ tokens. Separators between the instructions and each example take up $S$ tokens. No more than $T = 2,048$ tokens can be used.
\begin{enumerate}
\item $E = T - I - S$ tokens are allotted for the training examples and classification target.
\item The classification target is truncated to $\frac{1}{4} E$ tokens.
\item Each of the $N$ remaining training examples is truncated to $\frac{3}{4}\frac{E}{N}$ tokens.
\end{enumerate}
We truncate from a training example's data fields first, leaving the label intact.
\textbf{Field selection and sorting.} We exclude data fields that are unlikely to contribute substantially to GPT-3's performance. These fields either deal with the authors of the textual example or are URLs. Additionally, we sort the order in which the text fields occur to put the most important fields first. When examples are truncated, the most important information is preserved.
\begin{comment}
Don't think this table was adding much and we're low on space
\begin{table}[h]
\centering
\begin{tabular}{l l}
\toprule
\textbf{Labeled} & \textbf{Unlabeled} \\
\midrule
\makecell[l]{Query: Is it possible for me to change my PIN number? \\ Label: change\_pin} & \makecell[l]{Query: where did my funds come from? \\ Label:} \\
\bottomrule
\end{tabular}
\caption{A sample labeled and unlabeled example from \textit{B77}.}
\label{tab:prompt-examples}
\end{table}
\end{comment}
\textbf{Semantic selection.} To select training examples to include in the prompt for a given test example, we selected the most similar training examples as in \citet{liu2021makes}. To perform semantic search, we use the OpenAI API search endpoint with the \texttt{ada} engine.
\subsubsection{Classification}
With the prompt formed, we retrieve GPT-3's 100 most likely next tokens using the \texttt{davinci} engine. For each class, we assign the probability that its first token is generated. We then normalize the probabilities to sum to 1. For the \textit{B77} dataset, multiple labels share the same first token so we prepend a numerical prefix such as ``1. '' to each class.
\subsubsection{Parameter selection}
\label{sec:param-tuning}
We tune the GPT-3 baseline on the training set using \textit{leave-one-out cross validation} (LOOCV): $k$-fold cross validation with $k = n$ so that only one test example is used at a time for validation. While LOOCV isn't robust with as few as 50 examples as discussed in \citet{perez2021true}, it is one of the best options for parameter selection in the few-shot setting. Detailed LOOCV results are in Section \ref{sec:auto-baseline-details}.
\textbf{Instructions.} We test two modes of instruction: (a) a generic classification prompt: "Possible labels:" followed by a list of textual labels. (b) instructions similar to the ones given to human labelers, plus the list of textual labels. The instructions are taken whole when possible, and otherwise shortened and summarized manually to limit usage of the GPT-3 context window. Task-specific instructions outperform generic instructions by an $.04$ on averaged F1 score, thus we include task-specific instructions in the baseline.
\textbf{Semantic training example selection.} To select training examples for inclusion in the prompt from a larger set, we consider (a) selecting examples randomly and (b) using semantic search to identify the training examples most similar to the test example. Semantic selection outperforms random selection by $0.03$ on averaged F1, thus we include semantic selection in the baseline.
\textbf{Number of examples in the prompt.} We select the number of examples to include in the prompt on a per-dataset basis, as our truncation strategy induces a quality-quantity trade-off. For each dataset, we test performance with 5, 10, 25, and 50\footnote{49 rather than 50 training examples for LOO experiments} training examples and choose the number that performs best by F1. For datasets with long inputs, smaller numbers of more detailed samples often produce better performance, while datasets with smaller inputs can fit more complete labeled examples in the prompt.
\subsection{Other automatic baselines}
\textbf{In-context baselines.} We run further in-context baselines GPT-Neo \citep{gpt-neo} and GPT-2 \citep{radford2019language}. We provide code\footnote{\href{https://raft.elicit.org/baselines}{https://raft.elicit.org/baselines}{}} for generating predictions on \textsc{RAFT}{} using these models and any other causal language model available on the HuggingFace Hub. For semantic search, we use a MiniLM \citep{wang2020minilm} fine-tuned on sentence pairs via the sentence-transformers package\footnote{https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2}.
\textbf{Zero-shot baselines.} We run two transformers in the zero-shot setting:
\begin{itemize}
\item GPT-3, to judge to what extent training examples in the prompt aid performance
\item BART \citep{bart2019lewis} trained on MNLI \citep{mnli2017williams}, as suggested by \citet{zeroshot2019yin} and \citet{davison2020zeroshotlearning} as an effective zero-shot classification approach
\end{itemize}
\textbf{Non-neural baselines.} We run AdaBoost \citep{FREUND1997119} to establish a strong non-neural baseline. We construct feature vectors for each example based on the counts of $n$-grams of $1$-$5$ words as the input to a weighted ensemble of 100 depth-3 decision trees. These decision trees and weights are trained with AdaBoost with learning rate 1, and evaluated through weighted voting. We also include a plurality (most frequent) class baseline.
\iffalse
For Naive Bayes and AdaBoost (citation?) we used a different featurizer than our neural baselines. We concatenated all non-label data in every training example into a single string, separated by periods, then constructed $n$-grams from all words and adjacent sets of $n$ words in the dataset for $n \in [1, 5]$ after removing letter cases and certain special symbols. Each training or test example was vectorized as the count of each $n$-gram in the example.
\begin{itemize}
\item \textbf{Naive Bayes} -- We trained a Multinomial Naive Bayes model with Laplace smoothing parameter $\alpha = 0.05$.
\item \textbf{AdaBoost} -- For the base estimator, we used decision trees with a maximum depth of 3. We ensembled 100 estimators with a learning rate of 1.0.
\item \textbf{Plurality class} -- We simply classified every example with the most frequent class label in the training set.
\end{itemize}
\fi
\subsection{Human baseline}
To collect human baselines, we use the Surge\footnote{https://www.surgehq.ai/} crowdsourcing platform. Following \citet{wang2019superglue}, we randomly select 100 data points from each test set and use a 2-step labeling process: qualification then annotation. The crowdsourced label is the plurality vote of 5 labelers.
We put crowd workers in a similar situation to automated systems. We link to a sheet with the same 50 labeled examples, use the same textual labels, and give the same task-specific instructions that we are providing to practitioners to adapt for instructing language models.\footnote{For details on the human baseline gathering process, see Section \ref{sec:human-baseline-details}.}
\subsection{Analysis}
\begin{comment}
Wide-ranging results indicating varying levels of difficulty
[Insert things about in which areas LMs are still much worse than humans]
\end{comment}
\textbf{Humans generally outperform GPT-3.} Humans outperform GPT-3 on 8 out of 11 tasks, demonstrating room for improvement for models on real-world few-shot tasks. We expect that exceeding the crowdsourced baseline will require substantial advances in model performance, and even more so for a future expert human baseline.
Weaknesses of GPT-3 include:
\begin{itemize}
\item \textbf{Many classes}: Humans most outperform GPT-3 on \textit{B77}, which has by far the most classes in \textsc{RAFT}{}. With 77 classes and 50 labeled examples, many classes have no corresponding labeled examples. Additionally, just listing out the possible classes takes up a large portion of GPT-3's context window.
\item \textbf{Long inputs}: GPT-3 performs poorly on some tasks requiring reasoning over long inputs, such as \textit{NIS} and \textit{OSE}. GPT-3's context window may be a contributing factor.
\end{itemize}
\textbf{Crowd-sourced baselines struggle on domain-specific tasks.}
Crowd-sourced humans substantially outperform GPT-3 on only 1 of 4 tasks we identified as requiring domain expertise:
\begin{itemize}
\item Humans substantially outperform GPT-3 on \textit{ADE}, which requires medical expertise.
\item Humans outperformed GPT-3 by just .053 on \textit{ToS}, which requires parsing legal language.
\item GPT-3 outperforms humans on \textit{Over}, which requires greater legal expertise than \textit{ToS} \citep{zheng2021does}, and \textit{TAI}, which requires expertise in AI safety research.
\end{itemize}
\textbf{Zero-shot performance is weak.} GPT-3 zero-shot does poorly on \textsc{RAFT}{}, performing worse than the plurality class baseline. BART zero-shot exceeds the plurality class baseline but does not do so in every dataset, and it is not competitive with few-shot language models. We encourage future research on improving performance in the zero-shot setting, perhaps through improved prompt construction and transfer learning.
\textbf{Neural baselines besides few-shot GPT-3 perform worse than AdaBoost.} Generative language models smaller than GPT-3 comfortably outperform the plurality class baseline but remain below AdaBoost. We use the same amount of labelled examples in the prompt as with GPT-3 despite the context window being smaller; performance may improve with fewer (but longer) examples.
\section{Discussion}
\begin{comment}
Stronger human baselines
[Maybe: More experimental baselines]
[Maybe: More datasets]
More exploration on learning from natural language rather than examples
Instructions, feedback, explanations
Zero-shot benchmark?
tasks besides classification
Version with more robust evaluation
Varying levels of training examples, as in applied setting
Some bias in tasks toward Ought’s setting: AI-related tasks
Could decompose into limitations for
- tasks
- evaluation setup
- baselines
\end{comment}
\subsection{Limitations}
\textbf{Linguistic diversity.} The benchmark only includes English tasks. Dealing with multilingual corpora is a real-world challenge for many NLP systems, especially for those deployed in countries where there are multiple national languages. To fully capture the distribution of real-world tasks, additional languages will be needed.
\textbf{Possible biases in data collection.} While we attempted to execute our dataset selection process as described in Section \ref{sec:dataset-description} in an unbiased manner, the datasets we ended up selecting are part of a subjective human process that may be subject to biases. For example, the organizations we work with are disproportionately in technology and policy.
\subsection{Impact}
\textbf{Offensive content.} By including a hate-speech detection dataset, we include offensive content and may harm readers of the dataset. We believe the advantages from studying hate-speech detection are likely greater than the disadvantages of publicizing hate-speech datasets.
\textbf{Prohibitive costs.} The models best equipped to perform well on RAFT will often be the massive transformer models trained by private corporations. In advancing this benchmark as a means of evaluating models, we risk further widening the gap between what a dedicated individual or team can do, and what can only be done by industry research labs with sufficient funding.
\subsection{Future Work}
\textbf{Stronger human baselines.} Human baselines are intended to tell us how well the dataset would be labeled in the absence of automated systems. For many \textsc{RAFT}{} datasets, this process would involve a stronger baseline than is easily available via a crowd-worker platform: for example, the \textit{Over} dataset would be labeled by someone with law expertise. In addition to ML submissions, we welcome efforts to collect stronger human baselines for \textsc{RAFT}{}.
\textbf{Additional automatic baselines.} We expect that systems that use prompt-based fine-tuning rather than in-context learning may provide an even stronger automatic baseline. We further expect that models that leverage the open-domain information retrieval option can surpass models that don't.
\textbf{Application-specific metrics.} Different applications care about different metrics; e.g., in some applications it is more important to minimize false positives, whereas in others the focus is on false negatives. An ideal measure of real-world value would take that into account.
\textbf{Learning from natural language} In this work, we focused on instructions as a supplement to labeled examples. Similarly to \citet{mishra2021natural}, we found that including task-specific instructions improved performance. Like humans, NLP systems could also learn from other types of natural language. For example, could including explanations with each labeled example be used to further improve few-shot performance?
\section{Conclusion}
\textsc{RAFT}{} is a benchmark that tests language models across multiple domains on economically valuable classification tasks in the true few-shot setting. To our knowledge, this is the first multi-task benchmark designed to closely mirror how models are applied in both the task distribution and the evaluation setup. By complementing existing synthetic benchmarks designed to highlight where models fall short, it helps measure the gap between research and practice, incentivizes work that is valuable for deployed systems, and provides a template for future benchmarks that mirror deployment.
\begin{ack}
Our automatic baseline collection was subsidized by compute credits generously provided by OpenAI. Ethan Perez, Samuel Bowman, and Long Ouyang gave feedback on early versions of the \textsc{RAFT}{} concept and dataset lists. Douwe Kiela and Stella Biderman offered helpful advice on the project direction. Ross Gruetzemacher suggested inclusion of the Twitter Complaints dataset. We thank Thomas Wolf and Simon Brandeis for discussions and advice around the design of the benchmark’s infrastructure.
\end{ack}
\bibliographystyle{plainnat}
|
\section{Introduction}
\label{sec:intro}
Representation learning, which aims to learn informative and effective features of a task, is a key part of deep learning.
Naturally, deep reinforcement learning (DRL) is expected to benefit from the help of representation learning.
Many works \cite{lange2010deep, laskin2020curl, pacelli2020learning, franccois2019combined} have dedicated to boost DRL by learning a compact, discriminative or task-relevant representation from observations.
However, in multi-agent tasks, a good task-relevant representation also needs to be teammate-relevant or opponent-relevant.
It has been demonstrated in some works \cite{he2016opponent, jin2020stabilizing, foerster2017stabilising} that using independent DRL (ignore other agents' behaviors) may lead to unsatisfactory results.
Meanwhile, some works \cite{lowe2017multi, tesauro2004extending} indicate that inferring other agents' policies can improve cooperation between agents, but is prohibitively expensive for policies parameterized by deep neural networks.
These studies imply the demand for more efficient and effective representation learning of other agents in multi-agent DRL (MADRL).
Two core problems of representing other agents in MADRL are what to represent and how to combine the representation with MADRL.
Foerster et al. \cite{foerster2017stabilising} leverage low-dimensional fingerprints to represent other agents' policy changes, which forms succinct features, but is deficient in policy information completeness.
Jin et al. \cite{jin2020stabilizing} propose to represent other agents' behaviors implicitly using their positions at adjacent timesteps.
However, the behavior representation is learned via MADRL in an implicit and uninterpretable manner.
He et al. \cite{he2016opponent} propose deep reinforcement opponent network (DRON) to learn representations of other agents' actions explicitly by leveraging other agents' actions as supervision signals.
However, the compactness and information utility of representation are not considered.
In this paper, we leverage information bottleneck principle \cite{tishby2015deep, tishby2000information}
to learn an informative and compact representation relevant to other agents' behaviors to improve the performance of MADRL.
In particular,
we employ an encoder to extract features from each agent's positions at two adjacent timesteps, based on which a classifier is learned to estimate actions of each agent.
To filter out irrelevant information from observations and retain sufficient amount of information of other agents' actions used for cooperation decision, we follow the information bottleneck principle to minimize the mutual information between the representation and the observations, while maximizing the mutual information between the representation and other agents' actions.
To this end, we adopt a variational method \cite{belghazi2018mutual} to estimate the two mutual information and integrate this process into behavior representation learning.
We combine our proposed behavior representation learning method with our recent work, stabilized multi-agent deep Q learning (SMADQN) \cite{jin2020stabilizing} by multi-task learning and thereby the learned representation can also retain other information about the MADRL task in addition to other agents' actions. Experimental results demonstrate the superior performance of our proposed method compared to vanilla SMADQN and DRON-based SMADQN.
In summary, the main contributions of this paper are as follows:
1) We propose an information-bottleneck-based behavior representation learning method through which compact and informative features of other agents' behaviors are learned and exploited to facilitate MADRL.
2) We conduct extensive experiments in cooperative navigation tasks \cite{lowe2017multi, jin2020stabilizing, jin2019efficient}.
Experimental results demonstrate that compared to implicit behavior representation learning and the explicit behavior representation learning that does not consider information utility and compression, our method performs best in terms of both
learning speed and the success rates of the resulting policies.
\section{Method}
In this section, we first introduce the Markov game, SMADQN \cite{jin2020stabilizing} and DRON \cite{he2016opponent}.
Then, we present our method.
A Markov game with $N$ agents involves a set of states $s$, joint actions $(a_1,\cdots, a_N)$, transition probability function $p(s'|s,a_1,\cdots, a_N)$, and each agent's reward function $r_i(s,a_1,\cdots, a_N)$, $i\in [1,N]$.
At each timestep, each agent executes an action according to its policy $\pi_i$.
A problem of Markov game is to find the optimal policy $\pi_i^*$ for each agent so that
\footnotesize $\forall \pi_i, R_i(s^t, \pi_1^*, \cdots, \pi_i^*, \cdots, \pi_N^*)\geq R_i(s^t, \pi_1^*, \cdots, \pi_i, \cdots, \pi_N^*)$,
\normalsize where
\footnotesize $R_i(s^t, \pi_1, \cdots, \pi_N)=E[\sum\nolimits_{\tau=t}^T {{\gamma ^{\tau-t}}} {r_i(s^{\tau},a_1^{\tau}, \cdots, a_N^{\tau} )}]$
\normalsize denotes the expected total reward of agent $i$, $T$ is the time horizon, $\gamma\in[0,1]$ is a discount factor.
For convenience, we use $r_i^t$ to denote the reward of agent $i$ at timestep $t$.
SMADQN defines an extended action-value function $G$ for each agent to measure its expected total reward when it follows policy $\pi_{i}$. For agent $i$, $G$-function is defined as
\begin{equation}\label{action-value function}
\footnotesize
G_i^{\pi_{i}}(s^t,s_{-i}^t,s_{-i}^{t+1},a_i^t) = Q_i^{\pi_{i}}(s^t,f(s_{-i}^t, s_{-i}^{t+1}),a_i^t),
\end{equation}
where $s^t$ represents global states, $s_{-i}^t$ and $s_{-i}^{t+1}$ represent states of agents except agent $i$ at two adjacent timesteps, $f$ is an action estimation function of other agents' actions. $G$-function is a composite function that incorporates the action estimation function into the original action-value function $Q$ \cite{sutton2018reinforcement}.
An approximate extended Bellman equation for the optimal $G$-function is derived as:
\begin{equation}\label{Bellman_optimality_G}
\footnotesize
\begin{aligned}
&\mathbb{E}_{s_{-i}^{t+1}|s_{-i}^t,a_{-i}^t} G_i^*(s^t, s_{-i}^t, s_{-i}^{t+1},a^t_i) \approx\\ &\mathbb{E}_{s^{t+1}|s^t,a_i^t,a_{-i}^t}
\left[r_i^{t+1}+\gamma \mathop{\max}\limits_{a^{t+1}_i}G_i^*(s^{t+1},s_{-i}^t,s_{-i}^{t+1},a^{t+1}_i) \right].\\
\end{aligned}
\end{equation}
The optimal $G$-function is approximated by a neural network learned by minimizing the loss function given as:
\begin{equation}\label{loss_masddpg}
\footnotesize
\begin{aligned}
&L=\mathbb{E}_{s^t, s^{t+1},a^t_i}\left[\left(r_i^{t+1}+\gamma \max\limits_{a_i^{t+1}} G_i(s^{t+1},s_{-i}^t,s_{-i}^{t+1},a^{t+1}_i)\right.\right.\\
&\quad\quad\quad\quad \quad\quad \quad\quad \quad\quad \left.\left.-G_i(s^t, s_{-i}^t. s_{-i}^{t+1},a^t_i)\right)^2\right],
\end{aligned}
\end{equation}
SMADQN learns action estimation function of other agents' actions implicitly, which may lead to trivial estimation performance and thereby cause limited performance of the resulting policies.
Instead of merging action estimation learning into MADRL, DRON leverages other agents' actions as supervision signals and adopts supervised learning to learn action estimation.
It uses a classification network to estimate other agents' actions. The output of the last hidden layer of the network is used as other agents' action representation, and is fed into a decision network.
DRON can learn representations of other agents' behaviors explicitly.
However, it does not consider information compression and retention in the representation.
To facilitate and improve MADRL, extracting informative and compressed representation of other agents' behaviors is critical.
To this end, we propose Information-Bottleneck-based Other agents' behavior Representation learning for Multi-agent reinforcement learning (IBORM), which equips a behavior representation with the following capabilities, a) to extract features of other agents' actions, b) to filter out irrelevant information while retaining sufficient information about the actions and other potentially helpful information about the task to facilitate MADRL.
Specifically, we implement our idea in SMADQN.
We replace the implicit action representation learning of SMADQN with explicit action representation learning.
An encoder is employed to learn the representation using other agents' states at adjacent timesteps as inputs.
The encoder's output is a low-dimensional feature vector of other agents' actions, from which a classifier can predict the actions.
The representation learning and SMADQN are combined by leveraging multi-task learning
Additionally, to learn an informative and compressed representation, we leverage information bottleneck principle \cite{tishby2015deep, tishby2000information} to constrain the information contained in the representation.
To be specific, information bottleneck (IB) principle introduces an information theory principle for extracting an optimal representation $Z$ that captures the relevant information in a random variable $X$ about another correlated random variable $Y$ while minimizing the amount of irrelevant information, where $(Y,X,Z)$ forms a Markov chain, $Y\rightarrow X\rightarrow Z$. Namely, finding the optimal representation function is formulated as minimizing the following Lagrangian
\begin{equation}\label{loss_IB}
\footnotesize
\mathcal{L}(p(z|x)) = I(X; Z) - \kappa I(Z;Y),
\end{equation}
where $\kappa$ determines how much relevant information is contained in the representation.
Based on IB principle, we constrain the representation learning by the following terms
\begin{equation}\label{loss_IB_ours}
\footnotesize
\begin{aligned}
& \mathcal{L}(\alpha) \triangleq I(\phi_{s_{-i,j}}; ENC_i^{\alpha}(\phi_{s_{-i,j}}) - \kappa I(ENC_i^{\alpha}(\phi_{s_{-i,j}});a_{-i,j}), \\
\end{aligned}
\end{equation}
where $\alpha$ denotes the parameters of the encoder, $a_{-i,j}$ and $\phi_{s_{-i,j}}$ denote the action and agent $i$'s observation of the $j$th agent other than agent $i$, respectively.
Overall, the loss function of IBORM is given as:
\begin{equation}\label{loss function}
\footnotesize
\begin{aligned}
&L_i(\alpha, \beta, \theta)=J_i^{CE}(\alpha,\beta)+ \lambda_1 J_i^{DRL}(\alpha, \theta)+ \lambda_2 \mathcal{L}(\alpha),\\
\end{aligned}
\end{equation}
where $J^{CE}_i(\alpha, \beta)$ denotes the cross-entropy between classifier’s output and each agent’s true action , $\alpha$ and $\beta$ are parameters of the encoder and the classifier.
$J_i^{DRL}(\alpha, \theta)$ denotes a modified loss function of SMADQN, defined as
\begin{equation}\label{drl loss}
\footnotesize
\begin{aligned}
&J_i^{DRL}(\alpha, \theta) = \mathbb{E}_{\phi_{s_i}^t,a_i^t, \phi_{s_{-i,1}}^{t+1},\cdots,\phi_{s_{-i,N-1}}^{t+1}}\\ &\left[\left(y_i^t-G_i^{\theta}(\phi_{s_i}^t,ENC_i^{\alpha}(\phi_{s_{-i,1}}^{t+1}),
\cdots,ENC_i^{\alpha}(\phi_{s_{-i,N-1}}^{t+1}),a_i^t)\right)^2\right],
\end{aligned}
\end{equation}
where
\footnotesize $y_i^t=r_i^{t+1}+\gamma\mathop{\max}\nolimits_{a_i}G_i^{\theta_{tar}}(\phi_{s_i}^{t+1},ENC_i^{\alpha}(\phi_{s_{-i,1}}^{t+1}),
\cdots,\\
ENC_i^{\alpha}(\phi_{s_{-i,N-1}}^{t+1}),a_i)$.
\normalsize
$\theta$ and $\theta_{tar}$ are parameters of $G$ network and target network \cite{mnih2015human}, respectively.
Compared with (\ref{loss_masddpg}), we replace other agents' adjacent states with $N-1$ behavior representations.
For notation convenience, we rewrite (\ref{loss function}) as
\begin{equation}\label{loss function_new}
\footnotesize
\begin{aligned}
&L_i(\alpha, \beta, \theta)=J_i^{CE}(\alpha,\beta)+ \lambda_1 J_i^{DRL}(\alpha, \theta)+ \\
&\lambda_2 I(\phi_{s_{-i,j}}, ENC_i^{\alpha}(\phi_{s_{-i,j}})) - \lambda_3 I(ENC_i^{\alpha}(\phi_{s_{-i,j}}), a_{-i,j}),\\
\end{aligned}
\end{equation}
where $\lambda_1, \lambda_2, \lambda_3$ are positive weights.
From the perspective of information utility, the first and the last terms of (\ref{loss function_new}) are for extracting sufficient information of other agents' behaviors.
The second term extracts relevant information of the task.
The third term filters out irrelevant information.
Compared to IBORM, SMADQN only uses the DRL-based term, where the behavior representation learning is implicitly contained. DRON uses the cross-entropy term but does not constrain the amount and utility of the information in the representation.
The network architecture of IBORM is shown in Fig.~\ref{fig.DiagramIBOR2}.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\columnwidth]{DiagramIBOR2.png}
\captionsetup{font=footnotesize}
\caption{Network architecture diagram of IBORM.}
\vspace{-1mm}
\label{fig.DiagramIBOR2}
\end{figure}
\begin{figure}[!]
\centering
\includegraphics[width=0.7\columnwidth]{Algo_Framework.png}
\captionsetup{font=footnotesize}
\caption{Framework of IBORM algorithm.}
\vspace{-1mm}
\label{fig.AlgoFramework}
\end{figure}
An encoder-classifier is used to estimate each of other agents' actions.
The encoder is duplicated by $N-1$ times and generates bottleneck representations for $N-1$ other agents' actions, respectively.
Then, the representations are incorporated into a decision network to make interactive decisions.
To estimate the two mutual information terms in (\ref{loss function_new}), we employ the Mutual Information Neural Estimator (MINE) \cite{belghazi2018mutual} that estimates the mutual information between two variables $X$ and $Z$ as \footnotesize $\widehat{I(X,Z)}=\sup_{{\omega \in \Omega}} \mathbb{E}_{\mathbb{P}_{XZ}}[T_\omega(x,z)]-\log(\mathbb{E}_{\mathbb{P}_X \otimes \mathbb{P}_Z} [e^{T_\omega(x,z)}] )$ \normalsize by leveraging a trainable neural network $T_\omega$ with parameters $\omega$. Specifically, we use two networks corresponding to MINEs of the two mutual information terms in (\ref{loss function_new}). To integrate learning of MINEs and IBORM, we adopt an interlaced learning manner to update parameters of MINEs and IBORM alternately.
A framework of our algorithm is shown in Fig.~\ref{fig.AlgoFramework}.
The complete algorithm is shown in Algorithm \ref{algo}, where we denote
$\boldsymbol{\phi}_{s_{-i}}^t$ \footnotesize $ =[\phi_{s_{-i,1}}^{t},\cdots,\phi_{s_{-i,N-1}}^{t}]$
\normalsize
and
\footnotesize $\boldsymbol{a}_{-i}^t = [a_{-i,1}^t,\cdots,a_{-i,N-1}^t]$
\normalsize
for notation convenience.
\begin{algorithm}[!t]
\footnotesize
\captionsetup{font=small}
\caption{\footnotesize Stabilized multi-agent deep Q-learning with information-bottleneck-based other agents' behavior representation learning}
\begin{algorithmic}[1]
\raggedright
\For{agent $i=1$ to $N$}
\State
Initialize \\
\qquad networks \scriptsize $ENC_i^{\alpha}:\phi_{s_{-i,j}}\rightarrow z_{-i,j}$, $CLAS_i^{\beta}:z_{-i,j} \rightarrow \hat{a}_{-i,j}$, \\
\qquad $MINE_i^{\omega_1}(\phi_{s_{-i,j}}, z_{-i,j})$, $MINE_i^{\omega_2}(z_{-i,j},a_{-i,j})$, \\
\qquad $G_i^{\theta}(\phi_{s_i},ENC_i^{\alpha}(\phi_{s_{-i,1}}),\cdots,ENC_i^{\alpha}(\phi_{s_{-i,N-1}}),a_i)$,\\
\qquad \footnotesize target network \scriptsize $G_i^{\theta_{tar}}$ \footnotesize with \scriptsize $\theta_{tar}\leftarrow\theta$, \footnotesize replay buffer \scriptsize $\mathcal{D}_i$\\
\EndFor \footnotesize
\For{episode$=1$ to $Z$}
\State Receive $\phi_{s_i}^1, \boldsymbol{\phi}_{s_{-i}}^1$ for each agent
\For {t$=1$ to $T$}
\State Execute action for each agent $i$:\\
\scriptsize $a_i^t = \arg\max\limits_{a_i} G_i^{\theta}(\phi_{s_i}^t,ENC_i^{\alpha}(\phi_{s_{-i,1}}^t),\cdots,ENC_i^{\alpha}(\phi_{s_{-i,N-1}}^t),a_i)$
\footnotesize
\State Receive $r_i^{t+1}, \phi_{s_i}^{t+1}, \boldsymbol{\phi}_{s_{-i}}^{t+1}$ for each agent $i$
\State Record data: \scriptsize $\mathcal{D}_i \leftarrow \mathcal{D}_i \bigcup \{(\phi_{s_i}^t,a_i^t,r_i^{t+1},\phi_{s_i}^{t+1},\boldsymbol{\phi}_{s_{-i}}^{t+1},\boldsymbol{a}_{-i}^t)\}$\\
\qquad \quad ~ \footnotesize for each agent $i$
\For {agent $i=1$ to $N$}
\State Sample $M$ tuples \scriptsize \\
\qquad \qquad \quad $\{(\phi_{s_i}^{\tau_k},a_i^{\tau_k},r_i^{\tau_k+1},\phi_{s_i}^{\tau_k+1},\boldsymbol{\phi}_{s_{-i}}^{\tau_k+1},\boldsymbol{a}_{-i}^{\tau_k})\}_{k=1}^M$ from $\mathcal{D}_i$ \footnotesize
\State Compute \scriptsize~$\{z_{-i,j}^{\tau_k+1} = ENC_i^\alpha(\phi_{s_{-i,j}}^{\tau_k+1})\}_{
j=1}^{N-1} ~ $\footnotesize
\State Compute $y_i^{\tau_k}$ by\\
\scriptsize \quad \quad $y_i^{\tau_k}=r_i^{\tau_k+1}+\gamma\mathop{\max}\nolimits_{a_i}G_i^{\theta_{tar}}(\phi_{s_i}^{\tau_k+1},z_{-i,1}^{\tau_k+1},\cdots,z_{-i,N-1}^{\tau_k+1},a_i)$
\State \footnotesize Update $\omega_1, \omega_2$ according to \cite{belghazi2018mutual} using SGD
\State Update $\alpha, \beta, \theta$ to minimize (\ref{loss function_new}) using SGD
\State Update target network with soft update rate $\eta$:\\
\qquad \qquad ~~$\theta_{tar}\leftarrow\eta\theta+(1-\eta)\theta^{tar}$\\
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\label{algo}
\end{algorithm}
\section{Experiments}
In this section, we evaluate IBORM in multi-agent cooperative navigation task with the same settings used in \cite{jin2020stabilizing}.
In this task, agents need to cooperate through motions to reach the same number of targets using the minimum time.
An example containing three agents and targets is illustrated in Fig.~\ref{env}.
At each timestep, each agent selects a target and move a fixed distance toward the target.
The action of an agent is defined as $a_i \in [1,N]$ that indicates the index of the target selected by it. Agents' speed is $1 \ \text{m}/\text{timestep}$.
The current observation of agent $i$, i.e. $\phi_{s_i}$, is composed of the current positions of other entities (targets and the other agents) and agent $i$'s last action.
An agent's observation about other agents' states, i.e.
$\phi_{s_{-i}}$, includes the positions of targets and the current and last positions of other agents, which are necessary for an agent to predict other agents' actions.
The size of the environment is $15\times15~m^2$.
The maximum episode length is $30$ timesteps.
Agents are homogeneous. They share a common policy and reward function.
The reward function is aligned with \cite{jin2020stabilizing}.
\begin{figure}[!t]
\centering
\includegraphics[width=0.75\columnwidth]{Env.png}
\captionsetup{font=footnotesize}
\caption{Illustration of the cooperative navigation task involving three agents. }\vspace{-1mm}\label{fig:control scene}
\label{env}
\end{figure}
Network structure and hyperparameters of IBORM are as follows.
The decision network has two hidden layers containing $300$ and $200$ units, respectively.
The encoder and classifier both have two hidden layers. Encoder has $32$ and $16$ units in its two layers, respectively. Classifier has $16$ and $32$ units in its two layers, respectively.
Adam optimizer with a learning rate of 0.01 is applied to update parameters.
The size of the replay buffer is 1500, the batch size of SGD is 32, $\gamma=1$ and the target network is updated with a soft update rate \cite{lillicrap2015continuous} $\eta = 0.001$, which are all aligned with SMADQN.
$\lambda_1, \lambda_2$ and $\lambda_3$ are set as 10, 0.001 and 0.1 by grid search method.
We compare IBORM with SMADQN and a comparable DRON whose DRL loss are aligned with SMADQN.
We name the latter SMADRON.
For SMADQN,
its network has two hidden layers containing the the same number of units as that in IBORM's decision network.
For SMADRON, its network structure is the same as IBORM.
Hyperparameters used in SMADQN and SMADRON are the same as IBORM.
\begin{table}[t]
\renewcommand\arraystretch{1.1}
\scriptsize
\captionsetup{font=footnotesize}
\caption{Test results of different methods.}
\centering
\begin{tabular}{l l l l l l}
\toprule [1pt]
\multirow{2}*{Method} & \multicolumn{5}{c}{Success rate} \\
& N=3 & N=4 & N=5 & N=6 &N=7 \\
\hline
SMADQN &98.2\% &97.8\% &96.1\% &91.2\% &0.0\% \\
SMADRON &98.9\% &96.9\% &92.9\% &93.5\% &82.5\% \\
IBORM &\textbf{99.3}\% &\textbf{98.1}\% &\textbf{97.1}\% &\textbf{93.5}\% &\textbf{87.8}\% \\
\bottomrule [1pt]
\end{tabular}
\label{success rate}
\end{table}
We train each method by 10k episodes. At the beginning of each episode, positions of targets and agents are generated randomly.
Each method is evaluated with different numbers of targets and agents ($N=3,4,5,6,7$).
Convergence curves of average episode reward are shown in Fig.\ref{CovergenceCurves}.
As we can see from the results, when the number of agents increases, IBORM learns faster than the other two methods, which indicates the advantage of IBORM over implicit behavior representation learning (SMADQN) and explicit behavior representation learning without considering information utility (SMADRON).
To test the performance of the learned policies,
we generate $1000$ testing tasks with random positions of targets and agents.
Table \ref{success rate} shows success rate of cooperative navigation with different policies, from which we can see that IBORM outperforms the other two methods consistently.
\begin{figure}[t]
\centering
\subfigure[]
{
\begin{minipage}[t]{4cm}
\centering
\includegraphics[width=1.05\columnwidth]{vsSMADQN.png}
\end{minipage}
}
\subfigure[]
{
\begin{minipage}[t]{4cm}
\centering
\includegraphics[width=1.05\columnwidth]{vsSMADRON.png}
\end{minipage}
}
\captionsetup{font=footnotesize}
\caption {
Convergence curves of average episode reward of different methods.
(a) IBORM vs. SMADQN.
(b) IBORM vs. SMADRON.
}
\label{CovergenceCurves}
\end{figure}
We also investigate how learning performance of IBORM changes when $\lambda_1$, $\lambda_2$, and $\lambda_3$ vary in (\ref{loss function_new}).
Fig.\ref{Lambda} shows the results in tasks of $N=6$. Three subfigures corresponds to $\lambda_1$, $\lambda_2$ and $\lambda_3$, respectively.
As shown in the results, the suitable value range of each $\lambda$ is relatively wide.
Fig.\ref{Lambda}(b) indicates large $\lambda_2$ can lead to failure of IBORM. This is because large $\lambda_2$ causes much information to be discarded.
Fig.\ref{Lambda}(c) shows that large $\lambda_3$ reduces learning speed and gets less rewards,
because when $\lambda_3$ is enlarged, $\lambda_1$ is weakened relatively and thus the learned representation captures insufficient information regarding the task.
Too small $\lambda_3$ also slows down learning, because the learned representation captures insufficient information about other agents' behaviors and thus provides less help to MADRL. Additionally, when $\lambda_2$ or $\lambda_3$ equals zero, the learning performance degenerates, which indicates the importance of each mutual information constraint in IBORM.
\begin{figure}[t]
\centering
\subfigure[]
{
\begin{minipage}[t]{2.5cm}
\centering
\includegraphics[width=1.1\columnwidth]{lambda_1.png}
\end{minipage}
}
\subfigure[]
{
\begin{minipage}[t]{2.5cm}
\centering
\includegraphics[width=1.1\columnwidth]{lambda_2.png}
\end{minipage}
}
\subfigure[]
{
\begin{minipage}[t]{2.5cm}
\centering
\includegraphics[width=1.1\columnwidth]{lambda_3.png}
\end{minipage}
}
\captionsetup{font=footnotesize}
\caption {
Sensitivity of IBORM to $\lambda$ values on tasks containing six targets and agents. Subfigures (a), (b) and (c) corresponds to $\lambda_1$, $\lambda_2$ and $\lambda_3$, respectively.
}
\label{Lambda}
\end{figure}
\section{Conclusion}
\label{sec:page}
We propose IBORM to facilitate MADRL by learning a compact and informative representation regarding other agents' behaviors.
We implement IBORM based on our recently proposed MADRL algorithm, SMADQN, by replacing the implicit behavior representation learning of SMADQN with information-bottleneck-based explicit behavior representation learning.
Experimental results demonstrate that IBORM learns faster and the resulting policies can achieve higher success rate consistently, as compared with implicit behavior representation learning (SMADQN) and explicit behavior representation learning (SMADRON) without considering information compression and utility.
\clearpage
\vfill\pagebreak
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Let $F$ be a finite subset of $\mathbb Z^n$, which we will refer to as a \emph{cluster}.
A \emph{tiling} of $\mathbb Z^n$ by translates of $F$, or an $F$-\emph{tiling}, is, roughly, a covering of $\mathbb Z^n$ by non-overlapping translates of $F$.
The \emph{periodic tiling conjecture} (See \cite{lagarias_wang_tiling}) states that if there is an $F$-tiling of $\mathbb Z^n$ for some cluster $F$, then there is an $F$-tiling which is invariant under translations by a finite index subgroup of $\mathbb Z^n$, that is, there is a \emph{fully periodic} $F$-tiling.
It was proved in \cite{beauquier_nivat_1991} that if $F$ is a \emph{connected} cluster (which roughly means that there are no `gaps' in $F$) in $\mathbb Z^2$ then every $F$-tiling is periodic in at least one direction.
This is sufficient to establish the existence of a fully periodic tiling by a simple pigeon-hole argument.
An important progress towards the periodic tiling conjecture was made in \cite{szegedy_algorithms_to_tile} in which it was proved that the periodic tiling conjecture holds (in any dimension) for clusters of prime cardinality.
Another major progress was made in \cite{bhattacharya_tilings} where it was established that the periodic tiling conjecture holds in $\mathbb Z^2$, without any constraint on the cardinality or the geometry of the cluster.
The higher dimensional cases remain open.
The main result of \cite{bhattacharya_tilings} shows that the \emph{orbit closure} of any $F$-tiling, where $F$ is a cluster in $\mathbb Z^2$, has a tiling which satisfies a certain weak notion of periodicity, which was shown to be sufficient to guarantee the existence of a fully periodic tiling.
To make this precise, first we define a subset $S$ of $\mathbb Z^2$ as \emph{$1$-periodic} if there is a nonzero vector $g$ in $\mathbb Z^2$ such that $g+S=S$.
We say that a subset $S$ of $\mathbb Z^2$ is \emph{weakly periodic} if $S$ can be partitioned in to finitely many $1$-periodic subsets.
Bhattacharya \cite{bhattacharya_tilings} shows that the orbit closure of any $F$-tiling has a weakly periodic tiling in it.
A stronger version of this result was obtained in \cite{greenfeld_tao_2020} which says that \emph{every} $F$-tiling of a cluster in $\mathbb Z^2$ is weakly periodic, thereby removing the need to pass to the orbit closure.
Periodicity results in dimension higher than $2$ are not known for clusters having cardinality different from a prime.
In this paper we show that if $F\subseteq \mathbb Z^3$ is a cluster of cardinality the square of a prime such that there exists an $F$-tiling of $\mathbb Z^3$, then there exists a weakly periodic $F$-tiling of $\mathbb Z^3$ (See Theorem \ref{theorem:prime square theorem}).
This makes some progress towards periodicity results in higher dimension and is an extension of the result of Szegedy \cite{szegedy_algorithms_to_tile} in $\mathbb Z^3$.
Unlike in two dimensions, the existence of a weakly periodic tiling alone does not seem (to the author) to imply the existence of a fully periodic tiling.
Let us make a remark about the connection of tiling problems with the theory of computation.
Suppose $F$ is a cluster in $\mathbb Z^n$ and we want to decide if $F$ can tile $\mathbb Z^n$ by translations.
It was shown in \cite{wang_1961} that if the periodic tiling conjecture holds then this problem is indeed decidable.
To see this, we just need to decide if there is a fully periodic $F$-tiling of $\mathbb Z^n$ (assuming the periodic tiling conjecture holds).
It is easy to see that if there is a fully periodic $F$-tiling, then there exists an $F$-tiling which is periodic in the direction of the standard basis vectors, and in fact the periods in any of these standard directions can be assumed to be the same.
Thus, enumerating over the set of all the $n$-cubes centered at the origin, we can check for each cube if there is a subset of a given cube when repeated periodically tiles the plane, or we must reach a cube which cannot be tiled by translates of $F$.\footnote{This is because if $F$ tiles arbitrary large cubes then $F$ must tile the whose space, as can be seen by a compactness argument, or by the Konig Infinity Lemma.}
\subsection{Overview of the Proof}
In \cite{bhattacharya_tilings} the problem of finding a fully periodic $F$-tiling for a cluster $F\subseteq \mathbb Z^2$ was first reduced to showing the existence of a weakly periodic $F$-tiling by an elementary combinatorial argument.
The existence of a weakly periodic $F$-tiling was then shown to follow from the following dynamical statement: If $\mathbb Z^2$ acts ergodically on a probability space $(X, \mu)$ and $A$ is a subset of $X$ such that finitely may $\mathbb Z^2$-translates of $A$ partition $X$, then $A$ itself is weakly periodic.
To approach this problem, Bhattacharya uses the spectral theorem to transfer the problem to $L^2(\mathbb T^2, \nu)$, where $\nu$ is the spectral measure associated with the characteristic function of $A$.
Then using a dilation lemma and an averaging argument, it was shown that the spectral measure is supported on finitely many $1$-dimensional affine subtoriii of $\mathbb T^2$.
Then the averages of $1_A$ along the `directions' of these subtorii (when $\mathbb T^2$ is thought of as $\mathbf R^2/\mathbb Z^2$, each $1$-dimensional subtorii can be thought of as a `line' and hence has a direction) were shown to behave polynomially in $\mathbf R/\mathbb Z$.
Since a polynomial in $\mathbf R/\mathbb Z$ is either periodic or equidistributes, the space $(X, \mu)$ gets partitioned into ergodic components of a finite index subgroup of $\mathbb Z^2$, such that on each ergodic component, the averages either `equidistribute' or are constant.
The rest of the proof goes by carefully dealing with the equidistribution case.
Thus a key step in the argument was to show that the spectral measure is supported on a `thin' subset of $\mathbb T^2$.
We take this as a cue and study the problem for a cluster $F\subseteq \mathbb Z^3$.
If the corresponding spectral measure again happens to be supported on finitely many lines then Bhattacharya's proof goes through \emph{mutatis mutandis}.
To deal with the adverse cases we make use of the hypothesis that $|F| = p^2$ for some prime $p$.
\subsection{Acknowledgements}
The author would like to thank, Siddhartha Bhattacharya, Ankit Rai, Mohit Upmanu, Nishant Chandgotia, Etienne Mutot, and Pierre Guillon for helpful conversations and their encouragement.
Thanks are due to Prof. Jaikumar Radhakrishnan for helpful suggestions and comments.
\section{Preliminaries}
\subsection{Tilings and Periodicity}
A finite subset of $\mathbb Z^n$ will be referred to as a \textbf{cluster}.
Given a cluster $F\subseteq\mathbb Z^n$ we say that $T$ is an $F$-tiling if for each $p\in \mathbb Z^n$ there exist unique $a\in F$ and $t\in T$ such that $p=a+t$.
In other words, $T$ is an $F$-tiling if and only if $\mathbb Z^n$ can be partitioned by translates of $F$.
We say that $F$ is \textbf{exact} if there exists an $F$-tiling.
Let $S$ be a subset of $\mathbb Z^n$ and $i\in \set{1, 2, 3}$.
We say that $S$ is \textbf{$i$-periodic} if there exists a rank-$i$ subgroup $\Lambda$ of $\mathbb Z^n$ such that $S$ is invariant under $\Lambda$, that is, $g+S=S$ for all $g\in \Lambda$.
A subset $S$ of $\mathbb Z^n$ is called \textbf{$1$-weakly periodic}, or simply \textbf{weakly periodic} if there exist finitely many $1$-periodic subsets $S_1, \ldots, S_k$ such that $S=S_1\sqcup \cdots \sqcup S_k$.
A subset $S$ of $\mathbb Z^n$ is called \textbf{$2$-weakly periodic} if there exist finitely many $2$-periodic subsets $S_1, \ldots, S_k$ such that $S=S_1\sqcup \cdots \sqcup S_k$.
In other words, for $i=1, 2$, $S$ is called $i$-weakly periodic if it can be partitioned into finitely may $i$-periodic subsets.
We will use the following results about tilings of $\mathbb Z$ and $\mathbb Z^2$.
\begin{theorem}
\label{theorem:uniform periodicity of one dimensional tilings}
Let $F\subseteq \mathbb Z$ be an exact cluster.
Then there is a positive integer $n$ such that $n+T=T$ for all $F$-tilings $T$.
\end{theorem}
\begin{proof}
Note that if $T|_{[a, b]}$ is known for some interval $[a, b]$ of length exceeding $\text{diam}(F)$, then $T$ is determined entirely.
Now the assertion follows by a pigeonhole argument.
\end{proof}
\begin{theorem}
\label{theorem:uniform periodicity of prime cardinality clusters in two dimensions}
\emph{\cite{szegedy_algorithms_to_tile}, \cite[Example 4]{kari_szabados_alg_geom}}
Let $F\subseteq\mathbb Z^2$ be an exact cluster of prime cardinality such that the affine span of $F$ has rank $2$.
Then there is a finite index subgroup $\Lambda$ of $\mathbb Z^2$ such that every $F$-tiling $T$ is $\Lambda$-invariant.
\end{theorem}
\begin{theorem}
\emph{\cite{bhattacharya_tilings,greenfeld_tao_2020}}
\label{theorem:existence of biperiodic tilings in two dimensions}
Let $F\subseteq\mathbb Z^2$ be an exact cluster, then there is a $2$-periodic $F$-tiling.
\end{theorem}
\subsection{Dynamical Formulation}
In \cite{bhattacharya_tilings} Bhattacharya proved the periodic tiling conjecture (See \cite{lagarias_wang_tiling}) in two dimensions by developing a dynamical statement which allows to transfer the problem to an ergodic theoretic setting.
We discuss the relevant definitions and state the dynamical formulation here which we will use later.
Let $X$ be a subshift of $\set{0, 1}^{\mathbb Z^3}$ and $\mu$ be a $\mathbb Z^3$-invariant probability measure on $X$.
The action of $\mathbb Z^3$ on $X$ naturally leads to a unitary action of $\mathbb Z^3$ on $L^2(X, \mu)$ as follows.
For $\varphi\in L^2(X, \mu)$ and $g\in \mathbb Z^3$ we define $g\cdot \varphi$ by
\begin{equation}
(g\cdot \varphi)(x) = \varphi(g^{-1}\cdot x) = \varphi((-g)\cdot x)
\end{equation}
The action of $\mathbb Z^3$ on $L^2(X, \mu)$ can be extended to an action of $\mathbb Z[u_1^{\pm}, u_2^{\pm}, u_2^{\pm}]$--- the ring of Laurent polynomial in three variables with integer coefficients.
To describe this action, first let us set up a convenient notation.
For a vector $v\in \mathbb Z^3$, we denote the monomial $u_1^{v_1}u_2^{v_2}u_2^{v_3}$ as $U^v$, and denote the ring $\mathbb Z[u_1^\pm, u_2^\pm, u_3^\pm]$ as $\mathbb Z[U^\pm]$.
Now any element of $\mathbb Z[U^\pm]$ can be written as $\sum_{v\in \mathbb Z^3}a_vU^v$, where only finitely many $a_v$'s are nonzero.
We define, for a Laurent polynomial $p=\sum_{v\in \mathbb Z^3}a_vU^v$ and $\varphi\in L^2(X, \mu)$, the function $p\cdot \varphi$ in $L^2(X, \mu)$ as
\begin{equation}
p\cdot \varphi = (\sum_v a_vU^v)\cdot \varphi = \sum_{v\in \mathbb Z^3} a_v(v\cdot \varphi)
\end{equation}
We say that $p$ \textbf{annihilates} $\varphi$ if $p\cdot \varphi$ is $0$.
For $i\in \set{1, 2, 3}$, we say that $\varphi\in L^2(X, \mu)$ is \textbf{$i$-periodic} if there is a rank-$i$ subgroup $\Lambda$ of $\mathbb Z^3$ such that $v\cdot \varphi=\varphi$ for all $v\in \Lambda$.\footnote{We emphasize that this equation is written in $L^2$ and hence, when $f$ is an actual function, it only says that $f$ and $v\cdot f$ agree almost everywhere and not necessarily everywhere.}
A measurable subset $B$ of $X$ will be called $i$-periodic if $1_B$ is $i$-periodic.
Finally, we define a measurable subset $B$ of $X$ to be \textbf{$2$-weakly periodic} if there exists a partition $B=B_1\sqcup \cdots \sqcup B_k$ of $B$ into finitely many measurable subsets $B_1, \ldots, B_k$ such that each $B_i$ is $2$-periodic.
Similarly, we say that $B$ is $1$-\textbf{weakly periodic}, or simply \textbf{weakly periodic} if $B$ can be partitioned into finitely many $1$-periodic subsets.
The following lemma is the $3$-dimensional analog of \cite[Secton 2]{bhattacharya_tilings}.
\begin{lemma}
\emph{\cite[Section 2]{bhattacharya_tilings}}
\label{label:bhattacharya correspondence principle}
Write $A=\set{x\in X:\ x(0) = 1}$.
If $A$ is $i$-periodic then $\mu$-almost every point in $X$ is $i$-periodic.
If $A$ is $i$-weakly periodic, then $\mu$-almost every point in $X$ is $i$-weakly periodic.
\end{lemma}
\subsection{Spectral Theorem}
\label{section:spectral theorem}
Let $\mathbb T^n$ be the $n$-dimensional torus.
Let $\nu$ be a probability measure on $\mathbb T^n$.
There is a canonical representation of $\mathbb Z^n$ on $L^2(\mathbb T^n, \nu)$ which we now describe.
Recall that the characters on $\mathbb T^n$ are in bijection with $\mathbb Z^n$.
Let us write $\chi_g:\mathbb T^n\to \mathbf C$ to denote the character corresponding to $g\in \mathbb Z^n$.
Then we define a map $\sigma_g:L^2(\mathbb T^n, \nu)\to L^2(\mathbb T^n, \nu)$ by writing $\sigma_g(\phi) = \chi_g\phi$, where the latter is the pointwise product of $\chi_g$ and $\phi$.
It can be easily checked that each $\sigma_g$ is in fact a unitary linear map.
Thus we get a map $\sigma:\mathbb Z^n\to \mathcal U(L^2(\mathbb T^n, \nu))$ which takes $g$ to $\sigma_g$.
By the Stone-Weierstrass theorem we have the $\mathbf C$-span of the characters are dense in $C(\mathbb T^n)$, where $C(\mathbb T^n)$ is the set of all the complex valued continuous functions on $\mathbb T^n$ equipped with the sup-norm topology.
Also, since $\mathbb T^n$ is a compact metric space, we have $C(\mathbb T^n)$ is dense in $L^2(\mathbb T^n, \nu)$.
Therefore the $\mathbf C$-span of the characters are dense in $L^2(\mathbb T^n, \nu)$.
From this we see that, if $\mathbf 1$ denotes the constant map which takes the value $1$ everywhere, $\text{Span}\set{\sigma_g\mathbf 1:\ g\in \mathbb Z^n}$ is dense in $L^2(\mathbb T^n, \nu)$.
In other words, $\mathbf 1$ is a \emph{cyclic vector} for this representation.
We now want to state a theorem which dictates that this is a defining property of unitary representations of $\mathbb Z^n$.
\begin{theorem}
\label{theorem:spectral theorem}
\textit{\textbf{Spectral Theorem.}}
Let $H$ be a Hilbert space and $\tau:\mathbb Z^n\to \mathcal U(H)$ be a unitary representation of $\mathbb Z^n$.
Suppose $v\in H$ is a cyclic vector, that is, $\text{Span}\set{\tau_gv:\ g\in \mathbb Z^n}$ is dense in $H$, and assume that $v$ has unit norm.
Then there is a unique probability measure $\nu$ on $\mathbb T^n$ and a unitary isomorphism $\theta: H \to L^2(\mathbb T^n, \nu)$ with $\theta(v)= \mathbf 1$ such that the following diagram commutes for all $g\in \mathbb Z^n$
\begin{figure}[H]
\centering
\begin{tikzcd
H\ar[r, "\theta"]\ar[d, "\tau_g"] & L^2(\mathbb T^n, \nu)\ar[d, "\sigma_g"]\\%\ar[r, "\theta"]\ar[d, "\sigma_g"] \\
H\ar[r, "\theta"]\ar[r, "\theta"] & L^2(\mathbb T^n, \nu
\end{tikzcd}
\end{figure}
\noindent
\end{theorem}
So the above theorem says that the abstract representation $\tau$ can be thought of as the canonical concrete representation $\sigma$, at the cost of a probability measure $\nu$.
Thus understanding the measure $\nu$ is equivalent to understanding $\tau$.
\section{Preparatory Lemmas}
\label{section:elementary results}
\label{section:preparatory results}
\subsection{Some Combinatorial Lemmas}
A subset $S$ of $\mathbb Z^3$ is called a \textbf{standard prism} if there exists a subset $A$ of $\mathbb Z^2$, called the \textbf{base} of $S$, and a subset $B$ of $\mathbb Z$ such that $S=A\times B$.
A subset $S$ of $\mathbb Z^3$ is called a \textbf{prism} if it can be transformed into a standard prism by applying translations and elements of $GL_3(\mathbb Z)$.
\begin{lemma}
\label{lemma:first combonatorial lemma precursor}
Let $F$ be a finite subset of $\mathbb Z^3$ and $p$ be a prime.
Let $g$ be a nonzero vector in $\mathbb Z^3$.
Assume that whenever $\pi$ is a plane in $\mathbf R^3$ parallel to $g$ we have $|\pi\cap F|$ is divisible by $p$.
Then whenever $\ell$ is a line parallel to $g$ we have $|\ell\cap F|$ is also divisible by $p$. (The converse is also true.)
\end{lemma}
\begin{proof}
We may assume that $g$ is primitive.
After applying a suitable $GL_3(\mathbb Z)$ transformation, we may assume that $g=(0, 0, 1)$.
Let $S$ be the image of $F$ under the map $f:\mathbf R^3\to \mathbf R^2$ defined by
\begin{equation}
f(x, y, z) = (x, y)
\end{equation}
for all $(x, y, z)\in \mathbb Z^3$, that is, $f$ is the orthogonal projection on the $x, y$-plane.
Let $(a, b)$ be an extreme point of the convex hull of $S$ in $\mathbf R^2$ and let $\ell$ be a line in $\mathbf R^2$ such that $\ell\cap S = \set{(a, b)}$.
Let $\pi$ be a plane parallel to $g$ in $\mathbf R^3$ such that the image of $\pi$ under $f$ is $\ell$.
Then $\pi\cap F = \set{(a, b, z)\in \mathbb Z^3:\ (a, b, z)\in F}$.
By hypothesis, $|\pi\cap F|$ has size divisible by $p$.
Now the set $F\setminus \pi\cap F$ satisfies the same hypothesis as $F$ and now we can finish inductively.
\end{proof}
\begin{lemma}
\label{lemma:first combinatorial lemma}
Let $F\subseteq \mathbb Z^3$ be a set of size $p^2$, where $p$ is a prime.
Suppose that there is a nonzero vector $g$ in $\mathbb Z^3$ such that whenever $\pi$ is a plane parallel to $g$ in $\mathbf R^3$ we have $|\pi\cap F|$ is divisible by $p$.
Also, assume that there is a plane $\pi'$ not parallel to $g$ such that $|\pi''\cap F|$ is divisible by $p$ whenever $\pi''$ is a plane parallel to $\pi'$.
Then $F$ is a prism with base of size $p$.
\end{lemma}
\begin{proof}
After a suitable $GL_3(\mathbb Z)$ transformation, we may assume that $g=(0, 0, 1)$ and $\pi'$ is a plane orthogonal to $g$.
By Lemma \ref{lemma:first combonatorial lemma precursor} we know that whenever $\ell$ is a line in $\mathbf R^3$ parallel to the $z$-axis, we have that $\ell\cap F$ has size divisible by $p$.
Let
$$
S=\set{z\in \mathbb Z: (x, y, z)\in F \text{ for some } x, y\in \mathbb Z}
$$
and let $z_1< \cdots < z_r$ are all the elements of $S$.
Let $\pi_i'$ be the plane orthogonal to $g$ and passing through $z_i$.
Then each $\pi_i'\cap F$ is non-empty and has size divisible by $p$.
Also, for any line parallel to $g$, we have $|\ell\cap F|\leq r$ and $p$ divides $|\ell\cap F|$, which gives $r\geq p$.
At the same time $p^2=|F|=\sum_{i=1}^r |\pi_i'\cap F|\geq pr$, giving $p\geq r$ .
Thus we must have $r=p$.
This means that each $\pi'_i\cap F$ has size exactly $p$.
From here the desired result easily follows.
\end{proof}
\begin{lemma}
\label{lemma:tiling by a prism with prime base}
Let $F=A\times B$ be a standard prism with base $A\subseteq \mathbb Z^2$.
Assume that $|A|$ is a prime $p$.
Then there is a $2$-periodic $F$-tiling.
\footnote{Using this lemma, a simple pigeonhole argument shows that there exists a $3$-periodic $F$-tiling.}
\end{lemma}
\begin{proof}
If the affine span of $F$ has dimension smaller than $3$ then the result follows by Theorem \ref{theorem:existence of biperiodic tilings in two dimensions} .
Thus we may assume that the dimension of the affine span of $F$ is $3$.
Thus the affine span of $A$ has dimension $2$.
Let $T$ be an $F$-tiling of $\mathbb Z^3$.
It is easy to see that for any integer $c$ the set $T_c := T\cap (\mathbb Z^2\times \set{c})$ is a tiling of $\mathbb Z^2\times \set{c}\cong \mathbb Z^2$ by translates of $A\times \set{c}\cong A$.
By Theorem \ref{theorem:uniform periodicity of prime cardinality clusters in two dimensions} we know that there is a rank-$2$ subgroup $\Lambda$ of $\mathbb Z^3$ such that $T_c$ is $\Lambda$-invariant for each $c$.
Thus $T$ is $\Lambda$-invariant and hence $2$-periodic.
\end{proof}
\begin{corollary}
If $F$ is a prism in $\mathbb Z^3$ with a base of prime size, then there is a $2$-periodic $F$-tiling.
\end{corollary}
\begin{lemma}
\label{lemma:dilation lemma for higher level tilings}
\textit{\textbf{Dilation Lemma.}}
\emph{\cite[Corollary 11]{horak_kim_algebraic_method}}
Let $F\subseteq \mathbb Z^3$.
If $T$ is an $F$-tiling then for all $\alpha$ relatively prime with $|F|$ we have $T$ is also an $\alpha F$-tiling, where $\alpha F=\set{\alpha a:\ a\in F}$.
\end{lemma}
It should be noted that various authors (\cite{tijdeman_decomposition_of}, \cite{szegedy_algorithms_to_tile}, \cite{kari_szabados_alg_geom}, \cite{bhattacharya_tilings}, \cite{greenfeld_tao_2020}) had discovered the above lemma in one form or another.
\subsection{An Analytical Lemma}
An element $\gamma\in S^1$ is said to be \textbf{irrational} if it is equal to $e^{i\theta}$ for some $\theta$ which is an irrational multiple of $2\pi$ .
Equivalently, $\gamma\in S^1$ is irrational if there is no non-trivial character of $S^1$ in whose kernel $\gamma$ lies.
More generally, elements $\gamma_1, \ldots, \gamma_m$ in $S^1$ are called \textbf{rationally independent} if there is no non-trivial character $\chi$ of $(S^1)^m$ such that $\chi(\gamma_1, \ldots, \gamma_m) = 1$.
\begin{lemma}
\label{lemma:non measure preparatory lemma}
Let $\gamma_1, \ldots, \gamma_n$ be irrational elements in $S^1$ and $x_1, \ldots, x_n$ be complex numbers.
Assume that
\begin{equation}
(\gamma_1^k - 1) x_1 + \cdots + (\gamma_n^k - 1) x_n \in \mathbb Z
\end{equation}
for all non-negative integers $k$.
Then the above expression is $0$ for all $k$.
\end{lemma}
\begin{proof}
Let $m$ be the size of a maximal rationally independent subset of $\set{\gamma_1, \ldots, \gamma_n}$.
By renumbering the $\gamma_i$'s and $x_i$'s if required, we may assume that $\set{\gamma_1, \ldots, \gamma_m}$ is a maximal rationally independent subset of $\set{\gamma_1, \ldots, \gamma_n}$.
Thus we can find vectors $v_1, \ldots, v_n$ in $\mathbb Z^m$ such that
\begin{equation}
\label{equation:expression for mu}
\gamma_1^{v_i(1)} \cdots \gamma_m^{v_i(m)} = \gamma_i
\end{equation}
for $i=1, \ldots, n$, where $v(i)$ denotes the $i$-th coordinate of any $v\in \mathbb Z^m$.
Define the Laurent polynomial $f(U)$ in $\mathbf C[U_1^\pm, \ldots, U_m^\pm]$ as
\begin{equation}
f(U) = (U^{v_1} - 1)x_1 + \cdots + (U^{v_n} - 1)x_n
\end{equation}
Then we have
\begin{equation}
\label{equation:hook equation density}
f(\gamma_1^k, \ldots, \gamma_m^k)
=
(\gamma_1^k - 1)x_1 + \cdots + (\gamma_n^k - 1)x_n
\end{equation}
By the rational independence of $\gamma_1, \ldots, \gamma_m$, we know that the set $\set{(\gamma_1^k, \ldots, \gamma_m^k):\ k\geq 0}$ is dense in $(S^1)^m$.\footnote{See Theorem 4.14 in \cite{einsiedler_ward_ergodic_theory}.}
Thus the image of $f$ on a dense set of $(S^1)^m$ is contained in $\mathbb Z$.
Since $f$ is continuous, this implies that the image of $(S^1)^m$ under $f$ is contained in $\mathbb Z$.
The connectedness of $(S^1)^m$ now yields that the image of $(S^1)^m$ under $f$ is a singleton.
However, for any $\varepsilon>0$, we can find a $k$ such that $|\gamma_i^k-1|<\varepsilon$ for each $1\leq i\leq n$, and thus we must have that $f$ is identically zero on $(S^1)^m$, finishing the proof.
\end{proof}
\subsection{An Elementary Fact about the Kernel of Characters}
\begin{lemma}
\label{lemma:kernel of character second lemma}
Let $h$ be a primitive vector in $\mathbb Z^2$.
Then $\ker(\chi_h)$ is precisely the set
\begin{equation}
\bigcup_{v\in \mathbb Z^3:\ v\perp h} \set{vt\in \mathbf R^3/\mathbb Z^3:\ t\in \mathbf R}
\end{equation}
\end{lemma}
\begin{proof}
Recall that $GL_3(\mathbb Z)$ acts transitively on the set of the primitive vectors in $\mathbb Z^3\setminus\set{0}$.
This after applying a suitable $GL_3(\mathbb Z)$ transformation, we may assume that $h=(1, 0, 0)$, in which case the lemma easily follows.
\end{proof}
\begin{lemma}
\label{lemma:kernel of character final lemma}
Let $h$ be a primitive vector in $\mathbb Z^3$ and $n$ be a positive integer.
Then
\begin{equation}
\ker\chi_h
=
\bigcup_{v\in \mathbb Z^3:v\perp h}\set{vt\in \mathbf R^3/\mathbb Z^3:\ t\in \mathbf R}
\end{equation}
and
\begin{equation}
\ker(\chi_{n h})
=
\bigcup_{0\leq i, j, k< n} \lrb{\lrp{\frac{i}{n}, \frac{j}{n}, \frac{k}{n}} + \ker(\chi_h)}
\end{equation}
\end{lemma}
\begin{proof
A point $(x, y, z)\in \mathbf R^3/\mathbb Z^3$ is in $\ker(\chi_{nh})$ if and only if $(nx, ny, nz)$ is in $\ker\chi_h$.
Thus the set
\begin{equation}
\bigcup_{0\leq i, j, k< n} \lrp{\frac{i}{n}, \frac{j}{n}, \frac{k}{n}} + \ker(\chi_h)
\end{equation}
is contained in $\ker(\chi_{nh})$.
For the reverse containment, let $(x, y, z)$ be in $\ker(\chi_{nh})$, that is $(nx, ny, nz)$ in $\ker\chi_h$.
Then by Lemma \ref{lemma:kernel of character second lemma} there exists $t\in \mathbf R$ and $v\in \mathbb Z^3$ orthogonal to $h$ such that $(nx, ny, nz) = vt\pmod{\mathbb Z^3}$.
Thus there exists integers $i, j$ and $k$ such that $(nx, ny, nz) = (i, j, k) + vt$, whence the desired result is clear.
\end{proof}
\subsection{Algebraic Lemmas}
\begin{lemma}
\label{lemma:cyclotomy divisibility lemma}
Let $p$ be a prime and $k$ be a positive integer.
Let $\zeta$ be a primitive $p^k$-th root of unity.
Suppose there are integers $a_1, \ldots, a_m$ such that
\begin{equation}
\zeta^{a_1} + \cdots + \zeta^{a_m} = 0
\end{equation}
Then $p$ divides $m$.
\end{lemma}
\begin{proof}
Without loss of generality we may assume that the $a_i$'s are all non-negative.
Define the polynomial $Q(z) = z^{a_1} + \cdots + z^{a_m}$.
Then since $\zeta$ is a root of $Q(z)$, we have $Q(z)$ is divisible by the $p^k$-th cyclotomic polynomial $\Phi_{p^k}(z)$.
Let $\Psi(z)$ be an integer polynomial such that $Q(z) = \Phi_{p^k}(z)\Psi(z)$.
Now using the fact that $\Phi_{p^k}(z) = \Phi_p(z^{k-1})$ we have $Q(z) = \Phi_p(z^{p^{k-1}})\Psi(z)$.
Now substituting $z=1$ gives $m = p \Psi(1)$ and hence $p$ divides $m$.
\end{proof}
For any finite subset $S$ of $\mathbb Z^3$, we will write $Z(\sum_{g\in S} \chi_{g})$ to mean the set of all the points in $\mathbb T^3$ on which $\sum_{g\in S}\chi_{g}$ vanishes.
\begin{lemma}
\label{lemma:first spectral support lemma}
\emph{(See \cite[Lemma 3.2]{bhattacharya_tilings})}
Let $F\subseteq \mathbb Z^3$ be a finite set containing the origin.
Then there is a finite subset $\Delta$ of $\mathbb Z^3\setminus \set{0}$ with pairwise linearly independent elements such that
\begin{equation}
Z:=
\bigcap_{\alpha\text{ coprime to } |F|}
Z\lrb{\sum_{g\in F} \chi_{\alpha g}}
\subseteq
\bigcup_{h\in \Delta} \ker\chi_h
\end{equation}
\end{lemma}
\begin{proof}
Let $n$ be the product of all primes which divide $|F|$.
Then for each non-negative integer $k$ we have $nk + 1$ is relatively prime to $|F|$.
Fix $(a, b, c)\in Z$.
Then we have, for all non-negative integers $k$ that
\begin{equation}
\sum_{g\in F} \chi_{(nk+1)g}(a, b, c) = 0
\end{equation}
giving
\begin{equation}
\sum_{g\in F\setminus\set{0}} \chi_{(nk+1)g}(a, b, c)
=
\sum_{g\in F\setminus\set{0}} a^{g_1}b^{g_2}c^{g_3}(a^{ng_1}b^{ng_2}c^{ng_3})^k
=
-1
\end{equation}
Therefore
\begin{equation}
\sum_{g\in F\setminus\set{0}} a^{g_1}b^{g_2}c^{g_3}\lrb{ \frac{1}{N}\sum_{k=0}^{N-1}
(a^{ng_1}b^{ng_2}c^{ng_3})^k
}
=
-1
\end{equation}
Taking the limit $N\to \infty$ we see that there must exist $g\in F\setminus{\set{0}}$ such that $a^{ng_1}b^{ng_2}c^{ng_3}$ is $1$,\footnote{
This is because if $z\in S^1$ then the average $(z+z^2+ \cdots + z^N)/N$ does not converge to zero if and only if $z=1$.
}
and thus $(a, b, c)\in \ker(\chi_{ng})$.
Therefore $Z\subseteq \bigcup_{g\in F\setminus \set{0}} \ker\chi_{ng}$.
Let $\Delta$ be a minimal subset of $F\setminus \set{0}$ such that $Z\subseteq \bigcup_{g\in \Delta} \ker\chi_{ng}$.
It follows that the elements of $\Delta$ must be pairwise linearly independent and we are done.
\end{proof}
\begin{lemma}
\label{lemma:second spectral support lemma}
Let $F$ be a cluster in $\mathbb Z^3$ containing the origin.
Let $h$ be an arbitrary nonzero vector in $\mathbb Z^3$ and $h_0$ be a primitive vector such that $h$ is a scale of $h_0$.
Then at least one of the following must happen.
\begin{enumerate
\item Every line in $\mathbf R^3$ that is parallel to $h$ intersects $F$ in a set of size divisible by $p$.
\item There is a finite set $\Gamma\subseteq \mathbb Z^3\setminus \set{0}$ such that each element of $\Gamma$ is linearly independent with $h$ and
\begin{equation}
\ker\chi_h
\cap
\bigcap_{\alpha\text{ coprime to } |F|}
Z\lrb{\sum_{g\in F} \chi_{\alpha g}}
\subseteq
\bigcup_{v\in \Gamma} \ker\chi_h\cap \ker\chi_v
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
After applying a suitable $GL_3(\mathbb Z)$ transformation, we may assume that $h=(0, 0, m)$ for some positive integer $m$.\footnote{
This is because of the following.
For any linear map $T:\mathbf R^3\to \mathbf R^3$ such that $T\in GL_3(\mathbb Z)$, and any $F\subseteq\mathbb Z^3$ finite, we have $Tp\in Z\lrb{\sum_{g\in F} \chi_g}$ if and only if $(T^{-1})^*p\in Z\lrb{\sum_{g\in T(F)}\chi_{g}}$.
}
Then $h_0=(0, 0, 1)$ and any coset of $\ker\chi_{h_0}$ is equal to $(1, 1, \omega)\cdot \ker\chi_{h_0}$ for some $m$-th root of unity $\omega$.
Fix an $m$-th root of unity $\omega$ and let $d$ be a positive integer such that $\omega$ is a primitive $d$-th root of unity.
Let $d = p^r\beta$ where $\beta$ is relatively prime to $p$.
Let $(a, b, c) \in ((1, 1, \omega)\cdot \ker\chi_{h_0})\cap Z$, where
\begin{equation}
Z=\bigcap_{\alpha\text{ coprime to } |F|} Z\lrb{\sum_{g\in F} \chi_{\alpha g}}
\end{equation}
Then $c=\omega$ and
\begin{equation}
\sum_{g\in F} \chi_{\alpha g}(a, b, \omega) = 0
\end{equation}
for all $\alpha$ relatively prime with $|F|$.
Let $n$ be the product of all the primes dividing $|F|$.
For all non-negative integers $k$, we have $mn k + 1$ is relatively prime to $|F|$, and hence so is $(mnk+1)\beta$.
So we have
\begin{equation}
\label{equation:conundrum}
\sum_{g\in F} \chi_{(mnk+1)\beta g}(a, b, \omega)
=
\sum_{g\in F} a^{\beta g_1} b^{\beta g_2}\omega^{\beta g_3} \chi_{mnk\beta(g_1, g_2)}(a, b)
=
0
\end{equation}
Let $\omega^\beta = \zeta$, and hence $\zeta$ is a primitive $p^r$-th root of unity.
Let $\sim$ be a relation on $F$ defined as $g\sim g'$ for $g, g'\in F$ if and only if $(g_1, g_2) = (g_1', g_2')$.
Then $\sim$ is an equivalence relation, and let $\mathcal E$ be the set of all the equivalence classes of $\sim$.
For each $E\in \mathcal E$ choose an element $g^E$ of $E$.
For any $v\in \mathbb Z^3$ let $\hat v$ denote the vector in $\mathbb Z^2$ obtained by dropping the last coordinate of $v$.
Then Equation \ref{equation:conundrum} gives
\begin{equation}
\label{equation:second conundrum}
\sum_{E\in \mathcal E} \lrp{\sum_{g\in E} \zeta^{g_3}}\chi_{\beta \hat g^E}(a, b) \chi_{mnk\beta \hat g^E}(a, b)
=
0
\end{equation}
for all $k$ non-negative.
Let $\theta_E = \sum_{g\in E} \zeta^{g_3}$ and $\gamma_E= \theta_E \chi_{\beta \hat g^E}(a, b)$ for all $E$ in $\mathcal E$.
If $\theta_E$ is $0$ for all $E\in \mathcal E$, then by Lemma \ref{lemma:cyclotomy divisibility lemma} we have each $|E|$ is divisible by $p$.
This would mean precisely that (1) holds.
So we may assume that there is some $E_0\in \mathcal E$ such that $\theta_{E_0}$, or equivalently $\gamma_{E_0}$, is nonzero.
Then from Equation \ref{equation:second conundrum} we have
\begin{equation}
\chi_{-mnk \beta\hat g^{E_0}}(a, b)\lrb{\sum_{E\in \mathcal E} \gamma_E \chi_{mnk\beta \hat g^E}(a, b)}
=
0
\end{equation}
which gives
\begin{equation}
\gamma_{E_0} + \sum_{E\in \mathcal E\setminus\set{E_0}} \gamma_E \chi_{mnk\beta (\hat g^E - \hat g^{E_0})}(a, b)
=
\gamma_{E_0} + \sum_{E\in \mathcal E\setminus\set{E_0}} \gamma_E \chi_{mn\beta (\hat g^E - \hat g^{E_0})}(a, b)^k
=
0
\end{equation}
Averaging we have
\begin{equation}
\gamma_{E_0} + \sum_{E\in \mathcal E\setminus\set{E_0}} \gamma_E\lrb{ \frac{1}{N}\sum_{k=0}^{N-1}\chi_{mn\beta (\hat g^E - \hat g^{E_0})}(a, b)^k}
=
0
\end{equation}
Taking limits, we must have, since $\gamma_{E_0}\neq 0$, that $(a, b)\in \ker\chi_{mn\beta (\hat g^{E} - \hat g^{E_0})}$ for some $E\neq E_0$ in $\mathcal E$.
Define $h^E = mn\beta(g^{E}_1-g^{E_0}_1, g^{E}_2 - g^{E_0}_2, 1)$.
Then $h^E$ is linearly independent with $h$ since $\hat g^E - \hat g^{E_0}$ is nonzero and also $(a, b, c)\in \ker\chi_{h^E}$.
This gives the desired result.
\end{proof}
\begin{lemma}
\label{lemma:main technical precursor to main}
Let $F\subseteq \mathbb Z^3$ be a cluster containing the origin and $p$ be a prime.
Let $\rho$ be a rational point in $\mathbf R^3/\mathbb Z^3$ and $v$ be a nonzero vector in $\mathbb Z^3$ such that infinitely many points of the set
\begin{equation}
\set{\rho + vt\in \mathbf R^3/\mathbb Z^3:\ t\in \mathbf R}
\end{equation}
are in
\begin{equation}
\label{equation:partition equations}
Z = \bigcap_{\alpha\text{ coprime to } p} Z\lrb{\sum_{g\in F} \chi_{\alpha g}}
\end{equation}
Then any plane parallel to $\set{w\in \mathbf R^3:\ \ab{w, v} = 0}$ intersects $F$ in a set of size divisible by $p$.
\end{lemma}
\begin{proof}
Let $h_0$ be a primitive vector in $\mathbb Z^3\setminus \set{0}$ and $r$ be a rational number such that $\rho = r h_0$ in $\mathbf R^3/\mathbb Z^3$.
Then, by hypothesis, infinitely many points of
\begin{equation}
\set{rh_0+ vt\in \mathbf R^3/\mathbb Z^3:\ 0\leq t< 1}
\end{equation}
are in $Z$.
Let $r=c/d$, where $c$ and $d$ are relatively prime integers.
We get, for each $\alpha$ coprime to $p$,
\begin{equation}
\sum_{g\in F} e^{2\pi i r\alpha \ab{g, h_0}}e^{2\pi i \alpha \ab{g, v} t} = 0
\end{equation}
for infinitely many $t\in [0, 1)$.
Therefore, the (Laurent) polynomial
\begin{equation}
\label{equation:intermediate equation}
\sum_{g\in F} e^{2\pi i r \alpha \ab{g, h_0}} z^{\alpha \ab{g, v}}
\end{equation}
is satisfied by infinitely many $z\in S^1$ for each $\alpha$ coprime to $p$.
Let $d= p^m \beta$, where $m$ is a non-negative integer and $\beta$ is coprime to $p$.
Thus, by Equation \ref{equation:intermediate equation} the polynomial
\begin{equation}
\sum_{g\in F} e^{2\pi i r \beta \ab{g, h_0}} z^{\beta \ab{g, v}}
=
\sum_{g\in F} e^{2\pi i \ab{g, h_0} c/p^m} z^{\beta \ab{g, v}}
\end{equation}
has infinitely many solutions in $S^1$ and is hence identically zero.
Define an equivalence relation $\sim_v$ on $F$ by writing $g_1\sim_v g_2$ for $g_1, g_2\in F$ if $\ab{g_1, v}=\ab{g_2, v}$.
Let $F_1, \ldots, F_l$ be all the equivalence classes in $F$.
The coefficients of the above polynomial are
\begin{equation}
\label{equation:coefficients of the polynomial}
\sum_{g\in F_j} e^{2\pi i \ab{g, h_0} c/p^m}, \quad j=1, \ldots, l.
\end{equation}
and hence each of these terms are $0$.
If $m=0$ then this cannot happen and hence we must have $m\geq 1$.
This implies that $p$ divides $d$ and hence, since $c$ is relatively prime to $d$, we have $c$ is coprime to $p$.
Let $\zeta$ denote the complex number $e^{2\pi i/p^m}$.
By Equation \ref{equation:coefficients of the polynomial} we have $\zeta$ is a root of each of the (Laurent) polynomials
\begin{equation}
\sum_{g\in F_j} z^{\ab{g, h}c}, \quad j= 1, \ldots, l
\end{equation}
Therefore by Lemma \ref{lemma:cyclotomy divisibility lemma} we have that each $|F_j|$ is divisible by $p$.
This implies that any plane parallel to $\set{w\in \mathbf R^3:\ \ab{w, v} = 0}$ intersects $F$ in a set of cardinality divisible by $p$ and we are done.
\end{proof}
\section{The Periodicity Result}
\subsection{Main Theorem}
\label{subsection:main theorem}
\begin{theorem}
\label{theorem:prime square theorem}
Let $F$ be an exact cluster in $\mathbb Z^3$ with cardinality $p^2$, where $p$ is a prime.
Then there is a $1$-weakly periodic $F$-tiling.
\end{theorem}
\subsection{Proof}
\subsubsection{Notation}
\label{subsubsection:notation in the proof of main theorem}
Let $X\subseteq \set{0, 1}^{\mathbb Z^3}$ be the set of all the $F$-tilings and $\mu$ be a $\mathbb Z^3$-ergodic probability measure on $X$, which we may assume to be concentrated on the orbit closure of a given $F$-tiling $T$.
We define $A=\set{x\in X:\ x(0) = 1}$ and $f\in L^2(X, \mu)$ as the characteristic function on $A$.
Let $\nu$ be the spectral measure associated to the unit vector $f/\norm{f}_2$ in $L^2(X, \mu)$ and $\theta$ be the unitary isomorphism between $H_f$ --- the span closure of the orbit of $f/\norm{f}_2$ --- and $L^2(\mathbb T^3, \nu)$ as discussed in Theorem \ref{theorem:spectral theorem}.
\emph{We will use $\mathbb T$ to denote the unit circle and not $\mathbf R/\mathbb Z$.}
By Lemma \ref{lemma:dilation lemma for higher level tilings} we have $T$ is an $(\alpha F)$-tiling whenever $\alpha$ is relatively prime to $p$.
Thus
\begin{equation}
\sum_{g\in F} 1_{\alpha g A} = \sum_{g\in F} (\alpha g)\cdot 1_A = 1_X
\end{equation}
for all $\alpha$ relatively prime to $p$.
Now $1_X$ is $\mathbb Z^3$-invariant in $L^2(X, \mu)$, and hence $\theta(1_X)$ is $\mathbb Z^3$-invariant in $L^2(\mathbb T^3, \nu)$.
But the only $\mathbb Z^3$-invariant members of $L^2(\mathbb T^3, \nu)$ are the ones in the span of $\delta_{\set{1}}$---the Dirac function concentrated at the identity of $\mathbb T^3$.
Thus, applying $\theta$, we get
\begin{equation}
\sum_{g\in F} \chi_{\alpha g} = c \delta_{\set{1}}, \quad \alpha \textup{ coprime to } p
\end{equation}
for some constant $c$.
Therefore, the spectral measure $\nu$ is supported on $Z\cup\set{1}$, where
\begin{equation}
Z = \bigcap_{\alpha\text{ coprime to } p} \lrb{\sum_{g\in F} \chi_{\alpha g}}
\end{equation}
By Lemma \ref{lemma:first spectral support lemma} we know that there is a finite set $\Delta\subseteq \mathbb Z^3\setminus \set{0}$ whose elements are pairwise linearly independent such that the spectral measure $\nu$ associated to $f/\norm{f}_2$ is supported on $\bigcup_{g\in \Delta}\ker(\chi_g)$.
\subsubsection{Case 1: There exist at least two distinct members $g_0$ and $g_1$ in $\Delta$ such that every line parallel to either $g_0$ or $g_1$ intersects $F$ in a set of cardinality divisible by $p$
By Lemma \ref{lemma:first combinatorial lemma} we deduce that $F$ is a prism with prime base and hence by Lemma \ref{lemma:tiling by a prism with prime base} we see that there is a $2$-periodic $F$-tiling, which is in particular $1$-weakly periodic.
\subsubsection{Case 2: There is exactly one $g_0$ in $\Delta$ such that every line parallel to $g_0$ intersects $F$ in a set of cardinality divisible by $p$.}
We further divide this into two subcases.
First assume that there is $g_1$ distinct from $g_0$ in $\Delta$ such that
$$
(Z\cap \ker\chi_{g_1})\setminus \ker\chi_{g_0}
$$
is infinite.
Then by Lemma \ref{lemma:second spectral support lemma} we can find a nonzero vector $v\in \mathbb Z^3$ orthogonal to $g_1$ and a rational point $\rho\in \mathbf R^3/\mathbb Z^3$ such that infinitely many elements of
\begin{equation}
\label{equation:scary equation}
\set{\rho+vt\in \mathbf R^3/\mathbb Z^3:\ t\in \mathbf R}
\end{equation}
are in $Z$.
If $v$ is orthogonal to $g_0$ then the set in Equation \ref{equation:scary equation} is already contained in $\ker\chi_{g_0}$.
So we may assume that $v$ is not orthogonal to $g_0$.
Now by Lemma \ref{lemma:main technical precursor to main} we have that whenever $\pi$ is a plane parallel to $\pi_1:=\set{w\in \mathbf R^3:\ \ab{v, w} = 0}$, we have $|\pi\cap F|$ is divisible by $p$.
Also, by Lemma \ref{lemma:first combonatorial lemma precursor} we already know that whenever $\ell$ is a plane parallel to $g_0$ we have $|\pi\cap F|$ is divisible by $p$.
By the fact that $v$ is not orthogonal to $g_0$, we have that a plane parallel to $g_0$ cannot be parallel to $\pi_1$.
Then, again, by Lemma \ref{lemma:first combinatorial lemma} we deduce that $F$ is a prism with prime base and hence by Lemma \ref{lemma:tiling by a prism with prime base} we see that there is a $2$-periodic $F$-tiling, which is in particular $1$-weakly periodic.
Now assume that $(Z\cap \ker\chi_{h})\setminus \ker\chi_{g_0}$ is finite whenever $h\in \Delta$ is different from $g_0$.
For each $h\in \Delta\setminus\set{g_0}$, let $S_h=(Z\cap \ker\chi_h)\setminus \ker\chi_{g_0}$.
Then, by our assumption, each $S_h$ is finite.
Since $\supp(\nu)$ is contained in $Z\cap \bigcup_{g\in \Delta} \ker\chi_{g}$, we see that
\begin{equation}
\ker(\chi_{g_0}) \cup \bigcup_{h\in \Delta:\ h\neq g_0} S_h
\end{equation}
is a $\nu$-full measure set.
So we infer that there is a nonzero vector $g_0$ in $\mathbb Z^3$ and a finite set $S\subseteq \mathbb T^3$ disjoint with $\ker\chi_{g_0}$ such that $\nu$ is supported on $\ker(\chi_{g_0})\cup S$.
We may assume that $S$ has smallest size with this property and hence each element of $S$ has positive mass under $\nu$.
The minimality of the size of $S$ also implies that $\chi_{g_0}(s)$ is irrational for all $s\in S$, for otherwise we could replace $g_0$ by a scale of itself and reduce the size of $S$.
We will show that $S$ is empty.
Assume on the contrary that $S$ is non-empty.
Define an equivalence relation $\sim$ on $S$ by writing $p\sim q$ for $p, q\in S$ if $\chi_{g_0}(p)=\chi_{g_0}(q)$.
Let $\mathcal E$ be the set of all the equivalence classes.
Thus
\begin{equation}
1 = 1_{\ker(\chi_{g_0})} + \sum_{E\in \mathcal E} 1_E
\end{equation}
in $L^2(\mathbb T^3, \nu)$.
Choose a representative $p_E\in E$ for each $E\in \mathcal E$.
Now acting both sides of the above equation by $ng_0$, where $n$ is any non-negative integer, we get
\begin{equation}
\chi_{n g_0} = 1_{\ker(\chi_{g_0})} + \sum_{E\in \mathcal E}\chi_{g_0}(p_E)^n 1_E
\end{equation}
Going back into the $L^2(X, \mu)$ world by applying $\theta^{-1}$, we get that
\begin{equation}
\label{equation:hook super equation}
(n g_0)\cdot f = f^{g_0} + \sum_{E\in \mathcal E} \chi_{g_0}(p_E)^n \varphi_E
\end{equation}
where $\varphi_E = \norm{f}_2 \theta^{-1}(1_E)$, and $f^{g_0}$ is the orthogonal projection of $f$ onto the space of $g_0$-invariant functions $L^2(X, \mu)$.\footnote{By the Birkhoff ergodic theorem we see that $f^{g_0}$ lies in $H_f$.
Also, $1_{\ker\chi_{g_0}}$ is the orthogonal projection of $1$ onto the space of $g_0$-invariant functions in $L^2(\mathbb T^2, \nu)$. The fact that $\theta$ is a unitary isomorphism shows that $\theta^{-1}(1_{\ker\chi_{g_0}})$ is same as $f^{g_0}/\norm{f}_2$.}
Therefore, for each $n\geq 0$, we have
\begin{equation}
(ng_0)\cdot f - f = \sum_{E\in \mathcal E} (\chi_{g_0}(p_E)^n - 1)\varphi_E
\end{equation}
Let $Y\subseteq X$ be a $\mu$-full measure subset of $X$ such that
\begin{equation}
[(ng_0)\cdot f](y) - f(y) = \sum_{E\in \mathcal E}(\chi_{g_0}(p_E)^n - 1) \varphi_E(y)
\end{equation}
for all $y\in Y$ and all $n\geq 0$.
Therefore
\begin{equation}
\sum_{E\in \mathcal E} (\chi_{g_0}(p_E)^n - 1) \varphi_E(y) \in \mathbb Z
\end{equation}
for all $y\in Y$, and all $n\geq 0$.
But since each $\chi_{g_0}(p_E)$ is irrational, we may apply Lemma \ref{lemma:non measure preparatory lemma} to deduce that for any $y\in Y$ the only value the above expression can take, for any non-negative integer $n$, and hence in particular for $n=1$, is $0$.
Therefore $[g_0 \cdot f] (y) - f(y)$ is $0$ for each $y\in Y$.
But since $Y$ is a full measure set in $X$, we infer that $g_0\cdot f = f$ in $L^2(X, \mu)$.
Thus $f$ is $1$-periodic, and hence by \ref{label:bhattacharya correspondence principle} we deduce that the orbit closure of $T$ has a $1$-periodic point in it, which is in particular $1$-weakly periodic.
\subsubsection{Case 3: There is no $g$ in $\Delta$ such that every line parallel to $g$ intersects $F$ in a set of size divisible by $p$}
\label{subsubsection:last case in the proof}
In this case, by Lemma \ref{lemma:first spectral support lemma} and \ref{lemma:second spectral support lemma} we deduce that there exist nonzero vectors $v_1, \ldots, v_n\in \mathbb Z^3$ and finite subsets $S_1, \ldots, S_n\subseteq \mathbf R^3/\mathbb Z^3$ such that each member of each $S_i$ is a rational point and the measure $\nu$ is supported on\footnote{This is because for linearly independent vectors $g$ and $h$ in $\mathbb Z^3$ we have $\ker\chi_g\cap \ker\chi_h$ is equal to $S+\set{tv\in \mathbf R^3/\mathbb Z^3:\ t\in \mathbf R}$ for some nonzero $v\in \mathbb Z^3$ orthogonal to both $g$ and $h$ and some finite set $S$ of rational points in $\mathbf R^3/\mathbb Z^3$.}
\begin{equation}
\bigcup_{i=1}^n (S_i+ \set{t v_i\in \mathbf R^3/\mathbb Z^3:\ t\in\mathbf R})
\end{equation}
The proof now closely follows the proof of periodic tiling conjecture on $\mathbb Z^2$ given in \cite{bhattacharya_tilings}.
We give the full details in Section \ref{section:bhattachayas proof in three dimension} appendix.
|
\section{Introduction}
Machine Learning (ML) includes methods, tools, and techniques for inferring models from data and has provided successful applications of classification and prediction algorithms. In the area of software development and evolution, a recent study \cite{shafiq2020machine} revealed that there is a spectrum of applications of ML across the software development life-cycle, with most of the applications belonging to the category of \textit{Quality Assurance and Analytics}.
There exists an extensive variety of ML algorithms and this pool is growing steadily. A recent study \cite{shafiq2020machine} listed Decision Trees, Naive Bayes, and Random Forrest as the techniques most frequently applied in Software Engineering. However, it is important to determine which algorithm works well for a given problem and which are less effective. The performance of any ML technique is generally measured in terms of accuracy (or similar measures). However, the success of ML does not only depend on the algorithms used because ML is a process with various interdependent steps and the investments made in this process need to be related to the return gained from its results. This paper puts estimating the return-on-investment (ROI) of ML in the spotlight. ROI is most widely used in the context of business analysis, which we extend to ML classification problems. In particular, we focus on the decision-making of ML method selection, i.e., to determine when to stop the process and how much additional investment is needed to achieve a target goal (result).
The most important prerequisite for generating accurate ML models is high-quality training data, however securing such data is often an arduous task. Additionally, engineering and selecting appropriate features is especially time-consuming and requires a vast amount of effort and resources \cite{figalist2020end}. The benefits gained from the application of ML can be dramatically offset due to data collection and data pre-processing activities, which incur substantial costs and effort.
ROI is of great interest in engineering and business, where it is widely used as a guide for decision-making. This is true in Software Engineering (SE) as well. For example, Silverio et al. \cite{martinez2013rearm} evaluated cost-benefit analysis for the adoption of software reference architectures for optimizing architectural decision-making. Cleland et al. \cite{cleland2004heterogeneous} studied the ROI of heterogeneous solutions for the improvement of requirements traceability. However, the recent data explosion in the form of big data and advances in Machine Learning (ML) have posed questions on the efficiency and effectiveness of these processes that have become more relevant.
In this paper, we present two empirical studies from the field of requirements engineering. While it serves as one sample topic for a broader problem, Requirements Dependency Classification (RDC) has been a topic of interest for both researchers and practitioners. In particular, we study a fine-tuned \textit{BERT} (Bidirectional Encoder Representations from Transformers) \cite{devlin2018bert}, a recent technique published by researchers from Google, with Random Forest for solving RDC. BERT uses bidirectional training of transformer, a popular attention model, to language modelling, which claims to be state-of-the-art for NLP tasks. We compare BERT with Random Forest (RF), a widely used ML technique that serves as a baseline for comparison.
The objective of this study is to present an alternative method to evaluate ML algorithms. In that sense, we demonstrate the perspective of the returns ML algorithms would generate for the investment done while choosing a particular method for a given problem.
Our research contributions are as follows
\begin{itemize}
\item Describe an ML process model for ML classification and perform related ROI modeling.
\item Empirically evaluate Random Forest and fine-tuned BERT for textual classification in the context of requirement dependency classification (RDC) using accuracy and ROI.
\end{itemize}
The remainder of the paper is structured as follows: Section \ref{motivating} provides a motivating example of this study, followed by the description of related work in Section \ref{relatedwork}. Section \ref{RDE} explains requirement dependency, its extraction, practical relevance, and research questions. Section \ref{ROIofML} elaborates our ROI modeling of the ML process. Data used in this study are detailed in Section \ref{data} followed by empirical results in Section \ref{results}. The discussion Section \ref{discussion} details implications and limitations of this study before summarizing conclusions in Section \ref{conclusions}.
\section{Motivating Example}
\label{motivating}
\begin{figure}[t]
\centering
\includegraphics[scale=0.3]{Images/ROI_curve.png}
\caption{Break-even point from cost-benefit analysis of technology investment.}
\label{fig:2}
\end{figure}
Figure \ref{fig:2} shows a prototypical ROI curve for technology investment \cite{ruhe2020optimization}. When trying to achieve better results, the investment's cost (or effort) is growing over time, typically non-linearly. However, the benefit achieved from that investment eventually reaches some saturation point beyond which almost no further improvement is achieved. In total, a saturation point is achieved, after which further investment does not pay off anymore (i.e., point of diminishing returns).
Thus, the most crucial question arises-\textit{Do similar arguments apply for ML classification in Software Engineering?} While this could be true in general, we study it in the context of the requirements dependency classification problem.
Deshpande et al. \cite{deshpande_ruhe} report the results of a recent survey for requirements dependency classification and maintenance, with 76\% of responses (out of 70) from practitioners. More than 80\% of the participants agreed or strongly agreed that dependency type classification is difficult in practice; dependency information has implications for maintenance, and ignoring dependencies has a significant impact on project success \cite{deshpande_ruhe}.
Applying the advanced NLP technique BERT, we performed an ROI analysis on the requirements dependency classification. Automating this process saves time, and making the classification more effective helps better align the development process with the existing dependencies. For example, if a requirement r depends on another requirement s, then the implementation of s should precede implementing r. Violating this logical dependency will not only delay the usage of r but also decrease the effectiveness of testing.
\begin{figure}[H]
\centering
\includegraphics[scale=.43]{Images/21_BERT_ROI.png}
\caption{ROI vs F1 of BERT for Firefox dataset \cite{Deshpande_Ruhe_ROI}}
\label{fig:ESEM}
\end{figure}
Figure \ref{fig:ESEM} shows that there is an early peak in the ROI of using BERT. Since it is a very data-intensive technique, the ROI goes down with increasing training set size before the ROI reaches the global maximum. By comparison, considering only the harmonic mean (F1) of precision and recall gives a different recommendation for training set size. We discuss this in detail in Section 5.
\section{Related Work}
\label{relatedwork}
Although ROI is used in various contexts in Software Engineering and Data analytics, we discuss noted findings from the literature in the context of our proposed research.
\subsection{Exploration of ROI in Software Engineering}
Farbey et al. \cite{farbey_evaluation_nodate} explained that as a product moves through its life cycle, various evaluation methods such as ROI, Multi-Objective multi-criteria, Value analysis etc. play an important role in decision making. In this study, ROI was recommended either as a strategy to decrease uncertainty in the business area or to improve knowledge of how technology would operate.
Khoshgoftaar et al. \cite{khoshgoftaar_cost-benefit_2001} presented an interesting case study of a large telecommunication software system and demonstrated a methodology for
cost-benefit analysis of a software quality classification model. The cost and benefit computations were based on the type-I (FP) and type-II (FN) values of classification models. Although these cost-benefit models were ahead of their time, they did not consider the time and effort investment done on data and metrics gathering for cost computation. In another study on calculating ROI in the software product line, Bockle et al. \cite{bockle_calculating_2004} derived cost and benefit estimates based on organization level criteria, such as cost to the organization and cost of reuse. However, this did not involve data analytics of any form.
The guesswork could be eliminated from the decision-making process while evaluating the profitability of expenditure, which could help measure success over time. For instance, Erdogmus et al. \cite{erdogmus2004return} analyzed the ROI of quality investment to bring its value into perspective; posed an important question, "We generally want to increase a software product's quality because fixing existing software takes valuable time away from developing new software. But how much investment in software quality is desirable? When should we invest, and where?", which we think is difficult to quantify yet crucial for the success of software-based products.
Begel \& Zimmermann \cite{begel2014analyze} gathered and listed a set of 145 questions in a survey of 200 Microsoft developers and testers and termed them relevant for DA at Microsoft. One of the questions: "How important is it to have a software DA team answer this question?", expected answer on a five-point scale (\textit{Essential} to \textit{I don't understand}).
Although this analysis provides a sneak peek of the development and testing environments of Microsoft, it does not provide emphasis on any form of ROI. Essentially, we speculate that the ROI aspect was softened into asking for the perceived subjective importance through this question.
Boehm et al. \cite{boehm2008roi} \cite{boehm_roi_2004} presented quantitative results on the ROI of Systems Engineering based on the analysis of the 161 software projects in the COCOMO II database. Ruhe and Nayebi \cite{ruhe2016counts} proposed the \textit{Analytics Design Sheet} as a means to sketch the skeleton of the main components of the DA process. The four-quadrant template provides direction to brainstorm candidate DA methods and techniques in response to the problem statement and the available data. In its nature, the sheet is qualitative, while ROI analysis goes further and adds a quantitative perspective for outlining DA.
Ling et al. \cite{ling2005predicting} proposed a system to predict the escalation risk of current defect reports for maximum return on investment (ROI), based on mining historic defect report data from an online repository. ROI was computed by estimating the cost of not correcting an escalated defect (false negative) to be seven times the cost of correcting a non-escalated defect (false positive).
\begin{figure*}[!ht]
\centering
\includegraphics[scale=.5]{Images/RDE.PNG}
\caption{Requirements dependencies across various releases of project}
\label{fig:RDE}
\end{figure*}
\subsection{Exploration of ROI in Data Analytics}
Ferrari et al. \cite{ferrari2005roi} studied the ROI for text mining and showed that it not only has a tangible impact in terms of ROI but also intangible benefits, which arise from the investment in the knowledge management solution. This solution translates the returns directly that must be considered while integrating the financial perspective of analysis with the non-financial ones.
Weiss et al. \cite{weiss2008maximizing} emphasized how the quality of external data influence the results and quantified the effort of gathering and using such data when it is available at a premium into cost in terms of CPU time, even though the treatment of the subject is limited to a static setting. In a similar vein, Nagrecha et al. \cite{nagrecha2016quantifying} proposed a Net Present Value model to determine the cost and impact of analytics programs for an organization.
Taking inspiration from these studies in our research, we not only consider data pre-processing costs as an additional cost aspect but also transform machine learning metrics to dollar amounts, with derived costs and benefits being also validated by industry experts.
\subsection{Empirical Analysis for Requirements Dependency Classification}
Requirements dependencies classification is an active field of SE research.
The practical importance of the topic was confirmed by a survey \cite{deshpande_ruhe} of over 90 participants from the SE industry. Results showed that more than 80\% of the participants agreed or strongly agreed that (i) dependency type extraction is difficult in practice, (ii) dependency information has implications on maintenance, and (iii) ignoring dependencies has a significant negative impact on project success.
Several empirical studies have explored diverse computational methods that used natural language processing (NLP) \cite{och2002feasibility} \cite{atas2018automated}, semi-supervised technique \cite{deshpande_arora_ruhe_2019}, hybrid techniques \cite{deshpande_ruhe_2020} and deep learning \cite{guo2017semantically} in this context. Recently, Wang et al. \cite{9222252} explored a semi-automatic ML approach based on traceability to identify requirement dependencies to further identify security vulnerabilities. However, none of the approaches considered ROI to decide among techniques and the depth and breadth of their execution level.
\subsection{Exploration of Machine Learning process in Software Engineering}
We analyzed 96 papers from IEEE, Scopus, ScienceDirect, and ACM Digital Library which exclusively used ML, and data analytics within software engineering, and software development domains. Precision, Recall, Accuracy, and AUC were by far the most common performance measures used by researchers in these papers. Additionally, the choice of performance measure was generally not justified. Most studies did not present all steps of the ML process, and most of the papers formally present only 3 steps of the ML process such as data pre-processing, evaluation, and parameter tuning and all these steps are underestimated in terms of effort spent.
This study highlights the merits of simultaneously considering technical and business criteria when evaluating tradeoffs faced within machine learning approaches for requirements dependency classification (RDC). We extend prior work that focused on comparing various ML techniques based upon technical criteria of accuracy to include broader consideration of the impact i.e. Evaluating value generated by the analysis compared to the costs incurred for the analysis.
\section{Requirements Dependency Classification}
\label{RDE}
Similar to requirements elicitation \cite{lim2021data}, extraction of requirements dependencies is a cognitively difficult problem. These dependencies not only influence the development of software but also impact how requirements operate.
In this section, we provide the formal problem definition which serves as an example to demonstrate the value of looking beyond accuracy measure and investing in more general concepts of ROI analysis.
\subsection{Practical Relevance}
\par Figure \ref{fig:RDE} is an illustration of the practical relevance of considering requirements dependencies for incremental and iterative software development. Having multiple release cycles: $R_{i-1}, R_i, R_{i+1}$ defines the order of implementing and testing new or updated features. However, if a requirement is implemented in a release $R_i$ but requires a requirement implemented in a later release, then the requirement will not be usable. Similar arguments hold for two requirements that are related to each other but are implemented in different releases. Thus, identifying the dependencies early on is crucial as it drives the implementation as well as testing and rework efforts immensely.
\subsection{Problem Formulation}
While there are different types of dependencies between requirements \cite{Zhang14}, \cite{Carlshamre} we provide the definitions just for the ones used in the empirical study . For a set of requirements $R$ and a pair of requirements $(r,s)$ $\epsilon$ $R\times R$
\begin{itemize}
\item [1)] Two requirements r, s are called \textbf{\textit{INDEPENDENT}} if handling one of them has no logical or practical implication for handling the other one. Otherwise, they are called \textbf{\textit{DEPENDENT}}.
\item [2)] \textbf{\textit{REQUIRES}} is a form of \textit{DEPENDENT} relationship. If requirement $r$ requires the requirement $s$ to be implemented, then, $r$ and $s$ are in a \textit{REQUIRES} relationship. REQUIRES is an asymmetric relationship.
\item [3)] \textbf{\textit{RELATES\_TO}} is another specific form of \textit{DEPENDENT} relationship. Requirement $r$ relates\_to requirement $s$ if changing one of them has an impact on the other. \textit{RELATES\_TO} is a symmetric relationship\footnote{There are other types of dependencies such as \textit{DUPLICATES}, \textit{BLOCKS} etc. that also occur in the these datasets, however, we have considered the ones that occur most frequently}.
\end{itemize}
\noindent \textbf{Problem: Binary Requirements Dependency Classification (RDC)}
For a given set $R$ of requirements and their textual description, the binary Requirements Dependency Classification problem (RDC) is to decide for a given pair (r,s) $\epsilon$
$R\times R$ if (r,s) is in a \textit{REQUIRES} (called problem RDC\_1) or in a \textit{RELATES\_TO} (called problem RDC\_2) relationship.
\subsection{Research Questions}
In this paper, two research questions (RQs) are addressed:
\begin{enumerate}
\item []\textbf{RQ1:} How to model the ROI for ML classification? Specifically, how to instantiate the model for the problem of RDC?\\
\textbf{Rational:} The exclusive consideration of accuracy in the selection of ML classification techniques might be misleading. We consider ROI as an alternative and additional criterion. To study the cost and benefit of the ML classification in a specific context, it is essential to consider the complete process of ML classification and the impact of the results in the original problem space.
\newline
\item []\textbf{RQ2:} For RDC, how is the preference decision between RDC-BERT and RF impacted by the accuracy criteria F1 that includes ROI?\\
\textbf{Rational:} We evaluate the impact of the selection criteria through two empirical studies on two open-source software (OSS) datasets: Firefox, a software application from Mozilla family \cite{mozillawiki} and Typo3 \cite{typo3}, a content management software. Our goal is to evaluate two extraction techniques (RDC-BERT and RF) to demonstrate the impact of the consideration of ROI in addition to accuracy considerations.
\end{enumerate}
\section{ROI Modeling of ML Classification - RQ1}
\label{ROIofML}
Machine Learning classification is an iterative process comprising a series of steps. Aiming at ROI analysis of ML classification requires a look at the effort consumed for all these steps. In what follows, we describe various ML process steps, we estimate cost and benefit, and project the ROI of ML classification.
Although various ML workflow has been defined in the literature \cite{fayyad1996kdd} \cite{amershi2019software} \cite{mlworkflow}, in this section, we present the simplified version of it mainly focusing on the ML process.
\begin{figure*}[!ht]
\centering
\includegraphics[scale=.5]{Images/MLProcess.png}
\caption{Overview of the steps constituting the ML process}
\label{fig:arch}
\end{figure*}
\subsection{Modeling the Process}
The process steps are organized into four Phases: A, B, C, and D called Planning, Data Preparation, Execution, and Validation, respectively. Depending on the context, the effort allocated for these steps may vary. However, this approach parallels the process steps and guidelines for pragmatic optimization in software engineering by Ruhe et al. \cite{ruhe2020optimization}.
An overview of the steps is illustrated in Figure \ref{fig:arch}. Here, we did
not show all the possible arrows to indicate that loops
can, and do, occur between any two steps in the
process. The iterative and interactive ML process, involving various phases is summarized as:
\newline \newline
\noindent
\textbf{Phase A : Planning}
\begin{itemize}
\item[] \textbf{Step 1: Scoping and problem formulation}\\
Scoping defines the problem context and its boundaries. Problem formulation addresses the key independent and dependent attributes to be considered. As a result of later steps, the problem formulation eventually needs to be adjusted as asking the right question constitutes the largest effort for any application effort.
\newline
\item [] \textbf{Step 2: Evaluation of candidate machine learners}\\
A variety of ML algorithms exist and new ones are discovered regularly. Commonly used machine learning algorithms include Linear Regression, Logistic Regression, Decision Trees, K-means, Support Vector Machines, Naïve Bayes, Random Forest, and Neural Networks. There is no obvious preference in the sense that "One size fits all". However, there could be recommendations for a particular ML algorithm for a given problem based on its exemplary performance for a similar problem(s). An initial evaluation helps to select the most promising one(s). The selection is influenced by the success criterion of the classification (e.g., accuracy).\\
\end{itemize}
\noindent
\textbf{Phase B: Data Preparation}
\begin{enumerate}
\item []\textbf{Step 3: Data collection}\\
Different sources of data might exist for performing ML classification. Data collection looks into what is potentially relevant and checks the type and availability of the data.
\newline
\item []\textbf{Step 4: Data pre-processing}\\
Raw data would not be ready for processing through the ML algorithm as it could have duplicates, missing values, and contradictions that need to be tackled first for error-free results. Performing such pre-processing operations, for example, data
cleaning, normalization, transformation, feature extraction and
selection, etc. are essential for the success of ML classification, but these steps consume a considerable amount of human resources and processing time. The outcome of data pre-processing is the training
set which could be processed through ML algorithm further \cite{kotsiantis2006data}.
\newline
\item []\textbf{Step 5: Labeling}\\
Labeling is to assign labels to ground truth data \cite{amershi2019software}. Supervised ML methods need labeled data unlike unsupervised ML methods. Labeling is generally performed by domain experts who identify a set of samples (that are most likely representative of the real-world data) to train the ML models. Depending on the nature of the problem, online crowdsourced platforms could also be used for labeling tasks \cite{nayebi2017MAPFEAT}.
\end{enumerate}
\noindent \textbf{Phase C: Execution}
\begin{enumerate}
\item [] \textbf{Step 6: Training}\\
The key idea of ML is to learn from existing data and then apply the resulting model to new data. The quality and quantity of the training data are often as important as the actual machine learning algorithm. To learn from existing data also means that the data set is complete, with known input and output of the observations.
\newline
\item [] \textbf{Step 7: Hyper-parameter tuning}\\
ML algorithms depend upon several parameters such as named model parameters and named hyper-parameters. Named model parameters can be initialized and updated through the data learning process (e.g., the weights of neurons in neural networks). Named hyper-parameters cannot be directly estimated from data learning and should be set before training an ML model because they define the model architecture. Tuning these parameters means achieving settings that enable good algorithmic performance \cite{kuhn2013applied}.
\newline
\item [] \textbf{Step 8: Testing}\\
After training, the model is applied to the selected test set(s) (a small part of labeled data that is held out and excluded from the training process). The larger the number of variables in the real world, the bigger the training and test data should be. From performing testing, classification error counts are captured in the form of a confusion matrix.
\end{enumerate}
\noindent
\textbf{Phase D: Validation}
\begin{enumerate}
\item [] \textbf{Step 9: External validation}\\
Success from Step 8 does not automatically imply the success of the results in the context of the application. The validity of the problem formulation and the data might prevent the applicability of the results (i.e., not actionable within the organization resulting in significant wasted effort).
Internal validation approaches such as cross-validation can not guarantee the quality of a machine learning model due to potentially biased training data. External validation is critical for evaluating the generalization ability of the machine learning model, where independently derived datasets (external) are leveraged as validation datasets. While such independent validation is also sometimes used to refer to a validation study by other researchers that the researchers who developed the model \cite{ho2020extensions}.
\end{enumerate}
\subsection{Modeling Cost and Benefit}
Acknowledging that ML classification is a process of steps with possibly multiple iterations suggests the need to look at the estimated cost for all these steps. Cost estimation is known to be inherently difficult in software engineering \cite{shepperd2014cost}. The same is true for value prediction. Despite many factors influencing the costs and benefits of ML classifications, we provide a preliminary model to allow a rough estimate of the ROI.
For cost estimation, we make the assumption that the total cost of performing ML classification with any given ML technique is the sum of cost components of the four phases outlined in the previous section. To simplify the model, we focus on Phase B (Data Preparation) and Phase C (Execution) and ignore the other two phases. Finally, we assume an 80:20 effort (and cost) ratio between Phase B and Phase C, emphasizing the fact that the majority of effort is spent on data preparation.
For modeling the benefit of the classification results, we are looking at classification errors and their cost (penalty) created. A \textit{confusion matrix} CM is a matrix that contains information relating actual with predicted classifications. For \textit{n} classes, CM will be an $n \times n$ matrix
associated with a classifier. Table \ref{CMSample} shows the principal entries of CM for binary classification.
\begin{table}[!ht]
\centering
\renewcommand{\arraystretch}{1.5}
\caption{A confusion matrix of binary (two) class classification problem}
\begin{tabular}{p{1.5cm} |p{2cm}|p{2cm}|}
\cline{2-3}
& \textbf{Predicted Negative} & \textbf{Predicted Positive} \\ \cline{2-3}
\multicolumn{1}{l|}{\textbf{Actual Negative}} & True Negative (TN) & False Positive (FP) \\ \hline
\multicolumn{1}{l|}{\textbf{Actual Positive}} & False Negative (FN) & True Positive (TP) \\ \cline{2-3}
\end{tabular}
\label{CMSample}
\end{table}
\begin{table*}[!ht]
\centering
\caption{Parameters used for ROI computation}
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{l|l|l p{7cm} l}
\multicolumn{2}{l}{} & \textbf{Symbol} & \textbf{Meaning} & \textbf{Unit} \\ \hline
\multirow{7}{*}{\textbf{\textit{Cost factors}}{$^1$}} &
\multirow{1}{*}{\textbf{Phase A}} & $C_{pl}$$^2$ & Planning phase cost & \$ \\ \cline{2-5} &
\multirow{2}{*}{\textbf{Phase B}} & $C_{dg}$ & Data gathering cost & \$ \\
& & $C_{pp}$ & Pre-processing cost & \$ \\ & \multirow{3}{*}{\textbf{Phase C}} & $C_l$ & Labeling cost & \$ \\ \cline{2-5}
& & $C_t$$^2$ & Hyper-parameter tuning cost & \$ \\ &
& $C_{train/test}$ & Training and testing cost & \$ \\ \cline{2-5}
&
\textbf{Phase D} & $C_e$$^2$ & External Validation cost & \$ \\ \hline \hline
\multicolumn{2}{l|}{\multirow{2}{*}{\textbf{Classification Penalty}}} & $Cost_{FP}$ & Penalty per FP & \$ \\
\multicolumn{2}{l|}{} & $Cost_{FN}$ & Penalty per FN & \$ \\ \hline \hline
\multicolumn{2}{l|}{\multirow{6}{*}{\textbf{Others}}} & $N_{HR}$ & \#Human resources & Number \\
\multicolumn{2}{l|}{} & $C_{HR}$ & Human Resource cost & \$/hr \\
\multicolumn{2}{l|}{} & $N_{train}$ & Size of the training set & Number \\
\multicolumn{2}{l|}{} & $N_{test}$ & Size of the test set & Number \\
\multicolumn{2}{l|}{} & $N$ & $N_{train}$ + $N_{test}$ & Number \\
\multicolumn{2}{l|}{} & $Value_{prod}$$^3$ & Estimated value of the product for a release cycle & \$ \\ \hline
\end{tabular}
\label{Params}
\begin{flushleft}
\textbf{$^1$}\footnotesize{These are per sample cost factors. All the costs are computed by translating them from minutes to \$ by multiplying with resources and cost per hour of the resources} \\
\textbf{$^2$}\footnotesize{For simplicity few of the cost factors have been assumed to be zero}
\\
\textbf{$^3$}\footnotesize{This value was computed using various cost estimates for a period of one release cycle (= 18 months)}
\end{flushleft}
\end{table*}
The F1 score is a measure of the model's accuracy based on the training set and defined as the harmonic mean of the model's precision and recall in (\ref{eq:1}).
\begin{equation} \label{eq:1}
\centering
F1 = \frac{2 \times TP}{2 \times TP + FP + FN}
\end{equation}
In the context of dependency classification, the benefit could be modeled in terms of the ability of the ML model to produce the least amount of overhead by 1) Incorrectly classifying independency as a dependency (False Positive) 2) Incorrectly classifying dependency as independent (False Negative). So, using $Cost_{FP}$ and $Cost_{FN}$ as estimated re-work costs due to classification overhead, $Sum(Cost_{FN}, Cost_{FP})$ would be the cumulative expense that a company has to bear.
\par In a release cycle, if
\textit{estimated value} that a product could generate is :$Value_{prod}$ then the $Benefit$ would be the difference of the \textit{estimated value} and the \textit{classification overhead}. Table \ref{Params} lists the relevant cost components and their corresponding units.
\subsection{Modeling ROI}
During every classification, $Cost$ and $Benefit$ were computed using the parameters explained in Table \ref{Params}. \textit{Cost factors} are data processing costs (Phase B and Phase C) for all the train ($N_{train}$) and test ($N_{test}$) samples ($n$) in every iteration. This is further translated into dollar-cost by multiplying with hourly charges ($C_{HR}$) of $N_{HR}$ human resources.
\begin{equation} \label{eq:2}
Cost = n\times\sum_{all\hspace{1mm} applicable}{Cost\hspace{1mm} factors}\times N_{HR} \times C_{HR}
\end{equation}
$Return$ computations for RDA, assumes reward ($Cost_{FP}$) for misidentifying the independent requirements (FP) and heavily penalizing ($Cost_{FN}$) instances that were falsely identified as independent (FN).
\begin{equation} \label{eq:4}
\centering
Total Penalty = FP \times Cost_{FP} + FN \times Cost_{FN}
\end{equation}
\begin{equation} \label{eq:5}
\centering
Benefit = Value_{prod} - Total Penalty
\end{equation}
Return and investment are context-specific terms, and studying the ROI of Machine Learning classification needs tailoring to the context of the study. To determine the ROI, we follow the simplest form of its calculation relating to the difference between $Benefit$ and $Cost$ to the amount of $Cost$ as shown in (\ref{eq:6}). Both $Benefit$ and $Cost$ are measured as human effort in person-hours.
\begin{equation} \label{eq:6}
\centering
ROI = (Benefit - Cost)/Cost
\end{equation}
The core investigative focus of our study is to evaluate various conditions under which RDC-BERT (fine-tuned BERT using data specific to requirement dependency extraction) is preferable to the baseline ML method: Random Forrest (RF).
In this empirical analysis, beginning with a small train set, classifiers were created, and then the train set was incremented slowly by a fixed factor to generate new classifiers in every iteration until all the data available for training was exhausted. In every iteration, the classifiers were tested for a small fixed data set to capture the results.
\section{Data and Experiment Setup}
\label{data}
Online bug tracking systems such as Bugzilla \cite{bugzilla} and Redmine \cite{redmine} are widely used in open-source software development. Feature requests, tasks, bugs, epics, stories, features, enhancements, and new requirements are logged into these systems in the form issue reports \cite{shi2017understanding} \cite{bhowmik2015resolution} which help software developers to track them for effective implementation \cite{shin2015guidelines}, testing and release planning \cite{ruhe2005art}.
We mined data from Bugzilla and Redmine related to features for the two OSS projects namely, Firefox - a Mozilla web browser application and Typo3 - a content management system.
\begin{table*}[!ht]
\centering
\renewcommand{\arraystretch}{1.3}
\caption{Dependency pair samples from the two datasets}
\begin{tabular}{p{2.2cm}|p{1cm} p{5cm}|p{1cm} p{5cm}}
\hline
\textbf{Dependency type} & \textbf{ID} & \textbf{Description} & \textbf{ID} & \textbf{Description} \\ \hline
\multirow{8}{*}{\textit{REQUIRES}} & 1432952 & add ability to associate saved billing address with payment card in add/edit card form & 1429180 & option to use new billing address when adding new payment card \\
& 1394451 & update illustration for error connection failure & 1358293 & ux error connection failure copy design and illustration update \\
& 1524948 & introduce session group to allow to manage multiple session at same time & 1298912 & multiple snapshot perform periodic session backup and let user restore particular backup \\ \hline \hline
\multirow{8}{*}{\textit{RELATES\_TO}} & 92822 & ignore button for link targets & 92297 & make it possible to mark specific links to not get checked by linkvalidator \\
& 92576 & page tree filter: make it possible to explicitly filter by uid & 36075 & advanced filtering for the page-tree \\
& 91496 &differentiate between password reset "by user" and "by admin" & 89513 &provide password recovery for backend users \\ \hline
\end{tabular}
\label{table:samples}
\end{table*}
\subsection{Firefox}
\par In Bugzilla, feature requests are specific types of issues that are typically tagged as “enhancement” \cite{mozillawiki}. We retrieved these feature requests for the Firefox project using the search engine in the Bugzilla issue tracking system and exported all the related fields such as Title, Type, Priority, Product, Depends\_on, and Blocks. Each issue report contains dependency relationships with other issue reports as references metadata \cite{kim2019weighted}.
Using this information, 3,773 depends\_on (also interpreted as \textit{REQUIRES} dependency type) requirements pairs were retrieved. To generate negative samples, requirements that had no relationship were paired and 21,358 samples were generated.
\subsection{Typo3}
Redmine \cite{redmine} is a free and open-source web-based management and issue tracking tool website. It allows users to manage multiple projects and associated projects. Various issues across a range of projects are updated each day which helps software developers to track them for effective implementation. In Redmine, features are a specific type of issue that is extracted in this paper for further data analysis.
Typo3 Content Management System (CMS) is an Open Source Enterprise Content Management System\cite{typo3} with a large global community of approximately 900 members of the TYPO3 Association. We collected information such as issue\_links, description, the version found, the version released, issue\_id etc. for 5,017 features using Redmine's REST API through a Python script for this study.
\par All feature descriptions that had fewer than three words in them were filtered out, resulting in 1,324 feature pairs with dependency type \textit{RELATES\_TO}.
Using the rest of the features that were not in any type of dependency with others, 9,270 pairs were generated as a negative sample set.
Table \ref{table:samples} mentions sample pairs of requirements dependencies. For example, to be able to associate the address with payment card \textit{REQUIRES} ability to use a new billing address when adding a new payment card. For both data sets, to perform binary classification, both positive and negative samples are needed for training. Since we only had dependent (positive) samples in the data, we generated negative samples by pairing the requirements which were not related in the given snapshot of the dataset.
\subsection{Effort and Value Estimation}
Typo3, currently at released version 11, is a complex content management system that is developed as a hybrid OSS software product. It has a core team of 12 members with varying skills and expertise. They have a major release cycle of 18 months and they plan two or more releases ahead of time. Developers are encouraged to track the dependencies in Jira, however, a few of the team members utilize post-its to work and track them. Typo3 does not explicitly consider Requirements Engineering as a development phase, but they term the efforts towards identifying features and extracting dependencies as conceptual work or scoping. Over 15\% of the release, the cycle is identified as scoping effort and about 25\% of scoping in a release cycle is identified as dependency extraction and identification. Nine team members and the CTO are involved, mostly in identifying the dependencies.
The CEO confirmed that about 80 \% of the features are in some form of dependency with each other and missing the dependencies is more problematic than misidentifying them. As he puts this in words, \enquote{if you miss dependencies then it starts to ramp up quickly and this is when things go wrong, and breaks deadlines. we wanted to release in April (4 weeks ago) now deadline is mid October}.
Typo3 identifies and manages seven different types of dependencies and their inversions such as precedes, blocks, clones, caused\_by etc. Most of the dependency issues are identified rigorously through testing and the estimated re-work is about 12\%. They have minimal manual testing as they have test suits of over 75,000 test cases. The CEO estimated that the overwork caused by missing dependencies is about 10\% of the efforts. The average salary of the nine people involved in re-work is \$70 (CAD). A summary of all estimates is provided in Table \ref{Assumptions}.
\begin{figure*}[!htpb]
\centering
\includegraphics[scale=.5]{Images/Expsetup.PNG}
\caption{Overview of the experiment setup}
\label{fig:expSetup}
\end{figure*}
\subsection{Experiment Setup}
Figure \ref{fig:expSetup} depicts the overview of our experiment setup. The complete approach is multi layered as highlighted in the shades. Each one of these could be further expanded to include additional elements for solution space evaluation further.
In this study, to generate the results, RF, Naive Bayes and SVM ML algorithms were compared against RDC-BERT for the two datasets: Firefox and Typo3. Overall eight experiments were conducted. Since RF performed better among all the conventional ML algorithms \cite{Deshpande_Ruhe_ROI}, we report the results of RF and RDC-BERT (i.e. totally four experiments).
For each experiment, we computed ROI using False Negative and False Positive values (from Confusion matrix). In Section \ref{results} we present the insights to aid decision-making in algorithm selection based on these eight outcomes. For additional clarity, we list the names of the analysis of the results and their description in Table \ref{table:eightTest}.
\begin{table*}[h]
\centering
\renewcommand{\arraystretch}{1.25}
\caption{Overview of the various analyses done in Section \ref{results} }
\begin{tabular}{p{1cm} p{4.5cm} p{6.5cm} p{1cm}}
\textbf{} &\textbf{} & \textbf{Description} & \\ \hline
Fig \ref{fig:F1_Firefox} &F1\_Firefox & Firefox: Compare F1 of RF and RDC-BERT & RQ 2.1 \\
Fig \ref{fig:F1_typo3} & F1\_Typo3 & Typo3: Compare F1 of RF and RDC-BERT & RQ 2.1 \\
Fig \ref{fig:ROI_Firefox} & ROI\_Firefox & Firefox: Compare ROI of RF and RDC-BERT & RQ 2.1 \\
Fig \ref{fig:ROI_typo3} & ROI\_Typo3 & Typo3: Compare ROI of RF and RDC-BERT & RQ 2.1 \\ \hline \hline
Fig \ref{fig:BERT_Firefox} & F1\_ROI\_RDC-BERT\_Firefox & Firefox: F1 vs ROI of RDC-BERT & RQ 2.2 \\
Fig \ref{fig:RF_Firefox} & F1\_ROI\_RF\_Firefox & Firefox: F1 vs ROI of RF & RQ 2.2 \\
Fig \ref{fig:BERT_Typo3} & F1\_ROI\_RDC-BERT\_Typo3 & Typo3: F1 vs ROI of RDC-BERT & RQ 2.2 \\
Fig \ref{fig:RF_Typo3} & F1\_ROI\_RF\_Typo3 & Typo3: F1 vs ROI of RF & RQ 2.2 \\ \hline
\end{tabular}
\label{table:eightTest}
\end{table*}
Requirements pairs were pre-processed to eliminate noise such as spatial characters and numbers. The generated output is fed to RDC-BERT and RF for training. Care was taken to process the same data snapshot through RF and RDC-BERT models. Further, the fine-tuned BERT model (RDC-BERT) is then used for classification. The data was split (80:20) into train and test sets, and balanced between both classes.
In this empirical analysis, we conducted classification by utilizing a fraction of the whole dataset for training and testing for a small fixed data set. This was repeated by slowly increasing the training set and results were captured.
\newline \newline
\textbf{Random Forest: } For RF, we use TF-IDF to generate word vectors before training. Also, hyper-parameter tuning was performed and the results for 10-fold cross-validation were computed, followed by testing.
\newline \newline
\textbf{RDC-BERT:} For fine-tuning BERT, a pre-trained BERT model is used in combination with our RDC specific dataset. The result is a fine-tuning BERT model called RDC-BERT. To fine-tune the BERT model, we used \textit{NextSentencePrediction}\footnote{\url{https://huggingface.co/transformers/model_doc/bert.html#bertfornextsentenceprediction}}, a sentence pair classification pre-trained BERT model, and further fine-tuned it for the RDA specific dataset on Tesla K80 GPU on Google Colab\footnote{\url{https://colab.research.google.com/}}.
In every instance, for a given training set size, RDC-BERT was trained through three epochs with a batch size of 32, and a learning rate of 2e-5. In each epoch, the train set was divided into 90\% for training and 10\%for validation. Finally, RDC-BERT was used to classify the test set and the resulting F1-score and confusion matrix were captured.
BERT eliminates the need for feature extraction since it is a language model based on deep learning. BERT, pre-trained on a large text corpus, can be fine-tuned on specific tasks by providing only a small amount
of domain-specific data.
\section{Empirical Analysis - RQ2}
\label{results}
In this section, we report the results of our empirical analysis and answer RQ2. We structure results by the type of decisions to be made: (i) \textbf{RQ 2.1: }When comparing two techniques: Which one is preferable under conditions selected?, and (ii) \textbf{RQ 2.2: }When looking at one technique, when to stop the analysis? For both decisions, we present the results of the analysis for the two data sets introduced above and the two techniques under investigation using estimates from Table \ref{Assumptions}.
\begin{table}[!h]
\centering
\caption{Parameter settings for the two empirical analysis scenarios}
\renewcommand{\arraystretch}{1.25}
\begin{threeparttable}[t]
\centering
\begin{tabular}{p{4cm} p{3cm}}
\hline
\textbf{Parameters} & \textbf{Values}\\ \hline
Phase B: ($C_{dg} + C_{pp}+C_l$)$^1$ & 1.5 min/sample \\
Phase C: $C_{train/test}$ & 0.30 min/sample \\
$C_{HR}$ & \$70/hr \\
$N_{HR}$ & 10 \\ \hline
\multirow{2}{*}{$N$} &Firefox:7,546 \\ & Typo3: 2,648 \\\hline
$Cost_{FN}$ & \$25,000 \\
$Cost_{FP}$ & \$10,000 \\
$Value_{prod}$ & \$4,000,000 \\
\hline
\end{tabular}
\end{threeparttable}
\label{Assumptions}
\begin{flushleft}
\textbf{$^1$}\footnotesize{
$C_{dg}, C_{pp}$ and $C_l$ are weighed equally (= 0.5min/sample) each. Also ratio of Phase B:Phase C = 80:20 has been considered}
\end{flushleft}
\end{table}
\subsection{RQ 2.1: Comparison between RDC-BERT and RF}
\begin{figure*}[!ht]
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.5]{Images/F1_Firefox.png}
\caption{F1 of RDC-BERT vs RF for Firefox dataset}
\label{fig:F1_Firefox}
\end{minipage}
\hspace{10mm}
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.5]{Images/F1_typo3.png}
\caption{F1 of RDC-BERT vs RF for Typo3 dataset}
\label{fig:F1_typo3}
\end{minipage}
\end{figure*}
The traditional approach for comparing techniques is to look at just accuracy for some fixed training set. Figures \ref{fig:F1_Firefox} (F1\_Firefox) and \ref{fig:F1_typo3} (F1\_Typo3) show the comparison of the F1-scores for varying training set sizes for the two datasets. Results show that RF achieves a higher accuracy more quickly for even small-sized train sets respectively. However, with a training set greater than 40\% of the dataset for Firefox and 30\% for Typo3, RDC-BERT achieves better results overall.
Comparison of ROI for the two datasets and two methods (RDC-BERT and RF) is shown in Figures \ref{fig:ROI_Firefox} (ROI\_Firefox) and \ref{fig:ROI_typo3} (ROI\_Typo3) respectively.
\begin{figure*}[!h]
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.45]{Images/ROI_Firefox.png}
\caption{ROI of RDC-BERT vs RF for Firefox dataset}
\label{fig:ROI_Firefox}
\end{minipage}
\hspace{10mm}
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.5]{Images/ROI_typo3.png}
\caption{ROI of RDC-BERT vs RF for Typo3 dataset}
\label{fig:ROI_typo3}
\end{minipage}
\end{figure*}
For Firefox, with a smaller-sized train set, RF once again performs better comparatively, even though the ROI is negative. Similar results are evident for Typo3. RF performs marginally better ROI-wise for the smaller training set. ROI of RDC-BERT picks up pace only beyond 40\% and 30\% train set for Firefox and Typo3, respectively.
\subsection{RQ 2.2: Bi-criterion analysis of RDC-BERT and RF}
In the second part of the analysis for RQ2, we look at one technique at a time from the perspective of both F1-score and ROI. This will support decision-making towards the question of when does increase accuracy no longer pays off?
As illustrated in figures \ref{fig:F1_Firefox} and \ref{fig:F1_typo3}, increased training set does not yield better F1-score beyond 65\%. The F1-score hits a plateau and even starts to degrade for both of the methods and datasets.
However, if we look at the trade-off between the F1 and ROI for both datasets, the results become interesting. Figures \ref{fig:BERT_Firefox}: F1\_ROI\_RDC-BERT\_Firefox show that for RDC-BERT, F1-score increases linearly, however, max ROI is achieved when the train set is 70\% of the dataset. Whereas, for RF, in Figure \ref{fig:RF_Firefox} : F1\_ROI\_RF\_Firefox shows that F1 and ROI for the train set lower than 40\% is better than that of RDC-BERT. Chasing for a higher F1 score does not payoff and one needs to take a closer look at the benefits vs investment in more training data, eventually.
For Typo3, in Figure \ref{fig:BERT_Typo3}: F1\_ROI\_RDC-BERT\_Typo3 shows that F1-score and ROI grow steeply for RDC-BERT with the increasing train set. However, similar to Firefox, ROI and F1 of RF are stable and better than RDC-BERT for the train set smaller than 30\%. These findings once again emphasize the need to relook at how F1 and ROI together could aid in deciding on the ML selection.
\begin{figure*}[!h]
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.47]{Images/BERT_Firefox.png}
\caption{F1 vs ROI of RDC-BERT for Firefox dataset, utilizing values from Table \ref{Assumptions}}
\label{fig:BERT_Firefox}
\end{minipage}
\hspace{10mm}
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.43]{Images/RF_Firefox.png}
\caption{F1 vs ROI of RF for Firefox dataset, utilizing values from Table \ref{Assumptions}}
\label{fig:RF_Firefox}
\end{minipage}
\end{figure*}
\begin{figure*}[!ht]
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.47]{Images/BERT_Typo3.png}
\caption{F1 vs ROI of RDC-BERT for Typo3 dataset, utilizing values from Table \ref{Assumptions}}
\label{fig:BERT_Typo3}
\end{minipage}
\hspace{10mm}
\begin{minipage}{.45\textwidth}
\centering
\includegraphics[scale=.5]{Images/RF_Typo3.png}
\caption{F1 vs ROI of RF for Typo3 dataset, utilizing values from Table \ref{Assumptions}}
\label{fig:RF_Typo3}
\end{minipage}
\end{figure*}
In both datasets studies, it is evident that RDC-BERT models require large amounts of data (at least 30\% or more) to stabilize and show value (steady positive ROI). When comparing RDC-BERT with RF using ROI criteria (Fig 8 and 9) across the two data sets, RF outperforms RDC-BERT for the lower train set (incurring lower negative returns). However, positive ROIs are observed only at the larger train set at which RDC-BERT is consistently better than RF. Based upon the Firefox findings (Fig 10 and 11), RDC-BERT approaches the 80\% benchmark accuracy with approximately 50\% of the training data while RF requires 70\% training data to attain the same level of accuracy. However, both techniques can achieve positive ROI with as little 50\% training data but RBC-BERT achieves maximum ROI (30) with an accuracy of 0.87 with approximately 70\% training data, while RF achieves maximum ROI (2.2) with an accuracy of 0.75 with approximately 70\% training data.
Based upon the Typo3 findings (Fig 12 and 13), RDC-BERT approaches the 80\% benchmark accuracy with approximately 55\% of the training data while RF requires 70\% training data to reach the same level of accuracy. However, both RBC-BERT and RF can achieve positive ROI with as little 15\% training data, but RBC-BERT achieves maximum ROI (73.5) with an accuracy of 0.86 with approximately 65\% training data, while RF achieves maximum ROI (48) with an accuracy of 0.80 with approximately 70\% training data. Thus, RBC-BERT can deliver much higher ROI and similar levels of accuracy than RF given approximately the same amount of training data.
Finally, the parameter settings that seeded the initial model (Table 5) were based upon industry estimates, which were possible were verified by senior management in the respective firms. However, some of the findings may be sensitive to these initial conditions. Thus, these would need to be set for the specific context upon which the data sets are based. This is also the basis upon which scenario analysis could be conducted to evaluate the worst case, best case and most likely initial conditions to evaluate the impacts on subsequent decisions.
\section{Discussion}
\label{discussion}
\subsection{Implications}
ML is not simply a cost of doing business, rather it is a foundational activity that can provide value for the money invested.
Our proposed approach aligns this notion with the strategic direction of the organization. While return on investment (ROI) is a common approach used for business planning and decision making, it is not applied as widely within software engineering or specifically within applied ML.
In our study, we demonstrate how to instantiate ROI in the context of RDC. Our approach provides a pragmatic link between the business and technical aspects of the organization by providing a common language that incorporates both the technical aspects inherent in the evaluation of accuracy, with the business considerations of costs and benefits. We argue that this is an extremely powerful approach that provides evidence that is compelling and consistent for both technical and business decision-making.
In addition, we think that the ROI approach could sensitize the ML team to the entire process of ML classification and how that process fits into organizational processes. The ROI approach is essential for evaluating the possible tradeoffs between accuracy and the benefits. Mainly because without consideration of the key dependencies within the process, benefits in one part of the process (e.g., improved accuracy) can easily be undermined by excessive costs in another part of the process that would not typically be considered if focused exclusively on accuracy. Alternatively, lower levels of accuracy in the ML process might be acceptable if other benefits are accruing at reasonable costs. Thus, valuable ML investments are potentially being avoided based upon not meeting accuracy expectations, when those ML solutions could be sufficient to realize high payoffs for the organization.
Our approach increases the transparency of the decision-making process by adding diversity to the evaluation criteria that foreground the various tradeoffs being made. The development of AI tools that businesses and consumers can trust is essential for their continued adoption, especially as there is increasing regulatory scrutiny of the biases that arise in the ML algorithms or inherent in the data used for training.
While ML algorithms are generally trusted for relatively mechanical well-defined problems, this trust plummets when the decisions are subjective, and likely to vary by contextual variables that are not well understood. This in turn increases the pressure to adjust ML algorithms for variations in specific markets further driving up development costs. Such pressure directs the focus on customizing products and services based upon ML algorithms for specific markets while increasing costs further and undermines the benefits for certain markets or customers \cite{hbr}.
The proposed approach considers technical and business aspects simultaneously and provides a more traceable set of interconnected processes. This approach includes business and technical considerations to enable management to evaluate the risks of some undesirable decisions and the tradeoffs needed to realize the likely benefits.
\subsection{Limitations}
We have explored RF and RDC-BERT in the context of the RDC problem and presented our results. Since there is no single method which could work for any given problem, comparison of multiple approaches and their results remains out of the scope of this study.
Another threat to validity is the related to the conclusions made. Although we have taken care to randomize the data by shuffling and used stratified split to take care of balanced data in both training and validation, multiple runs with varying first iteration data sample are needed to be more confident on the conclusions made. However, we argue that the key observations made are valid from the restricted empirical validation performed.
\section{Conclusions and Future Work}
\label{conclusions}
ML classification is widely used in many disciplines of Science and Engineering. In this study, we demonstrate that just looking at performance measures such as accuracy could be misleading when, for example, deciding between two ML techniques evaluated for solving the same problem. Conversely, ignoring the cost and benefit of such a classification could cause the risk of unprecedented emphasis on improving accuracy that might not generate any value for the additional efforts spent. Additionally, in this research, we also provide a high-level ML process for classification (supervised machine learning). However, with minuscule changes, this process can be adapted to unsupervised or semi-supervised ML methods easily.
We use Requirements Dependency Classification as a sandbox to build a proof of the concept based on the two ML techniques used to solve RDC. In the future, we will extend the results in various dimensions. The concepts of this paper will be applied and evaluated for problems from other domains. However, the challenge is to project the benefit of achieving better accuracy results and estimating the total effort of data analysis. Also, depending on the problem, we will investigate other ML techniques and additional data sets.
With a broader data and knowledge base, we aim at developing a customized recommendation system that would support practitioners in their decision-making in terms of "How much Data Analytics is Enough".
\begin{acknowledgements}
We would like to thank graduate students Saipreetham Chakka and Aris Aristorenas for their assistance in generating results and conducting literature review for this study. \end{acknowledgements}
\section*{Conflict of interest}
The authors declare that they have no conflict of interest.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{sec:intro}
\baselineskip13pt
In this paper, we consider the forward Cauchy problem
\begin{equation}\label{eq:CPlin}
\begin{cases}
Q(\partial_t,\partial_x) u =0, \quad t\geq 0, x\in\R^n,\\
\partial_t^j u(0,x)=u_j(x),\qquad j=0,\ldots,m-1,\\
\end{cases}
\end{equation}
where $Q(\partial_t,\partial_x)$ is a hyperbolic operator of order~$m$, obtained by the sum of two or three homogeneous hyperbolic operators of order~$m$ and~$m-1$ or, respectively, $m$, $m-1$, $m-2$. We choose lower order polynomials which induce a partial dissipation of the energy for~\eqref{eq:CPlin}, that is, the roots of the full symbol $Q(\lambda,i\xi)$ have negative real part for any~$\xi\neq0$, but some roots vanish as $\xi\to0$. Under this assumption, we describe the asymptotic behavior of those roots as~$\xi\to0$ and as~$\xii\to\infty$. In particular, we study the vanishing speed of the real parts of those roots as $\xi\to0$ (and, in some cases, as $\xii\to\infty$). With this information, we obtain the decay rate for energy estimates for the solution to~\eqref{eq:CPlin} (Theorems~\ref{thm:CPQ2} and~\ref{thm:verygen}) and we describe the asymptotic profile of the solution, assuming a moment condition on the initial data (Theorem~\ref{thm:asymp}).
\begin{definition}\label{def:hyperbolic}
A complex polynomial $p(z)$ is hyperbolic if its roots are real-valued. It is strictly hyperbolic if its roots are real and simple. A complex polynomial $q(z)$ is strictly stable if $\Re z<0$ for any root of $q(z)$.
\end{definition}
\begin{definition}\label{def:stable}
A complex polynomial $q(z)$ is strictly stable if $\Re z<0$ for any root of $q(z)$.
\end{definition}
\begin{definition}\label{def:interlace}[see Definition 6.3.1 in~\cite{R-S}]
Let $p_{m-1}(z)$ be a hyperbolic polynomial of degree~$m-1$ with roots $\tilde\lambda_1\leq\ldots\leq\tilde\lambda_{m-1}$, and let $p_m(z)$ be a hyperbolic polynomial of degree~$m$ with roots $\lambda_1\leq\ldots\leq\lambda_m$.
We say that $p_{m-1}(z)$ and $p_m(z)$ interlace if
\begin{equation}\label{eq:weakinterlace}
\lambda_1 \leq \tilde\lambda_1 \leq \lambda_2 \leq \tilde\lambda_2 \leq \ldots \leq \tilde\lambda_{m-1}\leq \lambda_m.
\end{equation}
We say that $p_{m-1}(z)$ and $p_m(z)$ strictly interlace if
\begin{equation}\label{eq:interlace}
\lambda_1 < \tilde\lambda_1 < \lambda_2 < \tilde\lambda_2 < \ldots < \tilde\lambda_{m-1}< \lambda_{m}.
\end{equation}
\end{definition}
We stress that if two polynomials strictly interlace, then they are both strictly hyperbolic.
\begin{definition}\label{def:hyp}
Let
\begin{equation}\label{eq:P}
P_m(\partial_t,\partial_x)=\sum_{k+|\alpha|=m} c_{k,\alpha} \partial_t^k\partial_x^\alpha,
\end{equation}
be a homogeneous operator of order~$m$, with~$c_{m,0}>0$. We say that~$P(\partial_t,\partial_x)$ is hyperbolic if its symbol
\[ P(\lambda,\xi)=\sum_{k+|\alpha|=m} c_{k,\alpha} \lambda^k\xi^\alpha, \]
is hyperbolic, that is, it admits only real-valued roots~$\lambda_j(\xi')$, for any given~$\xi'\in S^{n-1}=\{\xi\in\R^n: \ \xii=1\}$. We say that~$P(\partial_t,\partial_x)$ is strictly hyperbolic if $P(\lambda,\xi')$ is strictly hyperbolic, that is, the roots~$\lambda_j(\xi')$ are distinct for any given~$\xi'\in S^{n-1}$.
\end{definition}
We choose $Q(\partial_t,\partial_x)=Q_1(\partial_t,\partial_x)$ or $Q(\partial_t,\partial_x)=Q_2(\partial_t,\partial_x)$, in~\eqref{eq:CPlin}, where
\begin{align}
\label{eq:Q1}
Q_1(\partial_t,\partial_x)
& = P_m(\partial_t,\partial_x)+P_{m-1}(\partial_t,\partial_x),\\%=\sum_{k+|\alpha|=m-1,m} c_{k,\alpha} \partial_t^k\partial_x^\alpha,\\
\label{eq:Q2}
Q_2(\partial_t,\partial_x) & = Q_1(\partial_t,\partial_x)+P_{m-2}(\partial_t,\partial_x)
\end{align}
with $P_{m-j}(\partial_t,\partial_x)$ hyperbolic operators of order~$m-j$. Without loss of generality, we assume in the following that~$c_{m,0}=1$ and $c_{m-1,0}>0$; moreover, $c_{m-2,0}>0$ if we consider~$Q_2$.
By the Hermite--Biehler theorem (see, for instance, \cite[Theorem 6.3.4]{R-S}), for any fixed~$\xi\neq0$, the strict stability of $Q_1(\lambda,i\xi)$ or $Q_2(\lambda,i\xi)$ is equivalent to the strict interlacing of $P_{m-1}(\lambda,\xi)$ and $P_m(\lambda,\xi)$, or, respectively, $P_m(\lambda,\xi)-P_{m-2}(\lambda,\xi)$. It is natural to write this condition for $Q_1(\lambda,i\xi)$ with $\xi\neq0$, by means of a condition on the roots of $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$, with $\xi'\in S^{n-1}$.
\begin{lemma}\label{lem:stable1}(see~\cite{VD}, see also~\cite[Proposition 2.2]{VV})
The polynomial $Q_1(\lambda,i\xi)$ is strictly stable for any $\xi\neq0$ if, and only if, $P_m(\partial_t,\partial_x)$ and $P_{m-1}(\partial_t,\partial_x)$ are strictly hyperbolic operators, and $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ strictly interlace, as in~\eqref{eq:interlace}, for any~$\xi'\in S^{n-1}$.
\end{lemma}
For $Q=Q_2$, the situation is more complicated.
\begin{hypothesis}\label{hyp:Q2}
We make the following assumption:
\begin{itemize}
\item $P_m(\partial_t,\partial_x)$ and~$P_{m-2}(\partial_t,\partial_x)$ are hyperbolic;
\item $P_{m-1}(\partial_t,\partial_x)$ is strictly hyperbolic;
\item $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ interlace, as in~\eqref{eq:weakinterlace}, for any~$\xi'\in S^{n-1}$;
\item $P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$ interlace, as in~\eqref{eq:weakinterlace}, for any~$\xi'\in S^{n-1}$;
\item there is no $(\lambda,\xi')\in\R\times S^{n-1}$ such that $P_m(\lambda,\xi')=P_{m-1}(\lambda,\xi')=P_{m-2}(\lambda,\xi')=0$.
\end{itemize}
\end{hypothesis}
\begin{lemma}\label{lem:stable2}(\cite[Theorem 2.1]{VV})
The polynomial $Q_2(\lambda,i\xi)$ is strictly stable if, and only if, Hypothesis~\ref{hyp:Q2} holds.
\end{lemma}
The strict stability of the polynomial $Q(\lambda,i\xi)$ for any $\xi\neq0$, with $Q=Q_1$ or $Q=Q_2$, guaranteed by Lemmas~\ref{lem:stable1} and~\ref{lem:stable2}, respectively, means that~$\Re \lambda_j(\xi)<0$ for any~$\xi\neq0$ and~$j=1,\ldots,m$, where $\lambda_j(\xi)$ are the roots of $Q(\lambda,i\xi)$. Since the roots of a polynomial continuously depend on its coefficients (see, for instance, \cite[Theorem 1.31]{R-S}), we get a uniform bound~$\Re\lambda_j(\xi)\leq -c$ in any compact subset~$K$ of~$\R^n\setminus\{0\}$, for some~$c=c(K)>0$. This property is sufficient to obtain exponential decay in time of $\hat u(t,\xi)$ in such compact subset of~$\R^n\setminus\{0\}$. To obtain decay estimates and to describe the asymptotic profile of the solution to~\eqref{eq:CPlin}, we need to study the asymptotic behavior of $\lambda_j(\xi)$ as $\xi\to0$ and as $\xii\to\infty$.
The strict interlacing of $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ for any $\xi'\in S^{n-1}$ is sufficient to obtain an easy asymptotic behavior of the roots of $Q_1(\lambda,i\xi)$ as $\xi\to0$ and as $\xii\to\infty$, see later, respectively, \eqref{eq:strictlow} and~\eqref{eq:stricthigh}. The situation is more complicated for $Q=Q_2$ under the assumption of Hypothesis~\ref{hyp:Q2}, but it is much simpler if we strengthen this assumption requiring the strict interlacing of $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$, and of $P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$, for any $\xi'\in S^{n-1}$.
\begin{theorem}\label{thm:CPQ2}
Assume that~$P_m(\partial_t,\partial_x)$ and~$P_{m-1}(\partial_t,\partial_x)$ are strictly hyperbolic, that the polynomials $P_{m-1}(\lambda,\xi')$ and $P_m(\lambda,\xi')$ strictly interlace, for any~$\xi'\in S^{n-1}$, and that
\[ u_j\in H^{s_0-j} \cap L^q \qquad \text{for some $s_0\geq m-1$, $q\in[1,2]$, for $j=0,\ldots,m-1$.} \]
Then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_1$ satisfies the following decay estimate
\begin{equation}
\label{eq:estQ1}\begin{split}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s}
& \leq C\,(1+t)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-(m-2)}2}\,\sum_{j=m-2,m-1}\big(\|u_j\|_{L^q}+\|u_j\|_{H^{k+s-j}}\big) \\
& \qquad + C\,\sum_{j=0}^{m-3}(1+t)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-j}2}\,\big(\|u_j\|_{L^q}+\|u_j\|_{H^{k+s-j}}\big),
\end{split}\end{equation}
for any $k+s\leq s_0$, such that $k+s\geq m-2$ if $q=2$, or $n(1/q-1/2)+k+s>m-2$, otherwise, with~$C>0$ independent of the initial data.
Assume, moreover, that~$P_{m-2}(\partial_t,\partial_x)$ is strictly hyperbolic and that the polynomials~$P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$ strictly interlace, for any~$\xi'\in S^{n-1}$. Then the solution to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies the following decay estimate
\begin{equation}
\label{eq:estQ2}\begin{split}
\|\partial_t^k u(t,\cdot)\|_{\dot H^s}
& \leq C\,(1+t)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-(m-3)}2}\,\sum_{j=m-3}^{m-1}\big(\|u_j\|_{L^q}+\|u_j\|_{H^{k+s-j}}\big)\\
& \qquad + C\,\sum_{j=0}^{m-4}(1+t)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-j}2}\,\big(\|u_j\|_{L^q}+\|u_j\|_{H^{k+s-j}}\big),
\end{split}
\end{equation}
for $k+s\leq s_0$, such that $k+s\geq m-3$ if $q=2$, or $n(1/q-1/2)+k+s>m-3$, otherwise, with~$C>0$ independent of the initial data.
\end{theorem}
We stress that in estimate~\eqref{eq:estQ2} an additional power $(1+t)^{-\frac12}$ appears, with respect to estimate~\eqref{eq:estQ1}. This extra decay rate is due to the presence of the additional dissipative term $P_{m-2}(\partial_t,\partial_x)$.
It is well-known that problem~\eqref{eq:CPlin} is $H^s$ well-posed, since $P_m(\partial_t,\partial_x)$ is strictly hyperbolic and has constant coefficients. The decay rate which appears taking higher order derivatives and assuming initial data in $L^q$, for some $q\in[1,2)$, is a consequence of the presence of the dissipative terms $P_{m-1}(\partial_t,\partial_x)$ and $P_{m-2}(\partial_t,\partial_x)$, and it is a typical phenomenon of parabolic problems.
\begin{remark}
Theorem~\ref{thm:CPQ2} for $Q=Q_1$ naturally extends results which are very well known for damped wave equations~\cite{M76} and analogous second-order equations; indeed, we may formally set $m=2$ and Theorem~\ref{thm:CPQ2}, as well as the other results of this paper (Lemmas~\ref{lem:low} and~\ref{lem:high}, Theorem~\ref{thm:asymp}), remains valid for $Q=Q_1$. However, in that case, the roots of $Q_1(\lambda,i\xi)$ may be explicitly computed, so there is no benefit from our analysis.
\end{remark}
When $q=1$, the optimality of the decay rate
in~\eqref{eq:estQ1} and of the decay rate
in~\eqref{eq:estQ2} follows as a consequence of the asymptotic profile of the solution to~\eqref{eq:CPlin} (see later, Theorem~\ref{thm:asymp}), under the moment condition $M\neq0$, where
\begin{equation}\label{eq:moment}
M = \int_{\R^n} \big( u_{m-1}(x) + c_{m-1,0}\, u_{m-2}(x) + c_{m-2,0}\,u_{m-3}(x)\big)\,dx\,,
\end{equation}
where we set $c_{m-2,0}=0$ if we consider $Q=Q_1$. If $M=0$, then
\[ \|\partial_t^k u(t,\cdot)\|_{\dot H^s}=\textit{o}\big((1+t)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-(m-3)}2}\big),\]
see Theorem~\ref{thm:asymp}.
\begin{remark}
In the case $Q=Q_1$, Theorem~\ref{thm:CPQ2} has been proved in~\cite{DAJ16} in the special case $P_{m-1}(\lambda,\xi) = c\,\partial_\lambda P_m(\lambda,\xi)$ for $c>0$. We emphasize that a strictly hyperbolic polynomial and its derivative always strictly interlace. In the case $Q=Q_2$, Theorem~\ref{thm:CPQ2} has been proved in~\cite{DAJ17} in the special case $P_{m-j}(\lambda,\xi) = c_j\,\partial_\lambda^j P_m(\lambda,\xi)$ for $c_j>0$, $j=1,2$. In~\cite[Theorem 1]{DAJ17}, also additional lower order terms are considered, under the assumption that the coefficients $\{c_j\}$ themselves are the coefficients of a strictly stable polynomial $\sum_{j=0}^{r} c_j z^{r-j}$. Moreover, weakly hyperbolic operators~$P_m$ are considered; when $Q=Q_2$, $P_m$ may have at most double roots. This scenario, studied in~\cite{DAJ17} for the special case $P_{m-j}(\lambda,\xi) = c_j\,\partial_\lambda^j P_m(\lambda,\xi)$, is generalized in \textsection~\ref{sec:weak} to more general interlacing assumptions for $P_m$, $P_{m-1}$, $P_{m-2}$, see Lemma~\ref{lem:highweak}, in particular.
\end{remark}
We stress that Hypothesis~\ref{hyp:Q2} is a weaker assumption than the assumption that $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ strictly interlace, and $P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$ strictly interlace, for any~$\xi'\in S^{n-1}$. However, under this stronger assumption, we have a simpler description of the asymptotic behavior of the roots $\lambda_j(\xi)$ of $Q_2(\lambda,i\xi)$, as $\xi\to0$ and as $\xii\to\infty$ (see later, \eqref{eq:strictlow}, \eqref{eq:lambdalow2} and~\eqref{eq:lambdahighsimple},\eqref{eq:stricthigh}). This simpler behavior is also better, in the sense that it produces better decay estimates for the solution to~\eqref{eq:CPlin}. For this reason, in Theorem~\ref{thm:CPQ2}, we derived estimates under the stronger assumption of strict interlacing polynomials.
In \textsection~\ref{sec:weak} we discuss the different, interesting, scenarios which may occur when $Q(\lambda,i\xi)$ is strictly stable for any~$\xi\neq0$,
i.e., Hypothesis~\ref{hyp:Q2} holds, but the interlacing of the polynomials is not strict. Four phenomena may appear, according to the ways in which the strict interlacing condition is weakened:
\begin{itemize}
\item the polynomial decay rate is related to a different scaling of the time and space variables if the interlacing of $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$ is not strict for some $\xi'\in S^{n-1}$ (see~\eqref{en:strictlowloss} in Theorem~\ref{thm:lowloss});
\item a loss of decay rate appears in the decay estimate, if $P_{m-2}(\partial_t,\partial_x)$ is not strictly hyperbolic (see~\eqref{en:weaklowloss} in Theorem~\ref{thm:lowloss});
\item a \emph{regularity-loss type decay} appears if the interlacing of $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$ is not strict for some $\xi'\in S^{n-1}$ (see Proposition~\ref{prop:lossdecay});
\item a loss of regularity appears if $P_m(\partial_t,\partial_x)$ is not strictly hyperbolic (see Proposition~\ref{prop:regloss}).
\end{itemize}
Assuming only Hypothesis~\ref{hyp:Q2}, i.e., the sufficient and necessary condition for the strict stability of $Q_2(\lambda,i\xi)$, all or some of the four phenomena may appear. On the one hand, the obtained decay estimate is weaker than~\eqref{eq:estQ2}, which holds if we assume the strict interlacing of $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$ for any $\xi'\in S^{n-1}$. On the other hand, if $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$ do not strictly interlace for any $\xi'\in S^{n-1}$, we shall require more initial data regularity. In particular, if $P_m(\partial_t,\partial_x)$ is hyperbolic, but not strictly, then~\eqref{eq:CPlin} is not well-posed in $H^s$, in general, but only in $H^\infty$,
since a loss of $1$ derivative may occur.
\begin{theorem}\label{thm:verygen}
Assume that Hypothesis~\ref{hyp:Q2} holds, and that
\[ u_j\in L^q \cap H^\infty \qquad \text{for some $q\in[1,2]$, for $j=0,\ldots,m-1$.} \]
Then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies the following decay estimate
\begin{equation}
\label{eq:estQ2verygen}\begin{split}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s}
& \leq C\,\sum_{j=m-3}^{m-1}(1+t)^{-\eta}\,\|u_j\|_{L^q} + C\,\sum_{j=0}^{m-1} (1+t)^{-\frac\nu2}\,\|u_j\|_{H^{k+s+\nu-j}}\\
& \qquad + C\,\sum_{j=0}^{m-4}(1+t)^{-\frac{n}4\left(\frac1q-\frac12\right)-\frac{k+s-j}4}\,\|u_j\|_{L^q},
\end{split}\end{equation}
for any~$\nu\geq1$, where
\begin{equation}
\label{eq:estQ2worst}
\eta=\min\left\{\frac{n}4\left(\frac1q-\frac12\right)+\frac{k+s-(m-3)}4, \ \frac{n}2\left(\frac1q-\frac12\right)+\frac{k+s-(m-2)}2\right\},
\end{equation}
provided that $k+s\geq m-3$ if $q=2$, or $n(1/q-1/2)+k+s>m-3$ otherwise, with~$C>0$ independent of the initial data.
\end{theorem}
The possible regularity loss of $1$ derivative is described by the requirement $\nu\geq1$ in Theorem~\ref{thm:verygen}; the regularity-loss type decay means that the decay rate $-\nu/2$ improves if additional regularity is taken on initial data. The decay rate in~\eqref{eq:estQ2worst} compares the influence of the possible source of loss of decay, with respect to~\eqref{eq:estQ2}. The possibility that $P_{m-2}(\partial_t,\partial_x)$ is not strictly hyperbolic leads to a loss of $(1+t)^{\frac12}$ power of decay rate, which become analogous to the decay rate in~\eqref{eq:estQ1}. On the other hand, the possibility that the interlacing of $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$ is not strict, in general, leads to a decay rate structure, in which the power $4$ appears in the denominator of $n(1/q-1/2)+k+s-(m-3)$, instead of the power $2$. This phenomenon is related to a faster vanishing speed of the real part of some root of $Q_2(\lambda,i\xi)$, as $\xi\to0$.
\bigskip
In the setting of long-time decay estimates for higher order inhomogeneous equations with constant coefficients, we address the interested reader to~\cite{RS10}, where dispersive and Strichartz estimates are obtained. In particular, under different hypotheses of geometric type on the roots of the full symbol of the operator, the authors are able to derive $L^q-L^{q'}$ estimates, $q\in[2,\infty]$, where~$q'=q/(q-1)$, for inhomogeneous hyperbolic equations. The decay rate in these estimates has the classical form~$(1+t)^{-\kappa\left(\frac1q-\frac1{q'}\right)}$, where~$\kappa>0$ depends on the assumptions on the roots of the full symbol.
Very weak dissipative effects have been considered for higher order hyperbolic equations with time-dependent coefficients in~\cite{DAE12}.
A huge literature exists for partially dissipative first-order hyperbolic systems with constant coefficients, under suitable assumptions on the term of order zero, and its relations with the first-order term. We address the interested reader to~\cite{SK85}, and to~\cite{BZ11} and the references therein, being aware that this cannot be an exhaustive list. Dissipative estimates for first-order hyperbolic systems with time-dependent coefficients have been obtained in~\cite{W14}.
An advantage of working with equations instead of systems is that one may possibly formulate more explicit assumptions to describe dissipative effects and investigate more in details the asymptotic profile of the solution to a problem.
\begin{Not}
We use the following notation:
\begin{itemize}
\item $\hat u(t,\xi)=\mathfrak{F}(u(t,\cdot))(\xi)$ denotes the Fourier transform of $u(t,x)$, with respect to the variable $x$;
\item for a given $\xi\neq0$, we put $\xi'=\xi/\xii$; $\xi'$ may also denote a point on the unit sphere $S^{n-1}=\{\xi\in\R^n: \ \xii=1\}$;
\item $L^q=L^q(\R^n)$, $q\in[1,\infty)$ denotes the usual space of function whose $q$-th power is integrable and $L^\infty=L^\infty(\R^n)$ denotes the usual space of essentially bounded functions;
\item $H^s$, $s\geq0$, denotes the Sobolev space of $L^2$ functions~$f$ such that $\xii^s\hat f$ is also in~$L^2$, with norm $\|f\|_{H^s}=\|(1+\xii^2)^{\frac{s}2}\hat f\|_{L^2}$; we define the quantity $\|f\|_{\dot H^s}=\|\xii^s\hat f\|_{L^2}$; for integer values of $s$, $H^s$ is the space of $L^2$ functions whose distributional derivatives of order not larger than $s$ are in $L^2$, and $\|f\|_{\dot H^s}\equiv\sum_{|\alpha|=s}\|\partial_x^\alpha f\|_{L^2}$; $H^\infty=\cap_{s>0} H^s$ is the subspace of $\mathcal C^\infty$ functions with all derivatives in $L^2$;
\item we write $f\lesssim g$ or $g\gtrsim f$ if there exists $C>0$ such that $f\leq Cg$;
\item $I_a$ denotes the Riesz potential of order $a$, i.e., $\mathfrak{F}(I_af)=\xii^{-a}\hat f$ in appropriate sense.
\item in this paper, it is crucial to distinguish whether a polynomial is evaluated at $\xi$ (for instance, for the homogeneous polynomial $P_{m-j}(\lambda,\xi)$) or at $i\xi$ (for instance, for the inhomogeneous polynomial $Q(\lambda,i\xi)$); the variable of the roots of a polynomial will always be $\xi$, or $\xi'$ for homogeneous polynomials, independently if the polynomial is real or complex.
\end{itemize}
For any $\xi'\in S^{n-1}$, $a_1(\xi')$, \ldots, $a_m(\xi')$ denote the $m$ roots of~$P_m(\lambda,\xi')$, $b_1(\xi')$, \ldots, $b_{m-1}(\xi')$ denote the $m-1$ roots of~$P_{m-1}(\lambda,\xi')$ and $d_1(\xi')$, \ldots, $d_{m-2}(\xi')$ denote the $m-2$ roots of~$P_{m-2}(\lambda,\xi')$. Namely,
\begin{align*}
P_m(\lambda,\xi')
& =\prod_{k=1}^m (\lambda-a_k(\xi')),\\
P_{m-1}(\lambda,\xi')
& =c_{m-1,0}\,\prod_{k=1}^{m-1} (\lambda-b_k(\xi')),\\
P_{m-2}(\lambda,\xi')
& =c_{m-2,0}\,\prod_{k=1}^{m-2} (\lambda-d_k(\xi')).
\end{align*}
On the other hand, $\lambda_1(\xi)$, \ldots, $\lambda_m(\xi)$ denote the $m$ complex-valued roots of $Q(\lambda,i\xi)$. These latter are not homogeneous, but to simplify some computation in the study of their asymptotic behavior, we introduce the auxiliary functions $\mu_j(\rho,\xi')=\rho^{-1}\lambda_j(\rho\xi')$, where $\rho=\xii$ and $\xi'=\xi/\xii$, for any $\xi\neq0$.
For a given homogeneous hyperbolic polynomial~$P_m(\lambda,\xi)$ of order $m$ with roots $\lambda_j(\xi')$ for any~$\xi'\in S^{n-1}$, we also introduce the following notation:
\begin{align*}
\check{P}_{m,j}(\lambda,\xi')
& = \frac{P_m(\lambda,\xi')}{\lambda-\lambda_j(\xi')},\\
\check{P}_{m,j,\ell}(\lambda,\xi')
& = \frac{P_m(\lambda,\xi')}{(\lambda-\lambda_j(\xi'))(\lambda-\lambda_\ell(\xi'))}.
\end{align*}
\end{Not}
\section{Asymptotic profiles of the roots of the symbol and of the solution to~\eqref{eq:CPlin}}\label{sec:asymp}
We may study the asymptotic behavior of the roots of $Q(\lambda,i\xi)$ as~$\xi\to0$ and as~$\xii\to\infty$ in a more general setting than the one in Theorem~\ref{thm:CPQ2}. In general, in the following, we put
\begin{equation}\label{eq:Q}
Q(\lambda,i\xi)= \sum_{j=0}^\ell P_{m-j}(\lambda,i\xi),
\end{equation}
where $P_{m-j}(\partial_t,\partial_x)$ are homogeneous operators of degree~$m-j$, with $c_{m,0}=1$ and $c_{m-\ell,0}>0$ (same notation as in~\eqref{eq:P}). Following as in~\cite{VV}, it is easy to see that a necessary condition for the strict stability of $Q(\lambda,i\xi)$ for any~$\xi\neq0$, is the interlacing of $P_{m-\ell}(\lambda,\xi')$ and $P_{m-\ell+1}(\lambda,\xi')$, and of $P_{m-1}(\lambda,\xi')$ and $P_m(\lambda,\xi')$, for any $\xi'\in S^{n-1}$.
Strengthening this assumption to the strict interlacing, the asymptotic behavior of the roots of $Q(\lambda,i\xi)$ is only determined by $P_{m-\ell}$ and $P_{m-\ell+1}$ as~$\xi\to0$, and by $P_m$ and $P_{m-1}$ as~$\xii\to\infty$, whereas the other polynomials come into play otherwise (see \textsection\ref{sec:weak}). The following two lemmas describe such behavior.
\begin{lemma}\label{lem:low}
Assume that~$P_{m-\ell}(\partial_t,\partial_x)$ and~$P_{m-\ell+1}(\partial_t,\partial_x)$ are strictly hyperbolic, and that~$P_{m-\ell}(\lambda,\xi')$ and~$P_{m-\ell+1}(\lambda,\xi')$ strictly interlace for any $\xi'\in S^{n-1}$. Then we may label the $m$ roots~$\lambda_j(\xi)$ of~$Q(\lambda,i\xi)=0$, $\xi\in\R^n$, in such a way that $\lambda_m(0)$, \ldots, $\lambda_{m-\ell+1}(0)$ are the solutions of
\begin{equation}\label{eq:lambda0}
\sum_{j=0}^{\ell} c_{m-j,0}\lambda^{\ell-j}=0,
\end{equation}
and
\begin{equation}\label{eq:lambdalow}
\lambda_j(\xi) = i\xii \beta_j(\xi') + \xii^2\,\frac{P_{m-\ell+1}(\beta_j(\xi'),\xi')}{\check{P}_{m-\ell,j}(\beta_j(\xi'),\xi')} + \textit{o}(\xii^2), \quad j=1,\ldots,m-\ell,
\end{equation}
as~$\xi\to0$, where~$\beta_j(\xi')$ are the $m-\ell$ real roots of~$P_{m-\ell}(\lambda,\xi')$, and~$\xi'=\xi/\xii\in S^{n-1}$.
\end{lemma}
The strict interlacing of~$P_{m-\ell}(\lambda,\xi')$ and~$P_{m-\ell+1}(\lambda,\xi')$ and the assumption that $c_{m-\ell},c_{m-\ell+1}>0$, guarantee that
\begin{equation}\label{eq:strictlow}
\frac{P_{m-\ell+1}(\beta_j(\xi'),\xi')}{\check{P}_{m-\ell,j}(\beta_j(\xi'),\xi')} = \frac{c_{m-\ell+1,0}\prod_{k=1}^{m-\ell+1} (\beta_j(\xi')-\alpha_k(\xi'))}{c_{m-\ell,0}\prod_{k\neq j}(\beta_j(\xi')-\beta_k(\xi'))} <0,
\end{equation}
in~\eqref{eq:lambdalow}, where $\alpha_j$ are the $m-\ell+1$ real roots of~$P_{m-\ell+1}(\lambda,\xi')$.
In particular, for~$\ell=1,2$ in Lemma~\ref{lem:low}, we obtain the following information:
\begin{itemize}
\item if~$Q=Q_1$ and~$P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ strictly interlace for any $\xi'\in S^{n-1}$, then $\lambda_m(0) = -c_{m-1,0}$ and
\begin{equation}\label{eq:lambdalow1}
\lambda_j(\xi) = i\xii b_j(\xi') + \xii^2\,\frac{P_{m-\ell+1}(b_j(\xi'),\xi')}{\check{P}_{m-\ell,j}(b_j(\xi'),\xi')} + \textit{o}(\xii^2), \quad j=1,\ldots,m-1,
\end{equation}
as~$\xi\to0$, where~$b_j(\xi')$ are the $m-1$ real roots of~$P_{m-1}(\lambda,\xi')$, and~$\xi'=\xi/\xii\in S^{n-1}$;
\item if~$Q=Q_2$ and~$P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$ strictly interlace for any $\xi'\in S^{n-1}$, then
\begin{equation}\label{eq:lambdam-1m}
\lambda_{m-1,m}(0) = \begin{cases}
\dfrac{-c_{m-1,0}\pm \sqrt{c_{m-1,0}^2-4c_{m-2,0}}}2 & \text{if~$c_{m-1,0}^2\geq4c_{m-2,0}$,}\\
\dfrac{-c_{m-1,0}\pm i \sqrt{4c_{m-2,0}-c_{m-1,0}^2}}2 & \text{if~$c_{m-1,0}^2<4c_{m-2,0}$,}
\end{cases}\end{equation}
and
\begin{equation}\label{eq:lambdalow2}
\lambda_j (\xi) = i\xii d_j(\xi') + \xii^2\,\frac{P_{m-1}(d_j(\xi'),\xi')}{\check{P}_{m-2,j}(d_j(\xi'),\xi')} + \textit{o}(\xii^2),\quad j=1,\ldots,m-2,
\end{equation}
as~$\xi\to0$, where~$d_j(\xi')$ are the $m-2$ roots of~$P_{m-2}(\lambda,\xi')$, and~$\xi'=\xi/\xii\in S^{n-1}$.
\end{itemize}
\begin{lemma}\label{lem:high}
Assume that~$P_{m-1}(\partial_t,\partial_x)$ and~$P_{m}(\partial_t,\partial_x)$ are strictly hyperbolic, and that~$P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ strictly interlace for any $\xi'\in S^{n-1}$. Then we may label the $m$ roots~$\lambda_j(\xi)$ of~$Q(\lambda,i\xi)$, $\xi\in\R^n$, in such a way that:
\begin{equation}\label{eq:lambdahighsimple}
\lambda_j (\xi) = i\xii a_j(\xi') - \frac{P_{m-1}(a_j(\xi'),\xi')}{\check{P}_{m,j}(a_j(\xi'),\xi')} + \textit{o}(1),\quad j=1,\ldots,m,
\end{equation}
as~$\xii\to\infty$, where~$a_j(\xi')$ are the $m$ real roots of~$P_m(\lambda,\xi')$ and~$\xi'=\xi/\xii\in S^{n-1}$.
\end{lemma}
The strict interlacing of~$P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ guarantees that
\begin{equation}\label{eq:stricthigh}
- \frac{P_{m-1}(a_j(\xi'),\xi')}{\check{P}_{m,j}(a_j(\xi'),\xi')} = - \frac{c_{m-1,0}\prod_{k=1}^{m-1}(a_j(\xi')-b_k(\xi'))}{\prod_{k\neq j}(a_j(\xi')-a_k(\xi'))} <0,
\end{equation}
in~\eqref{eq:lambdahighsimple}, where~$b_j(\xi')$ are the $m-1$ roots of~$P_{m-1}(\lambda,\xi')$.
\begin{remark}
In the special case $P_{m-j}(\lambda,\xi)=\partial_\lambda^jP_m(\lambda,\xi)$, a slightly different version of Lemmas~\ref{lem:low} and~\ref{lem:high} has been previously obtained in~\cite[Lemmas~2 and~3]{DAJ17}.
\end{remark}
\subsection*{Asymptotic profile of the solution to~\eqref{eq:CPlin}}
Thanks to Lemma~\ref{lem:low}, assuming initial data in $L^1$, we may refine Theorem~\ref{thm:CPQ2} to describe the asymptotic profile of the solution to~\eqref{eq:CPlin} as $t\to\infty$. We define $v(t,\cdot)$ and $w(t,\cdot)$ in $L^2$ by means of
\begin{align}
\label{eq:v}
\hat v(t,\xi)
& = M\,\sum_{j=1}^{m-1}\frac{e^{i\xii\,b_j(\xi')t+\xii^2\,\frac{P_m(b_j(\xi'),\xi')}{\check{P}_{m-1,j}(b_j(\xi'),\xi')}\,t}}{i^{m-2}\,\check{P}_{m-1,j}(b_j(\xi'),\xi')}\,,\\
\label{eq:w}
\hat w(t,\xi)
& = M\,\sum_{j=1}^{m-2}\frac{e^{i\xii\,d_j(\xi')t+\xii^2\,\frac{P_{m-1}(d_j(\xi'),\xi')}{\check{P}_{m-2,j}(d_j(\xi'),\xi')}t}}{i^{m-3}\,\check{P}_{m-2,j}(d_j(\xi'),\xi')}\,,
\end{align}
where $M$ is as in~\eqref{eq:moment} and $\xi'=\xi/\xii$. Then the asymptotic profile of the solution to~\eqref{eq:CPlin} is described by $I_{m-2}v$ if~$Q=Q_1$ and by $I_{m-3}w$ if $Q=Q_2$, where $I_a$ denotes the Riesz potential of order $a$, i.e., $\mathfrak{F}(I_af)=\xii^{-a}\hat f$ in appropriate sense. The expressions of $\hat v$ and $\hat w$ appear complicated but they become much easier in several cases of interest, see Examples~\ref{ex:MGT}, \ref{ex:BC}, \ref{ex:electric}.
\begin{theorem}\label{thm:asymp}
Assume that~$P_m(\partial_t,\partial_x)$ and~$P_{m-1}(\partial_t,\partial_x)$ are strictly hyperbolic, that the polynomials $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$ strictly interlace for any~$\xi'\in S^{n-1}$ and that
\[ u_j\in H^{s_0-j} \cap L^1 \qquad \text{for some $s_0\geq m-1$, for $j=0,\ldots,m-1$.} \]
Then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_1$ satisfies the following estimate:
\begin{equation}
\label{eq:Q1asymp}
\big\|\partial_t^k \big( u(t,\cdot) - I_{m-2} v(t,\cdot)\big)\big\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}4-\frac{k+s-(m-2)}2}\big)
\end{equation}
for $k+s\leq s_0$, such that $n/2+k+s>m-2$, where $M$ is as in~\eqref{eq:moment} and $v$ is as in~\eqref{eq:v}. Assume, moreover, that~$P_{m-2}(\partial_t,\partial_x)$ is strictly hyperbolic and that the polynomials~$P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$ strictly interlace for any~$\xi'\in S^{n-1}$. Then the solution to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies the following estimate:
\begin{equation}
\label{eq:Q2asymp}
\big\|\partial_t^k \big( u(t,\cdot) - I_{m-3}w(t,\cdot)\big)\big\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}4-\frac{k+s-(m-3)}2}\big)
\end{equation}
for $k+s\leq s_0$, such that $n/2+k+s>m-3$, where $M$ is as in~\eqref{eq:moment} and $w$ is as in~\eqref{eq:w}.
\end{theorem}
When $M\neq0$, Theorem~\ref{thm:asymp} describes the asymptotic profile of the solution to~\eqref{eq:CPlin}. In particular, it guarantees the optimality of the decay rate~\eqref{eq:estQ1} in Theorem~\ref{thm:CPQ2} when $q=1$ and $M\neq0$, due to
\begin{align*}
\| I_{m-2}v(t,\cdot)\|_{\dot H^s}
& = \Big\| \xii^{s-(m-2)}\,\sum_{j=1}^{m-1}\frac{e^{\xii^2\,\frac{P_m(b_j(\xi'),\xi')}{\check{P}_{m-1,j}(b_j(\xi'),\xi')}t}}{\check{P}_{m-1,j}(b_j(\xi'),\xi')} \Big\|_{L^2}\\
& = t^{-\frac{n}4-\frac{s-(m-2)}2}\,\Big\|\xii^{s-(m-2)}\,\sum_{j=1}^{m-1} \frac{e^{\xii^2\frac{P_m(b_j(\xi'),\xi')}{\check{P}_{m-1,j}(b_j(\xi'),\xi')}}}{\check{P}_{m-1,j}(b_j(\xi'),\xi')} \Big\|_{L^2},
\end{align*}
when~$n/2+s>m-2$. We proceed similarly for the decay rate in~\eqref{eq:estQ2}, using $w$. For $\partial_t^k v$ and $\partial_t^kw$, we may proceed in a similar way, but in this case we shall carefully treat the cases in which $b_j(\xi')=0$ for some $\xi'\in S^{n-1}$, due to the different behavior of $|\lambda_j^k|$.
On the other hand, when $M=0$, Theorem~\ref{thm:asymp} implies that the estimates in Theorem~\ref{thm:CPQ2} are not optimal. Indeed, in this case, the estimates may be improved, see Remark~\ref{rem:M0}, and the asymptotic profile of the solution to~\eqref{eq:CPlin} may be described by assuming further moment conditions.
\begin{remark}
In the setting of linear and semilinear damped wave equations, the asymptotic profile of the solution, as $t\to\infty$, is described by the fundamental solution to a heat equation or, more in general, to a diffusive problem, a property called \emph{diffusion phenomenon}, see~\cite{MN03, Ni, N11as, HM00, W14sp}. The diffusion phenomenon also appears in wave and evolution models with structural damping~\cite{DAE14JDE, DAE17NA, K00}. In the case of higher order equations, the diffusion phenomenon does not generally appear, due the presence of the oscillating terms $e^{i\xii b_j(\xi')t}$ and $e^{i\xii d_j(\xi')t}$, in~\eqref{eq:v} and~\eqref{eq:w}. However, in the special case $m=3$ and $P_1=\lambda$, Theorem~\ref{thm:asymp} for $Q=Q_2$ may be improved to show that
\[ \|\partial_t^k (u(t,\cdot)-w(t,\cdot))\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}4-\frac{s}2-k}\big), \
where $w$ is $M$ times the fundamental solution to
\[ \partial_t w - \sum_{|\alpha|=2}c_{0,\alpha}\partial_x^\alpha w=0. \]
In particular, each time derivative produces an additional decay~$t^{-1}$ in place of $t^{-\frac12}$.
We stress that $\sum_{|\alpha|=2}c_{0,\alpha}\xi^\alpha>0$ for any $\xi\neq0$, as a consequence of the strict hyperbolicity of $P_2(\partial_t,\partial_x)$.
\end{remark}
\section{Some examples of physical interest}\label{sec:examples}
In this section, we provide a few examples of models with physical interest to which Theorems~\ref{thm:CPQ2} and~\ref{thm:asymp} and, more in general, the results in \textsection~\ref{sec:asymp}, directly apply. However, it is clear that our results apply to any hyperbolic homogeneous equation of order $m\geq3$, to which a hyperbolic homogeneous operator of order $m-1$ and, possibly, a hyperbolic homogeneous operator of order $m-2$, are added to produce a dissipation. In \textsection\ref{sec:weak}, two other examples are provided to discuss the case of weak interlacing when the lowest order term has a double root.
\begin{example}\label{ex:MGT}
The Jordan-Moore-Gibson-Thompson equation (see~\cite{J,KLM,K}) is a hyperbolic model for acoustic waves introduced to preserve the finite speed of propagation. Its linear version, called Moore-Gibson-Thompson (MGT) equation, is $L(\partial_t,\partial_x)u=f$, with
\[ L(\partial_t,\partial_x) = \tau \partial_t^3 + \partial_t^2 - c^2\Delta -(\tau c^2+b)\Delta \partial_t, \]
where $\tau>0$ is the intrinsic relaxation time of the heat flux~\cite{CFO}, $c>0$ is the sound speed, and $b>0$ is the diffusivity of sound. As $\tau\to0$, the equation reduces to the linear Kuznetsov~\cite{Kunetsov} or Westervelt~\cite{Westervelt} equation, i.e., a wave equation with viscoelastic damping~\cite{Ponce, S00}:
\begin{equation}\label{eq:visco}
\partial_t^2u - c^2\Delta u -b\Delta \partial_tu=0.
\end{equation}
Setting $L(\partial_t,\partial_x)u=\tau Q_1(\partial_t,\partial_x)u$, this corresponds to fix
\begin{align*}
P_3(\lambda,\xi')
& = \lambda^3-(c^2+b\tau^{-1})\lambda=(\lambda+\sqrt{c^2+b\tau^{-1}}) \lambda (\lambda-\sqrt{c^2+b\tau^{-1}}),\\
P_2(\lambda,\xi')
& = \tau^{-1}(\lambda^2-c^2) = \tau^{-1}\,(\lambda+c)(\lambda-c).
\end{align*}
In particular, $P_3$ and $P_2$ strictly interlace for any $b>0$, since the roots of $P_2(\lambda,\xi')$ are $(b_1,b_2)=(-c,c)$, whereas the roots of $P_3(\lambda,\xi')$ are
\[ a_1=- \sqrt{c^2+\frac{b}{\tau}}, \quad a_2=0,\quad a_3= \sqrt{c^2+\frac{b}{\tau}}. \]
We may now compute
\[ P_3(\pm c,\xi')=\mp c b\tau^{-1}, \qquad \check{P}_{2,1}(-c,\xi')=-2c\tau^{-1},\quad \check{P}_{2,2}(c,\xi')=2c\tau^{-1}, \]
so that, by Lemma~\ref{lem:low}, the $3$ roots $\lambda_j(\xi)$ of $Q(\lambda,i\xi)$ have the following asymptotic behavior as $\xi\to0$:
\[ \lambda_3(0)=-\tau^{-1}, \qquad \lambda_{1,2}(\xi) = \pm ic\xii - \frac{b}2\,\xii^2 + \textit{o}(\xii^2). \]
We stress that $\lambda_{1,2}$ have the same asymptotic behaviors, as~$\xi\to0$, of the roots of the viscoelastic damped wave equation obtained setting $\tau=0$ in the MGT equation:
\[ \lambda_\pm (\xi) =
-\frac{b \xii^2 \pm i\xii \sqrt{4c^2-b^2\xii^2}}{2} = \pm ic\xii-\frac{b}2\,\xii^2+\textit{o}(\xii^2), \quad \xii\ll 1.
\]
On the other hand, $\lambda_3(0)\to-\infty$ as~$\tau\to0$.
We may write the asymptotic profile of the solution to the MGT equation as $I_1v$, where
\begin{align*}
\mathfrak{F}(I_1v)(t,\xi)
& = \xii^{-1}\hat v(t,\xi) = M\,\frac{e^{ic\xii t-\frac{b}2\xii^2 t}-e^{-ic\xii t-\frac{b}2\xii^2 t}}{2ic\tau^{-1}\xii}= M\tau\,\frac{\sin(c\xii t)}{c\xii}\,e^{-\frac{b}2\xii^2 t},\\
M\tau
& = \int_{\R^n} \big( \tau\,u_2(x) + u_1(x) \big) \,dx.
\end{align*}
By Theorem~\ref{thm:asymp}, we find
\[ \big\|\partial_t^k \big( u(t,\cdot) - I_1v(t,\cdot)\big)\big\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}4-\frac{k+s-1}2}\big),\
for $n/2+k+s>1$. As $\tau\to0$, we find the same kind of result obtained for the wave equation with viscoelastic damping~\eqref{eq:visco} in~\cite{I14}.
\end{example}
\begin{example}\label{ex:BC}
The Kuznetsov and Westervelt equations can formally be regarded as a simplification of the Blackstock-Crighton (BC) equation~\cite{Blackstock}. One of the model of the linear BC equation is
\begin{equation}\label{eq:BC}
\partial_t(\partial_t^2-c^2\Delta-b\Delta \partial_t)u-a\Delta(\partial_t^2-c^2\Delta)u = 0,
\end{equation}
where $a>0$ is the thermal conductivity, whereas $c>0$ and $b>0$ are the sound speed and the diffusivity of sound, as in the Kuznetsov, Westervelt and MGT equations in Example~\ref{ex:BC}. Replacing the first term in the BC equation, that is, the time-derivative of the wave equation with viscoelastic damping, by the time-derivative of the MGT equation, we obtain the fourth order hyperbolic equation $\tau Q_1(\partial_t,\partial_x)u=0$, where
\begin{align*}
P_4(\lambda,\xi')
& = \lambda^4-(c^2+(a+b)\tau^{-1})\lambda^2+ac^2\tau^{-1}=(\lambda+a_4)(\lambda+a_3)(\lambda-a_3)(\lambda-a_4), \\
P_3(\lambda,\xi')
& = \tau^{-1}\lambda(\lambda^2-c^2) = \tau^{-1}\,(\lambda+c)\lambda(\lambda-c),
\end{align*}
where the roots of $P_4(\lambda,\xi')$ are
\[ a_{3,4}=\sqrt{\frac{c^2+(a+b)\tau^{-1}\pm \sqrt{(c^2+(a+b)\tau^{-1})^2-4a\tau^{-1}c^2}}2}, \quad a_2=-a_3, \quad a_1=-a_4,\]
and the roots of $P_3(\lambda,\xi')$ are $(b_1,b_2,b_3)=(-c,0,c)$, for any $\xi'\in S^{n-1}$. The strict interlacing condition is verified if, and only if, $b>0$.
Due to
\begin{align*}
& P_4(\pm c,\xi')=bc^2\tau^{-1}, \quad P_4(0,\xi')=ac^2\tau^{-1},\\
& \check{P}_{3,1}(-c,\xi')=\check{P}_{3,3}(c,\xi')=2c^2\tau^{-1}, \quad \check{P}_{3,2}(0,\xi')=-c^2\tau^{-1},
\end{align*}
by Lemma~\ref{lem:low}, the four roots $\lambda_j(\xi)$ of $Q(\lambda,i\xi)$ have the following asymptotic behavior, as $\xi\to0$:
\begin{align*}
\lambda_4(0)
& = -\tau^{-1},\\
\lambda_{1,3}(\xi)
& = \pm ic\xii - \frac{b}2\,\xii^2 + \textit{o}(\xii^2), \\
\lambda_2(\xi)
& = -a\,\xii^2 + \textit{o}(\xii^2).
\end{align*}
In particular, exception given for $\lambda_2$, the asymptotic behavior is the same of the roots of the MGT equation, see Example~\ref{ex:MGT}.
Setting~$\tau=0$, we may compare those asymptotic behaviors with the asymptotic behaviors of the roots of the BC equation~\eqref{eq:BC}.
Recalling (see for instance, Routh-Hurwitz theorem) that a polynomial of third order with positive coefficients
\[ q(z)=z^3+\alpha_2z^2+\alpha_1z+\alpha_0 \]
is strictly stable if, and only if, $\alpha_0 < \alpha_1\alpha_2$,
we find that~$Q(\lambda,i\xi)$, with
\[ Q(\lambda,i\xi) = \lambda(\lambda^2+c^2\xii^2+b\xii^2\lambda)+a\xii^2(\lambda^2+c^2\xii^2), \]
as in~\eqref{eq:BC}, is strictly stable for any~$\xi\neq0$ if, and only if, $b>0$. Following as in the proof of Lemma~\ref{lem:low}, it is easy to show that the three roots~$\lambda_0(\xi)$ and~$\lambda_\pm(\xi)$ of $Q(\lambda,i\xi)$ verify
\[ \lambda_\pm(\xi)= \pm ic\xii - \frac{b}2\,\xii^2 + \textit{o}(\xii^2) = \lambda_{1,3}(\xi),\qquad \lambda_0(\xi)=-a\,\xii^2 + \textit{o}(\xii^2)=\lambda_2(\xi), \]
as~$\xi\to0$. On the other hand, $\lambda_4(0)\to-\infty$ as~$\tau\to0$.
We may write the asymptotic profile of the solution to the BC equation as $I_2v$, where
\begin{align*}
\mathfrak{F}(I_2v)(t,\xi)
& = \xii^{-2}\hat v(t,\xi) = M\,\frac{e^{ic\xii t-\frac{b}2\xii^2 t}+e^{-ic\xii t-\frac{b}2\xii^2 t}}{-2c^2\tau^{-1}\xii^2} + M\,\frac{e^{-a\xii^2 t}}{c^2\tau^{-1}\xii^2} \\
& = \frac{M\tau}{c^2\xii^2}\left(e^{-a\xii^2 t}-\cos(c\xii t)\,e^{-\frac{b}2\xii^2 t}\right)\,,\\
M\tau
& = \int_{\R^n} \big( \tau\,u_3(x) + u_2(x) \big) \,dx.
\end{align*}
By Theorem~\ref{thm:asymp}, we get
\[ \big\|\partial_t^k \big( u(t,\cdot) - I_2v(t,\cdot)\big)\big\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}4-\frac{k+s-2}2}\big),\
for $n/2+k+s>2$.
\end{example}
\begin{example}\label{ex:electric}
Let us consider the coupled system of elastic waves with Maxwell equations in~$\R^3$ (see, for instance, \cite{DLM09}), i.e.
\begin{equation}\label{eq:MHD3dplus}
\begin{cases}
\partial_t^2 u - \mu \Delta u - (\mu+\nu) \nabla \div u + \gamma \curl E =0,\\
\partial_t E +\sigma E - \curl H -\gamma\curl \partial_t u=0,\\
\partial_t H + c^2\curl E =0,
\end{cases}
\end{equation}
where~$\curl = \nabla\times$ denotes the curl operator, $\mu,\nu$ are the Lam\`e constants and verify $\mu>0$, $\mu+\nu>0$ and, after normalization, $\gamma>0$ is the coupling constant, $\sigma>0$ is the electric conductivity (we replaced $J=\sigma E$, by Ohm's law, in Maxwell's equations) and $c^2>0$ is the inverse of the product of the electric permittivity and magnetic permeability. Recalling that $\div \curl =0$, we notice that~$\div E$ verifies the scalar equation
\begin{equation}\label{eq:divE}
\partial_t \div E + \sigma \div E =0.
\end{equation}
Deriving with respect to $t$ the second equation in~\eqref{eq:MHD3dplus} and replacing the third one in it, recalling that
\begin{equation}\label{eq:curlcurl}
\curl \curl f = - \Delta f + \nabla \div f,
\end{equation}
we get the second-order $6\times 6$ system
\begin{equation}\label{eq:MHD3hom}
\begin{cases}
\partial_t^2 u - \mu \Delta u - (\mu+\nu) \nabla \div u + \gamma \curl E =0,\\
\partial_t^2 E +\sigma \partial_t E -c^2\Delta E + c^2\nabla \div E - \gamma \curl \partial_t^2 u=0.
\end{cases}
\end{equation}
Applying the operator~$ \partial_t^2-\mu\Delta-(\mu+\nu)\nabla \div $ to the second equation in~\eqref{eq:MHD3hom}, recalling~\eqref{eq:curlcurl}, we get a fourth-order $3\times3$ system for $E$:
\begin{equation}\label{eq:Evector}
(\partial_t^2 - \mu \Delta - (\mu+\nu) \nabla \div)(\partial_t^2+\sigma\partial_t-c^2\Delta+c^2\nabla \div)E + \gamma^2\partial_t^2 (-\Delta+\nabla\div) E=0.
\end{equation}
Since $\div E$ verifies~\eqref{eq:divE}, applying the operator~$\partial_t+\sigma$ to~\eqref{eq:Evector}, we may remove every term where $\div E$ appears and obtain the fifth-order scalar equation given by
\[ (\partial_t+\sigma) \big( (\partial_t^2-\mu\Delta)(\partial_t^2+\sigma\partial_t-c^2\Delta)-\gamma^2\partial_t^2\Delta\big) E=0, \]
which every component of $E$ shall satisfy. The equation above is $Q_2(\partial_t,\partial_x)E=0$, where $Q_2=P_5+P_4+P_3$ and
\begin{align*}
P_5(\lambda,\xi)
& = \lambda\,\big(\lambda^4-(\mu+c^2+\gamma^2)\lambda^2\xii^2+c^2\mu\xii^4\big), \\
P_4(\lambda,\xi)
& = \sigma\,\big(2\lambda^4-(2\mu+c^2+\gamma^2)\lambda^2\xii^2+c^2\mu\xii^4\big), \\
P_3(\lambda,\xi)
& = \sigma^2 \lambda (\lambda^2-\mu\xii^2).
\end{align*}
The roots of $P_3(\lambda,\xi')$ are $(d_1,d_2,d_3)=(-\sqrt{\mu},0,\sqrt{\mu})$, the roots of $P_4(\lambda,\xi')$ are
\[ b_{3,4} = \frac12\,\sqrt{2\mu+c^2+\gamma^2\pm \sqrt{(2\mu+c^2+\gamma^2)^2-8c^2\mu}}, \qquad b_{1,2}=-b_{3,4}, \]
and the roots of $P_5(\lambda,\xi')$ are
\[ a_{4,5} = \frac1{\sqrt{2}}\,\sqrt{\mu+c^2+\gamma^2\pm \sqrt{(\mu+c^2+\gamma^2)^2-4c^2\mu}}, \qquad a_3=0, \quad a_{1,2}=-a_{4,5}. \]
The strict interlacing condition of $P_3(\lambda,\xi')$ and $P_4(\lambda,\xi')$ is verified if $b_3<1<b_4$, that is,
\begin{align*}
2\mu+c^2+\gamma^2-\sqrt{(2\mu+c^2+\gamma^2)^2-8c^2\mu} < 4\mu < 2\mu+c^2+\gamma^2+\sqrt{(2\mu+c^2+\gamma^2)^2-8c^2\mu},
\end{align*}
which holds if, and only if,
\[ \big(2\mu-(c^2+\gamma^2)\big)^2 < (2\mu+c^2+\gamma^2)^2-8c^2\mu, \]
i.e., $\gamma^2>0$. With long but straightforward calculations, one may also check that $P_4(\lambda,\xi')$ and $P_5(\lambda,\xi')$ strictly interlace.
Therefore, Theorems~\ref{thm:CPQ2} and~\ref{thm:asymp} apply. Due to
\begin{align*}
& P_4(\pm\sqrt{\mu},\xi')=-\sigma\mu\gamma^2, \qquad P_4(0,\xi')=\sigma c^2\mu, \\
& \check{P}_{3,3}(\sqrt{\mu},\xi')=2\sigma^2\mu, \quad \check{P}_{3,2}(0,\xi')=-\sigma^2\mu, \quad \check{P}_{3,1}(-\sqrt{\mu},\xi')=2\sigma^2\mu,
\end{align*}
we may compute
\[ \lambda_{4,5}(0)=-\sigma, \qquad \lambda_{1,3}(\xi)=\pm i\xii\sqrt{\mu}-\frac{\gamma^2}{2\sigma}\xii^2+\textit{o}(\xii^2), \qquad \lambda_2(\xi)=-\frac{c^2}\sigma\,\xii^2+\textit{o}(\xii^2), \]
as $\xi\to0$. Therefore, the asymptotic profile of $E(t,\cdot)$, for each of its components, is described by $I_2w$, where
\begin{align*}
\mathfrak{F}(I_2w)(t,\xi)
& = \xii^{-2}\hat w(t,\xi) = -\,M\,e^{-\frac{\gamma^2}{2\sigma}\xii^2t}\,\frac{e^{i\xii\sqrt{\mu}\,t}+e^{-i\xii\sqrt{\mu}\,t}}{2\mu\sigma^2\xii^2} + M\,\frac{e^{-\frac{c^2}\sigma\,\xii^2t}}{\mu\sigma^2\xii^2}\\
& = \frac{M}{\mu\sigma^2\xii^2}\,\left(e^{-\frac{c^2}\sigma\,\xii^2t} - \cos(\xii\sqrt{\mu}\,t)\,e^{-\frac{\gamma^2}{2\sigma}\xii^2t} \right),
\intertext{if}
M
& = \int_{\R^3} \big(\partial_t^4 E(0,x)+\sigma \partial_t^3 E(0,x)+\sigma^2 \partial_t^2 E(0,x)\big)\,dx \neq0.
\end{align*}
Explicitly,
\[ \|\partial_t^k\big( E(t,\cdot)-I_2w(t,\cdot)\big)\|_{\dot H^s} =\textit{o}\big((1+t)^{-\frac34-\frac{k+s-2}2}\big),\
for any $k+s>1/2$.
\end{example}
\section{Proofs of the asymptotic profiles and of Theorems~\ref{thm:CPQ2} and~\ref{thm:asymp}}\label{sec:proofs}
We will first prove the asymptotic behavior of the roots of $Q(\lambda,i\xi)$ as $\xi\to0$ and as $\xii\to\infty$.
\subsection*{Proof of Lemmas~\ref{lem:low} and~\ref{lem:high}}
\begin{proof}[Proof of Lemma~\ref{lem:low}]
First of all, we notice that
\[ Q(\lambda,0)=\lambda^{m-\ell}\sum_{j=0}^{\ell} c_{m-j,0}\lambda^{\ell-j}, \]
so that~$\ell$ roots of $Q(\lambda,i\xi)$ solve~\eqref{eq:lambda0} at~$\xi=0$, whereas~$\lambda_j(0)=0$, for $j=1,\ldots,m-\ell$. We fix~$\xi'\in S^{n-1}$, and we set~$\xi=\rho\xi'$, with~$\rho>0$ and $\mu_j(\rho,\xi')=\rho^{-1}\lambda_j(\rho\xi')$. Since~$P_{m-\ell}$ is homogeneous,
\[ P_{m-\ell}(\lambda_j,i\xi) = \rho^{m-\ell} \, P_{m-\ell}(\mu_j,i\xi') = \rho^{m-\ell}\,c_{m-\ell,0}\,\prod_{k=1}^{m-\ell}(\mu_j-i\beta_k(\xi')). \]
Then we may write
\begin{equation}\label{eq:mainlow}
0 = \xii^{-(m-\ell)}\,Q(\lambda_j,i\xi)= P_{m-\ell}(\mu_j,i\xi') + \rho\,P_{m-\ell+1}(\mu_j,i\xi')+\textit{o}(\rho), \quad j=1,\ldots,m-\ell.
\end{equation}
In particular, as~$\rho\to0$, we obtain~$\mu_j = i\beta_j + \textit{o}(1)$, that is,
\[ \lambda_j(\xi)= i\xii \beta_j(\xi') + \textit{o}(\xii),\qquad j=1,\ldots,m-\ell.\]
On the other hand, since the roots $\beta_j(\xi')$ are distinct, due to the strict hyperbolicity of $P_{m-\ell}(\partial_t,\partial_x)$, the quantity
\[ \check{P}_{m-\ell,j}(\mu_j,i\xi') = c_{m-\ell,0} \prod_{k\neq j} (\mu_j-i\beta_k(\xi')), \]
is nonzero for sufficiently small $\xii$. Dividing~\eqref{eq:mainlow} by $\check P_{m-\ell,j}(\mu_j(\xi'),i\xi')$, we obtain
\[ \mu_j - i\beta_j(\xi') = -\rho\,\frac{P_{m-\ell+1}(\mu_j,i\xi')}{\check P_{m-\ell,j}(\mu_j,i\xi')} + \textit{o}(\rho) = \rho\,\frac{P_{m-\ell+1}(\beta_j(\xi'),\xi')}{\check P_{m-\ell,j}(\beta_j(\xi'),\xi')} + \textit{o}(\rho). \]
In the last equality we used that $P_{m-\ell+1}$ is homogeneous of degree $m-\ell+1$ and $\check{P}_{m-\ell,j}$ is homogeneous of degree $m-\ell-1$ to remove the imaginary unit $i$, with a sign change.
Multiplying by $\xii$ the above asymptotic behavior, we conclude the proof.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem:high}]
We fix~$\xi'\in S^{n-1}$, and we set~$\xi=\rho\xi'$, with~$\rho>0$. We also define $\mu_j(\rho,\xi')=\rho^{-1}\lambda_j(\xi)$ for $j=1,\ldots,m$. By the homogeneity of~$P_m(\lambda,\xi)$,
\[ P_m(\lambda_j,i\xi) = \rho^m \, P_m(\mu_j,i\xi') = \rho^m\,\prod_{k=1}^m(\mu_j-ia_k(\xi')). \]
Then we may write
\begin{equation}\label{eq:mainhigh}
0 = \xii^{-m}\,Q_2(\lambda_j,i\xi)= P_m(\mu_j,i\xi') + \rho^{-1}\,P_{m-1}(\mu_j,i\xi') + \textit{o}(\rho^{-1}).
\end{equation}
In particular, as~$\rho\to\infty$, we obtain~$\mu_j = ia_j(\xi') + \textit{o}(1)$, that is, $\lambda_j(\xi)= i\xii a_j(\xi') + \textit{o}(\xii)$. On the other hand, since the roots $a_j(\xi')$ are distinct, due to the strict hyperbolicity of $P_{m}$, the quantity
\[ \check{P}_{m,j}(\mu_j,i\xi') = \prod_{k\neq j} (\mu_j-ia_k(\xi')), \]
is nonzero for sufficiently large $\xii$. Dividing~\eqref{eq:mainhigh} by $\check P_{m,j}(\mu_j,i\xi')$, we obtain
\[ \mu_j - ia_j = -\rho^{-1}\,\frac{P_{m-1}(\mu_j,i\xi')}{\check P_{m,j}(\mu_j,i\xi')} + \textit{o}(\rho^{-1}) = -\rho^{-1}\,\frac{P_{m-1}(a_j,\xi')}{\check P_{m,j}(a_j(\xi'),\xi')} + \textit{o}(\rho^{-1}). \]
In the last equality we used that $P_{m-1}$ and $\check{P}_{m,j}$ are both homogeneous of degree $m-1$ to remove the imaginary unit $i$.
Multiplying by $\xii$ the above asymptotic behavior, we conclude the proof.
\end{proof}
\subsection*{Proof of Theorems~\ref{thm:CPQ2} and~\ref{thm:asymp}}
After performing the Fourier transform with respect to the space variable~$x$ in~\eqref{eq:CPlin}, $\hat u(t,\cdot)=\mathfrak{F}u(t,\cdot)$ is the solution to the ODE problem
\begin{equation}\label{eq:CPF}
\begin{cases}
Q(\partial_t,i\xi) \hat u =0, \quad t\geq 0,\ \xi\in\R^n,\\
\partial_t^j\hat u(0,x)=\hat u_j(\xi),\quad j=0,\ldots,m-1,
\end{cases}
\end{equation}
We may write an explicit representation of the solution~$\hat u$ to~\eqref{eq:CPF}, in terms of the complex-valued roots~$\lambda_1(\xi)$, \ldots, $\lambda_m(\xi)$, of the polynomial~$Q(\lambda,i\xi)$. We first consider $\xi\neq0$ such that the roots are all distinct. In this case,
\begin{equation}\label{eq:urepgen}
\hat u(t,\xi) = \sum_{j=1}^m \frac{\hat u_{m-1}-\hat u_{m-2}\sum_{k\neq j}\lambda_k
+ \ldots + (-1)^{m-1} \hat u_0 \prod_{k\neq j}\lambda_k}{\prod_{k\neq j}(\lambda_j-\lambda_k)}\,e^{\lambda_jt}.
\end{equation}
It is clear that it is impossible to compute, in general, the roots of a polynomial of order $m$ with complex-valued coefficients. For our purpose, it is sufficient to now that the real parts of the roots are negative, as a consequence of Lemmas~\ref{lem:stable1} and~\ref{lem:stable2}, together with the knowledge of the asymptotic profile of the roots as~$\xi\to0$ (Lemma~\ref{lem:low}) and~$\xii\to\infty$ (Lemma~\ref{lem:high}). This will provide us with pointwise estimates for
\begin{equation}\label{eq:urepgenest}
|\partial_t^k\hat u(t,\xi)| \leq \sum_{j=1}^m \frac{|\hat u_{m-1}|+|\hat u_{m-2}|\sum_{k\neq j}|\lambda_k|
+ \ldots + |\hat u_0| \prod_{k\neq j}|\lambda_k|}{\prod_{k\neq j}|\lambda_k-\lambda_j|}\,|\lambda_j|^k\,e^{\Re\lambda_jt},
\end{equation}
for any $\xi\neq0$.
\begin{remark}\label{rem:multiple}
If there is some multiple root, representation~\eqref{eq:urepgen} and estimate~\eqref{eq:urepgenest} are conveniently modified.
For instance, let $m=3$ and assume that $\lambda_1(\xi)\neq\lambda_2(\xi)=\lambda_3(\xi)$, for some~$\xi\neq0$. Then~\eqref{eq:urepgen} is replaced by
\begin{equation}\label{eq:urepdouble}\begin{split}
\hat u(t,\xi)
& = \frac{\hat u_2 - 2\lambda_3 \hat u_1 + \lambda_3^2\hat u_0}{(\lambda_3-\lambda_1)^2}\,e^{\lambda_1t} \\
& \qquad - \frac{\hat u_2-2\lambda_3 \hat u_1 +\lambda_1(2\lambda_3-\lambda_1)\hat u_0}{(\lambda_1-\lambda_3)^2}\,e^{\lambda_3t} \\
& \qquad - \frac{\hat u_2 - (\lambda_1+\lambda_3) \hat u_1+\lambda_1\lambda_3\hat u_0}{\lambda_1-\lambda_3}\, t\, e^{\lambda_3t}.
\end{split}\end{equation}
Formula~\eqref{eq:urepdouble} is the limit of~\eqref{eq:urepgen} as $\lambda_2\to\lambda_3$. Indeed, writing formula~\eqref{eq:urepgen} as
\begin{equation}\label{eq:urepdoublenear}\begin{split}
\hat u(t,\xi)
& = \frac{\hat u_2 - (\lambda_2+\lambda_3) \hat u_1 + \lambda_2\lambda_3\hat u_0}{(\lambda_2-\lambda_1)(\lambda_3-\lambda_1)}\,e^{\lambda_1t}\\
& \quad + \left(\frac{\hat u_2 - (\lambda_1+\lambda_3) \hat u_1 + \lambda_1\lambda_3\hat u_0}{(\lambda_1-\lambda_2)(\lambda_3-\lambda_2)}+\frac{\hat u_2 - (\lambda_1+\lambda_2) \hat u_1 + \lambda_1\lambda_2\hat u_0}{(\lambda_1-\lambda_3)(\lambda_2-\lambda_3)}\right)\,e^{\lambda_2t}\\
& \quad + \frac{\hat u_2 - (\lambda_1+\lambda_2) \hat u_1 + \lambda_1\lambda_2\hat u_0}{\lambda_1-\lambda_3}\,\frac{e^{\lambda_3t}-e^{\lambda_2t}}{\lambda_2-\lambda_3},
\end{split}\end{equation}
rewriting the second term as
\[ - \frac{\hat u_2 - (\lambda_2+\lambda_3) \hat u_1 + \lambda_1(\lambda_3+\lambda_2-\lambda_1)\hat u_0}{(\lambda_1-\lambda_2)(\lambda_1-\lambda_3)}\,e^{\lambda_2t}, \]
and taking the limit as $\lambda_2\to\lambda_3$, we find~\eqref{eq:urepdouble}. To treat such a case, estimate~\eqref{eq:urepgenest} may be replaced by
\begin{align*}
|\hat u(t,\xi)|
& \leq \frac{|\hat u_2|+|\lambda_2+\lambda_3|\,|\hat u_1| + |\lambda_2\lambda_3| |\hat u_0|}{|\lambda_2-\lambda_1|\,|\lambda_3-\lambda_1|}\,e^{\Re\lambda_1t}\\
& \qquad + \frac{|\hat u_2|+|\lambda_2+\lambda_3|\,|\hat u_1| + |\lambda_1|\,|\lambda_3+\lambda_2-\lambda_1|\,|\hat u_0|}{|\lambda_1-\lambda_2|\,|\lambda_1-\lambda_3|}\,e^{\Re\lambda_2t}\\
& \qquad + \frac{|\hat u_2|+|\lambda_1+\lambda_2|\,|\hat u_1| + |\lambda_1\lambda_2|\,|\hat u_0|}{|\lambda_1-\lambda_3|}\,t\,e^{\Re\lambda_3t}\,,
\end{align*}
and similarly for the time derivatives. For the sake of brevity, we omit a general formula for $m\geq3$ and multiple roots.
\end{remark}
We are now ready to prove Theorem~\ref{thm:CPQ2}.
\begin{proof}[Proof of Theorem~\ref{thm:CPQ2}]
The Cauchy problem~\eqref{eq:CPlin} is well-posed in $H^s$ since $P_m(\partial_t,\partial_x)$ is strictly hyperbolic, so we shall only prove the desired decay rate in~\eqref{eq:estQ1} and in~\eqref{eq:estQ2} as $t\to\infty$. Therefore, we assume $t\geq1$ in the following. We use Plancherel's theorem, Riemann-Lebesgue theorem and Hausdorff-Young inequality to estimate the $\dot H^s$ norm of $\partial_t^k u(t,\cdot)$, using pointwise estimates for $|\partial_t^k\hat u(t,\xi)|$.
Exception given for the case $Q=Q_2$ and $4c_{m-2,0}=c_{m-1,0}^2$, we may fix sufficiently small~$\delta>0$ and~$c>0$, such that the roots~$\lambda_j(\xi)$, described in Lemmas~\ref{lem:low} are distinct for~$0<\xii\leq\delta$, and, in view of \eqref{eq:strictlow}, for any~$\xii\leq\delta$, it holds:
\[ \Re\lambda_m(\xi)\leq -c, \qquad \Re\lambda_j(\xi)\leq -c\xii^2,\qquad j=1,\ldots,m-1, \]
if $Q=Q_1$, or
\[ \Re\lambda_{m-1,m}(\xi)\leq -c,\qquad \Re\lambda_j(\xi)\leq -c\xii^2, \qquad j=1,\ldots,m-2, \]
if $Q=Q_2$, for some $c>0$.
We first consider $Q=Q_1$. According to Lemma~\ref{lem:low}, we may estimate
\[ |\lambda_m(\xi)|\lesssim1, \qquad |\lambda_j(\xi)| \lesssim \xii, \qquad j=1,\ldots,m-1, \]
as well as
\begin{align*}
|\lambda_k(\xi)-\lambda_m(\xi)|
& = c_{m-1,0}+\textit{o}(1) \gtrsim 1, \qquad k=1,\ldots,m-1,\\
|\lambda_k(\xi)-\lambda_j(\xi)|
& =|b_k(\xi')-b_j(\xi')|\,\xii+\textit{o}(\xii)\gtrsim \xii,\qquad j,k=1,\ldots,m-1, \ k\neq j.
\end{align*}
As a consequence, we obtain
\begin{align*}
& \frac1{\prod_{k\neq m}|\lambda_k(\xi)-\lambda_m(\xi)|} \lesssim 1,\\
& \frac1{\prod_{k\neq j}|\lambda_k(\xi)-\lambda_j(\xi)|} \lesssim \xii^{-(m-2)}, \qquad j=1,\ldots,m-1,
\end{align*}
and
\begin{align*}
& \sum_{k\neq m}|\lambda_k(\xi)|=\sum_{k\neq m}|b_k(\xi)|+\textit{o}(\xii)\lesssim \xii, \qquad \ldots, \quad \prod_{k\neq m}|\lambda_k| \lesssim \xii^{m-1},\\
& \sum_{k\neq j}|\lambda_k|=c_{m-1,0}+\textit{o}(1)\lesssim 1, \qquad \ldots, \quad \prod_{k\neq j}|\lambda_k| \lesssim \xii^{m-2},\quad j=1,\ldots,m-1.
\end{align*}
Recalling~\eqref{eq:urepgen}, we get
\begin{align*}
|\partial_t^k\hat u(t,\xi)|
& \lesssim \xii^{k-(m-2)}\,\big(|\hat u_{m-1}(\xi)|+|\hat u_{m-2}(\xi)|\big)\,e^{-c\xii^2t} + \sum_{j=0}^{m-3} \xii^{k-j}\,|\hat u_j(\xi)|\,e^{-c\xii^2t} \\
& \qquad + \sum_{j=0}^{m-1} \xii^{m-1-j}\,|\hat u_j(\xi)|\,\,e^{-ct},
\end{align*}
for~$k\geq0$.
If~$q=2$, we immediately obtain
\begin{align*}
\|\xii^s\partial_t^k \hat u(t,\cdot)\|_{L^2(\xii\leq\delta)}
& \lesssim \big(\|u_{m-1}\|_{L^2}+\|u_{m-2}\|_{L^2}\big)\,\sup_{\xii\leq\delta} \xii^{k+s-(m-2)}\,e^{-c\xii^2t} \\
& \qquad + \sum_{j=0}^{m-3} \|u_j\|_{L^2}\,\sup_{\xii\leq\delta} \xii^{k+s-j}\,e^{-c\xii^2t} + e^{-ct}\,\sum_{j=0}^{m-1} \|u_j\|_{L^2}\\
& \lesssim t^{-\frac{k+s-(m-2)}2}\,\big(\|u_{m-1}\|_{L^2}+\|u_{m-2}\|_{L^2}\big)+ \sum_{j=0}^{m-3} t^{-\frac{k+s-j}2} \|u_j\|_{L^2},
\end{align*}
provided that $k+s\geq m-2$.
Let $q\in[1,2)$ and set $q'=q/(q-1)\in(2,\infty]$ and $r=1/2-1/q'=1/q-1/2$. By the change of variable~$\eta=\sqrt{t}\xi$ and by H\"older inequality, we easily get
\begin{align*}
\|\xii^s\partial_t^k\hat u(t,\cdot)\|_{L^2(\xii\leq\delta)}
& \lesssim \Big( \int_{\xii\leq\delta} \xii^{r(k+s-(m-2))}\,e^{-rc\xii^2t}\,d\xi\Big)^{\frac1r} \,\big(\|\hat u_{m-1}\|_{L^{q'}}+\|\hat u_{m-2}\|_{L^{q'}}\big) \\
& \qquad + \sum_{j=0}^{m-3} \Big( \int_{\xii\leq\delta} \xii^{r(k+s-j)}\,e^{-rc\xii^2t}\,d\xi\Big)^{\frac1r} \,\|\hat u_j\|_{L^{q'}} + e^{-ct}\,\sum_{j=0}^{m-1} \|u_j\|_{L^2} \\
& \lesssim t^{-\frac{n}r-(k+s-(m-2))}\,\big(\|u_{m-1}\|_{L^q}+\|u_{m-2}\|_{L^q}\big)+\sum_{j=0}^{m-3}t^{-\frac{n}r-(k+s-j)}\,\|u_j\|_{L^q},
\end{align*}
provided that~$n/r+k+s>m-2$, so that the power~$\xii^{r(k+s-(m-2))}$ is integrable near~$\xi=0$.
We now consider $Q=Q_2$. Let~$j_0=m-1,m$ in the following, for brevity. According to Lemma~\ref{lem:low}, we may estimate
\[ |\lambda_{j_0}(\xi)|\lesssim1, \qquad |\lambda_j(\xi)| \lesssim \xii, \qquad j=1,\ldots,m-2, \]
as well as
\[ |\lambda_k(\xi)-\lambda_{j_0}(\xi)| \gtrsim 1, \qquad |\lambda_k(\xi)-\lambda_j(\xi)|\gtrsim \xii,\qquad j,k=1,\ldots,m-2, \ k\neq j. \]
Moreover, $|\lambda_m(\xi)-\lambda_{m-1}(\xi)| \gtrsim 1$, since we assumed $4c_{m-2,0}\neq c_{m-1,0}^2$. As a consequence, we obtain
\begin{align*}
& \frac1{\prod_{k\neq j_0}|\lambda_k-\lambda_{j_0}|} \lesssim 1, \\
& \frac1{\prod_{k\neq j}|\lambda_k-\lambda_j|} \lesssim \xii^{-(m-3)}, \qquad j=1,\ldots,m-2,
\end{align*}
and
\begin{align*}
& \sum_{k\neq j_0}|\lambda_k|\lesssim 1, \quad \sum_{k,\ell\neq j_0, k<\ell}|\lambda_k\lambda_\ell|\lesssim \xii, \quad \ldots,\quad \prod_{k\neq j_0}|\lambda_k| \lesssim \xii^{m-2},\\
& \sum_{k\neq j}|\lambda_k|\lesssim 1, \quad \sum_{k,\ell\neq j, k<\ell}|\lambda_k\lambda_\ell|\lesssim 1, \quad \ldots, \quad \prod_{k\neq j}|\lambda_k| \lesssim \xii^{m-3},\quad j=1,\ldots,m-2.
\end{align*}
Recalling~\eqref{eq:urepgen}, we get
\begin{align*}
|\partial_t^k\hat u(t,\xi)|
& \lesssim \xii^{k-(m-3)}\,\big(|\hat u_{m-1}(\xi)|+|\hat u_{m-2}(\xi)|+|\hat u_{m-3}(\xi)|\big)\,e^{-c\xii^2t} + \sum_{j=0}^{m-4} \xii^{k-j}\,|\hat u_j(\xi)|\,e^{-c\xii^2t} \\
& \qquad + |\hat u_{m-1}(\xi)|\,\,e^{-ct} + \sum_{j=0}^{m-2} \xii^{m-2-j}\,|\hat u_j(\xi)|\,\,e^{-ct},
\end{align*}
for~$k\geq0$.
We proceed as we did for $Q_1$. If~$q=2$, we immediately obtain
\begin{align*}
\|\xii^s\partial_t^k \hat u(t,\cdot)\|_{L^2(\xii\leq\delta)}
& \lesssim \big(\|u_{m-1}\|_{L^2}+\|u_{m-2}\|_{L^2}+\|u_{m-3}\|_{L^2}\big)\,\sup_{\xii\leq\delta} \xii^{k+s-(m-3)}\,e^{-c\xii^2t} \\
& \qquad + \sum_{j=0}^{m-4} \|u_j\|_{L^2}\,\sup_{\xii\leq\delta} \xii^{k+s-j}\,e^{-c\xii^2t} + e^{-ct}\,\sum_{j=0}^{m-1} \|u_j\|_{L^2}\\
& \lesssim t^{-\frac{k+s-(m-3)}2}\,\big(\|u_{m-1}\|_{L^2}+\|u_{m-2}\|_{L^2}+\|u_{m-3}\|_{L^2}\big)+ \sum_{j=0}^{m-4} t^{-\frac{k+s-j}2} \|u_j\|_{L^2},
\end{align*}
provided that $k+s\geq m-3$. Let $q\in[1,2)$ and set $q'=q/(q-1)\in(2,\infty]$ and $r=1/2-1/q'=1/q-1/2$. By the change of variable~$\eta=\sqrt{t}\xi$ and by H\"older inequality, we may easily get
\begin{align*}
\|\xii^s\partial_t^k\hat u(t,\cdot)\|_{L^2(\xii\leq\delta)}
& \lesssim \Big( \int_{\xii\leq\delta} \xii^{r(k+s-(m-3))}\,e^{-rc\xii^2t}\,d\xi\Big)^{\frac1r} \,\big(\|\hat u_{m-1}\|_{L^{q'}}+\|\hat u_{m-2}\|_{L^{q'}}+\|\hat u_{m-3}\|_{L^{q'}}\big) \\
& \qquad + \sum_{j=0}^{m-4} \Big( \int_{\xii\leq\delta} \xii^{r(k+s-j)}\,e^{-rc\xii^2t}\,d\xi\Big)^{\frac1r} \,\|\hat u_j\|_{L^{q'}} + e^{-ct}\,\sum_{j=0}^{m-1} \|u_j\|_{L^2} \\
& \lesssim t^{-\frac{n}r-(k+s-(m-3))}\,\big(\|u_{m-1}\|_{L^q}+\|u_{m-2}\|_{L^q}+\|u_{m-3}\|_{L^q}\big)+\sum_{j=0}^{m-4}t^{-\frac{n}r-(k+s-j)}\,\|u_j\|_{L^q},
\end{align*}
provided that~$n/r+k+s>m-3$, so that the power~$\xii^{r(k+s-(m-3))}$ is integrable near~$\xi=0$.
\bigskip
We now consider the high frequencies. Since $P_m(\partial_t,\partial_x)$ is strictly hyperbolic, there exist~$M>0$ and~$c>0$ such that the roots~$\lambda_j$ in Lemma~\ref{lem:high} are distinct for any~$\xii\geq M$, and, for any~$\xii\geq M$, it holds~$\Re\lambda_j(\xi)\leq -c$ for some $c>0$. Moreover, we may estimate
\[ |\lambda_j(\xi)| \lesssim \xii,\qquad |\lambda_j(\xi)-\lambda_\ell(\xi)|=|a_j(\xi')-a_\ell(\xi')|\,\xii+\textit{o}(\xii) \gtrsim \xii,\quad j\neq \ell. \]
Recalling~\eqref{eq:urepgen}, we may now estimate
\[ |\partial_t^k\hat u(t,\xi)| \lesssim \sum_{j=0}^{m-1} \xii^{k-j} |\hat u_j(\xi)|\,e^{-ct}, \]
for~$k\geq0$. We immediately obtain that
\[ \Big(\int_{\xii\geq M} \xii^{2s} |\partial_t^k\hat u(t,\xi)|^2\,d\xi\Big)^{\frac12} \lesssim e^{-ct}\,\sum_{j=0}^{m-1} \|u_j\|_{H^{s+k-j}}. \]
At intermediate frequencies, $\delta\leq \xii\leq M$, we shall not worry about regularity of the initial data or about the decay rate, since~$\Re\lambda(\xi)\leq -c$, for some~$c>0$, uniformly in a compact subset of~$\R^n\setminus\{0\}$. The only difference with the previous analysis is that the roots~$\lambda_j$ may coincide for some~$\xi$. However, this possibility has no influence on the exponential decay rate. Indeed, following as in Remark~\ref{rem:multiple}, if at most $p$ roots coincide at some point $\xi\neq0$, we get the estimate
\[ |\partial_t^k \hat u(t,\xi)| \lesssim (1+t)^{p-1}\,e^{-ct}\sum_{j=0}^{m-1}|\hat u_j(\xi)|, \]
but it is clear that~$(1+t)^{m-1}\,e^{-ct} \leq C e^{-c't}$, for any fixed~$c'\in(0,c)$, for some $C>0$, so that the decay remains exponential even if multiple roots are present.
Summarizing, we may unify the estimate at high and intermediate frequencies, getting
\[ \Big(\int_{\xii\geq \delta} \xii^{2s} |\partial_t^k\hat u(t,\xi)|^2\,d\xi\Big)^{\frac12} \lesssim e^{-c't}\,\sum_{j=0}^{m-1} \|u_j\|_{H^{s+k-j}}. \]
and this concludes the proof of~\eqref{eq:estQ1} and~\eqref{eq:estQ2}.
We now go back to the case that we excluded, $4c_{m-2,0}=c_{m-1,0}^2$. In this case, we do not rely on~\eqref{eq:urepgen} for $\lambda_{m-1}$ and $\lambda_m$ near $\xi=0$, since $\lambda_m(0)=\lambda_{m-1}(0)=-c_{m-1,0}/2$. However, taking into account of Remark~\ref{rem:multiple}, the only difference is that $e^{-ct}$ is replaced by $(1+t)\,e^{-ct}$ in the estimate of $\hat u(t,\xi)$ related to these two roots, but as seen before, this is not a big deal, since $(1+t)\,e^{-ct} \leq C e^{-c't}$, for any fixed~$c'\in(0,c)$, for some $C>0$. This concludes the proof.
\end{proof}
With minor modifications in the first part of the proof of Theorem~\ref{thm:CPQ2}, we may prove Theorem~\ref{thm:asymp}.
\begin{proof}[Proof of Theorem~\ref{thm:asymp}]
We prove Theorem~\ref{thm:asymp} for $Q=Q_1$, being the proof for $Q=Q_2$ completely analogous.
In view of Theorem~\ref{thm:CPQ2}, we may assume with no loss of generality that $u_0=\ldots=u_{m-3}=0$. We claim that if
\[ u_{m-1},u_{m-2}\in L^{1,1}=L^1((1+|x|) dx), \]
then for sufficiently small $\delta>0$ and for any $t\geq1$, we have the improved estimate:
\begin{equation}\label{claim:11}
\big\|\xii^s\partial_t^k\big(\hat u(t,\xi)-\xii^{-(m-2)}\hat v(t,\xi)\big)\big\|_{L^2(\xii\leq\delta)} \lesssim t^{-\frac{n}4-(k+s-(m-2))-\frac12}\,\big(\|u_{m-1}\|_{L^{1,1}}+\|u_{m-2}\|_{L^{1,1}}\big),
\end{equation}
where an additional power $t^{-\frac12}$ appears, with respect to the decay rate in~\eqref{eq:estQ1}, provided that~$n/2+1+k+s>m-2$. By claim~\eqref{claim:11}, the proof of Theorem~\ref{thm:asymp} follows by a density argument, as in~\cite[Lemma 3.2]{K00}. Namely, for given $u_{m-1},u_{m-2}\in L^1$, for any $\varepsilon$, we choose $u_{m-1}^\varepsilon,u_{m-2}^\varepsilon\in L^{1,1}$ such that
\[ \|u_{m-1}-u_{m-1}^\varepsilon\|_{L^1}+\|u_{m-2}-u_{m-2}^\varepsilon\|_{L^1}\leq\varepsilon.\]
By combining Theorem~\ref{thm:CPQ2} with initial data $u_j-u_j^\varepsilon$ and claim~\eqref{claim:11}, for any sufficiently large $T_\varepsilon$, we derive
\begin{align*}
& t^{\frac{n}4+(k+s-(m-2))}\,\big\|\xii^s\partial_t^k\big(\hat u(t,\xi)-\xii^{-(m-2)}\hat v(t,\xi)\big)\big\|_{L^2(\xii\leq\delta)}\\
& \qquad \lesssim t^{-\frac12}\,\big(\|u_{m-1}^\varepsilon\|_{L^{1,1}}+\|u_{m-2}^\varepsilon\|_{L^{1,1}}\big) + \varepsilon\leq 2\varepsilon, \qquad t\geq T_\varepsilon;
\end{align*}
hence, the proof of Theorem~\ref{thm:asymp} follows. We now prove claim~\eqref{claim:11}. If $f\in L^{1,1}$, then $\hat f\in \mathcal C^1$; in particular, $|\hat f(\xi)-\hat f(0)|\lesssim\xii\,\|\hat f\|_{L^{1,1}}$, by Lagrange theorem. Due to
\[ M= \hat u_{m-1}(0) + c_{m-1,0}\hat u_{m-2}(0), \]
for any $j=1,\ldots,m-1$, we may estimate
\[ \big| \hat u_{m-1}(\xi) - \hat u_{m-2}(\xi)\sum_{k\neq j}\lambda_k(\xi) -M\big| \lesssim \xii\,\big(\|u_{m-1}\|_{L^{1,1}}+\|u_{m-2}\|_{L^{1,1}}\big), \]
as $\xi\to0$. Indeed, $|\lambda_k(\xi)|\lesssim \xii$ for $k\neq m$ and $\lambda_m(0)=-c_{m-1,0}$. On the other hand, for any $j=1,\ldots,m-1$, we may estimate
\begin{align*}
& \Big| e^{\lambda_jt}-e^{i\xii\,b_j(\xi')t+\xii^2\,\frac{P_m(b_j(\xi'),\xi')}{\check{P}_{m-1,j}(b_j(\xi'),\xi')}t}\Big| \lesssim t\xii^3\,e^{-c\xii^2t}, \\
& \left| \frac{1}{(i\xii)^{m-2}\check{P}_{m-1,j}(b_j(\xi'),\xi')}-\frac1{\prod_{k\neq j}(\lambda_j(\xi)-\lambda_k(\xi))}\right| \\
& \qquad = \left| \frac1{(i\xii)^{m-2}c_{m-1,0}\prod_{k\neq j}(b_j(\xi')-b_k(\xi'))}-\frac1{\prod_{k\neq j}(\lambda_j(\xi)-\lambda_k(\xi))}\right| \lesssim \xii^{1-(m-2)},
\end{align*}
where we replaced
\[ \check{P}_{m-1,j}(b_j(\xi'),\xi') = c_{m-1,0}\,\prod_{k\neq j}(b_j(\xi')-b_k(\xi')), \]
and we used Lemma~\ref{lem:low} to estimate
\begin{align*}
& |\lambda_j(\xi)-\lambda_m(\xi)-c_{m-1,0}|\lesssim \xii,\\
& |\lambda_j(\xi)-\lambda_k(\xi)-i\xii (b_j(\xi')-b_k(\xi'))|\lesssim \xii^2, \qquad k=1,\ldots,m-2, \ k\neq j.
\end{align*}
Therefore, for $\xii\leq\delta$, we get the improved estimate
\[ \big|\partial_t^k \big(\hat u(t,\xi)-\xii^{-(m-2)}\hat v(t,\xi)\big) \big|\lesssim \xii^{1+k-(m-2)}\,(1+t\xii^2)\,e^{-c\xii^2t}\,\big(\|u_{m-1}\|_{L^{1,1}}+\|u_{m-2}\|_{L^{1,1}}\big), \]
from which~\eqref{claim:11} follows, as in the proof of Theorem~\ref{thm:CPQ2}. We mention that when $Q=Q_2$,
\begin{align*}
& \left| \frac{1}{(i\xii)^{m-3}\check{P}_{m-2,j}(d_j(\xi'),\xi')}-\frac1{\prod_{k\neq j}(\lambda_j(\xi)-\lambda_k(\xi))}\right| \\
& \qquad = \left| \frac1{(i\xii)^{m-3}c_{m-2,0}\prod_{k\neq j}(d_j(\xi')-d_k(\xi'))}-\frac1{\prod_{k\neq j}(\lambda_j(\xi)-\lambda_k(\xi))}\right| \lesssim \xii^{1-(m-3)},
\end{align*}
for $j=1,\ldots,m-2$, where we estimated
\[ |(\lambda_j-\lambda_m)(\lambda_j-\lambda_{m-1})-c_{m-2,0}| \lesssim \xii. \]
\end{proof}
\begin{remark}\label{rem:M0}
The proof of Theorem~\ref{thm:asymp} shows that, when $M=0$, estimates~\eqref{eq:estQ1} and~\eqref{eq:estQ2} for $q=1$ may be improved, assuming some initial data in $L^{1,1}$. In particular, \eqref{eq:estQ1} is improved to
\begin{equation}
\label{eq:estQ1M0}\begin{split}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s}
& \leq C\,(1+t)^{-\frac{n}4-\frac{k+s-(m-3)}2}\,\sum_{j=m-2,m-1}\big(\|u_j\|_{L^{1,1}}+\|u_j\|_{H^{k+s-j}}\big) \\
& \qquad + C\,\sum_{j=0}^{m-3}(1+t)^{-\frac{n}4-\frac{k+s-j}2}\,\big(\|u_j\|_{L^1}+\|u_j\|_{H^{k+s-j}}\big),
\end{split}\end{equation}
for any $k+s\leq s_0$, such that $n/2+k+s>m-3$, and \eqref{eq:estQ2} is improved to
\begin{equation}
\label{eq:estQ2M0}\begin{split}
\|\partial_t^k u(t,\cdot)\|_{\dot H^s}
& \leq C\,(1+t)^{-\frac{n}4-\frac{k+s-(m-4)}2}\,\sum_{j=m-3}^{m-1}\big(\|u_j\|_{L^{1,1}}+\|u_j\|_{H^{k+s-j}}\big)\\
& \qquad + C\,\sum_{j=0}^{m-4}(1+t)^{-\frac{n}4-\frac{k+s-j}2}\,\big(\|u_j\|_{L^1}+\|u_j\|_{H^{k+s-j}}\big),
\end{split}
\end{equation}
for any $k+s\leq s_0$, such that $n/2+k+s>m-4$.
\end{remark}
\begin{remark}\label{rem:genQ}
In this paper, the choice to consider one or two lower order homogeneous hyperbolic operators is due to the possibility to use Lemmas~\ref{lem:stable1} and~\ref{lem:stable2} to express a condition equivalent to the strict stability of $Q(\lambda,i\xi)$. If one already knows that a more general operator $Q$ as in~\eqref{eq:Q} verifies the property that $Q(\lambda,i\xi)$ is strictly stable (as in the next Example~\ref{ex:3}), Lemmas~\ref{lem:low} and Lemma~\ref{lem:high} may be applied to obtain a result analogous to Theorems~\ref{thm:CPQ2} and~\ref{thm:asymp} for the solution to~\eqref{eq:CPlin}. For instance, the strict stability of $Q(\lambda,i\xi)$ could be checked by using Routh-Hurwitz criteria if all coefficients of $Q(\lambda,i\xi)$ are real.
\end{remark}
\begin{example}\label{ex:3}
Let us define
\[ Q(\partial_t,\partial_x)=\sum_{j=0}^3 P_{m-j}(\partial_t,\partial_x). \]
By Hermite--Biehler theorem, $Q(\lambda,i\xi)$ is strictly stable if, and only if, $P_{m-1}(\lambda,\xi)-P_{m-3}(\lambda,\xi)$ and $P_m(\lambda,\xi)-P_{m-2}(\lambda,\xi)$ strictly interlace, for any~$\xi\neq0$. Letting $\xi\to0$, we find the necessary condition that $c_{m-3,0}\leq c_{m-2,0}\,c_{m-1,0}$, which correspond to the weak stability (i.e., the real part of the roots is nonpositive) of
\[ Q(\lambda,0)=\lambda^{m-3}(\lambda^3+c_{m-1,0}\lambda^2+c_{m-2,0}\lambda+c_{m-3,0}).\]
We stress that the strict stability of $Q(\lambda,i\xi)$ is no longer only dependent on the roots of the homogeneous polynomials $P_{m-j}(\lambda,\xi')$, but it also involves the multiplicative coefficients. This is due to the fact that the polynomial has more than two lower order terms.
We provide an easy example, for which we may directly check the strict stability of $Q(\lambda,i\xi)$. We fix:
\begin{align*}
P_4(\partial_t,\partial_x)
& = \partial_t^2 (\partial_t^2-a^2\Delta),\\
P_3(\partial_t,\partial_x)
& = c_{3}\,\partial_t (\partial_t^2-a^2\Delta),\\
P_2(\partial_t,\partial_x)
& = c_{2}\,(\partial_t^2-b^2\Delta),\\
P_1(\partial_t,\partial_x)
& = c_{1}\,\partial_t,
\end{align*}
where $c_1,c_2,c_3,a,b>0$. Due to
\begin{align*}
P_4(\lambda,\xi)-P_2(\lambda,\xi)
& = \lambda^4 - (c_2+a^2\xii^2) \lambda^2 + c_2b^2\xii^2, \\
P_3(\lambda,\xi)-P_1(\lambda,\xi)
& = c_3 \lambda^3 - (c_1+c_3a^2\xii^2)\lambda,
\end{align*}
by straightforward computation, $P_3(\lambda,\xi)-P_1(\lambda,\xi)$ and $P_4(\lambda,\xi)-P_2(\lambda,\xi)$ strictly interlace for any~$\xi\neq0$ if, and only if,
$c_1<c_2c_3$ and
\[ b^2 < \left(1-\frac{c_1}{c_2c_3}\right)\,a^2. \]
Applying Lemma~\ref{lem:low}, we find that $\lambda_2(0)$, $\lambda_3(0)$, $\lambda_4(0)$ are the solutions to $\lambda^3+c_3\lambda^2+c_2\lambda+c_1=0$, whose real parts are negative, whereas
\[ \lambda_1(\xi)= - \frac{c_2\,b^2}{c_1}\,\xii^2 + \textit{o}(\xii^2), \]
as~$\xi\to0$. It is not difficult to extend Theorem~\ref{thm:asymp} and show that the asymptotic profile of the solution to~\eqref{eq:CPlin}, with~$Q$ as in this example, is described by~$M/c_1$ times the fundamental solution to the heat equation
\[ c_1 v_t -c_2\,b^2\,\Delta v=0,\]
where
\[ M=\int_{\R^n} \big(u_3(x)+c_3u_2(x)+c_2u_1(x)+c_1u_0(x)\big)\,dx. \]
More precisely, the estimate
\[ \big\|\partial_t^k \big( u(t,\cdot) - v(t,\cdot)\big)\big\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}4-\frac{s}2-k}\big), \]
holds.
\end{example}
\section{The influence of weak interlacing polynomials}\label{sec:weak}
In the first part of this section, we discuss what happens if the assumption of strict interlacing of the polynomials $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$, as in~\eqref{eq:interlace}, is weakened to non strict interlacing, as in~\eqref{eq:weakinterlace}, for some~$\xi'\in S^{n-1}$. In the second part of this section, we carry on the same analysis for the interlacing of the polynomials $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$. We use the same notation introduced in \textsection~\ref{sec:asymp}.
We stress that dropping the strict interlacing of the polynomials, we may also consider the case in which $P_m(\partial_t,\partial_x)$ and $P_{m-2}(\partial_t,\partial_x)$ are weakly hyperbolic. However, we cannot drop the assumption of strict hyperbolicity of $P_{m-1}(\partial_t,\partial_x)$, otherwise we lose the strict stability of $Q_2(\lambda,i\xi)$, see Lemma~\ref{lem:stable2}. More precisely, $P_m(\lambda,\xi')$ and $P_{m-2}(\lambda,\xi')$ may admit double roots for some $\xi'\in S^{n-1}$, since roots with multiplicity larger than two are prevented by the weak interlacing assumption with the strictly hyperbolic polynomial~$P_{m-1}(\lambda,\xi')$. However, a double root of $P_{m}(\lambda,\xi')$ or of $P_{m-2}(\lambda,\xi')$ is also a (simple) root of $P_{m-1}(\lambda,\xi')$, as a consequence of the interlacing condition:
\begin{align*}
a_j(\xi')=a_{j+1}(\xi') & \Rightarrow a_j(\xi')=b_j(\xi')=a_{j+1}(\xi'),\\
d_j(\xi')=d_{j+1}(\xi') & \Rightarrow d_j(\xi')=b_{j+1}(\xi')=d_{j+1}(\xi').
\end{align*}
If we allow the non strict interlacing of the polynomials $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$ for some $\xi'\in S^{n-1}$, then Lemma~\ref{lem:low} for $Q=Q_2$ is generalized as follows.
\begin{lemma}\label{lem:low2weak}
Let~$P_{m-2}(\partial_t,\partial_x)$ and $P_m(\partial_t,\partial_x)$ be (possibly weakly) hyperbolic and $P_{m-1}(\partial_t,\partial_x)$ be strictly hyperbolic. Assume that $P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$ interlace and that $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ interlace, for any $\xi'\in S^{n-1}$, and that there is no $(\lambda,\xi')\in\R\times S^{n-1}$ such that $P_{m-2}(\lambda,\xi')=P_{m-1}(\lambda,\xi')=P_m(\lambda,\xi')$. Let~$d_j(\xi')$ be the roots of $P_{m-2}(\lambda,\xi')$, $\xi'\in S^{n-1}$.
Then we may label the $m$ roots~$\lambda_j(\xi)$, $\xi\in\R^n$, of~$Q_2(\lambda,i\xi)$, in such a way that $\lambda_{m-1}$ and $\lambda_m$ are as in~\eqref{eq:lambdam-1m}, and $\lambda_j(\xi)$, $j=1,\ldots,m-2$, are described as follows, where we put $\xi'=\xi/\xii$.
\begin{enumerate}[(i)]
\item If $P_{m-1}(d_j(\xi'),\xi')\neq0$ and $\check{P}_{m-2,j}(d_j(\xi'),\xi')\neq0$, that is, $d_j(\xi')$ is a simple root of $P_{m-2}(\lambda,\xi')$ and it is not a root of $P_{m-1}(\lambda,\xi')$, then $\lambda_j(\xi)$ is as in~\eqref{eq:lambdalow2};
\item\label{en:lows} If $P_{m-1}(d_j(\xi'),\xi')=0$ and $\check{P}_{m-2,j}(d_j(\xi'),\xi')\neq0$, that is, $d_j(\xi')$ is a simple root of both $P_{m-1}(\lambda,\xi')$ and $P_{m-2}(\lambda,\xi')$, then
\begin{equation}\label{eq:asymplows}
\lambda_j(\xi) = i\xii d_j(\xi') + i\,\xii^3\,\frac{P_m(d_j(\xi'),\xi')}{\check{P}_{m-2,j}(d_j(\xi'),\xi')} + \xii^4\,\frac{P_m(d_j(\xi'),\xi')\,\tilde P_{m-1}(d_j(\xi'),\xi')}{\big(\check{P}_{m-2,j}(d_j(\xi'),\xi')\big)^2} + \textit{o}(\xii^4),
\end{equation}
as $\xi\to0$, where
\begin{equation}\label{eq:Ptilde}
\tilde P_{m-1}(\lambda,\xi')=\begin{cases}
\check{P}_{m-1,j}(\lambda,\xi') & \text{if $d_j(\xi')=b_j(\xi')$,}\\
\check{P}_{m-1,j+1}(\lambda,\xi') & \text{if $d_j(\xi')=b_{j+1}(\xi')$.}
\end{cases}
\end{equation}
\item\label{en:lowd} If $d_j(\xi')=b_{j+1}(\xi')=d_{j+1}(\xi')$, that is, $d_j(\xi')$ is a simple root of $P_{m-1}(\lambda,\xi')$ and a double root of $P_{m-2}(\lambda,\xi')$, then
\begin{equation}\label{eq:asymplowd}
\lambda_{j,j+1}(\xi) = i\xii d_j(\xi') + \xii^2\,\kappa_\pm(\xi') + \textit{o}(\xii^2),
\end{equation}
as $\xi\to0$, where $\kappa_\pm(\xi')$ are the two solutions to
\begin{equation}\label{eq:kappalow}
\kappa^2\check{P}_{m-2,j,j+1}(d_j(\xi'),\xi') - \kappa\,\check{P}_{m-1,j+1}(d_j(\xi'),\xi') + P_m(d_j(\xi'),\xi').
\end{equation}
\end{enumerate}
\end{lemma}
\begin{remark}
We stress that $P_m(d_j(\xi'),\xi')\,\tilde P_{m-1}(d_j(\xi'),\xi')\neq0$ in~\eqref{eq:asymplows}, since $d_j(\xi')$ is a simple root of $P_{m-1}(\lambda,\xi')$ and $P_{m-2}(\lambda,\xi')$, therefore, it cannot be also a root of $P_m(\lambda,\xi')$. Due to the interlacing condition,
\[ \frac{P_m(d_j(\xi'),\xi')\,\tilde P_{m-1}(d_j(\xi'),\xi')}{\big(\check{P}_{m-2,j}(d_j(\xi'),\xi')\big)^2}<0. \]
We stress that $\check{P}_{m-2,j,j+1}(d_j(\xi'),\xi')\neq0$ and $\check{P}_{m-1,j+1}(d_j(\xi'),\xi')\neq0$ in~\eqref{eq:asymplowd}, since $d_j(\xi')$ is a simple root of $P_{m-1}(\lambda,\xi')$ and a double root of $P_{m-2}(\lambda,\xi')$, therefore, it cannot be also a root of $P_m(\lambda,\xi')$. Moreover,
\[ \frac{\check{P}_{m-1,j+1}(d_j(\xi'),\xi')}{2\check{P}_{m-2,j,j+1}(d_j(\xi'),\xi')}<0, \qquad \check{P}_{m-2,j,j+1}(d_j(\xi'),\xi')\,P_m(d_j(\xi'),\xi')>0, \]
as a consequence of the interlacing condition; in particular, it follows that $\Re\kappa_\pm(\xi')<0$.
\end{remark}
\begin{proof}[Proof of Lemma~\ref{lem:low2weak}]
We follow the proof of Lemma~\ref{lem:low}, but we then take into account of the two new scenarios, \eqref{en:lows} and~\eqref{en:lowd}. In the both cases, $d_j(\xi')$ is a root of $P_{m-1}(\lambda,\xi')$. Since this latter is strictly hyperbolic, $\tilde{P}_{m-1}(d_j(\xi'),\xi')\neq0$. We first consider case~\eqref{en:lows}.
In this case, by
\begin{equation}\label{eq:mainlow2}
0 = \xii^{-(m-2)}\,Q_2(\lambda_j,i\xi)= P_{m-2}(\mu_j,i\xi') + \rho\,P_{m-1}(\mu_j,i\xi') + \rho^2\,P_m(\mu_j,i\xi'),
\end{equation}
we obtain:
\begin{align*}
0
& = P_{m-2}(\mu_j,i\xi') + \rho\,P_{m-1}(\mu_j,i\xi') + \rho^2\,P_m(\mu_j,i\xi') \\
& = (\mu_j-id_j(\xi')) \, \check{P}_{m-2,j}(\mu_j,i\xi') + \rho\,(\mu_j-id_j(\xi'))\,\tilde P_{m-1}(\mu_j,i\xi') + \rho^2\,P_m(\mu_j,i\xi')\\
& = (\mu_j-id_j(\xi')) \, \big( \check{P}_{m-2,j}(\mu_j,i\xi') + \rho\,\tilde P_{m-1}(\mu_j,i\xi')\big) + \rho^2\,P_m(\mu_j,i\xi').
\end{align*}
For sufficiently small $\rho$, $\check{P}_{m-2,j}(\mu_j,i\xi') + \rho\,\tilde P_{m-1}(\mu_j,i\xi')\neq0$, since $\check{P}_{m-2,j}(d_j(\xi'),\xi')\neq0$, therefore
\[ \mu_j-id_j(\xi') = - \frac{\rho^2\,P_m(\mu_j,i\xi')}{\check{P}_{m-2,j}(\mu_j,i\xi') + \rho\,\tilde P_{m-1}(\mu_j,i\xi')}. \]
The second-order expansion of the last expression gives:
\[ \mu_j - id_j(\xi') = i\,\rho^2\,\frac{P_m(d_j(\xi'),\xi')}{\check{P}_{m-2,j}(d_j(\xi'),\xi')} + \textit{o}(\rho^2), \]
where we used that $P_m$ is homogeneous of degree~$m$ and~$\check{P}_{m-2,j}$ is homogeneous of degree~$m-3$ to remove the imaginary unit, with a multiplication by $-i$. Since we are looking for the real part of $\mu_j$ we need one more expansion step. Writing
\[ -\frac{\rho^2\,P_m(\mu_j,i\xi')}{\check{P}_{m-2,j}(\mu_j,i\xi') + \rho\,\tilde P_{m-1}(\mu_j,i\xi')} = -\frac{\rho^2\,P_m(\mu_j,i\xi')\,\big(\check{P}_{m-2,j}(\mu_j,i\xi') - \rho\,\tilde P_{m-1}(\mu_j,i\xi')\big)}{\big(\check{P}_{m-2,j}(\mu_j,i\xi')\big)^2 - \rho^2\,\big(\tilde P_{m-1}(\mu_j,i\xi')\big)^2} \,,\]
and using the previously obtained information that $\mu_j=id_j(\xi')+\textit{O}(\rho^2)$, we easily obtain the further expansion:
\begin{align*}
\mu_j - id_j(\xi')
& = i\,\rho^2\,\frac{P_m(d_j(\xi'),\xi')}{\check{P}_{m-2,j}(d_j,\xi')} + \rho^3\,\frac{P_m(d_j(\xi'),\xi')\,\tilde P_{m-1}(d_j(\xi'),\xi')}{\big(\check{P}_{m-2,j}(d_j(\xi'),\xi')\big)^2} + \textit{o}(\rho^3),
\end{align*}
where in the last computation we used that $P_m\,\tilde P_{m-1}$ is homogeneous of degree $2m-2$ and $(\check{P}_{m-2,j})^2$ is homogeneous of degree $2m-6$ to remove the imaginary unity with no sign change. Multiplying by $\xii$, we prove~\eqref{eq:asymplows}.
We now consider case~\eqref{en:lowd}. In this case, by~\eqref{eq:mainlow2} we obtain:
\[ (\mu_j-id_j(\xi'))^2 \, \check{P}_{m-2,j,j+1}(\mu_j,i\xi') + \rho\,(\mu_j-id_j(\xi'))\,\tilde P_{m-1}(\mu_j,i\xi') + \rho^2\,P_m(\mu_j,i\xi').\]
Noticing that $\check{P}_{m-2,j,j+1}$ is homogeneous of degree $m-4$, $\tilde P_{m-1}$ is homogeneous of degree $m-2$ and $P_m$ is homogeneous of degree~$m$, due to $\mu_j=id_j(\xi')+\textit{o}(1)$, we find
\[ \mu_j-id_j(\xi') = \rho\,\kappa_\pm(\xi')+\textit{o}(\rho),\]
for sufficiently small $\rho$, where $\kappa_\pm$ are the two solutions to~\eqref{eq:kappalow}. Therefore,
\[ \mu_{j+1}-id_j(\xi') = \rho\,\kappa_\mp(\xi')+\textit{o}(\rho). \]
Multiplying by $\xii$, we prove~\eqref{eq:asymplowd}.
\end{proof}
If we allow the non strict interlacing of the polynomials $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$ for some $\xi'\in S^{n-1}$, then Lemma~\ref{lem:high} is generalized as follows.
\begin{lemma}\label{lem:highweak}
Let~$P_{m-2}(\partial_t,\partial_x)$ and $P_m(\partial_t,\partial_x)$ be (possibly weakly) hyperbolic and $P_{m-1}(\partial_t,\partial_x)$ be strictly hyperbolic. Assume that $P_{m-2}(\lambda,\xi')$ and~$P_{m-1}(\lambda,\xi')$ interlace and that $P_{m-1}(\lambda,\xi')$ and~$P_m(\lambda,\xi')$ interlace, for any $\xi'\in S^{n-1}$, and that there is no $(\lambda,\xi')\in\R\times S^{n-1}$ such that $P_{m-2}(\lambda,\xi')=P_{m-1}(\lambda,\xi')=P_m(\lambda,\xi')$. Let~$a_j(\xi')$ be the roots of $P_{m}(\lambda,\xi')$, $\xi'\in S^{n-1}$. Then we may label the $m$ roots~$\lambda_j(\xi)$ of~$Q_2(\lambda,i\xi)$, $\xi\in\R^n$, in such a way that they are described as follows, where we put $\xi'=\xi/\xii$.
\begin{enumerate}[(i)]
\item If $P_{m-1}(a_j(\xi'),\xi')\neq0$ and $\check{P}_{m,j}(a_j(\xi'),\xi')\neq0$, that is, $a_j(\xi')$ is a simple root of $P_m(\lambda,\xi')$ and it is not a root of $P_{m-1}(\lambda,\xi')$, then $\lambda_j(\xi)$ is as in~\eqref{eq:lambdahighsimple};
\item\label{en:highs} If $P_{m-1}(a_j(\xi'),\xi')=0$ and $\check{P}_{m,j}(a_j(\xi'),\xi')\neq0$, that is, $a_j(\xi')$ is a simple root of both $P_{m-1}(\lambda,\xi')$ and $P_m(\lambda,\xi')=0$, then
\begin{equation}\label{eq:asymphighs}
\lambda_j(\xi) = i\xii a_j(\xi') + i\,\xii^{-1}\,\frac{P_{m-2}(a_j(\xi'),\xi')}{\check{P}_{m,j}(a_j,\xi')} - \xii^{-2}\,\frac{P_{m-2}(a_j(\xi'),\xi')\,\tilde P_{m-1}(a_j(\xi'),\xi')}{\big(\check{P}_{m,j}(a_j(\xi'),\xi')\big)^2} + \textit{o}(\xii^{-2}),
\end{equation}
as $\xii\to\infty$, where
\[ \tilde P_{m-1}(\lambda,\xi')=\begin{cases}
\check{P}_{m-1,j-1}(\lambda,\xi') & \text{if $a_j(\xi')=b_{j-1}(\xi')$,}\\
\check{P}_{m-1,j}(\lambda,\xi') & \text{if $a_j(\xi')=b_{j}(\xi')$.}
\end{cases} \]
\item\label{en:highd} If $a_j(\xi')=b_j(\xi')=a_{j+1}(\xi')$, that is, $a_j(\xi')$ is a simple root of $P_{m-1}(\lambda,\xi')$ and a double root of $P_m(\lambda,\xi')$, then
\begin{equation}\label{eq:asymphighd}
\lambda_{j,j+1}(\xi) = i\xii a_j(\xi') + \kappa_\pm(\xi') + \textit{o}(1),
\end{equation}
as $\xii\to\infty$, where $\kappa_\pm(\xi')$ are the two solutions to
\begin{equation}\label{eq:kappahigh}
\kappa^2\check{P}_{m,j,j+1}(a_j(\xi'),\xi') + \kappa\,\check{P}_{m-1,j}(a_j(\xi'),\xi') + P_{m-2}(a_j(\xi'),\xi').
\end{equation}
\end{enumerate}
\end{lemma}
\begin{remark}
We stress that $P_{m-2}(a_j(\xi'),\xi')\,\tilde P_{m-1}(a_j(\xi'),\xi')\neq0$ in~\eqref{eq:asymphighs}, since $a_j(\xi')$ is a simple root of $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$, therefore, it cannot be also a root of $P_{m-2}(\lambda,\xi')$. Due to the interlacing condition,
\[ -\frac{P_{m-2}(a_j(\xi'),\xi')\,\tilde P_{m-1}(a_j(\xi'),\xi')}{\big(\check{P}_{m,j}(a_j(\xi'),\xi')\big)^2}<0. \]
We stress that $\check{P}_{m,j,j+1}(a_j(\xi'),\xi')\neq0$ and $\check{P}_{m-1,j}(a_j(\xi'),\xi')\neq0$ in~\eqref{eq:asymphighd}, since $a_j(\xi')$ is a simple root of $P_{m-1}(\lambda,\xi')$ and a double root of $P_{m}(\lambda,\xi')$, therefore, it cannot be also a root of $P_{m-2}(\lambda,\xi')$. Moreover,
\[ -\frac{\check{P}_{m-1,j}(a_j(\xi'),\xi')}{2\check{P}_{m,j,j+1}(a_j(\xi'),\xi')}<0, \qquad \check{P}_{m,j,j+1}(a_j(\xi'),\xi')\,P_{m-2}(a_j(\xi'),\xi')>0, \]
as a consequence of the interlacing condition; in particular, it follows that $\Re\kappa_\pm(\xi')<0$.
\end{remark}
\begin{proof}[Proof of Lemma~\ref{lem:highweak}]
The proof is similar to the proof of Lemma~\ref{lem:low2weak}. We follow the proof of Lemma~\ref{lem:high}, but we then take into account of the two new scenarios, \eqref{en:highs} and~\eqref{en:highd}. In the both cases, $a_j(\xi')$ is a root of $P_{m-1}(\lambda,\xi')$. Since this latter is strictly hyperbolic, $\tilde{P}_{m-1}(a_j(\xi'),\xi')\neq0$. We first consider case~\eqref{en:highs}. In this case, by
\begin{equation}\label{eq:mainhigh2}
0 = \xii^{-m}\,Q_2(\lambda_j,i\xi)= P_m(\mu_j,i\xi') + \rho^{-1}\,P_{m-1}(\mu_j,i\xi') + \rho^{-2}\,P_{m-2}(\mu_j,i\xi'),
\end{equation}
we find
\begin{align*}
0
& = (\mu_j-ia_j(\xi')) \check{P}_{m,j}(\mu_j,i\xi') + \rho^{-1}\,(\mu_j-ia_j(\xi'))\,\tilde{P}_{m-1}(\mu_j,i\xi') + \rho^{-2}\,P_{m-2}(\mu_j,i\xi')\\
& = (\mu_j-ia_j(\xi'))\big(\check{P}_{m,j}(\mu_j,i\xi') + \rho^{-1}\,\tilde{P}_{m-1}(\mu_j,i\xi')\big)+ \rho^{-2}\,P_{m-2}(\mu_j,i\xi').
\end{align*}
For sufficiently large $\rho$, $\check{P}_{m,j}(\mu_j,i\xi') + \rho^{-1}\,\tilde P_{m-1}(\mu_j,i\xi')\neq0$, since $\check{P}_{m,j}(a_j(\xi'),\xi')\neq0$, therefore
\[ \mu_j-ia_j(\xi') = -\frac{\rho^{-2}\,P_{m-2}(\mu_j,i\xi')}{\check{P}_{m,j}(\mu_j,i\xi') + \rho^{-1}\,\tilde P_{m-1}(\mu_j,i\xi')}. \]
The second-order expansion of the last expression gives:
\[ \mu_j - ia_j(\xi') = i\,\rho^{-2}\,\frac{P_{m-2}(a_j(\xi'),\xi')}{\check{P}_{m,j}(a_j,\xi')} + \textit{o}(\rho^{-2}), \]
where we used that $P_{m-2}$ is homogeneous of degree~$m-2$ and~$\check{P}_{m,j}$ is homogeneous of degree~$m-1$ to remove the imaginary unit, with a multiplication by $-i$. Since we are looking for the real part of $\mu_j$ we need one more expansion step. Writing
\[ -\frac{\rho^{-2}\,P_{m-2}(\mu_j,i\xi')}{\check{P}_{m,j}(\mu_j,i\xi') + \rho^{-1}\,\tilde P_{m-1}(\mu_j,i\xi')} = - \frac{\rho^{-2}\,P_{m-2}(\mu_j,i\xi')\,\big(\check{P}_{m,j}(\mu_j,i\xi') - \rho^{-1}\,\tilde P_{m-1}(\mu_j,i\xi')\big)}{\big(\check{P}_{m,j}(\mu_j,i\xi')\big)^2 - \rho^{-2}\,\big(\tilde P_{m-1}(\mu_j,i\xi')\big)^2} \,,\]
and using the previously obtained information that $\mu_j=ia_j(\xi')+\textit{O}(\rho^{-2})$, we easily obtain the further expansion:
\begin{align*}
\mu_j - ia_j(\xi')
& = i\,\rho^{-2}\,\frac{P_{m-2}(a_j(\xi'),\xi')}{\check{P}_{m,j}(a_j,\xi')} - \rho^{-3}\,\frac{P_{m-2}(a_j(\xi'),\xi')\,\tilde P_{m-1}(a_j(\xi'),\xi')}{\big(\check{P}_{m,j}(a_j(\xi'),\xi')\big)^2} + \textit{o}(\rho^{-3}),
\end{align*}
where in the last computation we used that $P_{m-2}\,\tilde P_{m-1}$ is homogeneous of degree $2m-4$ and $(\check{P}_{m,j})^2$ is homogeneous of degree $2m-2$ to remove the imaginary unity with a sign change. Multiplying by $\xii$, we prove~\eqref{eq:asymphighs}.
We now consider case~\eqref{en:highd}. In this case, by~\eqref{eq:mainhigh2} we obtain:
\[ (\mu_j-ia_j(\xi'))^2 \, \check{P}_{m,j,j+1}(\mu_j,i\xi') + \rho^{-1}\,(\mu_j-ia_j(\xi'))\,\check{P}_{m-1,j}(\mu_j,i\xi') + \rho^{-2}\,P_{m-2}(\mu_j,i\xi').\]
Noticing that $\check{P}_{m,j,j+1}$, $\check{P}_{m-1,j}$ and $P_{m-2}$ are homogeneous of degree $m-2$, due to $\mu_j=ia_j(\xi')+\textit{o}(1)$, we find
\[ \mu_j-ia_j(\xi') = \rho^{-1}\,\kappa_\pm(\xi')+\textit{o}(\rho^{-1}),\]
for sufficiently large $\rho$, where $\kappa_\pm$ are the two solutions to~\eqref{eq:kappahigh}. Consequently,
\[ \mu_{j+1}-ia_j(\xi') = \rho^{-1}\,\kappa_\mp(\xi')+\textit{o}(\rho^{-1}). \]
Multiplying by $\xii$, we prove~\eqref{eq:asymphighd}.
\end{proof}
To discuss the influence on the decay estimates for the solution to~\eqref{eq:CPlin}, coming from the four scenarios in Lemmas~\ref{lem:low2weak} and~\ref{lem:highweak}, it is convenient to localize $u$ at low and high frequencies to treat separately the scenarios at low and high frequencies. Without loss of generality, this localization may be more easily expressed, localizing the initial data.
\begin{theorem}\label{thm:lowloss}
Assume that Hypothesis~\ref{hyp:Q2} holds, and that
\[ u_j\in L^q \cap L^2 \qquad \text{for some $q\in[1,2]$, with $\hat u_j$ compactly supported, for $j=0,\ldots,m-1$.} \]
Then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies the following estimates:
\begin{equation}
\label{eq:estQ2weakgen}\begin{split}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s}
& \leq C\,\sum_{j=m-3}^{m-1}(1+t)^{-\eta}\,\big(\|u_j\|_{L^q}+\|u_j\|_{L^2}\big)\\
& \qquad +C\,\sum_{j=0}^{m-4}(1+t)^{-\frac{n}4\left(\frac1q-\frac12\right)-\frac{k+s-j}4}\,\big(\|u_j\|_{L^q}+\|u_j\|_{L^2}\big),
\end{split}\end{equation}
where~$\eta$ is as in~\eqref{eq:estQ2worst},
provided that $k+s\geq m-3$ if $q=2$, or $n(1/q-1/2)+k+s>m-3$ otherwise. The decay in~\eqref{eq:estQ2worst} may be improved in two cases.
\begin{enumerate}[(i)]
\item\label{en:strictlowloss} If $P_{m-2}(\partial_t,\partial_x)$ is strictly hyperbolic, then~\eqref{eq:estQ2weakgen} may be improved to
\begin{equation}
\label{eq:estQ2strict}\begin{split}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s}
& \leq C\,\sum_{j=m-3}^{m-1}(1+t)^{-\frac{n}4\left(\frac1q-\frac12\right)+\frac{k+s-(m-3)}4}\,\big(\|u_j\|_{L^q}+\|u_j\|_{L^2}\big) \\
& \qquad + C\,\sum_{j=0}^{m-4}(1+t)^{-\frac{n}4\left(\frac1q-\frac12\right)+\frac{k+s-j}4}\,\big(\|u_j\|_{L^q}+\|u_j\|_{L^2}\big).
\end{split}\end{equation}
\item\label{en:weaklowloss} Assume that for any $\xi'\in S^{n-1}$, the interlacing of $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$ is strict, exception given for the double roots of $P_{m-2}(\lambda,\xi')$, that is, $b_1(\xi')<d_1(\xi')$, $d_{m-2}(\xi')<b_{m-1}(\xi')$, and
\[ d_j(\xi')<d_{j+1}(\xi') \Rightarrow d_j(\xi')<b_{j+1}(\xi')<d_{j+1}(\xi'). \]
Then~\eqref{eq:estQ2weakgen} may be improved to
\begin{equation}
\label{eq:estQ2strong}\begin{split}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s}
& \leq C\,\sum_{j=m-3}^{m-1}(1+t)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-(m-2)}2}\,\big(\|u_j\|_{L^q}+\|u_j\|_{L^2}\big) \\
& \qquad + C\,\sum_{j=0}^{m-4}(1+t)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-j-1}2}\,\big(\|u_j\|_{L^q}+\|u_j\|_{L^2}\big).
\end{split}
\end{equation}
\end{enumerate}
\end{theorem}
\begin{proof}
We follow the proof of Theorem~\ref{thm:CPQ2}, but we take into account of the differences in the behavior of $\lambda_j(\xi)$ as $\xi\to0$. Let $\xii\leq\delta$ for sufficiently small $\delta>0$. For $j=m-1,m$, we may proceed as in the proof of Theorem~\ref{thm:CPQ2} and derive
\[ \frac{\big|\hat u_{m-1}-\hat u_{m-2}\sum_{k\neq j}\lambda_k + \ldots + (-1)^{m-1} \hat u_0 \prod_{k\neq j}\lambda_k\big|}{\prod_{k\neq j}|\lambda_k-\lambda_j|}\,e^{\lambda_jt} \lesssim e^{-ct}\,\sum_{k=0}^{m-1} |\hat u_k|, \]
if $4c_{m-2,0}\neq c_{m-1,0}^2$, and similarly if the equality holds.
For $j=1,\ldots,m-2$, for any~$\xi'\in\mathcal S^{n-1}$, we consider the three different scenarios in Lemma~\ref{lem:low2weak}.
If $P_{m-1}(d_j(\xi'),\xi')\neq0$,
then $\lambda_j(\xi)$ is as in~\eqref{eq:lambdalow2}, and we get the pointwise estimate
\begin{align*}
& \frac{\big|\hat u_{m-1}-\hat u_{m-2}\sum_{k\neq j}\lambda_k + \ldots + (-1)^{m-1} \hat u_0 \prod_{k\neq j}\lambda_k\big|}{\prod_{k\neq j}|\lambda_k-\lambda_j|}\,e^{\lambda_jt} \\
& \qquad \leq e^{-ct\xii^2}\, \left( \xii^{-(m-3)}\big(|\hat u_{m-1}|+|\hat u_{m-2}|+|\hat u_{m-3}|\big) + \sum_{k=0}^{m-4}\xii^{-j}\,|\hat u_k|\right),
\end{align*}
as in the proof of Theorem~\ref{thm:CPQ2}.
If $P_{m-1}(d_j(\xi'),\xi')=0$ and $\check{P}_{m-2,j}(d_j(\xi'),\xi')\neq0$, that is, $d_j(\xi')$ is a simple root of both $P_{m-1}(\lambda,\xi')$ and $P_{m-2}(\lambda,\xi')$, then $\lambda_j(\xi)$ is as in~\eqref{eq:asymplows}, and we get the pointwise estimate
\begin{align*}
& \frac{\big|\hat u_{m-1}-\hat u_{m-2}\sum_{k\neq j}\lambda_k + \ldots + (-1)^{m-1} \hat u_0 \prod_{k\neq j}\lambda_k\big|}{\prod_{k\neq j}|\lambda_k-\lambda_j|}\,e^{\lambda_jt} \\
& \qquad \leq e^{-ct\xii^4}\, \left( \xii^{-(m-3)}\big(|\hat u_{m-1}|+|\hat u_{m-2}|+|\hat u_{m-3}|\big) + \sum_{k=0}^{m-4}\xii^{-j}\,|\hat u_k|\right),
\end{align*}
where the only difference with the case treated in Theorem~\ref{thm:CPQ2} is that the exponential term $e^{-ct\xii^2}$ is replaced by $e^{-ct\xii^4}$. Indeed, it remains valid that
\begin{align*}
& |\lambda_k|\lesssim\xii, \qquad k=1,\ldots,m-2,\\
& |\lambda_k-\lambda_j|=|d_k(\xi')-d_j(\xi')|\,\xii\gtrsim \xii, \quad k=1,\ldots,m-2, \ k\neq j,
\end{align*}
since $d_j(\xi')$ is a simple root of $P_{m-2}(\lambda,\xi')$. In particular, if $P_{m-2}(\lambda,\xi')$ is strictly hyperbolic for any $\xi'\in S^{n-1}$, following as in the proof of Theorem~\ref{thm:CPQ2}, this just leads to replace the decay rate in~\eqref{eq:estQ2} with the decay rate in~\eqref{eq:estQ2strict}.
If $d_j(\xi')$ is a simple root of $P_{m-1}(\lambda,\xi')$ and a double root of $P_{m-2}(\lambda,\xi')$, then $\lambda_j(\xi)$ is as in~\eqref{eq:asymplowd}, in particular, $\Re\lambda_j(\xi)\leq-c\xii^2$, so that the power at the denominator in the decay rate remains as in~\eqref{eq:estQ2}. However, in this case, $\lambda_j(\xi)-\lambda_{j+1}(\xi)$ vanishes at a faster speed as~$\xi\to0$. Rewriting the representation of $\hat u(t,\xi)$ as in Remark~\ref{rem:multiple}, we get the pointwise estimate
\begin{align*}
& \frac{\big|\hat u_{m-1}-\hat u_{m-2}\sum_{k\neq j}\lambda_k + \ldots + (-1)^{m-1} \hat u_0 \prod_{k\neq j}\lambda_k\big|}{\prod_{k\neq j}|\lambda_k-\lambda_j|}\,e^{\lambda_jt} \\
& \qquad \leq (1+t\xii)\,e^{-ct\xii^2}\, \left( \xii^{-(m-3)}\big(|\hat u_{m-1}|+|\hat u_{m-2}|+|\hat u_{m-3}|\big) + \sum_{k=0}^{m-4}\xii^{-j}\,|\hat u_k|\right),
\end{align*}
where the only difference with the case treated in Theorem~\ref{thm:CPQ2} is the term $(1+t\xii)$; this term produces a loss of decay rate $\sqrt{t}$ for large $t$. In turn, if the interlacing of $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$ is strict, exception given for the double roots of $P_{m-2}(\lambda,\xi')$, for any $\xi'\in S^{n-1}$, following as in the proof of Theorem~\ref{thm:CPQ2}, this just leads to replace the decay rate in~\eqref{eq:estQ2} with the decay rate in~\eqref{eq:estQ2strong}, where the loss of decay $(1+t)^{\frac12}$ appears.
In the general case, if $P_{m-2}(\lambda,\xi')$ is not strictly hyperbolic, and the interlacing of $P_{m-2}(\lambda,\xi')$ and $P_{m-1}(\lambda,\xi')$ is weak, for some $\xi'\in S^{n-1}$, either of the previous scenario may happen for different roots $\lambda_j$ and/or at different points $\xi'\in S^{n-1}$, so that we shall consider the worst case scenario, that is, we shall replace the decay rate in~\eqref{eq:estQ2} with the one in~\eqref{eq:estQ2worst}.
\end{proof}
At high frequencies, we shall distinguish two cases. Indeed, if $P_m(\partial_t,\partial_x)$ is hyperbolic, but not strictly, then~\eqref{eq:CPlin} is not well-posed in $H^s$, in general, but only in $H^\infty$ (or $\mathcal C^\infty$), since a loss of $1$ derivative may occur. For this reason, we first discuss the case in which $P_m(\partial_t,\partial_x)$ is strictly hyperbolic, but the interlacing of $P_{m-1}(\lambda,\xi')$ and $P_m(\lambda,\xi')$ may be not strict for some $\xi'\in S^{n-1}$.
\begin{proposition}\label{prop:lossdecay}
Assume that Hypothesis~\ref{hyp:Q2} holds, and that $P_m(\partial_t,\partial_x)$ is strictly hyperbolic. Assume that
\[ u_j\in H^{s_0-j} \qquad \text{with $\hat u_j=0$ in a neighborhood of the origin, $j=0,\ldots,m-1$,} \]
for some $s_0\geq m-1$. Then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies the following regularity-loss type decay estimate:
\begin{equation}
\label{eq:estQ2loss}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s} \leq C\,(1+t)^{-\frac\nu2}\,\|u_j\|_{H^{k+s+\nu-j}},
\end{equation}
for any~$\nu\geq0$ such that $k+s+\nu\leq s_0$.
If, moreover, $P_{m-1}(\lambda,\xi')$ and $P_m(\lambda,\xi')$ strictly interlace for any $\xi'\in S^{n-1}$, then~\eqref{eq:estQ2loss} is improved to
\begin{equation}
\label{eq:estQ2noloss}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s} \leq C\,e^{-ct}\,\|u_j\|_{H^{k+s-j}},
\end{equation}
for any $k+s\leq s_0$.
\end{proposition}
The type of decay rate in~\eqref{eq:estQ2loss} is a \emph{regularity-loss decay,} in the sense that additional initial data regularity produces extra decay rate for the solution~\cite{HK06}. This phenomenon is known for some plate models under rotational inertia effects~\cite{CDI13p, SK10} and it also holds for wave equations with very strong damping~\cite{GGH} $\partial_t^2-\Delta u+(-\Delta)^ku_t=0$, $k>1$. Proposition~\ref{prop:lossdecay} shows that, as a consequence of Lemma~\ref{lem:highweak}, this phenomenon may also hold for pure hyperbolic equations.
\begin{proof}
We follow the proof of Theorem~\ref{thm:CPQ2} at high frequencies, but now, if $a_j(\xi')$ is a simple root of both $P_{m-1}(\lambda,\xi')$ and $P_m(\lambda,\xi')$ for some $\xi'\in S^{n-1}$ and $j=1,\ldots,m$, then, due to~\eqref{eq:asymphighs}, recalling~\eqref{eq:urepgen}, we may only estimate
\[ |\partial_t^k\hat u(t,\xi)| \lesssim \sum_{j=0}^{m-1} \xii^{k-j} |\hat u_j(\xi)|\,e^{-\frac{ct}{\xii^2}}, \]
for~$k\geq0$. We immediately obtain that, for any~$\nu\geq0$ such that $k+s+\nu\leq s_0$,
\[
\Big(\int_{\xii\geq M} \xii^{2s} |\partial_t^k\hat u(t,\xi)|^2\,d\xi\Big)^{\frac12} \lesssim \big(\sup_{\xii\geq M} \xii^{-\nu}\,e^{-\frac{ct}{\xii^2}}\big)\,\sum_{j=0}^{m-1} \|u_j\|_{H^{s+k+\nu-j}},
\]
and~\eqref{eq:estQ2loss} follows.
\end{proof}
When $P_m(\partial_t,\partial_x)$ is not strictly hyperbolic, Cauchy problem~\eqref{eq:CPlin} is well-posed in $H^\infty$.
\begin{proposition}\label{prop:regloss}
Assume that Hypothesis~\ref{hyp:Q2} holds, and that
\[ u_j\in H^\infty \qquad \text{with $\hat u_j=0$ in a neighborhood of the origin, $j=0,\ldots,m-1$,} \]
Then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies decay estimate~\eqref{eq:estQ2loss} for any~$\nu\geq1$. If, moreover, the interlacing of $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$ is strict, exception given for the double roots of $P_{m}(\lambda,\xi')$ for any $\xi'\in S^{n-1}$, that is,
\[ a_j(\xi')<a_{j+1}(\xi') \Rightarrow a_j(\xi')<b_{j}(\xi')<a_{j+1}(\xi'), \]
then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies the decay estimate
\begin{equation}
\label{eq:estQ2regloss}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s} \leq C\,e^{-ct}\,\|u_j\|_{H^{k+s+1-j}}.
\end{equation}
\end{proposition}
The loss of $1$ derivative in estimate~\eqref{eq:estQ2regloss}, and the restriction~$\nu\geq1$ in~\eqref{eq:estQ2loss} are consistent to the assumption of initial data in $H^\infty$, in order to get the well-posedness (in $H^\infty$) of Cauchy problem~\eqref{eq:CPlin}.
\begin{proof}
We follow the proof of Theorem~\ref{thm:CPQ2} at high frequencies, but now, $a_j(\xi')$ may be a double root of $P_m(\lambda,\xi')$ for some $\xi'\in S^{n-1}$. If $a_j(\xi')=a_{j+1}(\xi')$, the asymptotic behavior of $\lambda_j$ and $\lambda_{j+1}$ is described by~\eqref{eq:asymphighd}, and we may rely on Remark~\ref{rem:multiple} to estimate
\begin{align*}
& \left|\frac{\hat u_{m-1}-\hat u_{m-2}\sum_{k\neq j}\lambda_k + \ldots + (-1)^{m-1} \hat u_0 \prod_{k\neq j}\lambda_k}{\prod_{k\neq j}(\lambda_k-\lambda_j)}\,e^{\lambda_jt}\right.\\
& \qquad \left. + \frac{\hat u_{m-1}-\hat u_{m-2}\sum_{k\neq j+1}\lambda_k + \ldots + (-1)^{m-1} \hat u_0 \prod_{k\neq j+1}\lambda_k}{\prod_{k\neq j+1}(\lambda_k-\lambda_j)}\,e^{\lambda_{j+1}t}\right|\\
& \lesssim (1+t\xii)\,e^{-ct}\,\sum_{k=0}^{m-1}|\hat u_k|\,\xii^{-k}.
\end{align*}
If the interlacing of $P_{m-1}(\lambda,\xi')$ and $P_{m}(\lambda,\xi')$ is strict, exception given for the double roots of $P_{m}(\lambda,\xi')$, for any $\xi'\in S^{n-1}$, we immediately derive~\eqref{eq:estQ2regloss}. Otherwise, taking into account that $a_j(\xi')$ may be a simple root of both $P_{m-1}(\lambda,\xi')$ and $P_m(\lambda,\xi')$ for some $\xi'\in S^{n-1}$ and $j=1,\ldots,m$, we proceed as in the proof of Proposition~\ref{prop:regloss}. Considering both scenarios, we derive~\eqref{eq:estQ2loss} with~$\nu\geq1$.
\end{proof}
Theorem~\ref{thm:lowloss} and Propositions~\ref{prop:lossdecay} and~\ref{prop:regloss} may be combined in different ways, according to the interlacing assumptions of $P_{m-2}(\lambda,\xi')$ with $P_{m-1}(\lambda,\xi')$ and of $P_{m-1}(\lambda,\xi')$ with $P_{m}(\lambda,\xi')$, and to the strict or weak hyperbolicity of $P_{m-2}(\partial_t,\partial_x)$ and $P_{m}(\partial_t,\partial_x)$. In the very general case, with no extra assumption with respect to the necessary and sufficient condition for the strict stability, given in Hypothesis~\ref{hyp:Q2}, the proof of Theorem~\ref{thm:verygen} follows by combining estimate in~\eqref{eq:estQ2weakgen} in Theorem~\ref{thm:lowloss} and estimate~\eqref{eq:estQ2loss} with $\nu\geq1$ in Proposition~\ref{prop:regloss}.
An analogous result to Theorem~\ref{thm:asymp} may be obtained to describe the asymptotic behavior of the solution to~\eqref{eq:CPlin}, following the proof of Theorem~\ref{thm:asymp}, under the more general assumptions in this section. For the sake of brevity, we provide such result in the simpler anisotropic case, as in Examples \ref{ex:MGT}, \ref{ex:BC}, \ref{ex:electric}.
\begin{theorem}\label{thm:asympweak}
Assume that Hypothesis~\ref{hyp:Q2} holds, that
\[ u_j\in L^1\cap H^\infty \qquad \text{for $j=0,\ldots,m-1$,} \]
and that $P_{m-j}(\lambda,\xi')=P_{m-j}(\lambda)$ are anisotropic, i.e., independent of~$\xi'\in S^{n-1}$, $j=0,1,2,$. Let
\[ J_1=\{j: \ d_j=b_{j+1}=d_{j+1} \}, \qquad J_2=\{j: \ d_{j-1}<b_j=d_j \ \text{or} \ d_j=b_{j+1}<d_{j+1} \}, \]
and define
\begin{align}
\label{eq:vweak}
\hat v(t,\xi)
& = M\,\sum_{j\in J_1}\frac{e^{i\xii\,d_jt}}{i^{m-4}\check{P}_{m-2,j,j+1}(d_j)}\,\frac{e^{\xii^2\kappa_-t}-e^{\xii^2\kappa_+t}}{\kappa_+-\kappa_-}\,, \\
\label{eq:wweak}
\hat w(t,\xi)
& = M\,\sum_{j\in J_2}\frac{e^{i\xii\,d_jt + i\,\xii^3\,\frac{P_m(d_j)}{\check{P}_{m-2,j}(d_j)}\,t + \xii^4\,\frac{P_m(d_j)\,\tilde P_{m-1}(d_j)}{(\check{P}_{m-2,j}(d_j))^2}\,t}}{i^{m-3}\check{P}_{m-2,j}(d_j)}\,,
\end{align}
with $M$ as in~\eqref{eq:moment}, $\tilde P_{m-1}$ as in~\eqref{eq:Ptilde}, and $\kappa_\pm$ as in~\eqref{eq:kappalow}, provided that $\kappa_+\neq \kappa_-$. If $J_2$ is empty and $J_1$ is nonempty, then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies the following estimate
\begin{equation}
\label{eq:Q2asympweak2}
\big\|\partial_t^k \big( u(t,\cdot) - I_{m-2}v(t,\cdot)\big)\big\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}4-\frac{k+s-(m-2)}2}\big)
\end{equation}
for $n/2+k+s>m-2$. If $J_2$ is nonempty, then the solution $u$ to Cauchy problem~\eqref{eq:CPlin} with $Q=Q_2$ satisfies estimate~\eqref{eq:Q2asympweak2} for $m-2<n/2+k+s<m-1$ and the following estimate
\begin{equation}
\label{eq:Q2asympweak1}
\big\|\partial_t^k \big( u(t,\cdot) - I_{m-3}w(t,\cdot)\big)\big\|_{\dot H^s} = \textit{o}\big((1+t)^{-\frac{n}8-\frac{k+s-(m-3)}4}\big)
\end{equation}
for $n/2+k+s>m-1$. Moreover, \eqref{eq:Q2asympweak1} holds for $n/2+k+s>m-3$ if $J_1$ is empty.
\end{theorem}
The proof is analogous to the proof of Theorem~\ref{thm:asymp}, so we omit it.
\begin{example}\label{ex:electricdiss}
We go back to Example~\ref{ex:electric}, but we add a dissipation term $a\partial_tu$ to the system of elastic waves~\eqref{eq:MHD3dplus}:
\begin{equation}\label{eq:MHD3diss}
\begin{cases}
\partial_t^2 u + a \partial_t u - \mu \Delta u - (\mu+\nu) \nabla \div u + \gamma \curl E =0,\\
\partial_t E +\sigma E - \curl H -\gamma\curl \partial_t u=0,\\
\partial_t H + c^2\curl E =0.
\end{cases}
\end{equation}
Proceeding as in Example~\ref{ex:electric}, and defining $\varphi=\partial_tE+\sigma E$ for each component of $E$, $\varphi$ satisfies the fourth order equation
\[ \big( (\partial_t^2+a\partial_t-\mu\Delta)(\partial_t^2+\sigma\partial_t-c^2\Delta)-\gamma^2\partial_t^2\Delta\big) \varphi=0. \]
The equation above is $Q_2(\partial_t,\partial_x)\varphi=0$, where $Q_2=P_4+P_3+P_2$ and
\begin{align*}
P_4(\lambda,\xi)
& = \lambda^4-(\mu+c^2+\gamma^2)\lambda^2\xii^2+c^2\mu\xii^4, \\
P_3(\lambda,\xi)
& = (a+\sigma)\lambda^3 - (ac^2+\mu\sigma)\lambda\xii^2, \\
P_2(\lambda,\xi)
& = a\sigma\lambda^2.
\end{align*}
Now~$P_2(\lambda,\xi')$ has a double root~$0$, the roots of~$P_3(\lambda,\xi')$ are~$b_2=0$ and
\[ b_3=\sqrt{\frac{ac^2+\mu\sigma}{a+\sigma}}, \qquad b_1=-b_3, \]
and the roots of $P_4(\lambda,\xi')$ are
\[ a_{3,4} = \frac1{\sqrt{2}}\,\sqrt{\mu+c^2+\gamma^2\pm \sqrt{(\mu+c^2+\gamma^2)^2-4c^2\mu}}, \quad a_{1,2}=-a_{3,4}. \]
The strict interlacing condition of $P_3(\lambda,\xi')$ and $P_4(\lambda,\xi')$ may be checked by straightforward calculations. Therefore, Theorem~\ref{thm:lowloss}, \eqref{en:lowd} and Proposition~\ref{prop:lossdecay} apply.
By Lemma~\ref{lem:low2weak}, due to
\[ \check{P}_{2,1,2}(0,\xi')=a\sigma, \qquad \check{P}_{3,2}(0,\xi')=-(ac^2+\mu\sigma), \qquad P_4(0,\xi')=c^2\mu, \]
we may compute
\[ \lambda_4(0)=-a, \qquad \lambda_3(0)=-\sigma, \qquad \lambda_2(\xi)=-\frac{ac^2}{a+\sigma}\,\xii^2+\textit{o}(\xii^2), \qquad \lambda_1(\xi)=-\frac{\mu\sigma}{a+\sigma}\,\xii^2+\textit{o}(\xii^2),\]
as $\xi\to0$. By Theorem~\ref{thm:asympweak}, if $ac^2\neq \mu\sigma$, the asymptotic profile of $\varphi(t,\cdot)$ is described by $I_2v$, where
\begin{align*}
\mathfrak{F}(I_2v)
& = \xii^{-2}\hat v(t,\xi) = M\,\frac{a+\sigma}{a\sigma(\mu\sigma-ac^2)}\,\frac{e^{-\frac{ac^2}{a+\sigma}\,\xii^2 t}-e^{-\frac{\mu\sigma}{a+\sigma}\,\xii^2 t}}{\xii^2}\,,
\intertext{if}
M
& = \int_{\R^3} \big(\partial_t^3\varphi(0,x)+(a+\sigma)\partial_t^2 \varphi(0,x)+a\sigma \partial_t \varphi(0,x)\big)\,dx \neq0.
\end{align*}
Explicitly,
\[ \|\partial_t^k\big( \varphi(t,\cdot)-I_2v(t,\cdot)\big)\|_{\dot H^s} =\textit{o}\big((1+t)^{-\frac34-(k-1)-\frac{s}2}\big),\]
for~$k\geq1$ or~$s>1/2$.
\end{example}
We also provide an example of the application of Lemma~\ref{lem:low2weak} to the theory of systems of elastic waves with anisotropic dissipation.
\begin{example}\label{ex:anisotropic}
We consider the eigenvalues of a system of elastic waves in $\R^2$,
\[ \partial_t^2 u - \mu \Delta u - (\mu+\nu)\nabla \div u + A\partial_tu =0, \qquad A=\begin{pmatrix}
a_1 & 0 \\
0 & a_2
\end{pmatrix} \]
where $A$ with $a_1>a_2>0$, represents an anisotropic dissipation. The constant~$\mu,\nu$ are the Lam\`e constants and verify $\mu>0$, $\mu+\nu>0$. Let $U=(\xii \hat u, \partial_t \hat u)$. Then
\[ \partial_t U = MU, \quad M=\begin{pmatrix}
0 & \xii \mathrm{I_2}\\
-\xii^{-1}B & -A
\end{pmatrix}, \qquad B
\begin{pmatrix}
(2\mu+\nu)\xi_1^2+\mu\xi_2^2 & (\mu+\nu)\xi_1\xi_2\\
(\mu+\nu)\xi_1\xi_2 & \mu\xi_1^2+(2\mu+\nu)\xi_2^2 \end{pmatrix}. \]
By straightforward computation,
\[ \det (\lambda-M) = P_4(\lambda,i\xi)+P_3(\lambda,i\xi)+P_2(\lambda,i\xi), \]
where
\begin{align*}
P_4(\lambda,\xi)
& = (\lambda^2-\mu\xii^2)(\lambda^2-(2\mu+\nu)\xii^2),\\
P_3(\lambda,\xi)
& = \lambda^3 (a_1+a_2)- \lambda \big( a_1(\mu\xii^2+(\mu+\nu)\xi_2^2) + a_2(\mu\xii^2+(\mu+\nu)\xi_1^2)\big),\\
P_2(\lambda,\xi)
& = \lambda^2 a_1a_2.
\end{align*}
In particular, $P_4(\partial_t,\partial_x)$ and $P_3(\partial_t,\partial_x)$ are strictly hyperbolic, whereas $P_2(\partial_t,\partial_x)$ is weakly hyperbolic. The interlacing condition holds, since the roots of $P_4(\lambda,\xi)$ are $\pm \xii\sqrt{\mu}$ and $\pm \xii\sqrt{2\mu+\nu}$, whereas the roots of $P_3(\lambda,\xi)$ are $0$ and
\[ \pm \xii\,\sqrt{\mu + (\mu+\nu)\frac{a_1\xi_2^2+a_2\xi_1^2}{(a_1+a_2)\,\xii^2}}\,. \]
Therefore, the eigenvalues $\lambda_j(\xi)$ of~$M$ verify $\Re\lambda_j(\xi)<0$ for any $\xi\neq0$. Due to Lemma~\ref{lem:low2weak}, case~\eqref{en:lowd},
\[ \lambda_{4}(0)=-a_2, \quad \lambda_3(0)=-a_1, \quad \lambda_{1,2}(\xi)= \xii^2\,\kappa_\pm(\xi')+\textit{o}(\xii^2),\]
as $\xi\to0$, where $\kappa_\pm(\xi')$, $\xi'\in S^1$, are the two solutions to
\[ \kappa^2 a_1a_2 + \kappa \big( a_1(\mu+(\mu+\nu)(\xi_2')^2) + a_2(\mu+(\mu+\nu)(\xi_1')^2)\big) + \mu(2\mu+\nu) =0. \]
We stress that, as expected, in the limit case $a_2=0$, the polynomial $Q(\lambda,i\xi)=P_4(\lambda,i\xi)+P_3(\lambda,i\xi)$ is not strictly stable, since the interlacing of $P_3(\lambda,\xi)$ and $P_4(\lambda,\xi)$ is not strict at $(\pm1,0)$ and at $(0,\pm1)$.
\end{example}
Finally, we present two examples without a physical motivation, just to clarify the possible scenarios of weak interlacing from a mathematical point of view.
\begin{example}
If we replace the viscoelastic damping $-b\Delta\partial_t$ in the MGT equation in Example~\ref{ex:MGT} by a classical damping $b\partial_t$, then we obtain the third order operator $Q=Q_2(\partial_t,\partial_x)$, where
\[ P_3(\lambda,\xi')=\lambda(\lambda^2-c^2), \quad P_2(\lambda,\xi')=\tau^{-1}(\lambda^2-c^2), \quad P_1(\lambda,\xi')=\tau^{-1}b\lambda, \]
to which Lemma~\ref{lem:highweak} applies, and by~\eqref{eq:asymphighs} we derive the asymptotic behaviors
\[ \lambda_{1,3}(\xi) = \pm i c\xii \pm \frac{ib}{2c\tau\xii} - \frac{b}{2c^2\tau^2\xii^2} + \textit{o}(\xii^{-2}),\]
whereas by~\eqref{eq:lambdahighsimple} we get $\lambda_2(\xi)= -\tau^{-1} + \textit{o}(1)$, as $\xii\to\infty$.
\end{example}
\begin{example}
We consider the fourth order operator
\[ Q_2(\partial_t,\partial_x)= \partial_t^2(\partial_t^2-c^2\Delta) + \partial_t(\partial_t^2-\Delta)+\partial_t^2-\Delta,\]
with $c>1$, that is, we fix
\[ P_4(\lambda,\xi')=\lambda^2(\lambda^2-c^2), \quad P_3(\lambda,\xi')=\lambda(\lambda^2-1), \quad P_2(\lambda,\xi')=\lambda^2-1. \]
Then $\lambda_{3,4}(0)=(-1\pm 3i)/2$ and, by Lemma~\ref{lem:low2weak}, we obtain the asymptotic behaviors
\[ \lambda_{1,2}(\xi)= \pm i\xii \mp i\,\frac{c^2-1}{2} \, \xii^3 - \frac{c^2-1}{2} \,\xii^4 + \textit{o}(\xii^4), \]
as $\xi\to0$. Let $\kappa_\pm=(-1\pm i \sqrt{4c^2-1})/(2c^2)$ be the two solutions to~\eqref{eq:kappahigh}, that is, $c^2\kappa^2 + \kappa +1 =0$. By Lemma~\ref{lem:highweak}, we derive the asymptotic behaviors
\[ \lambda_{1,2}(\xi)= \pm ic\xii - \frac{c^2-1}{2c^2} + \textit{o}(1), \qquad \lambda_{3,4}(\xi)=\frac{-1\pm i \sqrt{4c^2-1}}{2c^2} + \textit{o}(1), \]
as $\xii\to\infty$. Applying Theorem~\ref{thm:asympweak}, we get the asymptotic behavior $I_1w$ for $u$, where
\begin{align*}
\mathfrak{F}(I_1w)
& =\xii^{-1}\hat w(t,\xi)= M\,\frac{e^{i\xii t-i\,\frac{c^2-1}{2} \, \xii^3 t- \frac{c^2-1}{2} \,\xii^4t}-e^{-i\xii t+i\,\frac{c^2-1}{2} \, \xii^3t - \frac{c^2-1}{2} \,\xii^4t}}{2i\xii} \\
& = M\,\xii^{-1}\sin \Big(\xii t- \frac{c^2-1}{2}\xii^3t\Big)\,e^{-\frac{c^2-1}{2} \,\xii^4t}\,.
\end{align*}
Explicitly,
\[ \|\partial_t^k\big( u(t,\cdot)-I_1w(t,\cdot)\big)\|_{\dot H^s} =\textit{o}\big((1+t)^{-\frac{n}8-\frac{k+s-1}4}\big),\]
for~$n/2+k+s>1$.
\end{example}
\section{Application of Theorem~\ref{thm:CPQ2} to semilinear problems}\label{sec:nonlinear}
The estimates in Theorem~\ref{thm:CPQ2} may be applied to study different nonlinear problems. In this section, we present an application to the problem perturbed by a power nonlinearity. We consider
\begin{equation}\label{eq:CPnl}
\begin{cases}
Q(\partial_t,\partial_x) u =f(D^\alpha u), \quad t\geq 0, x\in\R^n,\\
\partial_t^ju(0,x)=u_j(x),\quad j=0,\ldots,m-1,
\end{cases}
\end{equation}
where $D=(\nabla,\partial_t)$, $\nu=|\alpha|=0,\ldots,m-2$, and $f$ verifies the following local Lipschitz-type condition
\begin{equation}\label{eq:f}
|f(s_2)-f(s_1)|\leq C\,\big(|s_2|+|s_1|\big)^{p-1}\,|s_2-s_1|,
\end{equation}
for some $p>2$. For instance, $f=\pm |\partial_t^{\nu}u|^p$. We also define $\iota=0$ if we consider $Q=Q_1$, and $\iota=1$ if we consider $Q=Q_2$, in~\eqref{eq:CPnl}. Let the space dimension~$n$ be such that
\begin{equation}\label{eq:dimension}
(m-2-\iota-\nu)<n\leq 2(m-1-\iota-\nu).
\end{equation}
Then, assuming small initial data $u_j \in L^1\cap H^{m-1-j}$, $j=0,\ldots,m-1$, global small data solutions to~\eqref{eq:CPnl} exist for
\begin{equation}\label{eq:pcrit}
p> \bar p(n) = 1 + \frac{m-\iota-\nu}{n-(m-2-\iota-\nu)}\,.
\end{equation}
We notice that~\eqref{eq:dimension} and~\eqref{eq:pcrit} imply that $p>2$.
\begin{corollary}\label{cor:nl}
Let $Q=Q_1$ or $Q=Q_2$, as in Theorem~\ref{thm:CPQ2}, and fix $\iota=0$ if $Q=Q_1$ or $\iota=1$ if $Q=Q_2$. Assume that the space dimension~$n$ verifies~\eqref{eq:dimension} and that $p$ in~\eqref{eq:f} verifies~\eqref{eq:pcrit}. Then there exists $\varepsilon>0$ such that if
\[ u_j \in L^1\cap H^{m-1-j}, \quad j=0,\ldots,m-1, \qquad \mathcal A=\sum_{j=0}^{m-1} \big(\|u_j\|_{L^1}+\|u_j\|_{H^{m-1-j}}\big) \leq \varepsilon, \]
there exists a unique solution $u\in X=\cap_{j=0}^{m-1} \mathcal C^j([0,\infty),H^{m-1-j})$ to Cauchy problem~\eqref{eq:CPnl}. Moreover, the solution verifies the decay estimate
\begin{equation}
\label{eq:estnl}\begin{split}
\|\partial_t^ku(t,\cdot)\|_{\dot H^s}
& \leq C\,(1+t)^{-\frac{n}4-\frac{k+s-(m-2-\iota)}2}\,\mathcal A,
\end{split}\end{equation}
for any $k+s\leq m-1$ such that $n/2+k+s>m-2-\iota$.
\end{corollary}
\begin{proof}
The proof is very classical. The local existence of the solution follows by standard arguments and Sobolev embeddings, so we only prove the existence and uniqueness of the global solution. We prove Corollary~\ref{cor:nl} for $D^\alpha=\partial_t^\nu$, being the proof in the other cases analogous. For a given small $\delta>0$, we fix $\kappa=\delta+m-2-\iota-n/2$ and we define the norm
\begin{align*}
\|v\|_X
& = \sup_{t\geq0} (1+t)^{\frac{n}4+\frac{\kappa-(m-2-\iota)}2}\,\sum_{k+s=\kappa} \|\partial_t^ku(t,\cdot)\|_{\dot H^s}\\
& \qquad + \sup_{t\geq0} \sum_{k+s=[\kappa]+1}^{m-1} (1+t)^{\frac{n}4+\frac{k+s-(m-2-\iota)}2}\,\|\partial_t^ku(t,\cdot)\|_{\dot H^s},
\end{align*}
where $[\kappa]=\max\{j\in\N: \ j\leq \kappa\}$. In particular, we choose sufficiently small $\delta>0$ such that $\kappa\leq \nu+n(1/2-1/p)$. We stress that $m-2-\iota-n/2<\nu+n(1/2-1/p)$, thanks to~\eqref{eq:pcrit}.
By Theorem~\ref{thm:CPQ2}, the solution $\bar u$ to~\eqref{eq:CPlin} verifies the estimate $\|\bar u\|_X \leq C_1\mathcal A$, for some $C_1>0$. Moreover, if $\|v\|_X,\|w\|_X<\infty$, then, by~\eqref{eq:f} and H\"older's inequality,
\begin{equation}\label{eq:estFDv}\begin{split}
\|(f(\partial_t^\nu v)-f(\partial_t^\nu w))(s,\cdot)\|_{L^q}
& \leq C\,\|\partial_t^\nu(v-w)(s,\cdot)\|_{L^{pq}}\big(\|\partial_t^\nu v(s,\cdot)\|_{L^{pq}}^{p-1}+\|\partial_t^\nu w(s,\cdot)\|_{L^{pq}}^{p-1}\big)\\
& \lesssim (1+s)^{-\frac{n}{2q}\,(p-1)-\frac{\nu-(m-2-\iota)}2\,p}\,\|v-w\|_X\,\big(\|v\|_{X}^{p-1}+\|w\|_{X}^{p-1}\big),
\end{split}\end{equation}
for any $q\geq1$, where we used $\|\cdot\|_{L^{pq}}\lesssim \|\cdot\|_{\dot H^{n\left(\frac12-\frac1{qp}\right)}}$ and the definition of $\|\cdot\|_X$.
We define the integral operator
\[ Fv(t,\cdot) = \int_0^t K(t-s)\ast f(\partial_t^\nu v)(s,\cdot)\,ds, \]
where $K$ denotes the fundamental solution to~\eqref{eq:CPlin}, i.e., $K(t)\ast u_{m-1}$ solves~\eqref{eq:CPlin} with $u_0=\ldots=u_{m-2}=0$. If we prove that
\begin{equation}\label{eq:Fest}
\|Fv-Fw\|_X \leq C_2\,\|v-w\|_X\,\big(\|u\|_{X}^{p-1}+\|w\|_{X}^{p-1}\big),\quad \text{for any $v,w\in X$,}
\end{equation}
then we get the existence and uniqueness of the solution to~\eqref{eq:CPnl}, by Banach's contraction mapping theorem. Indeed, let~$R=2C_1\mathcal A$. Assuming $\mathcal A\leq\varepsilon$, it follows~$CR^{p-1}<1/2$ for sufficiently small $\varepsilon$. Then~$\|\bar u\|_X\leq R/2$ and~$F$ is a contraction on~$X_R= \{u\in X: \ \|u\|_{X}\leq R\}$. The solution~$u$ to~\eqref{eq:CPnl} is the unique fixed point for~$\bar u(t,x)+Fu(t,x)$. Moreover, $\|u\|_X\leq R= 2C_1\,\mathcal A$, so that~\eqref{eq:estnl} follows for any $k+s\leq m-1$ such that $n/2+k+s\geq\delta+m-2-\iota$. Since $\delta>0$ is arbitrarily small, we may obtain~\eqref{eq:estnl} for $n/2+k+s>m-2-\iota$.
To prove~\eqref{eq:Fest}, we apply Theorem~\ref{thm:CPQ2} to $K(t-s)\ast f(\partial_t^\nu v)(s,\cdot)$, with $q=1$ for $s\in[0,t/2]$ and with $q\in[1,2]$ such that
\[ 0<\frac{n}2\left(\frac1q-\frac12\right)+\frac{k+s-(m-2-\iota)}2 \leq 1, \]
for $s\in[t/2,t]$. Using~\eqref{eq:estFDv}, we obtain, for $k+s\geq\kappa$:
\begin{align*}
& \|\partial_t^k(Fv-Fw)(t,\cdot)\|_{\dot H^s} \lesssim (I_1(t)+I_2(t))\,\|v-w\|_X\,\big(\|u\|_{X}^{p-1}+\|w\|_{X}^{p-1}\big), \quad \text{where} \\
& I_1 = \int_0^{t/2} (1+t-s)^{-\frac{n}4-\frac{k+s-(m-2-\iota)}2}\,(1+s)^{-\frac{n}{2}\,(p-1)-\frac{\nu-(m-2-\iota)}2\,p}\,ds\\
& I_2 = \int_{t/2}^t (1+t-s)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-(m-2-\iota)}2}\,(1+s)^{-\frac{n}{2q}\,(p-1)-\frac{\nu-(m-2-\iota)}2\,p}\,ds.
\end{align*}
Noticing that
\[ -\frac{n}{2}\,(p-1)-\frac{\nu-(m-2-\iota)}2\,p<-1,\]
if, and only if, \eqref{eq:pcrit} holds, we may estimate
\begin{align*}
I_1
& \lesssim (1+t)^{-\frac{n}4-\frac{k+s-(m-2-\iota)}2}\,\int_0^{t/2} (1+s)^{-\frac{n}{2}\,(p-1)-\frac{\nu-(m-2-\iota)}2\,p}\,ds \\
& \lesssim (1+t)^{-\frac{n}4-\frac{k+s-(m-2-\iota)}2}\,,\\
I_2
& \lesssim (1+t)^{-\frac{n}{2q}\,(p-1)-\frac{\nu-(m-2-\iota)}2\,p}\,\int_{t/2}^t (1+t-s)^{-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-(m-2-\iota)}2}\,ds\\
& \lesssim (1+t)^{-\frac{n}{2q}\,(p-1)-\frac{\nu-(m-2-\iota)}2\,p+1-\frac{n}2\left(\frac1q-\frac12\right)-\frac{k+s-(m-2-\iota)}2}\,\log(1+t)\\
& \leq (1+t)^{-\frac{n}4-\frac{k+s-(m-2-\iota)}2}\,.
\end{align*}
Since we proved~\eqref{eq:Fest}, this concludes the proof.
\end{proof}
We may prove the nonexistence of nontrivial global-in-time solutions to~\eqref{eq:CPnl} when~$f=|u|^p$ if $M\geq0$ and $f=-|u|^p$ if $M\leq0$, for any~$p\in(1,\bar p(n)]$, if we add the assumption that $P_{m-1-\iota}(0,\xi)=0$. Indeed, we may prove that even $L^p_\lloc$ solutions do not exist globally-in-time. We say that $u\in L^p_\lloc([0,\infty)\times\R^n)$ is a global-in-time weak solution to~\eqref{eq:CPnl} with~$f=\pm |u|^p$ and $u_j\in L^1_\lloc(\R^n)$, $j=0,\ldots,m-1$, if
for any $\psi\in\mathcal C_c^\infty([0,\infty))$, with $\psi(t)=1$ in a neighborhood of the origin, and for any $\varphi\in\mathcal C_c^\infty(\R^n)$, the following integral equality holds:
\begin{equation}\label{eq:weakblow}\begin{split}
\pm \int_0^\infty \int_{\R^n} |u(t,x)|^p\,\psi(t)\,\varphi(x)\,dxdt
& = \int_0^\infty\int_{\R^n} u(t,x)\,Q(-\partial_t,-\partial_x)\big(\psi(t)\,\varphi(x)\big)\,dxdt \\
& \qquad - \sum_{k=1}^{m} \sum_{k+|\alpha|=m-1-\iota}^m c_{k,\alpha} \int_{\R^n} u_{k-1}(x)\,(-\partial_x)^\alpha\varphi(x)\,dx\,.
\end{split}\end{equation}
Integrating by parts, it is easy to prove that smooth solutions are weak solutions.
\begin{proposition}\label{prop:critical}
Let $Q=Q_1$ or $Q=Q_2$, as in Theorem~\ref{thm:CPQ2}, and fix $\iota=0$ if $Q=Q_1$ or $\iota=1$ if $Q=Q_2$. Assume that $P_{m-1-\iota}(0,\xi)=0$. Assume that $u_j\in L^1(\R^n)$ for any $j=0,\ldots,m-1$.
Let~$n\leq m-2-\iota$ and $p>1$, or~$1<p\leq \bar p(n)$, where $\bar p(n)=1 + (m-\iota)/(n-(m-2-\iota))$ as in~\eqref{eq:pcrit}. Then there is no global-in-time nontrivial weak solution to~\eqref{eq:CPnl}, in the sense of~\eqref{eq:weakblow}, with $f=|u|^p$ if $M\geq0$ and with $f=-|u|^p$, if $M\leq0$, where $M$ is as in~\eqref{eq:moment}.
\end{proposition}
With the notation in~\cite{DL03}, the assumption $P_{m-1-\iota}(0,\xi)=0$ guarantees that the operator $Q(\partial_t,\partial_x)$ is a sum of quasi-homogeneous operators of type $(h_i,2,1)$, with $\min h_i=m-\iota$, and the proof essentially follows by~\cite[Theorem 4.2]{DL03}, thanks to the assumption on the moment condition. However, for the ease of reading, we provide a self-contained proof of Proposition~\ref{prop:critical}.
\begin{proof}
Assume, by contradiction, that $u$ is a global-in-time nontrivial weak solution. Let $\Psi\in\mathcal C_c^\infty([0,\infty))$, nonincreasing, with $\psi(t)=1$ in~$[0,1]$. For a given integer $\ell\geq mp'$, where $p'=p/(p-1)$, we set $\psi(t)=(\Psi(t))^\ell$ and $\varphi(x)=\psi(|x|)$, then we define $\psi_R(t)=\psi(R^{-2}t)$ and $\varphi_R(x)=\varphi(R^{-1}x)$ for $R\gg1$. If we put
\begin{align*}
I_R
& =\int_0^\infty \int_{\R^n} |u(t,x)|^p\,\psi_R(t)\,\varphi_R(x)\,dxdt,\\
I_R^*
& =\int_{\Omega_R^*} |u(t,x)|^p\,\psi_R(t)\,\varphi_R(x)\,dxdt,
\end{align*}
where $\Omega_R^*=[0,\infty)\times\R^n\setminus \big([0,R]\times B_R\big)$, by H\"older and Young inequality, we derive
\begin{align*}
I_R
& \leq \frac12 I_R^* + C\,\int_0^\infty\int_{\R^n} |Q(-\partial_t,-\partial_x)\big(\psi_R(t)\,\varphi_R(x)\big)|^{p'}\,\big(\psi_R(t)\,\varphi_R(x)\big)^{-\frac{p'}{p}}\,dxdt \\
& \qquad \mp \sum_{k=1}^{m} \sum_{k+|\alpha|=m-1-\iota}^m c_{k,\alpha} \int_{\R^n} u_{k-1}(x)\,(-\partial_x)^\alpha\varphi_R(x)\,dx\,.
\end{align*}
We stress that the assumption $\ell\geq mp'$ is here used to obtain that
\[ |Q(-\partial_t,-\partial_x)\big(\psi(t)\,\varphi(x)\big)|^p \leq C\,\psi(t)\,\varphi(x), \]
whereas the fact that $I_R^*$ appears, instead of $I_R$, after H\"older and Young inequality, is related to the fact that $Q(-\partial_t,-\partial_x)\big(\psi(t)\,\varphi(x)\big)=0$ in $[0,R]\times B_R$.
Thanks to the homogeneity of the monomials in $Q(\partial_t,\partial_x)$, we easily get:
\begin{align*}
& \int_0^\infty\int_{\R^n} |Q(-\partial_t,-\partial_x)\big(\psi_R(t)\,\varphi_R(x)\big)|^{p'}\,\big(\psi_R(t)\,\varphi_R(x)\big)^{-\frac{p'}{p}}\,dxdt\\
& \qquad \leq C\,R^{-(m-\iota)p'+n+2} \leq C,
\end{align*}
with $C$ independent on~$R$, thanks to the assumption $p\leq \bar p(n)$, that is, $(m-\iota)p'\leq n+2$. On the other hand,
\[ \lim_{R\to\infty} \sum_{k=1}^{m} \sum_{k+|\alpha|=m-1-\iota}^m c_{k,\alpha} \int_{\R^n} u_{k-1}(x)\,(-\partial_x)^\alpha\varphi_R(x)\,dx = M, \]
by Lebesgue dominated convergence theorem, since $\varphi_R(x)\to1$ and $(-\partial_x)^\alpha\varphi_R(x)\to0$ when $\alpha\neq0$, for any $x\in\R^n$, as $R\to\infty$. As a consequence, $I_R$ is bounded, so that, by Beppo-Levi monotone convergence theorem,
\[ \int_0^\infty \int_{\R^n} |u(t,x)|^p\,dxdt = \lim_{R\to\infty} I_R\leq C, \]
that is, $u\in L^p([0,\infty)\times \R^n)$. As a consequence, by Lebesgue dominated convergence theorem, $I_R^*\to0$ as $R\to\infty$, so that we obtain the inequality
\[ \int_0^\infty \int_{\R^n} |u(t,x)|^p\,dxdt \leq \mp M. \]
The contradiction follows, since $\mp M\leq0$, so that $u\equiv0$.
\end{proof}
\begin{remark}
The results in this section may be generalized in several ways. When $p<\bar p(n)$, lifespan estimates for the solution to~\eqref{eq:CPnl} may be easily obtained. By assuming small initial data $u_j \in L^q\cap H^{m-1-j}$, for some $q\in(1,2]$, the critical exponent $\bar p(n)$ is replaced by $\bar p(n/q)$; moreover, in this case, global small data solutions exist for critical and supercritical powers $p\geq\bar p(n/q)$, and the nonexistence result in Proposition~\ref{prop:critical} may be extended to all subcritical powers~$1<p<\bar p(n/q)$, under a suitable sign assumption on the initial data.
We omit the details for brevity, addressing the reader to~\cite{DAE17NA} and the reference therein. Possibly, critical nonlinearities may be investigated as done in~\cite{EGR} for the damped wave equation.
\end{remark}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.